experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Move SIMD feature warnings to the first use of intensive CPU computation, rather than startup.

Browse Source 
Change: 143570811
This commit is contained in:
Pete Warden 2017-01-04 10:49:06 -08:00 committed by TensorFlower Gardener
parent ffd7338ecb
commit 354972df8a
4 changed files with 102 additions and 41 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
tensorflow/core/BUILD
View File

@ -239,6 +239,7 @@ cc_library(
"lib/strings/str_util.h",
"lib/strings/strcat.h",
"lib/strings/stringprintf.h",
"platform/cpu_feature_guard.h",
"platform/cpu_info.h",
"platform/dynamic_annotations.h",
"platform/env.h",

4
tensorflow/core/common_runtime/local_device.cc
View File

@ -19,6 +19,7 @@ limitations under the License.
#include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
#include "tensorflow/core/common_runtime/eigen_thread_pool.h"
#include "tensorflow/core/lib/core/threadpool.h"
#include "tensorflow/core/platform/cpu_feature_guard.h"
#include "tensorflow/core/platform/cpu_info.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/types.h"
@ -63,6 +64,9 @@ LocalDevice::LocalDevice(const SessionOptions& options,
Allocator* device_allocator)
: Device(options.env, attributes, device_allocator),
owned_tp_info_(nullptr) {
// If we're running on the CPU, log warnings if we're not compiled using the
// best flags for performance.
port::WarnAboutUnusedCPUFeatures();
LocalDevice::EigenThreadPoolInfo* tp_info;
if (use_global_threadpool_) {
// All ThreadPoolDevices in the process will use this single fixed

106
tensorflow/core/platform/cpu_feature_guard.cc
View File

@ -13,6 +13,10 @@ See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
#include "tensorflow/core/platform/cpu_feature_guard.h"
#include <mutex>
#include "tensorflow/core/platform/cpu_info.h"
#include "tensorflow/core/platform/logging.h"
@ -20,6 +24,31 @@ namespace tensorflow {
namespace port {
namespace {
// If the CPU feature isn't present, log a fatal error.
void CheckFeatureOrDie(CPUFeature feature, const string& feature_name) {
if (!TestCPUFeature(feature)) {
#ifdef __ANDROID__
// Some Android emulators seem to indicate they don't support SSE, so to
// avoid crashes when testing, switch this to a warning.
LOG(WARNING)
#else
LOG(FATAL)
#endif
<< "The TensorFlow library was compiled to use " << feature_name
<< " instructions, but these aren't available on your machine.";
}
}
// If the CPU feature is present, log warning.
void WarnIfFeatureUnused(CPUFeature feature, const string& feature_name) {
if (TestCPUFeature(feature)) {
LOG(WARNING) << "The TensorFlow library wasn't compiled to use "
<< feature_name
<< " instructions, but these are available on your machine "
"and could speed up CPU computations.";
}
}
// Raises an error if the binary has been compiled for a CPU feature (like AVX)
// that isn't available on the current machine. It also warns of performance
// loss if there's a feature available that's not being used.
@ -31,76 +60,71 @@ class CPUFeatureGuard {
CPUFeatureGuard() {
#ifdef __SSE__
CheckFeatureOrDie(CPUFeature::SSE, "SSE");
#else
WarnIfFeatureUnused(CPUFeature::SSE, "SSE");
#endif // __SSE__
#ifdef __SSE2__
CheckFeatureOrDie(CPUFeature::SSE2, "SSE2");
#else
WarnIfFeatureUnused(CPUFeature::SSE2, "SSE2");
#endif // __SSE2__
#ifdef __SSE3__
CheckFeatureOrDie(CPUFeature::SSE3, "SSE3");
#else
WarnIfFeatureUnused(CPUFeature::SSE3, "SSE3");
#endif // __SSE3__
#ifdef __SSE4_1__
CheckFeatureOrDie(CPUFeature::SSE4_1, "SSE4.1");
#else
WarnIfFeatureUnused(CPUFeature::SSE4_1, "SSE4.1");
#endif // __SSE4_1__
#ifdef __SSE4_2__
CheckFeatureOrDie(CPUFeature::SSE4_2, "SSE4.2");
#else
WarnIfFeatureUnused(CPUFeature::SSE4_2, "SSE4.2");
#endif // __SSE4_2__
#ifdef __AVX__
CheckFeatureOrDie(CPUFeature::AVX, "AVX");
#else
WarnIfFeatureUnused(CPUFeature::AVX, "AVX");
#endif // __AVX__
#ifdef __AVX2__
CheckFeatureOrDie(CPUFeature::AVX2, "AVX2");
#else
WarnIfFeatureUnused(CPUFeature::AVX2, "AVX2");
#endif // __AVX2__
#ifdef __AVX512F__
CheckFeatureOrDie(CPUFeature::AVX512F, "AVX512F");
#else
WarnIfFeatureUnused(CPUFeature::AVX512F, "AVX512F");
#endif // __AVX512F__
#ifdef __FMA__
CheckFeatureOrDie(CPUFeature::FMA, "FMA");
#else
WarnIfFeatureUnused(CPUFeature::FMA, "FMA");
#endif // __FMA__
}
void CheckFeatureOrDie(CPUFeature feature, const string& feature_name) {
if (!TestCPUFeature(feature)) {
#ifdef __ANDROID__
// Some Android emulators seem to indicate they don't support SSE, so to
// avoid crashes when testing, switch this to a warning.
LOG(WARNING)
#else
LOG(FATAL)
#endif
<< "The TensorFlow library was compiled to use " << feature_name
<< " instructions, but these aren't available on your machine.";
}
}
void WarnIfFeatureUnused(CPUFeature feature, const string& feature_name) {
if (TestCPUFeature(feature)) {
LOG(WARNING) << "The TensorFlow library wasn't compiled to use "
<< feature_name
<< " instructions, but these are available on your machine "
"and could speed up CPU computations.";
}
}
};
CPUFeatureGuard g_cpu_feature_guard_singleton;
std::once_flag g_cpu_feature_guard_warn_once_flag;
} // namespace
void WarnAboutUnusedCPUFeatures() {
std::call_once(g_cpu_feature_guard_warn_once_flag, [] {
#ifndef __SSE__
WarnIfFeatureUnused(CPUFeature::SSE, "SSE");
#endif // __SSE__
#ifndef __SSE2__
WarnIfFeatureUnused(CPUFeature::SSE2, "SSE2");
#endif // __SSE2__
#ifndef __SSE3__
WarnIfFeatureUnused(CPUFeature::SSE3, "SSE3");
#endif // __SSE3__
#ifndef __SSE4_1__
WarnIfFeatureUnused(CPUFeature::SSE4_1, "SSE4.1");
#endif // __SSE4_1__
#ifndef __SSE4_2__
WarnIfFeatureUnused(CPUFeature::SSE4_2, "SSE4.2");
#endif // __SSE4_2__
#ifndef __AVX__
WarnIfFeatureUnused(CPUFeature::AVX, "AVX");
#endif // __AVX__
#ifndef __AVX2__
WarnIfFeatureUnused(CPUFeature::AVX2, "AVX2");
#endif // __AVX2__
#ifndef __AVX512F__
WarnIfFeatureUnused(CPUFeature::AVX512F, "AVX512F");
#endif // __AVX512F__
#ifndef __FMA__
WarnIfFeatureUnused(CPUFeature::FMA, "FMA");
#endif // __FMA__
});
}
} // namespace port
} // namespace tensorflow

32
tensorflow/core/platform/cpu_feature_guard.h Normal file
View File

@ -0,0 +1,32 @@
/* Copyright 2017 The TensorFlow 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.
==============================================================================*/
#ifndef TENSORFLOW_PLATFORM_CPU_FEATURE_GUARD_H_
#define TENSORFLOW_PLATFORM_CPU_FEATURE_GUARD_H_
namespace tensorflow {
namespace port {
// Called by the framework when we expect heavy CPU computation and we want to
// be sure that the code has been compiled to run optimally on the current
// hardware. The first time it's called it will run lightweight checks of
// available SIMD acceleration features and log warnings about any that aren't
// used.
void WarnAboutUnusedCPUFeatures();
} // namespace port
} // namespace tensorflow
#endif // TENSORFLOW_PLATFORM_CPU_FEATURE_GUARD_H_