Skip to content
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

[backport v3.7] SDPA fixes for regular scale tensor and alignment for K tensor #2284

Open
wants to merge 2 commits into
base: rls-v3.7-pc
Choose a base branch
from
+7 −4
Open

[backport v3.7] SDPA fixes for regular scale tensor and alignment for K tensor #2284

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
6fec1cd
xe: sdpa: Allow non-transposed scalars for the KQ matmul
umar456 Dec 11, 2024
e4f5533
xe: sdpa: Fix KQ gemm alignment for the K tensor
umar456 Dec 18, 2024
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 src/gpu/intel/ocl/micro_sdpa.cl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -181,7 +181,7 @@ micro_sdpa(const global KEY_DATA_T *K, const global half *Q,
uint lda = DST_S2;

#if KEY_SCALES || KEY_ZERO_POINTS
uint ldkq = div_up(d, KEY_GROUP_SIZE);
uint ldkq = KEY_D3;
#endif
#if VAL_SCALES || VAL_ZERO_POINTS
uint ldvq = div_up(d, VAL_GROUP_SIZE);
Expand Down
9 changes: 6 additions & 3 deletions src/gpu/intel/ocl/micro_sdpa.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -242,14 +242,14 @@ status_t micro_sdpa_t::pd_t::init_microkernels(impl::engine_t *engine) {
auto scale_dt = key_scales_dt();
problem_kq.Ta_scale = jit::convert_dnnl_to_kernel_type(scale_dt);
problem_kq.A_scale.alignment = uint8_t(types::data_type_size(scale_dt));
problem_kq.A_scale.layout = MatrixLayout::T;
problem_kq.A_scale.layout = MatrixLayout::N;
problem_kq.aScale2D = true;
}
if (with_key_zp()) {
auto zp_dt = key_zp_dt();
problem_kq.Tao = jit::convert_dnnl_to_kernel_type(zp_dt);
problem_kq.AO.alignment = uint8_t(types::data_type_size(zp_dt));
problem_kq.AO.layout = MatrixLayout::T;
problem_kq.AO.layout = MatrixLayout::N;
problem_kq.aoPtrDims = kq_common_zp ? 0 : 2;
problem_kq.aOffset = ABOffset::Calc;
}
Expand All @@ -264,7 +264,10 @@ status_t micro_sdpa_t::pd_t::init_microkernels(impl::engine_t *engine) {

problem_kq.B.layout = MatrixLayout::Pr;
problem_kq.C.layout = MatrixLayout::T;
problem_kq.A.setAlignment(alignmentForLD(d->head_size() * problem.Ta));
const memory_desc_wrapper key_mdw(key_md());
auto ldk = static_cast<int>(
gemm_desc_t::get_ld(*key_md()) * key_mdw.data_type_size());
problem_kq.A.setAlignment(alignmentForLD(ldk));
problem_kq.B.setAlignment(64); // Q is packed in VNNI format in SLM
problem_kq.B.crosspack = 2;
problem_kq.B.tileR = into<uint16_t>(d_max());
Expand Down
Loading