IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

enable cpu map for mac (#17499)

Browse Source 
* enable cpu map for mac

enable nstreams and nthreads setting for mac

* keep streams=1 for M1

* add explicit type conversion

* remove definition of static cpu

* Update with master

* separate branches for __APPLE__ and __EMSCRIPTEN__

* modify the implementation of is_cpu_map_available function

---------

Co-authored-by: Wanglei Shen <wanglei.shen@intel.com>
This commit is contained in:
Fang Xu
2023-05-30 15:53:23 +08:00
committed by GitHub
co-authored by Wanglei Shen
parent 84f6deb757
commit 25865201ef
6 changed files with 135 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/inference/CMakeLists.txt
+8 -1
View File
@@ -40,7 +40,14 @@ if(WIN32)
file (GLOB LIBRARY_HEADERS
${LIBRARY_HEADERS}
${CMAKE_CURRENT_SOURCE_DIR}/src/os/win/*.hpp)
elseif(NOT (APPLE OR EMSCRIPTEN))
elseif(APPLE)
file (GLOB LIBRARY_SRC
${LIBRARY_SRC}
${CMAKE_CURRENT_SOURCE_DIR}/src/os/mac/*.cpp)
file (GLOB LIBRARY_HEADERS
${LIBRARY_HEADERS}
${CMAKE_CURRENT_SOURCE_DIR}/src/os/mac/*.hpp)
elseif(NOT EMSCRIPTEN)
file (GLOB LIBRARY_SRC
${LIBRARY_SRC}
${CMAKE_CURRENT_SOURCE_DIR}/src/os/lin/*.cpp)
src/inference/src/dev/threading/cpu_streams_executor.cpp
+1 -1
View File
@@ -148,7 +148,7 @@ struct CPUStreamsExecutor::Impl {
? custom::info::core_types().back()
: custom::info::core_types().front();
if (_impl->_config._cpu_pinning) {
# ifdef _WIN32
# if defined(_WIN32) || defined(__APPLE__)
_taskArena.reset(new custom::task_arena{custom::task_arena::constraints{}
.set_core_type(selected_core_type)
.set_max_concurrency(concurrency)});
src/inference/src/os/mac/mac_system_conf.cpp
+77
View File
@@ -0,0 +1,77 @@
// Copyright (C) 2018-2023 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//
#include <sys/sysctl.h>
#include <memory>
#include <vector>
#include "dev/threading/parallel_custom_arena.hpp"
#include "openvino/runtime/system_conf.hpp"
#include "streams_executor.hpp"
namespace ov {
CPU::CPU() {
_num_threads = parallel_get_max_threads();
parse_processor_info_macos(
_processors,
_numa_nodes,
_cores,
_proc_type_table);
}
int parse_processor_info_macos(int& _processors,
int& _numa_nodes,
int& _cores,
std::vector<std::vector<int>>& _proc_type_table) {
uint64_t output = 0;
size_t size = sizeof(output);
_processors = 0;
_numa_nodes = 0;
_cores = 0;
if (sysctlbyname("hw.ncpu", &output, &size, NULL, 0) < 0) {
return -1;
} else {
_processors = static_cast<int>(output);
}
if (sysctlbyname("hw.physicalcpu", &output, &size, NULL, 0) < 0) {
_processors = 0;
return -1;
} else {
_cores = static_cast<int>(output);
}
_numa_nodes = 1;
if (sysctlbyname("hw.optional.arm64", &output, &size, NULL, 0) < 0) {
_proc_type_table.resize(1, std::vector<int>(PROC_TYPE_TABLE_SIZE, 0));
_proc_type_table[0][ALL_PROC] = _processors;
_proc_type_table[0][MAIN_CORE_PROC] = _cores;
_proc_type_table[0][HYPER_THREADING_PROC] = _processors - _cores;
} else {
if (sysctlbyname("hw.perflevel0.physicalcpu", &output, &size, NULL, 0) < 0) {
_processors = 0;
_cores = 0;
_numa_nodes = 0;
return -1;
} else {
_proc_type_table.resize(1, std::vector<int>(PROC_TYPE_TABLE_SIZE, 0));
_proc_type_table[0][ALL_PROC] = _processors;
_proc_type_table[0][MAIN_CORE_PROC] = output;
}
if (sysctlbyname("hw.perflevel1.physicalcpu", &output, &size, NULL, 0) < 0) {
return 0;
} else {
_proc_type_table[0][EFFICIENT_CORE_PROC] = output;
}
}
return 0;
}
} // namespace ov
src/inference/src/streams_executor.hpp
+15
View File
@@ -73,4 +73,19 @@ void parse_processor_info_win(const char* base_ptr,
std::vector<std::vector<int>>& _cpu_mapping_table);
#endif
#if defined(__APPLE__)
/**
* @brief Parse processors infomation on Linux
* @param[in] _processors total number for processors in system.
* @param[out] _numa_nodes total number for sockets in system
* @param[out] _cores total number for physical CPU cores in system
* @param[out] _proc_type_table summary table of number of processors per type
* @return
*/
int parse_processor_info_macos(int& _processors,
int& _numa_nodes,
int& _cores,
std::vector<std::vector<int>>& _proc_type_table);
#endif
} // namespace ov
src/inference/src/system_conf.cpp
+33 -52
View File
@@ -163,7 +163,7 @@ CPU& cpu_info() {
return cpu;
}
#if defined(__APPLE__) || defined(__EMSCRIPTEN__)
#if defined(__EMSCRIPTEN__)
// for Linux and Windows the getNumberOfCPUCores (that accounts only for physical cores) implementation is OS-specific
// (see cpp files in corresponding folders), for __APPLE__ it is default :
int get_number_of_cpu_cores(bool) {
@@ -191,58 +191,39 @@ std::vector<std::vector<int>> reserve_available_cpus(const std::vector<std::vect
}
void set_cpu_used(const std::vector<int>& cpu_ids, const int used) {}
int parse_processor_info_macos(int& _processors,
int& _numa_nodes,
int& _cores,
std::vector<std::vector<int>>& _proc_type_table) {
uint64_t output = 0;
size_t size = sizeof(output);
_processors = 0;
_numa_nodes = 0;
_cores = 0;
if (sysctlbyname("hw.ncpu", &output, &size, NULL, 0) < 0) {
return -1;
} else {
_processors = output;
}
if (sysctlbyname("hw.physicalcpu", &output, &size, NULL, 0) < 0) {
_processors = 0;
return -1;
} else {
_cores = output;
}
_numa_nodes = 1;
if (sysctlbyname("hw.optional.arm64", &output, &size, NULL, 0) < 0) {
_proc_type_table.resize(1, std::vector<int>(PROC_TYPE_TABLE_SIZE, 0));
_proc_type_table[0][ALL_PROC] = _processors;
_proc_type_table[0][MAIN_CORE_PROC] = _cores;
_proc_type_table[0][HYPER_THREADING_PROC] = _processors - _cores;
} else {
if (sysctlbyname("hw.perflevel0.physicalcpu", &output, &size, NULL, 0) < 0) {
_processors = 0;
_cores = 0;
_numa_nodes = 0;
return -1;
} else {
_proc_type_table.resize(1, std::vector<int>(PROC_TYPE_TABLE_SIZE, 0));
_proc_type_table[0][ALL_PROC] = _processors;
_proc_type_table[0][MAIN_CORE_PROC] = output;
}
if (sysctlbyname("hw.perflevel1.physicalcpu", &output, &size, NULL, 0) < 0) {
return 0;
} else {
_proc_type_table[0][EFFICIENT_CORE_PROC] = output;
}
}
return 0;
#elif defined(__APPLE__)
// for Linux and Windows the getNumberOfCPUCores (that accounts only for physical cores) implementation is OS-specific
// (see cpp files in corresponding folders), for __APPLE__ it is default :
int get_number_of_cpu_cores(bool) {
return parallel_get_max_threads();
}
# if !((OV_THREAD == OV_THREAD_TBB) || (OV_THREAD == OV_THREAD_TBB_AUTO))
std::vector<int> get_available_numa_nodes() {
return {-1};
}
# endif
int get_number_of_logical_cpu_cores(bool) {
return parallel_get_max_threads();
}
bool is_cpu_map_available() {
CPU& cpu = cpu_info();
return cpu._proc_type_table.size() > 0;
}
std::vector<std::vector<int>> get_proc_type_table() {
CPU& cpu = cpu_info();
std::lock_guard<std::mutex> lock{cpu._cpu_mutex};
return cpu._proc_type_table;
}
int get_num_numa_nodes() {
return cpu_info()._numa_nodes;
}
std::vector<std::vector<int>> reserve_available_cpus(const std::vector<std::vector<int>> streams_info_table) {
return {{-1}};
}
void set_cpu_used(const std::vector<int>& cpu_ids, const int used) {}
#else
src/plugins/intel_cpu/src/config.cpp
+1
View File
@@ -274,6 +274,7 @@ void Config::readProperties(const std::map<std::string, std::string> &prop) {
#if defined(OPENVINO_ARCH_ARM) || defined(OPENVINO_ARCH_ARM64)
// TODO: multi-stream execution has functional issues on ARM target
streamExecutorConfig._streams = 1;
streamExecutorConfig._streams_changed = true;
#endif
CPU_DEBUG_CAP_ENABLE(applyDebugCapsProperties());