From 6eddde06a4f25d55d538b5d15628dcc2b6882147 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Mon, 13 Jul 2026 18:37:57 +0200 Subject: [PATCH] CUDA: refactor MMQ kernel configuration (#24127) * CUDA: refactor MMQ kernel configuration * fix Blackwell config * remove legacy code --- ggml/src/ggml-cuda/mmq-config-ampere.cuh | 366 ++ ggml/src/ggml-cuda/mmq-config-blackwell.cuh | 37 + ggml/src/ggml-cuda/mmq-config-cdna.cuh | 177 + ggml/src/ggml-cuda/mmq-config-pascal.cuh | 261 ++ ggml/src/ggml-cuda/mmq-config-rdna2.cuh | 261 ++ ggml/src/ggml-cuda/mmq-config-rdna4.cuh | 282 ++ ggml/src/ggml-cuda/mmq-load-tiles.cuh | 1671 +++++++ ggml/src/ggml-cuda/mmq-vec-dot.cuh | 1251 ++++++ ggml/src/ggml-cuda/mmq.cu | 55 +- ggml/src/ggml-cuda/mmq.cuh | 4300 ++++--------------- 10 files changed, 5121 insertions(+), 3540 deletions(-) create mode 100644 ggml/src/ggml-cuda/mmq-config-ampere.cuh create mode 100644 ggml/src/ggml-cuda/mmq-config-blackwell.cuh create mode 100644 ggml/src/ggml-cuda/mmq-config-cdna.cuh create mode 100644 ggml/src/ggml-cuda/mmq-config-pascal.cuh create mode 100644 ggml/src/ggml-cuda/mmq-config-rdna2.cuh create mode 100644 ggml/src/ggml-cuda/mmq-config-rdna4.cuh create mode 100644 ggml/src/ggml-cuda/mmq-load-tiles.cuh create mode 100644 ggml/src/ggml-cuda/mmq-vec-dot.cuh diff --git a/ggml/src/ggml-cuda/mmq-config-ampere.cuh b/ggml/src/ggml-cuda/mmq-config-ampere.cuh new file mode 100644 index 0000000000..0037bac3d0 --- /dev/null +++ b/ggml/src/ggml-cuda/mmq-config-ampere.cuh @@ -0,0 +1,366 @@ +static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_ampere(ggml_type type, int J, bool fallback) { + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + + return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true); +} diff --git a/ggml/src/ggml-cuda/mmq-config-blackwell.cuh b/ggml/src/ggml-cuda/mmq-config-blackwell.cuh new file mode 100644 index 0000000000..9fbe32b697 --- /dev/null +++ b/ggml/src/ggml-cuda/mmq-config-blackwell.cuh @@ -0,0 +1,37 @@ +static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_blackwell(ggml_type type, int J, bool fallback) { + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_MXFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + CASE(GGML_TYPE_NVFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, false); + + return ggml_cuda_mmq_get_config_ampere(type, J, fallback); +} diff --git a/ggml/src/ggml-cuda/mmq-config-cdna.cuh b/ggml/src/ggml-cuda/mmq-config-cdna.cuh new file mode 100644 index 0000000000..46ec6aa9d5 --- /dev/null +++ b/ggml/src/ggml-cuda/mmq-config-cdna.cuh @@ -0,0 +1,177 @@ +static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_cdna(ggml_type type, int J, bool fallback) { + CASE(GGML_TYPE_Q1_0, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q1_0, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q1_0, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q1_0, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q1_0, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q4_0, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_0, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_0, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_0, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_0, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q4_1, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_1, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_1, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_1, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_1, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q5_0, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_0, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_0, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_0, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_0, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q5_1, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_1, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_1, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_1, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_1, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q8_0, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q8_0, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q8_0, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q8_0, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q8_0, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_Q2_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q2_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q2_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q2_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q2_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q3_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q3_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q3_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q3_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q3_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q4_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q4_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q4_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q5_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q5_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q5_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_Q6_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q6_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q6_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_Q6_K, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_Q6_K, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_IQ1_S, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ1_S, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ1_S, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ1_S, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ1_S, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ2_XS, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XS, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XS, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_XS, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_XS, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ2_S, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_S, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_S, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ2_S, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ2_S, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ3_S, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_S, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_S, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ3_S, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ3_S, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ4_XS, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_XS, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_XS, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_XS, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_XS, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_IQ4_NL, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_NL, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_NL, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_IQ4_NL, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_IQ4_NL, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_MXFP4, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_MXFP4, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_MXFP4, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_MXFP4, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_MXFP4, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false); + + CASE(GGML_TYPE_NVFP4, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_NVFP4, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_NVFP4, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true); + CASE(GGML_TYPE_NVFP4, 512, 1, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 512, 1, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 512, 1, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + CASE(GGML_TYPE_NVFP4, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false); + + return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true); +} diff --git a/ggml/src/ggml-cuda/mmq-config-pascal.cuh b/ggml/src/ggml-cuda/mmq-config-pascal.cuh new file mode 100644 index 0000000000..8f0faac889 --- /dev/null +++ b/ggml/src/ggml-cuda/mmq-config-pascal.cuh @@ -0,0 +1,261 @@ +static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_pascal(ggml_type type, int J, bool fallback) { + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + + return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true); +} diff --git a/ggml/src/ggml-cuda/mmq-config-rdna2.cuh b/ggml/src/ggml-cuda/mmq-config-rdna2.cuh new file mode 100644 index 0000000000..de4db0a3db --- /dev/null +++ b/ggml/src/ggml-cuda/mmq-config-rdna2.cuh @@ -0,0 +1,261 @@ +static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_rdna2(ggml_type type, int J, bool fallback) { + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + + return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true); +} diff --git a/ggml/src/ggml-cuda/mmq-config-rdna4.cuh b/ggml/src/ggml-cuda/mmq-config-rdna4.cuh new file mode 100644 index 0000000000..6280e80ee4 --- /dev/null +++ b/ggml/src/ggml-cuda/mmq-config-rdna4.cuh @@ -0,0 +1,282 @@ +static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_rdna4(ggml_type type, int J, bool fallback) { + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q1_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q8_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q2_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q3_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q4_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q5_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_Q6_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false); + +// --------------------------------------------------------------------------------------------- + + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_MXFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false); + + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + CASE(GGML_TYPE_NVFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false); + + return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true); +} diff --git a/ggml/src/ggml-cuda/mmq-load-tiles.cuh b/ggml/src/ggml-cuda/mmq-load-tiles.cuh new file mode 100644 index 0000000000..3978b1baa7 --- /dev/null +++ b/ggml/src/ggml-cuda/mmq-load-tiles.cuh @@ -0,0 +1,1671 @@ +#pragma once + +#include "vecdotq.cuh" + +#include "mmq.cuh" + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q1_0( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int blocks_per_iter = MMQ_ITER_K / QK1_0; + constexpr int threads_per_row = blocks_per_iter * QI1_0; + constexpr int nrows = warp_size / threads_per_row; + constexpr int scale_entries_per_block = QK1_0 / QK8_1; + constexpr int scale_entries_per_row = blocks_per_iter * scale_entries_per_block; + + const int txi = threadIdx.x % threads_per_row; + const int kbx = txi / QI1_0; + const int kqsx = txi % QI1_0; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_q1_0 * bxi = (const block_q1_0 *) x + kbx0 + i*stride + kbx; + const int qs_offset = 4*kqsx; + const int qs0 = bxi->qs[qs_offset + 0] | (bxi->qs[qs_offset + 1] << 8) | + (bxi->qs[qs_offset + 2] << 16) | (bxi->qs[qs_offset + 3] << 24); + + int unpacked_bytes[8]; +#pragma unroll + for (int j = 0; j < 8; ++j) { + const int shift = j * 4; + const int bits4 = (qs0 >> shift) & 0x0F; + const int b0 = (bits4 & 0x01) ? 1 : -1; + const int b1 = (bits4 & 0x02) ? 1 : -1; + const int b2 = (bits4 & 0x04) ? 1 : -1; + const int b3 = (bits4 & 0x08) ? 1 : -1; + unpacked_bytes[j] = (b0 & 0xFF) | ((b1 & 0xFF) << 8) | ((b2 & 0xFF) << 16) | ((b3 & 0xFF) << 24); + } + + const int dst_offset = kbx*(scale_entries_per_block*QI8_0) + kqsx*QI8_0; +#pragma unroll + for (int j = 0; j < 8; ++j) { +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + dst_offset + j] = unpacked_bytes[j]; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j] = unpacked_bytes[j]; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + } + + const int ksx = threadIdx.x % scale_entries_per_row; + const int scale_block = ksx / scale_entries_per_block; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps) { + int i = i0 + threadIdx.y; + + if (fallback) { + i = min(i, i_max); + } + + const block_q1_0 * bxi = (const block_q1_0 *) x + kbx0 + i*stride + scale_block; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + ksx] = bxi->d; +#else + x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + ksx] = bxi->d; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q4_0( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_0); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + const int kbx = txi / QI4_0; + const int kqsx = txi % QI4_0; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbx; + const int qs0 = get_int_b2(bxi->qs, kqsx); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + kbx*(2*QI4_0) + kqsx + 0] = __vsubss4((qs0 >> 0) & 0x0F0F0F0F, 0x08080808); + x_qs[i*sram_stride + kbx*(2*QI4_0) + kqsx + QI4_0] = __vsubss4((qs0 >> 4) & 0x0F0F0F0F, 0x08080808); +#else + x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) + } + + constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_0; + constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; + const int kbxd = threadIdx.x % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) { + int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbxd; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + kbxd] = bxi->d; +#else + x_df[i*(MMQ_TILE_NE_K/QI4_0) + i/QI4_0 + kbxd] = bxi->d; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q4_1( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, I); + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_1); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + const int kbx = txi / QI4_1; + const int kqsx = txi % QI4_1; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbx; + const int qs0 = get_int_b4(bxi->qs, kqsx); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + kbx*(2*QI4_1) + kqsx + 0] = (qs0 >> 0) & 0x0F0F0F0F; + x_qs[i*sram_stride + kbx*(2*QI4_1) + kqsx + QI4_1] = (qs0 >> 4) & 0x0F0F0F0F; +#else + x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_1; + constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; + const int kbxd = threadIdx.x % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) { + int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbxd; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_dm[i*sram_stride + kbxd] = bxi->dm; +#else + x_dm[i*(MMQ_TILE_NE_K/QI4_1) + i/QI4_1 + kbxd] = bxi->dm; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q5_0( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_0, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_0); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + const int kbx = txi / QI5_0; + const int kqsx = txi % QI5_0; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbx; + + const int ql = get_int_b2(bxi->qs, kqsx); + const int qh = get_int_b2(bxi->qh, 0) >> (4 * kqsx); + + int qs0 = (ql >> 0) & 0x0F0F0F0F; + qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 + qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 + qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 + qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 + qs0 = __vsubss4(qs0, 0x10101010); // subtract 16 + + int qs1 = (ql >> 4) & 0x0F0F0F0F; + qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 + qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 + qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 + qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 + qs1 = __vsubss4(qs1, 0x10101010); // subtract 16 + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + kbx*(2*QI5_0) + kqsx + 0] = qs0; + x_qs[i*sram_stride + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_0) + kqsx + 0] = qs0; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI5_0; + constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; + const int kbxd = threadIdx.x % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) { + int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbxd; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + kbxd] = bxi->d; +#else + x_df[i*(MMQ_TILE_NE_K/QI5_0) + i/QI5_0 + kbxd] = bxi->d; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q5_1( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, I); + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_1); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + const int kbx = txi / QI5_1; + const int kqsx = txi % QI5_1; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbx; + + const int ql = get_int_b4(bxi->qs, kqsx); + const int qh = get_int_b4(bxi->qh, 0) >> (4 * kqsx); + + int qs0 = (ql >> 0) & 0x0F0F0F0F; + qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 + qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 + qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 + qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 + + int qs1 = (ql >> 4) & 0x0F0F0F0F; + qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 + qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 + qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 + qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + kbx*(2*QI5_1) + kqsx + 0] = qs0; + x_qs[i*sram_stride + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_1) + kqsx + 0] = qs0; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI5_1; + constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; + const int kbxd = threadIdx.x % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) { + int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbxd; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_dm[i*sram_stride + kbxd] = bxi->dm; +#else + x_dm[i*(MMQ_TILE_NE_K/QI5_1) + i/QI5_1 + kbxd] = bxi->dm; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q8_0( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_tile + 2*MMQ_TILE_NE_K); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + // MMQ_ITER_K / (4 * QR8_0) == 64 required. but NV has only 32 threads per warp + constexpr int threads_per_row = 32; + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + const int kbx = txi / QI8_0; + const int kqsx = txi % QI8_0; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbx; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + 0 + txi] = get_int_b2(bxi[0].qs, kqsx); + x_qs[i*sram_stride + MMQ_TILE_NE_K + txi] = get_int_b2(bxi[MMQ_TILE_NE_K/QI8_0].qs, kqsx); +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 0 + txi] = get_int_b2(bxi[0].qs, kqsx); + x_qs[i*(2*MMQ_TILE_NE_K + 1) + MMQ_TILE_NE_K + txi] = get_int_b2(bxi[MMQ_TILE_NE_K/QI8_0].qs, kqsx); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + constexpr int blocks_per_tile_x_row = 2*MMQ_TILE_NE_K / QI8_0; + constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; + const int kbxd = threadIdx.x % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) { + int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbxd; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + kbxd] = bxi->d; +#else + x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + kbxd] = bxi->d; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +// --------------------------------------------------------------------------------------------- + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q2_K( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, I); + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR2_K); + constexpr int nrows = ggml_cuda_get_physical_warp_size() / threads_per_row; + const int kqsx = threadIdx.x % threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_q2_K * bxi = (const block_q2_K *) x + kbx0 + i*stride; + + const int x_ql_0 = get_int_b2(bxi->qs, kqsx); + +#pragma unroll + for (int l = 0; l < QR2_K; ++l) { + const int k = (kqsx/8)*32 + l*8 + kqsx % 8; + + const int x_qs_k = (x_ql_0 >> (2*l)) & 0x03030303; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + k] = x_qs_k; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + k] = x_qs_k; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + const int sc_m = bxi->scales[kqsx]; +#ifdef FAST_FP16_AVAILABLE + const half2 x_dm_ik = __hmul2(bxi->dm, make_half2(sc_m & 0x0F, sc_m >> 4)); +#else + const float2 bxi_dmf = __half22float2(bxi->dm); + const half2 x_dm_ik = make_half2(bxi_dmf.x*(sc_m & 0x0F), bxi_dmf.y*(sc_m >> 4)); +#endif // FAST_FP16_AVAILABLE + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_dm[i*sram_stride + kqsx] = x_dm_ik; +#else + x_dm[i*(MMQ_TILE_NE_K + 1) + kqsx] = x_dm_ik; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q3_K( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); + int * x_sc = (int *) (x_df + txs.dm); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR3_K); + constexpr int nrows = warp_size / threads_per_row; + const int kqsx = threadIdx.x % threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride; + + const int x_ql_0 = get_int_b2(bxi->qs, kqsx); + const int x_qh_0 = get_int_b2(bxi->hmask, kqsx % (QI3_K/2)) >> (4 * (kqsx / (QI3_K/2))); + +#pragma unroll + for (int l = 0; l < QR3_K; ++l) { + const int k = (kqsx/8)*32 + l*8 + kqsx % 8; + + const int x_ql_k = (x_ql_0 >> (2*l)) & 0x03030303; + const int x_qh_k = ((x_qh_0 >> l) << 2) & 0x04040404; + + const int x_qs_k = __vsubss4(x_ql_k | x_qh_k, 0x04040404); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + k] = x_qs_k; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + k] = x_qs_k; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + } + + constexpr int rows_per_warp = warp_size / 4; +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) { + int i = i0 + threadIdx.y*rows_per_warp + threadIdx.x/4; + + if (fallback) { + i = min(i, i_max); + } + + const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride; + + const int ksc = threadIdx.x % 4; + + const int ksc_low = ksc % (QI3_K/8); + const int shift_low = 4 * (ksc / (QI3_K/8)); + const int sc_low = (get_int_b2(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F; + + const int ksc_high = QI3_K/8; + const int shift_high = 2 * ksc; + const int sc_high = ((get_int_b2(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030; + + const int sc = __vsubss4(sc_low | sc_high, 0x20202020); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + const int8_t * sc8 = (const int8_t *) ≻ + const float d = bxi->d; + +#pragma unroll + for (int l = 0; l < int(sizeof(int)); ++l) { + x_df[i*sram_stride + sizeof(int)*ksc + l] = d*sc8[l]; + } +#else + x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = sc; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + +#if !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)) +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*warp_size) { + int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % I; + + if (fallback) { + i = min(i, i_max); + } + + const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride; + + x_df[i] = bxi->d; + } +#endif // !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)) || defined(AMD_WMMA_AVAILABLE) +} + +static __device__ __forceinline__ int unpack_scales_q45_K(const int * scales, const int ksc) { + // scale arrangement after the following two lines: + // - ksc == 0: sc0, sc1, sc2, sc3 + // - ksc == 1: sc4, sc5, sc6, sc7 + // - ksc == 2: m0, m1, m2, m3 + // - ksc == 3: m4, m5, m6, m7 + return ((scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F) | // lower 4 bits + ((scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030); // upper 2 bits +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q4_K( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, I); + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + txs.qs); + int * x_sc = (int *) (x_dm + txs.dm); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_K); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride; + const int qs0 = get_int_b4(bxi->qs, txi); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + 16*(txi/8) + txi % 8 + 0] = (qs0 >> 0) & 0x0F0F0F0F; + x_qs[i*sram_stride + 16*(txi/8) + txi % 8 + 8] = (qs0 >> 4) & 0x0F0F0F0F; +#else + x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + constexpr int rows_per_warp = warp_size / 2; +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) { +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + // Need if on AMD instead of % because warp_size == 64 + // This causes double work and throughput loss (MI300X) + // H100 loses about 100 t/s with 'if' condition over '%' + int i = i0 + threadIdx.y*rows_per_warp + threadIdx.x/2; + if (i < I) { +#else + int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/2) % I; + { +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + if (fallback) { + i = min(i, i_max); + } + + const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride; + + const int * scales = (const int *) bxi->scales; + const int ksc = threadIdx.x % 2; + + const int sc32 = unpack_scales_q45_K(scales, ksc + 0); + const int m32 = unpack_scales_q45_K(scales, ksc + 2); + + const uint8_t * sc8 = (const uint8_t *) &sc32; + const uint8_t * m8 = (const uint8_t *) &m32; + + const half2 dm = bxi->dm * make_half2(1.0f, -1.0f); + + #pragma unroll + for (int l = 0; l < sizeof(int); ++l) { + x_dm[i*sram_stride + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]); + } + } + } +#else +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*warp_size) { + int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % I; + + if (fallback) { + i = min(i, i_max); + } + + const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride; + + x_dm[i] = bxi->dm; + } + constexpr int rows_per_warp = warp_size / 4; +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) { + int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/(MMQ_TILE_NE_K/8)) % I; + + if (fallback) { + i = min(i, i_max); + } + + const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride + (threadIdx.x % (MMQ_TILE_NE_K/8)) / (QI4_K/8); + + const int * scales = (const int *) bxi->scales; + + const int ksc = threadIdx.x % (MMQ_TILE_NE_K/8); + const int scales8 = unpack_scales_q45_K(scales, ksc); + + x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = scales8; + } +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q5_K( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, I); + int * x_qs = (int *) x_tile; + half2 * x_dm = (half2 *) (x_qs + txs.qs); + int * x_sc = (int *) (x_dm + txs.dm); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_K); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride; + const int ky = QR5_K*txi; + + const int ql = get_int_b4(bxi->qs, txi); + const int ql0 = (ql >> 0) & 0x0F0F0F0F; + const int ql1 = (ql >> 4) & 0x0F0F0F0F; + + const int qh = get_int_b4(bxi->qh, txi % (QI5_K/4)); + const int qh0 = ((qh >> (2 * (txi / (QI5_K/4)) + 0)) << 4) & 0x10101010; + const int qh1 = ((qh >> (2 * (txi / (QI5_K/4)) + 1)) << 4) & 0x10101010; + + const int kq0 = ky - ky % (QI5_K/2) + txi % (QI5_K/4) + 0; + const int kq1 = ky - ky % (QI5_K/2) + txi % (QI5_K/4) + QI5_K/4; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + kq0] = ql0 | qh0; + x_qs[i*sram_stride + kq1] = ql1 | qh1; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq0] = ql0 | qh0; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq1] = ql1 | qh1; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + constexpr int rows_per_warp = warp_size / 2; +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) { +#if defined(AMD_MFMA_AVAILABLE) + // Need if on AMD instead of % because warp_size == 64 + // This causes double work and throughput loss (MI300X) + // H100 loses about 100 t/s with 'if' condition over '%' + int i = i0 + threadIdx.y*rows_per_warp + threadIdx.x/2; + if (i < I) { +#else + int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/2) % I; + { +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + if (fallback) { + i = min(i, i_max); + } + + const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride; + + const int * scales = (const int *) bxi->scales; + const int ksc = threadIdx.x % 2; + + const int sc32 = unpack_scales_q45_K(scales, ksc + 0); + const int m32 = unpack_scales_q45_K(scales, ksc + 2); + + const uint8_t * sc8 = (const uint8_t *) &sc32; + const uint8_t * m8 = (const uint8_t *) &m32; + + const half2 dm = bxi->dm * make_half2(1.0f, -1.0f); + +#pragma unroll + for (int l = 0; l < int(sizeof(int)); ++l) { + x_dm[i*sram_stride + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]); + } + } + } +#else +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*warp_size) { + int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % I; + + if (fallback) { + i = min(i, i_max); + } + + const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride; + + x_dm[i] = bxi->dm; + } + + constexpr int rows_per_warp = warp_size / 4; +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) { + int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/(MMQ_TILE_NE_K/8)) % I; + + if (fallback) { + i = min(i, i_max); + } + + const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride; + + const int * scales = (const int *) bxi->scales; + + const int ksc = threadIdx.x % (MMQ_TILE_NE_K/8); + const int scales8 = unpack_scales_q45_K(scales, ksc); + + x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = scales8; + } +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q6_K( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); + int * x_sc = (int *) (x_df + MMQ_TILE_NE_K/QI6_K); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); + int * x_sc = (int *) (x_df + txs.dm); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR6_K); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride; + + const int ql = get_int_b2(bxi->ql, txi); + const int ql0 = (ql >> 0) & 0x0F0F0F0F; + const int ql1 = (ql >> 4) & 0x0F0F0F0F; + + const int qh = get_int_b2(bxi->qh, (QI6_K/4) * (txi / (QI6_K/2)) + txi % (QI6_K/4)); + const int qh0 = ((qh >> ((txi & 0x08) >> 2)) << 4) & 0x30303030; + const int qh1 = (qh >> ((txi & 0x08) >> 2)) & 0x30303030; + + const int kq0 = 2*txi - txi % (QI6_K/2) + 0; + const int kq1 = 2*txi - txi % (QI6_K/2) + QI6_K/2; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + kq0] = __vsubss4(ql0 | qh0, 0x20202020); + x_qs[i*sram_stride + kq1] = __vsubss4(ql1 | qh1, 0x20202020); +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq0] = __vsubss4(ql0 | qh0, 0x20202020); + x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq1] = __vsubss4(ql1 | qh1, 0x20202020); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*warp_size) { + int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % I; + + if (fallback) { + i = min(i, i_max); + } + + const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride] = bxi->d; +#else + x_df[i*(MMQ_TILE_NE_K/QI6_K) + i/QI6_K] = bxi->d; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + constexpr int rows_per_warp = warp_size / 4; +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) { + int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/(MMQ_TILE_NE_K/8)) % I; + + if (fallback) { + i = min(i, i_max); + } + + const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + (threadIdx.x % (MMQ_TILE_NE_K/8)) / 4; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_sc[i*sram_stride + threadIdx.x%4] = get_int_b2(bxi->scales, threadIdx.x % (MMQ_TILE_NE_K/8)); +#else + x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + threadIdx.x%(MMQ_TILE_NE_K/8)] = get_int_b2(bxi->scales, threadIdx.x%(QI6_K/8)); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +// --------------------------------------------------------------------------------------------- + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq1_s( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + half2 * x_ds = (half2 *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, I); + int * x_qs = (int *) x_tile; + half2 * x_ds = (half2 *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR1_S); + constexpr int nrows = warp_size / threads_per_row; + const int kqsx = threadIdx.x % threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * nrows) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_iq1_s * bxi = (const block_iq1_s *) x + kbx0 + i*stride; + + const int qs_packed = get_int_b2(bxi->qs, kqsx); + const uint8_t * qs = (const uint8_t *) &qs_packed; + + const int qh = bxi->qh[kqsx]; + + #pragma unroll + for (int l = 0; l < QR1_S/2; ++l) { + const int grid = iq1s_grid_gpu[qs[l] | (((qh >> (3*l)) & 0x07) << 8)]; + + const int grid0 = (grid >> 0) & 0x0F0F0F0F; + const int grid1 = (grid >> 4) & 0x0F0F0F0F; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + 8*kqsx + (2*l+0)] = grid0; + x_qs[i*sram_stride + 8*kqsx + (2*l+1)] = grid1; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+0)] = grid0; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+1)] = grid1; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + const float d1q = __half2float(bxi->d) * (((qh >> 11) & 0x0E) + 1); + const float delta = -1.0f + IQ1S_DELTA - (qh & 0x8000) * (2.0f*IQ1S_DELTA/0x8000); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_ds[i*sram_stride + kqsx] = make_half2(d1q, d1q*delta); +#else + x_ds[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = make_half2(d1q, d1q*delta); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq2_xxs( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_XXS, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_XXS)) / 2; + constexpr int nrows = warp_size / threads_per_row; + const int kqsx = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * nrows) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_iq2_xxs * bxi = (const block_iq2_xxs *) x + kbx0 + i*stride; + + const int q2 = get_int_b2(bxi->qs, 2*kqsx+0); + const uint8_t * aux8 = (const uint8_t *) &q2; + const uint32_t aux32 = get_int_b2(bxi->qs, 2*kqsx+1); + +#pragma unroll + for (int l = 0; l < QR2_XXS; ++l) { + const uint2 grid_pos = ((const uint2*)iq2xxs_grid)[aux8[l]]; + const uint32_t signs = unpack_ksigns(aux32 >> (7 * l)); + + const int signs0 = __vcmpne4(signs & 0x08040201, 0); + const int grid0 = __vsub4(grid_pos.x ^ signs0, signs0); + + const int signs1 = __vcmpne4(signs & 0x80402010, 0); + const int grid1 = __vsub4(grid_pos.y ^ signs1, signs1); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + 8*kqsx + (2*l + 0)] = grid0; + x_qs[i*sram_stride + 8*kqsx + (2*l + 1)] = grid1; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid0; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid1; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + const int ls = aux32 >> 27 | 1; // (scale * 2 + 1) + const float d = bxi->d; +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + kqsx] = d * ls / 8; // (d * scale + d / 2) / 4 +#else + x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = d * ls / 8; // (d * scale + d / 2) / 4 +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq2_xs( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_XS, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_XS)) / 2; + constexpr int nrows = warp_size / threads_per_row; + const int kqsx = threadIdx.x % threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * nrows) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_iq2_xs * bxi = (const block_iq2_xs *) x + kbx0 + i*stride; + + const int2 q2_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1)); + const uint16_t * q2 = (const uint16_t *) &q2_packed; + + #pragma unroll + for (int l = 0; l < QR2_XS; ++l) { + const uint2 grid_pos = ((const uint2*)iq2xs_grid)[q2[l] & 0x1FF]; + const uint32_t signs = unpack_ksigns(q2[l] >> 9); + + const int signs0 = __vcmpne4(signs & 0x08040201, 0); + const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0); + + const int signs1 = __vcmpne4(signs & 0x80402010, 0); + const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + 8*kqsx + (2*l + 0)] = grid_l; + x_qs[i*sram_stride + 8*kqsx + (2*l + 1)] = grid_h; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + const int ls = bxi->scales[kqsx]; + const float d = bxi->d; +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; + x_df[i*sram_stride + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; +#else + x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; + x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq2_s( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_S, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_S)) / 2; + constexpr int nrows = warp_size / threads_per_row; + const int kqsx = threadIdx.x % threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * nrows) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_iq2_s * bxi = (const block_iq2_s *) x + kbx0 + i*stride; + + const int qs_packed = get_int_b2(bxi->qs, kqsx); + const uint8_t * qs = (const uint8_t *) &qs_packed; + + const int qh = bxi->qh[kqsx]; + + const int signs_packed_32 = get_int_b2(bxi->qs, QK_K/32 + kqsx); + const uint8_t * signs_packed_8 = (const uint8_t *) &signs_packed_32; + +#pragma unroll + for (int l = 0; l < QR2_S; ++l) { + const int * grid_pos = (const int *)(iq2s_grid + (qs[l] | ((qh << (8-2*l)) & 0x300))); + + const int signs0 = __vcmpne4(((signs_packed_8[l] & 0x03) << 7) | ((signs_packed_8[l] & 0x0C) << 21), 0x00000000); + const int signs1 = __vcmpne4(((signs_packed_8[l] & 0x30) << 3) | ((signs_packed_8[l] & 0xC0) << 17), 0x00000000); + + const int grid_l = __vsub4(grid_pos[0] ^ signs0, signs0); + const int grid_h = __vsub4(grid_pos[1] ^ signs1, signs1); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + 8*kqsx + (2*l + 0)] = grid_l; + x_qs[i*sram_stride + 8*kqsx + (2*l + 1)] = grid_h; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + const int ls = bxi->scales[kqsx]; + const float d = bxi->d; +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; + x_df[i*sram_stride + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; +#else + x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; + x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq3_xxs( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_XXS, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR3_XXS)) / 2; + constexpr int nrows = warp_size / threads_per_row; + const int kqsx = threadIdx.x % threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * nrows) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_iq3_xxs * bxi = (const block_iq3_xxs *) x + kbx0 + i*stride; + + const int2 q3_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1)); + const uint8_t * q3 = (const uint8_t *) &q3_packed; + const uint32_t aux32 = get_int_b2(bxi->qs, QK_K/16 + kqsx); + +#pragma unroll + for (int l = 0; l < QR3_XXS; ++l) { + const int2 grid_pos = make_int2(iq3xxs_grid[q3[2*l+0]], iq3xxs_grid[q3[2*l+1]]); + const uint32_t signs = unpack_ksigns(aux32 >> (7*l)); + + const int signs0 = __vcmpne4(signs & 0x08040201, 0); + const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0); + + const int signs1 = __vcmpne4(signs & 0x80402010, 0); + const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + 8*kqsx + (2*l + 0)] = grid_l; + x_qs[i*sram_stride + 8*kqsx + (2*l + 1)] = grid_h; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + const int ls = aux32 >> 28; + const float d = bxi->d; +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + kqsx] = (ls*d + d/2)/2; +#else + x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = (ls*d + d/2)/2; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq3_s( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR3_S)) / 2; + constexpr int nrows = warp_size / threads_per_row; + const int kqsx = threadIdx.x % threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * nrows) { + int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_iq3_s * bxi = (const block_iq3_s *) x + kbx0 + i*stride; + + const int2 qs_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1)); + const uint8_t * qs = (const uint8_t *) &qs_packed; + + const int qh = bxi->qh[kqsx]; + + const int signs_packed_32 = get_int_b2(bxi->signs, kqsx); + const uint8_t * signs_packed_8 = (const uint8_t *) &signs_packed_32; + +#pragma unroll + for (int l = 0; l < QR3_S; ++l) { + const int2 grid_pos = make_int2( + iq3s_grid[qs[2*l+0] | ((qh << (8 - 2*l)) & 0x100)], + iq3s_grid[qs[2*l+1] | ((qh << (7 - 2*l)) & 0x100)]); + + const int signs0 = __vcmpne4(((signs_packed_8[l] & 0x03) << 7) | ((signs_packed_8[l] & 0x0C) << 21), 0x00000000); + const int signs1 = __vcmpne4(((signs_packed_8[l] & 0x30) << 3) | ((signs_packed_8[l] & 0xC0) << 17), 0x00000000); + + const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0); + const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + 8*kqsx + (2*l+0)] = grid_l; + x_qs[i*sram_stride + 8*kqsx + (2*l+1)] = grid_h; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+0)] = grid_l; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+1)] = grid_h; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + const int ls = 1 + 2*((bxi->scales[kqsx/2] >> (((2*kqsx) << 1) & 0x04)) & 0x0F); + const float d = bxi->d; +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + kqsx] = ls*d; +#else + x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = ls*d; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq4_xs( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_XS, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_XS); + constexpr int nrows = warp_size / threads_per_row; + const int kqsx = threadIdx.x % threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride; + + const int aux_q4 = get_int_b4(bxi->qs, kqsx); + const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl); + const int k0 = 8 * (kqsx / 4) + kqsx % 4; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + k0 + 0] = v.x; + x_qs[i*sram_stride + k0 + 4] = v.y; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 4] = v.y; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + constexpr int rows_per_warp = warp_size / 8; +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) { + int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / (MMQ_TILE_NE_K/4); + + if (fallback) { + i = min(i, i_max); + } + + const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride; + + const float d = __half2float(bxi->d); + + const int ls = ((bxi->scales_l[(threadIdx.x % 8)/2] >> (4*(threadIdx.x % 2))) & 0x0F) + | (((bxi->scales_h >> (2*(threadIdx.x % 8))) & 0x03) << 4); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + threadIdx.x % 8] = d * (ls - 32); +#else + x_df[i*(MMQ_TILE_NE_K/4) + i/4 + threadIdx.x % 8] = d * (ls - 32); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq4_nl( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_NL, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_NL); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + const int kbx = txi / QI4_NL; + const int kqsx = txi % QI4_NL; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbx; + + const int aux_q4 = get_int_b2(bxi->qs, kqsx); + const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl); + const int k0 = kbx * (2 * QI4_NL) + kqsx; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + k0 + 0] = v.x; + x_qs[i*sram_stride + k0 + QI4_NL] = v.y; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI4_NL] = v.y; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_NL; + constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; + const int kbxd = threadIdx.x % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) { + int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbxd; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + kbxd] = __half2float(bxi->d); +#else + x_df[i*(MMQ_TILE_NE_K/QI4_NL) + i/QI4_NL + kbxd] = __half2float(bxi->d); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +// --------------------------------------------------------------------------------------------- + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_mxfp4( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_MXFP4, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / (4 * QR_MXFP4); + constexpr int nrows = warp_size / threads_per_row; + const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; + const int kbx = txi / QI_MXFP4; + const int kqsx = txi % QI_MXFP4; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nrows*nwarps) { + int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); + + if (fallback) { + i = min(i, i_max); + } + + const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i*stride + kbx; + + const int aux_q4 = get_int_b1(bxi->qs, kqsx); + const int2 v = get_int_from_table_16(aux_q4, kvalues_mxfp4); + const int k0 = kbx * (2 * QI_MXFP4) + kqsx; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + k0 + 0] = v.x; + x_qs[i*sram_stride + k0 + QI_MXFP4] = v.y; +#else + x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x; + x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI_MXFP4] = v.y; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI_MXFP4; + constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; + const int kbxd = threadIdx.x % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) { + int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; + + if (fallback) { + i = min(i, i_max); + } + + const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i*stride + kbxd; + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_df[i*sram_stride + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f; +#else + x_df[i*(MMQ_TILE_NE_K/QI_MXFP4) + i/QI_MXFP4 + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_mxfp4_fp4( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + + int * x_qs = (int *) x_tile; + uint32_t * x_sc = (uint32_t *) (x_qs + 2 * MMQ_TILE_NE_K); + + const int txi = threadIdx.x; + + constexpr int iter_k = ggml_cuda_mmq_get_K_vram(type, J, fallback); + + constexpr int threads_per_row = iter_k / QK_MXFP4; // each thread processes 1 block + constexpr int rows_per_warp = warp_size / threads_per_row; + const int kbx = txi % threads_per_row; + const int row_in_warp = txi / threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += rows_per_warp * nwarps) { + int i = i0 + threadIdx.y * rows_per_warp + row_in_warp; + + if constexpr (fallback) { + i = min(i, i_max); + } + + const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i * stride + kbx; + + // quantize_mxfp4_mmq permutes nibbles to match the quantized format + const int k0 = kbx * 4; + memcpy(x_qs + i*sram_stride + k0, bxi->qs, 16); + + // Load E8M0 scales: pack 2 consecutive scales into one uint32 + if (kbx % 2 == 0) { + uint32_t e = bxi->e; + e |= ((bxi + 1)->e << 8); + x_sc[i*sram_stride + kbx / 2] = e; + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_nvfp4( + const char * __restrict__ x, int * __restrict__ x_tile, const int kb0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_NVFP4, I); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int threads_per_row = MMQ_ITER_K / QK_NVFP4; + constexpr int rows_per_warp = warp_size / threads_per_row; + const int kbx = threadIdx.x % threads_per_row; + const int row_in_warp = threadIdx.x / threads_per_row; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += rows_per_warp * nwarps) { + int i = i0 + threadIdx.y * rows_per_warp + row_in_warp; + + if constexpr (fallback) { + i = min(i, i_max); + } + + const block_nvfp4 * bxi = (const block_nvfp4 *) x + kb0 + i * stride + kbx; + const uint32_t * __restrict__ src_qs = reinterpret_cast(bxi->qs); + const int kqs = 16 * kbx; + const int ksc = 4 * kbx; + +#pragma unroll + for (int sub = 0; sub < QK_NVFP4 / QK_NVFP4_SUB; ++sub) { + const int2 q0 = get_int_from_table_16(src_qs[2 * sub + 0], kvalues_mxfp4); + const int2 q1 = get_int_from_table_16(src_qs[2 * sub + 1], kvalues_mxfp4); + +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + x_qs[i*sram_stride + kqs + 4 * sub + 0] = q0.x; + x_qs[i*sram_stride + kqs + 4 * sub + 1] = q1.x; + x_qs[i*sram_stride + kqs + 4 * sub + 2] = q0.y; + x_qs[i*sram_stride + kqs + 4 * sub + 3] = q1.y; + x_df[i*sram_stride + ksc + sub] = ggml_cuda_ue4m3_to_fp32(bxi->d[sub]); +#else + x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 0] = q0.x; + x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 1] = q1.x; + x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 2] = q0.y; + x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 3] = q1.y; + x_df[i * (2 * MMQ_TILE_NE_K * 2 / QI_NVFP4) + i / (QK_NVFP4_SUB / QI_NVFP4) + ksc + sub] = ggml_cuda_ue4m3_to_fp32(bxi->d[sub]); +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_nvfp4_nvfp4( + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int iter_k = ggml_cuda_mmq_get_K_vram(type, J, fallback); + constexpr int threads_per_row = iter_k / QK_NVFP4; // each thread processes 1 block + constexpr int rows_per_warp = warp_size / threads_per_row; + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + + uint32_t * x_u32 = (uint32_t *) x_tile; + + const int txi = threadIdx.x; + const int kbx = txi % threads_per_row; + const int row_in_warp = txi / threads_per_row; + + const block_nvfp4 * bxi_base = (const block_nvfp4 *) x + kbx0 + kbx; + uint32_t * x_u32_scale = x_u32 + 64 + kbx; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += rows_per_warp * nwarps) { + int i = i0 + threadIdx.y * rows_per_warp + row_in_warp; + + if constexpr (fallback) { + i = min(i, i_max); + } + + const block_nvfp4 * bxi = bxi_base + i * stride; + + const uint32_t * src_qs = reinterpret_cast(bxi->qs); + +#pragma unroll + for (int sub = 0; sub < QK_NVFP4 / QK_NVFP4_SUB; ++sub) { + x_u32[i*sram_stride + 8*kbx + 2 * sub + 0] = src_qs[2 * sub + 0]; + x_u32[i*sram_stride + 8*kbx + 2 * sub + 1] = src_qs[2 * sub + 1]; + } + + x_u32_scale[i*sram_stride] = get_int_b4(bxi->d, 0); + } +} diff --git a/ggml/src/ggml-cuda/mmq-vec-dot.cuh b/ggml/src/ggml-cuda/mmq-vec-dot.cuh new file mode 100644 index 0000000000..d573433865 --- /dev/null +++ b/ggml/src/ggml-cuda/mmq-vec-dot.cuh @@ -0,0 +1,1251 @@ +#pragma once + +#include "vecdotq.cuh" +#include "mma.cuh" + +using namespace ggml_cuda_mma; + +#include "mmq.cuh" + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q4_0_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, I); + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + txs.qs; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR4_0*VDR_Q4_0_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + const int kyqs = QI8_1 * ((k01/2) / (QI8_1/2)) + (k01/2) % (QI8_1/2); + + int u[2*VDR_Q4_0_Q8_1_MMQ]; + + constexpr int max_cpy = ggml_cuda_get_max_cpy_bytes(); + constexpr int mcpy_int = max_cpy / sizeof(int); + static_assert(VDR_Q4_0_Q8_1_MMQ == 4, "bad VDR_Q4_0_Q8_1_MMQ"); + + int tmp0[4], tmp1[4]; + + #pragma unroll + for (int l0 = 0; l0 < 4 / mcpy_int; ++l0) { + ggml_cuda_memcpy_1(tmp0 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + l0 * mcpy_int] ); + ggml_cuda_memcpy_1(tmp1 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + QI4_0 + l0 * mcpy_int]); + } + + u[0]=tmp0[0]; u[2]=tmp0[1]; u[4]=tmp0[2]; u[6]=tmp0[3]; + u[1]=tmp1[0]; u[3]=tmp1[1]; u[5]=tmp1[2]; u[7]=tmp1[3]; + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q4_0_q8_1_impl + (&x_qs[i*(MMQ_TILE_NE_K + 1) + k0/QR4_0], u, + x_df[i*(MMQ_TILE_NE_K/QI4_0) + i/QI4_0 + k0/(QR4_0*QI4_0)], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q4_1_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, I); + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + txs.qs; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR4_1*VDR_Q4_1_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + const int kyqs = QI8_1 * ((k01/2) / (QI8_1/2)) + (k01/2) % (QI8_1/2); + + int u[2*VDR_Q4_1_Q8_1_MMQ]; + + constexpr int max_cpy = ggml_cuda_get_max_cpy_bytes(); + constexpr int mcpy_int = max_cpy / sizeof(int); + static_assert(VDR_Q4_0_Q8_1_MMQ == 4, "bad VDR_Q4_0_Q8_1_MMQ"); + + int tmp0[4], tmp1[4]; + + #pragma unroll + for (int l0 = 0; l0 < 4 / mcpy_int; ++l0) { + ggml_cuda_memcpy_1(tmp0 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + l0 * mcpy_int] ); + ggml_cuda_memcpy_1(tmp1 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + QI4_1 + l0 * mcpy_int]); + } + + u[0]=tmp0[0]; u[2]=tmp0[1]; u[4]=tmp0[2]; u[6]=tmp0[3]; + u[1]=tmp1[0]; u[3]=tmp1[1]; u[5]=tmp1[2]; u[7]=tmp1[3]; + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q4_1_q8_1_impl + (&x_qs[i*(MMQ_TILE_NE_K + 1) + k0/QR4_1], u, + x_dm[i*(MMQ_TILE_NE_K/QI4_1) + i/QI4_1 + k0/(QR4_1*QI4_1)], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, I); + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + txs.qs; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += VDR_Q8_0_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q8_0_q8_1_impl + (&x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k0 % MMQ_TILE_NE_K], + x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + k0/QI8_0], y_df[j*MMQ_TILE_Y_K + (k0/QI8_1) % (MMQ_TILE_NE_K/QI8_1)]); + } + } + } +} + +template +static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + constexpr data_layout input_layout = get_input_data_layout(); + typedef tile<16, 8, int, input_layout> tile_A; + typedef tile<16, 8, int, input_layout> tile_B; + typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + 2*MMQ_TILE_NE_K; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + const half2 * y_ds = (const half2 *) y; + + const int i0 = (threadIdx.y / ntx) * rows_per_warp; + + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { + const int k0 = k00 + k01; + + tile_A A[ntx]; +#pragma unroll + for (int n = 0; n < ntx; ++n) { + load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { + tile_B B; + load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); + + float dB; + const int j = j0 + tile_C::get_j(0); + if (ds_layout == MMQ_Q8_1_DS_LAYOUT_D4) { + dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; + } else { + dB = __low2float(y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C; + mma(C, A[n], B); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + const int i = i0 + n*tile_A::I + tile_C::get_i(l); + const float dA = x_df[i*sram_stride + k0/QI8_0]; + sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l]*dA*dB; + } + } + } + } +#else + typedef tile<16, 8, int> tile_A; + typedef tile< 8, 8, int> tile_B; + typedef tile<16, 8, int> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + 2*MMQ_TILE_NE_K; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + const half2 * y_ds = (const half2 *) y; + + tile_A A[ntx][MMQ_TILE_NE_K/QI8_0]; + float dA[ntx][tile_C::ne/2][MMQ_TILE_NE_K/QI8_0]; + + const int i0 = (threadIdx.y/ntx)*rows_per_warp; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { + const int k0 = k00 + k01; + + load_ldmatrix(A[n][k01/QI8_0], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_A::I + tile_C::get_i(2*l); + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { + const int k0 = k00 + k01; + + dA[n][l][k01/QI8_0] = x_df[i*sram_stride + k0/QI8_0]; + } + } + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { + tile_B B; + float dB[tile_C::ne/2]; + + load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); // faster than load_ldmatrix + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); + + if (ds_layout == MMQ_Q8_1_DS_LAYOUT_D4) { + dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; + } else { + dB[l] = __low2float(y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C; + mma(C, A[n][k01/QI8_0], B); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l]*dA[n][l/2][k01/QI8_0]*dB[l%2]; + } + } + } + } +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) +} + + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q8_1_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, I); + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + txs.qs; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += VDR_Q8_0_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q8_1_q8_1_impl + (&x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], + x_dm[i*(MMQ_TILE_NE_K/QI5_1) + i/QI5_1 + k0/QI8_1], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + constexpr data_layout input_layout = get_input_data_layout(); + typedef tile<16, 8, int, input_layout> tile_A; + typedef tile<16, 8, int, input_layout> tile_B; + typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + 2*MMQ_TILE_NE_K; + const int * y_qs = (const int *) y + 4; + const half2 * y_dm = (const half2 *) y; + + const int i0 = (threadIdx.y / ntx) * rows_per_warp; + + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { + const int k0 = k00 + k01; + + tile_A A[ntx]; +#pragma unroll + for (int n = 0; n < ntx; ++n) { + load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { + tile_B B; + load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); + + const int j = j0 + tile_C::get_j(0); + const float2 dsB = __half22float2(y_dm[j*MMQ_TILE_Y_K + k01/QI8_1]); + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C; + mma(C, A[n], B); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + const int i = i0 + n*tile_A::I + tile_C::get_i(l); + float2 dmA = __half22float2(x_dm[i*sram_stride + k0/QI8_1]); + sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA.x*dsB.x*C.x[l]; + sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA.y*dsB.y; + } + } + } + } +#else + typedef tile<16, 8, int> tile_A; + typedef tile< 8, 8, int> tile_B; + typedef tile<16, 8, int> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + 2*MMQ_TILE_NE_K; + const int * y_qs = (const int *) y + 4; + const half2 * y_dm = (const half2 *) y; + + tile_A A[ntx][MMQ_TILE_NE_K/QI8_1]; + float2 dmA[ntx][tile_C::ne/2][MMQ_TILE_NE_K/QI8_1]; + + const int i0 = (threadIdx.y/ntx)*rows_per_warp; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { + const int k0 = k00 + k01; + + load_ldmatrix(A[n][k01/QI8_1], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_A::I + tile_C::get_i(2*l); + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { + const int k0 = k00 + k01; + + dmA[n][l][k01/QI8_1] = __half22float2(x_dm[i*sram_stride + k0/QI8_1]); + } + } + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { + tile_B B; + float2 dsB[tile_C::ne/2]; + + load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); // faster than load_ldmatrix + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); + + dsB[l] = __half22float2(y_dm[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C; + mma(C, A[n][k01/QI8_1], B); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA[n][l/2][k01/QI8_1].x*dsB[l%2].x*C.x[l]; + sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA[n][l/2][k01/QI8_1].y*dsB[l%2].y; + } + } + } + } +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) +} + +// Used for NVFP4, Q3_K, IQ2_S, and IQ2_XS +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(type, I); + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + txs.qs; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q8_0_16_q8_1_impl( + &x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], + &y_qs[j*MMQ_TILE_Y_K + k01], + &x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + k0/(QI8_0/2)], + y_df[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + } +} + +// Used for Q3_K, IQ2_S, and IQ2_XS: +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_mma( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + constexpr data_layout input_layout = get_input_data_layout(); + typedef tile<16, 4, int, input_layout> tile_A; + typedef tile<16, 4, int, input_layout> tile_B; + typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + + const int i0 = (threadIdx.y / ntx) * rows_per_warp; + + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { + const int k0 = k00 + k01; + + tile_A A[ntx]; +#pragma unroll + for (int n = 0; n < ntx; ++n) { + load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { + tile_B B; + load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); + + const int j = j0 + tile_C::get_j(0); + const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C; + mma(C, A[n], B); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(l); + sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * x_df[i*sram_stride + k0/4] * dB; + } + } + } + } +#elif defined(TURING_MMA_AVAILABLE) + + typedef tile<16, 4, int> tile_A; + typedef tile<16, 8, int> tile_A_8; + typedef tile< 8, 4, int> tile_B; + typedef tile<16, 8, int> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + + const int i0 = (threadIdx.y / ntx) * (ntx*tile_A::I); + + tile_A A[ntx][8]; + float dA[ntx][tile_C::ne/2][8]; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 8) { + const int k0 = k00 + k01; + + load_ldmatrix(((tile_A_8 *) A[n])[k01/8], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { + const int k0 = k00 + k01; + + dA[n][l][k01/4] = x_df[i*sram_stride + k0/4]; + } + } + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR3_K*VDR_Q3_K_Q8_1_MMQ) { + tile_B B[2]; + float dB[tile_C::ne/2]; + + // Here load_generic is faster than load_ldmatrix. + load_generic(B[0], y_qs + j0*MMQ_TILE_Y_K + (k01 + 0), MMQ_TILE_Y_K); + load_generic(B[1], y_qs + j0*MMQ_TILE_Y_K + (k01 + tile_B::J), MMQ_TILE_Y_K); + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); + + dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C[2]; + mma(C[0], A[n][k01/4 + 0], B[0]); + mma(C[1], A[n][k01/4 + 1], B[1]); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0/tile_C::J + n)*tile_C::ne + l] += dB[l%2]*(C[0].x[l]*dA[n][l/2][k01/4 + 0] + C[1].x[l]*dA[n][l/2][k01/4 + 1]); + } + } + } + } +#else + GGML_UNUSED_VARS(x, y, sum, k00); + NO_DEVICE_CODE; +#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q2_K_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, I); + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + txs.qs; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + + float2 y_df[J/nwarps]; +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + + y_df[j0/nwarps] = __half22float2(y_ds[j*MMQ_TILE_Y_K]); + } + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K/2; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + constexpr int ns = 2; + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q2_K_q8_1_impl_mmq( + &x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], + &x_dm[i*(MMQ_TILE_NE_K + 1) + k0/4], k01 < MMQ_TILE_NE_K/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y, + &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); + } + } + } + + // Some compilers fail to unroll the loop over k01 if there is a conditional statement for ns in the inner loop. + // As a workaround 2 separate loops are used instead. +#pragma unroll + for (int k01 = MMQ_TILE_NE_K/2; k01 < MMQ_TILE_NE_K; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + constexpr int ns = 1; + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q2_K_q8_1_impl_mmq( + &x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], + &x_dm[i*(MMQ_TILE_NE_K + 1) + k0/4], k01 < MMQ_TILE_NE_K/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y, + &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); + } + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q2_K_q8_1_mma( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + constexpr data_layout input_layout = get_input_data_layout(); + typedef tile<16, 4, int, input_layout> tile_A; + typedef tile<16, 4, int, input_layout> tile_B; + typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + MMQ_TILE_NE_K*2; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + + const int i0 = (threadIdx.y / ntx) * rows_per_warp; + + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { + const int k0 = k00 + k01; + + tile_A A[ntx]; +#pragma unroll + for (int n = 0; n < ntx; ++n) { + load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { + tile_B B; + load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); + + const int j = j0 + tile_C::get_j(0); + const float dB = (k01 < MMQ_TILE_NE_K/2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K]).x : __half22float2(y_ds[j*MMQ_TILE_Y_K]).y; + const float sB = (k01 >= MMQ_TILE_NE_K * 3/4) ? 0 + : (((k01/4)%2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).y + : __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).x); + + tile_C Cm; + if (k01 >= MMQ_TILE_NE_K * 3/4) { + tile_A A1; +#pragma unroll + for (int l = 0; l < tile_A::ne; ++l) { + A1.x[l] = 0x01010101; + } + mma(Cm, A1, B); + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C Cd; + mma(Cd, A[n], B); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(l); + const float2 dm = __half22float2(x_dm[i*sram_stride + k0/4]); + float tmp = Cd.x[l]*dm.x; + if (k01 >= MMQ_TILE_NE_K * 3/4) { + tmp -= Cm.x[l]*dm.y; + } + sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp*dB; + sum[(j0/tile_C::J + n)*tile_C::ne + l] -= dm.y*sB; + } + } + } + } +#elif defined(TURING_MMA_AVAILABLE) + + typedef tile<16, 4, int> tile_A; + typedef tile<16, 8, int> tile_A_8; + typedef tile< 8, 4, int> tile_B; + typedef tile<16, 8, int> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + MMQ_TILE_NE_K*2; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + + const int i0 = (threadIdx.y / ntx) * (ntx*tile_A::I); + + tile_A A[ntx][8]; + float dA[ntx][tile_C::ne/2][8]; + float mA[ntx][tile_C::ne/2][8]; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { + const int k0 = k00 + k01; + + load_ldmatrix(((tile_A_8 *) A[n])[k01/QI8_1], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1/2) { + const int k0 = k00 + k01; + + const float2 dm = __half22float2(x_dm[i*sram_stride + k0/(QI8_1/2)]); + + dA[n][l][k01/(QI8_1/2)] = dm.x; + mA[n][l][k01/(QI8_1/2)] = dm.y; + } + } + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { + float2 dB[tile_C::ne/2]; + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); + + dB[l] = __half22float2(y_ds[j*MMQ_TILE_Y_K]); + } + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { + tile_B B[2]; + + // Here load_generic is faster than load_ldmatrix. + load_generic(B[0], y_qs + j0*MMQ_TILE_Y_K + (k01 + 0), MMQ_TILE_Y_K); + load_generic(B[1], y_qs + j0*MMQ_TILE_Y_K + (k01 + tile_B::J), MMQ_TILE_Y_K); + + tile_C Cm[2]; + if (k01 >= MMQ_TILE_NE_K * 3/4) { + tile_A A1; + A1.x[0] = 0x01010101; + A1.x[1] = 0x01010101; + mma(Cm[0], A1, B[0]); + mma(Cm[1], A1, B[1]); + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C Cd[2]; + + mma(Cd[0], A[n][k01/4 + 0], B[0]); + mma(Cd[1], A[n][k01/4 + 1], B[1]); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + float tmp = Cd[0].x[l]*dA[n][l/2][k01/4 + 0] + Cd[1].x[l]*dA[n][l/2][k01/4 + 1]; + if (k01 >= MMQ_TILE_NE_K * 3/4) { + tmp -= Cm[0].x[l]*mA[n][l/2][k01/4 + 0] + Cm[1].x[l]*mA[n][l/2][k01/4 + 1]; + } + sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp*(k01 < MMQ_TILE_NE_K/2 ? dB[l%2].x : dB[l%2].y); + } + } + } + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K * 3/4; k01 += QI8_1) { + float2 sB[tile_C::ne/2]; + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); + + sB[l] = __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0/tile_C::J + n)*tile_C::ne + l] -= mA[n][l/2][k01/4 + 0]*sB[l%2].x; + sum[(j0/tile_C::J + n)*tile_C::ne + l] -= mA[n][l/2][k01/4 + 1]*sB[l%2].y; + } + } + } + } +#else + GGML_UNUSED_VARS(x, y, sum, k00); + NO_DEVICE_CODE; +#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q3_K_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, I); + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + txs.qs; + const int * x_sc = (const int *) x_df + txs.dm; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR3_K*VDR_Q3_K_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + const int8_t * scales = ((const int8_t *) (x_sc + i*(MMQ_TILE_NE_K/8) + i/8)) + k0/4; + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q3_K_q8_1_impl_mmq( + &x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], scales, + x_df[i], y_df[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q4_K_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, I); + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + txs.qs; + const int * x_sc = (const int *) x_dm + txs.dm; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR4_K*VDR_Q4_K_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + const uint8_t * sc = (const uint8_t *) &x_sc[i * (MMQ_TILE_NE_K/8) + i/8 + k0/32] + 2*(k01/16); + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q4_K_q8_1_impl_mmq( + &x_qs[i*(MMQ_TILE_NE_K + 1) + k0/2], &y_qs[j*MMQ_TILE_Y_K + k01], sc, sc+8, + x_dm[i], &y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q5_K_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, I); + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + txs.qs; + const int * x_sc = (const int *) x_dm + txs.dm; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR5_K*VDR_Q5_K_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + const uint8_t * sc = ((const uint8_t *) &x_sc[i * (MMQ_TILE_NE_K/8) + i/8 + k00/32]) + 2*(k01/16); + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q5_K_q8_1_impl_mmq( + &x_qs[i*(QR5_K*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], sc, sc+8, + x_dm[i], &y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q6_K_q8_1_dp4a( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, I); + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + txs.qs; + const int * x_sc = (const int *) x_df + txs.dm; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + +// #pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR6_K*VDR_Q6_K_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < I; i0 += warp_size) { + const int i = i0 + threadIdx.x; + + const int8_t * sc = ((const int8_t *) &x_sc[i * (MMQ_TILE_NE_K/8) + i/8 + k0/16]); + + sum[j0/nwarps*I/warp_size + i0/warp_size] += vec_dot_q6_K_q8_1_impl_mmq( + &x_qs[i*(QR6_K*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], sc, + x_df[i*(MMQ_TILE_NE_K/QI6_K) + i/QI6_K], &y_df[j*MMQ_TILE_Y_K + k01/QI8_1]); + } + } + } +} + +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q6_K_q8_1_mma( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + constexpr data_layout input_layout = get_input_data_layout(); + typedef tile<16, 4, int, input_layout> tile_A; + typedef tile<16, 4, int, input_layout> tile_B; + typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; + const int * x_sc = (const int *) x_df + MMQ_TILE_NE_K/QI6_K; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + + const int i0 = (threadIdx.y / ntx) * rows_per_warp; + + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { + const int k0 = k00 + k01; + + tile_A A[ntx]; +#pragma unroll + for (int n = 0; n < ntx; ++n) { + load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*sram_stride + k0, sram_stride); + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { + tile_B B; + load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); + + const int j = j0 + tile_C::get_j(0); + const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C; + mma(C, A[n], B); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(l); + const int8_t * sc = (const int8_t *) (x_sc + i*sram_stride + k00/16); + sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * sc[k01/4] * x_df[i*sram_stride] * dB; + } + } + } + } +#elif defined(TURING_MMA_AVAILABLE) + + typedef tile<16, 4, int> tile_A; + typedef tile< 8, 4, int> tile_B; + typedef tile<16, 8, int> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; + const int * x_sc = (const int *) x_df + MMQ_TILE_NE_K/QI6_K; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + + const int i0 = (threadIdx.y / ntx) * (ntx*tile_A::I); + + tile_A A[ntx][8]; + int scA[ntx][tile_C::ne/2][8]; + float dA[ntx][tile_C::ne/2]; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 8) { + const int k0 = k00 + k01; + + load_ldmatrix(A[n][k01/4 + 0], x_qs + (i0 + n*tile_A::I)*sram_stride + (k0 + 0), sram_stride); + load_ldmatrix(A[n][k01/4 + 1], x_qs + (i0 + n*tile_A::I)*sram_stride + (k0 + tile_A::J), sram_stride); + } + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 16) { + const int k0 = k00 + k01; + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); + + const int sc_packed = x_sc[i*sram_stride + k0/16]; + const int8_t * sc = (const int8_t *) &sc_packed; + +#pragma unroll + for (int ksc = 0; ksc < sizeof(int); ++ksc) { + scA[n][l][k01/4 + ksc] = sc[ksc]; + } + } + } + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); + + dA[n][l] = x_df[i*sram_stride]; + } + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { + float tmp[ntx][tile_C::ne] = {{0.0f}}; + +#pragma unroll + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 8) { + tile_B B[2]; + float dB[tile_C::ne/2]; + + // Here load_generic is faster than load_ldmatrix. + load_generic(B[0], y_qs + j0*MMQ_TILE_Y_K + 0 + k01, MMQ_TILE_Y_K); + load_generic(B[1], y_qs + j0*MMQ_TILE_Y_K + tile_B::J + k01, MMQ_TILE_Y_K); + +#pragma unroll + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); + + dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C[2]; + mma(C[0], A[n][k01/4 + 0], B[0]); + mma(C[1], A[n][k01/4 + 1], B[1]); + +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + tmp[n][l] += (C[0].x[l]*scA[n][l/2][k01/4 + 0] + C[1].x[l]*scA[n][l/2][k01/4 + 1])*dB[l%2]; + } + } + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp[n][l]*dA[n][l/2]; + } + } + } +#else + GGML_UNUSED_VARS(x, y, sum, k00); + NO_DEVICE_CODE; +#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE +} + +// --------------------------------------------------------------------------------------------- + +// Shared MMA kernel for MXFP4 and NVFP4 on Blackwell. +// Both quantizations encode values as e2m1 (FP4) and produce one uint32 scale per +// m16n8k64 MMA call; only the PTX kind (scale_vec::2X ue8m0 vs scale_vec::4X ue4m3) +// and the per-type stride constant differ. +template static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_fp4_fp4_mma( + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { + + typedef tile<16, 8, int> tile_A; + typedef tile<8, 8, int> tile_B; + typedef tile<16, 8, float> tile_C; + + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp / tile_C::I; + constexpr int nfrags = MMQ_TILE_NE_K / tile_A::J; + + y += (threadIdx.y % ntx) * (tile_C::J * MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const uint32_t * x_sc = (const uint32_t *) (x_qs + 2 * MMQ_TILE_NE_K); + const int * y_qs = (const int *) y + 4; + const uint32_t * y_sc = (const uint32_t *) y; + + // 2 threads per quad supply the packed scale register to the block_scale MMA, + // see https://docs.nvidia.com/cuda/parallel-thread-execution/#warp-level-block-scaling + const int tidx_A = threadIdx.x / 4 + (threadIdx.x % 2) * 8; + const int tidx_B = threadIdx.x / 4; + const int i0 = (threadIdx.y / ntx) * rows_per_warp; + + tile_A A[ntx][nfrags]; + uint32_t scaleA[ntx][nfrags]; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int frag = 0; frag < nfrags; ++frag) { + const int k0 = k00 + frag * tile_A::J; + load_ldmatrix(A[n][frag], x_qs + (i0 + n * tile_A::I) * sram_stride + k0, sram_stride); + scaleA[n][frag] = x_sc[(i0 + n * tile_A::I + tidx_A) * sram_stride + k0 / tile_A::J]; + } + } + +#pragma unroll + for (int j0 = 0; j0 < J; j0 += ntx * tile_C::J) { + tile_B B[nfrags]; + uint32_t scaleB[nfrags]; + +#pragma unroll + for (int frag = 0; frag < nfrags; ++frag) { + const int k0 = frag * tile_B::J; + load_generic(B[frag], y_qs + j0 * MMQ_TILE_Y_K + k0, MMQ_TILE_Y_K); + scaleB[frag] = y_sc[(j0 + tidx_B) * MMQ_TILE_Y_K + frag]; + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { +#pragma unroll + for (int frag = 0; frag < nfrags; ++frag) { + tile_C C = {}; + mma_block_scaled_fp4(C, A[n][frag], B[frag], scaleA[n][frag], scaleB[frag]); +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0 / tile_C::J + n) * tile_C::ne + l] += C.x[l]; + } + } + } + } +} + diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu index 6b3b0d064a..bf9f5d5264 100644 --- a/ggml/src/ggml-cuda/mmq.cu +++ b/ggml/src/ggml-cuda/mmq.cu @@ -3,6 +3,8 @@ #include "quantize.cuh" #include "mmid.cuh" +#include + static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) { switch (args.type_x) { case GGML_TYPE_Q1_0: @@ -118,15 +120,14 @@ void ggml_cuda_mul_mat_q( const int64_t s03 = src0->nb[3] / ts_src0; const int64_t s3 = dst->nb[3] / ts_dst; - const bool use_stream_k = (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) - || GGML_CUDA_CC_IS_CDNA(cc); + const bool fallback = ne01 % 128 != 0; // TODO: tighter pool buffer size vs q8 path const bool use_native_fp4 = blackwell_mma_available(cc) && (src0->type == GGML_TYPE_MXFP4 || src0->type == GGML_TYPE_NVFP4); if (!ids) { const size_t nbytes_src1_q8_1 = ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1 + - get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq); + ggml_cuda_mmq_get_J_max(src0->type, fallback, cc, ne11) * sizeof(block_q8_1_mmq); ggml_cuda_pool_alloc src1_q8_1(ctx.pool(), nbytes_src1_q8_1); { @@ -156,7 +157,7 @@ void ggml_cuda_mul_mat_q( ne00, ne01, ne1, s01, ne11, s1, ne02, ne12, s02, s12, s2, ne03, ne13, s03, s13, s3, - use_stream_k, ne1}; + ne1}; ggml_cuda_mul_mat_q_switch_type(ctx, args, stream); return; } @@ -184,7 +185,7 @@ void ggml_cuda_mul_mat_q( } const size_t nbytes_src1_q8_1 = ne12*n_expert_used*ne10_padded * sizeof(block_q8_1)/QK8_1 + - get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq); + ggml_cuda_mmq_get_J_max(src0->type, fallback, cc, ne11) * sizeof(block_q8_1_mmq); ggml_cuda_pool_alloc src1_q8_1(ctx.pool(), nbytes_src1_q8_1); const int64_t ne11_flat = ne12*n_expert_used; @@ -217,53 +218,11 @@ void ggml_cuda_mul_mat_q( ne00, ne01, ne_get_rows, s01, ne_get_rows, s1, ne02, ne02, s02, s12, s2, ne03, ne13, s03, s13, s3, - use_stream_k, ne12}; + ne12}; ggml_cuda_mul_mat_q_switch_type(ctx, args, stream); } -void ggml_cuda_op_mul_mat_q( - ggml_backend_cuda_context & ctx, - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, - const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, - const int64_t src1_padded_row_size, cudaStream_t stream) { - - const int64_t ne00 = src0->ne[0]; - - const int64_t ne10 = src1->ne[0]; - const int64_t ne11 = src1->ne[1]; - GGML_ASSERT(ne10 % QK8_1 == 0); - - const int64_t ne0 = dst->ne[0]; - - const int64_t row_diff = row_high - row_low; - const int64_t stride01 = ne00 / ggml_blck_size(src0->type); - - const int id = ggml_cuda_get_device(); - const int cc = ggml_cuda_info().devices[id].cc; - - // the main device has a larger memory buffer to hold the results from all GPUs - // nrows_dst == nrows of the matrix that the kernel writes into - const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff; - - // The stream-k decomposition is only faster for recent NVIDIA GPUs. - // Also its fixup needs to allocate a temporary buffer in the memory pool. - // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer. - const bool use_stream_k = ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) - || GGML_CUDA_CC_IS_CDNA(cc)) - && src1_ncols == ne11; - const mmq_args args = { - src0_dd_i, src0->type, (const int *) src1_ddq_i, nullptr, nullptr, dst_dd_i, - ne00, row_diff, src1_ncols, stride01, ne11, nrows_dst, - 1, 1, 0, 0, 0, - 1, 1, 0, 0, 0, - use_stream_k, src1_ncols}; - - ggml_cuda_mul_mat_q_switch_type(ctx, args, stream); - - GGML_UNUSED_VARS(src1, dst, src1_ddf_i, src1_padded_row_size); -} - bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t n_experts) { #ifdef GGML_CUDA_FORCE_CUBLAS return false; diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh index edf546d8f1..607e433bf1 100644 --- a/ggml/src/ggml-cuda/mmq.cuh +++ b/ggml/src/ggml-cuda/mmq.cuh @@ -1,22 +1,18 @@ #pragma once #include "common.cuh" -#include "vecdotq.cuh" -#include "mma.cuh" #include #include -using namespace ggml_cuda_mma; - #define MMQ_DP4A_MAX_BATCH_SIZE 64 // Max. batch size to use for dp4a MMQ kernels when FP16 tensor cores are available. #define MMQ_ITER_K 256 #define MMQ_ITER_K_FP4 512 #define MMQ_NWARPS 8 -typedef void (*load_tiles_mmq_t)(const char * __restrict__ x, int * x_tile, const int kbx0, const int i_max, const int stride); -typedef void (*vec_dot_mmq_t)(const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00); -typedef void (*mmq_write_back_t)(const float * __restrict__ sum, const int32_t * __restrict__ get_rows_to_sorted, +typedef void (*ggml_cuda_mmq_load_tiles_t)(const char * __restrict__ x, int * x_tile, const int kbx0, const int i_max, const int stride); +typedef void (*ggml_cuda_mmq_vec_dot_t)(const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00); +typedef void (*ggml_cuda_mmq_write_back_t)(const float * __restrict__ sum, const int32_t * __restrict__ get_rows_to_sorted, float * __restrict__ dst, const int stride, const int i_max, const int j_max); enum mmq_q8_1_ds_layout { @@ -106,69 +102,6 @@ struct tile_x_sizes { int sc; }; -static int get_mmq_x_max_host(const int cc) { - return (turing_mma_available(cc) || amd_wmma_available(cc)) ? 128 : - GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA ? -#ifdef GGML_CUDA_FORCE_MMQ - 128 : 64; -#else - MMQ_DP4A_MAX_BATCH_SIZE : 64; -#endif // GGML_CUDA_FORCE_MMQ -} - -static constexpr __device__ int get_mmq_x_max_device() { -#if defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - return 128; -#else // defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - -#if defined(GGML_USE_HIP) - return 64; -#else // defined(GGML_USE_HIP) - -#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA -#ifdef GGML_CUDA_FORCE_MMQ - return 128; -#else // GGML_CUDA_FORCE_MMQ - return MMQ_DP4A_MAX_BATCH_SIZE; -#endif // GGML_CUDA_FORCE_MMQ -#else // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA - return 64; -#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA - -#endif // defined(GGML_USE_HIP) -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) -} - -static int get_mmq_y_host(const int cc) { - return GGML_CUDA_CC_IS_AMD(cc) ? (GGML_CUDA_CC_IS_RDNA1(cc) ? 64 : 128) : - ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ? 128 : 64); -} - -static constexpr __device__ int get_iter_k([[maybe_unused]] const ggml_type type) { -#if defined(BLACKWELL_MMA_AVAILABLE) -if (type == GGML_TYPE_NVFP4 || type == GGML_TYPE_MXFP4) { - return MMQ_ITER_K_FP4; -} -#endif // defined(BLACKWELL_MMA_AVAILABLE) - return MMQ_ITER_K; -} - -static constexpr __device__ int get_mmq_y_device() { -#if defined(GGML_USE_HIP) -#if defined(RDNA1) - return 64; -#else - return 128; -#endif // defined RDNA1 -#else -#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA - return 128; -#else - return 64; -#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA -#endif // defined(GGML_USE_HIP) -} - // Decouple shared memory tile sizes from WARP_SIZE to allow for different warp sizes. // The K dimension of the tiles has either, // 1*MMQ_TILE_NE_K==32 (always for TILE_Y_K) or 2*MMQ_TILE_NE_K==64 (typically for TILE_X_K), @@ -178,18 +111,263 @@ static constexpr __device__ int get_mmq_y_device() { // in terms of 32 bit elements that means K % 2 == 1 for dp4a or K % 8 == 4 for mma. #define MMQ_TILE_NE_K 32 -#define MMQ_DP4A_TXS_Q4_0 tile_x_sizes{mmq_y*MMQ_TILE_NE_K + mmq_y, mmq_y*MMQ_TILE_NE_K/QI4_0 + mmq_y/QI4_0, 0} -#define MMQ_DP4A_TXS_Q4_1 tile_x_sizes{mmq_y*MMQ_TILE_NE_K + mmq_y, mmq_y*MMQ_TILE_NE_K/QI4_1 + mmq_y/QI4_1, 0} -#define MMQ_DP4A_TXS_Q8_0 tile_x_sizes{mmq_y*MMQ_TILE_NE_K*2 + mmq_y, mmq_y*MMQ_TILE_NE_K*2/QI8_0 + mmq_y/(QI8_0/2), 0} -#define MMQ_DP4A_TXS_Q8_0_16 tile_x_sizes{mmq_y*MMQ_TILE_NE_K*2 + mmq_y, mmq_y*MMQ_TILE_NE_K*4/QI8_0 + mmq_y/(QI8_0/4), 0} -#define MMQ_DP4A_TXS_Q8_1 tile_x_sizes{mmq_y*MMQ_TILE_NE_K*2 + mmq_y, mmq_y*MMQ_TILE_NE_K*2/QI8_1 + mmq_y/(QI8_1/2), 0} -#define MMQ_DP4A_TXS_Q2_K tile_x_sizes{mmq_y*MMQ_TILE_NE_K*2 + mmq_y, mmq_y*MMQ_TILE_NE_K + mmq_y, 0} -#define MMQ_DP4A_TXS_Q3_K tile_x_sizes{mmq_y*MMQ_TILE_NE_K*2 + mmq_y, mmq_y, mmq_y*MMQ_TILE_NE_K/8 + mmq_y/8} -#define MMQ_DP4A_TXS_Q4_K tile_x_sizes{mmq_y*MMQ_TILE_NE_K + mmq_y, mmq_y*MMQ_TILE_NE_K/QI4_K, mmq_y*MMQ_TILE_NE_K/8 + mmq_y/8} -#define MMQ_DP4A_TXS_Q5_K tile_x_sizes{mmq_y*MMQ_TILE_NE_K*2 + mmq_y, mmq_y*MMQ_TILE_NE_K/QI5_K + mmq_y/QI5_K, mmq_y*MMQ_TILE_NE_K/8 + mmq_y/8} -#define MMQ_DP4A_TXS_Q6_K tile_x_sizes{mmq_y*MMQ_TILE_NE_K*2 + mmq_y, mmq_y*MMQ_TILE_NE_K/QI6_K + mmq_y/QI6_K, mmq_y*MMQ_TILE_NE_K/8 + mmq_y/8} +// block_q8_1_mmq has (128 8-bit ints == 32 32-bit ints + 4 32-bit scales) +#define MMQ_TILE_Y_K (MMQ_TILE_NE_K + MMQ_TILE_NE_K / QI8_1) +#define MMQ_TILE_Y_FP4_K MMQ_TILE_Y_K -static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml_type type, int mmq_y) { +enum ggml_cuda_mmq_sram_layout { + GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, + GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, + GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, + GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, + GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, + GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, // MXFP4 and NVFP4 on Blackwell. + GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, // Generic NVFP4 +}; + +static constexpr __host__ __device__ int ggml_cuda_mmq_get_sram_stride(ggml_cuda_mmq_sram_layout sram_layout) { + switch (sram_layout) { + case GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0: + return 2*MMQ_TILE_NE_K + 2*MMQ_TILE_NE_K/QI8_0 + 4; + case GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1: + return 2*MMQ_TILE_NE_K + 2*MMQ_TILE_NE_K/QI8_1 + 4; + case GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K: + return 2*MMQ_TILE_NE_K + MMQ_TILE_NE_K + 4; + case GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K: + return 2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/2 + 4; + case GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K: + return 2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/QI6_K + MMQ_TILE_NE_K/8 + 7; + case GGML_CUDA_MMQ_SRAM_LAYOUT_FP4: + return 2*MMQ_TILE_NE_K + 8 + 4; + case GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4: + return 2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/2 + 4; + default: + return -1; + } +} + +static_assert(ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0) % 8 == 4, "Wrong padding."); +static_assert(ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1) % 8 == 4, "Wrong padding."); +static_assert(ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K) % 8 == 4, "Wrong padding."); +static_assert(ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K) % 8 == 4, "Wrong padding."); +static_assert(ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K) % 8 == 4, "Wrong padding."); +static_assert(ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_FP4) % 8 == 4, "Wrong padding."); +static_assert(ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4) % 8 == 4, "Wrong padding."); + +static_assert(ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_FP4) == ggml_cuda_mmq_get_sram_stride(GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1), "Wrong tile size for MXFP4"); + +// Config options for the MMQ kernel. +// Should not affect results, only speed/register pressure/shared memory use. +struct ggml_cuda_mmq_config { + ggml_type type; // src0->type + int nthreads; // Number of threads per CUDA block. + int occupancy; // Targeted occupancy for the MMA kernel. + int I; // SRAM tile width in src0->ne[1]/dst->ne[0] direction. + int J; // SRAM tile width in src1->ne[1]/dst->ne[1] direction. + ggml_cuda_mmq_sram_layout sram_layout; // SRAM tile length in src0->ne[0]/src1->ne[0] direction (physical 32 bit elements). + int K_vram; // VRAM tile length in src0->ne[0]/src1->ne[0] direction (logical elements). + bool stream_k; // Whether or not to use stream-k decomposition. + bool fallback; // Whether a fallback for out-of-bounds check in src0->ne[1] direction is needed. + + constexpr __host__ __device__ ggml_cuda_mmq_config( + ggml_type type, int nthreads, int occupancy, int I, int J, ggml_cuda_mmq_sram_layout sram_layout, int K_vram, bool stream_k, bool fallback) : + type(type), nthreads(nthreads), occupancy(occupancy), I(I), J(J), sram_layout(sram_layout), K_vram(K_vram), stream_k(stream_k), fallback(fallback) {} + + constexpr __device__ int rows_per_warp() const { +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + return 16; +#else + return J >= 48 && J % 16 == 0 ? 32 : 16; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + } + + // TODO transition all combinations of GPUs and quantizations to the MMA data layout. + __host__ int use_mma_data_layout(const int cc) const { + if (amd_mfma_available(cc) || amd_wmma_available(cc) || turing_mma_available(cc)) { + return true; + } + return false; + } + + constexpr __device__ bool use_mma_data_layout() const { +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) + return true; +#else + return false; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) + } + +}; + +#define CASE(type_, nthreads_, occupancy_, I_, J_, sram_layout_, K_vram_, stream_k_, fallback_) \ + if (type == (type_) && J == (J_) && fallback == (fallback_)) { \ + static_assert((nthreads_) % 32 == 0 && (nthreads_) <= 512, "bad nthreads"); \ + static_assert( (occupancy_) <= 8, "bad occupancy"); \ + static_assert((I_) % 32 == 0, "bad I"); \ + static_assert((J_) % 8 == 0, "bad J"); \ + static_assert((K_vram_) % 256 == 0, "bad K_vram"); \ + return ggml_cuda_mmq_config((type_), (nthreads_), (occupancy_), (I_), (J_), (sram_layout_), (K_vram_), (stream_k_), (fallback_)); \ + } \ + +#include "mmq-config-pascal.cuh" +#include "mmq-config-ampere.cuh" +#include "mmq-config-blackwell.cuh" + +#include "mmq-config-cdna.cuh" +#include "mmq-config-rdna2.cuh" +#include "mmq-config-rdna4.cuh" + +#undef CASE + +static __host__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config(const ggml_type type, const int J, const bool fallback, const int cc) { + if (GGML_CUDA_CC_IS_AMD(cc)) { + if (GGML_CUDA_CC_IS_CDNA(cc)) { + return ggml_cuda_mmq_get_config_cdna(type, J, fallback); + } + if (amd_wmma_available(cc)) { + return ggml_cuda_mmq_get_config_rdna4(type, J, fallback); + } + return ggml_cuda_mmq_get_config_rdna2(type, J, fallback); + } + if (blackwell_mma_available(cc)) { + return ggml_cuda_mmq_get_config_blackwell(type, J, fallback); + } + if (ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) { + return ggml_cuda_mmq_get_config_ampere(type, J, fallback); + } + return ggml_cuda_mmq_get_config_pascal(type, J, fallback); +} + +static constexpr __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config(ggml_type type, int J, bool fallback) { +#ifdef GGML_USE_HIP +#ifdef CDNA + return ggml_cuda_mmq_get_config_cdna(type, J, fallback); +#elif defined(AMD_WMMA_AVAILABLE) + return ggml_cuda_mmq_get_config_rdna4(type, J, fallback); +#else + return ggml_cuda_mmq_get_config_rdna2(type, J, fallback); +#endif // CDNA +#else +#ifdef BLACKWELL_MMA_AVAILABLE + return ggml_cuda_mmq_get_config_blackwell(type, J, fallback); +#elif __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA + return ggml_cuda_mmq_get_config_ampere(type, J, fallback); +#else + return ggml_cuda_mmq_get_config_pascal(type, J, fallback); +#endif // BLACKWELL_MMA_AVAILABLE +#endif // GGML_USE_HIP + GGML_UNUSED_VARS(type, J, fallback); +} + +static __host__ int ggml_cuda_mmq_get_type(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).type; +} + +static constexpr __device__ int ggml_cuda_mmq_get_type(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).type; +} + +static __host__ int ggml_cuda_mmq_get_nthreads(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).nthreads; +} + +static constexpr __device__ int ggml_cuda_mmq_get_nthreads(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).nthreads; +} + +static __host__ int ggml_cuda_mmq_get_occupancy(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).occupancy; +} + +static constexpr __device__ int ggml_cuda_mmq_get_occupancy(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).occupancy; +} + +static __host__ int ggml_cuda_mmq_get_I(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).I; +} + +static constexpr __device__ int ggml_cuda_mmq_get_I(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).I; +} + +static __host__ int ggml_cuda_mmq_get_J(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).J; +} + +static constexpr __device__ int ggml_cuda_mmq_get_J(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).J; +} + +static __host__ ggml_cuda_mmq_sram_layout ggml_cuda_mmq_get_sram_layout(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).sram_layout; +} + +static constexpr __device__ ggml_cuda_mmq_sram_layout ggml_cuda_mmq_get_sram_layout(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).sram_layout; +} + +static __host__ int ggml_cuda_mmq_get_K_vram(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).K_vram; +} + +static constexpr __device__ int ggml_cuda_mmq_get_K_vram(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).K_vram; +} + +static __host__ bool ggml_cuda_mmq_get_stream_k(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).stream_k; +} + +static constexpr __device__ bool ggml_cuda_mmq_get_stream_k(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).stream_k; +} + +static __host__ int ggml_cuda_mmq_get_fallback(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_config(type, J, fallback, cc).fallback; +} + +static constexpr __device__ int ggml_cuda_mmq_get_fallback(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).fallback; +} + +// --------------------------------------------------------------------------------------------- + +static __host__ int ggml_cuda_mmq_get_sram_stride(const ggml_type type, const int J, const bool fallback, const int cc) { + return ggml_cuda_mmq_get_sram_stride(ggml_cuda_mmq_get_sram_layout(type, J, fallback, cc)); +} + +static constexpr __device__ int ggml_cuda_mmq_get_sram_stride(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_sram_stride(ggml_cuda_mmq_get_sram_layout(type, J, fallback)); +} + +static __host__ int ggml_cuda_mmq_get_J_max(const ggml_type type, const bool fallback, const int cc, const int64_t ne11) { + int ret = std::min(ne11, int64_t(512)); + ret -= ret % 8; + for (;ret > 0; ret -= 8) { + if (ggml_cuda_mmq_get_config(type, ret, fallback, cc).type != GGML_TYPE_COUNT) { + return ret; + } + } + return ret; +} + +static constexpr __device__ int ggml_cuda_mmq_get_rows_per_warp(ggml_type type, int J, bool fallback) { + return ggml_cuda_mmq_get_config(type, J, fallback).rows_per_warp(); +} + +#define MMQ_DP4A_TXS_Q4_0 tile_x_sizes{I*MMQ_TILE_NE_K + I, I*MMQ_TILE_NE_K/QI4_0 + I/QI4_0, 0} +#define MMQ_DP4A_TXS_Q4_1 tile_x_sizes{I*MMQ_TILE_NE_K + I, I*MMQ_TILE_NE_K/QI4_1 + I/QI4_1, 0} +#define MMQ_DP4A_TXS_Q8_0 tile_x_sizes{I*MMQ_TILE_NE_K*2 + I, I*MMQ_TILE_NE_K*2/QI8_0 + I/(QI8_0/2), 0} +#define MMQ_DP4A_TXS_Q8_0_16 tile_x_sizes{I*MMQ_TILE_NE_K*2 + I, I*MMQ_TILE_NE_K*4/QI8_0 + I/(QI8_0/4), 0} +#define MMQ_DP4A_TXS_Q8_1 tile_x_sizes{I*MMQ_TILE_NE_K*2 + I, I*MMQ_TILE_NE_K*2/QI8_1 + I/(QI8_1/2), 0} +#define MMQ_DP4A_TXS_Q2_K tile_x_sizes{I*MMQ_TILE_NE_K*2 + I, I*MMQ_TILE_NE_K + I, 0} +#define MMQ_DP4A_TXS_Q3_K tile_x_sizes{I*MMQ_TILE_NE_K*2 + I, I, I*MMQ_TILE_NE_K/8 + I/8} +#define MMQ_DP4A_TXS_Q4_K tile_x_sizes{I*MMQ_TILE_NE_K + I, I*MMQ_TILE_NE_K/QI4_K, I*MMQ_TILE_NE_K/8 + I/8} +#define MMQ_DP4A_TXS_Q5_K tile_x_sizes{I*MMQ_TILE_NE_K*2 + I, I*MMQ_TILE_NE_K/QI5_K + I/QI5_K, I*MMQ_TILE_NE_K/8 + I/8} +#define MMQ_DP4A_TXS_Q6_K tile_x_sizes{I*MMQ_TILE_NE_K*2 + I, I*MMQ_TILE_NE_K/QI6_K + I/QI6_K, I*MMQ_TILE_NE_K/8 + I/8} + +static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml_type type, int I) { switch (type) { case GGML_TYPE_Q1_0: return MMQ_DP4A_TXS_Q8_0; case GGML_TYPE_Q4_0: return MMQ_DP4A_TXS_Q4_0; @@ -216,2982 +394,29 @@ static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml } } -#define MMQ_MMA_TILE_X_K_Q8_0 (2*MMQ_TILE_NE_K + 2*MMQ_TILE_NE_K/QI8_0 + 4) -#define MMQ_MMA_TILE_X_K_FP4 (2*MMQ_TILE_NE_K + 8 + 4) // MXFP4 and NVFP4 Blackwell -#define MMQ_MMA_TILE_X_K_NVFP4 (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/2 + 4) // NVFP4 Generic -#define MMQ_MMA_TILE_X_K_Q8_1 (2*MMQ_TILE_NE_K + 2*MMQ_TILE_NE_K/QI8_0 + 4) -#define MMQ_MMA_TILE_X_K_Q2_K (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K + 4) -#define MMQ_MMA_TILE_X_K_Q3_K (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/2 + 4) -#define MMQ_MMA_TILE_X_K_Q6_K (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/QI6_K + MMQ_TILE_NE_K/8 + 7) - -static_assert(MMQ_MMA_TILE_X_K_Q8_0 % 8 == 4, "Wrong padding."); -static_assert(MMQ_MMA_TILE_X_K_Q8_1 % 8 == 4, "Wrong padding."); -static_assert(MMQ_MMA_TILE_X_K_Q2_K % 8 == 4, "Wrong padding."); -static_assert(MMQ_MMA_TILE_X_K_Q3_K % 8 == 4, "Wrong padding."); -static_assert(MMQ_MMA_TILE_X_K_Q6_K % 8 == 4, "Wrong padding."); -static_assert(MMQ_MMA_TILE_X_K_FP4 % 8 == 4, "Wrong padding."); -static_assert(MMQ_MMA_TILE_X_K_FP4 == MMQ_MMA_TILE_X_K_Q8_1, "Wrong tile size for MXFP4"); -static_assert(MMQ_MMA_TILE_X_K_NVFP4 % 8 == 4, "Wrong padding."); - - -static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) { - switch (type) { - case GGML_TYPE_Q1_0: return MMQ_MMA_TILE_X_K_Q8_0; - case GGML_TYPE_Q4_0: return MMQ_MMA_TILE_X_K_Q8_0; - case GGML_TYPE_Q4_1: return MMQ_MMA_TILE_X_K_Q8_1; - case GGML_TYPE_Q5_0: return MMQ_MMA_TILE_X_K_Q8_0; - case GGML_TYPE_Q5_1: return MMQ_MMA_TILE_X_K_Q8_1; - case GGML_TYPE_Q8_0: return MMQ_MMA_TILE_X_K_Q8_0; - // tile sizes are the same for Q8_1 and FP4 for blackwell - case GGML_TYPE_MXFP4: return MMQ_MMA_TILE_X_K_Q8_1; -#if defined(BLACKWELL_MMA_AVAILABLE) - case GGML_TYPE_NVFP4: return MMQ_MMA_TILE_X_K_FP4; -#else - case GGML_TYPE_NVFP4: return MMQ_MMA_TILE_X_K_NVFP4; -#endif // defined(BLACKWELL_MMA_AVAILABLE) - case GGML_TYPE_Q2_K: return MMQ_MMA_TILE_X_K_Q2_K; - case GGML_TYPE_Q3_K: return MMQ_MMA_TILE_X_K_Q3_K; - case GGML_TYPE_Q4_K: return MMQ_MMA_TILE_X_K_Q8_1; - case GGML_TYPE_Q5_K: return MMQ_MMA_TILE_X_K_Q8_1; - case GGML_TYPE_Q6_K: return MMQ_MMA_TILE_X_K_Q6_K; - case GGML_TYPE_IQ2_XXS: return MMQ_MMA_TILE_X_K_Q8_0; - case GGML_TYPE_IQ2_XS: return MMQ_MMA_TILE_X_K_Q3_K; - case GGML_TYPE_IQ2_S: return MMQ_MMA_TILE_X_K_Q3_K; - case GGML_TYPE_IQ3_XXS: return MMQ_MMA_TILE_X_K_Q8_0; - case GGML_TYPE_IQ3_S: return MMQ_MMA_TILE_X_K_Q8_0; - case GGML_TYPE_IQ1_S: return MMQ_MMA_TILE_X_K_Q8_0; - case GGML_TYPE_IQ4_XS: return MMQ_MMA_TILE_X_K_Q8_0; - case GGML_TYPE_IQ4_NL: return MMQ_MMA_TILE_X_K_Q8_0; - default: return 0; +// FIXME temporary until all combinations of data types and GPUs can use the MMA data layout +static __host__ int ggml_cuda_mmq_get_nbytes_shared_x(const ggml_cuda_mmq_config & config, const int cc) { + if (config.use_mma_data_layout(cc)) { + return config.I * ggml_cuda_mmq_get_sram_stride(config.sram_layout) * 4; } -} - -// block_q8_1_mmq has (128 8-bit ints == 32 32-bit ints + 4 32-bit scales) -#define MMQ_TILE_Y_K (MMQ_TILE_NE_K + MMQ_TILE_NE_K / QI8_1) -#define MMQ_TILE_Y_FP4_K MMQ_TILE_Y_K - -static int mmq_get_granularity_host(const int mmq_x, const int cc) { - if (amd_mfma_available(cc) || amd_wmma_available(cc)) { - return mmq_x >= 128 ? 32 : 16; - } else if (turing_mma_available(cc) && mmq_x >= 48) { - return 16; - } else { - return 8; - } -} - -#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) -static constexpr __device__ int mmq_get_granularity_device(const int mmq_x) { - return mmq_x >= 128 ? 32 : 16; -} -#elif defined(TURING_MMA_AVAILABLE) -static constexpr __device__ int mmq_get_granularity_device(const int mmq_x) { - return mmq_x >= 48 ? 16 : 8; -} -#else -static constexpr __device__ int mmq_get_granularity_device(const int /*mmq_x*/) { - return 8; -} -#endif // AMD_MFMA_AVAILABLE - -#if defined(GGML_USE_HIP) -static int mmq_get_nwarps_host(const int cc, const int warp_size) { - return amd_mfma_available(cc) ? 8 : 256/warp_size; -} -#else -static int mmq_get_nwarps_host(const int /*cc*/, const int warp_size) { - return 256/warp_size; -} -#endif // (GGML_USE_HIP) - -static constexpr __device__ int mmq_get_nwarps_device() { -#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - return 8; -#else - return 256/ggml_cuda_get_physical_warp_size(); -#endif // AMD_MFMA_AVAILABLE + const tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(config.type, config.I); + return (txs.qs + txs.dm + txs.sc) * 4; } // ------------------------------------------------------------ -template static __device__ __forceinline__ void load_tiles_q1_0( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); +#include "mmq-load-tiles.cuh" +#include "mmq-vec-dot.cuh" -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int blocks_per_iter = MMQ_ITER_K / QK1_0; - constexpr int threads_per_row = blocks_per_iter * QI1_0; - constexpr int nrows = warp_size / threads_per_row; - constexpr int scale_entries_per_block = QK1_0 / QK8_1; - constexpr int scale_entries_per_row = blocks_per_iter * scale_entries_per_block; - - const int txi = threadIdx.x % threads_per_row; - const int kbx = txi / QI1_0; - const int kqsx = txi % QI1_0; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q1_0 * bxi = (const block_q1_0 *) x + kbx0 + i*stride + kbx; - const int qs_offset = 4*kqsx; - const int qs0 = bxi->qs[qs_offset + 0] | (bxi->qs[qs_offset + 1] << 8) | - (bxi->qs[qs_offset + 2] << 16) | (bxi->qs[qs_offset + 3] << 24); - - int unpacked_bytes[8]; -#pragma unroll - for (int j = 0; j < 8; ++j) { - const int shift = j * 4; - const int bits4 = (qs0 >> shift) & 0x0F; - const int b0 = (bits4 & 0x01) ? 1 : -1; - const int b1 = (bits4 & 0x02) ? 1 : -1; - const int b2 = (bits4 & 0x04) ? 1 : -1; - const int b3 = (bits4 & 0x08) ? 1 : -1; - unpacked_bytes[j] = (b0 & 0xFF) | ((b1 & 0xFF) << 8) | ((b2 & 0xFF) << 16) | ((b3 & 0xFF) << 24); - } - - const int dst_offset = kbx*(scale_entries_per_block*QI8_0) + kqsx*QI8_0; -#pragma unroll - for (int j = 0; j < 8; ++j) { -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + dst_offset + j] = unpacked_bytes[j]; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j] = unpacked_bytes[j]; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - } - - const int ksx = threadIdx.x % scale_entries_per_row; - const int scale_block = ksx / scale_entries_per_block; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + threadIdx.y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q1_0 * bxi = (const block_q1_0 *) x + kbx0 + i*stride + scale_block; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + ksx] = bxi->d; -#else - x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + ksx] = bxi->d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_q4_0( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_0); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - const int kbx = txi / QI4_0; - const int kqsx = txi % QI4_0; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbx; - const int qs0 = get_int_b2(bxi->qs, kqsx); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI4_0) + kqsx + 0] = __vsubss4((qs0 >> 0) & 0x0F0F0F0F, 0x08080808); - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI4_0) + kqsx + QI4_0] = __vsubss4((qs0 >> 4) & 0x0F0F0F0F, 0x08080808); -#else - x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) - } - - constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_0; - constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; - const int kbxd = threadIdx.x % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) { - int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbxd; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kbxd] = bxi->d; -#else - x_df[i*(MMQ_TILE_NE_K/QI4_0) + i/QI4_0 + kbxd] = bxi->d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template -static __device__ __forceinline__ void vec_dot_q4_0_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y); - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + txs.qs; - const int * y_qs = (const int *) y + 4; - const half2 * y_ds = (const half2 *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR4_0*VDR_Q4_0_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - const int kyqs = QI8_1 * ((k01/2) / (QI8_1/2)) + (k01/2) % (QI8_1/2); - - int u[2*VDR_Q4_0_Q8_1_MMQ]; - - constexpr int max_cpy = ggml_cuda_get_max_cpy_bytes(); - constexpr int mcpy_int = max_cpy / sizeof(int); - static_assert(VDR_Q4_0_Q8_1_MMQ == 4, "bad VDR_Q4_0_Q8_1_MMQ"); - - int tmp0[4], tmp1[4]; - - #pragma unroll - for (int l0 = 0; l0 < 4 / mcpy_int; ++l0) { - ggml_cuda_memcpy_1(tmp0 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + l0 * mcpy_int] ); - ggml_cuda_memcpy_1(tmp1 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + QI4_0 + l0 * mcpy_int]); - } - - u[0]=tmp0[0]; u[2]=tmp0[1]; u[4]=tmp0[2]; u[6]=tmp0[3]; - u[1]=tmp1[0]; u[3]=tmp1[1]; u[5]=tmp1[2]; u[7]=tmp1[3]; - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q4_0_q8_1_impl - (&x_qs[i*(MMQ_TILE_NE_K + 1) + k0/QR4_0], u, - x_df[i*(MMQ_TILE_NE_K/QI4_0) + i/QI4_0 + k0/(QR4_0*QI4_0)], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - } -} - -template static __device__ __forceinline__ void load_tiles_q4_1( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y); - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_1); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - const int kbx = txi / QI4_1; - const int kqsx = txi % QI4_1; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbx; - const int qs0 = get_int_b4(bxi->qs, kqsx); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI4_1) + kqsx + 0] = (qs0 >> 0) & 0x0F0F0F0F; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI4_1) + kqsx + QI4_1] = (qs0 >> 4) & 0x0F0F0F0F; -#else - x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_1; - constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; - const int kbxd = threadIdx.x % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) { - int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbxd; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + kbxd] = bxi->dm; -#else - x_dm[i*(MMQ_TILE_NE_K/QI4_1) + i/QI4_1 + kbxd] = bxi->dm; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template -static __device__ __forceinline__ void vec_dot_q4_1_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y); - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + txs.qs; - const int * y_qs = (const int *) y + 4; - const half2 * y_ds = (const half2 *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR4_1*VDR_Q4_1_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - const int kyqs = QI8_1 * ((k01/2) / (QI8_1/2)) + (k01/2) % (QI8_1/2); - - int u[2*VDR_Q4_1_Q8_1_MMQ]; - - constexpr int max_cpy = ggml_cuda_get_max_cpy_bytes(); - constexpr int mcpy_int = max_cpy / sizeof(int); - static_assert(VDR_Q4_0_Q8_1_MMQ == 4, "bad VDR_Q4_0_Q8_1_MMQ"); - - int tmp0[4], tmp1[4]; - - #pragma unroll - for (int l0 = 0; l0 < 4 / mcpy_int; ++l0) { - ggml_cuda_memcpy_1(tmp0 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + l0 * mcpy_int] ); - ggml_cuda_memcpy_1(tmp1 + l0 * mcpy_int, &y_qs[j*MMQ_TILE_Y_K + kyqs + QI4_1 + l0 * mcpy_int]); - } - - u[0]=tmp0[0]; u[2]=tmp0[1]; u[4]=tmp0[2]; u[6]=tmp0[3]; - u[1]=tmp1[0]; u[3]=tmp1[1]; u[5]=tmp1[2]; u[7]=tmp1[3]; - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q4_1_q8_1_impl - (&x_qs[i*(MMQ_TILE_NE_K + 1) + k0/QR4_1], u, - x_dm[i*(MMQ_TILE_NE_K/QI4_1) + i/QI4_1 + k0/(QR4_1*QI4_1)], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - } -} - -template static __device__ __forceinline__ void load_tiles_q5_0( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_0, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_0); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - const int kbx = txi / QI5_0; - const int kqsx = txi % QI5_0; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbx; - - const int ql = get_int_b2(bxi->qs, kqsx); - const int qh = get_int_b2(bxi->qh, 0) >> (4 * kqsx); - - int qs0 = (ql >> 0) & 0x0F0F0F0F; - qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 - qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 - qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 - qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 - qs0 = __vsubss4(qs0, 0x10101010); // subtract 16 - - int qs1 = (ql >> 4) & 0x0F0F0F0F; - qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 - qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 - qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 - qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 - qs1 = __vsubss4(qs1, 0x10101010); // subtract 16 - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI5_0) + kqsx + 0] = qs0; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_0) + kqsx + 0] = qs0; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI5_0; - constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; - const int kbxd = threadIdx.x % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) { - int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbxd; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kbxd] = bxi->d; -#else - x_df[i*(MMQ_TILE_NE_K/QI5_0) + i/QI5_0 + kbxd] = bxi->d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_q5_1( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y); - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_1); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - const int kbx = txi / QI5_1; - const int kqsx = txi % QI5_1; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbx; - - const int ql = get_int_b4(bxi->qs, kqsx); - const int qh = get_int_b4(bxi->qh, 0) >> (4 * kqsx); - - int qs0 = (ql >> 0) & 0x0F0F0F0F; - qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 - qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 - qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 - qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 - - int qs1 = (ql >> 4) & 0x0F0F0F0F; - qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 - qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 - qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 - qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI5_1) + kqsx + 0] = qs0; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_1) + kqsx + 0] = qs0; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI5_1; - constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; - const int kbxd = threadIdx.x % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) { - int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbxd; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + kbxd] = bxi->dm; -#else - x_dm[i*(MMQ_TILE_NE_K/QI5_1) + i/QI5_1 + kbxd] = bxi->dm; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_q8_0( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_tile + 2*MMQ_TILE_NE_K); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - // MMQ_ITER_K / (4 * QR8_0) == 64 required. but NV has only 32 threads per warp - constexpr int threads_per_row = 32; - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - const int kbx = txi / QI8_0; - const int kqsx = txi % QI8_0; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbx; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 0 + txi] = get_int_b2(bxi[0].qs, kqsx); - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + MMQ_TILE_NE_K + txi] = get_int_b2(bxi[MMQ_TILE_NE_K/QI8_0].qs, kqsx); -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 0 + txi] = get_int_b2(bxi[0].qs, kqsx); - x_qs[i*(2*MMQ_TILE_NE_K + 1) + MMQ_TILE_NE_K + txi] = get_int_b2(bxi[MMQ_TILE_NE_K/QI8_0].qs, kqsx); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - constexpr int blocks_per_tile_x_row = 2*MMQ_TILE_NE_K / QI8_0; - constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; - const int kbxd = threadIdx.x % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) { - int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbxd; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kbxd] = bxi->d; -#else - x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + kbxd] = bxi->d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_mxfp4( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_MXFP4, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR_MXFP4); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - const int kbx = txi / QI_MXFP4; - const int kqsx = txi % QI_MXFP4; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i*stride + kbx; - - const int aux_q4 = get_int_b1(bxi->qs, kqsx); - const int2 v = get_int_from_table_16(aux_q4, kvalues_mxfp4); - const int k0 = kbx * (2 * QI_MXFP4) + kqsx; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + k0 + 0] = v.x; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + k0 + QI_MXFP4] = v.y; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI_MXFP4] = v.y; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI_MXFP4; - constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; - const int kbxd = threadIdx.x % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) { - int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i*stride + kbxd; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_1 + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f; -#else - x_df[i*(MMQ_TILE_NE_K/QI_MXFP4) + i/QI_MXFP4 + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template -static __device__ __forceinline__ void load_tiles_mxfp4_fp4(const char * __restrict__ x, - int * __restrict__ x_tile, - const int kbx0, - const int i_max, - const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - int * x_qs = (int *) x_tile; - uint32_t * x_sc = (uint32_t *) (x_qs + 2 * MMQ_TILE_NE_K); - - const int txi = threadIdx.x; - - constexpr int iter_k = get_iter_k(GGML_TYPE_MXFP4); - - constexpr int threads_per_row = iter_k / QK_MXFP4; // each thread processes 1 block - constexpr int rows_per_warp = warp_size / threads_per_row; - const int kbx = txi % threads_per_row; - const int row_in_warp = txi / threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += rows_per_warp * nwarps) { - int i = i0 + threadIdx.y * rows_per_warp + row_in_warp; - - if constexpr (need_check) { - i = min(i, i_max); - } - - const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i * stride + kbx; - - // quantize_mxfp4_mmq permutes nibbles to match the quantized format - const int k0 = kbx * 4; - memcpy(x_qs + i * MMQ_MMA_TILE_X_K_FP4 + k0, bxi->qs, 16); - - // Load E8M0 scales: pack 2 consecutive scales into one uint32 - if (kbx % 2 == 0) { - uint32_t e = bxi->e; - e |= ((bxi + 1)->e << 8); - x_sc[i * MMQ_MMA_TILE_X_K_FP4 + kbx / 2] = e; - } - } -} - -#ifdef BLACKWELL_MMA_AVAILABLE -template -static __device__ __forceinline__ void load_tiles_nvfp4_nvfp4(const char * __restrict__ x, - int * __restrict__ x_tile, - const int kbx0, - const int i_max, - const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - constexpr int iter_k = get_iter_k(GGML_TYPE_NVFP4); - constexpr int threads_per_row = iter_k / QK_NVFP4; // each thread processes 1 block - constexpr int rows_per_warp = warp_size / threads_per_row; - - uint32_t * x_u32 = (uint32_t *) x_tile; - - const int txi = threadIdx.x; - const int kbx = txi % threads_per_row; - const int row_in_warp = txi / threads_per_row; - - const block_nvfp4 * bxi_base = (const block_nvfp4 *) x + kbx0 + kbx; - uint32_t * x_u32_scale = x_u32 + 64 + kbx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += rows_per_warp * nwarps) { - int i = i0 + threadIdx.y * rows_per_warp + row_in_warp; - - if constexpr (need_check) { - i = min(i, i_max); - } - - const block_nvfp4 * bxi = bxi_base + i * stride; - const int row_base = i * MMQ_MMA_TILE_X_K_FP4; - const int q_base = row_base + 8 * kbx; - - const uint32_t * src_qs = reinterpret_cast(bxi->qs); - -#pragma unroll - for (int sub = 0; sub < QK_NVFP4 / QK_NVFP4_SUB; ++sub) { - x_u32[q_base + 2 * sub + 0] = src_qs[2 * sub + 0]; - x_u32[q_base + 2 * sub + 1] = src_qs[2 * sub + 1]; - } - - x_u32_scale[row_base] = get_int_b4(bxi->d, 0); - } -} - -// Shared MMA kernel for MXFP4 and NVFP4 on Blackwell. -// Both quantizations encode values as e2m1 (FP4) and produce one uint32 scale per -// m16n8k64 MMA call; only the PTX kind (scale_vec::2X ue8m0 vs scale_vec::4X ue4m3) -// and the per-type stride constant differ. -template -static __device__ __forceinline__ void vec_dot_fp4_fp4_mma(const int * __restrict__ x, - const int * __restrict__ y, - float * __restrict__ sum, - const int k00) { - static_assert(type == GGML_TYPE_MXFP4 || type == GGML_TYPE_NVFP4, - "vec_dot_fp4_fp4_mma: type must be MXFP4 or NVFP4"); - - typedef tile<16, 8, int> tile_A; - typedef tile<8, 8, int> tile_B; - typedef tile<16, 8, float> tile_C; - - constexpr int stride = MMQ_MMA_TILE_X_K_FP4; - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp / tile_C::I; - constexpr int nfrags = MMQ_TILE_NE_K / tile_A::J; - - y += (threadIdx.y % ntx) * (tile_C::J * MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const uint32_t * x_sc = (const uint32_t *) (x_qs + 2 * MMQ_TILE_NE_K); - const int * y_qs = (const int *) y + 4; - const uint32_t * y_sc = (const uint32_t *) y; - - // 2 threads per quad supply the packed scale register to the block_scale MMA, - // see https://docs.nvidia.com/cuda/parallel-thread-execution/#warp-level-block-scaling - const int tidx_A = threadIdx.x / 4 + (threadIdx.x % 2) * 8; - const int tidx_B = threadIdx.x / 4; - const int i0 = (threadIdx.y / ntx) * rows_per_warp; - - tile_A A[ntx][nfrags]; - uint32_t scaleA[ntx][nfrags]; - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int frag = 0; frag < nfrags; ++frag) { - const int k0 = k00 + frag * tile_A::J; - load_ldmatrix(A[n][frag], x_qs + (i0 + n * tile_A::I) * stride + k0, stride); - scaleA[n][frag] = x_sc[(i0 + n * tile_A::I + tidx_A) * stride + k0 / tile_A::J]; - } - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx * tile_C::J) { - tile_B B[nfrags]; - uint32_t scaleB[nfrags]; - -#pragma unroll - for (int frag = 0; frag < nfrags; ++frag) { - const int k0 = frag * tile_B::J; - load_generic(B[frag], y_qs + j0 * MMQ_TILE_Y_K + k0, MMQ_TILE_Y_K); - scaleB[frag] = y_sc[(j0 + tidx_B) * MMQ_TILE_Y_K + frag]; - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int frag = 0; frag < nfrags; ++frag) { - tile_C C = {}; - mma_block_scaled_fp4(C, A[n][frag], B[frag], scaleA[n][frag], scaleB[frag]); -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - sum[(j0 / tile_C::J + n) * tile_C::ne + l] += C.x[l]; - } - } - } - } -} -#endif // BLACKWELL_MMA_AVAILABLE - - -template -static __device__ __forceinline__ void load_tiles_nvfp4(const char * __restrict__ x, - int * __restrict__ x_tile, - const int kb0, - const int i_max, - const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_NVFP4, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / QK_NVFP4; - constexpr int rows_per_warp = warp_size / threads_per_row; - const int kbx = threadIdx.x % threads_per_row; - const int row_in_warp = threadIdx.x / threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += rows_per_warp * nwarps) { - int i = i0 + threadIdx.y * rows_per_warp + row_in_warp; - - if constexpr (need_check) { - i = min(i, i_max); - } - - const block_nvfp4 * bxi = (const block_nvfp4 *) x + kb0 + i * stride + kbx; - const uint32_t * __restrict__ src_qs = reinterpret_cast(bxi->qs); - const int kqs = 16 * kbx; - const int ksc = 4 * kbx; - -#pragma unroll - for (int sub = 0; sub < QK_NVFP4 / QK_NVFP4_SUB; ++sub) { - const int2 q0 = get_int_from_table_16(src_qs[2 * sub + 0], kvalues_mxfp4); - const int2 q1 = get_int_from_table_16(src_qs[2 * sub + 1], kvalues_mxfp4); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i * MMQ_MMA_TILE_X_K_NVFP4 + kqs + 4 * sub + 0] = q0.x; - x_qs[i * MMQ_MMA_TILE_X_K_NVFP4 + kqs + 4 * sub + 1] = q1.x; - x_qs[i * MMQ_MMA_TILE_X_K_NVFP4 + kqs + 4 * sub + 2] = q0.y; - x_qs[i * MMQ_MMA_TILE_X_K_NVFP4 + kqs + 4 * sub + 3] = q1.y; - x_df[i * MMQ_MMA_TILE_X_K_NVFP4 + ksc + sub] = ggml_cuda_ue4m3_to_fp32(bxi->d[sub]); -#else - x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 0] = q0.x; - x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 1] = q1.x; - x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 2] = q0.y; - x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 3] = q1.y; - x_df[i * (2 * MMQ_TILE_NE_K * 2 / QI_NVFP4) + i / (QK_NVFP4_SUB / QI_NVFP4) + ksc + sub] = ggml_cuda_ue4m3_to_fp32(bxi->d[sub]); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - } -} - -template -static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y); - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + txs.qs; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += VDR_Q8_0_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q8_0_q8_1_impl - (&x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k0 % MMQ_TILE_NE_K], - x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + k0/QI8_0], y_df[j*MMQ_TILE_Y_K + (k0/QI8_1) % (MMQ_TILE_NE_K/QI8_1)]); - } - } - } -} - -template -static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { -#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr data_layout input_layout = get_input_data_layout(); - typedef tile<16, 8, int, input_layout> tile_A; - typedef tile<16, 8, int, input_layout> tile_B; - typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + 2*MMQ_TILE_NE_K; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - const half2 * y_ds = (const half2 *) y; - - const int i0 = (threadIdx.y / ntx) * rows_per_warp; - - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { - const int k0 = k00 + k01; - - tile_A A[ntx]; -#pragma unroll - for (int n = 0; n < ntx; ++n) { - load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_0 + k0, MMQ_MMA_TILE_X_K_Q8_0); - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { - tile_B B; - load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); - - float dB; - const int j = j0 + tile_C::get_j(0); - if (ds_layout == MMQ_Q8_1_DS_LAYOUT_D4) { - dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; - } else { - dB = __low2float(y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C C; - mma(C, A[n], B); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - const int i = i0 + n*tile_A::I + tile_C::get_i(l); - const float dA = x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + k0/QI8_0]; - sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l]*dA*dB; - } - } - } - } -#else - typedef tile<16, 8, int> tile_A; - typedef tile< 8, 8, int> tile_B; - typedef tile<16, 8, int> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + 2*MMQ_TILE_NE_K; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - const half2 * y_ds = (const half2 *) y; - - tile_A A[ntx][MMQ_TILE_NE_K/QI8_0]; - float dA[ntx][tile_C::ne/2][MMQ_TILE_NE_K/QI8_0]; - - const int i0 = (threadIdx.y/ntx)*rows_per_warp; - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { - const int k0 = k00 + k01; - - load_ldmatrix(A[n][k01/QI8_0], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_0 + k0, MMQ_MMA_TILE_X_K_Q8_0); - } - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int i = i0 + n*tile_A::I + tile_C::get_i(2*l); - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { - const int k0 = k00 + k01; - - dA[n][l][k01/QI8_0] = x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + k0/QI8_0]; - } - } - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { - tile_B B; - float dB[tile_C::ne/2]; - - load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); // faster than load_ldmatrix - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int j = j0 + tile_C::get_j(l); - - if (ds_layout == MMQ_Q8_1_DS_LAYOUT_D4) { - dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; - } else { - dB[l] = __low2float(y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C C; - mma(C, A[n][k01/QI8_0], B); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l]*dA[n][l/2][k01/QI8_0]*dB[l%2]; - } - } - } - } -#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) -} - - -template -static __device__ __forceinline__ void vec_dot_q8_1_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y); - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + txs.qs; - const int * y_qs = (const int *) y + 4; - const half2 * y_ds = (const half2 *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += VDR_Q8_0_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q8_1_q8_1_impl - (&x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], - x_dm[i*(MMQ_TILE_NE_K/QI5_1) + i/QI5_1 + k0/QI8_1], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - } -} - -template -static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { -#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr data_layout input_layout = get_input_data_layout(); - typedef tile<16, 8, int, input_layout> tile_A; - typedef tile<16, 8, int, input_layout> tile_B; - typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + 2*MMQ_TILE_NE_K; - const int * y_qs = (const int *) y + 4; - const half2 * y_dm = (const half2 *) y; - - const int i0 = (threadIdx.y / ntx) * rows_per_warp; - - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { - const int k0 = k00 + k01; - - tile_A A[ntx]; -#pragma unroll - for (int n = 0; n < ntx; ++n) { - load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_1 + k0, MMQ_MMA_TILE_X_K_Q8_1); - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { - tile_B B; - load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); - - const int j = j0 + tile_C::get_j(0); - const float2 dsB = __half22float2(y_dm[j*MMQ_TILE_Y_K + k01/QI8_1]); - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C C; - mma(C, A[n], B); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - const int i = i0 + n*tile_A::I + tile_C::get_i(l); - float2 dmA = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + k0/QI8_1]); - sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA.x*dsB.x*C.x[l]; - sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA.y*dsB.y; - } - } - } - } -#else - typedef tile<16, 8, int> tile_A; - typedef tile< 8, 8, int> tile_B; - typedef tile<16, 8, int> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + 2*MMQ_TILE_NE_K; - const int * y_qs = (const int *) y + 4; - const half2 * y_dm = (const half2 *) y; - - tile_A A[ntx][MMQ_TILE_NE_K/QI8_1]; - float2 dmA[ntx][tile_C::ne/2][MMQ_TILE_NE_K/QI8_1]; - - const int i0 = (threadIdx.y/ntx)*rows_per_warp; - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { - const int k0 = k00 + k01; - - load_ldmatrix(A[n][k01/QI8_1], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_1 + k0, MMQ_MMA_TILE_X_K_Q8_1); - } - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int i = i0 + n*tile_A::I + tile_C::get_i(2*l); - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { - const int k0 = k00 + k01; - - dmA[n][l][k01/QI8_1] = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + k0/QI8_1]); - } - } - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { - tile_B B; - float2 dsB[tile_C::ne/2]; - - load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); // faster than load_ldmatrix - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int j = j0 + tile_C::get_j(l); - - dsB[l] = __half22float2(y_dm[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C C; - mma(C, A[n][k01/QI8_1], B); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA[n][l/2][k01/QI8_1].x*dsB[l%2].x*C.x[l]; - sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA[n][l/2][k01/QI8_1].y*dsB[l%2].y; - } - } - } - } -#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) -} - -// Used for NVFP4, Q3_K, IQ2_S, and IQ2_XS -template -static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16; - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + txs.qs; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_0) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q8_0_16_q8_1_impl( - &x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], - &y_qs[j*MMQ_TILE_Y_K + k01], - &x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + k0/(QI8_0/2)], - y_df[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - } -} - -// Used for Q3_K, IQ2_S, and IQ2_XS: -template -static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { -#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr data_layout input_layout = get_input_data_layout(); - typedef tile<16, 4, int, input_layout> tile_A; - typedef tile<16, 4, int, input_layout> tile_B; - typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - - const int i0 = (threadIdx.y / ntx) * rows_per_warp; - - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { - const int k0 = k00 + k01; - - tile_A A[ntx]; -#pragma unroll - for (int n = 0; n < ntx; ++n) { - load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K); - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { - tile_B B; - load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); - - const int j = j0 + tile_C::get_j(0); - const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C C; - mma(C, A[n], B); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - const int i = i0 + n*tile_C::I + tile_C::get_i(l); - sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * x_df[i*MMQ_MMA_TILE_X_K_Q3_K + k0/4] * dB; - } - } - } - } -#elif defined(TURING_MMA_AVAILABLE) - - typedef tile<16, 4, int> tile_A; - typedef tile<16, 8, int> tile_A_8; - typedef tile< 8, 4, int> tile_B; - typedef tile<16, 8, int> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - - const int i0 = (threadIdx.y / ntx) * (ntx*tile_A::I); - - tile_A A[ntx][8]; - float dA[ntx][tile_C::ne/2][8]; - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 8) { - const int k0 = k00 + k01; - - load_ldmatrix(((tile_A_8 *) A[n])[k01/8], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K); - } - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { - const int k0 = k00 + k01; - - dA[n][l][k01/4] = x_df[i*MMQ_MMA_TILE_X_K_Q3_K + k0/4]; - } - } - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR3_K*VDR_Q3_K_Q8_1_MMQ) { - tile_B B[2]; - float dB[tile_C::ne/2]; - - // Here load_generic is faster than load_ldmatrix. - load_generic(B[0], y_qs + j0*MMQ_TILE_Y_K + (k01 + 0), MMQ_TILE_Y_K); - load_generic(B[1], y_qs + j0*MMQ_TILE_Y_K + (k01 + tile_B::J), MMQ_TILE_Y_K); - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int j = j0 + tile_C::get_j(l); - - dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C C[2]; - mma(C[0], A[n][k01/4 + 0], B[0]); - mma(C[1], A[n][k01/4 + 1], B[1]); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - sum[(j0/tile_C::J + n)*tile_C::ne + l] += dB[l%2]*(C[0].x[l]*dA[n][l/2][k01/4 + 0] + C[1].x[l]*dA[n][l/2][k01/4 + 1]); - } - } - } - } -#else - GGML_UNUSED_VARS(x, y, sum, k00); - NO_DEVICE_CODE; -#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE -} - -template static __device__ __forceinline__ void load_tiles_q2_K( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y); - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR2_K); - constexpr int nrows = ggml_cuda_get_physical_warp_size() / threads_per_row; - const int kqsx = threadIdx.x % threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q2_K * bxi = (const block_q2_K *) x + kbx0 + i*stride; - - const int x_ql_0 = get_int_b2(bxi->qs, kqsx); - -#pragma unroll - for (int l = 0; l < QR2_K; ++l) { - const int k = (kqsx/8)*32 + l*8 + kqsx % 8; - - const int x_qs_k = (x_ql_0 >> (2*l)) & 0x03030303; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q2_K + k] = x_qs_k; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + k] = x_qs_k; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - const int sc_m = bxi->scales[kqsx]; -#ifdef FAST_FP16_AVAILABLE - const half2 x_dm_ik = __hmul2(bxi->dm, make_half2(sc_m & 0x0F, sc_m >> 4)); -#else - const float2 bxi_dmf = __half22float2(bxi->dm); - const half2 x_dm_ik = make_half2(bxi_dmf.x*(sc_m & 0x0F), bxi_dmf.y*(sc_m >> 4)); -#endif // FAST_FP16_AVAILABLE - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + kqsx] = x_dm_ik; -#else - x_dm[i*(MMQ_TILE_NE_K + 1) + kqsx] = x_dm_ik; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template -static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y); - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + txs.qs; - const int * y_qs = (const int *) y + 4; - const half2 * y_ds = (const half2 *) y; - - float2 y_df[mmq_x/nwarps]; -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - - y_df[j0/nwarps] = __half22float2(y_ds[j*MMQ_TILE_Y_K]); - } - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K/2; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - constexpr int ns = 2; - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q2_K_q8_1_impl_mmq( - &x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], - &x_dm[i*(MMQ_TILE_NE_K + 1) + k0/4], k01 < MMQ_TILE_NE_K/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y, - &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); - } - } - } - - // Some compilers fail to unroll the loop over k01 if there is a conditional statement for ns in the inner loop. - // As a workaround 2 separate loops are used instead. -#pragma unroll - for (int k01 = MMQ_TILE_NE_K/2; k01 < MMQ_TILE_NE_K; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - constexpr int ns = 1; - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q2_K_q8_1_impl_mmq( - &x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], - &x_dm[i*(MMQ_TILE_NE_K + 1) + k0/4], k01 < MMQ_TILE_NE_K/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y, - &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); - } - } - } -} - -template -static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { -#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr data_layout input_layout = get_input_data_layout(); - typedef tile<16, 4, int, input_layout> tile_A; - typedef tile<16, 4, int, input_layout> tile_B; - typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + MMQ_TILE_NE_K*2; - const int * y_qs = (const int *) y + 4; - const half2 * y_ds = (const half2 *) y; - - const int i0 = (threadIdx.y / ntx) * rows_per_warp; - - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { - const int k0 = k00 + k01; - - tile_A A[ntx]; -#pragma unroll - for (int n = 0; n < ntx; ++n) { - load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K); - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { - tile_B B; - load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); - - const int j = j0 + tile_C::get_j(0); - const float dB = (k01 < MMQ_TILE_NE_K/2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K]).x : __half22float2(y_ds[j*MMQ_TILE_Y_K]).y; - const float sB = (k01 >= MMQ_TILE_NE_K * 3/4) ? 0 - : (((k01/4)%2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).y - : __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).x); - - tile_C Cm; - if (k01 >= MMQ_TILE_NE_K * 3/4) { - tile_A A1; -#pragma unroll - for (int l = 0; l < tile_A::ne; ++l) { - A1.x[l] = 0x01010101; - } - mma(Cm, A1, B); - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C Cd; - mma(Cd, A[n], B); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - const int i = i0 + n*tile_C::I + tile_C::get_i(l); - const float2 dm = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + k0/4]); - float tmp = Cd.x[l]*dm.x; - if (k01 >= MMQ_TILE_NE_K * 3/4) { - tmp -= Cm.x[l]*dm.y; - } - sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp*dB; - sum[(j0/tile_C::J + n)*tile_C::ne + l] -= dm.y*sB; - } - } - } - } -#elif defined(TURING_MMA_AVAILABLE) - - typedef tile<16, 4, int> tile_A; - typedef tile<16, 8, int> tile_A_8; - typedef tile< 8, 4, int> tile_B; - typedef tile<16, 8, int> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + MMQ_TILE_NE_K*2; - const int * y_qs = (const int *) y + 4; - const half2 * y_ds = (const half2 *) y; - - const int i0 = (threadIdx.y / ntx) * (ntx*tile_A::I); - - tile_A A[ntx][8]; - float dA[ntx][tile_C::ne/2][8]; - float mA[ntx][tile_C::ne/2][8]; - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { - const int k0 = k00 + k01; - - load_ldmatrix(((tile_A_8 *) A[n])[k01/QI8_1], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K); - } - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1/2) { - const int k0 = k00 + k01; - - const float2 dm = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + k0/(QI8_1/2)]); - - dA[n][l][k01/(QI8_1/2)] = dm.x; - mA[n][l][k01/(QI8_1/2)] = dm.y; - } - } - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { - float2 dB[tile_C::ne/2]; - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int j = j0 + tile_C::get_j(l); - - dB[l] = __half22float2(y_ds[j*MMQ_TILE_Y_K]); - } - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QI8_1) { - tile_B B[2]; - - // Here load_generic is faster than load_ldmatrix. - load_generic(B[0], y_qs + j0*MMQ_TILE_Y_K + (k01 + 0), MMQ_TILE_Y_K); - load_generic(B[1], y_qs + j0*MMQ_TILE_Y_K + (k01 + tile_B::J), MMQ_TILE_Y_K); - - tile_C Cm[2]; - if (k01 >= MMQ_TILE_NE_K * 3/4) { - tile_A A1; - A1.x[0] = 0x01010101; - A1.x[1] = 0x01010101; - mma(Cm[0], A1, B[0]); - mma(Cm[1], A1, B[1]); - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C Cd[2]; - - mma(Cd[0], A[n][k01/4 + 0], B[0]); - mma(Cd[1], A[n][k01/4 + 1], B[1]); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - float tmp = Cd[0].x[l]*dA[n][l/2][k01/4 + 0] + Cd[1].x[l]*dA[n][l/2][k01/4 + 1]; - if (k01 >= MMQ_TILE_NE_K * 3/4) { - tmp -= Cm[0].x[l]*mA[n][l/2][k01/4 + 0] + Cm[1].x[l]*mA[n][l/2][k01/4 + 1]; - } - sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp*(k01 < MMQ_TILE_NE_K/2 ? dB[l%2].x : dB[l%2].y); - } - } - } - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K * 3/4; k01 += QI8_1) { - float2 sB[tile_C::ne/2]; - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int j = j0 + tile_C::get_j(l); - - sB[l] = __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - sum[(j0/tile_C::J + n)*tile_C::ne + l] -= mA[n][l/2][k01/4 + 0]*sB[l%2].x; - sum[(j0/tile_C::J + n)*tile_C::ne + l] -= mA[n][l/2][k01/4 + 1]*sB[l%2].y; - } - } - } - } -#else - GGML_UNUSED_VARS(x, y, sum, k00); - NO_DEVICE_CODE; -#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE -} - -template static __device__ __forceinline__ void load_tiles_q3_K( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); - int * x_sc = (int *) (x_df + txs.dm); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR3_K); - constexpr int nrows = warp_size / threads_per_row; - const int kqsx = threadIdx.x % threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride; - - const int x_ql_0 = get_int_b2(bxi->qs, kqsx); - const int x_qh_0 = get_int_b2(bxi->hmask, kqsx % (QI3_K/2)) >> (4 * (kqsx / (QI3_K/2))); - -#pragma unroll - for (int l = 0; l < QR3_K; ++l) { - const int k = (kqsx/8)*32 + l*8 + kqsx % 8; - - const int x_ql_k = (x_ql_0 >> (2*l)) & 0x03030303; - const int x_qh_k = ((x_qh_0 >> l) << 2) & 0x04040404; - - const int x_qs_k = __vsubss4(x_ql_k | x_qh_k, 0x04040404); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + k] = x_qs_k; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + k] = x_qs_k; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - } - - constexpr int rows_per_warp = warp_size / 4; -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) { - int i = i0 + threadIdx.y*rows_per_warp + threadIdx.x/4; - - if (need_check) { - i = min(i, i_max); - } - - const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride; - - const int ksc = threadIdx.x % 4; - - const int ksc_low = ksc % (QI3_K/8); - const int shift_low = 4 * (ksc / (QI3_K/8)); - const int sc_low = (get_int_b2(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F; - - const int ksc_high = QI3_K/8; - const int shift_high = 2 * ksc; - const int sc_high = ((get_int_b2(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030; - - const int sc = __vsubss4(sc_low | sc_high, 0x20202020); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - const int8_t * sc8 = (const int8_t *) ≻ - const float d = bxi->d; - -#pragma unroll - for (int l = 0; l < int(sizeof(int)); ++l) { - x_df[i*MMQ_MMA_TILE_X_K_Q3_K + sizeof(int)*ksc + l] = d*sc8[l]; - } -#else - x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = sc; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - -#if !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)) -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*warp_size) { - int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride; - - x_df[i] = bxi->d; - } -#endif // !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)) || defined(AMD_WMMA_AVAILABLE) -} - -template -static __device__ __forceinline__ void vec_dot_q3_K_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, mmq_y); - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + txs.qs; - const int * x_sc = (const int *) x_df + txs.dm; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR3_K*VDR_Q3_K_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - const int8_t * scales = ((const int8_t *) (x_sc + i*(MMQ_TILE_NE_K/8) + i/8)) + k0/4; - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q3_K_q8_1_impl_mmq( - &x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], scales, - x_df[i], y_df[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - } -} - -static __device__ __forceinline__ int unpack_scales_q45_K(const int * scales, const int ksc) { - // scale arrangement after the following two lines: - // - ksc == 0: sc0, sc1, sc2, sc3 - // - ksc == 1: sc4, sc5, sc6, sc7 - // - ksc == 2: m0, m1, m2, m3 - // - ksc == 3: m4, m5, m6, m7 - return ((scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F) | // lower 4 bits - ((scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030); // upper 2 bits -} - -template static __device__ __forceinline__ void load_tiles_q4_K( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, mmq_y); - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + txs.qs); - int * x_sc = (int *) (x_dm + txs.dm); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_K); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride; - const int qs0 = get_int_b4(bxi->qs, txi); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 16*(txi/8) + txi % 8 + 0] = (qs0 >> 0) & 0x0F0F0F0F; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 16*(txi/8) + txi % 8 + 8] = (qs0 >> 4) & 0x0F0F0F0F; -#else - x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr int rows_per_warp = warp_size / 2; -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) { -#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - // Need if on AMD instead of % because warp_size == 64 - // This causes double work and throughput loss (MI300X) - // H100 loses about 100 t/s with 'if' condition over '%' - int i = i0 + threadIdx.y*rows_per_warp + threadIdx.x/2; - if (i < mmq_y) { -#else - int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/2) % mmq_y; - { -#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - if (need_check) { - i = min(i, i_max); - } - - const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride; - - const int * scales = (const int *) bxi->scales; - const int ksc = threadIdx.x % 2; - - const int sc32 = unpack_scales_q45_K(scales, ksc + 0); - const int m32 = unpack_scales_q45_K(scales, ksc + 2); - - const uint8_t * sc8 = (const uint8_t *) &sc32; - const uint8_t * m8 = (const uint8_t *) &m32; - - const half2 dm = bxi->dm * make_half2(1.0f, -1.0f); - - #pragma unroll - for (int l = 0; l < sizeof(int); ++l) { - x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]); - } - } - } -#else -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*warp_size) { - int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride; - - x_dm[i] = bxi->dm; - } - constexpr int rows_per_warp = warp_size / 4; -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) { - int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/(MMQ_TILE_NE_K/8)) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride + (threadIdx.x % (MMQ_TILE_NE_K/8)) / (QI4_K/8); - - const int * scales = (const int *) bxi->scales; - - const int ksc = threadIdx.x % (MMQ_TILE_NE_K/8); - const int scales8 = unpack_scales_q45_K(scales, ksc); - - x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = scales8; - } -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) -} - -template -static __device__ __forceinline__ void vec_dot_q4_K_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, mmq_y); - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + txs.qs; - const int * x_sc = (const int *) x_dm + txs.dm; - const int * y_qs = (const int *) y + 4; - const half2 * y_ds = (const half2 *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR4_K*VDR_Q4_K_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - const uint8_t * sc = (const uint8_t *) &x_sc[i * (MMQ_TILE_NE_K/8) + i/8 + k0/32] + 2*(k01/16); - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q4_K_q8_1_impl_mmq( - &x_qs[i*(MMQ_TILE_NE_K + 1) + k0/2], &y_qs[j*MMQ_TILE_Y_K + k01], sc, sc+8, - x_dm[i], &y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - } -} - -template static __device__ __forceinline__ void load_tiles_q5_K( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, mmq_y); - int * x_qs = (int *) x_tile; - half2 * x_dm = (half2 *) (x_qs + txs.qs); - int * x_sc = (int *) (x_dm + txs.dm); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_K); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride; - const int ky = QR5_K*txi; - - const int ql = get_int_b4(bxi->qs, txi); - const int ql0 = (ql >> 0) & 0x0F0F0F0F; - const int ql1 = (ql >> 4) & 0x0F0F0F0F; - - const int qh = get_int_b4(bxi->qh, txi % (QI5_K/4)); - const int qh0 = ((qh >> (2 * (txi / (QI5_K/4)) + 0)) << 4) & 0x10101010; - const int qh1 = ((qh >> (2 * (txi / (QI5_K/4)) + 1)) << 4) & 0x10101010; - - const int kq0 = ky - ky % (QI5_K/2) + txi % (QI5_K/4) + 0; - const int kq1 = ky - ky % (QI5_K/2) + txi % (QI5_K/4) + QI5_K/4; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kq0] = ql0 | qh0; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kq1] = ql1 | qh1; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq0] = ql0 | qh0; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq1] = ql1 | qh1; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr int rows_per_warp = warp_size / 2; -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) { -#if defined(AMD_MFMA_AVAILABLE) - // Need if on AMD instead of % because warp_size == 64 - // This causes double work and throughput loss (MI300X) - // H100 loses about 100 t/s with 'if' condition over '%' - int i = i0 + threadIdx.y*rows_per_warp + threadIdx.x/2; - if (i < mmq_y) { -#else - int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/2) % mmq_y; - { -#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - if (need_check) { - i = min(i, i_max); - } - - const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride; - - const int * scales = (const int *) bxi->scales; - const int ksc = threadIdx.x % 2; - - const int sc32 = unpack_scales_q45_K(scales, ksc + 0); - const int m32 = unpack_scales_q45_K(scales, ksc + 2); - - const uint8_t * sc8 = (const uint8_t *) &sc32; - const uint8_t * m8 = (const uint8_t *) &m32; - - const half2 dm = bxi->dm * make_half2(1.0f, -1.0f); - -#pragma unroll - for (int l = 0; l < int(sizeof(int)); ++l) { - x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]); - } - } - } -#else -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*warp_size) { - int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride; - - x_dm[i] = bxi->dm; - } - - constexpr int rows_per_warp = warp_size / 4; -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) { - int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/(MMQ_TILE_NE_K/8)) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride; - - const int * scales = (const int *) bxi->scales; - - const int ksc = threadIdx.x % (MMQ_TILE_NE_K/8); - const int scales8 = unpack_scales_q45_K(scales, ksc); - - x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = scales8; - } -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) -} - -template -static __device__ __forceinline__ void vec_dot_q5_K_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, mmq_y); - const int * x_qs = (const int *) x; - const half2 * x_dm = (const half2 *) x_qs + txs.qs; - const int * x_sc = (const int *) x_dm + txs.dm; - const int * y_qs = (const int *) y + 4; - const half2 * y_ds = (const half2 *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR5_K*VDR_Q5_K_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - const uint8_t * sc = ((const uint8_t *) &x_sc[i * (MMQ_TILE_NE_K/8) + i/8 + k00/32]) + 2*(k01/16); - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q5_K_q8_1_impl_mmq( - &x_qs[i*(QR5_K*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], sc, sc+8, - x_dm[i], &y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - } -} - -template static __device__ __forceinline__ void load_tiles_q6_K( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); - int * x_sc = (int *) (x_df + MMQ_TILE_NE_K/QI6_K); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); - int * x_sc = (int *) (x_df + txs.dm); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR6_K); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride; - - const int ql = get_int_b2(bxi->ql, txi); - const int ql0 = (ql >> 0) & 0x0F0F0F0F; - const int ql1 = (ql >> 4) & 0x0F0F0F0F; - - const int qh = get_int_b2(bxi->qh, (QI6_K/4) * (txi / (QI6_K/2)) + txi % (QI6_K/4)); - const int qh0 = ((qh >> ((txi & 0x08) >> 2)) << 4) & 0x30303030; - const int qh1 = (qh >> ((txi & 0x08) >> 2)) & 0x30303030; - - const int kq0 = 2*txi - txi % (QI6_K/2) + 0; - const int kq1 = 2*txi - txi % (QI6_K/2) + QI6_K/2; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q6_K + kq0] = __vsubss4(ql0 | qh0, 0x20202020); - x_qs[i*MMQ_MMA_TILE_X_K_Q6_K + kq1] = __vsubss4(ql1 | qh1, 0x20202020); -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq0] = __vsubss4(ql0 | qh0, 0x20202020); - x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq1] = __vsubss4(ql1 | qh1, 0x20202020); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*warp_size) { - int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q6_K] = bxi->d; -#else - x_df[i*(MMQ_TILE_NE_K/QI6_K) + i/QI6_K] = bxi->d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - constexpr int rows_per_warp = warp_size / 4; -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) { - int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/(MMQ_TILE_NE_K/8)) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + (threadIdx.x % (MMQ_TILE_NE_K/8)) / 4; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_sc[i*MMQ_MMA_TILE_X_K_Q6_K + threadIdx.x%4] = get_int_b2(bxi->scales, threadIdx.x % (MMQ_TILE_NE_K/8)); -#else - x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + threadIdx.x%(MMQ_TILE_NE_K/8)] = get_int_b2(bxi->scales, threadIdx.x%(QI6_K/8)); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template -static __device__ __forceinline__ void vec_dot_q6_K_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, mmq_y); - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + txs.qs; - const int * x_sc = (const int *) x_df + txs.dm; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - -// #pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += QR6_K*VDR_Q6_K_Q8_1_MMQ) { - const int k0 = k00 + k01; - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { - const int j = j0 + threadIdx.y; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { - const int i = i0 + threadIdx.x; - - const int8_t * sc = ((const int8_t *) &x_sc[i * (MMQ_TILE_NE_K/8) + i/8 + k0/16]); - - sum[j0/nwarps*mmq_y/warp_size + i0/warp_size] += vec_dot_q6_K_q8_1_impl_mmq( - &x_qs[i*(QR6_K*MMQ_TILE_NE_K + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], sc, - x_df[i*(MMQ_TILE_NE_K/QI6_K) + i/QI6_K], &y_df[j*MMQ_TILE_Y_K + k01/QI8_1]); - } - } - } -} - -template -static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { -#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr data_layout input_layout = get_input_data_layout(); - typedef tile<16, 4, int, input_layout> tile_A; - typedef tile<16, 4, int, input_layout> tile_B; - typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; - const int * x_sc = (const int *) x_df + MMQ_TILE_NE_K/QI6_K; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - - const int i0 = (threadIdx.y / ntx) * rows_per_warp; - - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { - const int k0 = k00 + k01; - - tile_A A[ntx]; -#pragma unroll - for (int n = 0; n < ntx; ++n) { - load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + k0, MMQ_MMA_TILE_X_K_Q6_K); - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { - tile_B B; - load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); - - const int j = j0 + tile_C::get_j(0); - const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C C; - mma(C, A[n], B); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - const int i = i0 + n*tile_C::I + tile_C::get_i(l); - const int8_t * sc = (const int8_t *) (x_sc + i*MMQ_MMA_TILE_X_K_Q6_K + k00/16); - sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * sc[k01/4] * x_df[i*MMQ_MMA_TILE_X_K_Q6_K] * dB; - } - } - } - } -#elif defined(TURING_MMA_AVAILABLE) - - typedef tile<16, 4, int> tile_A; - typedef tile< 8, 4, int> tile_B; - typedef tile<16, 8, int> tile_C; - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - - y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); - - const int * x_qs = (const int *) x; - const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; - const int * x_sc = (const int *) x_df + MMQ_TILE_NE_K/QI6_K; - const int * y_qs = (const int *) y + 4; - const float * y_df = (const float *) y; - - const int i0 = (threadIdx.y / ntx) * (ntx*tile_A::I); - - tile_A A[ntx][8]; - int scA[ntx][tile_C::ne/2][8]; - float dA[ntx][tile_C::ne/2]; - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 8) { - const int k0 = k00 + k01; - - load_ldmatrix(A[n][k01/4 + 0], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + (k0 + 0), MMQ_MMA_TILE_X_K_Q6_K); - load_ldmatrix(A[n][k01/4 + 1], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + (k0 + tile_A::J), MMQ_MMA_TILE_X_K_Q6_K); - } - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 16) { - const int k0 = k00 + k01; - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); - - const int sc_packed = x_sc[i*MMQ_MMA_TILE_X_K_Q6_K + k0/16]; - const int8_t * sc = (const int8_t *) &sc_packed; - -#pragma unroll - for (int ksc = 0; ksc < sizeof(int); ++ksc) { - scA[n][l][k01/4 + ksc] = sc[ksc]; - } - } - } - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); - - dA[n][l] = x_df[i*MMQ_MMA_TILE_X_K_Q6_K]; - } - } - -#pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { - float tmp[ntx][tile_C::ne] = {{0.0f}}; - -#pragma unroll - for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 8) { - tile_B B[2]; - float dB[tile_C::ne/2]; - - // Here load_generic is faster than load_ldmatrix. - load_generic(B[0], y_qs + j0*MMQ_TILE_Y_K + 0 + k01, MMQ_TILE_Y_K); - load_generic(B[1], y_qs + j0*MMQ_TILE_Y_K + tile_B::J + k01, MMQ_TILE_Y_K); - -#pragma unroll - for (int l = 0; l < tile_C::ne/2; ++l) { - const int j = j0 + tile_C::get_j(l); - - dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { - tile_C C[2]; - mma(C[0], A[n][k01/4 + 0], B[0]); - mma(C[1], A[n][k01/4 + 1], B[1]); - -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - tmp[n][l] += (C[0].x[l]*scA[n][l/2][k01/4 + 0] + C[1].x[l]*scA[n][l/2][k01/4 + 1])*dB[l%2]; - } - } - } - -#pragma unroll - for (int n = 0; n < ntx; ++n) { -#pragma unroll - for (int l = 0; l < tile_C::ne; ++l) { - sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp[n][l]*dA[n][l/2]; - } - } - } -#else - GGML_UNUSED_VARS(x, y, sum, k00); - NO_DEVICE_CODE; -#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE -} - -template static __device__ __forceinline__ void load_tiles_iq4_nl( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_NL, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_NL); - constexpr int nrows = warp_size / threads_per_row; - const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - const int kbx = txi / QI4_NL; - const int kqsx = txi % QI4_NL; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbx; - - const int aux_q4 = get_int_b2(bxi->qs, kqsx); - const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl); - const int k0 = kbx * (2 * QI4_NL) + kqsx; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 0] = v.x; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + QI4_NL] = v.y; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI4_NL] = v.y; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_NL; - constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row; - const int kbxd = threadIdx.x % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) { - int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbxd; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kbxd] = __half2float(bxi->d); -#else - x_df[i*(MMQ_TILE_NE_K/QI4_NL) + i/QI4_NL + kbxd] = __half2float(bxi->d); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_iq2_xxs( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_XXS, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_XXS)) / 2; - constexpr int nrows = warp_size / threads_per_row; - const int kqsx = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * nrows) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_iq2_xxs * bxi = (const block_iq2_xxs *) x + kbx0 + i*stride; - - const int q2 = get_int_b2(bxi->qs, 2*kqsx+0); - const uint8_t * aux8 = (const uint8_t *) &q2; - const uint32_t aux32 = get_int_b2(bxi->qs, 2*kqsx+1); - -#pragma unroll - for (int l = 0; l < QR2_XXS; ++l) { - const uint2 grid_pos = ((const uint2*)iq2xxs_grid)[aux8[l]]; - const uint32_t signs = unpack_ksigns(aux32 >> (7 * l)); - - const int signs0 = __vcmpne4(signs & 0x08040201, 0); - const int grid0 = __vsub4(grid_pos.x ^ signs0, signs0); - - const int signs1 = __vcmpne4(signs & 0x80402010, 0); - const int grid1 = __vsub4(grid_pos.y ^ signs1, signs1); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid0; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 1)] = grid1; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid0; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid1; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - const int ls = aux32 >> 27 | 1; // (scale * 2 + 1) - const float d = bxi->d; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx] = d * ls / 8; // (d * scale + d / 2) / 4 -#else - x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = d * ls / 8; // (d * scale + d / 2) / 4 -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_iq2_xs( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16; - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_XS)) / 2; - constexpr int nrows = warp_size / threads_per_row; - const int kqsx = threadIdx.x % threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * nrows) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_iq2_xs * bxi = (const block_iq2_xs *) x + kbx0 + i*stride; - - const int2 q2_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1)); - const uint16_t * q2 = (const uint16_t *) &q2_packed; - - #pragma unroll - for (int l = 0; l < QR2_XS; ++l) { - const uint2 grid_pos = ((const uint2*)iq2xs_grid)[q2[l] & 0x1FF]; - const uint32_t signs = unpack_ksigns(q2[l] >> 9); - - const int signs0 = __vcmpne4(signs & 0x08040201, 0); - const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0); - - const int signs1 = __vcmpne4(signs & 0x80402010, 0); - const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l; - x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 1)] = grid_h; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - const int ls = bxi->scales[kqsx]; - const float d = bxi->d; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; - x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; -#else - x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; - x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_iq2_s( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_S, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_S)) / 2; - constexpr int nrows = warp_size / threads_per_row; - const int kqsx = threadIdx.x % threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * nrows) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_iq2_s * bxi = (const block_iq2_s *) x + kbx0 + i*stride; - - const int qs_packed = get_int_b2(bxi->qs, kqsx); - const uint8_t * qs = (const uint8_t *) &qs_packed; - - const int qh = bxi->qh[kqsx]; - - const int signs_packed_32 = get_int_b2(bxi->qs, QK_K/32 + kqsx); - const uint8_t * signs_packed_8 = (const uint8_t *) &signs_packed_32; - -#pragma unroll - for (int l = 0; l < QR2_S; ++l) { - const int * grid_pos = (const int *)(iq2s_grid + (qs[l] | ((qh << (8-2*l)) & 0x300))); - - const int signs0 = __vcmpne4(((signs_packed_8[l] & 0x03) << 7) | ((signs_packed_8[l] & 0x0C) << 21), 0x00000000); - const int signs1 = __vcmpne4(((signs_packed_8[l] & 0x30) << 3) | ((signs_packed_8[l] & 0xC0) << 17), 0x00000000); - - const int grid_l = __vsub4(grid_pos[0] ^ signs0, signs0); - const int grid_h = __vsub4(grid_pos[1] ^ signs1, signs1); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l; - x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 1)] = grid_h; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - const int ls = bxi->scales[kqsx]; - const float d = bxi->d; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; - x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; -#else - x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; - x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_iq3_xxs( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_XXS, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR3_XXS)) / 2; - constexpr int nrows = warp_size / threads_per_row; - const int kqsx = threadIdx.x % threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * nrows) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_iq3_xxs * bxi = (const block_iq3_xxs *) x + kbx0 + i*stride; - - const int2 q3_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1)); - const uint8_t * q3 = (const uint8_t *) &q3_packed; - const uint32_t aux32 = get_int_b2(bxi->qs, QK_K/16 + kqsx); - -#pragma unroll - for (int l = 0; l < QR3_XXS; ++l) { - const int2 grid_pos = make_int2(iq3xxs_grid[q3[2*l+0]], iq3xxs_grid[q3[2*l+1]]); - const uint32_t signs = unpack_ksigns(aux32 >> (7*l)); - - const int signs0 = __vcmpne4(signs & 0x08040201, 0); - const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0); - - const int signs1 = __vcmpne4(signs & 0x80402010, 0); - const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid_l; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 1)] = grid_h; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - const int ls = aux32 >> 28; - const float d = bxi->d; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx] = (ls*d + d/2)/2; -#else - x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = (ls*d + d/2)/2; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_iq3_s( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR3_S)) / 2; - constexpr int nrows = warp_size / threads_per_row; - const int kqsx = threadIdx.x % threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * nrows) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_iq3_s * bxi = (const block_iq3_s *) x + kbx0 + i*stride; - - const int2 qs_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1)); - const uint8_t * qs = (const uint8_t *) &qs_packed; - - const int qh = bxi->qh[kqsx]; - - const int signs_packed_32 = get_int_b2(bxi->signs, kqsx); - const uint8_t * signs_packed_8 = (const uint8_t *) &signs_packed_32; - -#pragma unroll - for (int l = 0; l < QR3_S; ++l) { - const int2 grid_pos = make_int2( - iq3s_grid[qs[2*l+0] | ((qh << (8 - 2*l)) & 0x100)], - iq3s_grid[qs[2*l+1] | ((qh << (7 - 2*l)) & 0x100)]); - - const int signs0 = __vcmpne4(((signs_packed_8[l] & 0x03) << 7) | ((signs_packed_8[l] & 0x0C) << 21), 0x00000000); - const int signs1 = __vcmpne4(((signs_packed_8[l] & 0x30) << 3) | ((signs_packed_8[l] & 0xC0) << 17), 0x00000000); - - const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0); - const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l+0)] = grid_l; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l+1)] = grid_h; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+0)] = grid_l; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+1)] = grid_h; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - const int ls = 1 + 2*((bxi->scales[kqsx/2] >> (((2*kqsx) << 1) & 0x04)) & 0x0F); - const float d = bxi->d; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx] = ls*d; -#else - x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = ls*d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_iq1_s( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - half2 * x_ds = (half2 *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, mmq_y); - int * x_qs = (int *) x_tile; - half2 * x_ds = (half2 *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR1_S); - constexpr int nrows = warp_size / threads_per_row; - const int kqsx = threadIdx.x % threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * nrows) { - int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_iq1_s * bxi = (const block_iq1_s *) x + kbx0 + i*stride; - - const int qs_packed = get_int_b2(bxi->qs, kqsx); - const uint8_t * qs = (const uint8_t *) &qs_packed; - - const int qh = bxi->qh[kqsx]; - - #pragma unroll - for (int l = 0; l < QR1_S/2; ++l) { - const int grid = iq1s_grid_gpu[qs[l] | (((qh >> (3*l)) & 0x07) << 8)]; - - const int grid0 = (grid >> 0) & 0x0F0F0F0F; - const int grid1 = (grid >> 4) & 0x0F0F0F0F; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 8*kqsx + (2*l+0)] = grid0; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 8*kqsx + (2*l+1)] = grid1; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+0)] = grid0; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+1)] = grid1; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - const float d1q = __half2float(bxi->d) * (((qh >> 11) & 0x0E) + 1); - const float delta = -1.0f + IQ1S_DELTA - (qh & 0x8000) * (2.0f*IQ1S_DELTA/0x8000); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_ds[i*MMQ_MMA_TILE_X_K_Q8_1 + kqsx] = make_half2(d1q, d1q*delta); -#else - x_ds[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = make_half2(d1q, d1q*delta); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template static __device__ __forceinline__ void load_tiles_iq4_xs( - const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { - constexpr int nwarps = mmq_get_nwarps_device(); - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); -#else - constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_XS, mmq_y); - int * x_qs = (int *) x_tile; - float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - - constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_XS); - constexpr int nrows = warp_size / threads_per_row; - const int kqsx = threadIdx.x % threads_per_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) { - int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row); - - if (need_check) { - i = min(i, i_max); - } - - const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride; - - const int aux_q4 = get_int_b4(bxi->qs, kqsx); - const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl); - const int k0 = 8 * (kqsx / 4) + kqsx % 4; - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 0] = v.x; - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 4] = v.y; -#else - x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x; - x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 4] = v.y; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } - - constexpr int rows_per_warp = warp_size / 8; -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) { - int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / (MMQ_TILE_NE_K/4); - - if (need_check) { - i = min(i, i_max); - } - - const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride; - - const float d = __half2float(bxi->d); - - const int ls = ((bxi->scales_l[(threadIdx.x % 8)/2] >> (4*(threadIdx.x % 2))) & 0x0F) - | (((bxi->scales_h >> (2*(threadIdx.x % 8))) & 0x03) << 4); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + threadIdx.x % 8] = d * (ls - 32); -#else - x_df[i*(MMQ_TILE_NE_K/4) + i/4 + threadIdx.x % 8] = d * (ls - 32); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - } -} - -template -static __device__ __forceinline__ void mmq_write_back_dp4a( +template static __device__ __forceinline__ void ggml_cuda_mmq_write_back_dp4a( const float * __restrict__ sum, const int32_t * __restrict__ ids_dst, float * __restrict__ dst, const int stride, const int i_max, const int j_max) { - constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { + for (int j0 = 0; j0 < J; j0 += nwarps) { const int j = j0 + threadIdx.y; if (j > j_max) { @@ -3199,45 +424,38 @@ static __device__ __forceinline__ void mmq_write_back_dp4a( } #pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += warp_size) { + for (int i0 = 0; i0 < I; i0 += warp_size) { const int i = i0 + threadIdx.x; - if (need_check && i > i_max) { + if (fallback && i > i_max) { continue; } - dst[ids_dst[j]*stride + i] = sum[(j0/nwarps) * (mmq_y/warp_size) + i0/warp_size]; + dst[ids_dst[j]*stride + i] = sum[(j0/nwarps) * (I/warp_size) + i0/warp_size]; } } } -template -static __device__ __forceinline__ void mmq_write_back_mma( - const float * __restrict__ sum, const int * __restrict__ ids_dst, float * __restrict__ dst, - const int stride, const int i_max, const int j_max) { - - constexpr int granularity = mmq_get_granularity_device(mmq_x); - constexpr int nwarps = mmq_get_nwarps_device(); - +template +static __device__ __forceinline__ void ggml_cuda_mmq_write_back_mma( + const float * __restrict__ sum, const int * __restrict__ ids_dst, float * __restrict__ dst, + const int stride, const int i_max, const int j_max) { #if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr int tileC_IJ = mmq_get_granularity_device(0); - typedef tile tile_C; - constexpr int rows_per_warp = granularity; + typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C; #else - typedef tile<16, 8, int> tile_C; - constexpr int rows_per_warp = 2 * granularity; -#endif // defined(AMD_MFMA_AVAILABLE) - constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + typedef tile<16, 8, int> tile_C; +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback); + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. const int i0 = (threadIdx.y / ntx) * (ntx*tile_C::I); -#if defined(TURING_MMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - static_assert(nwarps*tile_C::I == mmq_y, "nwarps*tile_C::I != mmq_y"); -#else - GGML_UNUSED(nwarps); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { + for (int j0 = 0; j0 < J; j0 += ntx*tile_C::J) { #pragma unroll for (int n = 0; n < ntx; ++n) { #pragma unroll @@ -3250,7 +468,7 @@ static __device__ __forceinline__ void mmq_write_back_mma( const int i = i0 + n*tile_C::I + tile_C::get_i(l); - if (need_check && i > i_max) { + if (fallback && i > i_max) { continue; } @@ -3262,188 +480,339 @@ static __device__ __forceinline__ void mmq_write_back_mma( // ------------------------------------------------------------------------------------------------------------------------------------- -template -struct mmq_type_traits; +// TODO remove this struct and use ggml_cuda_mmq_sram_layout instead. +struct ggml_cuda_mmq_util_funcs { + int vdr; + ggml_cuda_mmq_load_tiles_t load_tiles; + ggml_cuda_mmq_vec_dot_t vec_dot; + ggml_cuda_mmq_write_back_t write_back; -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q1_0_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q1_0; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; + constexpr __host__ __device__ ggml_cuda_mmq_util_funcs( + int vdr, ggml_cuda_mmq_load_tiles_t load_tiles, ggml_cuda_mmq_vec_dot_t vec_dot, ggml_cuda_mmq_write_back_t write_back) : + vdr(vdr), load_tiles(load_tiles), vec_dot(vec_dot), write_back(write_back) {} }; -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q4_0_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q4_0; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q4_0_q8_1_dp4a; -}; +template +static constexpr __device__ ggml_cuda_mmq_util_funcs ggml_cuda_mmq_get_util_funcs() { + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q4_1_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q4_1; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_1_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q4_1_q8_1_dp4a; -}; + if (!ggml_cuda_mmq_get_config(type, J, fallback).use_mma_data_layout()) { + switch (type) { + case GGML_TYPE_Q1_0: + return ggml_cuda_mmq_util_funcs( + VDR_Q1_0_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q1_0, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q4_0: + return ggml_cuda_mmq_util_funcs( + VDR_Q4_0_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q4_0, + ggml_cuda_mmq_vec_dot_q4_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q4_1: + return ggml_cuda_mmq_util_funcs( + VDR_Q4_1_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q4_1, + ggml_cuda_mmq_vec_dot_q4_1_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q5_0: + return ggml_cuda_mmq_util_funcs( + VDR_Q5_0_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q5_0, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q5_1: + return ggml_cuda_mmq_util_funcs( + VDR_Q5_1_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q5_1, + ggml_cuda_mmq_vec_dot_q8_1_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q8_0: + return ggml_cuda_mmq_util_funcs( + VDR_Q8_0_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q8_0, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); +// --------------------------------------------------------------------------------------------- + case GGML_TYPE_Q2_K: + return ggml_cuda_mmq_util_funcs( + VDR_Q2_K_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q2_K, + ggml_cuda_mmq_vec_dot_q2_K_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q3_K: + return ggml_cuda_mmq_util_funcs( + VDR_Q3_K_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q3_K, + ggml_cuda_mmq_vec_dot_q3_K_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q4_K: + return ggml_cuda_mmq_util_funcs( + VDR_Q4_K_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q4_K, + ggml_cuda_mmq_vec_dot_q4_K_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q5_K: + return ggml_cuda_mmq_util_funcs( + VDR_Q5_K_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q5_K, + ggml_cuda_mmq_vec_dot_q5_K_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_Q6_K: + return ggml_cuda_mmq_util_funcs( + VDR_Q6_K_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_q6_K, + ggml_cuda_mmq_vec_dot_q6_K_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); +// --------------------------------------------------------------------------------------------- + case GGML_TYPE_IQ1_S: + return ggml_cuda_mmq_util_funcs( + VDR_IQ1_S_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_iq1_s, + ggml_cuda_mmq_vec_dot_q8_1_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_IQ2_XXS: + return ggml_cuda_mmq_util_funcs( + VDR_IQ2_XXS_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_iq2_xxs, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_IQ2_XS: + return ggml_cuda_mmq_util_funcs( + VDR_IQ2_XS_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_iq2_xs, + ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_IQ2_S: + return ggml_cuda_mmq_util_funcs( + VDR_IQ2_S_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_iq2_s, + ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_IQ3_XXS: + return ggml_cuda_mmq_util_funcs( + VDR_IQ3_XXS_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_iq3_xxs, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_IQ3_S: + return ggml_cuda_mmq_util_funcs( + VDR_IQ3_S_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_iq3_s, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_IQ4_XS: + return ggml_cuda_mmq_util_funcs( + VDR_IQ4_XS_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_iq4_xs, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_IQ4_NL: + return ggml_cuda_mmq_util_funcs( + VDR_IQ4_NL_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_iq4_nl, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); +// --------------------------------------------------------------------------------------------- + case GGML_TYPE_MXFP4: + return ggml_cuda_mmq_util_funcs( + VDR_MXFP4_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_mxfp4, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + case GGML_TYPE_NVFP4: + return ggml_cuda_mmq_util_funcs( + VDR_NVFP4_Q8_1_MMQ, + ggml_cuda_mmq_load_tiles_nvfp4, + ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_dp4a, + ggml_cuda_mmq_write_back_dp4a); + default: + return ggml_cuda_mmq_util_funcs(1, nullptr, nullptr, nullptr); + } + } -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q5_0_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q5_0; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; -}; +// --------------------------------------------------------------------------------------------- -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q5_1_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q5_1; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_1_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_1_q8_1_dp4a; -}; - -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q8_0_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q8_0; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; -}; - -template -struct mmq_type_traits { - static constexpr int vdr = VDR_MXFP4_Q8_1_MMQ; #ifdef BLACKWELL_MMA_AVAILABLE - static constexpr load_tiles_mmq_t load_tiles = load_tiles_mxfp4_fp4; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_fp4_fp4_mma; -#else - static constexpr load_tiles_mmq_t load_tiles = load_tiles_mxfp4; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; + switch (type) { + case GGML_TYPE_MXFP4: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_mxfp4_fp4, + ggml_cuda_mmq_vec_dot_fp4_fp4_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_NVFP4: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_nvfp4_nvfp4, + ggml_cuda_mmq_vec_dot_fp4_fp4_mma, + ggml_cuda_mmq_write_back_mma); + default: + break; + } #endif // BLACKWELL_MMA_AVAILABLE - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; -}; -template -struct mmq_type_traits { - static constexpr int vdr = VDR_NVFP4_Q8_1_MMQ; -#ifdef BLACKWELL_MMA_AVAILABLE - static constexpr load_tiles_mmq_t load_tiles = load_tiles_nvfp4_nvfp4; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_fp4_fp4_mma; -#else - static constexpr load_tiles_mmq_t load_tiles = load_tiles_nvfp4; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_16_q8_1_mma; -#endif // BLACKWELL_MMA_AVAILABLE - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_16_q8_1_dp4a; -}; +// --------------------------------------------------------------------------------------------- -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q2_K_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q2_K; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q2_K_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q2_K_q8_1_dp4a; -}; + switch (type) { + case GGML_TYPE_Q1_0: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q1_0, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q4_0: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q4_0, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q4_1: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q4_1, + ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q5_0: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q5_0, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q5_1: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q5_1, + ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q8_0: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q8_0, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); +// --------------------------------------------------------------------------------------------- + case GGML_TYPE_Q2_K: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q2_K, + ggml_cuda_mmq_vec_dot_q2_K_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q3_K: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q3_K, + ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q4_K: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q4_K, + ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q5_K: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q5_K, + ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_Q6_K: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_q6_K, + ggml_cuda_mmq_vec_dot_q6_K_q8_1_mma, + ggml_cuda_mmq_write_back_mma); +// --------------------------------------------------------------------------------------------- + case GGML_TYPE_IQ1_S: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_iq1_s, + ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_IQ2_XXS: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_iq2_xxs, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_IQ2_XS: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_iq2_xs, + ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_IQ2_S: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_iq2_s, + ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_IQ3_XXS: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_iq3_xxs, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_IQ3_S: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_iq3_s, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_IQ4_XS: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_iq4_xs, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_IQ4_NL: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_iq4_nl, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); +// --------------------------------------------------------------------------------------------- + case GGML_TYPE_MXFP4: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_mxfp4, + ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + case GGML_TYPE_NVFP4: + return ggml_cuda_mmq_util_funcs( + -1, + ggml_cuda_mmq_load_tiles_nvfp4, + ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_mma, + ggml_cuda_mmq_write_back_mma); + default: + return ggml_cuda_mmq_util_funcs(1, nullptr, nullptr, nullptr); + } +} -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q3_K_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q3_K; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_16_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q3_K_q8_1_dp4a; -}; +template +static constexpr __device__ int ggml_cuda_mmq_get_vdr() { + return ggml_cuda_mmq_get_util_funcs().vdr; +} -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q4_K_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q4_K; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_1_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q4_K_q8_1_dp4a; -}; +template +static constexpr __device__ ggml_cuda_mmq_load_tiles_t ggml_cuda_mmq_get_load_tiles() { + return ggml_cuda_mmq_get_util_funcs().load_tiles; +} -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q5_K_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q5_K; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_1_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q5_K_q8_1_dp4a; -}; +template +static constexpr __device__ ggml_cuda_mmq_vec_dot_t ggml_cuda_mmq_get_vec_dot() { + return ggml_cuda_mmq_get_util_funcs().vec_dot; +} -template -struct mmq_type_traits { - static constexpr int vdr = VDR_Q6_K_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_q6_K; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q6_K_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q6_K_q8_1_dp4a; -}; +template +static constexpr __device__ ggml_cuda_mmq_write_back_t ggml_cuda_mmq_get_write_back() { + return ggml_cuda_mmq_get_util_funcs().write_back; +} -template -struct mmq_type_traits { - static constexpr int vdr = VDR_IQ2_XXS_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq2_xxs; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; -}; +// --------------------------------------------------------------------------------------------- -template -struct mmq_type_traits { - static constexpr int vdr = VDR_IQ2_XS_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq2_xs; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_16_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_16_q8_1_dp4a; -}; - -template -struct mmq_type_traits { - static constexpr int vdr = VDR_IQ2_S_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq2_s; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_16_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_16_q8_1_dp4a; -}; - -template -struct mmq_type_traits { - static constexpr int vdr = VDR_IQ3_XXS_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq3_xxs; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; -}; - -template -struct mmq_type_traits { - static constexpr int vdr = VDR_IQ3_S_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq3_s; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; -}; - -template -struct mmq_type_traits { - static constexpr int vdr = VDR_IQ1_S_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq1_s; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_1_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_1_q8_1_dp4a; -}; - -template -struct mmq_type_traits { - static constexpr int vdr = VDR_IQ4_NL_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq4_nl; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; -}; - -template -struct mmq_type_traits { - static constexpr int vdr = VDR_IQ4_XS_Q8_1_MMQ; - static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq4_xs; - static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma; - static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a; -}; - -template +template static __device__ __forceinline__ void mul_mat_q_process_tile( const char * __restrict__ x, const int offset_x, const int * __restrict__ y, const int * __restrict__ ids_dst, float * __restrict__ dst, float * __restrict__ tmp_fixup, @@ -3451,22 +820,16 @@ static __device__ __forceinline__ void mul_mat_q_process_tile( const int tile_x_max_i, const int tile_y_max_j, const int kb0_start, const int kb0_stop) { constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - constexpr int nwarps = mmq_get_nwarps_device(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; constexpr int qk = ggml_cuda_type_traits::qk; - constexpr int mmq_y = get_mmq_y_device(); - constexpr load_tiles_mmq_t load_tiles = mmq_type_traits::load_tiles; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); + constexpr ggml_cuda_mmq_load_tiles_t load_tiles = ggml_cuda_mmq_get_load_tiles(); + constexpr ggml_cuda_mmq_vec_dot_t vec_dot = ggml_cuda_mmq_get_vec_dot(); + constexpr ggml_cuda_mmq_write_back_t write_back = ggml_cuda_mmq_get_write_back(); extern __shared__ int data_mul_mat_q[]; - int * tile_y = data_mul_mat_q + mmq_x; - int * tile_x = tile_y + GGML_PAD(mmq_x*MMQ_TILE_Y_K, nwarps*warp_size); - -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) - constexpr vec_dot_mmq_t vec_dot = mmq_type_traits::vec_dot_mma; - constexpr mmq_write_back_t write_back = mmq_write_back_mma; -#else - constexpr vec_dot_mmq_t vec_dot = mmq_type_traits::vec_dot_dp4a; - constexpr mmq_write_back_t write_back = mmq_write_back_dp4a; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) + int * tile_y = data_mul_mat_q + J; + int * tile_x = tile_y + GGML_PAD(J*MMQ_TILE_Y_K, nwarps*warp_size); #if defined(BLACKWELL_MMA_AVAILABLE) // FP4 tile stores 8 blocks @@ -3475,10 +838,10 @@ static __device__ __forceinline__ void mul_mat_q_process_tile( constexpr int ne_block = 4 * QK8_1; #endif // defined(BLACKWELL_MMA_AVAILABLE) - constexpr int ITER_K = get_iter_k(type); + constexpr int ITER_K = ggml_cuda_mmq_get_K_vram(type, J, fallback); constexpr int blocks_per_iter = ITER_K / qk; - float sum[mmq_x*mmq_y / (nwarps*warp_size)] = {0.0f}; + float sum[J*I / (nwarps*warp_size)] = {0.0f}; constexpr int sz = sizeof(block_q8_1_mmq) / sizeof(int); @@ -3487,7 +850,7 @@ static __device__ __forceinline__ void mul_mat_q_process_tile( { const int * by0 = y + ncols_y * (kb0 * qk / ne_block) * sz; #pragma unroll - for (int l0 = 0; l0 < mmq_x * MMQ_TILE_Y_K; l0 += nwarps * warp_size) { + for (int l0 = 0; l0 < J * MMQ_TILE_Y_K; l0 += nwarps * warp_size) { int l = l0 + threadIdx.y*warp_size + threadIdx.x; tile_y[l] = by0[l]; @@ -3503,7 +866,7 @@ static __device__ __forceinline__ void mul_mat_q_process_tile( { const int * by0 = y + ncols_y * ((kb0 * qk / ne_block) * sz + sz); #pragma unroll - for (int l0 = 0; l0 < mmq_x * MMQ_TILE_Y_K; l0 += nwarps * warp_size) { + for (int l0 = 0; l0 < J * MMQ_TILE_Y_K; l0 += nwarps * warp_size) { int l = l0 + threadIdx.y*warp_size + threadIdx.x; tile_y[l] = by0[l]; @@ -3518,7 +881,7 @@ static __device__ __forceinline__ void mul_mat_q_process_tile( } if (fixup) { - write_back(sum, ids_dst, tmp_fixup + blockIdx.x*(mmq_x*mmq_y), mmq_y, mmq_y, mmq_x); + write_back(sum, ids_dst, tmp_fixup + blockIdx.x*(J*I), I, I, J); } else { write_back(sum, ids_dst, dst, stride_col_dst, tile_x_max_i, tile_y_max_j); } @@ -3527,18 +890,8 @@ static __device__ __forceinline__ void mul_mat_q_process_tile( // The mul_mat_q kernel implements "stream-k" work partitioning as described in https://arxiv.org/abs/2301.03598 -template -#if defined(GGML_USE_HIP) -#if defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN) - __launch_bounds__(ggml_cuda_get_physical_warp_size()*mmq_get_nwarps_device(), 2) -#endif // defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN) -#else -#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA - __launch_bounds__(ggml_cuda_get_physical_warp_size()*mmq_get_nwarps_device(), 1) -#else - __launch_bounds__(ggml_cuda_get_physical_warp_size()*mmq_get_nwarps_device(), 2) -#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA -#endif // defined(GGML_USE_HIP) +template +__launch_bounds__(ggml_cuda_mmq_get_nthreads(type, J, fallback), ggml_cuda_mmq_get_occupancy(type, J, fallback)) static __global__ void mul_mat_q( const char * __restrict__ x, const int * __restrict__ y, const int32_t * __restrict__ ids_dst, const int32_t * __restrict__ expert_bounds, float * __restrict__ dst, float * __restrict__ tmp_fixup, @@ -3548,28 +901,27 @@ static __global__ void mul_mat_q( const uint3 ntx) { // Skip unused template specializations for faster compilation: - if (mmq_x > get_mmq_x_max_device() || mmq_x % mmq_get_granularity_device(mmq_x) != 0) { + if (ggml_cuda_mmq_get_config(type, J, fallback).type == GGML_TYPE_COUNT) { NO_DEVICE_CODE; return; } - constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size; + constexpr int qk = ggml_cuda_type_traits::qk; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); - constexpr int qk = ggml_cuda_type_traits::qk; - constexpr int mmq_y = get_mmq_y_device(); - - const uint32_t nty = (nrows_x + mmq_y - 1) / mmq_y; // Number of tiles y + const uint32_t nty = (nrows_x + I - 1) / I; // Number of tiles y // Initialize the ids for writing back data with just the index. // For regular matrix multiplications this is never changed. // For MoE the correct indices are loaded from ids_dst. extern __shared__ int ids_dst_shared[]; // Stored at beginning of shared memory. #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps*warp_size) { + for (int j0 = 0; j0 < J; j0 += nwarps*warp_size) { const int j = j0 + threadIdx.y*warp_size + threadIdx.x; - if (j0 + nwarps*warp_size > mmq_x && j >= mmq_x) { + if (j0 + nwarps*warp_size > J && j >= J) { break; } @@ -3577,9 +929,7 @@ static __global__ void mul_mat_q( } __syncthreads(); - // On non-CDNA AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead: -#if (defined(GGML_USE_HIP) && !defined(CDNA)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA - { + if constexpr (!ggml_cuda_mmq_get_stream_k(type, J, fallback)) { const uint2 tmp2 = fast_div_modulo(blockIdx.z, nchannels_y); const int wt = tmp2.x; const int zt = tmp2.y; @@ -3591,7 +941,7 @@ static __global__ void mul_mat_q( int col_high = ncols_dst; int col_diff = ncols_dst; int offset_y = wt*stride_sample_y + zt*stride_channel_y; - int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst; + int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*J*stride_col_dst; if (ids_dst) { col_low = expert_bounds[zt + 0]; @@ -3601,41 +951,40 @@ static __global__ void mul_mat_q( offset_y = 0; offset_dst = 0; - if (jt*mmq_x >= col_diff) { + if (jt*J >= col_diff) { return; } // __syncthreads(); // There is no previous tile that could cause a race condition. #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps*warp_size) { + for (int j0 = 0; j0 < J; j0 += nwarps*warp_size) { const int j = j0 + threadIdx.y*warp_size + threadIdx.x; - if (j0 + nwarps*warp_size > mmq_x && j >= mmq_x) { + if (j0 + nwarps*warp_size > J && j >= J) { break; } - ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j]; + ids_dst_shared[j] = ids_dst[col_low + jt*J + j]; } __syncthreads(); } - offset_y += (col_low + jt*mmq_x)*(sizeof(block_q8_1_mmq)/sizeof(int)); - offset_dst += it*mmq_y; + offset_y += (col_low + jt*J)*(sizeof(block_q8_1_mmq)/sizeof(int)); + offset_dst += it*I; - const int tile_x_max_i = nrows_x - it*mmq_y - 1; - const int tile_y_max_j = col_diff - jt*mmq_x - 1; + const int tile_x_max_i = nrows_x - it*I - 1; + const int tile_y_max_j = col_diff - jt*J - 1; - const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x; + const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*I*stride_row_x; constexpr bool fixup = false; - mul_mat_q_process_tile + mul_mat_q_process_tile (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, stride_row_x, ncols_y, stride_col_dst, tile_x_max_i, tile_y_max_j, 0, blocks_per_ne00.z); return; } -#endif // (defined(GGML_USE_HIP) && !defined(CDNA4) && !defined(CDNA3)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA - constexpr int ITER_K = get_iter_k(type); + constexpr int ITER_K = ggml_cuda_mmq_get_K_vram(type, J, fallback); constexpr int blocks_per_iter = ITER_K / qk; // kbc == k block continuous, current index in continuous ijk space. @@ -3665,7 +1014,7 @@ static __global__ void mul_mat_q( int col_high = ncols_dst; int col_diff = ncols_dst; int offset_y = wt*stride_sample_y + zt*stride_channel_y; - int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst; + int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*J*stride_col_dst; if (ids_dst) { col_low = expert_bounds[zt + 0]; @@ -3675,7 +1024,7 @@ static __global__ void mul_mat_q( offset_y = 0; offset_dst = 0; - if (jt*mmq_x >= col_diff) { + if (jt*J >= col_diff) { kbc += blocks_per_ne00.z; kbc -= fastmodulo(kbc, blocks_per_ne00); @@ -3687,28 +1036,28 @@ static __global__ void mul_mat_q( __syncthreads(); #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps*warp_size) { + for (int j0 = 0; j0 < J; j0 += nwarps*warp_size) { const int j = j0 + threadIdx.y*warp_size + threadIdx.x; - if (j0 + nwarps*warp_size > mmq_x && j >= mmq_x) { + if (j0 + nwarps*warp_size > J && j >= J) { break; } - ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j]; + ids_dst_shared[j] = ids_dst[col_low + jt*J + j]; } __syncthreads(); } - offset_y += (col_low + jt * mmq_x) * (sizeof(block_q8_1_mmq) / sizeof(int)); - offset_dst += it*mmq_y; + offset_y += (col_low + jt * J) * (sizeof(block_q8_1_mmq) / sizeof(int)); + offset_dst += it*I; - const int tile_x_max_i = nrows_x - it*mmq_y - 1; - const int tile_y_max_j = col_diff - jt*mmq_x - 1; + const int tile_x_max_i = nrows_x - it*I - 1; + const int tile_y_max_j = col_diff - jt*J - 1; - const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x; + const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*I*stride_row_x; constexpr bool fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer. - mul_mat_q_process_tile + mul_mat_q_process_tile (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, stride_row_x, ncols_y, stride_col_dst, tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop); @@ -3739,7 +1088,7 @@ static __global__ void mul_mat_q( int col_high = ncols_dst; int col_diff = ncols_dst; int offset_y = wt*stride_sample_y + zt*stride_channel_y; - int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst; + int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*J*stride_col_dst; if (ids_dst) { col_low = expert_bounds[zt + 0]; @@ -3749,17 +1098,17 @@ static __global__ void mul_mat_q( offset_y = 0; offset_dst = 0; - if (jt*mmq_x >= col_diff) { + if (jt*J >= col_diff) { return; } // The memory layout for the fixup buffer is always contiguous, therefore reset ids: __syncthreads(); #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps*warp_size) { + for (int j0 = 0; j0 < J; j0 += nwarps*warp_size) { const int j = j0 + threadIdx.y*warp_size + threadIdx.x; - if (j0 + nwarps*warp_size > mmq_x && j >= mmq_x) { + if (j0 + nwarps*warp_size > J && j >= J) { break; } @@ -3768,39 +1117,38 @@ static __global__ void mul_mat_q( __syncthreads(); } - offset_y += (col_low + jt * mmq_x) * (sizeof(block_q8_1_mmq) / sizeof(int)); - offset_dst += it*mmq_y; + offset_y += (col_low + jt * J) * (sizeof(block_q8_1_mmq) / sizeof(int)); + offset_dst += it*I; - const int tile_x_max_i = nrows_x - it*mmq_y - 1; - const int tile_y_max_j = col_diff - jt*mmq_x - 1; + const int tile_x_max_i = nrows_x - it*I - 1; + const int tile_y_max_j = col_diff - jt*J - 1; - const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x; + const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*I*stride_row_x; constexpr bool fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks. - mul_mat_q_process_tile + mul_mat_q_process_tile (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, stride_row_x, ncols_y, stride_col_dst, tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop); } -template -__launch_bounds__(ggml_cuda_get_physical_warp_size()*mmq_get_nwarps_device()/2, 1) +template +__launch_bounds__(ggml_cuda_mmq_get_nthreads(type, J, fallback)/2, 1) static __global__ void mul_mat_q_stream_k_fixup( const int32_t * __restrict__ ids_dst, const int32_t * __restrict__ expert_bounds, float * __restrict__ dst, float * __restrict__ tmp_last_tile, const uint3 blocks_per_ne00, const int nrows_x, const int ncols_dst, const int stride_col_dst, const uint3 nchannels_y, const int stride_channel_dst, const uint3 nsamples_y, const int stride_sample_dst, const uint3 ntx) { - constexpr int mmq_y = get_mmq_y_device(); + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + constexpr int nwarps = (ggml_cuda_mmq_get_nthreads(type, J, fallback) / 2) / warp_size; + constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback); constexpr int qk = ggml_cuda_type_traits::qk; - constexpr int ITER_K = get_iter_k(type); + constexpr int ITER_K = ggml_cuda_mmq_get_K_vram(type, J, fallback); constexpr int blocks_per_iter = ITER_K / qk; - constexpr int nwarps = mmq_get_nwarps_device()/2; - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - - float sum[mmq_x / nwarps] = {0.0f}; + float sum[J / nwarps] = {0.0f}; const int i = blockIdx.y*warp_size + threadIdx.x; - const int nty = (nrows_x + mmq_y - 1) / mmq_y; + const int nty = (nrows_x + I - 1) / I; const int bidx0 = blockIdx.x; @@ -3838,10 +1186,10 @@ static __global__ void mul_mat_q_stream_k_fixup( #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { + for (int j0 = 0; j0 < J; j0 += nwarps) { const int j = j0 + threadIdx.y; - sum[j0/nwarps] += tmp_last_tile[bidx*(mmq_x*mmq_y) + j*mmq_y + i]; + sum[j0/nwarps] += tmp_last_tile[bidx*(J*I) + j*I + i]; } // If this block started in a previous tile we are done and don't need to combine additional partial results. @@ -3868,17 +1216,17 @@ static __global__ void mul_mat_q_stream_k_fixup( const int it = tmp2.x; if (!ids_dst) { - const int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst + it*mmq_y; + const int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*J*stride_col_dst + it*I; dst += offset_dst; - const int i_max = nrows_x - it*mmq_y - 1; - const int j_max = ncols_dst - jt*mmq_x - 1; - if (need_check && i > i_max) { + const int i_max = nrows_x - it*I - 1; + const int j_max = ncols_dst - jt*J - 1; + if (fallback && i > i_max) { return; } #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { + for (int j0 = 0; j0 < J; j0 += nwarps) { const int j = j0 + threadIdx.y; if (j > j_max) { @@ -3890,27 +1238,27 @@ static __global__ void mul_mat_q_stream_k_fixup( return; } - __shared__ int ids_dst_shared[mmq_x]; + __shared__ int ids_dst_shared[J]; const int col_low = expert_bounds[zt + 0]; const int col_high = expert_bounds[zt + 1]; const int col_diff = col_high - col_low; - for (int j = threadIdx.y*warp_size + threadIdx.x; j < mmq_x; j += nwarps*warp_size) { - ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j]; + for (int j = threadIdx.y*warp_size + threadIdx.x; j < J; j += nwarps*warp_size) { + ids_dst_shared[j] = ids_dst[col_low + jt*J + j]; } __syncthreads(); - const int offset_dst = it*mmq_y; + const int offset_dst = it*I; dst += offset_dst; - const int i_max = nrows_x - it*mmq_y - 1; - const int j_max = col_diff - jt*mmq_x - 1; - if (need_check && i > i_max) { + const int i_max = nrows_x - it*I - 1; + const int j_max = col_diff - jt*J - 1; + if (fallback && i > i_max) { return; } #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { + for (int j0 = 0; j0 < J; j0 += nwarps) { const int j = j0 + threadIdx.y; if (j > j_max) { @@ -3926,37 +1274,35 @@ struct mmq_args { int64_t ncols_x; int64_t nrows_x; int64_t ncols_dst; int64_t stride_row_x; int64_t ncols_y; int64_t nrows_dst; int64_t nchannels_x; int64_t nchannels_y; int64_t stride_channel_x; int64_t stride_channel_y; int64_t stride_channel_dst; int64_t nsamples_x; int64_t nsamples_y; int64_t stride_sample_x; int64_t stride_sample_y; int64_t stride_sample_dst; - bool use_stream_k; int64_t ncols_max; + int64_t ncols_max; }; -template -static size_t mmq_get_nbytes_shared(const int mmq_x, const int mmq_y, const int cc, const int warp_size, const int nwarps) { - const tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(type, mmq_y); - const int mmq_tile_x_k = mmq_get_mma_tile_x_k(type); - const size_t nbs_ids = mmq_x*sizeof(int); - const size_t nbs_x = (turing_mma_available(cc) || amd_mfma_available(cc) || amd_wmma_available(cc)) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int); - const size_t nbs_y = mmq_x * (sizeof(block_q8_1_mmq)); - return nbs_ids + nbs_x + GGML_PAD(nbs_y, nwarps*warp_size*sizeof(int)); +static size_t mmq_get_nbytes_shared(const ggml_cuda_mmq_config & config, const int cc) { + const size_t nbs_ids = config.J*sizeof(int); + const size_t nbs_x = ggml_cuda_mmq_get_nbytes_shared_x(config, cc); + const size_t nbs_y = config.J * (sizeof(block_q8_1_mmq)); + return nbs_ids + nbs_x + GGML_PAD(nbs_y, config.nthreads*sizeof(int)); } -template +template static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) { const int id = ggml_cuda_get_device(); const int cc = ggml_cuda_info().devices[id].cc; const int nsm = ggml_cuda_info().devices[id].nsm; const int warp_size = ggml_cuda_info().devices[id].warp_size; - const int nwarps = mmq_get_nwarps_host(cc, warp_size); - const int mmq_y = get_mmq_y_host(cc); + + const ggml_cuda_mmq_config config = ggml_cuda_mmq_get_config(type, J, fallback, cc); + GGML_ASSERT(config.nthreads % warp_size == 0); + const int nwarps = config.nthreads / warp_size; + const int nbytes_shared = mmq_get_nbytes_shared(config, cc); const dim3 block_dims(warp_size, nwarps, 1); - const int nbytes_shared = mmq_get_nbytes_shared(mmq_x, mmq_y, cc, warp_size, nwarps); + CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q), nbytes_shared); + CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q), nbytes_shared); - CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q), nbytes_shared); - CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q), nbytes_shared); - - const int nty = (args.nrows_x + mmq_y - 1) / mmq_y; - const int ntx = (args.ncols_max + mmq_x - 1) / mmq_x; + const int nty = (args.nrows_x + config.I - 1) / config.I; + const int ntx = (args.ncols_max + config.J - 1) / config.J; const int ntzw = args.nchannels_y * args.nsamples_y; const dim3 block_nums_xy_tiling(nty, ntx, ntzw); @@ -3972,24 +1318,13 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a const uint3 channel_ratio_fd = init_fastdiv_values(channel_ratio); const uint3 sample_ratio_fd = init_fastdiv_values(sample_ratio); - if (!args.use_stream_k) { - if (args.nrows_x % mmq_y == 0) { - constexpr bool need_check = false; - mul_mat_q<<>> - (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr, - blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst, - channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst, - sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst, - ntx_fd); - } else { - constexpr bool need_check = true; - mul_mat_q<<>> - (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr, - blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst, - channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst, - sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst, - ntx_fd); - } + if (!ggml_cuda_mmq_get_stream_k(type, J, fallback, cc)) { + mul_mat_q<<>> + (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr, + blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst, + channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst, + sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst, + ntx_fd); return; } @@ -4007,133 +1342,121 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a ggml_cuda_pool & pool = ctx.pool(id); ggml_cuda_pool_alloc tmp_fixup(pool); if (fixup_needed) { - tmp_fixup.alloc(block_nums_stream_k.x * mmq_x*mmq_y); + tmp_fixup.alloc(block_nums_stream_k.x * config.J*config.I); } - const dim3 block_nums_fixup(block_nums_stream_k.x, mmq_y/warp_size, 1); + const dim3 block_nums_fixup(block_nums_stream_k.x, config.I/warp_size, 1); const dim3 block_dims_fixup(block_dims.x, block_dims.y/2, block_dims.z); - if (args.nrows_x % mmq_y == 0) { - constexpr bool need_check = false; - mul_mat_q<<>> - (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, - blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst, - channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst, - sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst, - ntx_fd); + mul_mat_q<<>> + (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, + blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst, + channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst, + sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst, + ntx_fd); - if (!fixup_needed) { - return; + if (!fixup_needed) { + return; + } + + CUDA_CHECK(cudaGetLastError()); + mul_mat_q_stream_k_fixup<<>> + (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, + args.nrows_dst, nchannels_y_fd, args.stride_channel_dst, nsamples_y_fd, args.stride_sample_dst, + ntx_fd); +} + +template +void mul_mat_q_switch_J(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) { + const int id = ggml_cuda_get_device(); + const int cc = ggml_cuda_info().devices[id].cc; + const size_t smpbo = ggml_cuda_info().devices[id].smpbo; + + int J_best = 0; + int ntiles_J_best = INT_MAX; + + for (int J = 8; J <= 128 && ntiles_J_best > 1; J += 8) { + const ggml_cuda_mmq_config config = ggml_cuda_mmq_get_config(type, J, fallback, cc); + if (config.type == GGML_TYPE_COUNT) { + continue; } - CUDA_CHECK(cudaGetLastError()); - mul_mat_q_stream_k_fixup<<>> - (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, - args.nrows_dst, nchannels_y_fd, args.stride_channel_dst, nsamples_y_fd, args.stride_sample_dst, - ntx_fd); - } else { - constexpr bool need_check = true; - mul_mat_q<<>> - (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, - blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst, - channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst, - sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst, - ntx_fd); - - if (!fixup_needed) { - return; + if (mmq_get_nbytes_shared(config, cc) > smpbo) { + continue; } - CUDA_CHECK(cudaGetLastError()); - mul_mat_q_stream_k_fixup<<>> - (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, - args.nrows_dst, nchannels_y_fd, args.stride_channel_dst, nsamples_y_fd, args.stride_sample_dst, - ntx_fd); + const int ntiles_x = (args.ncols_max + config.J - 1) / config.J; + + if (ntiles_x < ntiles_J_best) { + J_best = J; + ntiles_J_best = ntiles_x; + } + } + + switch (J_best) { + case 8: + launch_mul_mat_q(ctx, args, stream); + break; + case 16: + launch_mul_mat_q(ctx, args, stream); + break; + case 24: + launch_mul_mat_q(ctx, args, stream); + break; + case 32: + launch_mul_mat_q(ctx, args, stream); + break; + case 40: + launch_mul_mat_q(ctx, args, stream); + break; + case 48: + launch_mul_mat_q(ctx, args, stream); + break; + case 56: + launch_mul_mat_q(ctx, args, stream); + break; + case 64: + launch_mul_mat_q(ctx, args, stream); + break; + case 72: + launch_mul_mat_q(ctx, args, stream); + break; + case 80: + launch_mul_mat_q(ctx, args, stream); + break; + case 88: + launch_mul_mat_q(ctx, args, stream); + break; + case 96: + launch_mul_mat_q(ctx, args, stream); + break; + case 104: + launch_mul_mat_q(ctx, args, stream); + break; + case 112: + launch_mul_mat_q(ctx, args, stream); + break; + case 120: + launch_mul_mat_q(ctx, args, stream); + break; + case 128: + launch_mul_mat_q(ctx, args, stream); + break; + default: + fprintf(stderr, "J_best=%d\n", J_best); + GGML_ABORT("fatal error"); + break; } } template void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) { - const int id = ggml_cuda_get_device(); - const int cc = ggml_cuda_info().devices[id].cc; - const size_t smpbo = ggml_cuda_info().devices[id].smpbo; - const int warp_size = ggml_cuda_info().devices[id].warp_size; - const int nwarps = mmq_get_nwarps_host(cc, warp_size); - - const int mmq_x_max = get_mmq_x_max_host(cc); - const int mmq_y = get_mmq_y_host(cc); - - int mmq_x_best = 0; - int ntiles_x_best = INT_MAX; - - for (int mmq_x = 8; mmq_x <= mmq_x_max && ntiles_x_best > 1; mmq_x += 8) { - const int granularity = mmq_get_granularity_host(mmq_x, cc); - - if (mmq_x % granularity != 0 || mmq_get_nbytes_shared(mmq_x, mmq_y, cc, warp_size, nwarps) > smpbo) { - continue; - } - - const int ntiles_x = (args.ncols_max + mmq_x - 1) / mmq_x; - - if (ntiles_x < ntiles_x_best) { - mmq_x_best = mmq_x; - ntiles_x_best = ntiles_x; - } - } - - switch (mmq_x_best) { - case 8: - launch_mul_mat_q(ctx, args, stream); - break; - case 16: - launch_mul_mat_q(ctx, args, stream); - break; - case 24: - launch_mul_mat_q(ctx, args, stream); - break; - case 32: - launch_mul_mat_q(ctx, args, stream); - break; - case 40: - launch_mul_mat_q(ctx, args, stream); - break; - case 48: - launch_mul_mat_q(ctx, args, stream); - break; - case 56: - launch_mul_mat_q(ctx, args, stream); - break; - case 64: - launch_mul_mat_q(ctx, args, stream); - break; - case 72: - launch_mul_mat_q(ctx, args, stream); - break; - case 80: - launch_mul_mat_q(ctx, args, stream); - break; - case 88: - launch_mul_mat_q(ctx, args, stream); - break; - case 96: - launch_mul_mat_q(ctx, args, stream); - break; - case 104: - launch_mul_mat_q(ctx, args, stream); - break; - case 112: - launch_mul_mat_q(ctx, args, stream); - break; - case 120: - launch_mul_mat_q(ctx, args, stream); - break; - case 128: - launch_mul_mat_q(ctx, args, stream); - break; - default: - fprintf(stderr, "mmq_x_best=%d\n", mmq_x_best); - GGML_ABORT("fatal error"); - break; + if (args.nrows_x % 128 == 0) { + constexpr bool fallback = false; + mul_mat_q_switch_J(ctx, args, stream); + } else { + constexpr bool fallback = true; + mul_mat_q_switch_J(ctx, args, stream); } } @@ -4166,11 +1489,4 @@ extern DECL_MMQ_CASE(GGML_TYPE_IQ4_XS); void ggml_cuda_mul_mat_q( ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst); -void ggml_cuda_op_mul_mat_q( - ggml_backend_cuda_context & ctx, - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, - const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, - const int64_t src1_padded_row_size, cudaStream_t stream); - bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t n_experts); -