Add MatmulParams::cluster_dims parameter #3574
Merged
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
Following #3557 we can specify the cluster size for our fusions. Currently we don't do anything explicitly with CGAs, but this can help guarantee that tiles are scheduled onto GPCs in pairs. Each GPC has a number of TPCs, each of which holds 2 SMs, so this lets us take advantage of caching at the TPC and GPC level for operand loads, in addition to L2.
|
!test |
This reverts commit e3f611c.
|
!test |
rdspring1
approved these changes
Dec 11, 2024
tests/cpp/test_matmul.cpp
Outdated
| @@ -3663,7 +3663,7 @@ TEST_F(HopperMatmulTest, HSH_NT_128BSwizzle) { | |||
| const int64_t cta_m = 2 * getM(macro); | |||
| const int64_t cta_n = 1 * getN(macro); | |||
|
|
|||
| constexpr std::tuple<int64_t, int64_t, int64_t> cluster_dims{2, 1, 1}; | |||
| constexpr std::tuple<int, int, int> cluster_dims{2, 1, 1}; | |||
There was a problem hiding this comment.
super nitpick: can we stick with int64_t for consistency?
Suggested change
| constexpr std::tuple<int, int, int> cluster_dims{2, 1, 1}; | |
| constexpr std::tuple<int64_t, int64_t, int64_t> cluster_dims{2, 1, 1}; |
There was a problem hiding this comment.
Yeah, I was mostly doing that because the MatmulParams entries are int, but we should probably just change MatmulParams instead (in another PR).
There was a problem hiding this comment.
Done. Much smaller PR now...
|
!build |
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Outdated suggestions cannot be applied.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.
Following #3557 we can specify the cluster size for our fusions. Currently we don't do anything explicitly with CGAs, but this can help guarantee that tiles are scheduled onto GPCs in pairs. Each GPC has a number of TPCs, each of which holds 2 SMs, so this lets us take advantage of caching at the TPC and GPC level for operand loads, in addition to L2.
This PR enables this with a default size of
{2, 1, 1}for the Hopper scheduler. The parameter is ignored in the Ampere scheduler.It is not yet plumbed into the heuristic plugin API yet. I thought maybe we should wait until we have more parameters related to CGAs to do that.