p60975123
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[MSLITE] npu innerContext comment bug

This commit is contained in:
ling 2021-08-16 10:45:10 +08:00
parent a34d737858
commit abb35e6688
19 changed files with 44 additions and 45 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/base/slice_base.c
View File

@ -43,7 +43,7 @@ void PadSliceParameterTo8D(SliceParameter *param) {
param->param_length_ = DIMENSION_8D;
}
void DoSlice(const void *input, void *output, SliceParameter *param, int thread_id, int data_size) {
void DoSlice(const void *input, void *output, const SliceParameter *param, int thread_id, int data_size) {
int8_t *int8_in = (int8_t *)input;
int8_t *int8_out = (int8_t *)output;
@ -94,14 +94,14 @@ void DoSlice(const void *input, void *output, SliceParameter *param, int thread_
}
}
static bool WhetherCopyByAxis(int begin[], int end[], const int shape[], int dim) {
static bool WhetherCopyByAxis(const int begin[], const int end[], const int shape[], int dim) {
for (int i = dim + 1; i < DIMENSION_8D; ++i) {
if (begin[i] != 0 || end[i] != shape[i]) return false;
}
return true;
}
void DoSliceNoParallel(const void *input, void *output, SliceParameter *param, int data_size) {
void DoSliceNoParallel(const void *input, void *output, const SliceParameter *param, int data_size) {
int8_t *int8_in = (int8_t *)input;
int8_t *int8_out = (int8_t *)output;

4
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/base/slice_base.h
View File

@ -25,8 +25,8 @@ extern "C" {
#endif
void PadSliceParameterTo8D(SliceParameter *param);
void DoSlice(const void *input, void *output, SliceParameter *param, int thread_id, int data_size);
void DoSliceNoParallel(const void *input, void *output, SliceParameter *param, int data_size);
void DoSlice(const void *input, void *output, const SliceParameter *param, int thread_id, int data_size);
void DoSliceNoParallel(const void *input, void *output, const SliceParameter *param, int data_size);
#ifdef __cplusplus
}
#endif

4
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/base/split_base.c
View File

@ -20,12 +20,12 @@
#include "nnacl/errorcode.h"
int DoSplit(void *in_data, void **out_data, const int *input_shape, int offset, int num_unit,
SplitParameter *split_param, int data_size) {
const SplitParameter *split_param, int data_size) {
int8_t *int8_in = (int8_t *)in_data;
int num_split = split_param->num_split_;
int *split_sizes = split_param->split_sizes_;
int *strides = split_param->strides_;
const int *strides = split_param->strides_;
int split_dim = split_param->split_dim_;
int in_stride = strides[split_dim];

2
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/base/split_base.h
View File

@ -24,7 +24,7 @@
extern "C" {
#endif
int DoSplit(void *in_data, void **out_data, const int *input_shape, int offset, int num_unit,
SplitParameter *split_param, int data_size);
const SplitParameter *split_param, int data_size);
#ifdef __cplusplus
}
#endif

2
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/base/split_with_over_lap_base.c
View File

@ -18,7 +18,7 @@
#include <string.h>
#include "nnacl/errorcode.h"
int DoSplitWithOverlapParallel(char *in_data, char **out_data, int slice_idx, SplitWithOverlapParameter *param,
int DoSplitWithOverlapParallel(char *in_data, char **out_data, int slice_idx, const SplitWithOverlapParameter *param,
const int *start_indices, const int *end_indices) {
if (in_data == NULL || out_data == NULL) {
return NNACL_NULL_PTR;

2
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/base/split_with_over_lap_base.h
View File

@ -23,7 +23,7 @@
#ifdef __cplusplus
extern "C" {
#endif
int DoSplitWithOverlapParallel(char *in_data, char **out_data, int slice_idx, SplitWithOverlapParameter *param,
int DoSplitWithOverlapParallel(char *in_data, char **out_data, int slice_idx, const SplitWithOverlapParameter *param,
const int *start_indices, const int *end_indices);
#ifdef __cplusplus
}

2
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/base/unstack_base.c
View File

@ -16,7 +16,7 @@
#include "nnacl/base/unstack_base.h"
void Unstack(const void *input, void **output, UnstackParameter *para, int data_size) {
void Unstack(const void *input, void **output, const UnstackParameter *para, int data_size) {
const int8_t *in_addr = (int8_t *)input;
for (int j = 0; j < para->num_; j++) {
int8_t *out_addr = (int8_t *)output[j];

2
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/base/unstack_base.h
View File

@ -24,7 +24,7 @@
#ifdef __cplusplus
extern "C" {
#endif
void Unstack(const void *input, void **output, UnstackParameter *para, int data_size);
void Unstack(const void *input, void **output, const UnstackParameter *para, int data_size);
#ifdef __cplusplus
}
#endif

2
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/int8/add_int8.c
View File

@ -78,7 +78,7 @@ void AddInt8OutputRounding(int32x4_t *out1, int32x4_t *out2, int32x4_t *out3, in
}
#endif
void AddInt8(const int8_t *input0, const int8_t *input1, int8_t *output, int size, AddQuantParameter *params) {
void AddInt8(const int8_t *input0, const int8_t *input1, int8_t *output, int size, const AddQuantParameter *params) {
int in0_left_shift = (1 << params->left_shift_) * (1 << params->in0_args_.left_shift_);
int in1_left_shift = (1 << params->left_shift_) * (1 << params->in1_args_.left_shift_);
int index = 0;

2
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/int8/add_int8.h
View File

@ -50,7 +50,7 @@ typedef struct AddQuantParameter {
extern "C" {
#endif
void AddInt8(const int8_t *input0, const int8_t *input1, int8_t *output, int size, AddQuantParameter *params);
void AddInt8(const int8_t *input0, const int8_t *input1, int8_t *output, int size, const AddQuantParameter *params);
void AddOptInt8(const int8_t *ptr_in, const int8_t element_in, int8_t *output, int size,
const AddQuantParameter *params, const AddQuantQrgs *ptr_args, const AddQuantQrgs *ele_args);

6
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/int8/conv3x3_int8.c
View File

@ -867,9 +867,9 @@ void Conv3x3Int8Gemm(int32_t *dst, const int16_t *src, const int16_t *weight, in
}
// int8 convolution 3x3
void Conv3x3Int8(int16_t *input_data, int16_t *transed_weight, const int32_t *bias_data, int8_t *output_data,
int16_t *tile_buffer, int16_t *block_unit_buffer, int32_t *tmp_dst_buffer, int8_t *tmp_out,
int task_id, const ConvParameter *conv_param) {
void Conv3x3Int8(const int16_t *input_data, const int16_t *transed_weight, const int32_t *bias_data,
int8_t *output_data, int16_t *tile_buffer, int16_t *block_unit_buffer, int32_t *tmp_dst_buffer,
int8_t *tmp_out, int task_id, const ConvParameter *conv_param) {
int ic8 = UP_DIV(conv_param->input_channel_, C8NUM);
int out_w_block = UP_DIV(conv_param->output_w_, OUPUT_UNIT);
int out_h_block = UP_DIV(conv_param->output_h_, OUPUT_UNIT);

6
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/int8/conv3x3_int8.h
View File

@ -37,9 +37,9 @@ extern "C" {
void Conv3x3Int8FilterTransform(const int16_t *weight_data, int16_t *trans_weight, int iC8, int output_channel,
int kernel_plane);
void Conv3x3Int8(int16_t *input_data, int16_t *transed_weight, const int32_t *bias_data, int8_t *output_data,
int16_t *tile_buffer, int16_t *block_unit_buffer, int32_t *tmp_dst_buffer, int8_t *tmp_out,
int task_id, const ConvParameter *conv_param);
void Conv3x3Int8(const int16_t *input_data, const int16_t *transed_weight, const int32_t *bias_data,
int8_t *output_data, int16_t *tile_buffer, int16_t *block_unit_buffer, int32_t *tmp_dst_buffer,
int8_t *tmp_out, int task_id, const ConvParameter *conv_param);
#ifdef __cplusplus
}

6
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/int8/deconv_int8.c
View File

@ -130,9 +130,9 @@ void DeConvPackInputSum(const int8_t *src, int32_t *dst, int32_t filter_zp, size
return;
}
int DeConvInt8(const int8_t *input, const int8_t *weight, int32_t *output, int32_t *weight_sum, int32_t *input_sum,
size_t act_row, size_t act_col, size_t act_deep, ConvParameter *conv_param,
MATMUL_OPT_R4_FUNC matmul_func) {
int DeConvInt8(const int8_t *input, const int8_t *weight, int32_t *output, const int32_t *weight_sum,
const int32_t *input_sum, size_t act_row, size_t act_col, size_t act_deep,
const ConvParameter *conv_param, MATMUL_OPT_R4_FUNC matmul_func) {
if (matmul_func != NULL) {
matmul_func(input, weight, output, act_row, act_col, act_deep, input_sum, weight_sum);
} else {

6
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/int8/deconv_int8.h
View File

@ -34,9 +34,9 @@ void DeConvPackInputSum(const int8_t *src, int32_t *dst, int32_t filter_zp, size
void DeConvWeightTransInt8(const int8_t *src, int8_t *dst, int input_channel, int output_channel, int plane,
bool support_optimize_);
int DeConvInt8(const int8_t *input, const int8_t *weight, int32_t *output, int32_t *weight_sum, int32_t *input_sum,
size_t act_row, size_t act_col, size_t act_deep, ConvParameter *conv_param,
MATMUL_OPT_R4_FUNC matmul_func);
int DeConvInt8(const int8_t *input, const int8_t *weight, int32_t *output, const int32_t *weight_sum,
const int32_t *input_sum, size_t act_row, size_t act_col, size_t act_deep,
const ConvParameter *conv_param, MATMUL_OPT_R4_FUNC matmul_func);
int DeConvPostInt8(const int32_t *src, const int32_t *bias, int32_t *tmp, int8_t *out, int output_channel,
ConvParameter *conv_param, bool support_optimize);
#ifdef __cplusplus

14
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/int8/matmul_int8.c
View File

@ -17,7 +17,7 @@
#include "nnacl/int8/matmul_int8.h"
#include "nnacl/int8/fixed_point.h"
void RowMajor2Row2x16MajorInt8(int8_t *src_ptr, int8_t *dst_ptr, int row, int col) {
void RowMajor2Row2x16MajorInt8(const int8_t *src_ptr, int8_t *dst_ptr, int row, int col) {
int col16 = UP_ROUND(col, C16NUM);
for (int r = 0; r < row; r++) {
int rd2 = r / C2NUM;
@ -32,7 +32,7 @@ void RowMajor2Row2x16MajorInt8(int8_t *src_ptr, int8_t *dst_ptr, int row, int co
}
}
void RowMajor2Col16x2MajorInt8(int8_t *src_ptr, int8_t *dst_ptr, int row, int col) {
void RowMajor2Col16x2MajorInt8(const int8_t *src_ptr, int8_t *dst_ptr, int row, int col) {
int row16 = UP_ROUND(row, C16NUM);
int stride = sizeof(int8_t) * C16NUM * C2NUM;
for (int r = 0; r < row; ++r) {
@ -60,9 +60,9 @@ void RowMajor2Row8x4MajorInt8(const int8_t *src_ptr, int8_t *dst_ptr, int row, i
}
}
void MatrixPack4x16UnitInt8(int8_t *src, int8_t *dst, int row, int col, int stride) {
void MatrixPack4x16UnitInt8(const int8_t *src, int8_t *dst, int row, int col, int stride) {
for (int r = 0; r < row; r++) {
int8_t *src_r = src + r * stride;
const int8_t *src_r = src + r * stride;
int8_t *dst_r = dst + r * C16NUM;
memcpy(dst_r, src_r, col * sizeof(int8_t));
}
@ -196,9 +196,9 @@ void MatMulInt8_16x4(const int8_t *a, const int8_t *b, int *dst, int row_4, int
}
void MatMulInt8_4x2_r(const int8_t *a, const int8_t *b, int8_t *dst, size_t row, size_t col, size_t deep_16,
size_t stride, const int32_t *input_sum, const int32_t *bias, int32_t *left_shift,
int32_t *right_shift, int32_t *multiplier, int32_t output_zp, int32_t mini, int32_t maxi,
bool peroc) {
size_t stride, const int32_t *input_sum, const int32_t *bias, const int32_t *left_shift,
const int32_t *right_shift, const int32_t *multiplier, int32_t output_zp, int32_t mini,
int32_t maxi, bool peroc) {
/* support per-layer && weight per-channel */
/* row4x16-major * row16x2-major => (int8)row-major*/
for (int r = 0; r < row; r++) {

10
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/int8/matmul_int8.h
View File

@ -48,12 +48,12 @@ void MatMulInt8_8x8_r(const int8_t *a, const int8_t *b, int8_t *dst, size_t row,
/* 4x16 16x2 -> 4x2 */
/* arm32 conv1x1 */
void RowMajor2Row2x16MajorInt8(int8_t *src_ptr, int8_t *dst_ptr, int row, int col);
void RowMajor2Col16x2MajorInt8(int8_t *src_ptr, int8_t *dst_ptr, int row, int col);
void RowMajor2Row2x16MajorInt8(const int8_t *src_ptr, int8_t *dst_ptr, int row, int col);
void RowMajor2Col16x2MajorInt8(const int8_t *src_ptr, int8_t *dst_ptr, int row, int col);
void MatMulInt8_4x2_r(const int8_t *a, const int8_t *b, int8_t *dst, size_t row, size_t col, size_t deep_16,
size_t stride, const int32_t *input_sum, const int32_t *bias, int32_t *left_shift,
int32_t *right_shift, int32_t *multiplier, int32_t output_zp, int32_t mini, int32_t maxi,
bool peroc);
size_t stride, const int32_t *input_sum, const int32_t *bias, const int32_t *left_shift,
const int32_t *right_shift, const int32_t *multiplier, int32_t output_zp, int32_t mini,
int32_t maxi, bool peroc);
/* 4x4 4x16 -> 4x16 */
/* optimize conv1x1 */

6
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/int8/pack_int8.c
View File

@ -807,7 +807,7 @@ void Im2ColPackUnitInt8Opt(const int8_t *input_data, int8_t *packed_input, int8_
}
}
void PackInputToC8Int8(const int8_t *input_data, int16_t *packed_input, ConvParameter *conv_param) {
void PackInputToC8Int8(const int8_t *input_data, int16_t *packed_input, const ConvParameter *conv_param) {
int in_batch = conv_param->input_batch_;
int in_channel = conv_param->input_channel_;
int in_h = conv_param->input_h_;
@ -961,7 +961,7 @@ void PackDepthwiseInt8Input(const int8_t *src, int16_t *dst, const ConvParameter
}
void PackDepthwiseInt8Weight(const int8_t *origin_weight, int16_t *packed_weight_, int plane, int channel,
ConvQuantArg *quant_qrg) {
const ConvQuantArg *quant_qrg) {
int weight_zp = quant_qrg->filter_quant_args_[0].zp_;
for (int c = 0; c < channel; c++) {
if (quant_qrg->per_channel_ & FILTER_PER_CHANNEL) {
@ -980,7 +980,7 @@ void PackDepthwiseInt8Weight(const int8_t *origin_weight, int16_t *packed_weight
}
void PackDeconvDepthwiseInt8Weight(const int8_t *origin_weight, int16_t *packed_weight_, int plane, int channel,
ConvQuantArg *quant_qrg) {
const ConvQuantArg *quant_qrg) {
int weight_zp = quant_qrg->filter_quant_args_[0].zp_;
for (int c = 0; c < channel; c++) {
if (quant_qrg->per_channel_ & FILTER_PER_CHANNEL) {

6
mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/int8/pack_int8.h
View File

@ -39,7 +39,7 @@ void PackNHWCToNCHWInt8(const void *src, void *dst, int batch, int plane, int ch
void PackInputSum16x4Int8(const int8_t *input, int32_t *input_sum, const int32_t *filter_zp,
const ConvParameter *conv_param);
void PackInputSum16x4PerLayer(const int8_t *src, int32_t *dst, int32_t filter_zp, size_t row4, size_t col16);
void PackInputToC8Int8(const int8_t *input_data, int16_t *packed_input, ConvParameter *conv_param);
void PackInputToC8Int8(const int8_t *input_data, int16_t *packed_input, const ConvParameter *conv_param);
void PackWeightToC8Int8(const int8_t *origin_weight_data, int16_t *packed_weight_data, const ConvParameter *conv_param);
void Im2ColPackUnitInt8Opt(const int8_t *input_data, int8_t *packed_input, int8_t *matmul_input, int real_cal_num,
int block_index, const int32_t *filter_zp, int32_t *input_sum,
@ -52,9 +52,9 @@ void PreSum4x16Int8Peroc(const int8_t *src, int32_t *sum, const int32_t *zp, siz
void PackDepthwiseInt8Input(const int8_t *src, int16_t *dst, const ConvParameter *conv_param);
void PackDepthwiseInt8Weight(const int8_t *origin_weight, int16_t *packed_weight_, int plane, int channel,
ConvQuantArg *quant_qrg);
const ConvQuantArg *quant_qrg);
void PackDeconvDepthwiseInt8Weight(const int8_t *origin_weight, int16_t *packed_weight_, int plane, int channel,
ConvQuantArg *quant_qrg);
const ConvQuantArg *quant_qrg);
#ifdef __cplusplus
}

1
mindspore/lite/src/inner_context.cc
View File

@ -205,7 +205,6 @@ bool InnerContext::IsGpuEnabled() const {
bool InnerContext::IsNpuEnabled() const {
#ifdef SUPPORT_NPU
// return IsUserSetNpu() && npu_manager_->IsSupportNPU();
return IsUserSetNpu();
#else
return false;