From 87ee3e4f5438c567796e128b73eb7703aa56d2ec Mon Sep 17 00:00:00 2001 From: Zhangjingyu06 <92561254+Zhangjingyu06@users.noreply.github.com> Date: Fri, 14 Jan 2022 16:15:47 +0800 Subject: [PATCH] [XPU]add stack_grad op for kunlun2,*test=kunlun (#38674) * [XPU]add split op for kunlun2,*test=kunlun * [XPU]add split op for kunlun2,*test=kunlun * [XPU]add split op for kunlun,*test=kunlun * [XPU]add stack_grad op for kunlun2,*test=kunlun Co-authored-by: QingshuChen --- paddle/fluid/operators/stack_op_xpu.cc | 43 ++++++++++++++++--- .../fluid/platform/device/xpu/xpu1_op_list.h | 1 + .../fluid/platform/device/xpu/xpu2_op_list.h | 2 + .../tests/unittests/xpu/test_stack_op_xpu.py | 19 +++++++- 4 files changed, 58 insertions(+), 7 deletions(-) diff --git a/paddle/fluid/operators/stack_op_xpu.cc b/paddle/fluid/operators/stack_op_xpu.cc index 01ec4a2b16b..a2590e1180c 100644 --- a/paddle/fluid/operators/stack_op_xpu.cc +++ b/paddle/fluid/operators/stack_op_xpu.cc @@ -1,4 +1,4 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// 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. @@ -12,9 +12,12 @@ // See the License for the specific language governing permissions and // limitations under the License. +#ifdef PADDLE_WITH_XPU #include "paddle/fluid/operators/stack_op.h" #include -#ifdef PADDLE_WITH_XPU +#include +#include "paddle/fluid/operators/concat_op.h" +#include "paddle/fluid/platform/device/xpu/xpu_header.h" namespace paddle { namespace operators { @@ -59,14 +62,44 @@ class StackXPUKernel : public framework::OpKernel { } }; +template +class StackGradXPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* dy = ctx.Input(framework::GradVarName("Y")); + auto dx = ctx.MultiOutput(framework::GradVarName("X")); + auto axis = ctx.Attr("axis"); + auto& dev_ctx = ctx.template device_context(); + auto dy_dims = dy->dims(); + + if (axis < 0) axis += dy_dims.size() + 1; + auto dy_shape = framework::vectorize(dy_dims); + + std::vector dx_dims_list(dx.size(), 1); + std::vector dx_lists; + for (auto out : dx) { + dx_lists.push_back(out->mutable_data(ctx.GetPlace())); + } + + int r = xpu::split(dev_ctx.x_context(), dy->data(), dx_lists, + dy_shape, dx_dims_list, axis); + PADDLE_ENFORCE_EQ(r, XPU_SUCCESS, + platform::errors::External( + "The stack_grad XPU kernel return wrong value[%d %s]", + r, XPUAPIErrorMsg[r])); + } +}; + } // namespace operators } // namespace paddle namespace plat = paddle::platform; namespace ops = paddle::operators; - REGISTER_OP_XPU_KERNEL(stack, - ops::StackXPUKernel, + ops::StackXPUKernel, ops::StackXPUKernel, - ops::StackXPUKernel); + ops::StackXPUKernel); +REGISTER_OP_XPU_KERNEL(stack_grad, + ops::StackGradXPUKernel, + ops::StackGradXPUKernel); #endif diff --git a/paddle/fluid/platform/device/xpu/xpu1_op_list.h b/paddle/fluid/platform/device/xpu/xpu1_op_list.h index 26a1426bea0..a76bdd4ae96 100644 --- a/paddle/fluid/platform/device/xpu/xpu1_op_list.h +++ b/paddle/fluid/platform/device/xpu/xpu1_op_list.h @@ -300,6 +300,7 @@ XPUOpMap& get_kl1_ops() { pOpKernelType(vartype::UINT8, XPUPlace()), pOpKernelType(vartype::FP32, XPUPlace())})}, {"stack", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})}, + {"stack_grad", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})}, {"sum", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})}, {"tanh_grad", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})}, {"tanh", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})}, diff --git a/paddle/fluid/platform/device/xpu/xpu2_op_list.h b/paddle/fluid/platform/device/xpu/xpu2_op_list.h index 79261a5d7bc..3d140b4693a 100644 --- a/paddle/fluid/platform/device/xpu/xpu2_op_list.h +++ b/paddle/fluid/platform/device/xpu/xpu2_op_list.h @@ -333,6 +333,8 @@ XPUOpMap& get_kl2_ops() { {"stack", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace()), pOpKernelType(vartype::INT64, XPUPlace()), pOpKernelType(vartype::INT32, XPUPlace())})}, + {"stack_grad", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace()), + pOpKernelType(vartype::INT32, XPUPlace())})}, {"sum", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace()), pOpKernelType(vartype::FP16, XPUPlace())})}, {"tanh_grad", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace()), diff --git a/python/paddle/fluid/tests/unittests/xpu/test_stack_op_xpu.py b/python/paddle/fluid/tests/unittests/xpu/test_stack_op_xpu.py index 68e5a6ccdbf..20446aee41e 100644 --- a/python/paddle/fluid/tests/unittests/xpu/test_stack_op_xpu.py +++ b/python/paddle/fluid/tests/unittests/xpu/test_stack_op_xpu.py @@ -1,4 +1,4 @@ -# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved. +# 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. @@ -66,6 +66,15 @@ class TestStackOpBase(XPUOpTest): place = paddle.XPUPlace(0) self.check_output_with_place(place) + def test_check_grad(self): + if self.dtype == 'int64' or self.dtype == 'int32': + pass + else: + if paddle.is_compiled_with_xpu(): + paddle.enable_static() + place = paddle.XPUPlace(0) + self.check_grad_with_place(place, self.get_x_names(), 'Y') + class TestStackOp1(TestStackOpBase): def initParameters(self): @@ -81,11 +90,17 @@ class TestStackOp3(TestStackOpBase): def initParameters(self): self.axis = -1 + def test_check_grad(self): + pass + class TestStackOp4(TestStackOpBase): def initParameters(self): self.axis = -4 + def test_check_grad(self): + pass + class TestStackOp5(TestStackOpBase): def initParameters(self): @@ -113,7 +128,7 @@ class TestStackOpint(TestStackOpBase): self.num_inputs = 4 self.input_dim = (5, 6, 7) self.axis = 0 - self.dtype = 'int' + self.dtype = 'int32' def initParameters(self): self.num_inputs = 16 -- GitLab