apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-11-01 04:12:58 +08:00
Code Issues Actions 6 Packages Projects Releases Wiki Activity

[RKNPU2] RKYOLO Support FP32 return value (#898)

Browse Source 
* RKNPU2 Backend兼容其他模型的量化
fd_tensor正式移除zp和scale的量化参数

* 更新FP32返回值的RKYOLO

* 更新rkyolov5支持fp32格式

* 更新rkyolov5支持fp32格式

* 更新YOLOv5速度文档

Co-authored-by: DefTruth <31974251+DefTruth@users.noreply.github.com>
This commit is contained in:
Zheng_Bicheng
2022-12-19 10:03:18 +08:00
committed by GitHub
parent 1798ad69ed
commit 95beb2bbf6
10 changed files with 76 additions and 118 deletions
Download Patch File Download Diff File Expand all files Collapse all files

48
fastdeploy/backends/rknpu/rknpu2/rknpu2_backend.cc
View File

@@ -12,7 +12,7 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include "fastdeploy/backends/rknpu/rknpu2/rknpu2_backend.h"
#include "fastdeploy/utils/perf.h"
namespace fastdeploy {
RKNPU2Backend::~RKNPU2Backend() {
// Release memory uniformly here
@@ -190,7 +190,6 @@ bool RKNPU2Backend::GetModelInputOutputInfos() {
FDERROR << "rknpu2_backend only support input format is NHWC or UNDEFINED" << std::endl;
}
DumpTensorAttr(input_attrs_[i]);
// copy input_attrs_ to input tensor info
std::string temp_name = input_attrs_[i].name;
@@ -199,16 +198,13 @@ bool RKNPU2Backend::GetModelInputOutputInfos() {
for (int j = 0; j < input_attrs_[i].n_dims; j++) {
temp_shape[j] = (int)input_attrs_[i].dims[j];
}
FDDataType temp_dtype =
fastdeploy::RKNPU2Backend::RknnTensorTypeToFDDataType(
input_attrs_[i].type);
FDDataType temp_dtype = fastdeploy::RKNPU2Backend::RknnTensorTypeToFDDataType(input_attrs_[i].type);
TensorInfo temp_input_info = {temp_name, temp_shape, temp_dtype};
inputs_desc_[i] = temp_input_info;
}
// Get detailed output parameters
output_attrs_ =
(rknn_tensor_attr*)malloc(sizeof(rknn_tensor_attr) * io_num.n_output);
output_attrs_ = (rknn_tensor_attr*)malloc(sizeof(rknn_tensor_attr) * io_num.n_output);
memset(output_attrs_, 0, io_num.n_output * sizeof(rknn_tensor_attr));
outputs_desc_.resize(io_num.n_output);
@@ -225,19 +221,13 @@ bool RKNPU2Backend::GetModelInputOutputInfos() {
return false;
}
// If the output dimension is 3, the runtime will automatically change it to 4.
// If the output dimension is 3, the runtime will automatically change it to 4.
// Obviously, this is wrong, and manual correction is required here.
int n_dims = output_attrs_[i].n_dims;
if((n_dims == 4) && (output_attrs_[i].dims[3] == 1)){
n_dims--;
FDWARNING << "The output["
<< i
<< "].shape[3] is 1, remove this dim."
<< std::endl;
}
DumpTensorAttr(output_attrs_[i]);
// copy output_attrs_ to output tensor
std::string temp_name = output_attrs_[i].name;
std::vector<int> temp_shape{};
@@ -246,9 +236,8 @@ bool RKNPU2Backend::GetModelInputOutputInfos() {
temp_shape[j] = (int)output_attrs_[i].dims[j];
}
FDDataType temp_dtype =
fastdeploy::RKNPU2Backend::RknnTensorTypeToFDDataType(
output_attrs_[i].type);
// The data type of output data is changed to FP32
FDDataType temp_dtype = FDDataType::FP32;
TensorInfo temp_input_info = {temp_name, temp_shape, temp_dtype};
outputs_desc_[i] = temp_input_info;
}
@@ -265,11 +254,12 @@ bool RKNPU2Backend::GetModelInputOutputInfos() {
void RKNPU2Backend::DumpTensorAttr(rknn_tensor_attr& attr) {
printf("index=%d, name=%s, n_dims=%d, dims=[%d, %d, %d, %d], "
"n_elems=%d, size=%d, fmt=%s, type=%s, "
"qnt_type=%s, zp=%d, scale=%f\n",
"qnt_type=%s, zp=%d, scale=%f, pass_through=%d",
attr.index, attr.name, attr.n_dims, attr.dims[0], attr.dims[1],
attr.dims[2], attr.dims[3], attr.n_elems, attr.size,
get_format_string(attr.fmt), get_type_string(attr.type),
get_qnt_type_string(attr.qnt_type), attr.zp, attr.scale);
get_qnt_type_string(attr.qnt_type), attr.zp, attr.scale,
attr.pass_through);
}
TensorInfo RKNPU2Backend::GetInputInfo(int index) {
@@ -320,7 +310,12 @@ bool RKNPU2Backend::Infer(std::vector<FDTensor>& inputs,
input_attrs_[i].type = input_type;
input_attrs_[i].size = inputs[0].Nbytes();
input_attrs_[i].size_with_stride = inputs[0].Nbytes();
input_attrs_[i].pass_through = 0;
if(input_attrs_[i].type == RKNN_TENSOR_FLOAT16 ||
input_attrs_[i].type == RKNN_TENSOR_FLOAT32){
FDINFO << "The input model is not a quantitative model. "
"Close the normalize operation." << std::endl;
}
input_mems_[i] = rknn_create_mem(ctx, inputs[i].Nbytes());
if (input_mems_[i] == nullptr) {
FDERROR << "rknn_create_mem input_mems_ error." << std::endl;
@@ -345,11 +340,13 @@ bool RKNPU2Backend::Infer(std::vector<FDTensor>& inputs,
FDERROR << "rknn_create_mem output_mems_ error." << std::endl;
return false;
}
if(output_attrs_[i].type == RKNN_TENSOR_FLOAT16){
output_attrs_[i].type = RKNN_TENSOR_FLOAT32;
}
// The data type of output data is changed to FP32
output_attrs_[i].type = RKNN_TENSOR_FLOAT32;
// default output type is depend on model, this requires float32 to compute top5
ret = rknn_set_io_mem(ctx, output_mems_[i], &output_attrs_[i]);
// set output memory and attribute
if (ret != RKNN_SUCC) {
FDERROR << "output tensor memory rknn_set_io_mem fail! ret=" << ret
@@ -360,7 +357,7 @@ bool RKNPU2Backend::Infer(std::vector<FDTensor>& inputs,
this->infer_init = true;
}
// Copy input data to input tensor memory
for (uint32_t i = 0; i < io_num.n_input; i++) {
uint32_t width = input_attrs_[i].dims[2];
@@ -376,7 +373,6 @@ bool RKNPU2Backend::Infer(std::vector<FDTensor>& inputs,
return false;
}
}
// run rknn
ret = rknn_run(ctx, nullptr);
@@ -395,8 +391,6 @@ bool RKNPU2Backend::Infer(std::vector<FDTensor>& inputs,
}
(*outputs)[i].Resize(temp_shape, outputs_desc_[i].dtype,
outputs_desc_[i].name);
std::vector<float> output_scale = {output_attrs_[i].scale};
(*outputs)[i].SetQuantizationInfo(output_attrs_[i].zp, output_scale);
memcpy((*outputs)[i].MutableData(), (float*)output_mems_[i]->virt_addr,
(*outputs)[i].Nbytes());
}