maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

[mlir] Handle cases where transfer_read should turn into a scalar load 

The existing vector transforms reduce the dimension of transfer_read
ops.  However, beyond a certain point, the vector op actually has
to be reduced to a scalar load, since we can't load a zero-dimension
vector.  This handles this case.

Note that in the longer term, it may be preferaby to support
zero-dimension vectors.  see
https://llvm.discourse.group/t/should-we-have-0-d-vectors/3097.

Differential Revision: https://reviews.llvm.org/D103432
This commit is contained in:
Stephen Neuendorffer 2021-05-31 22:32:49 -07:00
parent 596e61c332
commit 432341d8a8
2 changed files with 20 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
mlir/lib/Dialect/Vector/VectorTransforms.cpp
View File

@ -2828,6 +2828,18 @@ struct TransferOpReduceRank : public OpRewritePattern<vector::TransferReadOp> {
// with broadasting. Otherwise we first want to permute the map.
if (!newMap.isMinorIdentityWithBroadcasting())
return failure();
// TODO: support zero-dimension vectors natively. See:
// https://llvm.discourse.group/t/should-we-have-0-d-vectors/3097.
// In the meantime, lower these to a scalar load when they pop up.
if (reducedShapeRank == 0) {
Value newRead = rewriter.create<memref::LoadOp>(
op.getLoc(), originalVecType.getElementType(), op.source(),
op.indices());
rewriter.replaceOpWithNewOp<vector::BroadcastOp>(op, originalVecType,
newRead);
return success();
}
SmallVector<int64_t> newShape = llvm::to_vector<4>(
originalVecType.getShape().take_back(reducedShapeRank));
// Vector rank cannot be zero. Handled by TransferReadToVectorLoadLowering.

11
mlir/test/Dialect/Vector/vector-transfer-lowering.mlir
View File

@ -228,6 +228,7 @@ func @transfer_broadcasting_complex(%mem : memref<10x20x30x8x8xf32>, %i : index)
#map3 = affine_map<(d0, d1) -> (d1, d0, 0, 0)>
#map4 = affine_map<(d0, d1) -> (0, d1, 0, d0)>
#map5 = affine_map<(d0, d1, d2, d3) -> (d2, d1, d3, d0)>
#map6 = affine_map<(d0, d1) -> (0)>
// CHECK-DAG: #[[$MAP0:.*]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, 0, 0)>
// CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d1, 0, d3)>
@ -235,7 +236,7 @@ func @transfer_broadcasting_complex(%mem : memref<10x20x30x8x8xf32>, %i : index)
// CHECK-LABEL: func @transfer_read_permutations
func @transfer_read_permutations(%arg0 : memref<?x?xf32>, %arg1 : memref<?x?x?x?xf32>)
-> (vector<7x14x8x16xf32>, vector<7x14x8x16xf32>, vector<7x14x8x16xf32>,
vector<7x14x8x16xf32>, vector<7x14x8x16xf32>, vector<7x14x8x16xf32>) {
vector<7x14x8x16xf32>, vector<7x14x8x16xf32>, vector<7x14x8x16xf32>, vector<8xf32>) {
// CHECK-DAG: %[[CF0:.*]] = constant 0.000000e+00 : f32
// CHECK-DAG: %[[C0:.*]] = constant 0 : index
%cst = constant 0.000000e+00 : f32
@ -275,9 +276,13 @@ func @transfer_read_permutations(%arg0 : memref<?x?xf32>, %arg1 : memref<?x?x?x?
// CHECK: vector.transfer_read %{{.*}}[%[[C0]], %[[C0]], %[[C0]], %[[C0]]], %[[CF0]] : memref<?x?x?x?xf32>, vector<16x14x7x8xf32>
// CHECK: vector.transpose %{{.*}}, [2, 1, 3, 0] : vector<16x14x7x8xf32> to vector<7x14x8x16xf32>
return %0, %1, %2, %3, %4, %5 : vector<7x14x8x16xf32>, vector<7x14x8x16xf32>,
%6 = vector.transfer_read %arg0[%c0, %c0], %cst {permutation_map = #map6} : memref<?x?xf32>, vector<8xf32>
// CHECK: memref.load %{{.*}}[%[[C0]], %[[C0]]] : memref<?x?xf32>
// CHECK: vector.broadcast %{{.*}} : f32 to vector<8xf32>
return %0, %1, %2, %3, %4, %5, %6 : vector<7x14x8x16xf32>, vector<7x14x8x16xf32>,
vector<7x14x8x16xf32>, vector<7x14x8x16xf32>, vector<7x14x8x16xf32>,
vector<7x14x8x16xf32>
vector<7x14x8x16xf32>, vector<8xf32>
}
// -----