提交 cb86440a 编写于 作者: Y yangfei

add some function

上级 1da3fcee
......@@ -15,41 +15,41 @@ limitations under the License. */
#include "operators/kernel/feed_kernel.h"
namespace paddle_mobile {
namespace operators {
template <>
bool FeedKernel<FPGA, float>::Init(FeedParam<FPGA> *param) {
Tensor *output = param->Out();
fpga::format_fp16_ofm(output);
return true;
}
template <>
void FeedKernel<FPGA, float>::Compute(const FeedParam<FPGA> &param) {
auto input = reinterpret_cast<Tensor *>(const_cast<LoDTensor *>(param.InputX()));
auto input_ptr = input->data<float>();
fpga::format_image(input);
Tensor *output = param.Out();
auto output_ptr = output->data<float>();
fpga::BypassArgs args = {fpga::DATA_TYPE_FP32};
args.input_data_type = fpga::DATA_TYPE_FP32;
args.output_data_type = fpga::DATA_TYPE_FP16;
args.input_layout_type = fpga::LAYOUT_CHW;
args.output_layout_type = fpga::LAYOUT_HWC;
args.image.address = reinterpret_cast<void *>(input_ptr);
args.image.channels = (uint32_t)input->dims()[1];
args.image.height = (uint32_t)input->dims()[2];
args.image.width = (uint32_t)input->dims()[3];
args.image.pad_height = 0;
args.image.pad_width = 0;
args.output.address = output_ptr;
args.output.scale_address = output->scale;
fpga::PerformBypass(args);
}
template class FeedKernel<FPGA, float>;
} // namespace operators
namespace operators {
template <>
bool FeedKernel<FPGA, float>::Init(FeedParam<FPGA> *param) {
Tensor *output = param->Out();
fpga::format_fp16_ofm(output);
return true;
}
template <>
void FeedKernel<FPGA, float>::Compute(const FeedParam<FPGA> &param) {
auto input =
reinterpret_cast<Tensor *>(const_cast<LoDTensor *>(param.InputX()));
auto input_ptr = input->data<float>();
fpga::format_image(input);
Tensor *output = param.Out();
auto output_ptr = output->data<float>();
fpga::BypassArgs args = {fpga::DATA_TYPE_FP32};
args.input_data_type = fpga::DATA_TYPE_FP32;
args.output_data_type = fpga::DATA_TYPE_FP16;
args.input_layout_type = fpga::LAYOUT_CHW;
args.output_layout_type = fpga::LAYOUT_HWC;
args.image.address = reinterpret_cast<void *>(input_ptr);
args.image.channels = (uint32_t)input->dims()[1];
args.image.height = (uint32_t)input->dims()[2];
args.image.width = (uint32_t)input->dims()[3];
args.image.pad_height = 0;
args.image.pad_width = 0;
args.output.address = output_ptr;
args.output.scale_address = output->scale;
fpga::PerformBypass(args);
}
template class FeedKernel<FPGA, float>;
} // namespace operators
} // namespace paddle_mobile
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册