[BACKEND][FRONTEND] Fix problems with test_matmul (#973)
1. Handle induction variable when step is negative 2. Restore async_wait that accidentally deleted 3. Add missing induction variable in prefetch 4. Add device property functions Co-authored-by: Philippe Tillet <Phil.Tillet@gmail.com>
This commit is contained in:
@@ -4783,10 +4783,15 @@ private:
|
||||
decomposed = true;
|
||||
});
|
||||
|
||||
// async wait is supported in Ampere and later
|
||||
mod.walk([&](triton::gpu::AsyncWaitOp asyncWaitOp) -> void {
|
||||
if (!triton::gpu::AsyncWaitOp::isSupported(computeCapability) ||
|
||||
decomposed) {
|
||||
if (!triton::gpu::AsyncWaitOp::isSupported(computeCapability)) {
|
||||
// async wait is supported in Ampere and later
|
||||
asyncWaitOp.erase();
|
||||
} else if (decomposed) {
|
||||
// Wait for all previous async ops
|
||||
OpBuilder builder(asyncWaitOp);
|
||||
auto newAsyncWaitOp =
|
||||
builder.create<triton::gpu::AsyncWaitOp>(asyncWaitOp.getLoc(), 0);
|
||||
asyncWaitOp.erase();
|
||||
}
|
||||
});
|
||||
|
Reference in New Issue
Block a user