[Mlir-commits] [mlir] [mlir][gpu] remove async requirement for conversion to LLVM (PR #88634)

[Maksim Levental]

https://github.com/makslevental created https://github.com/llvm/llvm-project/pull/88634

This check makes sense (as a current analysis limitation) if the op is in fact async but I don't really understand why all of these conversions require async? E.g., if I'm doing everything synchronously, I believe all of runtime APIs calls are still correct/legal?

>From 60037d4168eb392284c5e039cc3209f9e615c62d Mon Sep 17 00:00:00 2001
From: max <maksim.levental at gmail.com>
Date: Sat, 13 Apr 2024 14:23:29 -0500
Subject: [PATCH] [mlir][gpu] remove async requirement for conversion to LLVM

---
 mlir/lib/Conversion/GPUCommon/GPUToLLVMConversion.cpp | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/mlir/lib/Conversion/GPUCommon/GPUToLLVMConversion.cpp b/mlir/lib/Conversion/GPUCommon/GPUToLLVMConversion.cpp
index 78d4e806246872..c15f97ca51883d 100644
--- a/mlir/lib/Conversion/GPUCommon/GPUToLLVMConversion.cpp
+++ b/mlir/lib/Conversion/GPUCommon/GPUToLLVMConversion.cpp
@@ -757,13 +757,10 @@ static LogicalResult areAllLLVMTypes(Operation *op, ValueRange operands,
 static LogicalResult
 isAsyncWithOneDependency(ConversionPatternRewriter &rewriter,
                          gpu::AsyncOpInterface op) {
-  if (op.getAsyncDependencies().size() != 1)
+  if (op.getAsyncToken() && op.getAsyncDependencies().size() != 1)
     return rewriter.notifyMatchFailure(
         op, "Can only convert with exactly one async dependency.");
 
-  if (!op.getAsyncToken())
-    return rewriter.notifyMatchFailure(op, "Can convert only async version.");
-
   return success();
 }