[Mlir-commits] [mlir] [mlir][vector][nfc] Improve comments in `getCompressedMaskOp` (PR #115663)

[Andrzej Warzyński]

https://github.com/banach-space created https://github.com/llvm/llvm-project/pull/115663

This PR updates and expands the high-level comment for
`getCompressedMaskOp` and renames input variables with more descriptive
names.

The current variable names are somewhat unclear (e.g., `scale`) or
derived from `memref` terminology (e.g., `intraDataOffset` from
`LinearizedMemRefInfo`). The updated names in this PR aim to better
align with the context and usage in the vector domain.


>From 5967f2df77e5ec4630526a10b384238dbc92b4c8 Mon Sep 17 00:00:00 2001
From: Andrzej Warzynski <andrzej.warzynski at arm.com>
Date: Sun, 10 Nov 2024 15:51:15 +0000
Subject: [PATCH] [mlir][vector][nfc] Improve comments in `getCompressedMaskOp`

This PR updates and expands the high-level comment for
`getCompressedMaskOp` and renames input variables with more descriptive
names.

The current variable names are somewhat unclear (e.g., `scale`) or
derived from `memref` terminology (e.g., `intraDataOffset` from
`LinearizedMemRefInfo`). The updated names in this PR aim to better
align with the context and usage in the vector domain.
---
 .../Transforms/VectorEmulateNarrowType.cpp    | 37 ++++++++++++-------
 1 file changed, 23 insertions(+), 14 deletions(-)

diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorEmulateNarrowType.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorEmulateNarrowType.cpp
index 58841f29698e0d..ef6f270b44cd62 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorEmulateNarrowType.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorEmulateNarrowType.cpp
@@ -36,24 +36,33 @@ using namespace mlir;
 #define DBGSNL() (llvm::dbgs() << "\n")
 #define LDBG(X) LLVM_DEBUG(DBGS() << X << "\n")
 
-/// Returns a compressed mask. The mask value is set only if any mask is present
-/// in the scale range. E.g., if `scale` equals to 2, and `intraDataOffset`
-/// equals to 1 (intraDataOffset strictly smaller than scale), the following
-/// mask:
+/// Returns a compressed mask. For example, when emulating `i8` with `i32` and
+/// when the number of source elements spans two `i32` elements, this method
+/// will compress `vector<8xi1>` into `vector<2xi1>`.
+///
+/// The compressed/output mask value is set iff any mask in the corresponding
+/// `numSrcElemsPerDest` range of uncompressed/input masks is set. E.g., if
+/// `numSrcElemsPerDest` equals to 2, and `numFrontPadElems` equals to 1, the
+/// following mask:
 ///
 ///   %mask = [1, 1, 0, 0, 0, 0]
 ///
-/// will first be padded with number of `intraDataOffset` zeros:
+/// will first be padded with number of `numFrontPadElems` zeros:
 ///   %mask = [0, 1, 1, 0, 0, 0, 0, 0]
 ///
 /// then it will return the following new compressed mask:
 ///
 ///   %mask = [1, 1, 0, 0]
+///
+/// `numFrontPadElems` is assumed to be strictly smaller than
+/// `numSrcElemsPerDest`.
 static FailureOr<Operation *> getCompressedMaskOp(OpBuilder &rewriter,
                                                   Location loc, Value mask,
-                                                  int origElements, int scale,
-                                                  int intraDataOffset = 0) {
-  auto numElements = (intraDataOffset + origElements + scale - 1) / scale;
+                                                  int numSrcElems,
+                                                  int numSrcElemsPerDest,
+                                                  int numFrontPadElems = 0) {
+  auto numElements = (numFrontPadElems + numSrcElems + numSrcElemsPerDest - 1) /
+                     numSrcElemsPerDest;
 
   Operation *maskOp = mask.getDefiningOp();
   SmallVector<vector::ExtractOp, 2> extractOps;
@@ -81,8 +90,8 @@ static FailureOr<Operation *> getCompressedMaskOp(OpBuilder &rewriter,
     size_t numMaskOperands = maskOperands.size();
     AffineExpr s0;
     bindSymbols(rewriter.getContext(), s0);
-    s0 = s0 + scale - 1;
-    s0 = s0.floorDiv(scale);
+    s0 = s0 + numSrcElemsPerDest - 1;
+    s0 = s0.floorDiv(numSrcElemsPerDest);
     OpFoldResult origIndex =
         getAsOpFoldResult(maskOperands[numMaskOperands - 1]);
     OpFoldResult maskIndex =
@@ -96,18 +105,18 @@ static FailureOr<Operation *> getCompressedMaskOp(OpBuilder &rewriter,
     ArrayRef<int64_t> maskDimSizes = constantMaskOp.getMaskDimSizes();
     size_t numMaskOperands = maskDimSizes.size();
     int64_t origIndex = maskDimSizes[numMaskOperands - 1];
-    int64_t startIndex = intraDataOffset / scale;
-    int64_t maskIndex = llvm::divideCeil(intraDataOffset + origIndex, scale);
+    int64_t startIndex = numFrontPadElems / numSrcElemsPerDest;
+    int64_t maskIndex = llvm::divideCeil(numFrontPadElems + origIndex, numSrcElemsPerDest);
 
     // TODO: we only want the mask between [startIndex, maskIndex] to be true,
     // the rest are false.
-    if (intraDataOffset != 0 && maskDimSizes.size() > 1)
+    if (numFrontPadElems != 0 && maskDimSizes.size() > 1)
       return failure();
 
     SmallVector<int64_t> newMaskDimSizes(maskDimSizes.drop_back());
     newMaskDimSizes.push_back(maskIndex);
 
-    if (intraDataOffset == 0) {
+    if (numFrontPadElems == 0) {
       newMask = rewriter.create<vector::ConstantMaskOp>(loc, newMaskType,
                                                         newMaskDimSizes);
     } else {