From 68a7c001ad9ac20bdacbc4e406fa2b4f27a74b13 Mon Sep 17 00:00:00 2001 From: Eugene Zhulenev Date: Mon, 6 Dec 2021 11:23:53 -0800 Subject: [PATCH] [mlir] Improve async parallel for tests + fix typos Do load and store to verify that we process each element of the iteration space once. Reviewed By: cota Differential Revision: https://reviews.llvm.org/D115152 --- mlir/lib/Dialect/Async/Transforms/AsyncParallelFor.cpp | 5 +++-- .../Dialect/Async/CPU/test-async-parallel-for-1d.mlir | 18 +++++++++++++++--- .../Dialect/Async/CPU/test-async-parallel-for-2d.mlir | 18 +++++++++++++++--- 3 files changed, 33 insertions(+), 8 deletions(-) diff --git a/mlir/lib/Dialect/Async/Transforms/AsyncParallelFor.cpp b/mlir/lib/Dialect/Async/Transforms/AsyncParallelFor.cpp index c9514a9..8d80e67 100644 --- a/mlir/lib/Dialect/Async/Transforms/AsyncParallelFor.cpp +++ b/mlir/lib/Dialect/Async/Transforms/AsyncParallelFor.cpp @@ -252,6 +252,7 @@ createParallelComputeFunction(scf::ParallelOp op, PatternRewriter &rewriter) { Value blockFirstIndex = b.create(blockIndex, blockSize); // The last one-dimensional index in the block defined by the `blockIndex`: + // blockLastIndex = min(blockFirstIndex + blockSize, tripCount) - 1 Value blockEnd0 = b.create(blockFirstIndex, blockSize); Value blockEnd1 = b.create(blockEnd0, tripCount); Value blockLastIndex = b.create(blockEnd1, c1); @@ -279,7 +280,7 @@ createParallelComputeFunction(scf::ParallelOp op, PatternRewriter &rewriter) { // iteration coordinate using parallel operation bounds and step: // // computeBlockInductionVars[loopIdx] = - // lowerBound[loopIdx] + blockCoord[loopIdx] * step[loopDdx] + // lowerBound[loopIdx] + blockCoord[loopIdx] * step[loopIdx] SmallVector computeBlockInductionVars(op.getNumLoops()); // We need to know if we are in the first or last iteration of the @@ -329,7 +330,7 @@ createParallelComputeFunction(scf::ParallelOp op, PatternRewriter &rewriter) { // Keep building loop nest. if (loopIdx < op.getNumLoops() - 1) { - // Select nested loop lower/upper bounds depending on out position in + // Select nested loop lower/upper bounds depending on our position in // the multi-dimensional iteration space. auto lb = nb.create(isBlockFirstCoord[loopIdx], blockFirstCoord[loopIdx + 1], c0); diff --git a/mlir/test/Integration/Dialect/Async/CPU/test-async-parallel-for-1d.mlir b/mlir/test/Integration/Dialect/Async/CPU/test-async-parallel-for-1d.mlir index 4497fb0..4975826 100644 --- a/mlir/test/Integration/Dialect/Async/CPU/test-async-parallel-for-1d.mlir +++ b/mlir/test/Integration/Dialect/Async/CPU/test-async-parallel-for-1d.mlir @@ -68,11 +68,19 @@ func @entry() { %A = memref.alloc() : memref<9xf32> %U = memref.cast %A : memref<9xf32> to memref<*xf32> + // Initialize memref with zeros because we do load and store to in every test + // to verify that we process each element of the iteration space once. + scf.parallel (%i) = (%lb) to (%ub) step (%c1) { + memref.store %c0, %A[%i] : memref<9xf32> + } + // 1. %i = (0) to (9) step (1) scf.parallel (%i) = (%lb) to (%ub) step (%c1) { %0 = arith.index_cast %i : index to i32 %1 = arith.sitofp %0 : i32 to f32 - memref.store %1, %A[%i] : memref<9xf32> + %2 = memref.load %A[%i] : memref<9xf32> + %3 = arith.addf %1, %2 : f32 + memref.store %3, %A[%i] : memref<9xf32> } // CHECK: [0, 1, 2, 3, 4, 5, 6, 7, 8] call @print_memref_f32(%U): (memref<*xf32>) -> () @@ -85,7 +93,9 @@ func @entry() { scf.parallel (%i) = (%lb) to (%ub) step (%c2) { %0 = arith.index_cast %i : index to i32 %1 = arith.sitofp %0 : i32 to f32 - memref.store %1, %A[%i] : memref<9xf32> + %2 = memref.load %A[%i] : memref<9xf32> + %3 = arith.addf %1, %2 : f32 + memref.store %3, %A[%i] : memref<9xf32> } // CHECK: [0, 0, 2, 0, 4, 0, 6, 0, 8] call @print_memref_f32(%U): (memref<*xf32>) -> () @@ -102,7 +112,9 @@ func @entry() { %1 = arith.sitofp %0 : i32 to f32 %2 = arith.constant 20 : index %3 = arith.addi %i, %2 : index - memref.store %1, %A[%3] : memref<9xf32> + %4 = memref.load %A[%3] : memref<9xf32> + %5 = arith.addf %1, %4 : f32 + memref.store %5, %A[%3] : memref<9xf32> } // CHECK: [-20, 0, 0, -17, 0, 0, -14, 0, 0] call @print_memref_f32(%U): (memref<*xf32>) -> () diff --git a/mlir/test/Integration/Dialect/Async/CPU/test-async-parallel-for-2d.mlir b/mlir/test/Integration/Dialect/Async/CPU/test-async-parallel-for-2d.mlir index b9b08bb..b92dfba 100644 --- a/mlir/test/Integration/Dialect/Async/CPU/test-async-parallel-for-2d.mlir +++ b/mlir/test/Integration/Dialect/Async/CPU/test-async-parallel-for-2d.mlir @@ -59,13 +59,21 @@ func @entry() { %A = memref.alloc() : memref<8x8xf32> %U = memref.cast %A : memref<8x8xf32> to memref<*xf32> + // Initialize memref with zeros because we do load and store to in every test + // to verify that we process each element of the iteration space once. + scf.parallel (%i, %j) = (%lb, %lb) to (%ub, %ub) step (%c1, %c1) { + memref.store %c0, %A[%i, %j] : memref<8x8xf32> + } + // 1. (%i, %i) = (0, 8) to (8, 8) step (1, 1) scf.parallel (%i, %j) = (%lb, %lb) to (%ub, %ub) step (%c1, %c1) { %0 = arith.muli %i, %c8 : index %1 = arith.addi %j, %0 : index %2 = arith.index_cast %1 : index to i32 %3 = arith.sitofp %2 : i32 to f32 - memref.store %3, %A[%i, %j] : memref<8x8xf32> + %4 = memref.load %A[%i, %j] : memref<8x8xf32> + %5 = arith.addf %3, %4 : f32 + memref.store %5, %A[%i, %j] : memref<8x8xf32> } // CHECK: [0, 1, 2, 3, 4, 5, 6, 7] @@ -88,7 +96,9 @@ func @entry() { %1 = arith.addi %j, %0 : index %2 = arith.index_cast %1 : index to i32 %3 = arith.sitofp %2 : i32 to f32 - memref.store %3, %A[%i, %j] : memref<8x8xf32> + %4 = memref.load %A[%i, %j] : memref<8x8xf32> + %5 = arith.addf %3, %4 : f32 + memref.store %5, %A[%i, %j] : memref<8x8xf32> } // CHECK: [0, 1, 2, 3, 4, 5, 6, 7] @@ -111,7 +121,9 @@ func @entry() { %1 = arith.addi %j, %0 : index %2 = arith.index_cast %1 : index to i32 %3 = arith.sitofp %2 : i32 to f32 - memref.store %3, %A[%i, %j] : memref<8x8xf32> + %4 = memref.load %A[%i, %j] : memref<8x8xf32> + %5 = arith.addf %3, %4 : f32 + memref.store %5, %A[%i, %j] : memref<8x8xf32> } // CHECK: [0, 0, 2, 0, 4, 0, 6, 0] -- 2.7.4