From 57220f594db6bc5a7fbfae47d82fb55e9bb0862a Mon Sep 17 00:00:00 2001 From: oyxuan-11 <963650125@qq.com> Date: Mon, 15 Mar 2021 15:52:40 +0800 Subject: [PATCH] [NPU] Support NPU kernel cast op (#31635) Co-authored-by: frankwhzhang --- paddle/fluid/operators/cast_op_npu.cc | 91 +++++++++++++++++++ .../tests/unittests/npu/test_cast_op_npu.py | 75 +++++++++++++++ 2 files changed, 166 insertions(+) create mode 100755 paddle/fluid/operators/cast_op_npu.cc create mode 100755 python/paddle/fluid/tests/unittests/npu/test_cast_op_npu.py diff --git a/paddle/fluid/operators/cast_op_npu.cc b/paddle/fluid/operators/cast_op_npu.cc new file mode 100755 index 0000000000..5cfb152684 --- /dev/null +++ b/paddle/fluid/operators/cast_op_npu.cc @@ -0,0 +1,91 @@ +/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifdef PADDLE_WITH_ASCEND_CL +#include +#include + +#include "paddle/fluid/operators/cast_op.h" +#include "paddle/fluid/operators/npu_op_runner.h" + +namespace paddle { +namespace operators { + +static std::map + DTYPE_2_ACL_DTYPE = { + {framework::proto::VarType::BOOL, ACL_BOOL}, + {framework::proto::VarType::INT16, ACL_INT16}, + {framework::proto::VarType::INT32, ACL_INT32}, + {framework::proto::VarType::INT64, ACL_INT64}, + {framework::proto::VarType::FP16, ACL_FLOAT16}, + {framework::proto::VarType::FP32, ACL_FLOAT}, + {framework::proto::VarType::FP64, ACL_DOUBLE}, +}; + +using Tensor = framework::Tensor; + +template +class CastNPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + int dtype = ctx.Attr("out_dtype"); + + auto* out = ctx.Output("Out"); + + auto place = ctx.GetPlace(); + + auto iter = DTYPE_2_ACL_DTYPE.find(static_cast(dtype)); + int aclDtype = iter->second; + + if (dtype == framework::proto::VarType::FP32) { + out->mutable_data(place); + } else if (dtype == framework::proto::VarType::FP16) { + out->mutable_data(place); + } else if (dtype == framework::proto::VarType::INT16) { + out->mutable_data(place); + } else if (dtype == framework::proto::VarType::INT32) { + out->mutable_data(place); + } else if (dtype == framework::proto::VarType::INT64) { + out->mutable_data(place); + } else if (dtype == framework::proto::VarType::FP64) { + out->mutable_data(place); + } else if (dtype == framework::proto::VarType::BOOL) { + out->mutable_data(place); + } + + auto stream = + ctx.template device_context() + .stream(); + + auto runner = NpuOpRunner("Cast", {*x}, {*out}, {{"dst_type", static_cast(aclDtype)}}); + runner.Run(stream); + } +}; +} // namespace operators +} // namespace paddleaclDtype + +namespace ops = paddle::operators; + +REGISTER_OP_NPU_KERNEL( + cast, + ops::CastNPUKernel, + ops::CastNPUKernel, + ops::CastNPUKernel, + ops::CastNPUKernel, + ops::CastNPUKernel, + ops::CastNPUKernel, + ops::CastNPUKernel); +#endif diff --git a/python/paddle/fluid/tests/unittests/npu/test_cast_op_npu.py b/python/paddle/fluid/tests/unittests/npu/test_cast_op_npu.py new file mode 100755 index 0000000000..97a10f6657 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/npu/test_cast_op_npu.py @@ -0,0 +1,75 @@ +# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import numpy as np +import unittest +import sys +sys.path.append("..") +from op_test import OpTest +import paddle +import paddle.fluid as fluid +import paddle.fluid.core as core + +paddle.enable_static() +SEED = 2021 + + +@unittest.skipIf(not paddle.is_compiled_with_npu(), + "core is not compiled with NPU") +class TestCast1(OpTest): + def setUp(self): + self.set_npu() + self.op_type = "cast" + self.place = paddle.NPUPlace(0) + + ipt = np.random.random(size=[10, 10]) + 1 + self.inputs = {'X': ipt.astype('float32')} + self.outputs = {'Out': ipt.astype('float16')} + + self.attrs = { + 'in_dtype': int(core.VarDesc.VarType.FP32), + 'out_dtype': int(core.VarDesc.VarType.FP16) + } + + def set_npu(self): + self.__class__.use_npu = True + + def test_check_output(self): + self.check_output_with_place(self.place, check_dygraph=False) + +class TestCast2(OpTest): + def setUp(self): + self.set_npu() + self.op_type = "cast" + self.place = paddle.NPUPlace(0) + + ipt = np.random.random(size=[10, 10]) + 1 + self.inputs = {'X': ipt.astype('float16')} + self.outputs = {'Out': ipt.astype('float32')} + + self.attrs = { + 'in_dtype': int(core.VarDesc.VarType.FP16), + 'out_dtype': int(core.VarDesc.VarType.FP32) + } + + def set_npu(self): + self.__class__.use_npu = True + + def test_check_output(self): + self.check_output_with_place(self.place, check_dygraph=False, atol=1e-3) + +if __name__ == '__main__': + unittest.main() -- GitLab