[Mlir-commits] [mlir] 794979a - [mlir][gpu] Improve foreach_thread distribution
Thomas Raoux
llvmlistbot at llvm.org
Tue Jan 17 09:24:55 PST 2023
Author: Thomas Raoux
Date: 2023-01-17T17:12:55Z
New Revision: 794979ad8cc975b4417dcf5fcce9c9e8e28e52b3
URL: https://github.com/llvm/llvm-project/commit/794979ad8cc975b4417dcf5fcce9c9e8e28e52b3
DIFF: https://github.com/llvm/llvm-project/commit/794979ad8cc975b4417dcf5fcce9c9e8e28e52b3.diff
LOG: [mlir][gpu] Improve foreach_thread distribution
Replace Ids with 0 when block dim is 1 when distributing foreach_thread.
Differential Revision: https://reviews.llvm.org/D141718
Added:
Modified:
mlir/lib/Dialect/GPU/TransformOps/GPUTransformOps.cpp
mlir/test/Dialect/GPU/transform-gpu.mlir
Removed:
################################################################################
diff --git a/mlir/lib/Dialect/GPU/TransformOps/GPUTransformOps.cpp b/mlir/lib/Dialect/GPU/TransformOps/GPUTransformOps.cpp
index 6442523d624f0..69c87c433daf1 100644
--- a/mlir/lib/Dialect/GPU/TransformOps/GPUTransformOps.cpp
+++ b/mlir/lib/Dialect/GPU/TransformOps/GPUTransformOps.cpp
@@ -434,6 +434,12 @@ static DiagnosedSilenceableFailure rewriteOneForeachThreadToGpuThreads(
rewriter.create<ThreadIdOp>(loc, indexType, Dimension::x),
rewriter.create<ThreadIdOp>(loc, indexType, Dimension::y),
rewriter.create<ThreadIdOp>(loc, indexType, Dimension::z)};
+ // Replace ids of dimension size 1 by zero to simplify the IR.
+ Value zero = rewriter.create<arith::ConstantIndexOp>(loc, 0);
+ for (size_t i : llvm::seq(size_t(0), globalBlockDims.size())) {
+ if (globalBlockDims[i] == 1)
+ threadOps[i] = zero;
+ }
IRMapping bvm;
for (auto [blockIdx, blockDim] :
llvm::zip(foreachThreadOp.getThreadIndices(), threadMapping)) {
diff --git a/mlir/test/Dialect/GPU/transform-gpu.mlir b/mlir/test/Dialect/GPU/transform-gpu.mlir
index 97c9c19f6ec05..1c337fed5cc09 100644
--- a/mlir/test/Dialect/GPU/transform-gpu.mlir
+++ b/mlir/test/Dialect/GPU/transform-gpu.mlir
@@ -194,3 +194,39 @@ transform.sequence failures(propagate) {
%funcop = transform.structured.match ops{["gpu.launch"]} in %arg0
transform.gpu.map_nested_foreach_to_threads %funcop { blockDim = [32]}
}
+
+// -----
+
+!type = memref<3 x 2 x 32 x f32>
+!type1d = memref<32 x f32>
+
+// CHECK-LABEL: func.func @saxpy3d_fold_id_z(
+func.func @saxpy3d_fold_id_z(%x: !type, %y: !type, %t: !type1d, %alpha : f32, %stream : !gpu.async.token) -> !type {
+ %one = arith.constant 1 : index
+ %c12 = arith.constant 12 : index
+ %c9 = arith.constant 9 : index
+ %c7 = arith.constant 7 : index
+// CHECK: %[[C0:.+]] = arith.constant 0 : index
+// CHECK-NOT: gpu.thread_id z
+ %name = gpu.launch async[%stream] blocks(%arg3, %arg4, %arg5) in (%arg9 = %one, %arg10 = %one, %arg11 = %one)
+ threads(%arg6, %arg7, %arg8) in (%arg12 = %one, %arg13 = %one, %arg14 = %one)
+ {
+ scf.foreach_thread (%i, %j, %k) in (%one, %c7, %c9) {
+// CHECK: memref.load %{{.*}}[%[[C0]],
+// CHECK: memref.load %{{.*}}[%[[C0]],
+ %4 = memref.load %x[%i, %j, %k] : !type
+ %5 = memref.load %y[%i, %j, %k] : !type
+ %6 = math.fma %alpha, %4, %5 : f32
+// CHECK: memref.store %{{.*}}, %{{.*}}[%[[C0]]
+ memref.store %6, %y[%i, %j, %k] : !type
+ } { mapping = [#gpu.thread<z>, #gpu.thread<y>, #gpu.thread<x>] }
+ gpu.terminator
+ }
+ return %y : !type
+}
+
+transform.sequence failures(propagate) {
+^bb1(%arg0: !pdl.operation):
+ %funcop = transform.structured.match ops{["gpu.launch"]} in %arg0
+ transform.gpu.map_nested_foreach_to_threads %funcop { blockDim = [12, 9, 1], syncAfterDistribute = false }
+}
More information about the Mlir-commits
mailing list