423A35C7/pytorch3d
1
0
Fork 0
mirror of https://github.com/facebookresearch/pytorch3d.git synced 2026-03-04 03:05:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

Make cuda tensors contiguous in host function and remove contiguous check

Browse Source 
Summary:
Update the cuda kernels to:
- remove contiguous checks for the grad tensors and for cpu functions which use accessors
- for cuda implementations call `.contiguous()` on all tensors in the host function before invoking the kernel

Reviewed By: gkioxari

Differential Revision: D21598008

fbshipit-source-id: 9b97bda4582fd4269c8a00999874d4552a1aea2d
This commit is contained in:
Nikhila Ravi
2020-05-15 14:58:04 -07:00
committed by Facebook GitHub Bot
parent a8377f1f06
commit 3fef506895
21 changed files with 219 additions and 233 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
pytorch3d/csrc/compositing/alpha_composite.cu
View File

@@ -168,6 +168,8 @@ at::Tensor alphaCompositeCudaForward(
// doubles. Currently, support is for floats only.
alphaCompositeCudaForwardKernel<<<numBlocks, threadsPerBlock, 0, stream>>>(
// clang-format off
// As we are using packed accessors here the tensors
// do not need to be made contiguous.
result.packed_accessor64<float, 4, at::RestrictPtrTraits>(),
features.packed_accessor64<float, 2, at::RestrictPtrTraits>(),
alphas.packed_accessor64<float, 4, at::RestrictPtrTraits>(),
@@ -211,6 +213,8 @@ std::tuple<at::Tensor, at::Tensor> alphaCompositeCudaBackward(
// doubles. Currently, support is for floats only.
alphaCompositeCudaBackwardKernel<<<numBlocks, threadsPerBlock, 0, stream>>>(
// clang-format off
// As we are using packed accessors here the tensors
// do not need to be made contiguous.
grad_features.packed_accessor64<float, 2, at::RestrictPtrTraits>(),
grad_alphas.packed_accessor64<float, 4, at::RestrictPtrTraits>(),
grad_outputs.packed_accessor64<float, 4, at::RestrictPtrTraits>(),