apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-16 13:41:30 +08:00
Code Issues Actions 6 Packages Projects Releases Wiki Activity

[CVCUDA] Update CV-CUDA to v0.2.1, add vision processor C++ tutorial (#1678)

Browse Source 
* update cvcuda 0.2.0 -> 0.2.1

* add cpp tutorials demo

* fix reviewed problem
This commit is contained in:
guxukai
2023-03-24 16:57:35 +08:00
committed by GitHub
parent f0235a4c4e
commit 631f94d276
17 changed files with 170 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
fastdeploy/vision/common/processors/cast.cc
View File

@@ -96,7 +96,7 @@ bool Cast::ImplByCvCuda(FDMat* mat) {
auto dst_tensor =
CreateCvCudaTensorWrapData(*(mat->output_cache), mat->layout);
cvcuda_convert_op_(mat->Stream(), src_tensor, dst_tensor, 1.0f, 0.0f);
cvcuda_convert_op_(mat->Stream(), *src_tensor, *dst_tensor, 1.0f, 0.0f);
mat->SetTensor(mat->output_cache);
mat->mat_type = ProcLib::CVCUDA;

2
fastdeploy/vision/common/processors/center_crop.cc
View File

@@ -70,7 +70,7 @@ bool CenterCrop::ImplByCvCuda(FDMat* mat) {
int offset_x = static_cast<int>((mat->Width() - width_) / 2);
int offset_y = static_cast<int>((mat->Height() - height_) / 2);
NVCVRectI crop_roi = {offset_x, offset_y, width_, height_};
cvcuda_crop_op_(mat->Stream(), src_tensor, dst_tensor, crop_roi);
cvcuda_crop_op_(mat->Stream(), *src_tensor, *dst_tensor, crop_roi);
mat->SetTensor(mat->output_cache);
mat->SetWidth(width_);

6
fastdeploy/vision/common/processors/cvcuda_utils.cc
View File

@@ -43,8 +43,8 @@ nvcv::ImageFormat CreateCvCudaImageFormat(FDDataType type, int channel,
return nvcv::FMT_BGRf32;
}
nvcv::TensorWrapData CreateCvCudaTensorWrapData(const FDTensor& tensor,
Layout layout) {
std::shared_ptr<nvcv::TensorWrapData> CreateCvCudaTensorWrapData(
const FDTensor& tensor, Layout layout) {
FDASSERT(tensor.shape.size() == 3,
"When create CVCUDA tensor from FD tensor,"
"tensor shape should be 3-Dim,");
@@ -76,7 +76,7 @@ nvcv::TensorWrapData CreateCvCudaTensorWrapData(const FDTensor& tensor,
nvcv::TensorDataStridedCuda tensor_data(
nvcv::TensorShape{req.shape, req.rank, req.layout},
nvcv::DataType{req.dtype}, buf);
return nvcv::TensorWrapData(tensor_data);
return std::make_shared<nvcv::TensorWrapData>(tensor_data, nullptr);
}
void* GetCvCudaTensorDataPtr(const nvcv::TensorWrapData& tensor) {

2
fastdeploy/vision/common/processors/cvcuda_utils.h
View File

@@ -27,7 +27,7 @@ namespace vision {
nvcv::ImageFormat CreateCvCudaImageFormat(FDDataType type, int channel,
bool interleaved = true);
nvcv::TensorWrapData CreateCvCudaTensorWrapData(const FDTensor& tensor,
std::shared_ptr<nvcv::TensorWrapData> CreateCvCudaTensorWrapData(const FDTensor& tensor,
Layout layout = Layout::HWC);
void* GetCvCudaTensorDataPtr(const nvcv::TensorWrapData& tensor);
nvcv::ImageWrapData CreateImageWrapData(const FDTensor& tensor);

2
fastdeploy/vision/common/processors/hwc2chw.cc
View File

@@ -75,7 +75,7 @@ bool HWC2CHW::ImplByCvCuda(FDMat* mat) {
auto dst_tensor =
CreateCvCudaTensorWrapData(*(mat->output_cache), Layout::CHW);
cvcuda_reformat_op_(mat->Stream(), src_tensor, dst_tensor);
cvcuda_reformat_op_(mat->Stream(), *src_tensor, *dst_tensor);
mat->layout = Layout::CHW;
mat->SetTensor(mat->output_cache);

2
fastdeploy/vision/common/processors/pad.cc
View File

@@ -133,7 +133,7 @@ bool Pad::ImplByCvCuda(FDMat* mat) {
"output_cache", Device::GPU);
auto dst_tensor = CreateCvCudaTensorWrapData(*(mat->output_cache));
cvcuda_pad_op_(mat->Stream(), src_tensor, dst_tensor, top_, left_,
cvcuda_pad_op_(mat->Stream(), *src_tensor, *dst_tensor, top_, left_,
NVCV_BORDER_CONSTANT, value);
mat->SetTensor(mat->output_cache);

6
fastdeploy/vision/common/processors/pad_to_size.cc
View File

@@ -206,7 +206,7 @@ static bool PadHWCByCvCuda(cvcuda::CopyMakeBorder& pad_op, FDMat* mat,
"output_cache", Device::GPU);
auto dst_tensor = CreateCvCudaTensorWrapData(*(mat->output_cache));
pad_op(mat->Stream(), src_tensor, dst_tensor, 0, 0, NVCV_BORDER_CONSTANT,
pad_op(mat->Stream(), *src_tensor, *dst_tensor, 0, 0, NVCV_BORDER_CONSTANT,
border_value);
mat->SetTensor(mat->output_cache);
@@ -238,8 +238,8 @@ static bool PadCHWByCvCuda(cvcuda::CopyMakeBorder& pad_op, FDMat* mat,
input->device, input->device_id);
auto dst_tensor = CreateCvCudaTensorWrapData(dst);
pad_op(mat->Stream(), src_tensor, dst_tensor, 0, 0, NVCV_BORDER_CONSTANT,
border_value);
pad_op(mat->Stream(), (*src_tensor), (*dst_tensor), 0, 0,
NVCV_BORDER_CONSTANT, border_value);
}
mat->SetTensor(mat->output_cache);
mat->mat_type = ProcLib::CVCUDA;

2
fastdeploy/vision/common/processors/resize.cc
View File

@@ -146,7 +146,7 @@ bool Resize::ImplByCvCuda(FDMat* mat) {
auto dst_tensor = CreateCvCudaTensorWrapData(*(mat->output_cache));
// CV-CUDA Interp value is compatible with OpenCV
cvcuda_resize_op_(mat->Stream(), src_tensor, dst_tensor,
cvcuda_resize_op_(mat->Stream(), *src_tensor, *dst_tensor,
CreateCvCudaInterp(interp_));
mat->SetTensor(mat->output_cache);

2
fastdeploy/vision/common/processors/resize_by_short.cc
View File

@@ -95,7 +95,7 @@ bool ResizeByShort::ImplByCvCuda(FDMat* mat) {
"output_cache", Device::GPU);
auto dst_tensor = CreateCvCudaTensorWrapData(*(mat->output_cache));
cvcuda_resize_op_(mat->Stream(), src_tensor, dst_tensor,
cvcuda_resize_op_(mat->Stream(), *src_tensor, *dst_tensor,
CreateCvCudaInterp(interp_));
mat->SetTensor(mat->output_cache);

2
fastdeploy/vision/common/processors/stride_pad.cc
View File

@@ -167,7 +167,7 @@ bool StridePad::ImplByCvCuda(FDMat* mat) {
"output_cache", Device::GPU);
auto dst_tensor = CreateCvCudaTensorWrapData(*(mat->output_cache));
cvcuda_pad_op_(mat->Stream(), src_tensor, dst_tensor, 0, 0,
cvcuda_pad_op_(mat->Stream(), *src_tensor, *dst_tensor, 0, 0,
NVCV_BORDER_CONSTANT, value);
mat->SetTensor(mat->output_cache);