gpu: generic: deconvolution: move implementation from ocl folder #2074
Conversation
github-actions
bot
added
platform:gpu-intel
| GPU_INSTANCE_INTEL(intel::ocl::convolution_deconvolution_bwd_data_t) | ||
| GPU_INSTANCE_INTEL(intel::ocl::convolution_deconvolution_bwd_weights_t) | ||
| GPU_INSTANCE_NVIDIA(nvidia::cudnn_deconvolution_bwd_data_t) | ||
| GPU_INSTANCE_NVIDIA(nvidia::cudnn_deconvolution_bwd_weights_t) | ||
| GPU_INSTANCE_AMD(amd::miopen_deconvolution_bwd_data_t) | ||
| GPU_INSTANCE_AMD(amd::miopen_deconvolution_bwd_weights_t) | ||
| GPU_INSTANCE_GENERIC(generic::convolution_deconvolution_bwd_data_t) |
There was a problem hiding this comment.
Does this have a performance impact for Intel? If there's a case where both would match, it would now match weights before data.
There was a problem hiding this comment.
With these specific implementations, there should be no performance impact. They're mutually exclusive, looking for dnnl_backward_weights and dnnl_backward_data propagation kinds, respectively.
|
(nit) could you please use |
|
make test |
Done |
t4c1
changed the title
src/gpu/gpu_deconvolution_list.cpp
Outdated
| @@ -34,6 +34,9 @@ | |||
| #include "gpu/generic/sycl/ref_deconvolution.hpp" | |||
| #endif | |||
|
|
|||
| #include "gpu/generic/convolution_deconvolution.hpp" | |||
| #include "gpu/intel/ocl/convolution_deconvolution.hpp" | |||
There was a problem hiding this comment.
gpu/intel directory is not included in the build if gpu vendor is not intel. The header should always be guarded.
There was a problem hiding this comment.
Removed this include, as it was a duplicate of the guarded one.
| #include "gpu/intel/compute/utils.hpp" | ||
| #include "gpu/intel/gpu_primitive.hpp" | ||
| #include "gpu/intel/primitive_conf.hpp" |
There was a problem hiding this comment.
The generic implementations cannot include anything from vendor specific locations.
Move the fwd and bwd data implementations of deconv (convolution deconvolution) from opencl folder to generic one, as they are generic so they work with any convolution implementation. Also enables them for generic case.
This redoes the changes from #2020 that were reverted in 848b0f7 in a way that should be more acceptable.