furyhawk/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-07-21 10:15:53 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan/cpu: Support f16 as SET_ROWS src. (#25432)

Browse Source 
* vulkan/cpu: Support f16 as SET_ROWS src.

This adds full support for f16 SET_ROWS (equivalent to f32) to vulkan and CPU
backends, and adds more backend tests.

* Set DenormPreserve 16 when supported, to try to fix failures on Intel

* tune error threshold

* update metal supports_op
This commit is contained in:
Jeff Bolz
2026-07-14 08:26:55 -05:00
committed by GitHub
parent 8ff8c4299d
commit 7cbd61002d
7 changed files with 150 additions and 100 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cpu/ggml-cpu.c
+6
View File
@@ -2863,6 +2863,12 @@ struct ggml_cplan ggml_graph_plan(
cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
}
} break;
case GGML_OP_SET_ROWS:
{
if (node->src[0]->type == GGML_TYPE_F16 && node->type != GGML_TYPE_F16) {
cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
}
} break;
case GGML_OP_SOFT_MAX:
case GGML_OP_ROPE:
case GGML_OP_ROPE_BACK:
ggml/src/ggml-cpu/ops.cpp
+16 -11
View File
@@ -5041,7 +5041,7 @@ static void ggml_compute_forward_set_rows_impl(
assert(ne0 == nc);
assert(ne2 == ne02);
assert(ne3 == ne03);
GGML_ASSERT(src0->type == GGML_TYPE_F32 || (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16));
GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
assert(ne02 % ne11 == 0);
assert(ne03 % ne12 == 0);
@@ -5075,10 +5075,19 @@ static void ggml_compute_forward_set_rows_impl(
(const float *) ((char *) src0->data + i*nb01 + i02*nb02 + i03*nb03),
((char *) dst->data + i1*nb1 + i02*nb2 + i03*nb3), nc);
} else if constexpr (std::is_same_v<src_t, ggml_fp16_t>) {
memcpy(
if (dst->type == GGML_TYPE_F16) {
memcpy(
((char *) dst->data + i1*nb1 + i02*nb2 + i03*nb3),
((char *) src0->data + i*nb01 + i02*nb02 + i03*nb03),
rs);
} else {
float * wdata = (float *) params->wdata + (nc + CACHE_LINE_SIZE_F32) * ith;
ggml_fp16_to_fp32_row(
(const ggml_fp16_t *) ((char *) src0->data + i*nb01 + i02*nb02 + i03*nb03),
wdata, nc);
from_float(wdata,
((char *) dst->data + i1*nb1 + i02*nb2 + i03*nb3), nc);
}
} else {
GGML_ABORT("src0->type = %d (%s) not supported", src0->type, ggml_type_name(src0->type));
}
@@ -5107,16 +5116,12 @@ void ggml_compute_forward_set_rows(
} break;
case GGML_TYPE_F16:
{
if (dst->type == GGML_TYPE_F16) {
if (src1->type == GGML_TYPE_I64) {
ggml_compute_forward_set_rows_impl<ggml_fp16_t, int64_t>(params, dst);
} else if (src1->type == GGML_TYPE_I32) {
ggml_compute_forward_set_rows_impl<ggml_fp16_t, int32_t>(params, dst);
} else {
GGML_ABORT("src1->type = %d (%s) not supported", src1->type, ggml_type_name(src1->type));
}
if (src1->type == GGML_TYPE_I64) {
ggml_compute_forward_set_rows_impl<ggml_fp16_t, int64_t>(params, dst);
} else if (src1->type == GGML_TYPE_I32) {
ggml_compute_forward_set_rows_impl<ggml_fp16_t, int32_t>(params, dst);
} else {
GGML_ABORT("dst->type = %d (%s) not supported with src0->type = %d (%s)", dst->type, ggml_type_name(dst->type), src0->type, ggml_type_name(src0->type));
GGML_ABORT("src1->type = %d (%s) not supported", src1->type, ggml_type_name(src1->type));
}
} break;
default:
ggml/src/ggml-metal/ggml-metal-device.m
+5 -1
View File
@@ -1340,7 +1340,11 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
return op->src[0]->type != GGML_TYPE_NVFP4;
case GGML_OP_SET_ROWS:
{
if (op->src[0]->type != GGML_TYPE_F32 && op->src[0]->type != GGML_TYPE_F16) {
if (op->src[0]->type == GGML_TYPE_F16) {
return op->type == GGML_TYPE_F16;
}
if (op->src[0]->type != GGML_TYPE_F32) {
return false;
}
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+74 -45
View File
@@ -723,6 +723,7 @@ struct vk_device_struct {
bool uma;
bool prefer_host_memory;
bool float_controls_rte_fp16;
bool float_controls_denorm_preserve_fp16;
bool subgroup_basic;
bool subgroup_arithmetic;
bool subgroup_shuffle;
@@ -868,8 +869,9 @@ struct vk_device_struct {
vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT];
vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT];
vk_pipeline pipeline_cpy_transpose_16, pipeline_cpy_transpose_32;
vk_pipeline pipeline_set_rows_i32[GGML_TYPE_COUNT];
vk_pipeline pipeline_set_rows_i64[GGML_TYPE_COUNT];
// [src0 0=fp32,1=fp16][dst]
vk_pipeline pipeline_set_rows_i32[2][GGML_TYPE_COUNT];
vk_pipeline pipeline_set_rows_i64[2][GGML_TYPE_COUNT];
vk_pipeline pipeline_norm_f32;
vk_pipeline pipeline_group_norm_f32;
vk_pipeline pipeline_rms_norm_f32;
@@ -2595,10 +2597,10 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
vk::ShaderModuleCreateInfo shader_module_create_info({}, spv_size, reinterpret_cast<const uint32_t *>(spv_data));
// Patch SPIR-V to enable RTE rounding for FP16, avoiding the need for
// separate shader variants compiled with -DRTE16.
// Patch SPIR-V to enable supported FP16 float controls, avoiding the need
// for separate shader variants.
std::vector<uint32_t> spirv;
if (device->float_controls_rte_fp16) {
if (device->float_controls_rte_fp16 || device->float_controls_denorm_preserve_fp16) {
const uint32_t* spv_words = reinterpret_cast<const uint32_t *>(spv_data);
size_t word_count = spv_size / sizeof(uint32_t);
spirv.assign(spv_words, spv_words + word_count);
@@ -2635,9 +2637,17 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
// Insert from latest position first so earlier indices stay valid.
// OpExecutionMode %entrypoint RoundingModeRTE 16
uint32_t exec_mode[] = { (4u << spv::WordCountShift) | spv::OpExecutionMode, entry_point_id, spv::ExecutionModeRoundingModeRTE, 16 };
spirv.insert(spirv.begin() + exec_insert_pos, std::begin(exec_mode), std::end(exec_mode));
if (device->float_controls_rte_fp16) {
// OpExecutionMode %entrypoint RoundingModeRTE 16
uint32_t exec_mode[] = { (4u << spv::WordCountShift) | spv::OpExecutionMode, entry_point_id, spv::ExecutionModeRoundingModeRTE, 16 };
spirv.insert(spirv.begin() + exec_insert_pos, std::begin(exec_mode), std::end(exec_mode));
}
if (device->float_controls_denorm_preserve_fp16) {
// OpExecutionMode %entrypoint DenormPreserve 16
uint32_t exec_mode[] = { (4u << spv::WordCountShift) | spv::OpExecutionMode, entry_point_id, spv::ExecutionModeDenormPreserve, 16 };
spirv.insert(spirv.begin() + exec_insert_pos, std::begin(exec_mode), std::end(exec_mode));
}
// OpExtension "SPV_KHR_float_controls"
const char ext_str[] = "SPV_KHR_float_controls";
@@ -2647,9 +2657,17 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
memcpy(&extension[1], ext_str, sizeof(ext_str));
spirv.insert(spirv.begin() + ext_insert_pos, extension.begin(), extension.end());
// OpCapability RoundingModeRTE
uint32_t capability[] = { (2u << spv::WordCountShift) | spv::OpCapability, spv::CapabilityRoundingModeRTE };
spirv.insert(spirv.begin() + cap_insert_pos, std::begin(capability), std::end(capability));
if (device->float_controls_rte_fp16) {
// OpCapability RoundingModeRTE
uint32_t capability[] = { (2u << spv::WordCountShift) | spv::OpCapability, spv::CapabilityRoundingModeRTE };
spirv.insert(spirv.begin() + cap_insert_pos, std::begin(capability), std::end(capability));
}
if (device->float_controls_denorm_preserve_fp16) {
// OpCapability DenormPreserve
uint32_t capability[] = { (2u << spv::WordCountShift) | spv::OpCapability, spv::CapabilityDenormPreserve };
spirv.insert(spirv.begin() + cap_insert_pos, std::begin(capability), std::end(capability));
}
shader_module_create_info = vk::ShaderModuleCreateInfo({}, spirv.size() * sizeof(uint32_t), spirv.data());
}
@@ -5187,20 +5205,22 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_len, cpy_f32_q8_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
#define SET_ROWS(itype) \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_F32], "set_rows_f32" #itype, set_rows_f32 ## itype ## _len, set_rows_f32 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_F16], "set_rows_f16" #itype, set_rows_f16 ## itype ## _len, set_rows_f16 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_BF16], "set_rows_bf16" #itype, set_rows_bf16 ## itype ## _len, set_rows_bf16 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q1_0], "set_rows_q1_0" #itype, set_rows_q1_0 ## itype ## _len, set_rows_q1_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q4_0], "set_rows_q4_0" #itype, set_rows_q4_0 ## itype ## _len, set_rows_q4_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q4_1], "set_rows_q4_1" #itype, set_rows_q4_1 ## itype ## _len, set_rows_q4_1 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q5_0], "set_rows_q5_0" #itype, set_rows_q5_0 ## itype ## _len, set_rows_q5_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q5_1], "set_rows_q5_1" #itype, set_rows_q5_1 ## itype ## _len, set_rows_q5_1 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q8_0], "set_rows_q8_0" #itype, set_rows_q8_0 ## itype ## _len, set_rows_q8_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_IQ4_NL], "set_rows_iq4_nl" #itype, set_rows_iq4_nl ## itype ## _len, set_rows_iq4_nl ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
#define SET_ROWS(src_idx, src, itype) \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_F32], "set_rows_" #src "_f32" #itype, set_rows_ ## src ## _f32 ## itype ## _len, set_rows_ ## src ## _f32 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_F16], "set_rows_" #src "_f16" #itype, set_rows_ ## src ## _f16 ## itype ## _len, set_rows_ ## src ## _f16 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_BF16], "set_rows_" #src "_bf16" #itype, set_rows_ ## src ## _bf16 ## itype ## _len, set_rows_ ## src ## _bf16 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_Q1_0], "set_rows_" #src "_q1_0" #itype, set_rows_ ## src ## _q1_0 ## itype ## _len, set_rows_ ## src ## _q1_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_Q4_0], "set_rows_" #src "_q4_0" #itype, set_rows_ ## src ## _q4_0 ## itype ## _len, set_rows_ ## src ## _q4_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_Q4_1], "set_rows_" #src "_q4_1" #itype, set_rows_ ## src ## _q4_1 ## itype ## _len, set_rows_ ## src ## _q4_1 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_Q5_0], "set_rows_" #src "_q5_0" #itype, set_rows_ ## src ## _q5_0 ## itype ## _len, set_rows_ ## src ## _q5_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_Q5_1], "set_rows_" #src "_q5_1" #itype, set_rows_ ## src ## _q5_1 ## itype ## _len, set_rows_ ## src ## _q5_1 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_Q8_0], "set_rows_" #src "_q8_0" #itype, set_rows_ ## src ## _q8_0 ## itype ## _len, set_rows_ ## src ## _q8_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [src_idx][GGML_TYPE_IQ4_NL], "set_rows_" #src "_iq4_nl" #itype, set_rows_ ## src ## _iq4_nl ## itype ## _len, set_rows_ ## src ## _iq4_nl ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
SET_ROWS(_i32)
SET_ROWS(_i64)
SET_ROWS(0, f32, _i32)
SET_ROWS(0, f32, _i64)
SET_ROWS(1, f16, _i32)
SET_ROWS(1, f16, _i64)
#undef SET_ROWS
@@ -6031,6 +6051,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
device->shader_core_count = 0;
}
device->float_controls_rte_fp16 = vk12_props.shaderRoundingModeRTEFloat16;
device->float_controls_denorm_preserve_fp16 = vk12_props.shaderDenormPreserveFloat16;
device->subgroup_basic = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
(vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eBasic);
@@ -10843,10 +10864,17 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
case GGML_OP_DUP:
return ggml_vk_get_cpy_pipeline(ctx, src0, dst, dst->type);
case GGML_OP_SET_ROWS:
if (src1->type == GGML_TYPE_I64) {
return ctx->device->pipeline_set_rows_i64[dst->type];
} else {
return ctx->device->pipeline_set_rows_i32[dst->type];
{
if (src0->type != GGML_TYPE_F32 && src0->type != GGML_TYPE_F16) {
return nullptr;
}
const int src_idx = src0->type == GGML_TYPE_F16;
if (src1->type == GGML_TYPE_I64) {
return ctx->device->pipeline_set_rows_i64[src_idx][dst->type];
} else if (src1->type == GGML_TYPE_I32) {
return ctx->device->pipeline_set_rows_i32[src_idx][dst->type];
}
return nullptr;
}
case GGML_OP_SILU_BACK:
if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
@@ -17500,24 +17528,25 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32;
case GGML_OP_SET_ROWS:
{
if (op->src[0]->type == GGML_TYPE_F32) {
switch (op->type) {
case GGML_TYPE_F32:
case GGML_TYPE_F16:
case GGML_TYPE_BF16:
case GGML_TYPE_Q1_0:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_IQ4_NL:
return true;
default:
return false;
}
if ((op->src[0]->type != GGML_TYPE_F32 && op->src[0]->type != GGML_TYPE_F16) ||
(op->src[1]->type != GGML_TYPE_I32 && op->src[1]->type != GGML_TYPE_I64)) {
return false;
}
switch (op->type) {
case GGML_TYPE_F32:
case GGML_TYPE_F16:
case GGML_TYPE_BF16:
case GGML_TYPE_Q1_0:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_IQ4_NL:
return true;
default:
return false;
}
return false;
}
case GGML_OP_CONT:
case GGML_OP_CPY:
ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp
+26 -26
View File
@@ -10,7 +10,7 @@ layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
const uint BLOCK_SIZE = 32;
#endif
layout (binding = 0) readonly buffer S {float data_s[];};
layout (binding = 0) readonly buffer S {S_TYPE data_s[];};
#if defined(SET_ROWS)
#include "generic_binary_head.glsl"
@@ -35,7 +35,7 @@ void quantize(uint dst_idx, uint src_idx)
float vmax = 0.0;
[[unroll]] for (int j = 0; j < QUANT_K_Q4_0; ++j) {
const float v = data_s[src_idx + j];
const float v = float(data_s[src_idx + j]);
if (amax < abs(v)) {
amax = abs(v);
vmax = v;
@@ -48,8 +48,8 @@ void quantize(uint dst_idx, uint src_idx)
data_q[dst_idx].d = float16_t(d);
[[unroll]] for (int j = 0; j < QUANT_K_Q4_0/2; ++j) {
const float x0 = data_s[src_idx + 0 + j]*id;
const float x1 = data_s[src_idx + QUANT_K_Q4_0/2 + j]*id;
const float x0 = float(data_s[src_idx + 0 + j])*id;
const float x1 = float(data_s[src_idx + QUANT_K_Q4_0/2 + j])*id;
const uint xi0 = min(15, int(x0 + 8.5));
const uint xi1 = min(15, int(x1 + 8.5));
@@ -66,7 +66,7 @@ void quantize(uint dst_idx, uint src_idx)
float vmax = -vmin;
[[unroll]] for (int j = 0; j < QUANT_K_Q4_1; ++j) {
const float v = data_s[src_idx + j];
const float v = float(data_s[src_idx + j]);
if (v < vmin) vmin = v;
if (v > vmax) vmax = v;
@@ -79,8 +79,8 @@ void quantize(uint dst_idx, uint src_idx)
data_q[dst_idx].m = float16_t(vmin);
[[unroll]] for (int j = 0; j < QUANT_K_Q4_1/2; ++j) {
const float x0 = (data_s[src_idx + 0 + j] - vmin)*id;
const float x1 = (data_s[src_idx + QUANT_K_Q4_1/2 + j] - vmin)*id;
const float x0 = (float(data_s[src_idx + 0 + j]) - vmin)*id;
const float x1 = (float(data_s[src_idx + QUANT_K_Q4_1/2 + j]) - vmin)*id;
const uint xi0 = min(15, int(x0 + 0.5));
const uint xi1 = min(15, int(x1 + 0.5));
@@ -97,7 +97,7 @@ void quantize(uint dst_idx, uint src_idx)
float vmax = 0.0;
[[unroll]] for (int j = 0; j < QUANT_K_Q5_0; ++j) {
const float v = data_s[src_idx + j];
const float v = float(data_s[src_idx + j]);
if (amax < abs(v)) {
amax = abs(v);
vmax = v;
@@ -111,8 +111,8 @@ void quantize(uint dst_idx, uint src_idx)
uint32_t qh = 0;
[[unroll]] for (int j = 0; j < QUANT_K_Q5_0/2; ++j) {
const float x0 = data_s[src_idx + 0 + j]*id;
const float x1 = data_s[src_idx + QUANT_K_Q5_0/2 + j]*id;
const float x0 = float(data_s[src_idx + 0 + j])*id;
const float x1 = float(data_s[src_idx + QUANT_K_Q5_0/2 + j])*id;
const uint xi0 = min(31, int(x0 + 16.5));
const uint xi1 = min(31, int(x1 + 16.5));
@@ -129,11 +129,11 @@ void quantize(uint dst_idx, uint src_idx)
#if defined(DATA_A_Q5_1)
void quantize(uint dst_idx, uint src_idx)
{
float min = data_s[src_idx + 0];
float min = float(data_s[src_idx + 0]);
float max = min;
[[unroll]] for (int j = 1; j < QUANT_K_Q5_1; ++j) {
const float v = data_s[src_idx + j];
const float v = float(data_s[src_idx + j]);
min = v < min ? v : min;
max = v > max ? v : max;
}
@@ -146,8 +146,8 @@ void quantize(uint dst_idx, uint src_idx)
uint32_t qh = 0;
[[unroll]] for (int j = 0; j < QUANT_K_Q5_1/2; ++j) {
const float x0 = (data_s[src_idx + 0 + j] - min)*id;
const float x1 = (data_s[src_idx + QUANT_K_Q5_1/2 + j] - min)*id;
const float x0 = (float(data_s[src_idx + 0 + j]) - min)*id;
const float x1 = (float(data_s[src_idx + QUANT_K_Q5_1/2 + j]) - min)*id;
const uint xi0 = uint(x0 + 0.5);
const uint xi1 = uint(x1 + 0.5);
@@ -166,7 +166,7 @@ void quantize(uint dst_idx, uint src_idx)
float amax = 0.0; // absolute max
[[unroll]] for (int j = 0; j < QUANT_K_Q8_0; j++) {
const float v = data_s[src_idx + j];
const float v = float(data_s[src_idx + j]);
amax = max(amax, abs(v));
}
@@ -176,7 +176,7 @@ void quantize(uint dst_idx, uint src_idx)
data_q[dst_idx].d = float16_t(d);
[[unroll]] for (int j = 0; j < QUANT_K_Q8_0; ++j) {
const float x0 = data_s[src_idx + j]*id;
const float x0 = float(data_s[src_idx + j])*id;
data_q[dst_idx].qs[j] = int8_t(round(x0));
}
@@ -189,7 +189,7 @@ void quantize(uint dst_idx, uint src_idx)
float sum_abs = 0.0;
[[unroll]] for (int j = 0; j < QUANT_K_Q1_0; j++) {
sum_abs += abs(data_s[src_idx + j]);
sum_abs += abs(float(data_s[src_idx + j]));
}
const float d = sum_abs / QUANT_K_Q1_0;
@@ -201,7 +201,7 @@ void quantize(uint dst_idx, uint src_idx)
}
[[unroll]] for (int j = 0; j < QUANT_K_Q1_0; ++j) {
if (data_s[src_idx + j] >= 0.0) {
if (float(data_s[src_idx + j]) >= 0.0) {
data_q[dst_idx].qs[j / 8] |= uint8_t(1 << (j % 8));
}
}
@@ -226,7 +226,7 @@ void quantize(uint dst_idx, uint src_idx)
float vmax = 0.0;
[[unroll]] for (int j = 0; j < QUANT_K_IQ4_NL; ++j) {
const float v = data_s[src_idx + j];
const float v = float(data_s[src_idx + j]);
if (amax < abs(v)) {
amax = abs(v);
vmax = v;
@@ -238,16 +238,16 @@ void quantize(uint dst_idx, uint src_idx)
float sumqx = 0, sumq2 = 0;
[[unroll]] for (int j = 0; j < QUANT_K_IQ4_NL/2; ++j) {
const float x0 = data_s[src_idx + 0 + j]*id;
const float x1 = data_s[src_idx + QUANT_K_IQ4_NL/2 + j]*id;
const float x0 = float(data_s[src_idx + 0 + j])*id;
const float x1 = float(data_s[src_idx + QUANT_K_IQ4_NL/2 + j])*id;
const uint xi0 = best_index(x0);
const uint xi1 = best_index(x1);
data_q[dst_idx].qs[j] = uint8_t(xi0 | (xi1 << 4));
const float v0 = kvalues_iq4nl[xi0];
const float v1 = kvalues_iq4nl[xi1];
const float w0 = data_s[src_idx + 0 + j]*data_s[src_idx + 0 + j];
const float w1 = data_s[src_idx + QUANT_K_IQ4_NL/2 + j]*data_s[src_idx + QUANT_K_IQ4_NL/2 + j];
sumqx += w0*v0*data_s[src_idx + j] + w1*v1*data_s[src_idx + QUANT_K_IQ4_NL/2 + j];
const float w0 = float(data_s[src_idx + 0 + j])*float(data_s[src_idx + 0 + j]);
const float w1 = float(data_s[src_idx + QUANT_K_IQ4_NL/2 + j])*float(data_s[src_idx + QUANT_K_IQ4_NL/2 + j]);
sumqx += w0*v0*float(data_s[src_idx + j]) + w1*v1*float(data_s[src_idx + QUANT_K_IQ4_NL/2 + j]);
sumq2 += w0*v0*v0 + w1*v1*v1;
}
@@ -259,14 +259,14 @@ void quantize(uint dst_idx, uint src_idx)
#if defined(DATA_A_F32) || defined(DATA_A_F16)
void quantize(uint dst_idx, uint src_idx)
{
data_q[dst_idx] = A_TYPE(data_s[src_idx]);
data_q[dst_idx] = A_TYPE(float(data_s[src_idx]));
}
#endif
#if defined(DATA_A_BF16)
void quantize(uint dst_idx, uint src_idx)
{
data_q[dst_idx] = A_TYPE(fp32_to_bf16(data_s[src_idx]));
data_q[dst_idx] = A_TYPE(fp32_to_bf16(float(data_s[src_idx])));
}
#endif
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+6 -4
View File
@@ -824,13 +824,15 @@ void process_shaders() {
string_to_spv("cpy_transpose_32", "copy_transpose.comp", {{"A_TYPE", "uint"}, {"D_TYPE", "uint"}});
for (std::string t : {"q1_0", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"S_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("cpy_" + t + "_f32", "copy_from_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
}
for (std::string t : {"f32", "f16", "bf16", "q1_0", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
string_to_spv("set_rows_" + t + "_i32", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uint"}, {"B_SIZE", "32"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("set_rows_" + t + "_i64", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uvec2"}, {"B_SIZE", "64"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
for (auto src : {std::pair{"f32", "float"}, std::pair{"f16", "float16_t"}}) {
for (std::string dst : {"f32", "f16", "bf16", "q1_0", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
string_to_spv("set_rows_" + std::string(src.first) + "_" + dst + "_i32", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(dst), "1"}, {"B_TYPE", "uint"}, {"B_SIZE", "32"}, {"S_TYPE", src.second}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("set_rows_" + std::string(src.first) + "_" + dst + "_i64", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(dst), "1"}, {"B_TYPE", "uvec2"}, {"B_SIZE", "64"}, {"S_TYPE", src.second}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
}
}
auto get_type_str = [](bool f16) {
tests/test-backend-ops.cpp
+17 -13
View File
@@ -2423,11 +2423,10 @@ struct test_set_rows : public test_case {
void initialize_tensors(ggml_context * ctx) override {
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
if (ggml_is_view_op(t->op)) {
continue;
}
if (t->type == GGML_TYPE_I64 || t->type == GGML_TYPE_I32) {
if (ggml_is_view_op(t->op)) {
continue;
}
init_set_rows_row_ids(t, ne[1]);
} else {
init_tensor_uniform(t);
@@ -2450,6 +2449,9 @@ struct test_set_rows : public test_case {
err_estimate /= 8.0f;
}
err_estimate *= err_estimate;
if (type_src == GGML_TYPE_F16) {
err_estimate *= 16.0f;
}
err_estimate /= 0.25f*float(ne[0] * r * ne[2]*nr23[0] * ne[3]*nr23[1]);
return err_estimate;
}
@@ -7928,17 +7930,19 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_I64, { 1, 8, 1, 3 }, { 1, 1 }, 2, false));
test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_I32, { 1, 8, 1, 3 }, { 1, 1 }, 2, false));
test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, GGML_TYPE_Q8_0, GGML_TYPE_I32, { 256, 5, 1, 3 }, { 1, 1, }, 1, false));
for (ggml_type type : all_types) {
for (int b : {1, 7}) {
for (bool v : {false, true}) {
test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, type, GGML_TYPE_I64, { 256, 5, b, 3 }, { 1, 1, }, 1, v));
test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, type, GGML_TYPE_I64, { 256, 11, 1, b }, { 2, 3, }, 7, v));
for (ggml_type src_type : {GGML_TYPE_F16, GGML_TYPE_F32}) {
for (ggml_type type : all_types) {
for (int b : {1, 7}) {
for (bool v : {false, true}) {
test_cases.emplace_back(new test_set_rows(src_type, type, GGML_TYPE_I64, { 256, 5, b, 3 }, { 1, 1, }, 1, v));
test_cases.emplace_back(new test_set_rows(src_type, type, GGML_TYPE_I64, { 256, 11, 1, b }, { 2, 3, }, 7, v));
test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, type, GGML_TYPE_I64, { 3*ggml_blck_size(type), 3, b, 1 }, { 2, 3, }, 2, v));
test_cases.emplace_back(new test_set_rows(src_type, type, GGML_TYPE_I64, { 3*ggml_blck_size(type), 3, b, 1 }, { 2, 3, }, 2, v));
if (ggml_blck_size(type) == 1) {
test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, type, GGML_TYPE_I64, { 31, 3, b, 1 }, { 2, 3, }, 2, v));
test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, type, GGML_TYPE_I64, { 33, 5, 1, b }, { 2, 3, }, 1, v));
if (ggml_blck_size(type) == 1) {
test_cases.emplace_back(new test_set_rows(src_type, type, GGML_TYPE_I64, { 31, 3, b, 1 }, { 2, 3, }, 2, v));
test_cases.emplace_back(new test_set_rows(src_type, type, GGML_TYPE_I64, { 33, 5, 1, b }, { 2, 3, }, 1, v));
}
}
}
}