Skip to content

[mlir][vector] Distribute all non-permutation or broadcasted masked transfer reads #73539

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 1 commit into from
Nov 27, 2023
Merged

[mlir][vector] Distribute all non-permutation or broadcasted masked transfer reads #73539

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
2 changes: 1 addition & 1 deletion mlir/lib/Dialect/Vector/Transforms/VectorDistribute.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -837,7 +837,7 @@ struct WarpOpTransferRead : public OpRewritePattern<WarpExecuteOnLane0Op> {
// of which lane is responsible for which element is captured strictly
// by shape information on the warp op, and thus requires materializing
// the permutation in IR.
if (!read.getPermutationMap().isMinorIdentity())
if (!mlir::compressUnusedDims(read.getPermutationMap()).isIdentity())
return failure();
VectorType maskType =
getDistributedType(read.getMaskType(), map, warpOp.getWarpSize());
Expand Down
26 changes: 26 additions & 0 deletions mlir/test/Dialect/Vector/vector-warp-distribute.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1351,6 +1351,32 @@ func.func @warp_propagate_masked_transfer_read(%laneid: index, %src: memref<4096

// -----

func.func @warp_propagate_nontrivial_map_masked_transfer_read(%laneid: index, %src: memref<4096x4096xf32>, %index: index) -> vector<2xf32> {
%f0 = arith.constant 0.000000e+00 : f32
%c0 = arith.constant 0 : index
%r = vector.warp_execute_on_lane_0(%laneid)[64] -> (vector<2xf32>) {
%mask = "mask_def_0"() : () -> (vector<128xi1>)
%0 = vector.transfer_read %src[%index, %c0], %f0, %mask {in_bounds = [true], permutation_map = affine_map<(d0, d1) -> (d0)>} : memref<4096x4096xf32>, vector<128xf32>
vector.yield %0 : vector<128xf32>
}
return %r : vector<2xf32>
}

// CHECK-PROP-DAG: #[[$MAP0:.+]] = affine_map<()[s0, s1] -> (s0 + s1 * 2)>
// CHECK-PROP-DAG: #[[$MAP1:.+]] = affine_map<(d0, d1) -> (d0)>
// CHECK-PROP-LABEL: func.func @warp_propagate_nontrivial_map_masked_transfer_read
// CHECK-PROP-SAME: %[[ARG0:.+]]: index, {{.*}}, %[[ARG2:.+]]: index
// CHECK-PROP: %[[C0:.*]] = arith.constant 0 : index
// CHECK-PROP: %[[R:.*]] = vector.warp_execute_on_lane_0(%{{.*}})[64] -> (vector<2xi1>) {
// CHECK-PROP: %[[M0:.*]] = "mask_def_0"
// CHECK-PROP: vector.yield %[[M0]] : vector<128xi1>
// CHECK-PROP: }
// CHECK-PROP: %[[DIST_READ_IDX0:.+]] = affine.apply #[[$MAP0]]()[%[[ARG2]], %[[ARG0]]]
// CHECK-PROP: vector.transfer_read {{.*}}[%[[DIST_READ_IDX0]], %[[C0]]], {{.*}}, %[[R]]
// CHECK-PROP-SAME: permutation_map = #[[$MAP1]]} {{.*}} vector<2xf32>

// -----

func.func @warp_propagate_masked_transfer_read_shared_mask(%laneid: index, %src: memref<4096x4096xf32>, %index: index, %index2: index, %mask_ub: index) -> (vector<2xf32>, vector<2xf32>) {
%f0 = arith.constant 0.000000e+00 : f32
%c0 = arith.constant 0 : index
Expand Down