/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve. 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. */ #pragma once /** * TestUtils.h is used to automatically compare CPU and GPU code is consistent. * This file provides a class(AutoCompare) and a template * function(BaseMatrixCompare) to simplify the comparison * of CPU and GPU member functions. * Refer test_Matrix.cpp and test_BaseMatrix.cpp for how to use autotest. */ #include #include "paddle/math/Matrix.h" #include "paddle/math/SparseMatrix.h" #include "TensorCheck.h" namespace autotest { using paddle::BaseMatrix; using paddle::CpuIVector; using paddle::GpuIVector; using paddle::CpuSparseMatrix; using paddle::GpuSparseMatrix; template class ReplaceType { public: typedef T1 type; }; template <> class ReplaceType { public: typedef CpuMatrix type; }; template <> class ReplaceType { public: typedef GpuMatrix type; }; template <> class ReplaceType { public: typedef CpuMatrix type; }; template <> class ReplaceType { public: typedef GpuMatrix type; }; // construct a argument template T construct(int height, int width); template <> float construct(int height, int width) { return 0.5; } template <> size_t construct(int height, int width) { size_t offset = std::rand() % (height < width ? height : width); return offset; } template <> CpuMatrix construct(int height, int width) { CpuMatrix a(height, width); return a; } template <> GpuMatrix construct(int height, int width) { GpuMatrix a(height, width); return a; } // init a argument template void init(T& v) { return; } template <> void init(CpuMatrix& v) { v.randomizeUniform(); } template <> void init(GpuMatrix& v) { v.randomizeUniform(); } // init a tuple which contains a set of arguments. template inline typename std::enable_if::type initTuple( std::tuple& t) {} template inline typename std::enable_if < I::type initTuple(std::tuple& t) { init(std::get(t)); initTuple(t); } // copy a argument, copy src to dest template void copy(T1& dest, T2& src) { dest = src; } template <> void copy(GpuMatrix& dest, CpuMatrix& src) { dest.copyFrom(src); } // copy a tuple, copy src to dest template inline typename std::enable_if::type copyTuple( std::tuple& dest, std::tuple& src) {} template inline typename std::enable_if < I::type copyTuple(std::tuple& dest, std::tuple& src) { copy(std::get(dest), std::get(src)); copyTuple(dest, src); } // call member function template R call(C& obj, R (FC::*f)(FArgs...), Args&&... args) { return (obj.*f)(args...); } template void BaseMatrixCompare(R (C::*f)(Args...), AssertEq compare) { for (auto height : {1, 11, 73, 128, 200, 330}) { for (auto width : {1, 3, 32, 100, 512, 1000}) { CpuMatrix obj1(AsRowVector ? 1 : height, AsColVector ? 1 : width); GpuMatrix obj2(AsRowVector ? 1 : height, AsColVector ? 1 : width); init(obj1); copy(obj2, obj1); auto tuple1 = std::make_tuple( construct>::type>::type, CpuMatrix>::type>(height, width)...); auto tuple2 = std::make_tuple( construct>::type>::type, GpuMatrix>::type>(height, width)...); initTuple(tuple1); copyTuple(tuple2, tuple1); call(obj1, f, std::get(tuple1)...); call(obj2, f, std::get(tuple2)...); TensorCheck(compare, obj1, obj2); } } } // AutoCompare template class ReturnType { public: typedef T type; }; template <> class ReturnType { public: typedef GpuMatrix type; }; template <> class ReturnType { public: typedef GpuIVector type; }; template <> class ReturnType { public: typedef GpuSparseMatrix type; }; template typename ReturnType::type autoArgs(T& v) { return v; } template <> GpuMatrix autoArgs(CpuMatrix& v) { GpuMatrix a(v.getHeight(), v.getWidth()); a.copyFrom(v); return a; } template <> GpuIVector autoArgs(CpuIVector& v) { GpuIVector a(v.getSize()); a.copyFrom(v); return a; } template <> GpuSparseMatrix autoArgs(CpuSparseMatrix& v) { GpuSparseMatrix a(v.getHeight(), v.getWidth(), v.getElementCnt(), v.getValueType(), v.getFormat()); a.copyFrom(v, HPPL_STREAM_DEFAULT); hl_stream_synchronize(HPPL_STREAM_DEFAULT); return a; } class AutoCompare { public: AutoCompare(size_t height, size_t width) : cpu(height, width), gpu(height, width) { init(cpu); copy(gpu, cpu); } template void operator()(R (C::*f)(FArgs...), Args&&... args) { call(cpu, f, args...); call(gpu, f, autoArgs(args)...); TensorCheckErr(cpu, gpu); } protected: CpuMatrix cpu; GpuMatrix gpu; }; } // namespace autotest template void BaseMatrixCompare(R (C::*f)(Args...)) { static_assert(sizeof...(I) == sizeof...(Args), "size of parameter packs are not equal"); #ifndef PADDLE_TYPE_DOUBLE autotest::AssertEqual compare(1e-5); #else autotest::AssertEqual compare(1e-10); #endif autotest::BaseMatrixCompare(f, compare); } template void BaseMatrixAsColVector(R (C::*f)(Args...)) { static_assert(sizeof...(I) == sizeof...(Args), "size of parameter packs are not equal"); #ifndef PADDLE_TYPE_DOUBLE autotest::AssertEqual compare(1e-3); #else autotest::AssertEqual compare(1e-8); #endif autotest::BaseMatrixCompare(f, compare); } template void BaseMatrixAsRowVector(R (C::*f)(Args...)) { static_assert(sizeof...(I) == sizeof...(Args), "size of parameter packs are not equal"); #ifndef PADDLE_TYPE_DOUBLE autotest::AssertEqual compare(1e-3); #else autotest::AssertEqual compare(1e-8); #endif autotest::BaseMatrixCompare(f, compare); }