Star1321
2073
代码
代码
Issues742
Pull Requests635
Actions
项目讨论
Wiki
分析
Star1321
2073
已合并
修复GRU、GRUGard算子cleancode问题 #8851
chenyifeng创建于 8月18日
修复GRU、GRUGard算子cleancode问题 #8851
已合并
chenyifeng创建于 8月18日
从chen_0715/ops-nnfix_gru_codecheck_0818合入到cann/ops-nnmaster
共 5 个文件变更+30-31
Mrnn/gru/op_api/aclnn_gru.cpp+3-4
@@ -241,7 +241,7 @@ static bool CheckDims(const aclTensor* input, const aclTensorList* params, const
241 uint64_t bScale = hasBias ? 2 : 1;241 uint64_t bScale = hasBias ? 2 : 1;
242 uint64_t dScale = bidirection ? 2 : 1;242 uint64_t dScale = bidirection ? 2 : 1;
243 uint64_t oneLayerParams = 2 * bScale * dScale;243 uint64_t oneLayerParams = 2 * bScale * dScale;
244- for (uint64_t i = 0; i < (uint64_t)numLayers; i++) {244+ for (uint64_t i = 0; i < static_cast<uint64_t>(numLayers); i++) {
245 for (uint64_t j = 0; j < dScale; j++) {245 for (uint64_t j = 0; j < dScale; j++) {
246 uint64_t offsets = i * oneLayerParams + j * oneLayerParams / 2;246 uint64_t offsets = i * oneLayerParams + j * oneLayerParams / 2;
247 OP_CHECK_WRONG_DIMENSION((*params)[offsets], WEIGHT_DIMS, return false);247 OP_CHECK_WRONG_DIMENSION((*params)[offsets], WEIGHT_DIMS, return false);
@@ -297,7 +297,7 @@ static bool CheckShape(const aclTensor* input, const aclTensorList* params, cons
297 uint64_t oneLayerParams = 2 * bScale * dScale;297 uint64_t oneLayerParams = 2 * bScale * dScale;
298 298 
299 // 逐层校验299 // 逐层校验
300- for (uint64_t i = 0; i < (uint64_t)numLayers; i++) {300+ for (uint64_t i = 0; i < static_cast<uint64_t>(numLayers); i++) {
301 op::Shape expWiShape = {GRU_GATE_NUM * hiddenSize, curLayerInputSize};301 op::Shape expWiShape = {GRU_GATE_NUM * hiddenSize, curLayerInputSize};
302 op::Shape expWhShape = {GRU_GATE_NUM * hiddenSize, hiddenSize};302 op::Shape expWhShape = {GRU_GATE_NUM * hiddenSize, hiddenSize};
303 op::Shape expBShape = {GRU_GATE_NUM * hiddenSize};303 op::Shape expBShape = {GRU_GATE_NUM * hiddenSize};
@@ -632,8 +632,7 @@ static aclnnStatus CheckDataDtypes(const GruDataParamsIn& inputs, const GruDataP
632static aclnnStatus CheckDataParamsValid(const GruDataParamsIn& inputs, const GruDataParamsOut& outputs,632static aclnnStatus CheckDataParamsValid(const GruDataParamsIn& inputs, const GruDataParamsOut& outputs,
633 GruDataInfo& info)633 GruDataInfo& info)
634{634{
635- aclnnStatus ret;635+ aclnnStatus ret = CheckDataParamsNullptr(inputs, outputs);
636- ret = CheckDataParamsNullptr(inputs, outputs);
637 OP_CHECK(ret == ACLNN_SUCCESS,636 OP_CHECK(ret == ACLNN_SUCCESS,
638 OP_LOGE(ret, "CheckDataParamsNullptr failed, certain incoming nullptrs may be invalid."), return ret);637 OP_LOGE(ret, "CheckDataParamsNullptr failed, certain incoming nullptrs may be invalid."), return ret);
639 638 
Mrnn/gru/op_host/gru_tiling.cpp+0-1
@@ -64,7 +64,6 @@ ge::graphStatus GruTilingOp::GetOpInfo()
64 64 
65 auto outputHShape = context_->GetOutputShape(1)->GetStorageShape(); // [B,H]65 auto outputHShape = context_->GetOutputShape(1)->GetStorageShape(); // [B,H]
66 gruParams_.batch = outputHShape.GetDim(0);66 gruParams_.batch = outputHShape.GetDim(0);
67- 
68 } else {67 } else {
69 OP_LOGE(context_->GetNodeName(), "input_x only support 2D or 3D.");68 OP_LOGE(context_->GetNodeName(), "input_x only support 2D or 3D.");
70 return ge::GRAPH_FAILED;69 return ge::GRAPH_FAILED;
Mrnn/gru_grad/op_api/aclnn_gru_backward.cpp+14-12
@@ -46,6 +46,7 @@ constexpr int64_t DIM_ONE = 1;
46constexpr int64_t DIM_TWO = 2;46constexpr int64_t DIM_TWO = 2;
47constexpr int64_t DIM_THREE = 3;47constexpr int64_t DIM_THREE = 3;
48constexpr int64_t GATE_COUNT = 3; // GRU: reset, update, new48constexpr int64_t GATE_COUNT = 3; // GRU: reset, update, new
49+constexpr int64_t GATE_LIST_COUNT = 5;
49 50 
50// 参数列表中每层的参数个数51// 参数列表中每层的参数个数
51constexpr int64_t PARAMS_PER_LAYER_NO_BIAS_NO_BIDI = 2; // w_ih, w_hh52constexpr int64_t PARAMS_PER_LAYER_NO_BIAS_NO_BIDI = 2; // w_ih, w_hh
@@ -112,12 +113,12 @@ static bool CheckTensorListDtype(const aclTensorList* list, const char* name, ge
112 OP_CHECK_DTYPE_NOT_SUPPORT(tensor, DTYPE_SUPPORT_LIST, return false);113 OP_CHECK_DTYPE_NOT_SUPPORT(tensor, DTYPE_SUPPORT_LIST, return false);
113 if (tensor->GetDataType() != baseDtype) {114 if (tensor->GetDataType() != baseDtype) {
114 OP_LOGE(ACLNN_ERR_PARAM_INVALID, "%s[%llu] dtype inconsistent, expected %s, actual %s.", name,115 OP_LOGE(ACLNN_ERR_PARAM_INVALID, "%s[%llu] dtype inconsistent, expected %s, actual %s.", name,
115- (unsigned long long)i, op::ToString(baseDtype).GetString(),116+ static_cast<unsigned long long>(i), op::ToString(baseDtype).GetString(),
116 op::ToString(tensor->GetDataType()).GetString());117 op::ToString(tensor->GetDataType()).GetString());
117 return false;118 return false;
118 }119 }
119 if (!IsFormatSupported(tensor, name)) {120 if (!IsFormatSupported(tensor, name)) {
120- OP_LOGE(ACLNN_ERR_PARAM_INVALID, "%s[%llu] format invalid", name, (unsigned long long)i);121+ OP_LOGE(ACLNN_ERR_PARAM_INVALID, "%s[%llu] format invalid", name, static_cast<unsigned long long>(i));
121 return false;122 return false;
122 }123 }
123 }124 }
@@ -131,7 +132,6 @@ static bool CheckDtypeValid(const aclTensor* input, const aclTensor* hx, const a
131 aclTensorList* dparamsOut)132 aclTensorList* dparamsOut)
132{133{
133 ge::DataType baseDtype = input->GetDataType();134 ge::DataType baseDtype = input->GetDataType();
134- 
135 if (!CheckDtypeConsistent(input, "input", baseDtype))135 if (!CheckDtypeConsistent(input, "input", baseDtype))
136 return false;136 return false;
137 if (!CheckDtypeConsistent(hx, "hx", baseDtype))137 if (!CheckDtypeConsistent(hx, "hx", baseDtype))
@@ -187,9 +187,9 @@ static bool ValidateShape(const aclTensor* tensor, const std::vector<int64_t>& e
187 187 
188static bool CheckShapeValid(const aclTensor* input, const aclTensor* hx, const aclTensorList* params,188static bool CheckShapeValid(const aclTensor* input, const aclTensor* hx, const aclTensorList* params,
189 const aclTensor* dy, const aclTensor* dh, const aclTensorList* r, const aclTensorList* z,189 const aclTensor* dy, const aclTensor* dh, const aclTensorList* r, const aclTensorList* z,
190- const aclTensorList* n, const aclTensorList* hn, const aclTensorList* h,190+ const aclTensorList* n, const aclTensorList* hn, const aclTensorList* h, bool hasBias,
191- const aclTensor* batchSizes, bool hasBias, int64_t numLayers, bool bidirection,191+ int64_t numLayers, bool bidirection, bool batchFirst, const aclTensor* dxOut,
192- bool batchFirst, aclTensor* dxOut, aclTensor* dhPrevOut, aclTensorList* dparamsOut)192+ const aclTensor* dhPrevOut, const aclTensorList* dparamsOut)
193{193{
194 // 从 input 推断 T, B, I194 // 从 input 推断 T, B, I
195 auto xShape = input->GetViewShape();195 auto xShape = input->GetViewShape();
@@ -235,10 +235,11 @@ static bool CheckShapeValid(const aclTensor* input, const aclTensor* hx, const a
235 uint64_t expectedGateLen = static_cast<uint64_t>(D * numLayers);235 uint64_t expectedGateLen = static_cast<uint64_t>(D * numLayers);
236 const char* gateNames[] = {"r", "z", "n", "hn", "h"};236 const char* gateNames[] = {"r", "z", "n", "hn", "h"};
237 const aclTensorList* gateLists[] = {r, z, n, hn, h};237 const aclTensorList* gateLists[] = {r, z, n, hn, h};
238- for (int g = 0; g < 5; ++g) {238+ for (int g = 0; g < GATE_LIST_COUNT; ++g) {
239 if (gateLists[g]->Size() != expectedGateLen) {239 if (gateLists[g]->Size() != expectedGateLen) {
240 OP_LOGE(ACLNN_ERR_PARAM_INVALID, "%s list size mismatch, expected %llu, actual %llu.", gateNames[g],240 OP_LOGE(ACLNN_ERR_PARAM_INVALID, "%s list size mismatch, expected %llu, actual %llu.", gateNames[g],
241- (unsigned long long)expectedGateLen, (unsigned long long)gateLists[g]->Size());241+ static_cast<unsigned long long>(expectedGateLen),
242+ static_cast<unsigned long long>(gateLists[g]->Size()));
242 return false;243 return false;
243 }244 }
244 std::vector<int64_t> gateExpected = {T, B, H};245 std::vector<int64_t> gateExpected = {T, B, H};
@@ -253,7 +254,7 @@ static bool CheckShapeValid(const aclTensor* input, const aclTensor* hx, const a
253 uint64_t expectedParamsLen = static_cast<uint64_t>(paramsPerLayer * numLayers);254 uint64_t expectedParamsLen = static_cast<uint64_t>(paramsPerLayer * numLayers);
254 if (params->Size() != expectedParamsLen) {255 if (params->Size() != expectedParamsLen) {
255 OP_LOGE(ACLNN_ERR_PARAM_INVALID, "params list size mismatch, expected %llu, actual %llu.",256 OP_LOGE(ACLNN_ERR_PARAM_INVALID, "params list size mismatch, expected %llu, actual %llu.",
256- (unsigned long long)expectedParamsLen, (unsigned long long)params->Size());257+ static_cast<unsigned long long>(expectedParamsLen), static_cast<unsigned long long>(params->Size()));
257 return false;258 return false;
258 }259 }
259 260 
@@ -270,7 +271,8 @@ static bool CheckShapeValid(const aclTensor* input, const aclTensor* hx, const a
270 271 
271 if (dparamsOut->Size() != expectedParamsLen) {272 if (dparamsOut->Size() != expectedParamsLen) {
272 OP_LOGE(ACLNN_ERR_PARAM_INVALID, "dparamsOut list size mismatch, expected %llu, actual %llu.",273 OP_LOGE(ACLNN_ERR_PARAM_INVALID, "dparamsOut list size mismatch, expected %llu, actual %llu.",
273- (unsigned long long)expectedParamsLen, (unsigned long long)dparamsOut->Size());274+ static_cast<unsigned long long>(expectedParamsLen),
275+ static_cast<unsigned long long>(dparamsOut->Size()));
274 return false;276 return false;
275 }277 }
276 278 
@@ -289,8 +291,8 @@ static aclnnStatus CheckParams(const aclTensor* input, const aclTensor* hx, cons
289 CHECK_RET(CheckDtypeValid(input, hx, params, dy, dh, r, z, n, hn, h, batchSizes, dxOut, dhPrevOut, dparamsOut),291 CHECK_RET(CheckDtypeValid(input, hx, params, dy, dh, r, z, n, hn, h, batchSizes, dxOut, dhPrevOut, dparamsOut),
290 ACLNN_ERR_PARAM_INVALID);292 ACLNN_ERR_PARAM_INVALID);
291 293 
292- CHECK_RET(CheckShapeValid(input, hx, params, dy, dh, r, z, n, hn, h, batchSizes, hasBias, numLayers, bidirection,294+ CHECK_RET(CheckShapeValid(input, hx, params, dy, dh, r, z, n, hn, h, hasBias, numLayers, bidirection, batchFirst,
293- batchFirst, dxOut, dhPrevOut, dparamsOut),295+ dxOut, dhPrevOut, dparamsOut),
294 ACLNN_ERR_PARAM_INVALID);296 ACLNN_ERR_PARAM_INVALID);
295 297 
296 return ACLNN_SUCCESS;298 return ACLNN_SUCCESS;
Mrnn/gru_grad/op_host/gru_grad_tiling.cpp+12-12
@@ -179,30 +179,30 @@ ge::graphStatus GruGradTiling::GetWorkspaceSize()
179 return ge::GRAPH_SUCCESS;179 return ge::GRAPH_SUCCESS;
180}180}
181 181 
182-bool GruGradTiling::ValidateInputShape(int idx, const std::vector<int64_t>& e)182+bool GruGradTiling::ValidateInputShape(int index, const std::vector<int64_t>& expected_dims)
183{183{
184- auto in = context_->GetInputShape(idx);184+ auto in = context_->GetInputShape(index);
185 if (!in)185 if (!in)
186 return false;186 return false;
187 auto s = in->GetStorageShape();187 auto s = in->GetStorageShape();
188- if (s.GetDimNum() != (int64_t)e.size())188+ if (s.GetDimNum() != static_cast<int64_t>(expected_dims.size()))
189 return false;189 return false;
190- for (size_t i = 0; i < e.size(); i++)190+ for (size_t i = 0; i < expected_dims.size(); i++)
191- if (e[i] != s.GetDim(i))191+ if (expected_dims[i] != s.GetDim(i))
192 return false;192 return false;
193 return true;193 return true;
194}194}
195 195 
196-bool GruGradTiling::ValidateOutputShape(int idx, const std::vector<int64_t>& e)196+bool GruGradTiling::ValidateOutputShape(int index, const std::vector<int64_t>& expected_dims)
197{197{
198- auto out = context_->GetOutputShape(idx);198+ auto out = context_->GetOutputShape(index);
199 if (!out)199 if (!out)
200 return false;200 return false;
201 auto s = out->GetStorageShape();201 auto s = out->GetStorageShape();
202- if (s.GetDimNum() != (int64_t)e.size())202+ if (s.GetDimNum() != static_cast<int64_t>(expected_dims.size()))
203 return false;203 return false;
204- for (size_t i = 0; i < e.size(); i++)204+ for (size_t i = 0; i < expected_dims.size(); i++)
205- if (e[i] != s.GetDim(i))205+ if (expected_dims[i] != s.GetDim(i))
206 return false;206 return false;
207 return true;207 return true;
208}208}
@@ -221,7 +221,7 @@ bool GruGradTiling::CheckParamsShape()
221 tilingData_.inputSize, tilingData_.hiddenSize);221 tilingData_.inputSize, tilingData_.hiddenSize);
222 222 
223 bool ret = true;223 bool ret = true;
224- auto ci = [&](int i, auto& e, const char* n) {224+ auto ci = [this](int i, auto& e, const char* n) {
225 bool ok = ValidateInputShape(i, e);225 bool ok = ValidateInputShape(i, e);
226 auto in = context_->GetInputShape(i);226 auto in = context_->GetInputShape(i);
227 if (in) {227 if (in) {
@@ -239,7 +239,7 @@ bool GruGradTiling::CheckParamsShape()
239 }239 }
240 return ok;240 return ok;
241 };241 };
242- auto co = [&](int i, auto& e, const char* n) {242+ auto co = [this](int i, auto& e, const char* n) {
243 bool ok = ValidateOutputShape(i, e);243 bool ok = ValidateOutputShape(i, e);
244 auto out = context_->GetOutputShape(i);244 auto out = context_->GetOutputShape(i);
245 if (out) {245 if (out) {
Mrnn/gru_grad/op_host/gru_grad_tiling.h+1-2
@@ -58,8 +58,7 @@ private:
58 void ConcatXhBlockCalculate();58 void ConcatXhBlockCalculate();
59 void SetTilingData();59 void SetTilingData();
60 60 
61- CutBatchTiling CalculateCutBatchTiling(int64_t ubParaNum, int64_t alignedSize, int64_t actualSize,61+ CutBatchTiling CalculateCutBatchTiling(int64_t ub, int64_t al, int64_t act, int64_t ml, int64_t b);
62- int64_t copyMLinesMax, int64_t batch);
63 62 
64 bool ValidateInputShape(int index, const std::vector<int64_t>& expected_dims);63 bool ValidateInputShape(int index, const std::vector<int64_t>& expected_dims);
65 bool ValidateOutputShape(int index, const std::vector<int64_t>& expected_dims);64 bool ValidateOutputShape(int index, const std::vector<int64_t>& expected_dims);