More improvements to Conv2d speed #98
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This should allow each patch to be used for many filters. TODO: - Play around with index order. Should filters be smallest? - Generalize: take out hardcoded lsize
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Trying computing multiple filters per group to reduce the number of times the image has to be loaded. So far, it hasn't seemed to help that much, though.
Maybe it makes sense that this doesn't help. For global filters (convolution), the limiting factor seems to be FLOPS, not memory access, so reducing image loads wouldn't make a difference.
For local filters, the amount of memory in the filters (nf * ni * nj * nc * si * sj) is much greater than in the image (ni * nj * nc), assuming the filter stride is 1, so reducing image loads won't make a difference. It would only be the case when the stride is about the size of the kernel width that the image data would be on par with the filter data for one workgroup, and so in this case computing multiple kernels per group might help.