Skip to content

Avoid small MatMul batch parameter heap allocations #29085

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

Open
GopalakrishnanN wants to merge 1 commit into
base: main
Choose a base branch
from
+47 −28
Open

Avoid small MatMul batch parameter heap allocations #29085

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
75 changes: 47 additions & 28 deletions onnxruntime/core/providers/cpu/math/matmul.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,7 @@
// Licensed under the MIT License.

#include "core/providers/cpu/math/matmul.h"
#include "core/common/inlined_containers.h"
#include "core/providers/cpu/math/gemm_matmul_common.h"
#include "core/providers/cpu/math/matmul_helper.h"
#include "core/util/math.h"
Expand Down Expand Up @@ -329,39 +330,57 @@
}

if (can_use_fastmath_sbgemm) {
std::vector<MLAS_SBGEMM_DATA_PARAMS> data(max_len);
for (size_t i = 0; i < max_len; i++) {
data[i].BIsfp32 = !(bool(packed_b_));
data[i].AIsfp32 = true;
data[i].A = a_data + helper.LeftOffsets()[i];
data[i].lda = lda;
data[i].B = data[i].BIsfp32 ? b_data + helper.RightOffsets()[i] : (float*)packed_b_.get();
data[i].ldb = ldb;
data[i].C = y_data + helper.OutputOffsets()[i];
data[i].ldc = N;
data[i].Bias = nullptr;
data[i].OutputProcessor = nullptr;
data[i].BIsPacked = static_cast<bool>(packed_b_);
auto gemm_batch = [&](auto& data) {
for (size_t i = 0; i < max_len; i++) {
data[i].BIsfp32 = !(bool(packed_b_));

Check warning on line 335 in onnxruntime/core/providers/cpu/math/matmul.cc

View workflow job for this annotation

GitHub Actions / Optional Lint C++ 

[cpplint] reported by reviewdog 🐶
 Using deprecated casting style.  Use static_cast<bool>(...) instead  [readability/casting] [4]

Raw Output:
onnxruntime/core/providers/cpu/math/matmul.cc:335:  Using deprecated casting style.  Use static_cast<bool>(...) instead  [readability/casting] [4]
data[i].AIsfp32 = true;
data[i].A = a_data + helper.LeftOffsets()[i];
data[i].lda = lda;
data[i].B = data[i].BIsfp32 ? b_data + helper.RightOffsets()[i] : (float*)packed_b_.get();

Check warning on line 339 in onnxruntime/core/providers/cpu/math/matmul.cc

View workflow job for this annotation

GitHub Actions / Optional Lint C++ 

[cpplint] reported by reviewdog 🐶
 Using C-style cast.  Use reinterpret_cast<float*>(...) instead  [readability/casting] [4]

Raw Output:
onnxruntime/core/providers/cpu/math/matmul.cc:339:  Using C-style cast.  Use reinterpret_cast<float*>(...) instead  [readability/casting] [4]
data[i].ldb = ldb;
data[i].C = y_data + helper.OutputOffsets()[i];
data[i].ldc = N;
data[i].Bias = nullptr;
data[i].OutputProcessor = nullptr;
data[i].BIsPacked = static_cast<bool>(packed_b_);
}
MlasSBGemmBatch(trans_a ? CblasTrans : CblasNoTrans, trans_b ? CblasTrans : CblasNoTrans,
M, N, K, max_len, data.data(), thread_pool, &mlas_backend_kernel_selector_config_);
};

if (max_len <= 2) {
InlinedVector<MLAS_SBGEMM_DATA_PARAMS, 2> data(max_len);
gemm_batch(data);
} else {
std::vector<MLAS_SBGEMM_DATA_PARAMS> data(max_len);
gemm_batch(data);
}
Comment on lines +351 to 357
MlasSBGemmBatch(trans_a ? CblasTrans : CblasNoTrans, trans_b ? CblasTrans : CblasNoTrans,
M, N, K, max_len, data.data(), thread_pool, &mlas_backend_kernel_selector_config_);
} else
#endif
{
std::vector<MLAS_SGEMM_DATA_PARAMS> data(max_len);
for (size_t i = 0; i < max_len; i++) {
data[i].BIsPacked = bool(packed_b_);
data[i].A = a_data + helper.LeftOffsets()[i];
data[i].lda = lda;
data[i].B = data[i].BIsPacked ? (float*)packed_b_.get() : b_data + helper.RightOffsets()[i];
data[i].ldb = ldb;
data[i].C = y_data + helper.OutputOffsets()[i];
data[i].ldc = N;
data[i].alpha = alpha_attr_;
data[i].beta = 0.0f;
auto gemm_batch = [&](auto& data) {
for (size_t i = 0; i < max_len; i++) {
data[i].BIsPacked = bool(packed_b_);

Check warning on line 363 in onnxruntime/core/providers/cpu/math/matmul.cc

View workflow job for this annotation

GitHub Actions / Optional Lint C++ 

[cpplint] reported by reviewdog 🐶
 Using deprecated casting style.  Use static_cast<bool>(...) instead  [readability/casting] [4]

Raw Output:
onnxruntime/core/providers/cpu/math/matmul.cc:363:  Using deprecated casting style.  Use static_cast<bool>(...) instead  [readability/casting] [4]
data[i].A = a_data + helper.LeftOffsets()[i];
data[i].lda = lda;
data[i].B = data[i].BIsPacked ? (float*)packed_b_.get() : b_data + helper.RightOffsets()[i];

Check warning on line 366 in onnxruntime/core/providers/cpu/math/matmul.cc

View workflow job for this annotation

GitHub Actions / Optional Lint C++ 

[cpplint] reported by reviewdog 🐶
 Using C-style cast.  Use reinterpret_cast<float*>(...) instead  [readability/casting] [4]

Raw Output:
onnxruntime/core/providers/cpu/math/matmul.cc:366:  Using C-style cast.  Use reinterpret_cast<float*>(...) instead  [readability/casting] [4]
data[i].ldb = ldb;
data[i].C = y_data + helper.OutputOffsets()[i];
data[i].ldc = N;
data[i].alpha = alpha_attr_;
data[i].beta = 0.0f;
}
MlasGemmBatch(trans_a ? CblasTrans : CblasNoTrans, trans_b ? CblasTrans : CblasNoTrans,
M, N, K, data.data(), max_len, thread_pool, &mlas_backend_kernel_selector_config_);
};

if (max_len <= 2) {
InlinedVector<MLAS_SGEMM_DATA_PARAMS, 2> data(max_len);
gemm_batch(data);
} else {
std::vector<MLAS_SGEMM_DATA_PARAMS> data(max_len);

Check warning on line 381 in onnxruntime/core/providers/cpu/math/matmul.cc

View workflow job for this annotation

GitHub Actions / Optional Lint C++ 

[cpplint] reported by reviewdog 🐶
 Add #include <vector> for vector<>  [build/include_what_you_use] [4]

Raw Output:
onnxruntime/core/providers/cpu/math/matmul.cc:381:  Add #include <vector> for vector<>  [build/include_what_you_use] [4]
gemm_batch(data);
}
Comment on lines +377 to 383
MlasGemmBatch(trans_a ? CblasTrans : CblasNoTrans, trans_b ? CblasTrans : CblasNoTrans,
M, N, K, data.data(), max_len, thread_pool, &mlas_backend_kernel_selector_config_);
}
return Status::OK();
}
Expand Down
Loading