你好，我还有一个问题，就是bangc输出的tensor传回pytorch（libtorch）的时候会经过一次自动的NHWC到NCHW的转换，有什么方法能不进行这一步，或者是在libtorch里让NCHW转回NHWC吗？因为我的bangc已经输出正确结果了，然后经过这个自动转换输出就错了。展开
全部回复
-
-
-
-
-
-
-
-
-
-
LV.2 #5 幺白幺木回复#4踏雪寻梅回复尊敬的开发者您好，请提供您使用的板卡型号及SDK版本。另外请问是否方便将您这边验证出错的代码发出，我们分析后给您反馈。
展开Mlu版本信息：cat /proc/driver/cambricon/mlus/0000:17:00.0/information
Device name: MLU270-F4
Device inode path: /dev/cambricon_dev0
Device Major: 234
Device Minor: 0
Driver Version: v4.9.2
MCU Version: v1.1.4
Board Serial Number: SN/182101100027
MLU Firmware Version: 4.9.2
Board CV: 1
IPU Freq: 1000MHz
Interrupt Mode: MSI
Bus Location: 17_0_0
Bus Type: PCIE
LnkCap: Speed 8.0GT/s, Width x16
Region 0: Memory at 80000000 [size=256M]
Region 2: Memory at 78000000 [size=64M]
Region 4: Memory at 7c000000 [size=64M]
验证代码为加粗加大的cout<<input<<endl
#include "aten/operators/cnml/internal/cnml_internal.h"#include "iostream"
using namespace std;
namespace torch_mlu {
namespace cnml {
namespace ops {
at::Tensor cnml_pixelshuffle_internal(const at::Tensor& input, const int64_t input_r) {
auto options =
at::TensorOptions(c10::ScalarType::Float).device(at::DeviceType::MLU);
auto input_size = input.sizes().vec();
int r = input_r;
std::vector<int64_t> output_shape(4);
output_shape[0] = input_size[0];
output_shape[1] = input_size[1] / (r * r);
output_shape[2] = input_size[2] * r;
output_shape[3] = input_size[3] * r;
auto output = at::empty(output_shape, input.options());
// prepare input cnml tensor
cnmlTensor_t input_cnml_tensor[1];
auto* input_impl = getMluTensorImpl(input);
input_cnml_tensor[0] = input_impl->CreateCnmlTensor(CNML_TENSOR, toCnmlDataType(input.dtype()));
// prepare output cnml tensor
cnmlTensor_t output_cnml_tensor[1];
auto* output_impl = getMluTensorImpl(output);
output_cnml_tensor[0] = output_impl->CreateCnmlTensor(CNML_TENSOR, toCnmlDataType(output.dtype()));
// End the execution flow if not MLU device
CHECK_MLU_DEVICE(output);
int total_number = input.numel();
int dtype = 0;
switch (toCnmlDataType(input.dtype())) {
case CNML_DATA_FLOAT32:
dtype = 4;
break;
case CNML_DATA_FLOAT16:
dtype = 2;
break;
default:
CNLOG(ERROR) << "Pixelshuffle op only supports FLOAT32 & FLOAT16";
}
// for(int i = 0; i < total_number; i++)
// std::cout<<*((float*)(input+i))<<std::endl;
cout<<input<<endl
// setup operator
cnmlBaseOp_t pixelshuffle_op;
cnmlPluginPixelshuffleOpParam_t param;
TORCH_CNML_CHECK(cnmlCreatePluginPixelshuffleOpParam(¶m,
total_number,
input_size[0],
input_size[2],
input_size[3],
input_size[1],
input_r,
dtype,
GET_CORE_VERSION));
TORCH_CNML_CHECK(cnmlCreatePluginPixelshuffleOp(&pixelshuffle_op,
param,
input_cnml_tensor,
output_cnml_tensor));
// return to JIT if running mode is fuse
CHECK_RETURN_TO_FUSE(pixelshuffle_op, output);
auto queue = getCurQueue();
// compile op
TORCH_CNML_CHECK(cnmlCompileBaseOp(pixelshuffle_op, GET_CORE_VERSION, GET_CORE_NUMBER));
// compute operator
void* input_ptrs[1];
input_ptrs[0] = input_impl->raw_mutable_data();
cout<<input_ptrs[0]<<endl;
void* output_ptrs[1];
output_ptrs[0] = output_impl->raw_mutable_data();
TORCH_CNML_CHECK(cnmlComputePluginPixelshuffleOpForward(pixelshuffle_op,
input_ptrs,
1,
output_ptrs,
1,
queue));
syncQueue(queue);
// destroy params and ops
if (param != nullptr) {
TORCH_CNML_CHECK(cnmlDestroyPluginPixelshuffleOpParam(¶m));
param = nullptr;
}
if (pixelshuffle_op != nullptr) {
TORCH_CNML_CHECK(cnmlDestroyBaseOp(&pixelshuffle_op));
pixelshuffle_op = nullptr;
}
// std::cout<<output<<std::endl;
// output.print();
return output;
}
} // namespace ops
} // namespace cnml
} // namespace torch_mlu
-
-
-
-
请登录后评论