[MLIR] Modify lowering of gpu.alloc op to llvm (#69969)

If gpu.alloc has no asyn deependency ( in case if gpu.alloc has
hostShared allocation), create a new stream & synchronize. This PR is
follow up to #66401

Author: nbpatel

mlir/lib/Conversion/GPUCommon/GPUToLLVMConversion.cpp

@@ -836,7 +836,11 @@ LogicalResult ConvertAllocOpToGpuRuntimeCallPattern::matchAndRewrite(
   // Allocate the underlying buffer and store a pointer to it in the MemRef
   // descriptor.
   Type elementPtrType = this->getElementPtrType(memRefType);
-  auto stream = adaptor.getAsyncDependencies().front();
+
+  auto nullPtr = rewriter.create<mlir::LLVM::ZeroOp>(loc, llvmPointerType);
+  Value stream = adaptor.getAsyncDependencies().empty()
+                     ? nullPtr
+                     : adaptor.getAsyncDependencies().front();
 
   auto isHostShared = rewriter.create<mlir::LLVM::ConstantOp>(
       loc, llvmInt8Type, rewriter.getI8IntegerAttr(isShared));
@@ -855,7 +859,12 @@ LogicalResult ConvertAllocOpToGpuRuntimeCallPattern::matchAndRewrite(
   auto memRefDescriptor = this->createMemRefDescriptor(
       loc, memRefType, allocatedPtr, alignedPtr, shape, strides, rewriter);
 
-  rewriter.replaceOp(allocOp, {memRefDescriptor, stream});
+  if (allocOp.getAsyncToken()) {
+    // Async alloc: make dependent ops use the same stream.
+    rewriter.replaceOp(allocOp, {memRefDescriptor, stream});
+  } else {
+    rewriter.replaceOp(allocOp, {memRefDescriptor});
+  }
 
   return success();
 }

mlir/test/Conversion/GPUCommon/lower-alloc-to-gpu-runtime-calls.mlir

@@ -19,4 +19,22 @@ module attributes {gpu.container_module} {
     gpu.wait [%3]
     return
   }
+
+  // CHECK-LABEL: llvm.func @alloc_sync
+  // CHECK-SAME: %[[size:.*]]: i64
+  func.func @alloc_sync(%size : index) {
+    // CHECK: %[[gep:.*]] = llvm.getelementptr {{.*}}[%[[size]]]
+    // CHECK: %[[size_bytes:.*]] = llvm.ptrtoint %[[gep]]
+    // CHECK: %[[nullptr:.*]] = llvm.mlir.zero
+    // CHECK: %[[isHostShared:.*]] = llvm.mlir.constant 
+    // CHECK: llvm.call @mgpuMemAlloc(%[[size_bytes]], %[[nullptr]], %[[isHostShared]])
+    %0 = gpu.alloc host_shared (%size) : memref<?xf32>
+    // CHECK: %[[stream:.*]] = llvm.call @mgpuStreamCreate()
+    %1 = gpu.wait async
+    %2 = gpu.dealloc async [%1] %0 : memref<?xf32>
+    // CHECK: llvm.call @mgpuStreamSynchronize(%[[stream]])
+    // CHECK: llvm.call @mgpuStreamDestroy(%[[stream]])
+    gpu.wait [%2]
+    return
+  }
 }