Skip to content

Fix sparse mask handling in softmax kernel #33814

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
maxnick merged 4 commits into from
Feb 4, 2026
Merged

Fix sparse mask handling in softmax kernel #33814

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
5ae6469
Fix sparse mask handling in softmax kernel
mangguo321 Jan 26, 2026
cdd68e6
Add softmax kernel unit test with sparse mask
mangguo321 Jan 30, 2026
2bcf7f8
Fix typo
mangguo321 Jan 30, 2026
6301843
Skip the test on non-x86_64 platforms as xattention is supported only…
mangguo321 Feb 1, 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
39 changes: 24 additions & 15 deletions src/plugins/intel_cpu/src/nodes/kernels/scaled_attn/softmax_kernel.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -318,8 +318,9 @@ inline void scale_add2_reduce_max(float* a,
if (has_sparse_mask) { \
size_t mask_idx = (i + n * vec_len_f32_avx512) / sparse_block_size; \
uint8_t mask_val = sparse_mask[mask_idx]; \
__m512 v_mask_block = _mm512_set1_ps(mask_val ? 0.f : -FLT_MAX); \
v_a = _mm512_add_ps(v_a, v_mask_block); \
if (!mask_val) { \
v_a = v_nfltmax; \
} \
} \
if (has_causal_mask) { \
auto v_maski8 = \
Expand Down Expand Up @@ -355,8 +356,9 @@ inline void scale_add2_reduce_max(float* a,
if (has_sparse_mask) {
size_t mask_idx = i / sparse_block_size;
uint8_t mask_val = sparse_mask[mask_idx];
__m512 v_mask_block = _mm512_set1_ps(mask_val ? 0.f : -FLT_MAX);
v_a = _mm512_add_ps(v_a, v_mask_block);
if (!mask_val) {
v_a = v_nfltmax;
}
}

if (has_causal_mask) {
Expand Down Expand Up @@ -390,8 +392,9 @@ inline void scale_add2_reduce_max(float* a,
if (has_sparse_mask) {
size_t mask_idx = i / sparse_block_size;
uint8_t mask_val = sparse_mask[mask_idx];
__m512 v_mask_block = _mm512_set1_ps(mask_val ? 0.f : -FLT_MAX);
v_a = _mm512_add_ps(v_a, v_mask_block);
if (!mask_val) {
v_a = v_nfltmax;
}
}

if (has_causal_mask) {
Expand Down Expand Up @@ -439,8 +442,9 @@ inline void scale_add2_reduce_max(float* a,
if (has_sparse_mask) { \
size_t mask_idx = (i + n * vec_len_f32_avx2) / sparse_block_size; \
uint8_t mask_val = sparse_mask[mask_idx]; \
__m256 v_mask_block = _mm256_set1_ps(mask_val ? 0.f : -FLT_MAX); \
v_a = _mm256_add_ps(v_a, v_mask_block); \
if (!mask_val) { \
v_a = v_nfltmax; \
} \
} \
if (has_causal_mask) { \
auto v_maski8 = _mm_loadu_si128(reinterpret_cast<__m128i const*>(causal_mask + i + n * vec_len_f32_avx2)); \
Expand Down Expand Up @@ -476,8 +480,9 @@ inline void scale_add2_reduce_max(float* a,
if (has_sparse_mask) {
size_t mask_idx = i / sparse_block_size;
uint8_t mask_val = sparse_mask[mask_idx];
__m256 v_mask_block = _mm256_set1_ps(mask_val ? 0.f : -FLT_MAX);
v_a = _mm256_add_ps(v_a, v_mask_block);
if (!mask_val) {
v_a = v_nfltmax;
}
}

if (has_causal_mask) {
Expand Down Expand Up @@ -512,8 +517,9 @@ inline void scale_add2_reduce_max(float* a,
if (has_sparse_mask) {
size_t mask_idx = i / sparse_block_size;
uint8_t mask_val = sparse_mask[mask_idx];
__m256 v_mask_block = _mm256_set1_ps(mask_val ? 0.f : -FLT_MAX);
v_a = _mm256_add_ps(v_a, v_mask_block);
if (!mask_val) {
v_a = v_nfltmax;
}
}

if (has_causal_mask) {
Expand Down Expand Up @@ -560,8 +566,9 @@ inline void scale_add2_reduce_max(float* a,
if (has_sparse_mask) {
size_t mask_idx = i / sparse_block_size;
uint8_t mask_val = sparse_mask[mask_idx];
float32x4_t v_mask_block = vdupq_n_f32(mask_val ? 0.0F : -FLT_MAX);
v_a = vaddq_f32(v_a, v_mask_block);
if (!mask_val) {
v_a = v_nfltmax;
}
}

if (has_causal_mask) {
Expand Down Expand Up @@ -596,7 +603,9 @@ inline void scale_add2_reduce_max(float* a,
if (has_sparse_mask) {
size_t mask_idx = i / sparse_block_size;
uint8_t mask_val = sparse_mask[mask_idx];
a[i] += (mask_val ? 0.0F : -FLT_MAX);
if (!mask_val) {
a[i] = -FLT_MAX;
}
}

if (has_causal_mask) {
Expand Down
87 changes: 87 additions & 0 deletions src/plugins/intel_cpu/tests/unit/softmax_kernel_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,87 @@
// Copyright (C) 2018-2026 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//

#include "nodes/kernels/scaled_attn/softmax_kernel.hpp"

#include <cmath>
#include <vector>

#include "gtest/gtest.h"

namespace {
TEST(SoftmaxKernelTest, AttnSoftmaxKernelWithSparseMask) {
std::vector<float> input = {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f};
std::vector<float> output(input.size(), 0.0f);
std::vector<uint8_t> sparse_mask = {1, 0, 1, 0}; // Masking some elements, block size 2
float scale = 1.0f;
float* alibi = nullptr;
void* attn_mask = nullptr;
uint8_t* causal_mask = nullptr;
bool select_nfltmax_at_0 = false;
size_t len = input.size();
size_t total_size = input.size();
ov::element::Type attn_mask_prec = ov::element::f32;
ov::element::Type dst_precision = ov::element::f32;
const float* sink = nullptr;
float alibi_slope = 0.0f;
size_t sparse_block_size = 2;
ov::Extensions::Cpu::XARCH::attn_softmax_kernel<float>(input.data(),
output.data(),
scale,
alibi,
attn_mask,
causal_mask,
select_nfltmax_at_0,
len,
total_size,
attn_mask_prec,
dst_precision,
sink,
alibi_slope,
sparse_mask.data(),
sparse_block_size);
std::vector<float> expect_output = {0.00483724f, 0.013149f, 0.0f, 0.0f, 0.264104f, 0.71791f, 0.0f, 0.0f};
for (size_t i = 0; i < output.size(); ++i) {
EXPECT_NEAR(output[i], expect_output[i], 1e-5f);
}
}

TEST(SoftmaxKernelTest, AttnSoftmaxKernelWithNaNInputAndSparseMask) {
std::vector<float> input = {1.0f, 2.0f, std::nanf(""), 4.0f, 5.0f, 6.0f, 7.0f, 8.0f};
std::vector<float> output(input.size(), 0.0f);
std::vector<uint8_t> sparse_mask = {1, 0, 1, 0}; // Masking some elements, block size 2
float scale = 1.0f;
float* alibi = nullptr;
void* attn_mask = nullptr;
uint8_t* causal_mask = nullptr;
bool select_nfltmax_at_0 = false;
size_t len = input.size();
size_t total_size = input.size();
ov::element::Type attn_mask_prec = ov::element::f32;
ov::element::Type dst_precision = ov::element::f32;
const float* sink = nullptr;
float alibi_slope = 0.0f;
size_t sparse_block_size = 2;
ov::Extensions::Cpu::XARCH::attn_softmax_kernel<float>(input.data(),
output.data(),
scale,
alibi,
attn_mask,
causal_mask,
select_nfltmax_at_0,
len,
total_size,
attn_mask_prec,
dst_precision,
sink,
alibi_slope,
sparse_mask.data(),
sparse_block_size);
std::vector<float> expect_output = {0.00483724f, 0.013149f, 0.0f, 0.0f, 0.264104f, 0.71791f, 0.0f, 0.0f};
for (size_t i = 0; i < output.size(); ++i) {
EXPECT_NEAR(output[i], expect_output[i], 1e-5f);
}
}

} // namespace
Loading