[Mlir-commits] [mlir] [mlir][Vector] Update VectorEmulateNarrowType.cpp (2/N) (PR #123527)
Andrzej Warzyński
llvmlistbot at llvm.org
Sun Feb 2 07:46:16 PST 2025
https://github.com/banach-space updated https://github.com/llvm/llvm-project/pull/123527
>From 3954c8f5dafae1f3435d490b3f114420dabbe925 Mon Sep 17 00:00:00 2001
From: Andrzej Warzynski <andrzej.warzynski at arm.com>
Date: Sun, 2 Feb 2025 15:21:41 +0000
Subject: [PATCH 1/2] [mlir][vector][nfc] Fix typos in
"VectorEmulateNarrowType.cpp"
Updates `emulatedVectorLoad` that was introduced in #115922.
Specifically, ATM `emulatedVectorLoad` mixes "emulated type" and
"container type". This only became clear after #123526 in which the
concepts of "emulated" and "container" types were introduced.
This is an NFC change and simply updates the variable naming.
---
.../Transforms/VectorEmulateNarrowType.cpp | 24 ++++++++++---------
1 file changed, 13 insertions(+), 11 deletions(-)
diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorEmulateNarrowType.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorEmulateNarrowType.cpp
index 63365cb5446124..0d310dc8be2fe9 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorEmulateNarrowType.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorEmulateNarrowType.cpp
@@ -278,23 +278,25 @@ static Value dynamicallyInsertSubVector(RewriterBase &rewriter, Location loc,
return dest;
}
-/// Returns the op sequence for an emulated sub-byte data type vector load.
-/// specifically, use `emulatedElemType` for loading a vector of `origElemType`.
-/// The load location is given by `base` and `linearizedIndices`, and the
-/// load size is given by `numEmulatedElementsToLoad`.
+/// Emulate a vector load for `emulatedElemTy` using `containerElemTy`
+///
+/// Specifically, use `containerElemTy` for loading a vector of
+/// `emulatedElemTy`. The load location is given by `base` and
+/// `linearizedIndices`, and the load size is given by
+/// `numEmulatedElementsToLoad`.
static VectorValue emulatedVectorLoad(OpBuilder &rewriter, Location loc,
Value base,
OpFoldResult linearizedIndices,
- int64_t numEmultedElementsToLoad,
- Type origElemType,
- Type emulatedElemType) {
- auto scale = emulatedElemType.getIntOrFloatBitWidth() /
- origElemType.getIntOrFloatBitWidth();
+ int64_t numContainerElemsToLoad,
+ Type emulatedElemTy,
+ Type containerElemTy) {
+ auto scale = containerElemTy.getIntOrFloatBitWidth() /
+ emulatedElemTy.getIntOrFloatBitWidth();
auto newLoad = rewriter.create<vector::LoadOp>(
- loc, VectorType::get(numEmultedElementsToLoad, emulatedElemType), base,
+ loc, VectorType::get(numContainerElemsToLoad, containerElemTy), base,
getValueOrCreateConstantIndexOp(rewriter, loc, linearizedIndices));
return rewriter.create<vector::BitCastOp>(
- loc, VectorType::get(numEmultedElementsToLoad * scale, origElemType),
+ loc, VectorType::get(numContainerElemsToLoad * scale, emulatedElemTy),
newLoad);
}
>From 9fab1bb9988cacea6075690f74dcec0c6740b2f8 Mon Sep 17 00:00:00 2001
From: Andrzej Warzynski <andrzej.warzynski at arm.com>
Date: Sun, 2 Feb 2025 15:36:33 +0000
Subject: [PATCH 2/2] [mlir][Vector] Update VectorEmulateNarrowType.cpp (2/N)
This is PR 2 in a series of N patches aimed at improving
"VectorEmulateNarrowType.cpp". This is mainly minor refactoring, no
major functional changes are made/added.
This PR renames the variable "scale". Note, "scale" could mean either:
* "original-elements-per-emulated-type", or
* "emulated-elements-per-original-type".
While from the context it is clear that it's always the former (original
type is always a sub-byte type and the emulated type is usually `i8`),
this PR reduces the cognitive load by making this clear.
**DEPENDS ON:**
* #123526 123526
Please only review the [top
commit](https://github.com/llvm/llvm-project/pull/123527/commits/d40b31bb098e874be488182050c68b887e8d091a).
**GitHub issue to track this work**:
https://github.com/llvm/llvm-project/issues/123630
---
.../Transforms/VectorEmulateNarrowType.cpp | 52 +++++++++----------
1 file changed, 26 insertions(+), 26 deletions(-)
diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorEmulateNarrowType.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorEmulateNarrowType.cpp
index 0d310dc8be2fe9..a0e170c7a04125 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorEmulateNarrowType.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorEmulateNarrowType.cpp
@@ -290,13 +290,13 @@ static VectorValue emulatedVectorLoad(OpBuilder &rewriter, Location loc,
int64_t numContainerElemsToLoad,
Type emulatedElemTy,
Type containerElemTy) {
- auto scale = containerElemTy.getIntOrFloatBitWidth() /
+ auto emulatedPerContainerElem = containerElemTy.getIntOrFloatBitWidth() /
emulatedElemTy.getIntOrFloatBitWidth();
auto newLoad = rewriter.create<vector::LoadOp>(
loc, VectorType::get(numContainerElemsToLoad, containerElemTy), base,
getValueOrCreateConstantIndexOp(rewriter, loc, linearizedIndices));
return rewriter.create<vector::BitCastOp>(
- loc, VectorType::get(numContainerElemsToLoad * scale, emulatedElemTy),
+ loc, VectorType::get(numContainerElemsToLoad * emulatedPerContainerElem, emulatedElemTy),
newLoad);
}
@@ -388,10 +388,10 @@ static Value extractSliceIntoByte(ConversionPatternRewriter &rewriter,
"sliceNumElements * vector element size must be less than or equal to 8");
assert(8 % vectorElementType.getIntOrFloatBitWidth() == 0 &&
"vector element must be a valid sub-byte type");
- auto scale = 8 / vectorElementType.getIntOrFloatBitWidth();
+ auto emulatedPerContainerElem = 8 / vectorElementType.getIntOrFloatBitWidth();
auto emptyByteVector = rewriter.create<arith::ConstantOp>(
- loc, VectorType::get({scale}, vectorElementType),
- rewriter.getZeroAttr(VectorType::get({scale}, vectorElementType)));
+ loc, VectorType::get({emulatedPerContainerElem}, vectorElementType),
+ rewriter.getZeroAttr(VectorType::get({emulatedPerContainerElem}, vectorElementType)));
auto extracted = staticallyExtractSubvector(rewriter, loc, vector,
extractOffset, sliceNumElements);
return staticallyInsertSubvector(rewriter, loc, extracted, emptyByteVector,
@@ -656,9 +656,9 @@ struct ConvertVectorMaskedStore final
"(bit-wise misalignment)");
}
- int scale = containerBits / emulatedBits;
+ int emulatedPerContainerElem = containerBits / emulatedBits;
int origElements = op.getValueToStore().getType().getNumElements();
- if (origElements % scale != 0)
+ if (origElements % emulatedPerContainerElem != 0)
return failure();
auto stridedMetadata =
@@ -708,11 +708,11 @@ struct ConvertVectorMaskedStore final
// FIXME: Make an example based on the comment above work (see #115460 for
// reproducer).
FailureOr<Operation *> newMask =
- getCompressedMaskOp(rewriter, loc, op.getMask(), origElements, scale);
+ getCompressedMaskOp(rewriter, loc, op.getMask(), origElements, emulatedPerContainerElem);
if (failed(newMask))
return failure();
- auto numElements = (origElements + scale - 1) / scale;
+ auto numElements = (origElements + emulatedPerContainerElem - 1) / emulatedPerContainerElem;
auto newType = VectorType::get(numElements, containerElemTy);
auto passThru = rewriter.create<arith::ConstantOp>(
loc, newType, rewriter.getZeroAttr(newType));
@@ -721,7 +721,7 @@ struct ConvertVectorMaskedStore final
loc, newType, adaptor.getBase(), linearizedIndices,
newMask.value()->getResult(0), passThru);
- auto newBitCastType = VectorType::get(numElements * scale, emulatedElemTy);
+ auto newBitCastType = VectorType::get(numElements * emulatedPerContainerElem, emulatedElemTy);
Value valueToStore =
rewriter.create<vector::BitCastOp>(loc, newBitCastType, newLoad);
valueToStore = rewriter.create<arith::SelectOp>(
@@ -765,7 +765,7 @@ struct ConvertVectorLoad final : OpConversionPattern<vector::LoadOp> {
op, "impossible to pack emulated elements into container elements "
"(bit-wise misalignment)");
}
- int scale = containerBits / emulatedBits;
+ int emulatedPerContainerElem = containerBits / emulatedBits;
// Adjust the number of elements to load when emulating narrow types,
// and then cast back to the original type with vector.bitcast op.
@@ -797,7 +797,7 @@ struct ConvertVectorLoad final : OpConversionPattern<vector::LoadOp> {
// compile time as they must be constants.
auto origElements = op.getVectorType().getNumElements();
- bool isAlignedEmulation = origElements % scale == 0;
+ bool isAlignedEmulation = origElements % emulatedPerContainerElem == 0;
auto stridedMetadata =
rewriter.create<memref::ExtractStridedMetadataOp>(loc, op.getBase());
@@ -818,9 +818,9 @@ struct ConvertVectorLoad final : OpConversionPattern<vector::LoadOp> {
: getConstantIntValue(linearizedInfo.intraDataOffset);
// Always load enough elements which can cover the original elements.
- int64_t maxintraDataOffset = foldedIntraVectorOffset.value_or(scale - 1);
+ int64_t maxintraDataOffset = foldedIntraVectorOffset.value_or(emulatedPerContainerElem - 1);
auto numElements =
- llvm::divideCeil(maxintraDataOffset + origElements, scale);
+ llvm::divideCeil(maxintraDataOffset + origElements, emulatedPerContainerElem);
Value result =
emulatedVectorLoad(rewriter, loc, adaptor.getBase(), linearizedIndices,
numElements, emulatedElemTy, containerElemTy);
@@ -870,7 +870,7 @@ struct ConvertVectorMaskedLoad final
op, "impossible to pack emulated elements into container elements "
"(bit-wise misalignment)");
}
- int scale = containerBits / emulatedBits;
+ int emulatedPerContainerElem = containerBits / emulatedBits;
// Adjust the number of elements to load when emulating narrow types,
// and then cast back to the original type with vector.bitcast op.
@@ -916,7 +916,7 @@ struct ConvertVectorMaskedLoad final
// subvector at the proper offset after bit-casting.
auto origType = op.getVectorType();
auto origElements = origType.getNumElements();
- bool isAlignedEmulation = origElements % scale == 0;
+ bool isAlignedEmulation = origElements % emulatedPerContainerElem == 0;
auto stridedMetadata =
rewriter.create<memref::ExtractStridedMetadataOp>(loc, op.getBase());
@@ -935,18 +935,18 @@ struct ConvertVectorMaskedLoad final
? 0
: getConstantIntValue(linearizedInfo.intraDataOffset);
- int64_t maxIntraDataOffset = foldedIntraVectorOffset.value_or(scale - 1);
+ int64_t maxIntraDataOffset = foldedIntraVectorOffset.value_or(emulatedPerContainerElem - 1);
FailureOr<Operation *> newMask = getCompressedMaskOp(
- rewriter, loc, op.getMask(), origElements, scale, maxIntraDataOffset);
+ rewriter, loc, op.getMask(), origElements, emulatedPerContainerElem, maxIntraDataOffset);
if (failed(newMask))
return failure();
Value passthru = op.getPassThru();
auto numElements =
- llvm::divideCeil(maxIntraDataOffset + origElements, scale);
+ llvm::divideCeil(maxIntraDataOffset + origElements, emulatedPerContainerElem);
auto loadType = VectorType::get(numElements, containerElemTy);
- auto newBitcastType = VectorType::get(numElements * scale, emulatedElemTy);
+ auto newBitcastType = VectorType::get(numElements * emulatedPerContainerElem, emulatedElemTy);
auto emptyVector = rewriter.create<arith::ConstantOp>(
loc, newBitcastType, rewriter.getZeroAttr(newBitcastType));
@@ -974,7 +974,7 @@ struct ConvertVectorMaskedLoad final
Value mask = op.getMask();
auto newSelectMaskType =
- VectorType::get(numElements * scale, rewriter.getI1Type());
+ VectorType::get(numElements * emulatedPerContainerElem, rewriter.getI1Type());
// TODO: try to fold if op's mask is constant
auto emptyMask = rewriter.create<arith::ConstantOp>(
loc, newSelectMaskType, rewriter.getZeroAttr(newSelectMaskType));
@@ -1033,11 +1033,11 @@ struct ConvertVectorTransferRead final
op, "impossible to pack emulated elements into container elements "
"(bit-wise misalignment)");
}
- int scale = containerBits / emulatedBits;
+ int emulatedPerContainerElem = containerBits / emulatedBits;
auto origElements = op.getVectorType().getNumElements();
- bool isAlignedEmulation = origElements % scale == 0;
+ bool isAlignedEmulation = origElements % emulatedPerContainerElem == 0;
auto newPadding = rewriter.create<arith::ExtUIOp>(loc, containerElemTy,
adaptor.getPadding());
@@ -1060,9 +1060,9 @@ struct ConvertVectorTransferRead final
? 0
: getConstantIntValue(linearizedInfo.intraDataOffset);
- int64_t maxIntraDataOffset = foldedIntraVectorOffset.value_or(scale - 1);
+ int64_t maxIntraDataOffset = foldedIntraVectorOffset.value_or(emulatedPerContainerElem - 1);
auto numElements =
- llvm::divideCeil(maxIntraDataOffset + origElements, scale);
+ llvm::divideCeil(maxIntraDataOffset + origElements, emulatedPerContainerElem);
auto newRead = rewriter.create<vector::TransferReadOp>(
loc, VectorType::get(numElements, containerElemTy), adaptor.getSource(),
@@ -1070,7 +1070,7 @@ struct ConvertVectorTransferRead final
newPadding);
auto bitCast = rewriter.create<vector::BitCastOp>(
- loc, VectorType::get(numElements * scale, emulatedElemTy), newRead);
+ loc, VectorType::get(numElements * emulatedPerContainerElem, emulatedElemTy), newRead);
Value result = bitCast->getResult(0);
if (!foldedIntraVectorOffset) {
More information about the Mlir-commits
mailing list