p85971406
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

!48240 [MS][LITE][parallel predict] pack weight: bug fix , not support grop convolution 

Merge pull request !48240 from yefeng/513-fix_grop_convolution_packweight
This commit is contained in:
i-robot 2023-01-31 07:05:11 +00:00 committed by Gitee
parent bfe7af1f8a f0b7bf0554
commit c63e226f0c
No known key found for this signature in database
GPG Key ID: 173E9B9CA92EEF8F
6 changed files with 41 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
mindspore/lite/src/litert/kernel/cpu/base/convolution_base.cc
View File

@ -50,7 +50,11 @@ ConvolutionBaseCPUKernel::~ConvolutionBaseCPUKernel() {
if (addr_map.find(reinterpret_cast<uintptr_t>(packed_weight_)) != addr_map.end()) {
FreeAlignedData(reinterpret_cast<void **>(&packed_weight_));
} else if (!op_parameter_->is_train_session_) {
lite::PackWeightManager::GetInstance()->Free(packed_weight_);
if (!is_sharing_pack_) {
free(packed_weight_);
} else {
lite::PackWeightManager::GetInstance()->Free(packed_weight_);
}
packed_weight_ = nullptr;
}
if (addr_map.find(reinterpret_cast<uintptr_t>(bias_data_)) != addr_map.end()) {

1
mindspore/lite/src/litert/kernel/cpu/base/convolution_base.h
View File

@ -82,6 +82,7 @@ class ConvolutionBaseCPUKernel : public LiteKernel {
std::unordered_map<uintptr_t, void *> addr_map;
void *packed_weight_ = nullptr;
bool weight_is_packed_ = false;
bool is_sharing_pack_ = true;
void *bias_data_ = nullptr;
const InnerContext *ctx_ = nullptr;
ConvParameter *conv_param_ = nullptr;

1
mindspore/lite/src/litert/kernel/cpu/fp32/convolution_sw_1x1_fp32.h
View File

@ -33,6 +33,7 @@ class ConvolutionSW1x1CPUKernel : public LiteKernel {
float *origin_weight, float *origin_bias)
: LiteKernel(parameter, inputs, outputs, ctx), origin_weight_(origin_weight), origin_bias_(origin_bias) {
matmul_base_ = CreateMatmulFp32CPUKernel(parameter, inputs, outputs, ctx);
matmul_base_->SetSharingPack(false);
}
~ConvolutionSW1x1CPUKernel() {
if (matmul_base_ != nullptr) {

10
mindspore/lite/src/litert/kernel/cpu/fp32/convolution_winograd_base_fp32.cc
View File

@ -260,8 +260,14 @@ int ConvolutionWinogradBaseCPUKernel::MallocWeightBiasData() {
if (!op_parameter_->is_train_session_) {
if (packed_weight_ == nullptr) {
CHECK_LESS_RETURN(MAX_MALLOC_SIZE, trans_matrix_data_size);
packed_weight_ = lite::PackWeightManager::GetInstance()->GetPackData(in_tensors_[1]->data(),
trans_matrix_data_size, &weight_is_packed_);
if (reinterpret_cast<ConvParameter *>(op_parameter_)->group_ > 1) {
packed_weight_ = malloc(trans_matrix_data_size);
weight_is_packed_ = false;
is_sharing_pack_ = false;
} else {
packed_weight_ = lite::PackWeightManager::GetInstance()->GetPackData(
in_tensors_[1]->data(), trans_matrix_data_size, &weight_is_packed_);
}
if (packed_weight_ == nullptr) {
MS_LOG(ERROR) << "malloc matrix_buffer failed.";
return RET_MEMORY_FAILED;

30
mindspore/lite/src/litert/kernel/cpu/fp32/matmul_fp32_base.cc
View File

@ -54,10 +54,18 @@ MatmulFp32BaseCPUKernel::~MatmulFp32BaseCPUKernel() {
matrix_c_.pack_ptr = nullptr;
}
if (params_->a_const_) {
lite::PackWeightManager::GetInstance()->Free(matrix_a_.pack_ptr);
if (is_sharing_pack_) {
lite::PackWeightManager::GetInstance()->Free(matrix_a_.pack_ptr);
} else {
free(matrix_a_.pack_ptr);
}
}
if (params_->b_const_) {
lite::PackWeightManager::GetInstance()->Free(matrix_b_.pack_ptr);
if (is_sharing_pack_) {
lite::PackWeightManager::GetInstance()->Free(matrix_b_.pack_ptr);
} else {
free(matrix_b_.pack_ptr);
}
}
}
@ -172,8 +180,13 @@ int MatmulFp32BaseCPUKernel::PackMatrixA() {
}
} else {
bool is_packed = false;
void *data = lite::PackWeightManager::GetInstance()->GetPackData(
in_tensors()[FIRST_INPUT]->data(), static_cast<size_t>(matrix_a_.pack_size) * sizeof(float), &is_packed);
void *data = nullptr;
if (is_sharing_pack_) {
data = lite::PackWeightManager::GetInstance()->GetPackData(
in_tensors()[FIRST_INPUT]->data(), static_cast<size_t>(matrix_a_.pack_size) * sizeof(float), &is_packed);
} else {
data = malloc(static_cast<size_t>(matrix_a_.pack_size) * sizeof(float));
}
matrix_a_.pack_ptr = reinterpret_cast<float *>(data);
if (matrix_a_.pack_ptr == nullptr) {
MS_LOG(ERROR) << "matrix a pack ptr is nullptr.";
@ -228,8 +241,13 @@ int MatmulFp32BaseCPUKernel::PackMatrixB() {
}
} else {
bool is_packed = false;
void *data = lite::PackWeightManager::GetInstance()->GetPackData(
in_tensors()[SECOND_INPUT]->data(), static_cast<size_t>(matrix_b_.pack_size) * sizeof(float), &is_packed);
void *data = nullptr;
if (is_sharing_pack_) {
data = lite::PackWeightManager::GetInstance()->GetPackData(
in_tensors()[SECOND_INPUT]->data(), static_cast<size_t>(matrix_b_.pack_size) * sizeof(float), &is_packed);
} else {
data = malloc(static_cast<size_t>(matrix_b_.pack_size) * sizeof(float));
}
matrix_b_.pack_ptr = reinterpret_cast<float *>(data);
if (matrix_b_.pack_ptr == nullptr) {
MS_LOG(ERROR) << "matrix b pack ptr is nullptr.";

2
mindspore/lite/src/litert/kernel/cpu/fp32/matmul_fp32_base.h
View File

@ -55,6 +55,7 @@ class MatmulFp32BaseCPUKernel : public LiteKernel {
static int InitBroadcastParams(const std::vector<int> &a_shape_const, const std::vector<int> &b_shape_const,
MatMulParameter *params, std::vector<int> *a_offsets, std::vector<int> *b_offsets);
int PackMatrixBParallelRunByBatch(int task_id) const;
inline void SetSharingPack(bool is_sharing) { is_sharing_pack_ = is_sharing; }
using ParallelRun = int (MatmulFp32BaseCPUKernel::*)(int task_id) const;
ParallelRun parallel_fun_ = nullptr;
@ -138,6 +139,7 @@ class MatmulFp32BaseCPUKernel : public LiteKernel {
MatrixPackFun matrix_b_pack_fun_ = nullptr;
float *conv1x1_origin_weight_ = nullptr;
float *conv1x1_origin_bias_ = nullptr;
bool is_sharing_pack_ = true;
};
} // namespace mindspore::kernel
#endif // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_CPU_FP32_MATMUL_FP32_BASE_H_