[mlir][linalg] Implement Winograd Conv2D.

This patch implements the Winograd Conv2D algorithm. It supports several
configurations of Winograd Conv2D, including F(2, 3), F(4, 3) and F(2,
5). These configurations show that the implementation can support
different kernel size (3 and 5) and different output size (2 and 4).
Besides symetric kernel size 3x3 and 5x5, this patch also supports 1x3,
3x1, 1x5, and 5x1 kernels.

The implementation is based on the paper, Fast Algorithm for
Convolutional Neural Networks. (https://arxiv.org/abs/1509.09308)

Author: Hsiangkai

mlir/include/mlir/Dialect/Linalg/Transforms/Transforms.h

@@ -1692,6 +1692,9 @@ void populateTransposeMatmulPatterns(RewritePatternSet &patterns,
 void populateBlockPackMatmulPatterns(RewritePatternSet &patterns,
                                      const ControlBlockPackMatmulFn &controlFn);
 
+/// Patterns to apply Winograd Conv2D algorithm.
+void populateWinogradConv2DPatterns(RewritePatternSet &patterns);
+
 } // namespace linalg
 } // namespace mlir
 

mlir/lib/Dialect/Linalg/Transforms/CMakeLists.txt

@@ -38,6 +38,7 @@ add_mlir_dialect_library(MLIRLinalgTransforms
   Transforms.cpp
   TransposeConv2D.cpp
   Vectorization.cpp
+  WinogradConv2D.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/Linalg