Disable minimize_shared_allocs pass by default #574
Conversation
This file contains hidden or 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
Signed-off-by: Aurore De Spirlet <[email protected]>
adedespirlet
force-pushed
the
disable_minimize_shared_allocs
branch
from
7666153 to
e3a0a76
Compare
Contributor
|
I suspect tests need to be updated, and we also want to keep it on in the options for extend attention and the likes. We can also think about a heuristic approach where we compute the total LDS footprint required for the kernel and dynamically turn this pass on if it exceeds the available amount, but that requires some target information to be available and may be better done with a MLIR-based pass where arch information is easier to come by. So not block on this. |
Signed-off-by: Tim Gymnich <[email protected]>
tgymnich
reviewed
Dec 19, 2025
Comment on lines
-124
to
+126
| # CHECK-COUNT-1: memref.alloc | ||
| # CHECK: scf.for %[[ARG3:.*]] = %[[C0]] to %[[C10]] step %[[C1]] { | ||
| # CHECK: scf.for %arg4 = %[[C0]] to %[[C4]] step %[[C1]] | ||
| # CHECK: memref.alloc |
Comment on lines
+217
to
+218
| # CHECK: memref.alloc | ||
| # CHECK: scf.for %[[ARG5:.*]] = %[[C0]] to %[[C4]] step %[[C1]] iter_args(%[[ARG6:.*]] = %[[CST:.*]]) -> (vector<4xf32>) { |
Comment on lines
-309
to
-332
| # CHECK-DAG: %[[C1:.*]] = arith.constant 1 : index | ||
| # CHECK-DAG: %[[C4:.*]] = arith.constant 4 : index | ||
| # CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index | ||
| # CHECK-DAG: %[[C10:.*]] = arith.constant 10 : index | ||
| # CHECK-DAG: %[[CST_0:.*]] = arith.constant dense<0.000000e+00> : vector<4xf32> | ||
| # CHECK-COUNT-1: memref.alloc | ||
| # CHECK: %[[CAST_INIT_B:.*]] = arith.index_cast | ||
| # CHECK: %[[WHILE:.*]] = scf.while (%[[ACC:.*]] = %[[CST_0]], %[[B:.*]] = %[[CAST_INIT_B]]) : (vector<4xf32>, index) -> (vector<4xf32>, index) { | ||
| # CHECK: %[[COND:.*]] = arith.cmpi slt, %[[B]], %[[C10]] : index | ||
| # CHECK: scf.condition(%[[COND]]) %[[ACC]], %[[B]] : vector<4xf32>, index | ||
| # CHECK: } do { | ||
| # CHECK: ^bb0(%[[ACC:.*]]: vector<4xf32>, %[[B:.*]]: index): | ||
| # CHECK: %[[FOR:.*]] = scf.for %[[ARG7:.*]] = %[[C0]] to %[[C4]] step %[[C1]] iter_args(%[[ARG8:.*]] = %[[CST_0]]) -> (vector<4xf32>) { | ||
| # CHECK-COUNT-1: vector.load | ||
| # CHECK: amdgpu.lds_barrier | ||
| # CHECK-COUNT-1: vector.store | ||
| # CHECK-COUNT-1: vector.load | ||
| # CHECK-COUNT-1: vector.store | ||
| # CHECK: amdgpu.lds_barrier | ||
| # CHECK-COUNT-2: vector.load | ||
| # CHECK: amdgpu.mfma | ||
| # CHECK: scf.yield | ||
| # CHECK-COUNT-4: vector.store | ||
|
|
Contributor
There was a problem hiding this comment.
AI removed a bit too much here..
Signed-off-by: Tim Gymnich <[email protected]>
Signed-off-by: Tim Gymnich <[email protected]>
Contributor
|
This is a prime example why we should do more unit-style tests and less integration tests. We should still have some integration tests, but avoid tight coupling in them: just check that key things happen. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
This change sets the default behavior of the minimize_shared_allocs pass to False.
Currently this pass reduces peak LDS memory usage by combining distinct memref.alloc operations into a single allocation with views. However this prevents the backend from inferring aliasing information. As a result the backend conservatively inserts unnecessary wait count instructions (vmcnt(0)), which breaks software pipelining.
While this optimization is beneficial for kernels with disjoint buffer lifecycles (like extended attention) it causes significant performance regressions in standard pipelined kernels.
Let's disable it by default and enable it explicitly only for the specific cases where it is required.