[Mlir-commits] [mlir] [mlir][linalg] Enable masked vectorisation for depthwise convolutions (PR #81625)

[Cullen Rhodes]

================
@@ -0,0 +1,187 @@
+// RUN: mlir-opt -split-input-file -transform-interpreter -cse %s | FileCheck %s
+
+func.func @depthwise_conv1d_nwc_wc_1x8x3xi8_tensor(%input: tensor<1x8x?xi8>,
+                                                   %filter: tensor<1x?xi8>,
+                                                   %output: tensor<1x8x?xi8>) -> (tensor<1x8x?xi8>) {
+  %res = linalg.depthwise_conv_1d_nwc_wc
+    {dilations = dense<1> : vector<1xi64>,
+    strides = dense<1> : vector<1xi64>}
+    ins(%input, %filter : tensor<1x8x?xi8>, tensor<1x?xi8>)
+    outs(%output : tensor<1x8x?xi8>) -> tensor<1x8x?xi8>
+  return %res : tensor<1x8x?xi8>
+}
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
+    %0 = transform.structured.match ops{["linalg.depthwise_conv_1d_nwc_wc"]} in %arg0 : (!transform.any_op) -> !transform.any_op
+    transform.structured.vectorize %0 vector_sizes [1, 8, 4, 1] : !transform.any_op
+    transform.yield
+  }
+}
+
+// CHECK-LABEL:   func.func @depthwise_conv1d_nwc_wc_1x8x3xi8_tensor(
+// CHECK-SAME:      %[[INPUT:.*]]: tensor<1x8x?xi8>,
+// CHECK-SAME:      %[[FILTER:.*]]: tensor<1x?xi8>,
+// CHECK-SAME:      %[[OUTPUT:.*]]: tensor<1x8x?xi8>) -> tensor<1x8x?xi8> {
+
+// CHECK:           %[[C1:.*]] = arith.constant 1 : index
+// CHECK:           %[[C0:.*]] = arith.constant 0 : index
+// CHECK:           %[[PAD:.*]] = arith.constant 0 : i8
+
+/// Create a mask for the input tensor
+// CHECK:           %[[C2:.*]] = arith.constant 2 : index
+// CHECK:           %[[CH_DIM_IN:.*]] = tensor.dim %[[INPUT]], %[[C2]] : tensor<1x8x?xi8>
+// CHECK:           %[[C8:.*]] = arith.constant 8 : index
+// CHECK:           %[[MASK_IN:.*]] = vector.create_mask %[[C1]], %[[C8]], %[[CH_DIM_IN]] : vector<1x8x4xi1>
+/// Read the input tensor
+// CHECK:           %[[VEC_IN:.*]] = vector.mask %[[MASK_IN]] { vector.transfer_read %[[INPUT]]{{\[}}%[[C0]], %[[C0]], %[[C0]]], %[[PAD]] : tensor<1x8x?xi8>, vector<1x8x4xi8> } : vector<1x8x4xi1> -> vector<1x8x4xi8>
+
+/// Create a mask for the filter tensor
+// CHECK:           %[[CH_DIM_FLT:.*]] = tensor.dim %[[FILTER]], %[[C1]] : tensor<1x?xi8>
+// CHECK:           %[[MASK_FLT:.*]] = vector.create_mask %[[C1]], %[[CH_DIM_FLT]] : vector<1x4xi1>
+/// Read the filter tensor
+// CHECK:           %[[VEC_FLT:.*]] = vector.mask %[[MASK_FLT]] { vector.transfer_read %[[FILTER]]{{\[}}%[[C0]], %[[C0]]], %[[PAD]] : tensor<1x?xi8>, vector<1x4xi8> } : vector<1x4xi1> -> vector<1x4xi8>
+
+/// Create a mask for the output tensor
+// CHECK:           %[[CH_DIM_OUT:.*]] = tensor.dim %[[OUTPUT]], %[[C2]] : tensor<1x8x?xi8>
+// CHECK:           %[[MASK_OUT:.*]] = vector.create_mask %[[C1]], %[[C8]], %[[CH_DIM_OUT]] : vector<1x8x4xi1>
+// CHECK:           %[[VEC_OUT:.*]] = vector.mask %[[MASK_OUT]] { vector.transfer_read %[[OUTPUT]]{{\[}}%[[C0]], %[[C0]], %[[C0]]], %[[PAD]] : tensor<1x8x?xi8>, vector<1x8x4xi8> } : vector<1x8x4xi1> -> vector<1x8x4xi8>
+
+/// Convolution
+// CHECK:           %[[IN_1:.*]] = vector.extract_strided_slice %[[VEC_IN]] {offsets = [0, 0, 0], sizes = [1, 8, 4], strides = [1, 1, 1]} : vector<1x8x4xi8> to vector<1x8x4xi8>
----------------
c-rhodes wrote:

the insert/extract strided slice operations in this test and the scalable equivalent are not necessary, with canonicalization these are removed which I think makes it easier to follow, as well as making it more obvious what is happening for the final test where the dilation is 2 and the `extract_strided_slice` is necessary.

https://github.com/llvm/llvm-project/pull/81625