[Mlir-commits] [mlir] [mlir][x86vector] Lower BF16 vector.contract to FMA using AVX2 BF16 packed ops. (PR #170267)

[Adam Siemieniuk]

================
@@ -0,0 +1,290 @@
+//===- VectorContractBF16ToFMA.cpp
+//--------------------------------------------===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/Vector/IR/VectorOps.h"
+#include "mlir/Dialect/Vector/Utils/VectorUtils.h"
+#include "mlir/Dialect/X86Vector/Transforms.h"
+#include "mlir/Dialect/X86Vector/Utils/X86VectorUtils.h"
+#include "mlir/Dialect/X86Vector/X86VectorDialect.h"
+
+#include "mlir/IR/BuiltinAttributes.h"
+#include "mlir/IR/Dominance.h"
+#include "mlir/IR/PatternMatch.h"
+
+#include "mlir/Pass/Pass.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+#include "llvm/Support/Casting.h"
+
+using namespace mlir;
+using namespace mlir::vector;
+using namespace mlir::x86vector;
+
+// This function retrives the source operation of the load or transfer
+// reads and creates subviews for the BF16 packed-operations to
+// broadcast or load BF16 elements as F32 packed elements.
+//
+// For example:
+// ```
+//   vector.load %arg0[%c0, %c0, %c0]:memref<4x1x2xbf16>,vector<1x1x2xbf16>
+//   vector.load %arg0[%c0, %c0, %c0]:memref<1x32x2xbf16>,vector<1x8x2xbf16>
+// ```
+// to
+// ```
+//   memref.subview %arg0[%c0,%c0,%c1]:memref<4x1x2xbf16> to memref<1x1x1xbf16>
+//   memref.subview %arg1[%c0,%c0,%c0]:memref<1x32x2xbf16> to memref<1x8x2xbf16>
+//   memref.subview %arg0[%c0,%c0,%c0]:memref<4x1x2xbf16> to memref<1x1x1xbf16>
+// ```
+static FailureOr<llvm::SmallVector<mlir::memref::SubViewOp>>
+getSubviewFromVectorInput(Location loc, PatternRewriter &rewriter,
+                          mlir::Value prodOp, int64_t mnDim, int64_t vnniDim,
----------------
adam-smnk wrote:

nit: `mnDim` and `vnniDim` intuitively make me think about their positions but here it indicates their target sizes
I'd suggest adding either clarification to docs or some suffix like `Size` to names.

https://github.com/llvm/llvm-project/pull/170267