Abseil Common Libraries (C++) (grcp 依赖) https://abseil.io/
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

160 lines
5.5 KiB

Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
//
// Copyright 2017 The Abseil Authors.
//
// 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
//
// https://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.
//
// -----------------------------------------------------------------------------
// File: cycleclock.h
// -----------------------------------------------------------------------------
//
// This header file defines a `CycleClock`, which yields the value and frequency
// of a cycle counter that increments at a rate that is approximately constant.
//
// NOTE:
//
// The cycle counter frequency is not necessarily related to the core clock
// frequency and should not be treated as such. That is, `CycleClock` cycles are
// not necessarily "CPU cycles" and code should not rely on that behavior, even
// if experimentally observed.
//
// An arbitrary offset may have been added to the counter at power on.
//
// On some platforms, the rate and offset of the counter may differ
// slightly when read from different CPUs of a multiprocessor. Usually,
// we try to ensure that the operating system adjusts values periodically
// so that values agree approximately. If you need stronger guarantees,
// consider using alternate interfaces.
//
// The CPU is not required to maintain the ordering of a cycle counter read
// with respect to surrounding instructions.
#ifndef ABSL_BASE_INTERNAL_CYCLECLOCK_H_
#define ABSL_BASE_INTERNAL_CYCLECLOCK_H_
Export of internal Abseil changes -- d3b99682554d339c42556680f4d65f83226005e2 by Martijn Vels <mvels@google.com>: Inline CycleClock code and remove branch for x86 CycleClockSource function This CL removes the relaxed load for x86 as there is no acquire price to pay on x86. It inlines the UnscaledCycleClock::Now() which is a single RTDSC op for x86, and likewise inlines CycleClock::Now() for x86. The inlining should mostly have secondary benefits such as reducing spills on outlined calls. LTO may eventually hoist these functions inline for the hotspots, but it doesn't hurt to default inline this for all builds and let the compiler decide on the first pass. The perlab benchmark is noisy for the plain BM_Now, but the other benchmarks and the run on my local machine are clear. ------------- Local Benchy Benchmark name old cpu/op new cpu/op delta BM_Now 3.41ns ± 1% 2.30ns ± 2% -32.52% (p=0.000 n=50+50) BM_NowWithRegisterPresure 4.96ns ± 2% 4.19ns ± 2% -15.57% (p=0.000 n=56+55) BM_NowWithCallback 3.30ns ± 2% 1.91ns ± 2% -42.00% (p=0.000 n=47+60) ------------- Perflab Benchy Benchmark name old cpu/op new cpu/op delta BM_Now 8.20ns ±13% 4.32ns ±83% ~ (p=0.413 n=4+5) BM_NowWithRegisterPresure 7.91ns ± 1% 3.68ns ± 2% -53.45% (p=0.029 n=4+4) BM_NowWithCallback 2.66ns ±13% 1.58ns ± 0% -40.51% (p=0.008 n=5+5) PiperOrigin-RevId: 434474766 Change-Id: I991d987ae9233e50f09606c874055cf4c5a56300 -- b38330686a0af176a2679163e4d2fa1b90e2f667 by Laramie Leavitt <lar@google.com>: Style, comment, and test updates * Remove a redundant assert in uniform_real_distribution. * Update comment in internal/generate_real.h * Style updates to uniform_real_distribution_test mainly replacing TypeParam with real_type, using aliases for some limits, etc. * Add a few more minor tests. PiperOrigin-RevId: 433902174 Change-Id: Id75be8e24be2fb8f6aea05feec13e3ef320a7254 -- ab2da6047ff7f5dae3add3779fcddf73b03feabf by Abseil Team <absl-team@google.com>: Remove declaration of method whose definition was previously removed. PiperOrigin-RevId: 433507828 Change-Id: I0130b689813125250f7de2664e767e181f676c89 -- df0c87f4ec2c010691931c1bef9d26470a6e63a2 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 433289136 Change-Id: Iba157dc83ed99dafd17a2223d2504e49f8afbb9e -- 7445fa312f2995772900eda82467325b3401a17d by Martijn Vels <mvels@google.com>: Optimize CordReader logic now that CONCAT is removed This CL cleans up various helper functions and logic remaining from previous complex CONCAT logic that is no longer needed, simplifying the CordReader logic. PiperOrigin-RevId: 433208748 Change-Id: I5f7b1883573c44e7c6f8af12c3cddbd197cb134d GitOrigin-RevId: d3b99682554d339c42556680f4d65f83226005e2
3 years ago
#include <atomic>
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
#include <cstdint>
Export of internal Abseil changes -- d3b99682554d339c42556680f4d65f83226005e2 by Martijn Vels <mvels@google.com>: Inline CycleClock code and remove branch for x86 CycleClockSource function This CL removes the relaxed load for x86 as there is no acquire price to pay on x86. It inlines the UnscaledCycleClock::Now() which is a single RTDSC op for x86, and likewise inlines CycleClock::Now() for x86. The inlining should mostly have secondary benefits such as reducing spills on outlined calls. LTO may eventually hoist these functions inline for the hotspots, but it doesn't hurt to default inline this for all builds and let the compiler decide on the first pass. The perlab benchmark is noisy for the plain BM_Now, but the other benchmarks and the run on my local machine are clear. ------------- Local Benchy Benchmark name old cpu/op new cpu/op delta BM_Now 3.41ns ± 1% 2.30ns ± 2% -32.52% (p=0.000 n=50+50) BM_NowWithRegisterPresure 4.96ns ± 2% 4.19ns ± 2% -15.57% (p=0.000 n=56+55) BM_NowWithCallback 3.30ns ± 2% 1.91ns ± 2% -42.00% (p=0.000 n=47+60) ------------- Perflab Benchy Benchmark name old cpu/op new cpu/op delta BM_Now 8.20ns ±13% 4.32ns ±83% ~ (p=0.413 n=4+5) BM_NowWithRegisterPresure 7.91ns ± 1% 3.68ns ± 2% -53.45% (p=0.029 n=4+4) BM_NowWithCallback 2.66ns ±13% 1.58ns ± 0% -40.51% (p=0.008 n=5+5) PiperOrigin-RevId: 434474766 Change-Id: I991d987ae9233e50f09606c874055cf4c5a56300 -- b38330686a0af176a2679163e4d2fa1b90e2f667 by Laramie Leavitt <lar@google.com>: Style, comment, and test updates * Remove a redundant assert in uniform_real_distribution. * Update comment in internal/generate_real.h * Style updates to uniform_real_distribution_test mainly replacing TypeParam with real_type, using aliases for some limits, etc. * Add a few more minor tests. PiperOrigin-RevId: 433902174 Change-Id: Id75be8e24be2fb8f6aea05feec13e3ef320a7254 -- ab2da6047ff7f5dae3add3779fcddf73b03feabf by Abseil Team <absl-team@google.com>: Remove declaration of method whose definition was previously removed. PiperOrigin-RevId: 433507828 Change-Id: I0130b689813125250f7de2664e767e181f676c89 -- df0c87f4ec2c010691931c1bef9d26470a6e63a2 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 433289136 Change-Id: Iba157dc83ed99dafd17a2223d2504e49f8afbb9e -- 7445fa312f2995772900eda82467325b3401a17d by Martijn Vels <mvels@google.com>: Optimize CordReader logic now that CONCAT is removed This CL cleans up various helper functions and logic remaining from previous complex CONCAT logic that is no longer needed, simplifying the CordReader logic. PiperOrigin-RevId: 433208748 Change-Id: I5f7b1883573c44e7c6f8af12c3cddbd197cb134d GitOrigin-RevId: d3b99682554d339c42556680f4d65f83226005e2
3 years ago
#include "absl/base/attributes.h"
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
#include "absl/base/config.h"
Export of internal Abseil changes -- d3b99682554d339c42556680f4d65f83226005e2 by Martijn Vels <mvels@google.com>: Inline CycleClock code and remove branch for x86 CycleClockSource function This CL removes the relaxed load for x86 as there is no acquire price to pay on x86. It inlines the UnscaledCycleClock::Now() which is a single RTDSC op for x86, and likewise inlines CycleClock::Now() for x86. The inlining should mostly have secondary benefits such as reducing spills on outlined calls. LTO may eventually hoist these functions inline for the hotspots, but it doesn't hurt to default inline this for all builds and let the compiler decide on the first pass. The perlab benchmark is noisy for the plain BM_Now, but the other benchmarks and the run on my local machine are clear. ------------- Local Benchy Benchmark name old cpu/op new cpu/op delta BM_Now 3.41ns ± 1% 2.30ns ± 2% -32.52% (p=0.000 n=50+50) BM_NowWithRegisterPresure 4.96ns ± 2% 4.19ns ± 2% -15.57% (p=0.000 n=56+55) BM_NowWithCallback 3.30ns ± 2% 1.91ns ± 2% -42.00% (p=0.000 n=47+60) ------------- Perflab Benchy Benchmark name old cpu/op new cpu/op delta BM_Now 8.20ns ±13% 4.32ns ±83% ~ (p=0.413 n=4+5) BM_NowWithRegisterPresure 7.91ns ± 1% 3.68ns ± 2% -53.45% (p=0.029 n=4+4) BM_NowWithCallback 2.66ns ±13% 1.58ns ± 0% -40.51% (p=0.008 n=5+5) PiperOrigin-RevId: 434474766 Change-Id: I991d987ae9233e50f09606c874055cf4c5a56300 -- b38330686a0af176a2679163e4d2fa1b90e2f667 by Laramie Leavitt <lar@google.com>: Style, comment, and test updates * Remove a redundant assert in uniform_real_distribution. * Update comment in internal/generate_real.h * Style updates to uniform_real_distribution_test mainly replacing TypeParam with real_type, using aliases for some limits, etc. * Add a few more minor tests. PiperOrigin-RevId: 433902174 Change-Id: Id75be8e24be2fb8f6aea05feec13e3ef320a7254 -- ab2da6047ff7f5dae3add3779fcddf73b03feabf by Abseil Team <absl-team@google.com>: Remove declaration of method whose definition was previously removed. PiperOrigin-RevId: 433507828 Change-Id: I0130b689813125250f7de2664e767e181f676c89 -- df0c87f4ec2c010691931c1bef9d26470a6e63a2 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 433289136 Change-Id: Iba157dc83ed99dafd17a2223d2504e49f8afbb9e -- 7445fa312f2995772900eda82467325b3401a17d by Martijn Vels <mvels@google.com>: Optimize CordReader logic now that CONCAT is removed This CL cleans up various helper functions and logic remaining from previous complex CONCAT logic that is no longer needed, simplifying the CordReader logic. PiperOrigin-RevId: 433208748 Change-Id: I5f7b1883573c44e7c6f8af12c3cddbd197cb134d GitOrigin-RevId: d3b99682554d339c42556680f4d65f83226005e2
3 years ago
#include "absl/base/internal/unscaledcycleclock.h"
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
namespace absl {
ABSL_NAMESPACE_BEGIN
namespace base_internal {
Export of internal Abseil changes -- d3b99682554d339c42556680f4d65f83226005e2 by Martijn Vels <mvels@google.com>: Inline CycleClock code and remove branch for x86 CycleClockSource function This CL removes the relaxed load for x86 as there is no acquire price to pay on x86. It inlines the UnscaledCycleClock::Now() which is a single RTDSC op for x86, and likewise inlines CycleClock::Now() for x86. The inlining should mostly have secondary benefits such as reducing spills on outlined calls. LTO may eventually hoist these functions inline for the hotspots, but it doesn't hurt to default inline this for all builds and let the compiler decide on the first pass. The perlab benchmark is noisy for the plain BM_Now, but the other benchmarks and the run on my local machine are clear. ------------- Local Benchy Benchmark name old cpu/op new cpu/op delta BM_Now 3.41ns ± 1% 2.30ns ± 2% -32.52% (p=0.000 n=50+50) BM_NowWithRegisterPresure 4.96ns ± 2% 4.19ns ± 2% -15.57% (p=0.000 n=56+55) BM_NowWithCallback 3.30ns ± 2% 1.91ns ± 2% -42.00% (p=0.000 n=47+60) ------------- Perflab Benchy Benchmark name old cpu/op new cpu/op delta BM_Now 8.20ns ±13% 4.32ns ±83% ~ (p=0.413 n=4+5) BM_NowWithRegisterPresure 7.91ns ± 1% 3.68ns ± 2% -53.45% (p=0.029 n=4+4) BM_NowWithCallback 2.66ns ±13% 1.58ns ± 0% -40.51% (p=0.008 n=5+5) PiperOrigin-RevId: 434474766 Change-Id: I991d987ae9233e50f09606c874055cf4c5a56300 -- b38330686a0af176a2679163e4d2fa1b90e2f667 by Laramie Leavitt <lar@google.com>: Style, comment, and test updates * Remove a redundant assert in uniform_real_distribution. * Update comment in internal/generate_real.h * Style updates to uniform_real_distribution_test mainly replacing TypeParam with real_type, using aliases for some limits, etc. * Add a few more minor tests. PiperOrigin-RevId: 433902174 Change-Id: Id75be8e24be2fb8f6aea05feec13e3ef320a7254 -- ab2da6047ff7f5dae3add3779fcddf73b03feabf by Abseil Team <absl-team@google.com>: Remove declaration of method whose definition was previously removed. PiperOrigin-RevId: 433507828 Change-Id: I0130b689813125250f7de2664e767e181f676c89 -- df0c87f4ec2c010691931c1bef9d26470a6e63a2 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 433289136 Change-Id: Iba157dc83ed99dafd17a2223d2504e49f8afbb9e -- 7445fa312f2995772900eda82467325b3401a17d by Martijn Vels <mvels@google.com>: Optimize CordReader logic now that CONCAT is removed This CL cleans up various helper functions and logic remaining from previous complex CONCAT logic that is no longer needed, simplifying the CordReader logic. PiperOrigin-RevId: 433208748 Change-Id: I5f7b1883573c44e7c6f8af12c3cddbd197cb134d GitOrigin-RevId: d3b99682554d339c42556680f4d65f83226005e2
3 years ago
using CycleClockSourceFunc = int64_t (*)();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
// -----------------------------------------------------------------------------
// CycleClock
// -----------------------------------------------------------------------------
class CycleClock {
public:
// CycleClock::Now()
//
// Returns the value of a cycle counter that counts at a rate that is
// approximately constant.
static int64_t Now();
// CycleClock::Frequency()
//
// Returns the amount by which `CycleClock::Now()` increases per second. Note
// that this value may not necessarily match the core CPU clock frequency.
static double Frequency();
private:
Export of internal Abseil changes -- d3b99682554d339c42556680f4d65f83226005e2 by Martijn Vels <mvels@google.com>: Inline CycleClock code and remove branch for x86 CycleClockSource function This CL removes the relaxed load for x86 as there is no acquire price to pay on x86. It inlines the UnscaledCycleClock::Now() which is a single RTDSC op for x86, and likewise inlines CycleClock::Now() for x86. The inlining should mostly have secondary benefits such as reducing spills on outlined calls. LTO may eventually hoist these functions inline for the hotspots, but it doesn't hurt to default inline this for all builds and let the compiler decide on the first pass. The perlab benchmark is noisy for the plain BM_Now, but the other benchmarks and the run on my local machine are clear. ------------- Local Benchy Benchmark name old cpu/op new cpu/op delta BM_Now 3.41ns ± 1% 2.30ns ± 2% -32.52% (p=0.000 n=50+50) BM_NowWithRegisterPresure 4.96ns ± 2% 4.19ns ± 2% -15.57% (p=0.000 n=56+55) BM_NowWithCallback 3.30ns ± 2% 1.91ns ± 2% -42.00% (p=0.000 n=47+60) ------------- Perflab Benchy Benchmark name old cpu/op new cpu/op delta BM_Now 8.20ns ±13% 4.32ns ±83% ~ (p=0.413 n=4+5) BM_NowWithRegisterPresure 7.91ns ± 1% 3.68ns ± 2% -53.45% (p=0.029 n=4+4) BM_NowWithCallback 2.66ns ±13% 1.58ns ± 0% -40.51% (p=0.008 n=5+5) PiperOrigin-RevId: 434474766 Change-Id: I991d987ae9233e50f09606c874055cf4c5a56300 -- b38330686a0af176a2679163e4d2fa1b90e2f667 by Laramie Leavitt <lar@google.com>: Style, comment, and test updates * Remove a redundant assert in uniform_real_distribution. * Update comment in internal/generate_real.h * Style updates to uniform_real_distribution_test mainly replacing TypeParam with real_type, using aliases for some limits, etc. * Add a few more minor tests. PiperOrigin-RevId: 433902174 Change-Id: Id75be8e24be2fb8f6aea05feec13e3ef320a7254 -- ab2da6047ff7f5dae3add3779fcddf73b03feabf by Abseil Team <absl-team@google.com>: Remove declaration of method whose definition was previously removed. PiperOrigin-RevId: 433507828 Change-Id: I0130b689813125250f7de2664e767e181f676c89 -- df0c87f4ec2c010691931c1bef9d26470a6e63a2 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 433289136 Change-Id: Iba157dc83ed99dafd17a2223d2504e49f8afbb9e -- 7445fa312f2995772900eda82467325b3401a17d by Martijn Vels <mvels@google.com>: Optimize CordReader logic now that CONCAT is removed This CL cleans up various helper functions and logic remaining from previous complex CONCAT logic that is no longer needed, simplifying the CordReader logic. PiperOrigin-RevId: 433208748 Change-Id: I5f7b1883573c44e7c6f8af12c3cddbd197cb134d GitOrigin-RevId: d3b99682554d339c42556680f4d65f83226005e2
3 years ago
#if ABSL_USE_UNSCALED_CYCLECLOCK
static CycleClockSourceFunc LoadCycleClockSource();
#ifdef NDEBUG
#ifdef ABSL_INTERNAL_UNSCALED_CYCLECLOCK_FREQUENCY_IS_CPU_FREQUENCY
// Not debug mode and the UnscaledCycleClock frequency is the CPU
// frequency. Scale the CycleClock to prevent overflow if someone
// tries to represent the time as cycles since the Unix epoch.
static constexpr int32_t kShift = 1;
#else
// Not debug mode and the UnscaledCycleClock isn't operating at the
// raw CPU frequency. There is no need to do any scaling, so don't
// needlessly sacrifice precision.
static constexpr int32_t kShift = 0;
#endif
#else // NDEBUG
// In debug mode use a different shift to discourage depending on a
// particular shift value.
static constexpr int32_t kShift = 2;
#endif // NDEBUG
static constexpr double kFrequencyScale = 1.0 / (1 << kShift);
ABSL_CONST_INIT static std::atomic<CycleClockSourceFunc> cycle_clock_source_;
#endif // ABSL_USE_UNSCALED_CYCLECLOC
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
CycleClock() = delete; // no instances
CycleClock(const CycleClock&) = delete;
CycleClock& operator=(const CycleClock&) = delete;
Export of internal Abseil changes -- d3b99682554d339c42556680f4d65f83226005e2 by Martijn Vels <mvels@google.com>: Inline CycleClock code and remove branch for x86 CycleClockSource function This CL removes the relaxed load for x86 as there is no acquire price to pay on x86. It inlines the UnscaledCycleClock::Now() which is a single RTDSC op for x86, and likewise inlines CycleClock::Now() for x86. The inlining should mostly have secondary benefits such as reducing spills on outlined calls. LTO may eventually hoist these functions inline for the hotspots, but it doesn't hurt to default inline this for all builds and let the compiler decide on the first pass. The perlab benchmark is noisy for the plain BM_Now, but the other benchmarks and the run on my local machine are clear. ------------- Local Benchy Benchmark name old cpu/op new cpu/op delta BM_Now 3.41ns ± 1% 2.30ns ± 2% -32.52% (p=0.000 n=50+50) BM_NowWithRegisterPresure 4.96ns ± 2% 4.19ns ± 2% -15.57% (p=0.000 n=56+55) BM_NowWithCallback 3.30ns ± 2% 1.91ns ± 2% -42.00% (p=0.000 n=47+60) ------------- Perflab Benchy Benchmark name old cpu/op new cpu/op delta BM_Now 8.20ns ±13% 4.32ns ±83% ~ (p=0.413 n=4+5) BM_NowWithRegisterPresure 7.91ns ± 1% 3.68ns ± 2% -53.45% (p=0.029 n=4+4) BM_NowWithCallback 2.66ns ±13% 1.58ns ± 0% -40.51% (p=0.008 n=5+5) PiperOrigin-RevId: 434474766 Change-Id: I991d987ae9233e50f09606c874055cf4c5a56300 -- b38330686a0af176a2679163e4d2fa1b90e2f667 by Laramie Leavitt <lar@google.com>: Style, comment, and test updates * Remove a redundant assert in uniform_real_distribution. * Update comment in internal/generate_real.h * Style updates to uniform_real_distribution_test mainly replacing TypeParam with real_type, using aliases for some limits, etc. * Add a few more minor tests. PiperOrigin-RevId: 433902174 Change-Id: Id75be8e24be2fb8f6aea05feec13e3ef320a7254 -- ab2da6047ff7f5dae3add3779fcddf73b03feabf by Abseil Team <absl-team@google.com>: Remove declaration of method whose definition was previously removed. PiperOrigin-RevId: 433507828 Change-Id: I0130b689813125250f7de2664e767e181f676c89 -- df0c87f4ec2c010691931c1bef9d26470a6e63a2 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 433289136 Change-Id: Iba157dc83ed99dafd17a2223d2504e49f8afbb9e -- 7445fa312f2995772900eda82467325b3401a17d by Martijn Vels <mvels@google.com>: Optimize CordReader logic now that CONCAT is removed This CL cleans up various helper functions and logic remaining from previous complex CONCAT logic that is no longer needed, simplifying the CordReader logic. PiperOrigin-RevId: 433208748 Change-Id: I5f7b1883573c44e7c6f8af12c3cddbd197cb134d GitOrigin-RevId: d3b99682554d339c42556680f4d65f83226005e2
3 years ago
friend class CycleClockSource;
};
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
class CycleClockSource {
private:
// CycleClockSource::Register()
//
// Register a function that provides an alternate source for the unscaled CPU
// cycle count value. The source function must be async signal safe, must not
// call CycleClock::Now(), and must have a frequency that matches that of the
// unscaled clock used by CycleClock. A nullptr value resets CycleClock to use
// the default source.
static void Register(CycleClockSourceFunc source);
};
Export of internal Abseil changes -- d3b99682554d339c42556680f4d65f83226005e2 by Martijn Vels <mvels@google.com>: Inline CycleClock code and remove branch for x86 CycleClockSource function This CL removes the relaxed load for x86 as there is no acquire price to pay on x86. It inlines the UnscaledCycleClock::Now() which is a single RTDSC op for x86, and likewise inlines CycleClock::Now() for x86. The inlining should mostly have secondary benefits such as reducing spills on outlined calls. LTO may eventually hoist these functions inline for the hotspots, but it doesn't hurt to default inline this for all builds and let the compiler decide on the first pass. The perlab benchmark is noisy for the plain BM_Now, but the other benchmarks and the run on my local machine are clear. ------------- Local Benchy Benchmark name old cpu/op new cpu/op delta BM_Now 3.41ns ± 1% 2.30ns ± 2% -32.52% (p=0.000 n=50+50) BM_NowWithRegisterPresure 4.96ns ± 2% 4.19ns ± 2% -15.57% (p=0.000 n=56+55) BM_NowWithCallback 3.30ns ± 2% 1.91ns ± 2% -42.00% (p=0.000 n=47+60) ------------- Perflab Benchy Benchmark name old cpu/op new cpu/op delta BM_Now 8.20ns ±13% 4.32ns ±83% ~ (p=0.413 n=4+5) BM_NowWithRegisterPresure 7.91ns ± 1% 3.68ns ± 2% -53.45% (p=0.029 n=4+4) BM_NowWithCallback 2.66ns ±13% 1.58ns ± 0% -40.51% (p=0.008 n=5+5) PiperOrigin-RevId: 434474766 Change-Id: I991d987ae9233e50f09606c874055cf4c5a56300 -- b38330686a0af176a2679163e4d2fa1b90e2f667 by Laramie Leavitt <lar@google.com>: Style, comment, and test updates * Remove a redundant assert in uniform_real_distribution. * Update comment in internal/generate_real.h * Style updates to uniform_real_distribution_test mainly replacing TypeParam with real_type, using aliases for some limits, etc. * Add a few more minor tests. PiperOrigin-RevId: 433902174 Change-Id: Id75be8e24be2fb8f6aea05feec13e3ef320a7254 -- ab2da6047ff7f5dae3add3779fcddf73b03feabf by Abseil Team <absl-team@google.com>: Remove declaration of method whose definition was previously removed. PiperOrigin-RevId: 433507828 Change-Id: I0130b689813125250f7de2664e767e181f676c89 -- df0c87f4ec2c010691931c1bef9d26470a6e63a2 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 433289136 Change-Id: Iba157dc83ed99dafd17a2223d2504e49f8afbb9e -- 7445fa312f2995772900eda82467325b3401a17d by Martijn Vels <mvels@google.com>: Optimize CordReader logic now that CONCAT is removed This CL cleans up various helper functions and logic remaining from previous complex CONCAT logic that is no longer needed, simplifying the CordReader logic. PiperOrigin-RevId: 433208748 Change-Id: I5f7b1883573c44e7c6f8af12c3cddbd197cb134d GitOrigin-RevId: d3b99682554d339c42556680f4d65f83226005e2
3 years ago
#if ABSL_USE_UNSCALED_CYCLECLOCK
inline CycleClockSourceFunc CycleClock::LoadCycleClockSource() {
#if !defined(__x86_64__)
// Optimize for the common case (no callback) by first doing a relaxed load;
// this is significantly faster on non-x86 platforms.
if (cycle_clock_source_.load(std::memory_order_relaxed) == nullptr) {
return nullptr;
}
#endif // !defined(__x86_64__)
// This corresponds to the store(std::memory_order_release) in
// CycleClockSource::Register, and makes sure that any updates made prior to
// registering the callback are visible to this thread before the callback
// is invoked.
return cycle_clock_source_.load(std::memory_order_acquire);
}
// Accessing globals in inlined code in Window DLLs is problematic.
#ifndef _WIN32
inline int64_t CycleClock::Now() {
auto fn = LoadCycleClockSource();
if (fn == nullptr) {
return base_internal::UnscaledCycleClock::Now() >> kShift;
}
return fn() >> kShift;
}
#endif
inline double CycleClock::Frequency() {
return kFrequencyScale * base_internal::UnscaledCycleClock::Frequency();
}
#endif // ABSL_USE_UNSCALED_CYCLECLOCK
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
} // namespace base_internal
ABSL_NAMESPACE_END
} // namespace absl
#endif // ABSL_BASE_INTERNAL_CYCLECLOCK_H_