diff --git a/lib/Conversion/TorchOnnxToTorch/DefaultDomainGtoP.cpp b/lib/Conversion/TorchOnnxToTorch/DefaultDomainGtoP.cpp index d6b4e5bf10465..27ec21537bf87 100644 --- a/lib/Conversion/TorchOnnxToTorch/DefaultDomainGtoP.cpp +++ b/lib/Conversion/TorchOnnxToTorch/DefaultDomainGtoP.cpp @@ -264,10 +264,9 @@ void mlir::torch::onnx_c::populateDefaultDomainGtoP( for (uint64_t i = 1; i < operands.size(); i++) { result = rewriter.create( binder.getLoc(), resultType, result, operands[i]); - } - rewriter.replaceOp( - binder.op, result.getDefiningOp()); - return success(); + } + rewriter.replaceOp(binder.op, result.getDefiningOp()); + return success(); }); patterns.onOp("Min", 1, [](OpBinder binder, ConversionPatternRewriter &rewriter) { @@ -323,6 +322,154 @@ void mlir::torch::onnx_c::populateDefaultDomainGtoP( binder.op, resultType, lhs, rhs); return success(); }); + patterns.onOp( + "Gather", 13, [](OpBinder binder, ConversionPatternRewriter &rewriter) { + Torch::ValueTensorType resultType; + Value data, indices; + int64_t axis; + if (binder.tensorOperandAtIndex(data, 0) || + binder.tensorOperandAtIndex(indices, 1) || + binder.tensorResultType(resultType) || + binder.s64IntegerAttr(axis, "axis", 0)) + return failure(); + Location loc = binder.getLoc(); + + // 1. Get data shape and rank. + auto dataTensorType = data.getType().cast(); + if (!dataTensorType || !dataTensorType.hasSizes()) { + return rewriter.notifyMatchFailure(binder.op, + "Expect non empty input data"); + } + ArrayRef dataShape = dataTensorType.getSizes(); + unsigned dataRank = dataShape.size(); + + // 2. Get indices shape and rank. + auto indexType = indices.getType().cast(); + if (!indexType || !indexType.hasSizes()) { + return rewriter.notifyMatchFailure(binder.op, + "Expect non empty index tensor"); + } + ArrayRef indexShape = indexType.getSizes(); + unsigned indexRank = indexShape.size(); + + // 3. Compute total elements in the indices tensor, as we will collapse + // the indices tensor to a unary tensor. Also compute index shape and + // data shape tensors as they will be used for creating output types. + int64_t indexElemCount = 1; + for (int64_t dim : indexShape) { + if (dim == -1) { + indexElemCount = Torch::kUnknownSize; + break; + } + indexElemCount *= dim; + } + + Value constOne = rewriter.create( + loc, rewriter.getI64IntegerAttr(1)); + SmallVector indexShapeTensor; + Value indexElemCountVal = constOne; + for (unsigned i = 0; i < indexRank; ++i) { + Value indexDimVal = rewriter.create( + loc, indices, + rewriter.create( + loc, rewriter.getI64IntegerAttr(i))); + indexShapeTensor.emplace_back(indexDimVal); + indexElemCountVal = rewriter.create( + loc, indexElemCountVal, indexDimVal); + } + + SmallVector dataShapeTensor; + for (unsigned i = 0; i < dataRank; ++i) { + dataShapeTensor.emplace_back(rewriter.create( + loc, data, + rewriter.create( + loc, rewriter.getI64IntegerAttr(i)))); + } + + // 4. We can not directly perform torch.gather as the onnx.gather op + // collects the input data at different location of output compared to + // torch.gather op. The output of torch.gather and onnx.gather ops are + // indexed differently. + // check https://onnx.ai/onnx/operators/onnx__Gather.html for more + // details. So we will collapse indices tensor to a unary tensor and + // materialize to non-axis dimension of data tensor. For example, + // assuming indices is of shape (4, 5, 6), data is (8, 10, 11, 12) and + // axis=1. we will collapse indices into a (120,) unary tensor, + // materialize to non-axis dimension of data i.e. reshaping the unary + // indices tensor to (1, 120, 1, 1) and then perform the torch.gather + // operation. Now broadcast the output of gather operation to non-axis + // dimensions of data tensor. This would make the result of shape (8, + // 10, 120, 12). Post the broadcasting, expand the indices dimensions by + // reshaping (8, 10, 120, 12) to (8, 10, 4, 5, 6, 12) tensor, which is + // our expected final result. + SmallVector collapsedIndexShape(dataRank, 1); + collapsedIndexShape[axis] = indexElemCount; + Type collapsedIndexType = Torch::ValueTensorType::get( + indexType.getContext(), llvm::ArrayRef(collapsedIndexShape), + indexType.getOptionalDtype()); + + SmallVector collapsedIndexSize(dataRank, constOne); + collapsedIndexSize[axis] = indexElemCountVal; + auto collapsedIndexSizeList = + rewriter.create( + loc, + Torch::ListType::get(Torch::IntType::get(indices.getContext())), + collapsedIndexSize); + + auto collapsedIndices = rewriter.create( + loc, collapsedIndexType, indices, collapsedIndexSizeList); + + // 5. Compute gather result type and perform gather operation. + Type gatherResultType = Torch::ValueTensorType::get( + dataTensorType.getContext(), llvm::ArrayRef(collapsedIndexShape), + dataTensorType.getOptionalDtype()); + Value constAxis = rewriter.create( + binder.getLoc(), rewriter.getType(), + rewriter.getIntegerAttr(rewriter.getIntegerType(64), axis)); + Value constFalse = rewriter.create( + binder.getLoc(), rewriter.getType(), + rewriter.getBoolAttr(false)); + auto gatherOp = rewriter.create( + loc, gatherResultType, data, constAxis, collapsedIndices, + /*sparseGrad=*/constFalse); + + // 6. Broadcast the gather output to non-axis dimensions of data tensor. + SmallVector dataShapeVector(dataShape); + dataShapeVector[axis] = indexElemCount; + Type expandResultType = Torch::ValueTensorType::get( + dataTensorType.getContext(), llvm::ArrayRef(dataShapeVector), + dataTensorType.getOptionalDtype()); + + dataShapeTensor[axis] = indexElemCountVal; + auto expandSizeList = rewriter.create( + loc, Torch::ListType::get(Torch::IntType::get(data.getContext())), + dataShapeTensor); + auto expandedGather = rewriter.create( + loc, expandResultType, gatherOp, expandSizeList, + /*implicit=*/constFalse); + + // 7. Compute the result type of reshape op which expands the collapsed + // indices shapes back to the original indices shapes and reshape the + // output produced at step 6. This will produce our expected result of + // onnx.gather op. + SmallVector resultShapeTensor; + for (unsigned i = 0; i < dataRank; ++i) { + if (i == axis) { + resultShapeTensor.insert(resultShapeTensor.end(), + indexShapeTensor.begin(), + indexShapeTensor.end()); + continue; + } + resultShapeTensor.emplace_back(dataShapeTensor[i]); + } + auto resultSizeList = rewriter.create( + loc, Torch::ListType::get(Torch::IntType::get(data.getContext())), + resultShapeTensor); + + rewriter.replaceOpWithNewOp( + binder.op, resultType, expandedGather, resultSizeList); + return success(); + }); patterns.onOp( "GatherElements", 13, [](OpBinder binder, ConversionPatternRewriter &rewriter) { diff --git a/test/Conversion/TorchOnnxToTorch/simple_ops_g_to_p.mlir b/test/Conversion/TorchOnnxToTorch/simple_ops_g_to_p.mlir index 5d6b86172597f..b816fe0c6920e 100644 --- a/test/Conversion/TorchOnnxToTorch/simple_ops_g_to_p.mlir +++ b/test/Conversion/TorchOnnxToTorch/simple_ops_g_to_p.mlir @@ -37,6 +37,43 @@ func.func @test_less(%arg0: !torch.vtensor<[3,4,5],f32>, %arg1: !torch.vtensor<[ // ----- +// CHECK-LABEL: func.func @test_gather +func.func @test_gather(%arg0: !torch.vtensor<[3,4,5],f32>, %arg1: !torch.vtensor<[8,10,20,40], si64>) -> !torch.vtensor<[8,10,20,40,4,5],f32> attributes {torch.onnx_meta.opset_version = 13 : si64} { + // CHECK-DAG: %[[INT1:.+]] = torch.constant.int 1 + // CHECK-DAG: %[[INT0:.+]] = torch.constant.int 0 + // CHECK: %[[ARG1_SIZE0:.+]] = torch.aten.size.int %arg1, %[[INT0]] + // CHECK: %[[MUL1:.+]] = torch.aten.mul.int %[[INT1]], %[[ARG1_SIZE0]] + // CHECK: %[[INT1_0:.+]] = torch.constant.int 1 + // CHECK: %[[ARG1_SIZE1:.+]] = torch.aten.size.int %arg1, %[[INT1_0]] + // CHECK: %[[MUL2:.+]] = torch.aten.mul.int %[[MUL1]], %[[ARG1_SIZE1]] + // CHECK-DAG: %[[INT2:.+]] = torch.constant.int 2 + // CHECK: %[[ARG1_SIZE2:.+]] = torch.aten.size.int %arg1, %[[INT2]] : !torch.vtensor<[8,10,20,40],si64>, !torch.int -> !torch.int + // CHECK: %[[MUL3:.+]] = torch.aten.mul.int %[[MUL2]], %[[ARG1_SIZE2]] : !torch.int, !torch.int -> !torch.int + // CHECK-DAG: %[[INT3:.+]] = torch.constant.int 3 + // CHECK: %[[ARG1_SIZE3:.+]] = torch.aten.size.int %arg1, %[[INT3]] : !torch.vtensor<[8,10,20,40],si64>, !torch.int -> !torch.int + // CHECK: %[[MUL4:.+]] = torch.aten.mul.int %[[MUL3]], %[[ARG1_SIZE3]] : !torch.int, !torch.int -> !torch.int + // CHECK: %[[INT0_2:.+]] = torch.constant.int 0 + // CHECK: %[[ARG0_SIZE0:.+]] = torch.aten.size.int %arg0, %[[INT0_2]] : !torch.vtensor<[3,4,5],f32>, !torch.int -> !torch.int + // CHECK: %[[INT1_3:.+]] = torch.constant.int 1 + // CHECK: %[[ARG0_SIZE1:.+]] = torch.aten.size.int %arg0, %[[INT1_3]] : !torch.vtensor<[3,4,5],f32>, !torch.int -> !torch.int + // CHECK: %[[INT2_4:.+]] = torch.constant.int 2 + // CHECK: %[[ARG0_SIZE2:.+]] = torch.aten.size.int %arg0, %[[INT2_4]] : !torch.vtensor<[3,4,5],f32>, !torch.int -> !torch.int + // CHECK: %[[LIST1:.+]] = torch.prim.ListConstruct %[[MUL4]], %[[INT1]], %[[INT1]] : (!torch.int, !torch.int, !torch.int) -> !torch.list + // CHECK: %[[VIEW1:.+]] = torch.aten.view %arg1, %[[LIST1]] : !torch.vtensor<[8,10,20,40],si64>, !torch.list -> !torch.vtensor<[64000,1,1],si64> + // CHECK: %[[INT0_1:.+]] = torch.constant.int 0 + // CHECK: %[[FALSE:.+]] = torch.constant.bool false + // CHECK: %[[GATHER:.+]] = torch.aten.gather %arg0, %[[INT0_1]], %[[VIEW1]], %[[FALSE]] : !torch.vtensor<[3,4,5],f32>, !torch.int, !torch.vtensor<[64000,1,1],si64>, !torch.bool -> !torch.vtensor<[64000,1,1],f32> + // CHECK: %[[LIST2:.+]] = torch.prim.ListConstruct %[[MUL4]], %[[ARG0_SIZE1]], %[[ARG0_SIZE2]] : (!torch.int, !torch.int, !torch.int) -> !torch.list + // CHECK: %[[EXPAND:.+]] = torch.aten.expand %[[GATHER]], %[[LIST2]], %[[FALSE]] : !torch.vtensor<[64000,1,1],f32>, !torch.list, !torch.bool -> !torch.vtensor<[64000,4,5],f32> + // CHECK: %[[LIST3:.+]] = torch.prim.ListConstruct %[[ARG1_SIZE0]], %[[ARG1_SIZE1]], %[[ARG1_SIZE2]], %[[ARG1_SIZE3]], %[[ARG0_SIZE1]], %[[ARG0_SIZE2]] : (!torch.int, !torch.int, !torch.int, !torch.int, !torch.int, !torch.int) -> !torch.list + // CHECK: %[[RES:.+]] = torch.aten.view %[[EXPAND]], %[[LIST3]] : !torch.vtensor<[64000,4,5],f32>, !torch.list -> !torch.vtensor<[8,10,20,40,4,5],f32> + // CHECK: return %[[RES]] : !torch.vtensor<[8,10,20,40,4,5],f32> + %0 = torch.operator "onnx.Gather"(%arg0, %arg1) {torch.onnx.axis = 0 : si64} : (!torch.vtensor<[3,4,5],f32>, !torch.vtensor<[8,10,20,40], si64>) -> !torch.vtensor<[8,10,20,40,4,5],f32> + return %0 : !torch.vtensor<[8,10,20,40,4,5],f32> +} + +// ----- + // CHECK-LABEL: func.func @test_gather_elements func.func @test_gather_elements(%arg0: !torch.vtensor<[3,4,5],f32>, %arg1: !torch.vtensor<[3,4,5], si64>) -> !torch.vtensor<[3,4,5],f32> attributes {torch.onnx_meta.opset_version = 13 : si64} { // CHECK-DAG: %[[INT0:.+]] = torch.constant.int 0