AT_DISPATCH_FLOATING_TYPES_AND2 fails with ScalarType::Byte

[r-zenine]

🐛 Bug

Using AT_DISPATCH_FLOATING_TYPES_AND2, in cuda code is causing a compilation error:

I am trying to enable support for uint8 for nearest Neighbor upsampling.
Therefore, I tried to replace AT_DISPATCH_FLOATING_TYPES_AND_HALF with AT_DISPATCH_FLOATING_TYPES_AND2(at::ScalarType::Half, at::ScalarType::Byte, ...
Which caused a compilation. On the other hand, using AT_DISPATCH_ALL_TYPES_AND(at::ScalarType::Half ... works fine.

To Reproduce

Steps to reproduce the behavior:

1. Checkout to bdd7dbfd4b
2. Change /home/ryad/workspace/pytorch/aten/src/ATen/native/cuda/UpSampleNearest3d.cu(176) to use AT_DISPATCH_FLOATING_TYPES_AND2
3. compile

[1/3] Building NVCC (Device) object caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/torch_cuda_generated_UpSampleNearest3d.cu.o
FAILED: caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/torch_cuda_generated_UpSampleNearest3d.cu.o 
cd /home/ryad/workspace/pytorch/build/caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda && /home/ryad/miniconda3/envs/pytorch-dev/bin/cmake -E make_directory /home/ryad/workspace/pytorch/build/caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/. && /home/ryad/miniconda3/envs/pytorch-dev/bin/cmake -D verbose:BOOL=OFF -D build_configuration:STRING=Release -D generated_file:STRING=/home/ryad/workspace/pytorch/build/caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/./torch_cuda_generated_UpSampleNearest3d.cu.o -D generated_cubin_file:STRING=/home/ryad/workspace/pytorch/build/caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/./torch_cuda_generated_UpSampleNearest3d.cu.o.cubin.txt -P /home/ryad/workspace/pytorch/build/caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/torch_cuda_generated_UpSampleNearest3d.cu.o.Release.cmake
/home/ryad/workspace/pytorch/aten/src/ATen/native/cuda/UpSampleNearest3d.cu(176): error: incomplete type is not allowed

/home/ryad/workspace/pytorch/aten/src/ATen/AccumulateType.h(48): error: class "at::AccumulateType<<error-type>, true>" has no member "type"
          detected during instantiation of type "at::acc_type<<error-type>, true>" 
/home/ryad/workspace/pytorch/aten/src/ATen/native/cuda/UpSampleNearest3d.cu(176): here

/home/ryad/workspace/pytorch/aten/src/ATen/native/cuda/UpSampleNearest3d.cu(269): error: incomplete type is not allowed

3 errors detected in the compilation of "/tmp/tmpxft_00007f4c_00000000-6_UpSampleNearest3d.cpp1.ii".
CMake Error at torch_cuda_generated_UpSampleNearest3d.cu.o.Release.cmake:281 (message):
  Error generating file
  /home/ryad/workspace/pytorch/build/caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/./torch_cuda_generated_UpSampleNearest3d.cu.o


[2/3] Building NVCC (Device) object caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/torch_cuda_generated_UpSampleNearest2d.cu.o
/home/ryad/workspace/pytorch/aten/src/ATen/native/cuda/UpSampleNearest2d.cu: In function ‘at::Tensor at::native::upsample_nearest2d_cuda(const at::Tensor&, c10::IntArrayRef, c10::optional<double>, c10::optional<double>)’:
/home/ryad/workspace/pytorch/aten/src/ATen/native/cuda/UpSampleNearest2d.cu:305:44: warning: ‘c10::MemoryFormat c10::get_contiguous_memory_format()’ is deprecated [-Wdeprecated-declarations]
   Tensor output = at::empty_like(input, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
                                            ^~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/ryad/workspace/pytorch/c10/core/MemoryFormat.h:34:36: note: declared here
 C10_DEPRECATED inline MemoryFormat get_contiguous_memory_format() {
                                    ^~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/ryad/workspace/pytorch/aten/src/ATen/native/cuda/UpSampleNearest2d.cu:305:73: warning: ‘c10::MemoryFormat c10::get_contiguous_memory_format()’ is deprecated [-Wdeprecated-declarations]
   Tensor output = at::empty_like(input, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
                                                                         ^
/home/ryad/workspace/pytorch/c10/core/MemoryFormat.h:34:36: note: declared here
 C10_DEPRECATED inline MemoryFormat get_contiguous_memory_format() {
                                    ^~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/ryad/workspace/pytorch/aten/src/ATen/native/cuda/UpSampleNearest2d.cu: In function ‘at::Tensor at::native::upsample_nearest2d_backward_cuda(const at::Tensor&, c10::IntArrayRef, c10::IntArrayRef, c10::optional<double>, c10::optional<double>)’:
/home/ryad/workspace/pytorch/aten/src/ATen/native/cuda/UpSampleNearest2d.cu:328:54: warning: ‘c10::MemoryFormat c10::get_contiguous_memory_format()’ is deprecated [-Wdeprecated-declarations]
   Tensor grad_input = at::empty_like(grad_output, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
                                                      ^~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/ryad/workspace/pytorch/c10/core/MemoryFormat.h:34:36: note: declared here
 C10_DEPRECATED inline MemoryFormat get_contiguous_memory_format() {
                                    ^~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/ryad/workspace/pytorch/aten/src/ATen/native/cuda/UpSampleNearest2d.cu:328:83: warning: ‘c10::MemoryFormat c10::get_contiguous_memory_format()’ is deprecated [-Wdeprecated-declarations]
   Tensor grad_input = at::empty_like(grad_output, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
                                                                                   ^
/home/ryad/workspace/pytorch/c10/core/MemoryFormat.h:34:36: note: declared here
 C10_DEPRECATED inline MemoryFormat get_contiguous_memory_format() {
                                    ^~~~~~~~~~~~~~~~~~~~~~~~~~~~
ninja: build stopped: subcommand failed.

Expected behavior

My understanding is that the code should compile since using AT_ALL_TYPES_AND is compiling.

cc @yf225

[colesbury]

You probably also need to add a ScalarTypeToCPPType mapping from Byte to uint8_t here:

pytorch/c10/core/ScalarType.h

Lines 100 to 109 in e93e7b2

	template<>
	struct ScalarTypeToCPPType<c10::ScalarType::Bool> {
	using type = bool;
	// This is a workaround for the CUDA bug which prevents ::detail::ScalarTypeToCType<T>::type being used directly
	// due to ambiguous reference which can't to be resolved. For some reason it cant pick between at::detail and at::cuda::detail.
	// For repro example, please see: https://gist.github.com/izdeby/952ae7cf256ddb740a73776d39a7e7ba
	// TODO: remove once the bug is fixed.
	static type t;
	};

If that's not sufficient, there might be other similar mappings missing.

[r-zenine]

Hi @colesbury,
Thanks for the tip. I'll give it a try.

[colesbury]

PyTorch devs: we might want to ensure that the AT_DISPATCH macros work with all the built-in scalar types

[r-zenine]

Hi @colesbury ,

I'll submit the fix for Byte.