Skip to content

vulkan: add SHMEM_STRIDE_PAD/APPLY_SLM_A_RESHAPE for coopmat1 on Intel Xe #25380

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
0cc4m merged 5 commits into from
Aug 15, 2026
+163 −141
Merged

vulkan: add SHMEM_STRIDE_PAD/APPLY_SLM_A_RESHAPE for coopmat1 on Intel Xe #25380

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
b758072
vulkan: add SHMEM_STRIDE_PAD/APPLY_SLM_A_RESHAPE for coopmat mul_mm o…
fish-jiang Jul 7, 2026
940b575
vulkan: fix shmem estimate for Intel SHMEM_STRIDE_PAD=0 in matmul_shm…
fish-jiang Jul 8, 2026
7690d5b
cacheline aligned for shared kvalues_mxfp4
fish-jiang Jul 10, 2026
2dbd936
vulkan: fix OOB read in kvalues_mxfp4 init after cacheline padding
fish-jiang Jul 10, 2026
06cf97d
vulkan: restrict SLM-A reshape to Intel Windows driver, revert mxfp4 …
fish-jiang Jul 17, 2026
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
14 changes: 11 additions & 3 deletions ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3961,7 +3961,10 @@ static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vec
}

// Needs to be kept up to date on shader changes
const uint32_t bank_conflict_offset = device->coopmat_support ? 8 : 1;
// Needs to stay aligned with ggml_vk_mul_mm_spec.
const bool intel_shmem_stride_pad_zero = device->vendor_id == VK_VENDOR_ID_INTEL && device->coopmat_support &&
device->driver_id == vk::DriverId::eIntelProprietaryWindows;
const uint32_t bank_conflict_offset = intel_shmem_stride_pad_zero ? 0 : (device->coopmat_support ? 8 : 1);
const uint32_t type_size = device->fp16 ? sizeof(ggml_fp16_t) : sizeof(float);
const uint32_t warps = warptile[0] / warptile[10];

Expand Down Expand Up @@ -4578,8 +4581,13 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
}
#endif

auto const &ggml_vk_mul_mm_spec = [](std::vector<uint32_t> spec, bool aligned) {
spec.push_back(aligned ? 1u : 0u);
auto const &ggml_vk_mul_mm_spec = [&device](std::vector<uint32_t> spec, bool aligned) {
spec.push_back(aligned ? 1u : 0u); // constantID=11: ALIGNED
if (device->vendor_id == VK_VENDOR_ID_INTEL && device->coopmat_support &&
device->driver_id == vk::DriverId::eIntelProprietaryWindows) {
spec.push_back(0u); // constantID=12: SHMEM_STRIDE_PAD = 0
spec.push_back(1u); // constantID=13: APPLY_SLM_A_RESHAPE = true
}
return spec;
};

Expand Down
9 changes: 6 additions & 3 deletions ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -119,10 +119,13 @@ layout (constant_id = 3) const uint BK = 16; // Assumed to be 32 if working wit
#endif

#ifdef COOPMAT
#define SHMEM_STRIDE (BK / 2 + 4)
layout(constant_id = 12) const uint SHMEM_STRIDE_PAD = 4;
layout(constant_id = 13) const bool APPLY_SLM_A_RESHAPE = false;
#else
#define SHMEM_STRIDE (BK / 2 + 1)
const uint SHMEM_STRIDE_PAD = 1;
const bool APPLY_SLM_A_RESHAPE = false;
#endif
#define SHMEM_STRIDE (BK / 2 + SHMEM_STRIDE_PAD)

shared FLOAT_TYPEV2 buf_a[BM * SHMEM_STRIDE];
shared FLOAT_TYPEV2 buf_b[BN * SHMEM_STRIDE];
Expand Down Expand Up @@ -302,7 +305,7 @@ void main() {
[[unroll]] for (uint i = 0; i < BK; i += TK) {
[[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
// Load from shared into cache
coopMatLoad(cache_a, buf_a, (warp_r * WM + cm_row * TM) * SHMEM_STRIDE + i / 2, SHMEM_STRIDE, gl_CooperativeMatrixLayoutRowMajor);
coopMatLoad(cache_a, buf_a, a_shmem_index(warp_r * WM + cm_row * TM, i / 2), a_shmem_stride(), gl_CooperativeMatrixLayoutRowMajor);

[[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
coopMatLoad(cache_b, buf_b, (warp_c * WN + cm_col * TN) * SHMEM_STRIDE + i / 2, SHMEM_STRIDE, gl_CooperativeMatrixLayoutColumnMajor);
Expand Down
Loading
Loading