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

[Oleksandr Alex Zinenko]

================
@@ -0,0 +1,321 @@
+//===- WinogradConv2D.cpp - Winograd Conv2D implementation ----------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// Implement Winograd Conv2D algorithm. The implementation is based on the
+// paper: Fast Algorithms for Convolutional Neural Networks
+// (https://arxiv.org/abs/1509.09308)
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/Linalg/IR/Linalg.h"
+#include "mlir/Dialect/Tensor/IR/Tensor.h"
+#include "mlir/Dialect/Tosa/Utils/ConversionUtils.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+#include "llvm/Support/MathExtras.h"
+
+namespace mlir {
+namespace linalg {
+
+namespace {
+
+using TransformMapKeyTy = std::pair<int, int>;
+
+// We use F(m, r) to define the size of minimal filtering algorithms.
+// m is the output dimension and r is the filter dimension. We can get
+// the input dimension, alpha, from the formula, alpha = m + r - 1.
+//
+// For example, when m = 2 and r = 3, we know its input size is 4.
+// The Conv2D will operate on 4x4 input data with 3x3 filter and get
+// 2x2 output result.
+constexpr TransformMapKeyTy F_2_3{2, 3};
+constexpr TransformMapKeyTy F_4_3{4, 3};
+constexpr TransformMapKeyTy F_2_5{2, 5};
+
+Value collapse2DData(RewriterBase &rewriter, Location loc, Value data) {
+  auto type = cast<ShapedType>(data.getType());
+  auto elementType = type.getElementType();
+  auto shape = type.getShape();
+  auto collapseType = RankedTensorType::get(
+      {shape[0] * shape[1] * shape[2] * shape[3], shape[4], shape[5]},
+      elementType);
+  SmallVector<ReassociationIndices> reassociation = {{0, 1, 2, 3}, {4}, {5}};
+  return rewriter.create<tensor::CollapseShapeOp>(loc, collapseType, data,
+                                                  reassociation);
+}
+
+// This function generates linalg.batch_matmul to multiply input with filter.
+// linalg.batch_matmul only supports 3-dimension data sets. We can treat
+// tileH x tileW x H x W data as the 1-dimension data array. That is to convert
+// [tileH, tileW, H, W, N, C] to [tileH x tileW x H x W, N, C]. In this way, we
+// can convert 6-dimension input data to 3-dimension representation that is
+// suitable for linalg.batch_matmul.
+//
+// Batched matmul will do the matrix multiply with the reduction on channel.
+//
+// We get
+//
+// %collapsed_input = tensor.collapse_shape %input
+// %collapsed_filter = tensor.collapse_shape %filter
+// %ret = linalg.batch_matmul %collapsed_input, %collapsed_filter
+// %expanded_ret = tensor.expand_shape %ret
----------------
ftynse wrote:

Note that you can use `linalg.generic` with explicit indexing maps and multiple batch dimensions that will remove the need to manipulate shapes. Not insisting on this change, but something to consider and maybe comment as to why the choice was made one way or another.

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