Skip to content

[mlir][amdgpu] fold memref.subview/expand_shape/collapse_shape into amdgpu.gather_to_lds for DST operand #152277

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 6 commits into from
Aug 8, 2025
Merged

[mlir][amdgpu] fold memref.subview/expand_shape/collapse_shape into amdgpu.gather_to_lds for DST operand #152277

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
d63ecf5
[AMDGPU] fold dst operand of gather to lds
sebvince Aug 4, 2025
b4908bd
Fix cast issue thanks to [email protected]
sebvince Aug 5, 2025
2af05fa
Update tests
sebvince Aug 6, 2025
2ae5b13
Fix formatting and removed unnecessary include file
sebvince Aug 6, 2025
fea2ca8
Add more tests
sebvince Aug 6, 2025
7c02f9f
Formatting on amdgpu-fold-memrefs.mlir
sebvince Aug 6, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
109 changes: 63 additions & 46 deletions mlir/lib/Dialect/AMDGPU/Transforms/FoldMemRefsOps.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,62 +28,79 @@ struct AmdgpuFoldMemRefOpsPass final
}
};

static LogicalResult foldMemrefViewOp(PatternRewriter &rewriter, Location loc,
Value view, mlir::OperandRange indices,
SmallVectorImpl<Value> &resolvedIndices,
Value &memrefBase, StringRef role) {
Operation *defOp = view.getDefiningOp();
if (!defOp) {
return failure();
}
return llvm::TypeSwitch<Operation *, LogicalResult>(defOp)
.Case<memref::SubViewOp>([&](memref::SubViewOp subviewOp) {
mlir::affine::resolveIndicesIntoOpWithOffsetsAndStrides(
rewriter, loc, subviewOp.getMixedOffsets(),
subviewOp.getMixedStrides(), subviewOp.getDroppedDims(), indices,
resolvedIndices);
memrefBase = subviewOp.getSource();
return success();
})
.Case<memref::ExpandShapeOp>([&](memref::ExpandShapeOp expandShapeOp) {
if (failed(mlir::memref::resolveSourceIndicesExpandShape(
loc, rewriter, expandShapeOp, indices, resolvedIndices,
false))) {
return failure();
}
memrefBase = expandShapeOp.getViewSource();
return success();
})
.Case<memref::CollapseShapeOp>(
[&](memref::CollapseShapeOp collapseShapeOp) {
if (failed(mlir::memref::resolveSourceIndicesCollapseShape(
loc, rewriter, collapseShapeOp, indices,
resolvedIndices))) {
return failure();
}
memrefBase = collapseShapeOp.getViewSource();
return success();
})
.Default([&](Operation *op) {
return rewriter.notifyMatchFailure(
op, (role + " producer is not one of SubViewOp, ExpandShapeOp, or "
"CollapseShapeOp")
.str());
});
}

struct FoldMemRefOpsIntoGatherToLDSOp final : OpRewritePattern<GatherToLDSOp> {
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(GatherToLDSOp op,
PatternRewriter &rewriter) const override {
Location loc = op.getLoc();

Value memrefSource;
SmallVector<Value> sourceIndices;
auto foldResult =
llvm::TypeSwitch<Operation *, LogicalResult>(
op.getSrc().getDefiningOp())
.Case<memref::SubViewOp>([&](memref::SubViewOp subviewOp) {
// If the source is a SubViewOp, we can directly rewrite the
// GatherToLDSOp.
mlir::affine::resolveIndicesIntoOpWithOffsetsAndStrides(
rewriter, loc, subviewOp.getMixedOffsets(),
subviewOp.getMixedStrides(), subviewOp.getDroppedDims(),
op.getSrcIndices(), sourceIndices);
memrefSource = subviewOp.getSource();
return success();
})
.Case<memref::ExpandShapeOp>(
[&](memref::ExpandShapeOp expandShapeOp) {
if (failed(mlir::memref::resolveSourceIndicesExpandShape(
loc, rewriter, expandShapeOp, op.getSrcIndices(),
sourceIndices, false))) {
return failure();
}
memrefSource = expandShapeOp.getViewSource();
return success();
})
.Case<memref::CollapseShapeOp>(
[&](memref::CollapseShapeOp collapseShapeOp) {
if (failed(mlir::memref::resolveSourceIndicesCollapseShape(
loc, rewriter, collapseShapeOp, op.getSrcIndices(),
sourceIndices))) {
return failure();
}
memrefSource = collapseShapeOp.getViewSource();
return success();
})
.Default([&](Operation *op) {
// If the source is not a SubViewOp, ExpandShapeOp, or
// CollapseShapeOp, we cannot fold the GatherToLDSOp.
return rewriter.notifyMatchFailure(
op,
"source producer is not one of SubViewOp, ExpandShapeOp, or "
"CollapseShapeOp");
});
SmallVector<Value> sourceIndices, destIndices;
Value memrefSource, memrefDest;

auto foldSrcResult =
foldMemrefViewOp(rewriter, loc, op.getSrc(), op.getSrcIndices(),
sourceIndices, memrefSource, "source");

if (failed(foldSrcResult)) {
memrefSource = op.getSrc();
sourceIndices = op.getSrcIndices();
}

auto foldDstResult =
foldMemrefViewOp(rewriter, loc, op.getDst(), op.getDstIndices(),
destIndices, memrefDest, "destination");

if (failed(foldResult)) {
return failure();
if (failed(foldDstResult)) {
memrefDest = op.getDst();
destIndices = op.getDstIndices();
}

rewriter.replaceOpWithNewOp<GatherToLDSOp>(op, memrefSource, sourceIndices,
op.getDst(), op.getDstIndices(),
memrefDest, destIndices,
op.getTransferType());

return success();
Expand Down
91 changes: 80 additions & 11 deletions mlir/test/Dialect/AMDGPU/amdgpu-fold-memrefs.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,18 +54,20 @@ func.func @subview_folding_offset(%offset_i: index, %offset_j: index) {
// CHECK: func @test_expand_shape
// CHECK-SAME: %[[ARG0:.*]]: index, %[[ARG1:.*]]: index
func.func @test_expand_shape(%offset_i: index, %offset_j: index) {
// CHECK: %[[LOCAL:.*]] = memref.alloc() : memref<64x64xf16, 3>
// CHECK: %[[LOCAL:.*]] = memref.alloc() : memref<4096xf16, 3>
// CHECK: %[[MEM:.*]] = memref.alloc() : memref<8192xf16>
// CHECK: %[[C0:.*]] = arith.constant 0 : index
// CHECK: %[[IDX:.*]] = affine.linearize_index [%[[ARG0]], %[[ARG1]]] by (64, 128) : index
// CHECK: amdgpu.gather_to_lds %[[MEM]][%[[IDX]]], %[[LOCAL]][%[[C0]], %[[C0]]]
// CHECK-SAME: vector<8xf16>, memref<8192xf16>, memref<64x64xf16, 3>
// CHECK: %[[IDXM:.*]] = affine.linearize_index [%[[ARG0]], %[[ARG1]]] by (64, 128) : index
// CHECK: %[[IDXL:.*]] = affine.linearize_index [%[[C0]], %[[C0]]] by (64, 64) : index
// CHECK: amdgpu.gather_to_lds %[[MEM]][%[[IDXM]]], %[[LOCAL]][%[[IDXL]]]
// CHECK-SAME: vector<8xf16>, memref<8192xf16>, memref<4096xf16, 3>

%alloc = memref.alloc() : memref<64x64xf16, #gpu_lds_addrspace>
%alloc = memref.alloc() : memref<4096xf16, #gpu_lds_addrspace>
%mem = memref.alloc() : memref<8192xf16>
%expand = memref.expand_shape %mem [[0, 1]] output_shape [64, 128] : memref<8192xf16> into memref<64x128xf16>
%expand_mem = memref.expand_shape %mem [[0, 1]] output_shape [64, 128] : memref<8192xf16> into memref<64x128xf16>
%expand_alloc = memref.expand_shape %alloc [[0, 1]] output_shape [64, 64] : memref<4096xf16, #gpu_lds_addrspace> into memref<64x64xf16, #gpu_lds_addrspace>
%c0 = arith.constant 0 : index
amdgpu.gather_to_lds %expand[%offset_i, %offset_j], %alloc[%c0, %c0]
amdgpu.gather_to_lds %expand_mem[%offset_i, %offset_j], %expand_alloc[%c0, %c0]
: vector<8xf16>, memref<64x128xf16>, memref<64x64xf16, #gpu_lds_addrspace>
func.return
}
Expand All @@ -80,15 +82,82 @@ func.func @test_collapse_shape(%offset_i: index, %offset_j: index) {
// CHECK: %[[LOCAL:.*]] = memref.alloc() : memref<64x64xf16, 3>
// CHECK: %[[MEM:.*]] = memref.alloc() : memref<64x128xf16>
// CHECK: %[[C0:.*]] = arith.constant 0 : index
// CHECK: %[[INDICES:.*]]:2 = affine.delinearize_index %[[ARG0]] into (64, 128) : index, index
// CHECK: amdgpu.gather_to_lds %[[MEM]][%[[INDICES]]#0, %[[INDICES]]#1], %[[LOCAL]][%[[C0]], %[[C0]]]
// CHECK: %[[INDICES_MEM:.*]]:2 = affine.delinearize_index %[[ARG0]] into (64, 128) : index, index
// CHECK: %[[INDICES_LDS:.*]]:2 = affine.delinearize_index %[[ARG1]] into (64, 64) : index, index
// CHECK: amdgpu.gather_to_lds %[[MEM]][%[[INDICES_MEM]]#0, %[[INDICES_MEM]]#1], %[[LOCAL]][%[[INDICES_LDS]]#0, %[[INDICES_LDS]]#1]
// CHECK-SAME: vector<8xf16>, memref<64x128xf16>, memref<64x64xf16, 3>

%alloc = memref.alloc() : memref<64x64xf16, #gpu_lds_addrspace>
%collapse_alloc = memref.collapse_shape %alloc [[0, 1]] : memref<64x64xf16, #gpu_lds_addrspace> into memref<4096xf16, #gpu_lds_addrspace>
%mem = memref.alloc() : memref<64x128xf16>
%collapse = memref.collapse_shape %mem [[0, 1]] : memref<64x128xf16> into memref<8192xf16>
%collapse_mem = memref.collapse_shape %mem [[0, 1]] : memref<64x128xf16> into memref<8192xf16>
%c0 = arith.constant 0 : index
amdgpu.gather_to_lds %collapse[%offset_i], %alloc[%c0, %c0]
amdgpu.gather_to_lds %collapse_mem[%offset_i], %collapse_alloc[%offset_j]
: vector<8xf16>, memref<8192xf16>, memref<4096xf16, #gpu_lds_addrspace>
func.return
}


// -----

#gpu_lds_addrspace = 3


// CHECK: func @test_expand_shape_src_raw_buffer
// CHECK-SAME: %[[ARG0:.*]]: memref<8192xf16, #amdgpu.address_space<fat_raw_buffer>>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: index
func.func @test_expand_shape_src_raw_buffer(%mem : memref<8192xf16, #amdgpu.address_space<fat_raw_buffer>>, %offset_i: index, %offset_j: index) {
// CHECK: %[[LOCAL:.*]] = memref.alloc() : memref<4096xf16, 3>
// CHECK: %[[C0:.*]] = arith.constant 0 : index
// CHECK: %[[IDXM:.*]] = affine.linearize_index [%[[ARG1]], %[[ARG2]]] by (64, 128) : index
// CHECK: amdgpu.gather_to_lds %[[ARG0]][%[[IDXM]]], %[[LOCAL]][%[[C0]]]
// CHECK-SAME: vector<8xf16>, memref<8192xf16, #amdgpu.address_space<fat_raw_buffer>>, memref<4096xf16, 3>

%alloc = memref.alloc() : memref<4096xf16, #gpu_lds_addrspace>
%expand_mem = memref.expand_shape %mem [[0, 1]] output_shape [64, 128] : memref<8192xf16, #amdgpu.address_space<fat_raw_buffer>> into memref<64x128xf16, #amdgpu.address_space<fat_raw_buffer>>

%c0 = arith.constant 0 : index
amdgpu.gather_to_lds %expand_mem[%offset_i, %offset_j], %alloc[%c0]
: vector<8xf16>, memref<64x128xf16, #amdgpu.address_space<fat_raw_buffer>>, memref<4096xf16, #gpu_lds_addrspace>
func.return
}

// -----

#gpu_lds_addrspace = 3

// CHECK: func @test_expand_shape_dst_only
// CHECK-SAME: %[[ARG0:.*]]: index, %[[ARG1:.*]]: index
func.func @test_expand_shape_dst_only(%offset_i: index, %offset_j: index) {
// CHECK: %[[LOCAL:.*]] = memref.alloc() : memref<4096xf16, 3>
// CHECK: %[[MEM:.*]] = memref.alloc() : memref<8192xf16>
// CHECK: %[[C0:.*]] = arith.constant 0 : index
// CHECK: %[[IDX_LDS:.*]] = affine.linearize_index [%[[ARG1]], %[[C0]]] by (64, 64) : index
// CHECK: amdgpu.gather_to_lds %[[MEM]][%[[ARG0]]], %[[LOCAL]][%[[IDX_LDS]]]
// CHECK-SAME: vector<8xf16>, memref<8192xf16>, memref<4096xf16, 3>

%alloc = memref.alloc() : memref<4096xf16, #gpu_lds_addrspace>
%mem = memref.alloc() : memref<8192xf16>
%expand_alloc = memref.expand_shape %alloc [[0, 1]] output_shape [64, 64] : memref<4096xf16, #gpu_lds_addrspace> into memref<64x64xf16, #gpu_lds_addrspace>

%c0 = arith.constant 0 : index
amdgpu.gather_to_lds %mem[%offset_i], %expand_alloc[%offset_j, %c0]
: vector<8xf16>, memref<8192xf16>, memref<64x64xf16, #gpu_lds_addrspace>
func.return
}

// -----

#gpu_lds_addrspace = 3

// CHECK: func @test_nop
// CHECK-SAME: %[[ARG0:.*]]: memref<8192xf16, #amdgpu.address_space<fat_raw_buffer>>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: index
func.func @test_nop(%mem : memref<8192xf16, #amdgpu.address_space<fat_raw_buffer>>, %offset_i: index, %offset_j: index) {
// CHECK: %[[LOCAL:.*]] = memref.alloc() : memref<4096xf16, 3>
// CHECK: amdgpu.gather_to_lds %[[ARG0]][%[[ARG1]]], %[[LOCAL]][%[[ARG2]]]
// CHECK-SAME: vector<8xf16>, memref<8192xf16, #amdgpu.address_space<fat_raw_buffer>>, memref<4096xf16, 3>

%alloc = memref.alloc() : memref<4096xf16, #gpu_lds_addrspace>
amdgpu.gather_to_lds %mem[%offset_i], %alloc[%offset_j]
: vector<8xf16>, memref<8192xf16, #amdgpu.address_space<fat_raw_buffer>>, memref<4096xf16, #gpu_lds_addrspace>
func.return
}