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

opencl: handle OOB write in noshuffle GEMV kernels (odd ne01) (#25640)

Browse Source
This commit is contained in:
Hongqiang Wang
2026-07-14 13:46:54 -07:00
committed by GitHub
parent a4ce2595c5
commit 00fa7cb284
11 changed files with 63 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl
+6 -1
View File
@@ -296,7 +296,12 @@ kernel void kernel_gemv_noshuffle_iq4_nl_f32(
// 2 outputs per fiber in wave 0
if (groupId == 0) {
dst = (global float*)((global char*)dst + offsetd);
vstore2(totalSum, 0, &(dst[gid * 2]));
// Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64,
// so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01
// and would overrun dst into the adjacent tensor. No-op / byte-identical when
// ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding).
if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0;
if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1;
}
}
ggml/src/ggml-opencl/kernels/gemv_noshuffle_q1_0_f32.cl
+5 -1
View File
@@ -116,6 +116,10 @@ __kernel void kernel_gemv_noshuffle_q1_0_f32(
if (groupId == 0) {
dst = (global float*)((global char*)dst + offsetd);
dst[gid] = totalSum;
// Guard the output row. The x-grid is padded to CEIL_DIV(M,wavesize)*wavesize,
// so when ne01 is not a multiple of the wave size the tail work-items run past
// row ne01 and would overrun dst into the adjacent tensor. No-op / byte-identical
// when ne01 is wave-aligned (no padding).
if (gid < M) dst[gid] = totalSum;
}
}
ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32.cl
+6 -1
View File
@@ -268,7 +268,12 @@ __kernel void kernel_gemv_noshuffle_q4_0_f32(
// 2 outputs per fiber in wave 0
if (groupId == 0) {
dst = (global float*)((global char*)dst + offsetd);
vstore2(totalSum, 0, &(dst[gid * 2]));
// Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64,
// so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01
// and would overrun dst into the adjacent tensor. No-op / byte-identical when
// ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding).
if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0;
if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1;
}
}
ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32_spec.cl
+5 -1
View File
@@ -262,7 +262,11 @@ __kernel void kernel_gemv_noshuffle_q4_0_f32(
// 2 outputs per fiber in wave 0
if (groupId == 0) {
dst = (global float*)((global char*)dst + offsetd);
vstore2(totalSum, 0, &(dst[gid * 2]));
// Guard the two output rows against the padded x-grid tail overrunning dst.
// The current shape specializations are all ne01 % 128 == 0 (no padding), so
// this is a no-op / byte-identical today; keep it in lockstep with the base kernel.
if (gid * 2 + 0 < ne01) dst[gid * 2 + 0] = totalSum.s0;
if (gid * 2 + 1 < ne01) dst[gid * 2 + 1] = totalSum.s1;
}
}
ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl
+6 -1
View File
@@ -277,7 +277,12 @@ kernel void kernel_gemv_noshuffle_q4_1_f32(
// 2 outputs per fiber in wave 0
if (groupId == 0) {
dst = (global float*)((global char*)dst + offsetd);
vstore2(totalSum, 0, &(dst[gid * 2]));
// Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64,
// so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01
// and would overrun dst into the adjacent tensor. No-op / byte-identical when
// ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding).
if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0;
if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1;
}
}
ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl
+6 -1
View File
@@ -312,7 +312,12 @@ kernel void kernel_gemv_noshuffle_q4_k_f32(
// 2 outputs per fiber in wave 0
if (groupId == 0) {
dst = (global float*)((global char*)dst + offsetd);
vstore2(totalSum, 0, &(dst[gid * 2]));
// Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64,
// so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01
// and would overrun dst into the adjacent tensor. No-op / byte-identical when
// ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding).
if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0;
if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1;
}
}
ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl
+6 -1
View File
@@ -285,7 +285,12 @@ __kernel void kernel_gemv_noshuffle_q5_0_f32(
// 2 outputs per fiber in wave 0
if (groupId == 0) {
dst = (global float*)((global char*)dst + offsetd);
vstore2(totalSum, 0, &(dst[gid * 2]));
// Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64,
// so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01
// and would overrun dst into the adjacent tensor. No-op / byte-identical when
// ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding).
if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0;
if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1;
}
}
ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl
+6 -1
View File
@@ -288,7 +288,12 @@ __kernel void kernel_gemv_noshuffle_q5_1_f32(
// 2 outputs per fiber in wave 0
if (groupId == 0) {
dst = (global float*)((global char*)dst + offsetd);
vstore2(totalSum, 0, &(dst[gid * 2]));
// Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64,
// so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01
// and would overrun dst into the adjacent tensor. No-op / byte-identical when
// ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding).
if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0;
if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1;
}
}
ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl
+6 -1
View File
@@ -321,6 +321,11 @@ kernel void kernel_gemv_noshuffle_q5_k_f32(
// 2 outputs per fiber in wave 0
if (groupId == 0) {
dst = (global float*)((global char*)dst + offsetd);
vstore2(totalSum, 0, &(dst[gid * 2]));
// Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64,
// so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01
// and would overrun dst into the adjacent tensor. No-op / byte-identical when
// ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding).
if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0;
if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1;
}
}
ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl
+6 -1
View File
@@ -288,6 +288,11 @@ kernel void kernel_gemv_noshuffle_q6_K_f32(
if (grp == 0) {
dst = (global float*)((global char*)dst + offsetd);
vstore2(total_sum, 0, &(dst[gid * 2]));
// Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64,
// so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01
// and would overrun dst into the adjacent tensor (garbage downstream).
// No-op / byte-identical when ne01 % 128 == 0 (no padding).
if (gid * 2 + 0 < ne01) dst[gid * 2 + 0] = total_sum.s0;
if (gid * 2 + 1 < ne01) dst[gid * 2 + 1] = total_sum.s1;
}
}
ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl
+5 -1
View File
@@ -190,6 +190,10 @@ __kernel void kernel_gemv_noshuffle_q8_0_f32(
// 1 outputs per fiber in wave 0
if (groupId == 0) {
dst = (global float*)((global char*)dst + offsetd);
dst[gid] = totalSum;
// Guard the output row. The x-grid is padded to CEIL_DIV(M,wavesize)*wavesize,
// so when ne01 is not a multiple of the wave size the tail work-items run past
// row ne01 and would overrun dst into the adjacent tensor. No-op / byte-identical
// when ne01 is wave-aligned (no padding).
if (gid < M) dst[gid] = totalSum;
}
}