Added CUDA support for interpolation AA bilinear/bicubic #3842
Conversation
| // Setup local buffers | ||
| scalar_t wx[256]; | ||
| scalar_t wy[256]; | ||
| scalar_t buffer1[256]; | ||
| scalar_t buffer2[256]; |
There was a problem hiding this comment.
Here is an assumption that resize scale is less than 127.
Otherwise, we can specify dynamic shared memory size before calling the cuda code, however we should then ensure that device has enough shared memory...
There was a problem hiding this comment.
This is fine for now as long as we raise an error if this configuration is not satisfied, which is what we currently do.
We will need to fix this for PyTorch though
| // Setup local buffers | ||
| scalar_t wx[256]; | ||
| scalar_t wy[256]; | ||
| scalar_t buffer1[256]; | ||
| scalar_t buffer2[256]; |
There was a problem hiding this comment.
This is fine for now as long as we raise an error if this configuration is not satisfied, which is what we currently do.
We will need to fix this for PyTorch though
| TORCH_CHECK(rheight < 127, "Max supported scale factor is 127"); | ||
| TORCH_CHECK(rwidth < 127, "Max supported scale factor is 127"); |
There was a problem hiding this comment.
Is the max scale factor supported different if we use bicubic?
There was a problem hiding this comment.
Thanks for catching that, yes for bicubic max scale should be at most 63. In general the check should be
TORCH_CHECK(rheight < 255 / interp_size, "Max supported scale factor is 127 (bilinear), 63 (bicubic)");
vfdev-5
force-pushed
the
vfdev-5/interpolate-aa-cuda
branch
from
c76e8fa to
4e43b5c
Compare
) Summary: * Added CUDA support for interpolation AA bilinear/bicubic * Fixed code formatting error * Fixed cuda tests on cpu-only builds * Fixed cuda internal torch check for bicubic mode Reviewed By: cpuhrsch Differential Revision: D28538753 fbshipit-source-id: 8b6c16db07a05fdba4b35cb8f5ec077b50691fb7
4 participants
Description: