[Mlir-commits] [mlir] [mlir][linalg] Implement Conv2D using Winograd Conv2D algorithm (PR #96181)

[Oleksandr Alex Zinenko]

================
@@ -0,0 +1,248 @@
+// RUN: mlir-opt %s -split-input-file -test-linalg-transform-patterns=test-winograd-conv2d | FileCheck %s
+
+func.func @conv2d_4x4_3x3(%arg0: tensor<2x6x6x5xf32>, %arg1: tensor<2x3x3x5xf32>, %arg2: tensor<1xf32>) -> tensor<2x4x4x2xf32> {
+  %0 = tensor.empty() : tensor<2x4x4x2xf32>
+  %1 = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2, d3) -> (0)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2 : tensor<1xf32>) outs(%0 : tensor<2x4x4x2xf32>) {
+  ^bb0(%in: f32, %out: f32):
+    linalg.yield %in : f32
+  } -> tensor<2x4x4x2xf32>
+  %2 = linalg.conv_2d_nhwc_fhwc {dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %arg1 : tensor<2x6x6x5xf32>, tensor<2x3x3x5xf32>) outs(%1 : tensor<2x4x4x2xf32>) -> tensor<2x4x4x2xf32>
+  return %2 : tensor<2x4x4x2xf32>
+}
+
+// CHECK: #[[$MAP0:.+]] = affine_map<(d0, d1, d2, d3) -> (0)>
+// CHECK: #[[$MAP1:.+]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
+// CHECK-LABEL: func.func @conv2d_4x4_3x3
+// CHECK-SAME:  (%[[ARG0:.*]]: tensor<2x6x6x5xf32>, %[[ARG1:.*]]: tensor<2x3x3x5xf32>, %[[ARG2:.*]]: tensor<1xf32>) -> tensor<2x4x4x2xf32> {
+// CHECK:       %[[S0:.*]] = tensor.empty() : tensor<2x4x4x2xf32>
+// CHECK-NEXT:  %[[S1:.*]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%[[ARG2]] : tensor<1xf32>) outs(%[[S0]] : tensor<2x4x4x2xf32>) {
----------------
ftynse wrote:

We don't care about these operations being strictly on the next line, let's not overconstrain tests.

https://github.com/llvm/llvm-project/pull/96181