CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-29 08:41:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: Add missing bounds checking to scalar/coopmat1 mul_mat_id (#15334)

Browse Source
This commit is contained in:
Jeff Bolz
2025-08-16 03:58:38 -05:00
committed by GitHub
parent 912ff8c119
commit 2e2b22ba66
2 changed files with 9 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
View File

@@ -801,7 +801,7 @@ void main() {
}
#else
const uint row_i = ic * BN + loadc_b + l;
if (row_i < _ne1) {
if (row_i < _ne1 && block + loadr_b < end_k) {
const u16vec2 row_idx = row_ids[row_i];
buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]);
} else {
@@ -875,10 +875,12 @@ void main() {
const u16vec2 row_idx = row_ids[row_i];
if (dr + cm_row * TM + store_r < p.M) {
data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
}
}
}
}
#else
const bool is_aligned = p.stride_d % 4 == 0; // Assumption: D_TYPE == float
@@ -925,7 +927,9 @@ void main() {
#endif // MUL_MAT_ID
[[unroll]] for (uint cr = 0; cr < TM; cr++) {
#ifdef MUL_MAT_ID
if (dr_warp + cr < p.M) {
data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
}
#else
if (dr_warp + cr < p.M && dc_warp + cc < p.N) {
data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);

2
tests/test-backend-ops.cpp
View File

@@ -5824,6 +5824,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 16, 16, b, 32, 1024, 16));
}
test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 1, 1, false, 8, 16, 1));
for (ggml_type type_a : base_types) {
for (ggml_type type_b : {GGML_TYPE_F32 /*, GGML_TYPE_F16 */}) {
for (int n_mats : {4, 8}) {