================
@@ -233,3 +233,41 @@ func.func @fat_buffer_load_to_rocdl_f32(%global :
memref<128x72xf32, #amdgpu.add
: f32, memref<128x72xf32, #amdgpu.address_space<fat_raw_buffer>>,
memref<64x64xf32, #gpu.address_space<workgroup>>
func.return
}
+
+// CHECK-LABEL: func @global_load_to_rocdl_async_f32
+// CHECK-SAME: (%[[ARG0:.*]]: memref<128x72xf32, #gpu.address_space<global>>)
+func.func @global_load_to_rocdl_async_f32(%global : memref<128x72xf32,
#gpu.address_space<global>>) {
+ %c0 = arith.constant 0 : index
+ %c12 = arith.constant 12 : index
+ %c32 = arith.constant 32 : index
+ %alloc = memref.alloc() : memref<64x64xf32, #gpu.address_space<workgroup>>
+ // CHECK: %[[GLOBAL_DESC:.*]] = builtin.unrealized_conversion_cast %[[ARG0]]
+
+ // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
+ // CHECK-DAG: %[[IC0:.*]] = builtin.unrealized_conversion_cast %c0 : index
to i64
+ // CHECK-DAG: %[[C12:.*]] = arith.constant 12 : index
+ // CHECK-DAG: %[[IC12:.*]] = builtin.unrealized_conversion_cast %[[C12]]
+ // CHECK-DAG: %[[C32:.*]] = arith.constant 32 : index
+ // CHECK-DAG: %[[IC32:.*]] = builtin.unrealized_conversion_cast %[[C32]]
+
+ // CHECK: %[[ALLOC:.*]] = memref.alloc()
+ // CHECK: %[[LDS_DESC:.*]] = builtin.unrealized_conversion_cast
+ // CHECK: %[[GLOBAL_BASE:.*]] = llvm.extractvalue %[[GLOBAL_DESC]][1]
+
+ // CHECK: %[[C72:.*]] = llvm.mlir.constant(72 : index) : i64
+ // CHECK: %[[MUL:.*]] = llvm.mul %[[IC12]], %[[C72]] : i64
+ // CHECK: %[[SRC_OFFSET:.*]] = llvm.add %[[MUL]], %[[IC0]] : i64
+
+ // CHECK: %[[GLOBAL_PTR:.*]] = llvm.getelementptr
%[[GLOBAL_BASE]][%[[SRC_OFFSET]]]
+ // CHECK: %[[LDS_BASE:.*]] = llvm.extractvalue %[[LDS_DESC]][1]
+
+ // CHECK: %[[C64:.*]] = llvm.mlir.constant(64 : index) : i64
+ // CHECK: %[[MUL_2:.*]] = llvm.mul %[[IC32]], %[[C64]] : i64
+ // CHECK: %[[DST_OFFSET:.*]] = llvm.add %[[MUL_2]], %[[IC0]] : i64
+
+ // CHECK: %[[LDS_PTR:.*]] = llvm.getelementptr %[[LDS_BASE]][%[[DST_OFFSET]]]
+ // CHECK: rocdl.load.async.to.lds %[[GLOBAL_PTR]], %[[LDS_PTR]], 4
+ amdgpu.gather_to_lds async %global[%c12, %c0], %alloc[%c32, %c0]
+ : f32, memref<128x72xf32, #gpu.address_space<global>>, memref<64x64xf32,
#gpu.address_space<workgroup>>
----------------
lialan wrote:
nit: let's add a fat_raw_buffer test as well.
https://github.com/llvm/llvm-project/pull/181082
_______________________________________________
llvm-branch-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
