From 46a1f69b3c0c4ec6808bbe9af9f5e436437b5612 Mon Sep 17 00:00:00 2001 From: pangyoki Date: Mon, 19 Oct 2020 19:22:22 +0800 Subject: [PATCH] cherry pick 27861 Add truncated_gaussian_random XPU kernel, test=kunlun (#28060) * Add truncated_gaussian_random_op XPU kernel * Add truncated_gaussian_random_op XPU kernel, test=kunlun * little change, test=kunlun * change boost_get to BOOST_GET_CONST * change boost_get to BOOST_GET_CONST, test=kunlun * little change, test=kunlun * use Generator to generate random number and optimize format, test=kunlun * little change, test=kunlun * add TODO, test=kunlun --- .../operators/truncated_gaussian_random_op.cc | 135 +-------------- .../operators/truncated_gaussian_random_op.h | 160 ++++++++++++++++++ .../truncated_gaussian_random_op_xpu.cc | 65 +++++++ .../test_truncated_gaussian_random_op_xpu.py | 39 +++++ 4 files changed, 266 insertions(+), 133 deletions(-) create mode 100644 paddle/fluid/operators/truncated_gaussian_random_op.h create mode 100644 paddle/fluid/operators/truncated_gaussian_random_op_xpu.cc create mode 100644 python/paddle/fluid/tests/unittests/xpu/test_truncated_gaussian_random_op_xpu.py diff --git a/paddle/fluid/operators/truncated_gaussian_random_op.cc b/paddle/fluid/operators/truncated_gaussian_random_op.cc index 419f0f7a2a5..b9f7ba30922 100644 --- a/paddle/fluid/operators/truncated_gaussian_random_op.cc +++ b/paddle/fluid/operators/truncated_gaussian_random_op.cc @@ -14,146 +14,15 @@ limitations under the License. */ #include #include +#include #include "paddle/fluid/framework/generator.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/truncated_gaussian_random_op.h" namespace paddle { namespace operators { -// reference: https://gist.github.com/lakshayg/d80172fe5ae3c5d2c2aedb53c250320e -template -T Erfinv(T x) { - if (x < -1 || x > 1) { - return std::numeric_limits::quiet_NaN(); - } else if (x == 1.0) { - return std::numeric_limits::infinity(); - } else if (x == -1.0) { - return -std::numeric_limits::infinity(); - } - - const T LN2 = 6.931471805599453094172321214581e-1; - - const T A0 = 1.1975323115670912564578e0; - const T A1 = 4.7072688112383978012285e1; - const T A2 = 6.9706266534389598238465e2; - const T A3 = 4.8548868893843886794648e3; - const T A4 = 1.6235862515167575384252e4; - const T A5 = 2.3782041382114385731252e4; - const T A6 = 1.1819493347062294404278e4; - const T A7 = 8.8709406962545514830200e2; - - const T B0 = 1.0000000000000000000e0; - const T B1 = 4.2313330701600911252e1; - const T B2 = 6.8718700749205790830e2; - const T B3 = 5.3941960214247511077e3; - const T B4 = 2.1213794301586595867e4; - const T B5 = 3.9307895800092710610e4; - const T B6 = 2.8729085735721942674e4; - const T B7 = 5.2264952788528545610e3; - - const T C0 = 1.42343711074968357734e0; - const T C1 = 4.63033784615654529590e0; - const T C2 = 5.76949722146069140550e0; - const T C3 = 3.64784832476320460504e0; - const T C4 = 1.27045825245236838258e0; - const T C5 = 2.41780725177450611770e-1; - const T C6 = 2.27238449892691845833e-2; - const T C7 = 7.74545014278341407640e-4; - - const T D0 = 1.4142135623730950488016887e0; - const T D1 = 2.9036514445419946173133295e0; - const T D2 = 2.3707661626024532365971225e0; - const T D3 = 9.7547832001787427186894837e-1; - const T D4 = 2.0945065210512749128288442e-1; - const T D5 = 2.1494160384252876777097297e-2; - const T D6 = 7.7441459065157709165577218e-4; - const T D7 = 1.4859850019840355905497876e-9; - - const T E0 = 6.65790464350110377720e0; - const T E1 = 5.46378491116411436990e0; - const T E2 = 1.78482653991729133580e0; - const T E3 = 2.96560571828504891230e-1; - const T E4 = 2.65321895265761230930e-2; - const T E5 = 1.24266094738807843860e-3; - const T E6 = 2.71155556874348757815e-5; - const T E7 = 2.01033439929228813265e-7; - - const T F0 = 1.414213562373095048801689e0; - const T F1 = 8.482908416595164588112026e-1; - const T F2 = 1.936480946950659106176712e-1; - const T F3 = 2.103693768272068968719679e-2; - const T F4 = 1.112800997078859844711555e-3; - const T F5 = 2.611088405080593625138020e-5; - const T F6 = 2.010321207683943062279931e-7; - const T F7 = 2.891024605872965461538222e-15; - - T abs_x = abs(x); - - if (abs_x <= 0.85) { - T r = 0.180625 - 0.25 * x * x; - T num = - (((((((A7 * r + A6) * r + A5) * r + A4) * r + A3) * r + A2) * r + A1) * - r + - A0); - T den = - (((((((B7 * r + B6) * r + B5) * r + B4) * r + B3) * r + B2) * r + B1) * - r + - B0); - return x * num / den; - } - - T r = sqrt(LN2 - log(1.0 - abs_x)); - - T num, den; - if (r <= 5.0) { - r = r - 1.6; - num = - (((((((C7 * r + C6) * r + C5) * r + C4) * r + C3) * r + C2) * r + C1) * - r + - C0); - den = - (((((((D7 * r + D6) * r + D5) * r + D4) * r + D3) * r + D2) * r + D1) * - r + - D0); - } else { - r = r - 5.0; - num = - (((((((E7 * r + E6) * r + E5) * r + E4) * r + E3) * r + E2) * r + E1) * - r + - E0); - den = - (((((((F7 * r + F6) * r + F5) * r + F4) * r + F3) * r + F2) * r + F1) * - r + - F0); - } - - if (x < 0) { - return -num / den; - } else { - return num / den; - } -} - -template -struct TruncatedNormal { - T mean, std; - T a_normal_cdf; - T b_normal_cdf; - TruncatedNormal(T mean, T std) : mean(mean), std(std) { - auto normal_cdf = [](T x) { - return (1.0 + std::erf(x / std::sqrt(2.0))) / 2.0; - }; - a_normal_cdf = normal_cdf(-2.0); - b_normal_cdf = normal_cdf(2.0); - } - - T operator()(T value) const { - auto p = a_normal_cdf + (b_normal_cdf - a_normal_cdf) * value; - return std::sqrt(2.0) * Erfinv(2 * p - 1) * std + mean; - } -}; - template class CPUTruncatedGaussianRandomKernel : public framework::OpKernel { public: diff --git a/paddle/fluid/operators/truncated_gaussian_random_op.h b/paddle/fluid/operators/truncated_gaussian_random_op.h new file mode 100644 index 00000000000..a6ff2f686cb --- /dev/null +++ b/paddle/fluid/operators/truncated_gaussian_random_op.h @@ -0,0 +1,160 @@ +/* Copyright (c) 2020 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. */ + +#pragma once + +#include +#include + +#include "paddle/fluid/framework/generator.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +// reference: https://gist.github.com/lakshayg/d80172fe5ae3c5d2c2aedb53c250320e +template +T Erfinv(T x) { + if (x < -1 || x > 1) { + return std::numeric_limits::quiet_NaN(); + } else if (x == 1.0) { + return std::numeric_limits::infinity(); + } else if (x == -1.0) { + return -std::numeric_limits::infinity(); + } + + const T LN2 = 6.931471805599453094172321214581e-1; + + const T A0 = 1.1975323115670912564578e0; + const T A1 = 4.7072688112383978012285e1; + const T A2 = 6.9706266534389598238465e2; + const T A3 = 4.8548868893843886794648e3; + const T A4 = 1.6235862515167575384252e4; + const T A5 = 2.3782041382114385731252e4; + const T A6 = 1.1819493347062294404278e4; + const T A7 = 8.8709406962545514830200e2; + + const T B0 = 1.0000000000000000000e0; + const T B1 = 4.2313330701600911252e1; + const T B2 = 6.8718700749205790830e2; + const T B3 = 5.3941960214247511077e3; + const T B4 = 2.1213794301586595867e4; + const T B5 = 3.9307895800092710610e4; + const T B6 = 2.8729085735721942674e4; + const T B7 = 5.2264952788528545610e3; + + const T C0 = 1.42343711074968357734e0; + const T C1 = 4.63033784615654529590e0; + const T C2 = 5.76949722146069140550e0; + const T C3 = 3.64784832476320460504e0; + const T C4 = 1.27045825245236838258e0; + const T C5 = 2.41780725177450611770e-1; + const T C6 = 2.27238449892691845833e-2; + const T C7 = 7.74545014278341407640e-4; + + const T D0 = 1.4142135623730950488016887e0; + const T D1 = 2.9036514445419946173133295e0; + const T D2 = 2.3707661626024532365971225e0; + const T D3 = 9.7547832001787427186894837e-1; + const T D4 = 2.0945065210512749128288442e-1; + const T D5 = 2.1494160384252876777097297e-2; + const T D6 = 7.7441459065157709165577218e-4; + const T D7 = 1.4859850019840355905497876e-9; + + const T E0 = 6.65790464350110377720e0; + const T E1 = 5.46378491116411436990e0; + const T E2 = 1.78482653991729133580e0; + const T E3 = 2.96560571828504891230e-1; + const T E4 = 2.65321895265761230930e-2; + const T E5 = 1.24266094738807843860e-3; + const T E6 = 2.71155556874348757815e-5; + const T E7 = 2.01033439929228813265e-7; + + const T F0 = 1.414213562373095048801689e0; + const T F1 = 8.482908416595164588112026e-1; + const T F2 = 1.936480946950659106176712e-1; + const T F3 = 2.103693768272068968719679e-2; + const T F4 = 1.112800997078859844711555e-3; + const T F5 = 2.611088405080593625138020e-5; + const T F6 = 2.010321207683943062279931e-7; + const T F7 = 2.891024605872965461538222e-15; + + T abs_x = abs(x); + + if (abs_x <= 0.85) { + T r = 0.180625 - 0.25 * x * x; + T num = + (((((((A7 * r + A6) * r + A5) * r + A4) * r + A3) * r + A2) * r + A1) * + r + + A0); + T den = + (((((((B7 * r + B6) * r + B5) * r + B4) * r + B3) * r + B2) * r + B1) * + r + + B0); + return x * num / den; + } + + T r = sqrt(LN2 - log(1.0 - abs_x)); + + T num, den; + if (r <= 5.0) { + r = r - 1.6; + num = + (((((((C7 * r + C6) * r + C5) * r + C4) * r + C3) * r + C2) * r + C1) * + r + + C0); + den = + (((((((D7 * r + D6) * r + D5) * r + D4) * r + D3) * r + D2) * r + D1) * + r + + D0); + } else { + r = r - 5.0; + num = + (((((((E7 * r + E6) * r + E5) * r + E4) * r + E3) * r + E2) * r + E1) * + r + + E0); + den = + (((((((F7 * r + F6) * r + F5) * r + F4) * r + F3) * r + F2) * r + F1) * + r + + F0); + } + + if (x < 0) { + return -num / den; + } else { + return num / den; + } +} + +template +struct TruncatedNormal { + T mean, std; + T a_normal_cdf; + T b_normal_cdf; + TruncatedNormal(T mean, T std) : mean(mean), std(std) { + auto normal_cdf = [](T x) { + return (1.0 + std::erf(x / std::sqrt(2.0))) / 2.0; + }; + a_normal_cdf = normal_cdf(-2.0); + b_normal_cdf = normal_cdf(2.0); + } + + T operator()(T value) const { + auto p = a_normal_cdf + (b_normal_cdf - a_normal_cdf) * value; + return std::sqrt(2.0) * Erfinv(2 * p - 1) * std + mean; + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/truncated_gaussian_random_op_xpu.cc b/paddle/fluid/operators/truncated_gaussian_random_op_xpu.cc new file mode 100644 index 00000000000..b2ff91a3745 --- /dev/null +++ b/paddle/fluid/operators/truncated_gaussian_random_op_xpu.cc @@ -0,0 +1,65 @@ +/* Copyright (c) 2020 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_XPU + +#include "paddle/fluid/operators/truncated_gaussian_random_op.h" +#include +#include +#include "paddle/fluid/framework/generator.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class XPUTruncatedGaussianRandomKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + float mean = context.Attr("mean"); + float std = context.Attr("std"); + auto* tensor = context.Output("Out"); + T* data = tensor->mutable_data(context.GetPlace()); + + std::uniform_real_distribution dist(std::numeric_limits::min(), + 1.0); + TruncatedNormal truncated_normal(mean, std); + int64_t size = tensor->numel(); + + unsigned int seed = static_cast(context.Attr("seed")); + // TODO(pangyoki): implement GetXPURandomEngine to set different seeds on + // corresponding XPU device. + auto engine = framework::GetCPURandomEngine(seed); + + std::unique_ptr data_cpu(new T[size]); + + for (int64_t i = 0; i < size; ++i) { + data_cpu[i] = truncated_normal(dist(*engine)); + } + + memory::Copy(BOOST_GET_CONST(platform::XPUPlace, context.GetPlace()), data, + platform::CPUPlace(), reinterpret_cast(data_cpu.get()), + size * sizeof(T)); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_XPU_KERNEL(truncated_gaussian_random, + ops::XPUTruncatedGaussianRandomKernel< + paddle::platform::XPUDeviceContext, float>); + +#endif // PADDLE_WITH_XPU diff --git a/python/paddle/fluid/tests/unittests/xpu/test_truncated_gaussian_random_op_xpu.py b/python/paddle/fluid/tests/unittests/xpu/test_truncated_gaussian_random_op_xpu.py new file mode 100644 index 00000000000..d096cb8ec13 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/xpu/test_truncated_gaussian_random_op_xpu.py @@ -0,0 +1,39 @@ +# Copyright (c) 2020 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 sys +sys.path.append("..") +import unittest +import numpy + +import paddle +import paddle.fluid as fluid +import paddle.fluid.core as core +from paddle.fluid.op import Operator +from paddle.fluid.executor import Executor +from test_truncated_gaussian_random_op import TestTrunctedGaussianRandomOp + +paddle.enable_static() + + +class TestXPUTrunctedGaussianRandomOp(TestTrunctedGaussianRandomOp): + def test_xpu(self): + if paddle.is_compiled_with_xpu(): + self.gaussian_random_test(place=fluid.XPUPlace(0)) + + +if __name__ == "__main__": + unittest.main() -- GitLab