Star873
1772
代码
代码
Issues228
Pull Requests330
Actions
项目讨论
Wiki
分析
Star873
1772
已合并
[CANNBOT]codecheck扫描修改 #5907
wangweidong创建于 6月9日
[CANNBOT]codecheck扫描修改 #5907
已合并
wangweidong创建于 6月9日
从gcw_YBIAEfqJ/ops-nncodecheck合入到cann/ops-nnmaster
共 18 个文件变更+91-51
Mactivation/bnll/examples/arch35/test_geir_bnll.cpp+2-2
@@ -67,11 +67,11 @@ using std::vector;
67 input.push_back(tensor_placeholder##intputIndex); \67 input.push_back(tensor_placeholder##intputIndex); \
68 graph.AddOp(placeholder##intputIndex); \68 graph.AddOp(placeholder##intputIndex); \
69 op1.set_input_##intputName(placeholder##intputIndex); \69 op1.set_input_##intputName(placeholder##intputIndex); \
70- inputs.push_back(placeholder##intputIndex);70+ inputs.push_back(placeholder##intputIndex)
71 71 
72#define ADD_OUTPUT(outputIndex, outputName, outputDtype, outputShape) \72#define ADD_OUTPUT(outputIndex, outputName, outputDtype, outputShape) \
73 TensorDesc outputName##outputIndex##_desc = TensorDesc(ge::Shape(outputShape), FORMAT_ND, outputDtype); \73 TensorDesc outputName##outputIndex##_desc = TensorDesc(ge::Shape(outputShape), FORMAT_ND, outputDtype); \
74- op1.update_output_desc_##outputName(outputName##outputIndex##_desc);74+ op1.update_output_desc_##outputName(outputName##outputIndex##_desc)
75 75 
76string GetTime()76string GetTime()
77{77{
Mactivation/bnll/examples/test_geir_bnll.cpp+2-2
@@ -67,11 +67,11 @@ using std::vector;
67 input.push_back(tensor_placeholder##intputIndex); \67 input.push_back(tensor_placeholder##intputIndex); \
68 graph.AddOp(placeholder##intputIndex); \68 graph.AddOp(placeholder##intputIndex); \
69 op1.set_input_##intputName(placeholder##intputIndex); \69 op1.set_input_##intputName(placeholder##intputIndex); \
70- inputs.push_back(placeholder##intputIndex);70+ inputs.push_back(placeholder##intputIndex)
71 71 
72#define ADD_OUTPUT(outputIndex, outputName, outputDtype, outputShape) \72#define ADD_OUTPUT(outputIndex, outputName, outputDtype, outputShape) \
73 TensorDesc outputName##outputIndex##_desc = TensorDesc(ge::Shape(outputShape), FORMAT_ND, outputDtype); \73 TensorDesc outputName##outputIndex##_desc = TensorDesc(ge::Shape(outputShape), FORMAT_ND, outputDtype); \
74- op1.update_output_desc_##outputName(outputName##outputIndex##_desc);74+ op1.update_output_desc_##outputName(outputName##outputIndex##_desc)
75 75 
76string GetTime()76string GetTime()
77{77{
Mactivation/bnll/op_host/arch35/bnll_tiling_arch35.cpp+7-4
@@ -29,6 +29,8 @@ constexpr uint32_t WS_SYS_SIZE = 0U;
29constexpr int64_t COMPUTE_TYPE_SIZE = 4;29constexpr int64_t COMPUTE_TYPE_SIZE = 4;
30constexpr int64_t MIN_SPLIT_THRESHOLD = 1024;30constexpr int64_t MIN_SPLIT_THRESHOLD = 1024;
31constexpr int64_t COMPARE_ALIGN_ELEMENTS = 256 / COMPUTE_TYPE_SIZE;31constexpr int64_t COMPARE_ALIGN_ELEMENTS = 256 / COMPUTE_TYPE_SIZE;
32+constexpr int64_t BUFFER_NUM_DB = 9; // 双缓冲 UB 划分份数
33+constexpr int64_t BUFFER_NUM_SB = 7; // 单缓冲 UB 划分份数
32 34 
33static const gert::Shape g_vec_1_shape = {1};35static const gert::Shape g_vec_1_shape = {1};
34 36 
@@ -76,15 +78,15 @@ static ge::graphStatus BNLLTilingFunc(gert::TilingContext* context)
76 OP_CHECK_NULL_WITH_CONTEXT(context, currentWorkspace);78 OP_CHECK_NULL_WITH_CONTEXT(context, currentWorkspace);
77 79 
78 int64_t usedCoreNum = 1;80 int64_t usedCoreNum = 1;
79- uint64_t useDoubleBuffer = 0;81+ bool useDoubleBuffer = false;
80 82 
81 if (totalIdx > 0) {83 if (totalIdx > 0) {
82 int64_t ubBlockSize = GetUbBlockSize(context);84 int64_t ubBlockSize = GetUbBlockSize(context);
83 tiling->totalNum = totalIdx;85 tiling->totalNum = totalIdx;
84 tiling->blockFactor = CeilAlign(CeilDiv(totalIdx, coreNum), ubBlockSize);86 tiling->blockFactor = CeilAlign(CeilDiv(totalIdx, coreNum), ubBlockSize);
85 usedCoreNum = CeilDiv(totalIdx, tiling->blockFactor);87 usedCoreNum = CeilDiv(totalIdx, tiling->blockFactor);
86- useDoubleBuffer = (totalIdx > MIN_SPLIT_THRESHOLD) ? 1 : 0;88+ useDoubleBuffer = (totalIdx > MIN_SPLIT_THRESHOLD);
87- int64_t bufferNum = useDoubleBuffer ? 9 : 7;89+ int64_t bufferNum = useDoubleBuffer ? BUFFER_NUM_DB : BUFFER_NUM_SB;
88 int64_t alignUnit = (ubBlockSize > COMPARE_ALIGN_ELEMENTS) ? ubBlockSize : COMPARE_ALIGN_ELEMENTS;90 int64_t alignUnit = (ubBlockSize > COMPARE_ALIGN_ELEMENTS) ? ubBlockSize : COMPARE_ALIGN_ELEMENTS;
89 tiling->ubFactor = FloorAlign(91 tiling->ubFactor = FloorAlign(
90 FloorDiv(static_cast<int64_t>(ubSize) / COMPUTE_TYPE_SIZE, bufferNum), alignUnit);92 FloorDiv(static_cast<int64_t>(ubSize) / COMPUTE_TYPE_SIZE, bufferNum), alignUnit);
@@ -94,7 +96,8 @@ static ge::graphStatus BNLLTilingFunc(gert::TilingContext* context)
94 }96 }
95 97 
96 context->SetBlockDim(usedCoreNum);98 context->SetBlockDim(usedCoreNum);
97- ASCENDC_TPL_SEL_PARAM(context, static_cast<uint32_t>(dataType), useDoubleBuffer);99+ uint32_t doubleBufferKey = useDoubleBuffer ? 1U : 0U;
100+ ASCENDC_TPL_SEL_PARAM(context, static_cast<uint32_t>(dataType), doubleBufferKey);
98 return ge::GRAPH_SUCCESS;101 return ge::GRAPH_SUCCESS;
99}102}
100 103 
Mactivation/bnll/op_kernel/arch35/bnll_tiling_data.h+2-0
@@ -13,6 +13,8 @@
13#ifndef _BNLL_TILING_DATA_H_13#ifndef _BNLL_TILING_DATA_H_
14#define _BNLL_TILING_DATA_H_14#define _BNLL_TILING_DATA_H_
15 15 
16+#include <cstdint>
17+ 
16struct BNLLTilingData {18struct BNLLTilingData {
17 int64_t totalNum = 0;19 int64_t totalNum = 0;
18 int64_t blockFactor = 0;20 int64_t blockFactor = 0;
Mactivation/celu/examples/arch35/test_geir_celu.cpp+2-3
@@ -57,7 +57,7 @@ using std::vector;
57 input.push_back(tensor_placeholder##intputIndex); \57 input.push_back(tensor_placeholder##intputIndex); \
58 graph.AddOp(placeholder##intputIndex); \58 graph.AddOp(placeholder##intputIndex); \
59 celu_op.set_input_##intputName(placeholder##intputIndex); \59 celu_op.set_input_##intputName(placeholder##intputIndex); \
60- inputs.push_back(placeholder##intputIndex);60+ inputs.push_back(placeholder##intputIndex)
61 61 
62#define LOG_PRINT(message, ...) \62#define LOG_PRINT(message, ...) \
63 do { \63 do { \
@@ -120,8 +120,7 @@ int32_t GenOnesDataFloat32(vector<int64_t> shapes, Tensor &input_tensor, TensorD
120 120 
121int32_t WriteDataToFile(string bin_file, uint64_t data_size, uint8_t *inputData)121int32_t WriteDataToFile(string bin_file, uint64_t data_size, uint8_t *inputData)
122{122{
123- FILE *fp;123+ FILE *fp = fopen(bin_file.c_str(), "w");
124- fp = fopen(bin_file.c_str(), "w");
125 if (fp == nullptr) {124 if (fp == nullptr) {
126 return FAILED;125 return FAILED;
127 }126 }
Mactivation/celu/examples/test_geir_celu.cpp+2-3
@@ -57,7 +57,7 @@ using std::vector;
57 input.push_back(tensor_placeholder##intputIndex); \57 input.push_back(tensor_placeholder##intputIndex); \
58 graph.AddOp(placeholder##intputIndex); \58 graph.AddOp(placeholder##intputIndex); \
59 celu_op.set_input_##intputName(placeholder##intputIndex); \59 celu_op.set_input_##intputName(placeholder##intputIndex); \
60- inputs.push_back(placeholder##intputIndex);60+ inputs.push_back(placeholder##intputIndex)
61 61 
62#define LOG_PRINT(message, ...) \62#define LOG_PRINT(message, ...) \
63 do { \63 do { \
@@ -120,8 +120,7 @@ int32_t GenOnesDataFloat32(vector<int64_t> shapes, Tensor &input_tensor, TensorD
120 120 
121int32_t WriteDataToFile(string bin_file, uint64_t data_size, uint8_t *inputData)121int32_t WriteDataToFile(string bin_file, uint64_t data_size, uint8_t *inputData)
122{122{
123- FILE *fp;123+ FILE *fp = fopen(bin_file.c_str(), "w");
124- fp = fopen(bin_file.c_str(), "w");
125 if (fp == nullptr) {124 if (fp == nullptr) {
126 return FAILED;125 return FAILED;
127 }126 }
Mactivation/fast_gelu_v2/examples/arch35/test_geir_fast_gelu_v2.cpp+6-3
@@ -37,6 +37,9 @@ using std::map;
37using std::string;37using std::string;
38using std::vector;38using std::vector;
39 39 
40+constexpr uint32_t FP32_BYTE_SIZE = 4;
41+constexpr uint32_t FP16_BYTE_SIZE = 2;
42+ 
40string GetTime()43string GetTime()
41{44{
42 time_t timep;45 time_t timep;
@@ -49,11 +52,11 @@ string GetTime()
49uint32_t GetDataTypeSize(DataType dt)52uint32_t GetDataTypeSize(DataType dt)
50{53{
51 if (dt == ge::DT_FLOAT) {54 if (dt == ge::DT_FLOAT) {
52- return 4;55+ return FP32_BYTE_SIZE;
53 } else if (dt == ge::DT_FLOAT16 || dt == ge::DT_BF16) {56 } else if (dt == ge::DT_FLOAT16 || dt == ge::DT_BF16) {
54- return 2;57+ return FP16_BYTE_SIZE;
55 }58 }
56- return 4;59+ return FP32_BYTE_SIZE;
57}60}
58 61 
59int32_t GenOnesData(vector<int64_t> shapes, Tensor &input_tensor, TensorDesc &input_tensor_desc,62int32_t GenOnesData(vector<int64_t> shapes, Tensor &input_tensor, TensorDesc &input_tensor_desc,
Mactivation/fast_gelu_v2/examples/test_geir_fast_gelu_v2.cpp+6-3
@@ -37,6 +37,9 @@ using std::map;
37using std::string;37using std::string;
38using std::vector;38using std::vector;
39 39 
40+constexpr uint32_t FP32_BYTE_SIZE = 4;
41+constexpr uint32_t FP16_BYTE_SIZE = 2;
42+ 
40string GetTime()43string GetTime()
41{44{
42 time_t timep;45 time_t timep;
@@ -49,11 +52,11 @@ string GetTime()
49uint32_t GetDataTypeSize(DataType dt)52uint32_t GetDataTypeSize(DataType dt)
50{53{
51 if (dt == ge::DT_FLOAT) {54 if (dt == ge::DT_FLOAT) {
52- return 4;55+ return FP32_BYTE_SIZE;
53 } else if (dt == ge::DT_FLOAT16 || dt == ge::DT_BF16) {56 } else if (dt == ge::DT_FLOAT16 || dt == ge::DT_BF16) {
54- return 2;57+ return FP16_BYTE_SIZE;
55 }58 }
56- return 4;59+ return FP32_BYTE_SIZE;
57}60}
58 61 
59int32_t GenOnesData(vector<int64_t> shapes, Tensor &input_tensor, TensorDesc &input_tensor_desc,62int32_t GenOnesData(vector<int64_t> shapes, Tensor &input_tensor, TensorDesc &input_tensor_desc,
Mactivation/fast_gelu_v2/op_host/arch35/fast_gelu_v2_tiling_arch35.cpp+12-4
@@ -32,6 +32,12 @@ using Ops::Base::GetUbBlockSize;
32 32 
33constexpr uint32_t WS_SYS_SIZE = 0U;33constexpr uint32_t WS_SYS_SIZE = 0U;
34constexpr int64_t MIN_SPLIT_THRESHOLD = 1024;34constexpr int64_t MIN_SPLIT_THRESHOLD = 1024;
35+constexpr int64_t FP32_BYTE_SIZE = 4;
36+constexpr int64_t FP16_BYTE_SIZE = 2;
37+constexpr int64_t IO_BUF_NUM = 2; // input + output 队列各 1 块
38+constexpr int64_t TMP_BUF_FP32_NUM = 5; // FP16/BF16 路径 5 块 fp32 中间 buffer
39+constexpr int64_t FP32_BUF_NUM_DB = 7; // FP32 路径双缓冲总块数
40+constexpr int64_t FP32_BUF_NUM_SB = 5; // FP32 路径单缓冲总块数
35 41 
36static const gert::Shape g_vec_1_shape = {1};42static const gert::Shape g_vec_1_shape = {1};
37 43 
@@ -104,16 +110,17 @@ static ge::graphStatus CalcTilingParams(gert::TilingContext* context, FastGeluV2
104 int64_t totalIdx, ge::DataType dataType, int64_t coreNum, uint64_t ubSize)110 int64_t totalIdx, ge::DataType dataType, int64_t coreNum, uint64_t ubSize)
105{111{
106 int64_t ubBlockSize = GetUbBlockSize(context);112 int64_t ubBlockSize = GetUbBlockSize(context);
107- uint64_t useDoubleBuffer = (totalIdx > MIN_SPLIT_THRESHOLD) ? 1 : 0;113+ bool useDoubleBuffer = (totalIdx > MIN_SPLIT_THRESHOLD);
108 114 
109 tiling->totalNum = totalIdx;115 tiling->totalNum = totalIdx;
110 tiling->blockFactor = CeilAlign(CeilDiv(totalIdx, coreNum), ubBlockSize);116 tiling->blockFactor = CeilAlign(CeilDiv(totalIdx, coreNum), ubBlockSize);
111 117 
112 int64_t bytesPerElement;118 int64_t bytesPerElement;
113 if (dataType == ge::DT_FLOAT16 || dataType == ge::DT_BF16) {119 if (dataType == ge::DT_FLOAT16 || dataType == ge::DT_BF16) {
114- bytesPerElement = (useDoubleBuffer ? 4 : 2) * 2 + 5 * 4;120+ int64_t ioBufNum = useDoubleBuffer ? IO_BUF_NUM * 2 : IO_BUF_NUM;
121+ bytesPerElement = ioBufNum * FP16_BYTE_SIZE + TMP_BUF_FP32_NUM * FP32_BYTE_SIZE;
115 } else {122 } else {
116- bytesPerElement = (useDoubleBuffer ? 7 : 5) * 4;123+ bytesPerElement = (useDoubleBuffer ? FP32_BUF_NUM_DB : FP32_BUF_NUM_SB) * FP32_BYTE_SIZE;
117 }124 }
118 tiling->ubFactor = FloorAlign(FloorDiv(static_cast<int64_t>(ubSize), bytesPerElement), ubBlockSize);125 tiling->ubFactor = FloorAlign(FloorDiv(static_cast<int64_t>(ubSize), bytesPerElement), ubBlockSize);
119 126 
@@ -124,7 +131,8 @@ static ge::graphStatus CalcTilingParams(gert::TilingContext* context, FastGeluV2
124 OP_LOGE(context, "FastGeluV2: blockFactor is 0"), return ge::GRAPH_FAILED);131 OP_LOGE(context, "FastGeluV2: blockFactor is 0"), return ge::GRAPH_FAILED);
125 132 
126 context->SetBlockDim(CeilDiv(totalIdx, tiling->blockFactor));133 context->SetBlockDim(CeilDiv(totalIdx, tiling->blockFactor));
127- ASCENDC_TPL_SEL_PARAM(context, static_cast<uint32_t>(dataType), static_cast<uint32_t>(useDoubleBuffer));134+ uint32_t doubleBufferKey = useDoubleBuffer ? 1U : 0U;
135+ ASCENDC_TPL_SEL_PARAM(context, static_cast<uint32_t>(dataType), doubleBufferKey);
128 return ge::GRAPH_SUCCESS;136 return ge::GRAPH_SUCCESS;
129}137}
130 138 
Mactivation/fast_gelu_v2/op_kernel/arch35/fast_gelu_v2_tiling_data.h+2-0
@@ -22,6 +22,8 @@
22#ifndef _FAST_GELU_V2_TILING_DATA_H_22#ifndef _FAST_GELU_V2_TILING_DATA_H_
23#define _FAST_GELU_V2_TILING_DATA_H_23#define _FAST_GELU_V2_TILING_DATA_H_
24 24 
25+#include <cstdint>
26+ 
25struct FastGeluV2TilingData {27struct FastGeluV2TilingData {
26 int64_t totalNum = 0; // Total number of input elements across all dimensions28 int64_t totalNum = 0; // Total number of input elements across all dimensions
27 int64_t blockFactor = 0; // Number of elements assigned to each AI Core (aligned to ubBlockSize)29 int64_t blockFactor = 0; // Number of elements assigned to each AI Core (aligned to ubBlockSize)
Mactivation/relu6/examples/arch35/test_geir_relu6.cpp+2-3
@@ -57,7 +57,7 @@ using std::vector;
57 input.push_back(tensor_placeholder##intputIndex); \57 input.push_back(tensor_placeholder##intputIndex); \
58 graph.AddOp(placeholder##intputIndex); \58 graph.AddOp(placeholder##intputIndex); \
59 relu6_op.set_input_##intputName(placeholder##intputIndex); \59 relu6_op.set_input_##intputName(placeholder##intputIndex); \
60- inputs.push_back(placeholder##intputIndex);60+ inputs.push_back(placeholder##intputIndex)
61 61 
62#define LOG_PRINT(message, ...) \62#define LOG_PRINT(message, ...) \
63 do { \63 do { \
@@ -120,8 +120,7 @@ int32_t GenOnesDataFloat32(vector<int64_t> shapes, Tensor &input_tensor, TensorD
120 120 
121int32_t WriteDataToFile(string bin_file, uint64_t data_size, uint8_t *inputData)121int32_t WriteDataToFile(string bin_file, uint64_t data_size, uint8_t *inputData)
122{122{
123- FILE *fp;123+ FILE *fp = fopen(bin_file.c_str(), "w");
124- fp = fopen(bin_file.c_str(), "w");
125 if (fp == nullptr) {124 if (fp == nullptr) {
126 return FAILED;125 return FAILED;
127 }126 }
Mactivation/relu6/examples/test_geir_relu6.cpp+2-3
@@ -57,7 +57,7 @@ using std::vector;
57 input.push_back(tensor_placeholder##intputIndex); \57 input.push_back(tensor_placeholder##intputIndex); \
58 graph.AddOp(placeholder##intputIndex); \58 graph.AddOp(placeholder##intputIndex); \
59 relu6_op.set_input_##intputName(placeholder##intputIndex); \59 relu6_op.set_input_##intputName(placeholder##intputIndex); \
60- inputs.push_back(placeholder##intputIndex);60+ inputs.push_back(placeholder##intputIndex)
61 61 
62#define LOG_PRINT(message, ...) \62#define LOG_PRINT(message, ...) \
63 do { \63 do { \
@@ -120,8 +120,7 @@ int32_t GenOnesDataFloat32(vector<int64_t> shapes, Tensor &input_tensor, TensorD
120 120 
121int32_t WriteDataToFile(string bin_file, uint64_t data_size, uint8_t *inputData)121int32_t WriteDataToFile(string bin_file, uint64_t data_size, uint8_t *inputData)
122{122{
123- FILE *fp;123+ FILE *fp = fopen(bin_file.c_str(), "w");
124- fp = fopen(bin_file.c_str(), "w");
125 if (fp == nullptr) {124 if (fp == nullptr) {
126 return FAILED;125 return FAILED;
127 }126 }
Mactivation/selu/op_host/arch35/selu_tiling_arch35.cpp+13-7
@@ -47,6 +47,12 @@ using Ops::Base::GetAivCoreNum;
47using Ops::Base::GetUbSize;47using Ops::Base::GetUbSize;
48 48 
49constexpr uint32_t WS_SYS_SIZE = 0U;49constexpr uint32_t WS_SYS_SIZE = 0U;
50+constexpr int64_t BLOCK_ALIGN_BYTES = 32; // UB 32 字节对齐
51+constexpr int64_t FP32_BYTE_SIZE = 4; // cast-to-fp32 中间计算字节宽
52+constexpr int64_t FP16_BYTE_SIZE = 2;
53+constexpr int64_t INT8_BYTE_SIZE = 1;
54+constexpr int64_t IO_BUF_NUM = 2; // inputQueue + outputQueue
55+constexpr int64_t TMP_BUF_NUM = 2; // tmpBuf1 + tmpBuf2
50 56 
51static const gert::Shape g_vec_1_shape = {1};57static const gert::Shape g_vec_1_shape = {1};
52 58 
@@ -107,30 +113,30 @@ static ge::graphStatus ComputeTiling(gert::TilingContext* context, SeluTilingDat
107 uint64_t ubSize, int64_t coreNum)113 uint64_t ubSize, int64_t coreNum)
108{114{
109 // 用户约定：所有非 fp32 dtype 中间都走 cast-to-fp32，computeTypeSize 固定为 4115 // 用户约定：所有非 fp32 dtype 中间都走 cast-to-fp32，computeTypeSize 固定为 4
110- int64_t typeSize = 4;116+ int64_t typeSize = FP32_BYTE_SIZE;
111- int64_t computeTypeSize = 4;117+ int64_t computeTypeSize = FP32_BYTE_SIZE;
112 switch (dataType) {118 switch (dataType) {
113 case ge::DT_FLOAT:119 case ge::DT_FLOAT:
114 case ge::DT_INT32:120 case ge::DT_INT32:
115- typeSize = 4;121+ typeSize = FP32_BYTE_SIZE;
116 break;122 break;
117 case ge::DT_FLOAT16:123 case ge::DT_FLOAT16:
118 case ge::DT_BF16:124 case ge::DT_BF16:
119- typeSize = 2;125+ typeSize = FP16_BYTE_SIZE;
120 break;126 break;
121 case ge::DT_INT8:127 case ge::DT_INT8:
122- typeSize = 1;128+ typeSize = INT8_BYTE_SIZE;
123 break;129 break;
124 default:130 default:
125 OP_LOGE(context, "Selu: unexpected dtype %d", static_cast<int>(dataType));131 OP_LOGE(context, "Selu: unexpected dtype %d", static_cast<int>(dataType));
126 return ge::GRAPH_FAILED;132 return ge::GRAPH_FAILED;
127 }133 }
128 134 
129- int64_t ubBlockSize = 32 / typeSize; // 32-byte alignment in elements135+ int64_t ubBlockSize = BLOCK_ALIGN_BYTES / typeSize; // 32-byte alignment in elements
130 int64_t blockFactor = CeilDiv(totalElements, coreNum);136 int64_t blockFactor = CeilDiv(totalElements, coreNum);
131 blockFactor = ((blockFactor + ubBlockSize - 1) / ubBlockSize) * ubBlockSize;137 blockFactor = ((blockFactor + ubBlockSize - 1) / ubBlockSize) * ubBlockSize;
132 int64_t usedCoreNum = CeilDiv(totalElements, blockFactor);138 int64_t usedCoreNum = CeilDiv(totalElements, blockFactor);
133- int64_t ubDivisor = (2 * typeSize + 2 * computeTypeSize) / typeSize;139+ int64_t ubDivisor = (IO_BUF_NUM * typeSize + TMP_BUF_NUM * computeTypeSize) / typeSize;
134 int64_t ubFactor = FloorAlign(140 int64_t ubFactor = FloorAlign(
135 FloorDiv(static_cast<int64_t>(ubSize) / typeSize, ubDivisor),141 FloorDiv(static_cast<int64_t>(ubSize) / typeSize, ubDivisor),
136 ubBlockSize);142 ubBlockSize);
Mactivation/selu/op_kernel/arch35/selu_tiling_data.h+2-0
@@ -21,6 +21,8 @@
21#ifndef _SELU_TILING_DATA_H_21#ifndef _SELU_TILING_DATA_H_
22#define _SELU_TILING_DATA_H_22#define _SELU_TILING_DATA_H_
23 23 
24+#include <cstdint>
25+ 
24struct SeluTilingData {26struct SeluTilingData {
25 int64_t totalElements = 0; // Total number of elements27 int64_t totalElements = 0; // Total number of elements
26 int64_t blockFactor = 0; // Number of elements per AI Core28 int64_t blockFactor = 0; // Number of elements per AI Core
Mactivation/softsign_grad/examples/arch35/test_geir_softsign_grad.cpp+8-5
@@ -46,6 +46,9 @@ using std::map;
46using std::string;46using std::string;
47using std::vector;47using std::vector;
48 48 
49+constexpr uint32_t FP32_BYTE_SIZE = 4;
50+constexpr uint32_t FP16_BYTE_SIZE = 2;
51+ 
49#define ADD_INPUT(inputIndex, inputName, inputDtype, inputShape) \52#define ADD_INPUT(inputIndex, inputName, inputDtype, inputShape) \
50 vector<int64_t> placeholder##inputIndex##_shape = inputShape; \53 vector<int64_t> placeholder##inputIndex##_shape = inputShape; \
51 auto placeholder##inputIndex = op::Data("placeholder" #inputIndex).set_attr_index(inputIndex - 1); \54 auto placeholder##inputIndex = op::Data("placeholder" #inputIndex).set_attr_index(inputIndex - 1); \
@@ -65,11 +68,11 @@ using std::vector;
65 input.push_back(tensor_placeholder##inputIndex); \68 input.push_back(tensor_placeholder##inputIndex); \
66 graph.AddOp(placeholder##inputIndex); \69 graph.AddOp(placeholder##inputIndex); \
67 softsignGradOp.set_input_##inputName(placeholder##inputIndex); \70 softsignGradOp.set_input_##inputName(placeholder##inputIndex); \
68- inputs.push_back(placeholder##inputIndex);71+ inputs.push_back(placeholder##inputIndex)
69 72 
70#define ADD_OUTPUT(outputIndex, outputName, outputDtype, outputShape) \73#define ADD_OUTPUT(outputIndex, outputName, outputDtype, outputShape) \
71 TensorDesc outputName##outputIndex##_desc = TensorDesc(ge::Shape(outputShape), FORMAT_ND, outputDtype); \74 TensorDesc outputName##outputIndex##_desc = TensorDesc(ge::Shape(outputShape), FORMAT_ND, outputDtype); \
72- softsignGradOp.update_output_desc_##outputName(outputName##outputIndex##_desc);75+ softsignGradOp.update_output_desc_##outputName(outputName##outputIndex##_desc)
73 76 
74string GetTime()77string GetTime()
75{78{
@@ -83,11 +86,11 @@ string GetTime()
83uint32_t GetDataTypeSize(DataType dt)86uint32_t GetDataTypeSize(DataType dt)
84{87{
85 if (dt == ge::DT_FLOAT) {88 if (dt == ge::DT_FLOAT) {
86- return 4;89+ return FP32_BYTE_SIZE;
87 } else if (dt == ge::DT_FLOAT16 || dt == ge::DT_BF16) {90 } else if (dt == ge::DT_FLOAT16 || dt == ge::DT_BF16) {
88- return 2;91+ return FP16_BYTE_SIZE;
89 }92 }
90- return 4;93+ return FP32_BYTE_SIZE;
91}94}
92 95 
93int32_t GenInputData(96int32_t GenInputData(
Mactivation/softsign_grad/examples/test_geir_softsign_grad.cpp+8-5
@@ -46,6 +46,9 @@ using std::map;
46using std::string;46using std::string;
47using std::vector;47using std::vector;
48 48 
49+constexpr uint32_t FP32_BYTE_SIZE = 4;
50+constexpr uint32_t FP16_BYTE_SIZE = 2;
51+ 
49#define ADD_INPUT(inputIndex, inputName, inputDtype, inputShape) \52#define ADD_INPUT(inputIndex, inputName, inputDtype, inputShape) \
50 vector<int64_t> placeholder##inputIndex##_shape = inputShape; \53 vector<int64_t> placeholder##inputIndex##_shape = inputShape; \
51 auto placeholder##inputIndex = op::Data("placeholder" #inputIndex).set_attr_index(inputIndex - 1); \54 auto placeholder##inputIndex = op::Data("placeholder" #inputIndex).set_attr_index(inputIndex - 1); \
@@ -65,11 +68,11 @@ using std::vector;
65 input.push_back(tensor_placeholder##inputIndex); \68 input.push_back(tensor_placeholder##inputIndex); \
66 graph.AddOp(placeholder##inputIndex); \69 graph.AddOp(placeholder##inputIndex); \
67 softsignGradOp.set_input_##inputName(placeholder##inputIndex); \70 softsignGradOp.set_input_##inputName(placeholder##inputIndex); \
68- inputs.push_back(placeholder##inputIndex);71+ inputs.push_back(placeholder##inputIndex)
69 72 
70#define ADD_OUTPUT(outputIndex, outputName, outputDtype, outputShape) \73#define ADD_OUTPUT(outputIndex, outputName, outputDtype, outputShape) \
71 TensorDesc outputName##outputIndex##_desc = TensorDesc(ge::Shape(outputShape), FORMAT_ND, outputDtype); \74 TensorDesc outputName##outputIndex##_desc = TensorDesc(ge::Shape(outputShape), FORMAT_ND, outputDtype); \
72- softsignGradOp.update_output_desc_##outputName(outputName##outputIndex##_desc);75+ softsignGradOp.update_output_desc_##outputName(outputName##outputIndex##_desc)
73 76 
74string GetTime()77string GetTime()
75{78{
@@ -83,11 +86,11 @@ string GetTime()
83uint32_t GetDataTypeSize(DataType dt)86uint32_t GetDataTypeSize(DataType dt)
84{87{
85 if (dt == ge::DT_FLOAT) {88 if (dt == ge::DT_FLOAT) {
86- return 4;89+ return FP32_BYTE_SIZE;
87 } else if (dt == ge::DT_FLOAT16 || dt == ge::DT_BF16) {90 } else if (dt == ge::DT_FLOAT16 || dt == ge::DT_BF16) {
88- return 2;91+ return FP16_BYTE_SIZE;
89 }92 }
90- return 4;93+ return FP32_BYTE_SIZE;
91}94}
92 95 
93int32_t GenInputData(96int32_t GenInputData(
Mactivation/softsign_grad/op_host/arch35/softsign_grad_tiling_arch35.cpp+11-4
@@ -29,6 +29,12 @@ using Ops::NN::OpTiling::EnsureNotScalar;
29constexpr int64_t FP32_SIZE = 4;29constexpr int64_t FP32_SIZE = 4;
30constexpr int64_t MIN_SPLIT_THRESHOLD = 1024;30constexpr int64_t MIN_SPLIT_THRESHOLD = 1024;
31 31 
32+// UB 划分份数：按是否需 Cast、是否开启 double buffer 区分缓冲块数
33+constexpr int64_t BUFFER_NUM_CAST_DB = 6;
34+constexpr int64_t BUFFER_NUM_CAST_SB = 5;
35+constexpr int64_t BUFFER_NUM_NOCAST_DB = 7;
36+constexpr int64_t BUFFER_NUM_NOCAST_SB = 4;
37+ 
32static ge::graphStatus GetPlatformInfo(gert::TilingContext* context, uint64_t& ubSize, int64_t& coreNum)38static ge::graphStatus GetPlatformInfo(gert::TilingContext* context, uint64_t& ubSize, int64_t& coreNum)
33{39{
34 fe::PlatFormInfos* platformInfoPtr = context->GetPlatformInfo();40 fe::PlatFormInfos* platformInfoPtr = context->GetPlatformInfo();
@@ -132,14 +138,14 @@ static ge::graphStatus SoftsignGradTilingFunc(gert::TilingContext* context)
132 int64_t usedCoreNum = CeilDiv(totalNum, tiling->blockFactor);138 int64_t usedCoreNum = CeilDiv(totalNum, tiling->blockFactor);
133 139 
134 int64_t ubBlockSize = GetUbBlockSize(context);140 int64_t ubBlockSize = GetUbBlockSize(context);
135- uint64_t useDoubleBuffer = (totalNum > MIN_SPLIT_THRESHOLD) ? 1 : 0;141+ bool useDoubleBuffer = (totalNum > MIN_SPLIT_THRESHOLD);
136 142 
137 int64_t bufferNum;143 int64_t bufferNum;
138 bool needsCast = (dataType != ge::DT_FLOAT);144 bool needsCast = (dataType != ge::DT_FLOAT);
139 if (needsCast) {145 if (needsCast) {
140- bufferNum = useDoubleBuffer ? 6 : 5;146+ bufferNum = useDoubleBuffer ? BUFFER_NUM_CAST_DB : BUFFER_NUM_CAST_SB;
141 } else {147 } else {
142- bufferNum = useDoubleBuffer ? 7 : 4;148+ bufferNum = useDoubleBuffer ? BUFFER_NUM_NOCAST_DB : BUFFER_NUM_NOCAST_SB;
143 }149 }
144 150 
145 tiling->ubFactor = FloorAlign(151 tiling->ubFactor = FloorAlign(
@@ -148,7 +154,8 @@ static ge::graphStatus SoftsignGradTilingFunc(gert::TilingContext* context)
148 154 
149 context->SetBlockDim(usedCoreNum);155 context->SetBlockDim(usedCoreNum);
150 156 
151- ASCENDC_TPL_SEL_PARAM(context, useDoubleBuffer);157+ uint64_t bufferModeKey = useDoubleBuffer ? 1U : 0U;
158+ ASCENDC_TPL_SEL_PARAM(context, bufferModeKey);
152 159 
153 return ge::GRAPH_SUCCESS;160 return ge::GRAPH_SUCCESS;
154}161}
Mactivation/softsign_grad/op_kernel/arch35/softsign_grad_tiling_data.h+2-0
@@ -18,6 +18,8 @@
18#ifndef _SOFTSIGN_GRAD_TILING_DATA_H_18#ifndef _SOFTSIGN_GRAD_TILING_DATA_H_
19#define _SOFTSIGN_GRAD_TILING_DATA_H_19#define _SOFTSIGN_GRAD_TILING_DATA_H_
20 20 
21+#include <cstdint>
22+ 
21struct SoftsignGradTilingData {23struct SoftsignGradTilingData {
22 int64_t totalNum = 0; // 总元素数量24 int64_t totalNum = 0; // 总元素数量
23 int64_t blockFactor = 0; // 每个核处理的元素数量25 int64_t blockFactor = 0; // 每个核处理的元素数量