Lower SplitOp to Krnl dialect (#155)

* Fix importing variadic output

* Lower splitop

* Support unknown dimension and add lit tests

Co-authored-by: Tian Jin <tjingrant@gmail.com>
This commit is contained in:
Tung D. Le 2020-06-11 11:57:20 +09:00 committed by GitHub
parent 4ab96fbc6c
commit 8c4d527eea
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
6 changed files with 216 additions and 0 deletions

View File

@ -20,6 +20,7 @@ add_library(OMONNXToKrnl
Tensor/Unsqueeze.cpp
Tensor/Constant.cpp
Tensor/Concat.cpp
Tensor/Split.cpp
ConvertONNXToKrnl.cpp)
target_link_libraries(OMONNXToKrnl
onnx)

View File

@ -99,6 +99,7 @@ void FrontendToKrnlLoweringPass::runOnOperation() {
populateLoweringONNXIdentityOpPattern(patterns, &getContext());
populateLoweringONNXConstantOpPattern(patterns, &getContext());
populateLoweringONNXConcatOpPattern(patterns, &getContext());
populateLoweringONNXSplitOpPattern(patterns, &getContext());
// Neural network
populateLoweringONNXConvOpPattern(patterns, &getContext());
populateLoweringONNXNormalizationOpPattern(patterns, &getContext());

View File

@ -250,3 +250,6 @@ void populateLoweringONNXConstantOpPattern(
void populateLoweringONNXConcatOpPattern(
OwningRewritePatternList &patterns, MLIRContext *ctx);
void populateLoweringONNXSplitOpPattern(
OwningRewritePatternList &patterns, MLIRContext *ctx);

View File

@ -0,0 +1,106 @@
//===---------------- Split.cpp - Lowering Split Op -----------------------===//
//
// Copyright 2019 The IBM Research Authors.
//
// =============================================================================
//
// This file lowers the ONNX Split Operator to Krnl dialect.
//
//===----------------------------------------------------------------------===//
#include "src/Conversion/ONNXToKrnl/ONNXToKrnlCommon.hpp"
using namespace mlir;
struct ONNXSplitOpLowering : public ConversionPattern {
ONNXSplitOpLowering(MLIRContext *ctx)
: ConversionPattern(mlir::ONNXSplitOp::getOperationName(), 1, ctx) {}
LogicalResult matchAndRewrite(Operation *op, ArrayRef<Value> operands,
ConversionPatternRewriter &rewriter) const final {
// Gather info.
auto loc = op->getLoc();
ONNXSplitOp splitOp = llvm::dyn_cast<ONNXSplitOp>(op);
auto axis = splitOp.axis().getSExtValue();
auto split = splitOp.split().getValue();
SmallVector<int64_t, 4> splitOffset;
int64_t offset = 0;
for (int i = 0; i < split.size(); ++i) {
splitOffset.emplace_back(offset);
offset += ArrayAttrIntVal(split, i);
}
auto rank = splitOp.input().getType().cast<ShapedType>().getRank();
auto outputNum = splitOp.getNumResults();
// Alloc and dealloc.
SmallVector<Value, 4> allocs;
for (int i = 0; i < outputNum; ++i) {
Value alloc;
bool insertDealloc = checkInsertDealloc(op, i);
auto memRefType = convertToMemRefType(splitOp.outputs()[i].getType());
if (hasAllConstantDimensions(memRefType))
alloc = insertAllocAndDealloc(memRefType, loc, rewriter, insertDealloc);
else {
SmallVector<Value, 4> allocOperands;
auto shape = memRefType.getShape();
for (decltype(rank) r = 0; r < rank; ++r) {
if (shape[r] < 0) {
Value dim;
if (r != axis)
dim = rewriter.create<DimOp>(loc, operands[0], r);
else
dim = emitConstantOp(rewriter, loc, rewriter.getIndexType(),
ArrayAttrIntVal(split, i));
allocOperands.push_back(dim);
}
}
alloc = rewriter.create<AllocOp>(loc, memRefType, allocOperands);
if (insertDealloc) {
auto *parentBlock = alloc.getDefiningOp()->getBlock();
auto dealloc = rewriter.create<DeallocOp>(loc, alloc);
dealloc.getOperation()->moveBefore(&parentBlock->back());
}
}
allocs.emplace_back(alloc);
}
// Creates loops, one for each output.
for (int i = 0; i < outputNum; ++i) {
OpBuilder::InsertionGuard insertGuard(rewriter);
// Create loop.
BuildKrnlLoop outputLoops(rewriter, loc, rank);
outputLoops.createDefineOptimizeAndIterateOp(allocs[i]);
outputLoops.createIterateOp();
rewriter.setInsertionPointToStart(outputLoops.getIterateBlock());
// Indices for the read and write.
SmallVector<Value, 4> readIndices;
SmallVector<Value, 4> writeIndices;
for (int r = 0; r < rank; ++r) {
// Same index for read and write if the dimension is:
// - the first dimension, or
// - not the split axis.
if (i == 0 || r != axis) {
readIndices.emplace_back(outputLoops.getInductionVar(r));
} else {
auto index = rewriter.getAffineDimExpr(0);
auto indexMap = AffineMap::get(1, 0, index + splitOffset[i]);
auto indexWithOffset = rewriter.create<AffineApplyOp>(loc, indexMap,
ArrayRef<Value>{/*index=*/outputLoops.getInductionVar(r)});
readIndices.emplace_back(indexWithOffset);
}
writeIndices.emplace_back(outputLoops.getInductionVar(r));
}
// Insert copy.
auto loadData = rewriter.create<LoadOp>(loc, operands[0], readIndices);
rewriter.create<StoreOp>(loc, loadData, allocs[i], writeIndices);
}
rewriter.replaceOp(op, allocs);
return success();
}
};
void populateLoweringONNXSplitOpPattern(
OwningRewritePatternList &patterns, MLIRContext *ctx) {
patterns.insert<ONNXSplitOpLowering>(ctx);
}

View File

@ -352,8 +352,16 @@ test_to_enable = [
"test_lstm_with_initial_bias_cpu",
"test_lstm_with_peepholes_cpu",
# Split
"test_split_equal_parts_1d_cpu",
"test_split_equal_parts_2d_cpu",
"test_split_equal_parts_default_axis_cpu",
"test_split_variable_parts_1d_cpu",
"test_split_variable_parts_2d_cpu",
"test_split_variable_parts_default_axis_cpu",
]
# Extract name of all test cases.
import inspect
all_tests = inspect.getmembers(

View File

@ -2134,3 +2134,100 @@ func @test_lstm_bidirectional_mode(%arg0: tensor<4x3x2xf32>, %arg1: tensor<1x12x
// CHECK: %[[REVERSE_SEQUENCE_IV:.+]] = affine.apply [[REVERSE_IV_MAP]](%arg3)[%[[SEQUENCE_LEN]]{{]}}
// CHECK: [[Xt_LOAD:%.+]] = load %arg0[%[[REVERSE_SEQUENCE_IV]], {{.*}}, {{.*}}] : memref<4x3x2xf32>
}
// -----
func @test_split_equal(%arg0 : tensor<16x32x64xf32>) -> (tensor<*xf32>, tensor<*xf32>) {
%0, %1 = "onnx.Split"(%arg0) { axis = 0} : (tensor<16x32x64xf32>) -> (tensor<*xf32>, tensor<*xf32>)
"std.return"(%0, %1) : (tensor<*xf32>, tensor<*xf32>) -> ()
// CHECK: [[INDEX_MAP:#.+]] = affine_map<(d0) -> (d0 + 8)>
// CHECK-LABEL: @test_split_equal
// CHECK: [[RES_1:%.+]] = alloc() : memref<8x32x64xf32>
// CHECK: [[RES_0:%.+]] = alloc() : memref<8x32x64xf32>
// CHECK: [[DEF_LOOP_0:%.+]]:3 = krnl.define_loops 3
// CHECK: [[OPT_LOOP_0:%.+]]:3 = krnl.optimize_loops {
// CHECK: krnl.return_loops [[DEF_LOOP_0]]#0, [[DEF_LOOP_0]]#1, [[DEF_LOOP_0]]#2
// CHECK: } : () -> (!krnl.loop, !krnl.loop, !krnl.loop)
// CHECK: krnl.iterate([[OPT_LOOP_0]]#0, [[OPT_LOOP_0]]#1, [[OPT_LOOP_0]]#2) with ([[DEF_LOOP_0]]#0 -> %arg1 = 0 to 8, [[DEF_LOOP_0]]#1 -> %arg2 = 0 to 32, [[DEF_LOOP_0]]#2 -> %arg3 = 0 to 64) {
// CHECK: [[LOAD_0:%.+]] = load %arg0[%arg1, %arg2, %arg3] : memref<16x32x64xf32>
// CHECK: store [[LOAD_0]], [[RES_0]][%arg1, %arg2, %arg3] : memref<8x32x64xf32>
// CHECK: }
// CHECK: [[DEF_LOOP_1:%.+]]:3 = krnl.define_loops 3
// CHECK: [[OPT_LOOP_1:%.+]]:3 = krnl.optimize_loops {
// CHECK: krnl.return_loops [[DEF_LOOP_1]]#0, [[DEF_LOOP_1]]#1, [[DEF_LOOP_1]]#2
// CHECK: } : () -> (!krnl.loop, !krnl.loop, !krnl.loop)
// CHECK: krnl.iterate([[OPT_LOOP_1]]#0, [[OPT_LOOP_1]]#1, [[OPT_LOOP_1]]#2) with ([[DEF_LOOP_1]]#0 -> %arg1 = 0 to 8, [[DEF_LOOP_1]]#1 -> %arg2 = 0 to 32, [[DEF_LOOP_1]]#2 -> %arg3 = 0 to 64) {
// CHECK: %[[INDEX:.+]] = affine.apply [[INDEX_MAP]](%arg1)
// CHECK: [[LOAD_1:%.+]] = load %arg0[%[[INDEX]], %arg2, %arg3] : memref<16x32x64xf32>
// CHECK: store [[LOAD_1]], [[RES_1]][%arg1, %arg2, %arg3] : memref<8x32x64xf32>
// CHECK: }
// CHECK: return [[RES_0]], [[RES_1]] : memref<8x32x64xf32>, memref<8x32x64xf32>
}
// -----
func @test_split_variable(%arg0 : tensor<16x32x64xf32>) -> (tensor<*xf32>, tensor<*xf32>) {
%0, %1 = "onnx.Split"(%arg0) { axis = 1, split = [2, 30]} : (tensor<16x32x64xf32>) -> (tensor<*xf32>, tensor<*xf32>)
"std.return"(%0, %1) : (tensor<*xf32>, tensor<*xf32>) -> ()
// CHECK: [[INDEX_MAP:#.+]] = affine_map<(d0) -> (d0 + 2)>
// CHECK-LABEL: @test_split_variable
// CHECK: [[RES_1:%.+]] = alloc() : memref<16x30x64xf32>
// CHECK: [[RES_0:%.+]] = alloc() : memref<16x2x64xf32>
// CHECK: [[DEF_LOOP_0:%.+]]:3 = krnl.define_loops 3
// CHECK: [[OPT_LOOP_0:%.+]]:3 = krnl.optimize_loops {
// CHECK: krnl.return_loops [[DEF_LOOP_0]]#0, [[DEF_LOOP_0]]#1, [[DEF_LOOP_0]]#2
// CHECK: } : () -> (!krnl.loop, !krnl.loop, !krnl.loop)
// CHECK: krnl.iterate([[OPT_LOOP_0]]#0, [[OPT_LOOP_0]]#1, [[OPT_LOOP_0]]#2) with ([[DEF_LOOP_0]]#0 -> %arg1 = 0 to 16, [[DEF_LOOP_0]]#1 -> %arg2 = 0 to 2, [[DEF_LOOP_0]]#2 -> %arg3 = 0 to 64) {
// CHECK: [[LOAD_0:%.+]] = load %arg0[%arg1, %arg2, %arg3] : memref<16x32x64xf32>
// CHECK: store [[LOAD_0]], [[RES_0]][%arg1, %arg2, %arg3] : memref<16x2x64xf32>
// CHECK: }
// CHECK: [[DEF_LOOP_1:%.+]]:3 = krnl.define_loops 3
// CHECK: [[OPT_LOOP_1:%.+]]:3 = krnl.optimize_loops {
// CHECK: krnl.return_loops [[DEF_LOOP_1]]#0, [[DEF_LOOP_1]]#1, [[DEF_LOOP_1]]#2
// CHECK: } : () -> (!krnl.loop, !krnl.loop, !krnl.loop)
// CHECK: krnl.iterate([[OPT_LOOP_1]]#0, [[OPT_LOOP_1]]#1, [[OPT_LOOP_1]]#2) with ([[DEF_LOOP_1]]#0 -> %arg1 = 0 to 16, [[DEF_LOOP_1]]#1 -> %arg2 = 0 to 30, [[DEF_LOOP_1]]#2 -> %arg3 = 0 to 64) {
// CHECK: %[[INDEX:.+]] = affine.apply [[INDEX_MAP]](%arg2)
// CHECK: [[LOAD_1:%.+]] = load %arg0[%arg1, %[[INDEX]], %arg3] : memref<16x32x64xf32>
// CHECK: store [[LOAD_1]], [[RES_1]][%arg1, %arg2, %arg3] : memref<16x30x64xf32>
// CHECK: }
// CHECK: return [[RES_0]], [[RES_1]] : memref<16x2x64xf32>, memref<16x30x64xf32>
}
// -----
func @test_split_unknown_dimension(%arg0 : tensor<?x?x64xf32>) -> (tensor<*xf32>, tensor<*xf32>) {
%0, %1 = "onnx.Split"(%arg0) { axis = 1, split = [2, 30]} : (tensor<?x?x64xf32>) -> (tensor<*xf32>, tensor<*xf32>)
"std.return"(%0, %1) : (tensor<*xf32>, tensor<*xf32>) -> ()
// CHECK: [[INDEX_MAP:#.+]] = affine_map<(d0) -> (d0 + 2)>
// CHECK-LABEL: @test_split_unknown_dimension
// CHECK: [[DIM_0:%.+]] = dim %arg0, 0 : memref<?x?x64xf32>
// CHECK: [[RES_0:%.+]] = alloc([[DIM_0]]) : memref<?x2x64xf32>
// CHECK: [[DIM_1:%.+]] = dim %arg0, 0 : memref<?x?x64xf32>
// CHECK: [[RES_1:%.+]] = alloc([[DIM_1]]) : memref<?x30x64xf32>
// CHECK: [[DEF_LOOP_0:%.+]]:3 = krnl.define_loops 3
// CHECK: [[OPT_LOOP_0:%.+]]:3 = krnl.optimize_loops {
// CHECK: krnl.return_loops [[DEF_LOOP_0]]#0, [[DEF_LOOP_0]]#1, [[DEF_LOOP_0]]#2
// CHECK: } : () -> (!krnl.loop, !krnl.loop, !krnl.loop)
// CHECK: [[DIM_0:%.+]] = dim [[RES_0]], 0 : memref<?x2x64xf32>
// CHECK: krnl.iterate([[OPT_LOOP_0]]#0, [[OPT_LOOP_0]]#1, [[OPT_LOOP_0]]#2) with ([[DEF_LOOP_0]]#0 -> %arg1 = 0 to [[DIM_0]], [[DEF_LOOP_0]]#1 -> %arg2 = 0 to 2, [[DEF_LOOP_0]]#2 -> %arg3 = 0 to 64) {
// CHECK: [[LOAD_0:%.+]] = load %arg0[%arg1, %arg2, %arg3] : memref<?x?x64xf32>
// CHECK: store [[LOAD_0]], [[RES_0]][%arg1, %arg2, %arg3] : memref<?x2x64xf32>
// CHECK: }
// CHECK: [[DEF_LOOP_1:%.+]]:3 = krnl.define_loops 3
// CHECK: [[OPT_LOOP_1:%.+]]:3 = krnl.optimize_loops {
// CHECK: krnl.return_loops [[DEF_LOOP_1]]#0, [[DEF_LOOP_1]]#1, [[DEF_LOOP_1]]#2
// CHECK: } : () -> (!krnl.loop, !krnl.loop, !krnl.loop)
// CHECK: [[DIM_1:%.+]] = dim [[RES_1]], 0 : memref<?x30x64xf32>
// CHECK: krnl.iterate([[OPT_LOOP_1]]#0, [[OPT_LOOP_1]]#1, [[OPT_LOOP_1]]#2) with ([[DEF_LOOP_1]]#0 -> %arg1 = 0 to [[DIM_1]], [[DEF_LOOP_1]]#1 -> %arg2 = 0 to 30, [[DEF_LOOP_1]]#2 -> %arg3 = 0 to 64) {
// CHECK: %[[INDEX:.+]] = affine.apply [[INDEX_MAP]](%arg2)
// CHECK: [[LOAD_1:%.+]] = load %arg0[%arg1, %[[INDEX]], %arg3] : memref<?x?x64xf32>
// CHECK: store [[LOAD_1]], [[RES_1]][%arg1, %arg2, %arg3] : memref<?x30x64xf32>
// CHECK: }
// CHECK: return [[RES_0]], [[RES_1]] : memref<?x2x64xf32>, memref<?x30x64xf32>
}