[AMDGPU] fold memref.subview/expand_shape/collapse_shape into amdgpu.gather_to_lds
#149851
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
Co-authored-by: Copilot <[email protected]>
Co-authored-by: Copilot <[email protected]>
There was a problem hiding this comment.
Pull Request Overview
This PR implements a new optimization pass to fold memref.subview operations into amdgpu.gather_to_lds operations, which can simplify IR and improve performance by eliminating intermediate subview operations.
- Adds
AmdgpuFoldSubviewOpsPasswith patternFoldSubviewIntoGatherToLDSOpthat identifies and folds subview sources - Implements index resolution using affine maps to adjust indices when folding subviews with offsets
- Adds comprehensive test coverage for both zero-offset and non-zero offset subview folding scenarios
Reviewed Changes
Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.
Show a summary per file
| File | Description |
|---|---|
| mlir/lib/Dialect/AMDGPU/Transforms/FoldSubviewOps.cpp | Core implementation of the folding pass and pattern |
| mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td | Pass definition and documentation |
| mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h | Pass declarations and pattern population function |
| mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt | Build system integration for new source file |
| mlir/test/Dialect/AMDGPU/amdgpu-fold-subviews.mlir | Test cases validating the folding optimization |
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
|
@llvm/pr-subscribers-mlir-memref @llvm/pr-subscribers-mlir-gpu Author: Alan Li (lialan) ChangesThis PR adds a new optimization pass to fold memref.subview operations into amdgpu.gather_to_lds operations, simplifying the overall operation and potentially improving performance. The pass identifies when a GatherToLDSOp has a memref.subview as its source and attempts to fold the subview by adjusting the indices accordingly.
Full diff: https://github.com/llvm/llvm-project/pull/149851.diff 5 Files Affected:
diff --git a/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h b/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h
index cc2f543e79f69..a61903609aaff 100644
--- a/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h
+++ b/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h
@@ -22,8 +22,9 @@ class ConversionTarget;
namespace amdgpu {
#define GEN_PASS_DECL_AMDGPUEMULATEATOMICSPASS
-#define GEN_PASS_DECL_AMDGPURESOLVESTRIDEDMETADATAPASS
+#define GEN_PASS_DECL_AMDGPUFOLDSUBVIEWOPSPASS
#define GEN_PASS_DECL_AMDGPUMASKEDLOADTOLOADPASS
+#define GEN_PASS_DECL_AMDGPURESOLVESTRIDEDMETADATAPASS
#define GEN_PASS_REGISTRATION
#include "mlir/Dialect/AMDGPU/Transforms/Passes.h.inc"
@@ -38,6 +39,9 @@ void populateAmdgpuResolveStridedMetadataPatterns(RewritePatternSet &patterns,
void populateAmdgpuMaskedloadToLoadPatterns(RewritePatternSet &patterns,
PatternBenefit benefit = 1);
+void populateAmdgpuFoldSubviewOpsPatterns(RewritePatternSet &patterns,
+ PatternBenefit benefit = 1);
+
} // namespace amdgpu
} // namespace mlir
diff --git a/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td b/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td
index 8d0e6829ab0cc..fad939ced9877 100644
--- a/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td
+++ b/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td
@@ -70,4 +70,16 @@ def AmdgpuMaskedloadToLoadPass : Pass<"amdgpu-maskedload-to-load"> {
"memref::MemRefDialect"
];
}
+
+def AmdgpuFoldSubviewOpsPass : Pass<"amdgpu-fold-subview-ops"> {
+ let summary = "Fold subview operations into their parent operations";
+ let description = [{
+ This pass identifies `memref.subview` sources of `GatherToLDSOp` and
+ attempts to fold the source ops, potentially simplifying the overall
+ operation and improving performance.
+ }];
+ let dependentDialects = [
+ "memref::MemRefDialect"
+ ];
+}
#endif // MLIR_DIALECT_AMDGPU_TRANSFORMS_PASSES_TD_
diff --git a/mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt b/mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt
index 17bbe54ea6c0c..20621ec0d55a4 100644
--- a/mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt
+++ b/mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt
@@ -1,7 +1,8 @@
add_mlir_dialect_library(MLIRAMDGPUTransforms
EmulateAtomics.cpp
- ResolveStridedMetadata.cpp
+ FoldSubviewOps.cpp
MaskedloadToLoad.cpp
+ ResolveStridedMetadata.cpp
ADDITIONAL_HEADER_DIRS
{$MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/AMDGPU/Transforms
diff --git a/mlir/lib/Dialect/AMDGPU/Transforms/FoldSubviewOps.cpp b/mlir/lib/Dialect/AMDGPU/Transforms/FoldSubviewOps.cpp
new file mode 100644
index 0000000000000..adbdf4b856bd5
--- /dev/null
+++ b/mlir/lib/Dialect/AMDGPU/Transforms/FoldSubviewOps.cpp
@@ -0,0 +1,67 @@
+//===- FoldSubviewOps.cpp - AMDGPU fold subview ops ---------------------===//
+//
+// 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/AMDGPU/Transforms/Passes.h"
+
+#include "mlir/Dialect/AMDGPU/IR/AMDGPUDialect.h"
+#include "mlir/Dialect/Affine/ViewLikeInterfaceUtils.h"
+#include "mlir/Dialect/MemRef/IR/MemRef.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+
+namespace mlir::amdgpu {
+#define GEN_PASS_DEF_AMDGPUFOLDSUBVIEWOPSPASS
+#include "mlir/Dialect/AMDGPU/Transforms/Passes.h.inc"
+} // namespace mlir::amdgpu
+
+using namespace mlir;
+using namespace mlir::amdgpu;
+
+namespace {
+struct AmdgpuFoldSubviewOpsPass
+ : public amdgpu::impl::AmdgpuFoldSubviewOpsPassBase<
+ AmdgpuFoldSubviewOpsPass> {
+ void runOnOperation() override {
+ RewritePatternSet patterns(&getContext());
+ populateAmdgpuFoldSubviewOpsPatterns(patterns);
+ if (failed(applyPatternsGreedily(getOperation(), std::move(patterns))))
+ signalPassFailure();
+ }
+};
+
+struct FoldSubviewIntoGatherToLDSOp : public OpRewritePattern<GatherToLDSOp> {
+ using OpRewritePattern<GatherToLDSOp>::OpRewritePattern;
+ LogicalResult matchAndRewrite(GatherToLDSOp op,
+ PatternRewriter &rewriter) const override {
+ Location loc = op.getLoc();
+
+ // Check if the source is a subview operation:
+ auto subviewOp = dyn_cast<memref::SubViewOp>(op.getSrc().getDefiningOp());
+ if (!subviewOp)
+ return rewriter.notifyMatchFailure(
+ loc, "GatherToLDSOp folding is currently supported only when the "
+ "source is a SubviewOp. This is one specific pattern, and other "
+ "scenarios may be added in the future.");
+
+ SmallVector<Value> sourceIndices;
+ mlir::affine::resolveIndicesIntoOpWithOffsetsAndStrides(
+ rewriter, loc, subviewOp.getMixedOffsets(), subviewOp.getMixedStrides(),
+ subviewOp.getDroppedDims(), op.getSrcIndices(), sourceIndices);
+
+ rewriter.replaceOpWithNewOp<GatherToLDSOp>(
+ op, subviewOp.getSource(), sourceIndices, op.getDst(),
+ op.getDstIndices(), op.getTransferType());
+
+ return success();
+ }
+};
+} // namespace
+
+void mlir::amdgpu::populateAmdgpuFoldSubviewOpsPatterns(
+ RewritePatternSet &patterns, PatternBenefit benefit) {
+ patterns.add<FoldSubviewIntoGatherToLDSOp>(patterns.getContext(), benefit);
+}
diff --git a/mlir/test/Dialect/AMDGPU/amdgpu-fold-subviews.mlir b/mlir/test/Dialect/AMDGPU/amdgpu-fold-subviews.mlir
new file mode 100644
index 0000000000000..d582991c3622f
--- /dev/null
+++ b/mlir/test/Dialect/AMDGPU/amdgpu-fold-subviews.mlir
@@ -0,0 +1,50 @@
+// RUN: mlir-opt -amdgpu-fold-subview-ops -split-input-file %s | FileCheck %s
+
+#gpu_lds_addrspace = 3
+
+// CHECK: func @test_memref
+// CHECK-SAME: %[[ARG0:.*]]: index, %[[ARG1:.*]]: index
+func.func @test_memref(%offset_i: index, %offset_j: index) {
+ // CHECK: %[[C0:.*]] = arith.constant 0 : index
+ // CHECK: %[[LOCAL:.*]] = memref.alloc() : memref<64x64xf16, 3>
+ // CHECK: %[[MEM:.*]] = memref.alloc() : memref<64x128xf16>
+ // CHECK: %[[MEM]][%arg0, %arg1], %[[LOCAL]][%[[C0]], %[[C0]]]
+ // CHECK-SAME: vector<8xf16>, memref<64x128xf16>, memref<64x64xf16, 3>
+
+ %alloc = memref.alloc() : memref<64x64xf16, #gpu_lds_addrspace>
+ %mem = memref.alloc() : memref<64x128xf16>
+ %subview = memref.subview %mem[0, 0][32, 64][1, 1] : memref<64x128xf16> to memref<32x64xf16, strided<[128, 1]>>
+ %c0 = arith.constant 0 : index
+ amdgpu.gather_to_lds %subview[%offset_i, %offset_j], %alloc[%c0, %c0]
+ : vector<8xf16>, memref<32x64xf16, strided<[128, 1]>>, memref<64x64xf16, #gpu_lds_addrspace>
+ func.return
+}
+
+// -----
+
+#gpu_lds_addrspace = 3
+
+// CHECK: #[[MAP:.*]] = affine_map<()[s0] -> (s0 + 32)>
+// CHECK: #[[MAP1:.*]] = affine_map<()[s0] -> (s0 + 64)>
+
+// CHECK: func @subview_folding_offset
+// CHECK-SAME: %[[ARG0:.*]]: index, %[[ARG1:.*]]: index
+func.func @subview_folding_offset(%offset_i: index, %offset_j: index) {
+ // CHECK: %[[C0:.*]] = arith.constant 0 : index
+ // CHECK: %[[LOCAL:.*]] = memref.alloc() : memref<64x64xf16, 3>
+ // CHECK: %[[MEM:.*]] = memref.alloc() : memref<64x128xf16>
+
+ // CHECK: %[[IDX0:.*]] = affine.apply #[[MAP]]()[%[[ARG0]]]
+ // CHECK: %[[IDX1:.*]] = affine.apply #[[MAP1]]()[%[[ARG1]]]
+
+ // CHECK: %[[MEM]][%[[IDX0]], %[[IDX1]]], %[[LOCAL]][%[[C0]], %[[C0]]]
+ // CHECK-SAME: vector<8xf16>, memref<64x128xf16>, memref<64x64xf16, 3>
+
+ %alloc = memref.alloc() : memref<64x64xf16, #gpu_lds_addrspace>
+ %mem = memref.alloc() : memref<64x128xf16>
+ %subview = memref.subview %mem[32, 64][32, 64][1, 1] : memref<64x128xf16> to memref<32x64xf16, strided<[128, 1], offset: 4160>>
+ %c0 = arith.constant 0 : index
+ amdgpu.gather_to_lds %subview[%offset_i, %offset_j], %alloc[%c0, %c0]
+ : vector<8xf16>, memref<32x64xf16, strided<[128, 1], offset: 4160>>, memref<64x64xf16, #gpu_lds_addrspace>
+ func.return
+}
|
|
@llvm/pr-subscribers-backend-amdgpu Author: Alan Li (lialan) ChangesThis PR adds a new optimization pass to fold memref.subview operations into amdgpu.gather_to_lds operations, simplifying the overall operation and potentially improving performance. The pass identifies when a GatherToLDSOp has a memref.subview as its source and attempts to fold the subview by adjusting the indices accordingly.
Full diff: https://github.com/llvm/llvm-project/pull/149851.diff 5 Files Affected:
diff --git a/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h b/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h
index cc2f543e79f69..a61903609aaff 100644
--- a/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h
+++ b/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h
@@ -22,8 +22,9 @@ class ConversionTarget;
namespace amdgpu {
#define GEN_PASS_DECL_AMDGPUEMULATEATOMICSPASS
-#define GEN_PASS_DECL_AMDGPURESOLVESTRIDEDMETADATAPASS
+#define GEN_PASS_DECL_AMDGPUFOLDSUBVIEWOPSPASS
#define GEN_PASS_DECL_AMDGPUMASKEDLOADTOLOADPASS
+#define GEN_PASS_DECL_AMDGPURESOLVESTRIDEDMETADATAPASS
#define GEN_PASS_REGISTRATION
#include "mlir/Dialect/AMDGPU/Transforms/Passes.h.inc"
@@ -38,6 +39,9 @@ void populateAmdgpuResolveStridedMetadataPatterns(RewritePatternSet &patterns,
void populateAmdgpuMaskedloadToLoadPatterns(RewritePatternSet &patterns,
PatternBenefit benefit = 1);
+void populateAmdgpuFoldSubviewOpsPatterns(RewritePatternSet &patterns,
+ PatternBenefit benefit = 1);
+
} // namespace amdgpu
} // namespace mlir
diff --git a/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td b/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td
index 8d0e6829ab0cc..fad939ced9877 100644
--- a/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td
+++ b/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td
@@ -70,4 +70,16 @@ def AmdgpuMaskedloadToLoadPass : Pass<"amdgpu-maskedload-to-load"> {
"memref::MemRefDialect"
];
}
+
+def AmdgpuFoldSubviewOpsPass : Pass<"amdgpu-fold-subview-ops"> {
+ let summary = "Fold subview operations into their parent operations";
+ let description = [{
+ This pass identifies `memref.subview` sources of `GatherToLDSOp` and
+ attempts to fold the source ops, potentially simplifying the overall
+ operation and improving performance.
+ }];
+ let dependentDialects = [
+ "memref::MemRefDialect"
+ ];
+}
#endif // MLIR_DIALECT_AMDGPU_TRANSFORMS_PASSES_TD_
diff --git a/mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt b/mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt
index 17bbe54ea6c0c..20621ec0d55a4 100644
--- a/mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt
+++ b/mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt
@@ -1,7 +1,8 @@
add_mlir_dialect_library(MLIRAMDGPUTransforms
EmulateAtomics.cpp
- ResolveStridedMetadata.cpp
+ FoldSubviewOps.cpp
MaskedloadToLoad.cpp
+ ResolveStridedMetadata.cpp
ADDITIONAL_HEADER_DIRS
{$MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/AMDGPU/Transforms
diff --git a/mlir/lib/Dialect/AMDGPU/Transforms/FoldSubviewOps.cpp b/mlir/lib/Dialect/AMDGPU/Transforms/FoldSubviewOps.cpp
new file mode 100644
index 0000000000000..adbdf4b856bd5
--- /dev/null
+++ b/mlir/lib/Dialect/AMDGPU/Transforms/FoldSubviewOps.cpp
@@ -0,0 +1,67 @@
+//===- FoldSubviewOps.cpp - AMDGPU fold subview ops ---------------------===//
+//
+// 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/AMDGPU/Transforms/Passes.h"
+
+#include "mlir/Dialect/AMDGPU/IR/AMDGPUDialect.h"
+#include "mlir/Dialect/Affine/ViewLikeInterfaceUtils.h"
+#include "mlir/Dialect/MemRef/IR/MemRef.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+
+namespace mlir::amdgpu {
+#define GEN_PASS_DEF_AMDGPUFOLDSUBVIEWOPSPASS
+#include "mlir/Dialect/AMDGPU/Transforms/Passes.h.inc"
+} // namespace mlir::amdgpu
+
+using namespace mlir;
+using namespace mlir::amdgpu;
+
+namespace {
+struct AmdgpuFoldSubviewOpsPass
+ : public amdgpu::impl::AmdgpuFoldSubviewOpsPassBase<
+ AmdgpuFoldSubviewOpsPass> {
+ void runOnOperation() override {
+ RewritePatternSet patterns(&getContext());
+ populateAmdgpuFoldSubviewOpsPatterns(patterns);
+ if (failed(applyPatternsGreedily(getOperation(), std::move(patterns))))
+ signalPassFailure();
+ }
+};
+
+struct FoldSubviewIntoGatherToLDSOp : public OpRewritePattern<GatherToLDSOp> {
+ using OpRewritePattern<GatherToLDSOp>::OpRewritePattern;
+ LogicalResult matchAndRewrite(GatherToLDSOp op,
+ PatternRewriter &rewriter) const override {
+ Location loc = op.getLoc();
+
+ // Check if the source is a subview operation:
+ auto subviewOp = dyn_cast<memref::SubViewOp>(op.getSrc().getDefiningOp());
+ if (!subviewOp)
+ return rewriter.notifyMatchFailure(
+ loc, "GatherToLDSOp folding is currently supported only when the "
+ "source is a SubviewOp. This is one specific pattern, and other "
+ "scenarios may be added in the future.");
+
+ SmallVector<Value> sourceIndices;
+ mlir::affine::resolveIndicesIntoOpWithOffsetsAndStrides(
+ rewriter, loc, subviewOp.getMixedOffsets(), subviewOp.getMixedStrides(),
+ subviewOp.getDroppedDims(), op.getSrcIndices(), sourceIndices);
+
+ rewriter.replaceOpWithNewOp<GatherToLDSOp>(
+ op, subviewOp.getSource(), sourceIndices, op.getDst(),
+ op.getDstIndices(), op.getTransferType());
+
+ return success();
+ }
+};
+} // namespace
+
+void mlir::amdgpu::populateAmdgpuFoldSubviewOpsPatterns(
+ RewritePatternSet &patterns, PatternBenefit benefit) {
+ patterns.add<FoldSubviewIntoGatherToLDSOp>(patterns.getContext(), benefit);
+}
diff --git a/mlir/test/Dialect/AMDGPU/amdgpu-fold-subviews.mlir b/mlir/test/Dialect/AMDGPU/amdgpu-fold-subviews.mlir
new file mode 100644
index 0000000000000..d582991c3622f
--- /dev/null
+++ b/mlir/test/Dialect/AMDGPU/amdgpu-fold-subviews.mlir
@@ -0,0 +1,50 @@
+// RUN: mlir-opt -amdgpu-fold-subview-ops -split-input-file %s | FileCheck %s
+
+#gpu_lds_addrspace = 3
+
+// CHECK: func @test_memref
+// CHECK-SAME: %[[ARG0:.*]]: index, %[[ARG1:.*]]: index
+func.func @test_memref(%offset_i: index, %offset_j: index) {
+ // CHECK: %[[C0:.*]] = arith.constant 0 : index
+ // CHECK: %[[LOCAL:.*]] = memref.alloc() : memref<64x64xf16, 3>
+ // CHECK: %[[MEM:.*]] = memref.alloc() : memref<64x128xf16>
+ // CHECK: %[[MEM]][%arg0, %arg1], %[[LOCAL]][%[[C0]], %[[C0]]]
+ // CHECK-SAME: vector<8xf16>, memref<64x128xf16>, memref<64x64xf16, 3>
+
+ %alloc = memref.alloc() : memref<64x64xf16, #gpu_lds_addrspace>
+ %mem = memref.alloc() : memref<64x128xf16>
+ %subview = memref.subview %mem[0, 0][32, 64][1, 1] : memref<64x128xf16> to memref<32x64xf16, strided<[128, 1]>>
+ %c0 = arith.constant 0 : index
+ amdgpu.gather_to_lds %subview[%offset_i, %offset_j], %alloc[%c0, %c0]
+ : vector<8xf16>, memref<32x64xf16, strided<[128, 1]>>, memref<64x64xf16, #gpu_lds_addrspace>
+ func.return
+}
+
+// -----
+
+#gpu_lds_addrspace = 3
+
+// CHECK: #[[MAP:.*]] = affine_map<()[s0] -> (s0 + 32)>
+// CHECK: #[[MAP1:.*]] = affine_map<()[s0] -> (s0 + 64)>
+
+// CHECK: func @subview_folding_offset
+// CHECK-SAME: %[[ARG0:.*]]: index, %[[ARG1:.*]]: index
+func.func @subview_folding_offset(%offset_i: index, %offset_j: index) {
+ // CHECK: %[[C0:.*]] = arith.constant 0 : index
+ // CHECK: %[[LOCAL:.*]] = memref.alloc() : memref<64x64xf16, 3>
+ // CHECK: %[[MEM:.*]] = memref.alloc() : memref<64x128xf16>
+
+ // CHECK: %[[IDX0:.*]] = affine.apply #[[MAP]]()[%[[ARG0]]]
+ // CHECK: %[[IDX1:.*]] = affine.apply #[[MAP1]]()[%[[ARG1]]]
+
+ // CHECK: %[[MEM]][%[[IDX0]], %[[IDX1]]], %[[LOCAL]][%[[C0]], %[[C0]]]
+ // CHECK-SAME: vector<8xf16>, memref<64x128xf16>, memref<64x64xf16, 3>
+
+ %alloc = memref.alloc() : memref<64x64xf16, #gpu_lds_addrspace>
+ %mem = memref.alloc() : memref<64x128xf16>
+ %subview = memref.subview %mem[32, 64][32, 64][1, 1] : memref<64x128xf16> to memref<32x64xf16, strided<[128, 1], offset: 4160>>
+ %c0 = arith.constant 0 : index
+ amdgpu.gather_to_lds %subview[%offset_i, %offset_j], %alloc[%c0, %c0]
+ : vector<8xf16>, memref<32x64xf16, strided<[128, 1], offset: 4160>>, memref<64x64xf16, #gpu_lds_addrspace>
+ func.return
+}
|
|
@llvm/pr-subscribers-mlir Author: Alan Li (lialan) ChangesThis PR adds a new optimization pass to fold memref.subview operations into amdgpu.gather_to_lds operations, simplifying the overall operation and potentially improving performance. The pass identifies when a GatherToLDSOp has a memref.subview as its source and attempts to fold the subview by adjusting the indices accordingly.
Full diff: https://github.com/llvm/llvm-project/pull/149851.diff 5 Files Affected:
diff --git a/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h b/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h
index cc2f543e79f69..a61903609aaff 100644
--- a/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h
+++ b/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h
@@ -22,8 +22,9 @@ class ConversionTarget;
namespace amdgpu {
#define GEN_PASS_DECL_AMDGPUEMULATEATOMICSPASS
-#define GEN_PASS_DECL_AMDGPURESOLVESTRIDEDMETADATAPASS
+#define GEN_PASS_DECL_AMDGPUFOLDSUBVIEWOPSPASS
#define GEN_PASS_DECL_AMDGPUMASKEDLOADTOLOADPASS
+#define GEN_PASS_DECL_AMDGPURESOLVESTRIDEDMETADATAPASS
#define GEN_PASS_REGISTRATION
#include "mlir/Dialect/AMDGPU/Transforms/Passes.h.inc"
@@ -38,6 +39,9 @@ void populateAmdgpuResolveStridedMetadataPatterns(RewritePatternSet &patterns,
void populateAmdgpuMaskedloadToLoadPatterns(RewritePatternSet &patterns,
PatternBenefit benefit = 1);
+void populateAmdgpuFoldSubviewOpsPatterns(RewritePatternSet &patterns,
+ PatternBenefit benefit = 1);
+
} // namespace amdgpu
} // namespace mlir
diff --git a/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td b/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td
index 8d0e6829ab0cc..fad939ced9877 100644
--- a/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td
+++ b/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td
@@ -70,4 +70,16 @@ def AmdgpuMaskedloadToLoadPass : Pass<"amdgpu-maskedload-to-load"> {
"memref::MemRefDialect"
];
}
+
+def AmdgpuFoldSubviewOpsPass : Pass<"amdgpu-fold-subview-ops"> {
+ let summary = "Fold subview operations into their parent operations";
+ let description = [{
+ This pass identifies `memref.subview` sources of `GatherToLDSOp` and
+ attempts to fold the source ops, potentially simplifying the overall
+ operation and improving performance.
+ }];
+ let dependentDialects = [
+ "memref::MemRefDialect"
+ ];
+}
#endif // MLIR_DIALECT_AMDGPU_TRANSFORMS_PASSES_TD_
diff --git a/mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt b/mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt
index 17bbe54ea6c0c..20621ec0d55a4 100644
--- a/mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt
+++ b/mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt
@@ -1,7 +1,8 @@
add_mlir_dialect_library(MLIRAMDGPUTransforms
EmulateAtomics.cpp
- ResolveStridedMetadata.cpp
+ FoldSubviewOps.cpp
MaskedloadToLoad.cpp
+ ResolveStridedMetadata.cpp
ADDITIONAL_HEADER_DIRS
{$MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/AMDGPU/Transforms
diff --git a/mlir/lib/Dialect/AMDGPU/Transforms/FoldSubviewOps.cpp b/mlir/lib/Dialect/AMDGPU/Transforms/FoldSubviewOps.cpp
new file mode 100644
index 0000000000000..adbdf4b856bd5
--- /dev/null
+++ b/mlir/lib/Dialect/AMDGPU/Transforms/FoldSubviewOps.cpp
@@ -0,0 +1,67 @@
+//===- FoldSubviewOps.cpp - AMDGPU fold subview ops ---------------------===//
+//
+// 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/AMDGPU/Transforms/Passes.h"
+
+#include "mlir/Dialect/AMDGPU/IR/AMDGPUDialect.h"
+#include "mlir/Dialect/Affine/ViewLikeInterfaceUtils.h"
+#include "mlir/Dialect/MemRef/IR/MemRef.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+
+namespace mlir::amdgpu {
+#define GEN_PASS_DEF_AMDGPUFOLDSUBVIEWOPSPASS
+#include "mlir/Dialect/AMDGPU/Transforms/Passes.h.inc"
+} // namespace mlir::amdgpu
+
+using namespace mlir;
+using namespace mlir::amdgpu;
+
+namespace {
+struct AmdgpuFoldSubviewOpsPass
+ : public amdgpu::impl::AmdgpuFoldSubviewOpsPassBase<
+ AmdgpuFoldSubviewOpsPass> {
+ void runOnOperation() override {
+ RewritePatternSet patterns(&getContext());
+ populateAmdgpuFoldSubviewOpsPatterns(patterns);
+ if (failed(applyPatternsGreedily(getOperation(), std::move(patterns))))
+ signalPassFailure();
+ }
+};
+
+struct FoldSubviewIntoGatherToLDSOp : public OpRewritePattern<GatherToLDSOp> {
+ using OpRewritePattern<GatherToLDSOp>::OpRewritePattern;
+ LogicalResult matchAndRewrite(GatherToLDSOp op,
+ PatternRewriter &rewriter) const override {
+ Location loc = op.getLoc();
+
+ // Check if the source is a subview operation:
+ auto subviewOp = dyn_cast<memref::SubViewOp>(op.getSrc().getDefiningOp());
+ if (!subviewOp)
+ return rewriter.notifyMatchFailure(
+ loc, "GatherToLDSOp folding is currently supported only when the "
+ "source is a SubviewOp. This is one specific pattern, and other "
+ "scenarios may be added in the future.");
+
+ SmallVector<Value> sourceIndices;
+ mlir::affine::resolveIndicesIntoOpWithOffsetsAndStrides(
+ rewriter, loc, subviewOp.getMixedOffsets(), subviewOp.getMixedStrides(),
+ subviewOp.getDroppedDims(), op.getSrcIndices(), sourceIndices);
+
+ rewriter.replaceOpWithNewOp<GatherToLDSOp>(
+ op, subviewOp.getSource(), sourceIndices, op.getDst(),
+ op.getDstIndices(), op.getTransferType());
+
+ return success();
+ }
+};
+} // namespace
+
+void mlir::amdgpu::populateAmdgpuFoldSubviewOpsPatterns(
+ RewritePatternSet &patterns, PatternBenefit benefit) {
+ patterns.add<FoldSubviewIntoGatherToLDSOp>(patterns.getContext(), benefit);
+}
diff --git a/mlir/test/Dialect/AMDGPU/amdgpu-fold-subviews.mlir b/mlir/test/Dialect/AMDGPU/amdgpu-fold-subviews.mlir
new file mode 100644
index 0000000000000..d582991c3622f
--- /dev/null
+++ b/mlir/test/Dialect/AMDGPU/amdgpu-fold-subviews.mlir
@@ -0,0 +1,50 @@
+// RUN: mlir-opt -amdgpu-fold-subview-ops -split-input-file %s | FileCheck %s
+
+#gpu_lds_addrspace = 3
+
+// CHECK: func @test_memref
+// CHECK-SAME: %[[ARG0:.*]]: index, %[[ARG1:.*]]: index
+func.func @test_memref(%offset_i: index, %offset_j: index) {
+ // CHECK: %[[C0:.*]] = arith.constant 0 : index
+ // CHECK: %[[LOCAL:.*]] = memref.alloc() : memref<64x64xf16, 3>
+ // CHECK: %[[MEM:.*]] = memref.alloc() : memref<64x128xf16>
+ // CHECK: %[[MEM]][%arg0, %arg1], %[[LOCAL]][%[[C0]], %[[C0]]]
+ // CHECK-SAME: vector<8xf16>, memref<64x128xf16>, memref<64x64xf16, 3>
+
+ %alloc = memref.alloc() : memref<64x64xf16, #gpu_lds_addrspace>
+ %mem = memref.alloc() : memref<64x128xf16>
+ %subview = memref.subview %mem[0, 0][32, 64][1, 1] : memref<64x128xf16> to memref<32x64xf16, strided<[128, 1]>>
+ %c0 = arith.constant 0 : index
+ amdgpu.gather_to_lds %subview[%offset_i, %offset_j], %alloc[%c0, %c0]
+ : vector<8xf16>, memref<32x64xf16, strided<[128, 1]>>, memref<64x64xf16, #gpu_lds_addrspace>
+ func.return
+}
+
+// -----
+
+#gpu_lds_addrspace = 3
+
+// CHECK: #[[MAP:.*]] = affine_map<()[s0] -> (s0 + 32)>
+// CHECK: #[[MAP1:.*]] = affine_map<()[s0] -> (s0 + 64)>
+
+// CHECK: func @subview_folding_offset
+// CHECK-SAME: %[[ARG0:.*]]: index, %[[ARG1:.*]]: index
+func.func @subview_folding_offset(%offset_i: index, %offset_j: index) {
+ // CHECK: %[[C0:.*]] = arith.constant 0 : index
+ // CHECK: %[[LOCAL:.*]] = memref.alloc() : memref<64x64xf16, 3>
+ // CHECK: %[[MEM:.*]] = memref.alloc() : memref<64x128xf16>
+
+ // CHECK: %[[IDX0:.*]] = affine.apply #[[MAP]]()[%[[ARG0]]]
+ // CHECK: %[[IDX1:.*]] = affine.apply #[[MAP1]]()[%[[ARG1]]]
+
+ // CHECK: %[[MEM]][%[[IDX0]], %[[IDX1]]], %[[LOCAL]][%[[C0]], %[[C0]]]
+ // CHECK-SAME: vector<8xf16>, memref<64x128xf16>, memref<64x64xf16, 3>
+
+ %alloc = memref.alloc() : memref<64x64xf16, #gpu_lds_addrspace>
+ %mem = memref.alloc() : memref<64x128xf16>
+ %subview = memref.subview %mem[32, 64][32, 64][1, 1] : memref<64x128xf16> to memref<32x64xf16, strided<[128, 1], offset: 4160>>
+ %c0 = arith.constant 0 : index
+ amdgpu.gather_to_lds %subview[%offset_i, %offset_j], %alloc[%c0, %c0]
+ : vector<8xf16>, memref<32x64xf16, strided<[128, 1], offset: 4160>>, memref<64x64xf16, #gpu_lds_addrspace>
+ func.return
+}
|
|
@llvm/pr-subscribers-mlir-amdgpu Author: Alan Li (lialan) ChangesThis PR adds a new optimization pass to fold memref.subview operations into amdgpu.gather_to_lds operations, simplifying the overall operation and potentially improving performance. The pass identifies when a GatherToLDSOp has a memref.subview as its source and attempts to fold the subview by adjusting the indices accordingly.
Full diff: https://github.com/llvm/llvm-project/pull/149851.diff 5 Files Affected:
diff --git a/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h b/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h
index cc2f543e79f69..a61903609aaff 100644
--- a/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h
+++ b/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h
@@ -22,8 +22,9 @@ class ConversionTarget;
namespace amdgpu {
#define GEN_PASS_DECL_AMDGPUEMULATEATOMICSPASS
-#define GEN_PASS_DECL_AMDGPURESOLVESTRIDEDMETADATAPASS
+#define GEN_PASS_DECL_AMDGPUFOLDSUBVIEWOPSPASS
#define GEN_PASS_DECL_AMDGPUMASKEDLOADTOLOADPASS
+#define GEN_PASS_DECL_AMDGPURESOLVESTRIDEDMETADATAPASS
#define GEN_PASS_REGISTRATION
#include "mlir/Dialect/AMDGPU/Transforms/Passes.h.inc"
@@ -38,6 +39,9 @@ void populateAmdgpuResolveStridedMetadataPatterns(RewritePatternSet &patterns,
void populateAmdgpuMaskedloadToLoadPatterns(RewritePatternSet &patterns,
PatternBenefit benefit = 1);
+void populateAmdgpuFoldSubviewOpsPatterns(RewritePatternSet &patterns,
+ PatternBenefit benefit = 1);
+
} // namespace amdgpu
} // namespace mlir
diff --git a/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td b/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td
index 8d0e6829ab0cc..fad939ced9877 100644
--- a/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td
+++ b/mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td
@@ -70,4 +70,16 @@ def AmdgpuMaskedloadToLoadPass : Pass<"amdgpu-maskedload-to-load"> {
"memref::MemRefDialect"
];
}
+
+def AmdgpuFoldSubviewOpsPass : Pass<"amdgpu-fold-subview-ops"> {
+ let summary = "Fold subview operations into their parent operations";
+ let description = [{
+ This pass identifies `memref.subview` sources of `GatherToLDSOp` and
+ attempts to fold the source ops, potentially simplifying the overall
+ operation and improving performance.
+ }];
+ let dependentDialects = [
+ "memref::MemRefDialect"
+ ];
+}
#endif // MLIR_DIALECT_AMDGPU_TRANSFORMS_PASSES_TD_
diff --git a/mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt b/mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt
index 17bbe54ea6c0c..20621ec0d55a4 100644
--- a/mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt
+++ b/mlir/lib/Dialect/AMDGPU/Transforms/CMakeLists.txt
@@ -1,7 +1,8 @@
add_mlir_dialect_library(MLIRAMDGPUTransforms
EmulateAtomics.cpp
- ResolveStridedMetadata.cpp
+ FoldSubviewOps.cpp
MaskedloadToLoad.cpp
+ ResolveStridedMetadata.cpp
ADDITIONAL_HEADER_DIRS
{$MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/AMDGPU/Transforms
diff --git a/mlir/lib/Dialect/AMDGPU/Transforms/FoldSubviewOps.cpp b/mlir/lib/Dialect/AMDGPU/Transforms/FoldSubviewOps.cpp
new file mode 100644
index 0000000000000..adbdf4b856bd5
--- /dev/null
+++ b/mlir/lib/Dialect/AMDGPU/Transforms/FoldSubviewOps.cpp
@@ -0,0 +1,67 @@
+//===- FoldSubviewOps.cpp - AMDGPU fold subview ops ---------------------===//
+//
+// 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/AMDGPU/Transforms/Passes.h"
+
+#include "mlir/Dialect/AMDGPU/IR/AMDGPUDialect.h"
+#include "mlir/Dialect/Affine/ViewLikeInterfaceUtils.h"
+#include "mlir/Dialect/MemRef/IR/MemRef.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+
+namespace mlir::amdgpu {
+#define GEN_PASS_DEF_AMDGPUFOLDSUBVIEWOPSPASS
+#include "mlir/Dialect/AMDGPU/Transforms/Passes.h.inc"
+} // namespace mlir::amdgpu
+
+using namespace mlir;
+using namespace mlir::amdgpu;
+
+namespace {
+struct AmdgpuFoldSubviewOpsPass
+ : public amdgpu::impl::AmdgpuFoldSubviewOpsPassBase<
+ AmdgpuFoldSubviewOpsPass> {
+ void runOnOperation() override {
+ RewritePatternSet patterns(&getContext());
+ populateAmdgpuFoldSubviewOpsPatterns(patterns);
+ if (failed(applyPatternsGreedily(getOperation(), std::move(patterns))))
+ signalPassFailure();
+ }
+};
+
+struct FoldSubviewIntoGatherToLDSOp : public OpRewritePattern<GatherToLDSOp> {
+ using OpRewritePattern<GatherToLDSOp>::OpRewritePattern;
+ LogicalResult matchAndRewrite(GatherToLDSOp op,
+ PatternRewriter &rewriter) const override {
+ Location loc = op.getLoc();
+
+ // Check if the source is a subview operation:
+ auto subviewOp = dyn_cast<memref::SubViewOp>(op.getSrc().getDefiningOp());
+ if (!subviewOp)
+ return rewriter.notifyMatchFailure(
+ loc, "GatherToLDSOp folding is currently supported only when the "
+ "source is a SubviewOp. This is one specific pattern, and other "
+ "scenarios may be added in the future.");
+
+ SmallVector<Value> sourceIndices;
+ mlir::affine::resolveIndicesIntoOpWithOffsetsAndStrides(
+ rewriter, loc, subviewOp.getMixedOffsets(), subviewOp.getMixedStrides(),
+ subviewOp.getDroppedDims(), op.getSrcIndices(), sourceIndices);
+
+ rewriter.replaceOpWithNewOp<GatherToLDSOp>(
+ op, subviewOp.getSource(), sourceIndices, op.getDst(),
+ op.getDstIndices(), op.getTransferType());
+
+ return success();
+ }
+};
+} // namespace
+
+void mlir::amdgpu::populateAmdgpuFoldSubviewOpsPatterns(
+ RewritePatternSet &patterns, PatternBenefit benefit) {
+ patterns.add<FoldSubviewIntoGatherToLDSOp>(patterns.getContext(), benefit);
+}
diff --git a/mlir/test/Dialect/AMDGPU/amdgpu-fold-subviews.mlir b/mlir/test/Dialect/AMDGPU/amdgpu-fold-subviews.mlir
new file mode 100644
index 0000000000000..d582991c3622f
--- /dev/null
+++ b/mlir/test/Dialect/AMDGPU/amdgpu-fold-subviews.mlir
@@ -0,0 +1,50 @@
+// RUN: mlir-opt -amdgpu-fold-subview-ops -split-input-file %s | FileCheck %s
+
+#gpu_lds_addrspace = 3
+
+// CHECK: func @test_memref
+// CHECK-SAME: %[[ARG0:.*]]: index, %[[ARG1:.*]]: index
+func.func @test_memref(%offset_i: index, %offset_j: index) {
+ // CHECK: %[[C0:.*]] = arith.constant 0 : index
+ // CHECK: %[[LOCAL:.*]] = memref.alloc() : memref<64x64xf16, 3>
+ // CHECK: %[[MEM:.*]] = memref.alloc() : memref<64x128xf16>
+ // CHECK: %[[MEM]][%arg0, %arg1], %[[LOCAL]][%[[C0]], %[[C0]]]
+ // CHECK-SAME: vector<8xf16>, memref<64x128xf16>, memref<64x64xf16, 3>
+
+ %alloc = memref.alloc() : memref<64x64xf16, #gpu_lds_addrspace>
+ %mem = memref.alloc() : memref<64x128xf16>
+ %subview = memref.subview %mem[0, 0][32, 64][1, 1] : memref<64x128xf16> to memref<32x64xf16, strided<[128, 1]>>
+ %c0 = arith.constant 0 : index
+ amdgpu.gather_to_lds %subview[%offset_i, %offset_j], %alloc[%c0, %c0]
+ : vector<8xf16>, memref<32x64xf16, strided<[128, 1]>>, memref<64x64xf16, #gpu_lds_addrspace>
+ func.return
+}
+
+// -----
+
+#gpu_lds_addrspace = 3
+
+// CHECK: #[[MAP:.*]] = affine_map<()[s0] -> (s0 + 32)>
+// CHECK: #[[MAP1:.*]] = affine_map<()[s0] -> (s0 + 64)>
+
+// CHECK: func @subview_folding_offset
+// CHECK-SAME: %[[ARG0:.*]]: index, %[[ARG1:.*]]: index
+func.func @subview_folding_offset(%offset_i: index, %offset_j: index) {
+ // CHECK: %[[C0:.*]] = arith.constant 0 : index
+ // CHECK: %[[LOCAL:.*]] = memref.alloc() : memref<64x64xf16, 3>
+ // CHECK: %[[MEM:.*]] = memref.alloc() : memref<64x128xf16>
+
+ // CHECK: %[[IDX0:.*]] = affine.apply #[[MAP]]()[%[[ARG0]]]
+ // CHECK: %[[IDX1:.*]] = affine.apply #[[MAP1]]()[%[[ARG1]]]
+
+ // CHECK: %[[MEM]][%[[IDX0]], %[[IDX1]]], %[[LOCAL]][%[[C0]], %[[C0]]]
+ // CHECK-SAME: vector<8xf16>, memref<64x128xf16>, memref<64x64xf16, 3>
+
+ %alloc = memref.alloc() : memref<64x64xf16, #gpu_lds_addrspace>
+ %mem = memref.alloc() : memref<64x128xf16>
+ %subview = memref.subview %mem[32, 64][32, 64][1, 1] : memref<64x128xf16> to memref<32x64xf16, strided<[128, 1], offset: 4160>>
+ %c0 = arith.constant 0 : index
+ amdgpu.gather_to_lds %subview[%offset_i, %offset_j], %alloc[%c0, %c0]
+ : vector<8xf16>, memref<32x64xf16, strided<[128, 1], offset: 4160>>, memref<64x64xf16, #gpu_lds_addrspace>
+ func.return
+}
|
|
High-level comment: this doesn't need to be a new pass. Just go add logic to |
|
(That is, tentative reject for being overcomplicated) (This'll all work nicely if I ever have the time to get my interfaces for DMA-like ops and load/store-like ops up, but for now, this goes in with the rest of the subview folding logic) |
kuhar
reviewed
Jul 21, 2025
qedawkins
approved these changes
Jul 21, 2025
Ahh they have NVGPU patterns in the same file so I think it is okay to just do it there. I will move things there. |
No, we should keep it out. The NVGPU patterns should not be there either. |
lialan
changed the title
memref.subview into amdgpu.gather_to_ldsmemref.subview/expand_shape/collapse_shape into amdgpu.gather_to_lds
kuhar
reviewed
Jul 23, 2025
Comment on lines
22
to
23
| struct AmdgpuFoldMemRefOpsPass | ||
| : public amdgpu::impl::AmdgpuFoldSubviewOpsPassBase< |
There was a problem hiding this comment.
Suggested change
| struct AmdgpuFoldMemRefOpsPass | |
| : public amdgpu::impl::AmdgpuFoldSubviewOpsPassBase< | |
| struct AmdgpuFoldMemRefOpsPass final | |
| : amdgpu::impl::AmdgpuFoldSubviewOpsPassBase< |
| void runOnOperation() override { | ||
| RewritePatternSet patterns(&getContext()); | ||
| populateAmdgpuFoldSubviewOpsPatterns(patterns); | ||
| if (failed(applyPatternsGreedily(getOperation(), std::move(patterns)))) |
There was a problem hiding this comment.
Do we need the greedy rewriter or is walkAndApplyPatterns enough?
There was a problem hiding this comment.
TIL. It is a crime that walkAndApplyPatterns is not being used widely.
kuhar
approved these changes
Jul 23, 2025
updated as well. |
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/138/builds/16495 Here is the relevant piece of the build log for the reference |
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/80/builds/14755 Here is the relevant piece of the build log for the reference |
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/204/builds/16486 Here is the relevant piece of the build log for the reference |
|
@lialan can you look at the link errors? |
definitely missing including the MemRefUtils in some backend, I will take a look. |
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/203/builds/17673 Here is the relevant piece of the build log for the reference |
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/205/builds/16463 Here is the relevant piece of the build log for the reference |
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/130/builds/14426 Here is the relevant piece of the build log for the reference |
rupprecht
added a commit
to rupprecht/llvm-project
that referenced
this pull request
Jul 23, 2025
rupprecht
added a commit
to rupprecht/llvm-project
that referenced
this pull request
Jul 23, 2025
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/153/builds/39096 Here is the relevant piece of the build log for the reference |
lialan
added a commit
that referenced
this pull request
Jul 24, 2025
…nto `amdgpu.gather_to_lds`" (#150334) This is a reapply of patch #149851. The reapply also fixes a CMake/Bazel build issue, which was the reason of the revert. (Thanks @rupprecht ) Original patch (#149851) message: ----- This PR adds a new optimization pass to fold `memref.subview/expand_shape/collapse_shape` ops into consumer `amdgpu.gather_to_lds` operations. * Implements a new pass `AmdgpuFoldMemRefOpsPass` with pattern `FoldMemRefOpsIntoGatherToLDSOp` * Adds corresponding folding tests
mahesh-attarde
pushed a commit
to mahesh-attarde/llvm-project
that referenced
this pull request
Jul 28, 2025
…pu.gather_to_lds` (llvm#149851) This PR adds a new optimization pass to fold `memref.subview/expand_shape/collapse_shape` ops into consumer `amdgpu.gather_to_lds` operations. * Implements a new pass `AmdgpuFoldMemRefOpsPass` with pattern `FoldMemRefOpsIntoGatherToLDSOp` * Adds corresponding folding tests --------- Co-authored-by: Copilot <[email protected]>
mahesh-attarde
pushed a commit
to mahesh-attarde/llvm-project
that referenced
this pull request
Jul 28, 2025
llvm#150256) …to `amdgpu.gather_to_lds` (llvm#149851)" This reverts commit dbc63f1. Having build deps issue.
mahesh-attarde
pushed a commit
to mahesh-attarde/llvm-project
that referenced
this pull request
Jul 28, 2025
…nto `amdgpu.gather_to_lds`" (llvm#150334) This is a reapply of patch llvm#149851. The reapply also fixes a CMake/Bazel build issue, which was the reason of the revert. (Thanks @rupprecht ) Original patch (llvm#149851) message: ----- This PR adds a new optimization pass to fold `memref.subview/expand_shape/collapse_shape` ops into consumer `amdgpu.gather_to_lds` operations. * Implements a new pass `AmdgpuFoldMemRefOpsPass` with pattern `FoldMemRefOpsIntoGatherToLDSOp` * Adds corresponding folding tests
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR adds a new optimization pass to fold
memref.subview/expand_shape/collapse_shapeops into consumeramdgpu.gather_to_ldsoperations.AmdgpuFoldMemRefOpsPasswith patternFoldMemRefOpsIntoGatherToLDSOp