[Mlir-commits] [mlir] [mlir][vector] Determine vector sizes from the result shape in the ca… (PR #88249)

[Han-Chung Wang]

================
@@ -812,3 +812,37 @@ func.func @test_vectorize_unpack_no_masks(%source: tensor<8x8x32x16xf32>, %dest:
     transform.yield
   } 
 }
+
+  // -----
+
+// CHECK-LABEL: test_vectorize_padded_pack_no_vector_sizes
+func.func @test_vectorize_padded_pack_no_vector_sizes(%arg0: tensor<32x7x15xf32>, %arg1: tensor<32x4x1x16x2xf32>) -> tensor<32x4x1x16x2xf32> {
+  %pad = arith.constant 0.000000e+00 : f32
+  %pack = tensor.pack %arg0 padding_value(%pad : f32) inner_dims_pos = [2, 1] inner_tiles = [16, 2] into %arg1 : tensor<32x7x15xf32> -> tensor<32x4x1x16x2xf32>
+  return %pack : tensor<32x4x1x16x2xf32>
+}
+//  CHECK-DAG: %[[cst:.*]] = arith.constant 0.000000e+00 : f32
+//  CHECK-DAG: %[[c0:.*]] = arith.constant 0 : index
+//  CHECK-DAG: %[[c32:.*]] = arith.constant 32 : index
+//  CHECK-DAG: %[[c7:.*]] = arith.constant 7 : index
+//  CHECK-DAG: %[[c15:.*]] = arith.constant 15 : index
+//      CHECK: %[[mask:.*]] = vector.create_mask %[[c32]], %[[c7]], %[[c15]] : vector<32x8x16xi1>
+//      CHECK: %[[masked_read:.*]] = vector.mask %[[mask]] {
+// CHECK-SAME:   vector.transfer_read %{{.*}}[%[[c0]], %[[c0]], %[[c0]]], %[[cst]]
+// CHECK-SAME:   {in_bounds = [true, true, true]} : tensor<32x7x15xf32>, vector<32x8x16xf32>
+// CHECK-SAME: } : vector<32x8x16xi1> -> vector<32x8x16xf32>
----------------
hanhanW wrote:

@dcaballe I'm a little outdated here. When the vector input sizes are not provided, do we prefer `transfer_read` with some `in_bounds` are false? Or do we prefer creating masks like this?

Some targets might not support masking, so I'd prefer the former one. My understanding is that we trigger masking only when input vector sizes are provided. Is the current one (i.e., transfer_read with masking) okay for the targets without masking support?

https://github.com/llvm/llvm-project/pull/88249