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

!40122 [MS][LITE]fix scale tensorrt bug 

Merge pull request !40122 from 张学同/applyproximaladagrad_gpu
This commit is contained in:
i-robot 2022-08-09 11:55:59 +00:00 committed by Gitee
parent db092bda4a 3c45bcddb1
commit acf281a026
No known key found for this signature in database
GPG Key ID: 173E9B9CA92EEF8F
14 changed files with 9 additions and 241 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
mindspore/lite/src/extendrt/delegate/tensorrt/op/batchtospace_tensorrt.cc
View File

@ -45,17 +45,6 @@ int BatchToSpaceTensorRT::IsSupport(const BaseOperatorPtr &base_operator, const
int BatchToSpaceTensorRT::AddInnerOp(TensorRTContext *ctx) {
nvinfer1::ITensor *input_tensor = input(ctx, 0).trt_tensor_;
if (input(ctx, 0).trt_tensor_->getDimensions().nbDims == DIMENSION_4D && input(ctx, 0).format_ == Format::NHWC) {
// transpose: NHWC->NCHW
nvinfer1::IShuffleLayer *transpose_layer_in = NHWC2NCHW(ctx, *input(ctx, 0).trt_tensor_);
if (transpose_layer_in == nullptr) {
MS_LOG(ERROR) << "transpose: NHWC->NCHW failed";
return RET_ERROR;
}
transpose_layer_in->setName((op_name_ + "_transpose2NCHW").c_str());
this->transpose_layer_ = transpose_layer_in;
input_tensor = transpose_layer_in->getOutput(0);
}
const int *block_size_ptr = reinterpret_cast<const int *>(in_tensors_[1].Data());
int bh = *(block_size_ptr + 0);
int bw = *(block_size_ptr + 1);

35
mindspore/lite/src/extendrt/delegate/tensorrt/op/concate_tensorrt.cc
View File

@ -66,15 +66,6 @@ int ConcateTensorRT::AddInnerOp(TensorRTContext *ctx) {
MS_LOG(ERROR) << "PreProcessInputs failed for " << op_name_;
return ret;
}
if (!same_format_) {
if (trt_input_tensors[0]->getDimensions().nbDims == DIMENSION_4D && out_format_ == Format::NCHW) {
// when inputs all NCHW, change axis
axis_ = ConvertAxisFromNHWC2NCHW(axis_);
MS_LOG(DEBUG) << "concate axis change to " << axis_ << " when using NCHW format.";
} else {
MS_LOG(WARNING) << "input tensor format needs check, convert concat axis failed for " << op_name_;
}
}
if (type_ == ops::kNameStack) {
for (size_t i = 0; i != in_tensors_.size(); ++i) {
@ -127,29 +118,9 @@ int ConcateTensorRT::PreProcessInputs(TensorRTContext *ctx, nvinfer1::ITensor *t
}
}
// make sure all inputs are same format
if (input_nbDims == DIMENSION_4D) {
for (size_t i = 0; i < in_tensors_.size(); i++) {
if (input(ctx, i).format_ == out_format_) {
trt_input_tensors[i] = input(ctx, i).trt_tensor_;
MS_LOG(DEBUG) << "concate input " << GetTensorFormat(input(ctx, i));
} else {
nvinfer1::IShuffleLayer *transpose_layer = NHWC2NCHW(ctx, *input(ctx, i).trt_tensor_);
if (transpose_layer == nullptr) {
MS_LOG(ERROR) << "op action convert failed";
return RET_ERROR;
}
trt_input_tensors[i] = transpose_layer->getOutput(0);
this->transpose_layer_ = transpose_layer;
same_format_ = true;
MS_LOG(DEBUG) << "concate input " << GetTensorFormat(trt_input_tensors[i], Format::NHWC, true);
}
}
} else {
for (size_t i = 0; i < in_tensors_.size(); i++) {
trt_input_tensors[i] = input(ctx, i).trt_tensor_;
MS_LOG(DEBUG) << "concate input " << GetTensorFormat(input(ctx, i));
}
for (size_t i = 0; i < in_tensors_.size(); i++) {
trt_input_tensors[i] = input(ctx, i).trt_tensor_;
MS_LOG(DEBUG) << "concate input " << GetTensorFormat(input(ctx, i));
}
return RET_OK;
}

11
mindspore/lite/src/extendrt/delegate/tensorrt/op/convolution_tensorrt.cc
View File

@ -54,17 +54,6 @@ int ConvolutionTensorRT::AddInnerOp(TensorRTContext *ctx) {
}
nvinfer1::ITensor *conv_input = input(ctx, 0).trt_tensor_;
if (input(ctx, 0).trt_tensor_->getDimensions().nbDims == DIMENSION_4D && input(ctx, 0).format_ == Format::NHWC) {
// transpose: NHWC->NCHW
nvinfer1::IShuffleLayer *transpose_layer_in = NHWC2NCHW(ctx, *input(ctx, 0).trt_tensor_);
if (transpose_layer_in == nullptr) {
MS_LOG(ERROR) << "transpose: NHWC->NCHW failed";
return RET_ERROR;
}
transpose_layer_in->setName((op_name_ + "_transpose2NCHW").c_str());
this->transpose_layer_ = transpose_layer_in;
conv_input = transpose_layer_in->getOutput(0);
}
// transpose weight
const auto &weight_tensor = in_tensors_[1];

11
mindspore/lite/src/extendrt/delegate/tensorrt/op/deconvolution_tensorrt.cc
View File

@ -52,17 +52,6 @@ int DeconvolutionTensorRT::AddInnerOp(TensorRTContext *ctx) {
return RET_ERROR;
}
nvinfer1::ITensor *deconv_input = input(ctx, 0).trt_tensor_;
if (input(ctx, 0).trt_tensor_->getDimensions().nbDims == DIMENSION_4D && input(ctx, 0).format_ == Format::NHWC) {
// transpose: NHWC->NCHW
nvinfer1::IShuffleLayer *transpose_layer_in = NHWC2NCHW(ctx, *input(ctx, 0).trt_tensor_);
if (transpose_layer_in == nullptr) {
MS_LOG(ERROR) << "transpose: NHWC->NCHW failed";
return RET_ERROR;
}
transpose_layer_in->setName((op_name_ + "_transpose2NCHW").c_str());
this->transpose_layer_ = transpose_layer_in;
deconv_input = transpose_layer_in->getOutput(0);
}
// transpose weight
const auto &weight_tensor = in_tensors_[1];

19
mindspore/lite/src/extendrt/delegate/tensorrt/op/elementwise_tensorrt.cc
View File

@ -179,26 +179,7 @@ int ElementWiseTensorRT::PreprocessInputTensors(TensorRTContext *ctx, ITensorHel
}
*x_input = input(ctx, 0);
*y_input = input(ctx, 1);
int const_tensor_index = (in_tensors_[0].Data() != nullptr && in_tensors_[0].IsConst()) ? 0 : 1;
auto input_helper = const_tensor_index == 0 ? y_input : x_input;
auto const_helper = const_tensor_index == 0 ? x_input : y_input;
MS_LOG(DEBUG) << "before transpose " << GetTensorFormat(*x_input);
MS_LOG(DEBUG) << "before transpose " << GetTensorFormat(*y_input);
if (input_helper->trt_tensor_->getDimensions().nbDims == DIMENSION_4D &&
input_helper->format_ != const_helper->format_) {
// when inputs format are different, change to NHWC
auto need_trans = input_helper->format_ == Format::NCHW ? input_helper : const_helper;
nvinfer1::IShuffleLayer *transpose_layer = NCHW2NHWC(ctx, *need_trans->trt_tensor_);
if (transpose_layer == nullptr) {
MS_LOG(ERROR) << "op action convert failed";
return RET_ERROR;
}
transpose_layer->setName((op_name_ + "_input_transpose2NHWC").c_str());
need_trans->trt_tensor_ = transpose_layer->getOutput(0);
need_trans->format_ = Format::NHWC;
need_trans->same_format_ = true;
}
MS_LOG(DEBUG) << "after transpose " << GetTensorFormat(*x_input);
MS_LOG(DEBUG) << "after transpose " << GetTensorFormat(*y_input);
if (GetDimsVolume(x_input->trt_tensor_->getDimensions()) == GetDimsVolume(y_input->trt_tensor_->getDimensions()) &&

12
mindspore/lite/src/extendrt/delegate/tensorrt/op/pad_tensorrt.cc
View File

@ -69,18 +69,6 @@ int PadTensorRT::AddInnerOp(TensorRTContext *ctx) {
nvinfer1::ITensor *pad_input = input(ctx, 0).trt_tensor_;
MS_LOG(DEBUG) << "before transpose " << GetTensorFormat(pad_input, input(ctx, 0).format_, input(ctx, 0).same_format_);
if (input(ctx, 0).trt_tensor_->getDimensions().nbDims == DIMENSION_4D && input(ctx, 0).format_ == Format::NHWC) {
// transpose: NHWC->NCHW
nvinfer1::IShuffleLayer *transpose_layer_in = NHWC2NCHW(ctx, *input(ctx, 0).trt_tensor_);
if (transpose_layer_in == nullptr) {
MS_LOG(ERROR) << "transpose: NHWC->NCHW failed";
return RET_ERROR;
}
transpose_layer_in->setName((op_name_ + "_transpose2NCHW").c_str());
this->transpose_layer_ = transpose_layer_in;
pad_input = transpose_layer_in->getOutput(0);
MS_LOG(DEBUG) << "after transpose " << GetTensorFormat(pad_input, Format::NCHW, false);
}
// trt 6 only support 2D padding
const int *padding_data = reinterpret_cast<const int *>(in_tensors_[1].Data());

11
mindspore/lite/src/extendrt/delegate/tensorrt/op/pool_tensorrt.cc
View File

@ -51,17 +51,6 @@ int PoolTensorRT::AddInnerOp(TensorRTContext *ctx) {
}
nvinfer1::ITensor *pool_input = input(ctx, 0).trt_tensor_;
if (input(ctx, 0).trt_tensor_->getDimensions().nbDims == DIMENSION_4D && input(ctx, 0).format_ == Format::NHWC) {
// transpose: NHWC->NCHW
nvinfer1::IShuffleLayer *transpose_layer_in = NHWC2NCHW(ctx, *input(ctx, 0).trt_tensor_);
if (transpose_layer_in == nullptr) {
MS_LOG(ERROR) << "transpose: NHWC->NCHW failed";
return RET_ERROR;
}
transpose_layer_in->setName((op_name_ + "_transpose2NCHW").c_str());
this->transpose_layer_ = transpose_layer_in;
pool_input = transpose_layer_in->getOutput(0);
}
// global version pooling
if (kernel_size_.empty()) {

31
mindspore/lite/src/extendrt/delegate/tensorrt/op/reduce_tensorrt.cc
View File

@ -48,36 +48,7 @@ int ReduceTensorRT::AddInnerOp(TensorRTContext *ctx) {
out_format_ = input(ctx, 0).format_;
nvinfer1::ITensor *reduce_input = input(ctx, 0).trt_tensor_;
MS_LOG(DEBUG) << "origin input " << GetTensorFormat(input(ctx, 0));
if (input(ctx, 0).trt_tensor_->getDimensions().nbDims == DIMENSION_4D &&
!SameDims(input(ctx, 0).trt_tensor_->getDimensions(), in_tensors_[0].Shape())) {
/*
if (input(ctx, 0).format_ == Format::NCHW) {
// NCHW->NHWC
nvinfer1::IShuffleLayer *transpose_layer = NCHW2NHWC(ctx, *input(ctx, 0).trt_tensor_);
if (transpose_layer == nullptr) {
MS_LOG(ERROR) << "create transpose layer failed for " << op_name_;
return RET_ERROR;
}
transpose_layer->setName((op_name_ + "_transpose_in").c_str());
reduce_input = transpose_layer->getOutput(0);
out_format_ = Format::NHWC;
this->transpose_layer_ = transpose_layer;
} else if (input(ctx, 0).format_ == Format::NHWC) {
// NHWC->NCHW
nvinfer1::IShuffleLayer *transpose_layer = NHWC2NCHW(ctx, *input(ctx, 0).trt_tensor_);
if (transpose_layer == nullptr) {
MS_LOG(ERROR) << "create transpose layer failed for " << op_name_;
return RET_ERROR;
}
transpose_layer->setName((op_name_ + "_transpose_in").c_str());
reduce_input = transpose_layer->getOutput(0);
out_format_ = Format::NCHW;
this->transpose_layer_ = transpose_layer;
} else {
MS_LOG(WARNING) << "input tensor format needs check: " << op_name_;
}
*/
}
MS_LOG(DEBUG) << "after transpose input " << GetTensorFormat(reduce_input, out_format_, true);
if (reduce_op->get_mode() == ReduceMode::Reduce_L2) {
// x^2

11
mindspore/lite/src/extendrt/delegate/tensorrt/op/resize_tensorrt.cc
View File

@ -53,17 +53,6 @@ int ResizeTensorRT::AddInnerOp(TensorRTContext *ctx) {
nvinfer1::ITensor *resize_in_tensor = input(ctx, 0).trt_tensor_;
MS_LOG(DEBUG) << "origin input " << GetTensorFormat(input(ctx, 0));
if (resize_in_tensor->getDimensions().nbDims == DIMENSION_4D && input(ctx, 0).format_ == Format::NHWC) {
// NHWC->NCHW
nvinfer1::IShuffleLayer *transpose_layer = NHWC2NCHW(ctx, *input(ctx, 0).trt_tensor_);
if (transpose_layer == nullptr) {
MS_LOG(ERROR) << "create transpose layer failed for " << op_name_;
return RET_ERROR;
}
transpose_layer->setName((op_name_ + "_transpose_in").c_str());
resize_in_tensor = transpose_layer->getOutput(0);
this->transpose_layer_ = transpose_layer;
}
MS_LOG(DEBUG) << "after transpose input " << GetTensorFormat(resize_in_tensor, Format::NCHW, true);
auto method = resize_op_->get_method();

33
mindspore/lite/src/extendrt/delegate/tensorrt/op/scale_tensorrt.cc
View File

@ -69,7 +69,7 @@ int ScaleTensorRT::AddInnerOp(TensorRTContext *ctx) {
MS_LOG(DEBUG) << "after transpose " << GetTensorFormat(scale_in_tensor, out_format_, out_same_format_);
nvinfer1::ITensor *op_out_tensor{nullptr};
if (scale_in_tensor->getDimensions().nbDims == DIMENSION_4D) {
if (scale_in_tensor->getDimensions().nbDims == DIMENSION_4D && mode_ != nvinfer1::ScaleMode::kCHANNEL) {
op_out_tensor = RunAs4DimsScale(ctx, scale_in_tensor);
} else {
op_out_tensor = RunAsMutiDimsScale(ctx, scale_in_tensor);
@ -92,37 +92,6 @@ int ScaleTensorRT::AddInnerOp(TensorRTContext *ctx) {
nvinfer1::ITensor *ScaleTensorRT::PreProcessInputTensor(TensorRTContext *ctx) {
nvinfer1::ITensor *scale_in_tensor = input(ctx, 0).trt_tensor_;
if (input(ctx, 0).trt_tensor_->getDimensions().nbDims == DIMENSION_4D && mode_ == nvinfer1::ScaleMode::kCHANNEL) {
// per channel input format should be nchw, otherwise should be same with scale nhwc
// transpose: NHWC->NCHW
if ((input(ctx, 0).format_ == Format::NHWC && axis_ == kNHWC_C) ||
(input(ctx, 0).same_format_ == true && axis_ == kNHWC_C)) {
nvinfer1::IShuffleLayer *transpose_layer_in = NHWC2NCHW(ctx, *input(ctx, 0).trt_tensor_);
if (transpose_layer_in == nullptr) {
MS_LOG(ERROR) << "op action convert failed";
return nullptr;
}
transpose_layer_in->setName((op_name_ + "_transpose2NCHW").c_str());
scale_in_tensor = transpose_layer_in->getOutput(0);
out_format_ = Format::NCHW;
out_same_format_ = !out_same_format_;
} else if (out_format_ != Format::NCHW && axis_ != kNCHW_C) {
MS_LOG(WARNING) << op_name_ << " out format (NHWC:1, NCHW:0) infer as " << out_format_ << ", and axis is "
<< axis_;
}
} else if (input(ctx, 0).trt_tensor_->getDimensions().nbDims == DIMENSION_4D &&
input(ctx, 0).format_ == Format::NCHW && mode_ == nvinfer1::ScaleMode::kELEMENTWISE) {
// transpose: NCHW->NHWC
nvinfer1::IShuffleLayer *transpose_layer_in = NCHW2NHWC(ctx, *input(ctx, 0).trt_tensor_);
if (transpose_layer_in == nullptr) {
MS_LOG(ERROR) << "op action convert failed";
return nullptr;
}
transpose_layer_in->setName((op_name_ + "_transpose2NHWC").c_str());
scale_in_tensor = transpose_layer_in->getOutput(0);
out_format_ = Format::NHWC;
out_same_format_ = true;
}
return scale_in_tensor;
}

12
mindspore/lite/src/extendrt/delegate/tensorrt/op/shape_tensorrt.cc
View File

@ -42,17 +42,7 @@ int ShapeTensorRT::AddInnerOp(TensorRTContext *ctx) {
return RET_ERROR;
}
nvinfer1::ITensor *shape_input = input(ctx, 0).trt_tensor_;
if (input(ctx, 0).trt_tensor_->getDimensions().nbDims == DIMENSION_4D && input(ctx, 0).format_ == Format::NCHW) {
// transpose: NCHW->NHWC
nvinfer1::IShuffleLayer *transpose_layer_in = NCHW2NHWC(ctx, *input(ctx, 0).trt_tensor_);
if (transpose_layer_in == nullptr) {
MS_LOG(ERROR) << "transpose: NCHW->NHWC failed for " << op_name_;
return RET_ERROR;
}
transpose_layer_in->setName((op_name_ + "_transpose2NHWC").c_str());
shape_input = transpose_layer_in->getOutput(0);
this->transpose_layer_ = transpose_layer_in;
}
nvinfer1::IShapeLayer *shape_layer = ctx->network()->addShape(*shape_input);
if (shape_layer == nullptr) {

37
mindspore/lite/src/extendrt/delegate/tensorrt/op/shuffle_tensorrt.cc
View File

@ -133,32 +133,7 @@ int ShuffleTensorRT::InputTensorPreprocess(TensorRTContext *ctx) {
shuffler_input_ = input(ctx, 0).trt_tensor_;
MS_LOG(DEBUG) << "before transpose " << GetTensorFormat(input(ctx, 0));
out_format_ = input(ctx, 0).format_;
if (shuffler_input_->getDimensions().nbDims == DIMENSION_4D && !input(ctx, 0).same_format_) {
// input tensor support NCHW format input
/*
if (input(ctx, 0).format_ == Format::NCHW) {
// for transpose op, if tensor has same dim with ms tensor, keep origin dims
nvinfer1::IShuffleLayer *transpose_layer = NCHW2NHWC(ctx_, *shuffler_input_);
if (transpose_layer == nullptr) {
MS_LOG(ERROR) << "create transpose layer failed for " << op_name_;
return RET_ERROR;
}
transpose_layer->setName((op_name_ + "_transpose_in").c_str());
shuffler_input_ = transpose_layer->getOutput(0);
out_format_ = Format::NHWC;
} else if (input(ctx, 0).format_ == Format::NHWC) {
// infer format may error, correct here
nvinfer1::IShuffleLayer *transpose_layer = NHWC2NCHW(ctx_, *shuffler_input_);
if (transpose_layer == nullptr) {
MS_LOG(ERROR) << "create transpose layer failed for " << op_name_;
return RET_ERROR;
}
transpose_layer->setName((op_name_ + "_transpose_in").c_str());
shuffler_input_ = transpose_layer->getOutput(0);
out_format_ = Format::NCHW;
}
*/
}
MS_LOG(DEBUG) << "after transpose " << GetTensorFormat(shuffler_input_, out_format_, true);
return RET_OK;
}
@ -244,15 +219,7 @@ int ShuffleTensorRT::AddTransposeOp(nvinfer1::IShuffleLayer *shuffle_layer) {
perm_data++;
}
shuffle_layer->setFirstTranspose(perm);
if (perm_ternsor.ElementNum() == DIMENSION_4D) {
if (perm.order[kNCHW_C] == kNHWC_C && perm.order[kNCHW_H] == kNHWC_H && perm.order[kNCHW_W] == kNHWC_W) {
out_format_ = Format::NCHW;
} else if (perm.order[kNHWC_H] == kNCHW_H && perm.order[kNHWC_W] == kNCHW_W && perm.order[kNHWC_C] == kNCHW_C) {
out_format_ = Format::NHWC;
} else {
MS_LOG(INFO) << "input format and perm order is not NHWC or NCHW: " << op_name_;
}
}
shuffler_output_ = shuffle_layer->getOutput(0);
return RET_OK;
}

5
mindspore/lite/src/extendrt/delegate/tensorrt/op/softmax_tensorrt.cc
View File

@ -80,10 +80,7 @@ nvinfer1::ISoftMaxLayer *SoftMaxTensorRT::AddSoftMaxOp(TensorRTContext *ctx) {
return nullptr;
}
int64_t axis_format_value = (axis_val[0] == -1) ? input(ctx, 0).trt_tensor_->getDimensions().nbDims - 1 : axis_val[0];
if (input(ctx, 0).trt_tensor_->getDimensions().nbDims == DIMENSION_4D && input(ctx, 0).format_ == Format::NHWC) {
// transpose axis to NCHW
axis_format_value = ConvertAxisFromNHWC2NCHW(axis_format_value);
}
uint32_t axis_bit = 1 << axis_format_value;
MS_LOG(DEBUG) << op_name_ << " axis_value is " << axis_format_value << ", set axis to " << axis_bit;
current_layer_->setAxes(axis_bit);

11
mindspore/lite/src/extendrt/delegate/tensorrt/op/spacetobatch_tensorrt.cc
View File

@ -45,17 +45,6 @@ int SpaceToBatchTensorRT::IsSupport(const BaseOperatorPtr &base_operator, const
int SpaceToBatchTensorRT::AddInnerOp(TensorRTContext *ctx) {
nvinfer1::ITensor *input_tensor = input(ctx, 0).trt_tensor_;
if (input(ctx, 0).trt_tensor_->getDimensions().nbDims == DIMENSION_4D && input(ctx, 0).format_ == Format::NHWC) {
// transpose: NHWC->NCHW
nvinfer1::IShuffleLayer *transpose_layer_in = NHWC2NCHW(ctx, *input(ctx, 0).trt_tensor_);
if (transpose_layer_in == nullptr) {
MS_LOG(ERROR) << "transpose: NHWC->NCHW failed";
return RET_ERROR;
}
transpose_layer_in->setName((op_name_ + "_transpose2NCHW").c_str());
this->transpose_layer_ = transpose_layer_in;
input_tensor = transpose_layer_in->getOutput(0);
}
const int *block_size_ptr = reinterpret_cast<const int *>(in_tensors_[1].Data());
int bh = *(block_size_ptr + 0);
int bw = *(block_size_ptr + 1);