Skip to content

vulkan: mul_mat_id coopmat2 optimizations #15546

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Aug 31, 2025
Merged

vulkan: mul_mat_id coopmat2 optimizations #15546

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
73e07ad
vulkan: mul_mat_id coopmat2 optimizations
jeffbolznv Aug 24, 2025
e5f97df
Also add a path for BN/4 - worth a couple more percent
jeffbolznv Aug 29, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2225,7 +2225,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
s_mmq_wg_denoms_k = { 32, 64, 1 };

// spec constants and tile sizes for quant matmul_id
l_warptile_mmqid = { 256, 128, 128, 16, 0, device->subgroup_size };
l_warptile_mmqid = { 256, 128, 128, 16, 1, device->subgroup_size };
m_warptile_mmqid = { 256, 128, 64, 16, 0, device->subgroup_size };
s_warptile_mmqid = { 256, 128, 64, 16, 0, device->subgroup_size };
l_mmqid_wg_denoms = { 128, 128, 1 };
Expand Down
97 changes: 92 additions & 5 deletions ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -444,18 +444,105 @@ void main() {

tensorLayoutBClamp = setTensorLayoutStrideNV(tensorLayoutBClamp, stride_b, 1);

coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator> sum;
sum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator>(0.0);

uint k_iters = (end_k - start_k + BK - 1) / BK;

fetch_scales(ir * BM, pos_a, stride_a, start_k, tid, false);
store_scales(tid);

#ifdef MUL_MAT_ID
if (enable_smaller_matrices && ic * BN + BNover4 >= _ne1) {
coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BNover4, gl_MatrixUseAccumulator> sum;
sum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BNover4, gl_MatrixUseAccumulator>(0.0);

[[dont_unroll]]
for (uint block_k = start_k, i = 0; i < k_iters; block_k += BK, ++i) {

if ((block_k % QUANT_K) == 0) {
store_scales(tid);
}
if (block_k + BK < end_k && ((block_k + BK) % QUANT_K) == 0) {
fetch_scales(ir * BM, pos_a, stride_a, block_k + BK, tid, false);
}

if ((ir + 1) * BM <= p.M && block_k + BK <= end_k) {
coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;

coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB);

sum = coopMatMulAdd(mat_a, mat_b, sum);
} else {
coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;

coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB);

sum = coopMatMulAdd(mat_a, mat_b, sum);
}
}

// Convert from ACC_TYPE to D_TYPE
coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover4, gl_MatrixUseAccumulator> mat_d;
mat_d = coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover4, gl_MatrixUseAccumulator>(sum);

// Call callback to store each element, remapping row through shared memory
coopMatPerElementNV(mat_d, mat_d, perElemOpD, ir, ic);
return;
}
if (enable_smaller_matrices && ic * BN + BNover2 >= _ne1) {
coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BNover2, gl_MatrixUseAccumulator> sum;
sum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BNover2, gl_MatrixUseAccumulator>(0.0);

[[dont_unroll]]
for (uint block_k = start_k, i = 0; i < k_iters; block_k += BK, ++i) {

if ((block_k % QUANT_K) == 0) {
store_scales(tid);
}
if (block_k + BK < end_k && ((block_k + BK) % QUANT_K) == 0) {
fetch_scales(ir * BM, pos_a, stride_a, block_k + BK, tid, false);
}

if ((ir + 1) * BM <= p.M && block_k + BK <= end_k) {
coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;

coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);

sum = coopMatMulAdd(mat_a, mat_b, sum);
} else {
coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;

coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);

sum = coopMatMulAdd(mat_a, mat_b, sum);
}
}

// Convert from ACC_TYPE to D_TYPE
coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover2, gl_MatrixUseAccumulator> mat_d;
mat_d = coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover2, gl_MatrixUseAccumulator>(sum);

// Call callback to store each element, remapping row through shared memory
coopMatPerElementNV(mat_d, mat_d, perElemOpD, ir, ic);
return;
}
#endif
coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator> sum;
sum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator>(0.0);

[[dont_unroll]]
for (uint block_k = start_k, i = 0; i < k_iters; block_k += BK, ++i) {

store_scales(tid);
if (block_k + BK < end_k) {
if ((block_k % QUANT_K) == 0) {
store_scales(tid);
}
if (block_k + BK < end_k && ((block_k + BK) % QUANT_K) == 0) {
fetch_scales(ir * BM, pos_a, stride_a, block_k + BK, tid, false);
}

Expand Down
Loading