aboutsummaryrefslogtreecommitdiff
path: root/src/common/cpuinfo/CpuInfo.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/common/cpuinfo/CpuInfo.cpp')
-rw-r--r--src/common/cpuinfo/CpuInfo.cpp164
1 files changed, 109 insertions, 55 deletions
diff --git a/src/common/cpuinfo/CpuInfo.cpp b/src/common/cpuinfo/CpuInfo.cpp
index 436e7ea803..93f51e599a 100644
--- a/src/common/cpuinfo/CpuInfo.cpp
+++ b/src/common/cpuinfo/CpuInfo.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021-2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -25,7 +25,7 @@
#include "arm_compute/core/Error.h"
#include "arm_compute/core/Log.h"
-#include "src/common/cpuinfo/target/CpuInfoSveUtils.h"
+
#include "support/StringSupport.h"
#include "support/ToolchainSupport.h"
@@ -35,27 +35,35 @@
#include <algorithm>
#include <cstring>
#include <fstream>
+#if !defined(_WIN64)
#include <regex.h> /* C++ std::regex takes up a lot of space in the standalone builds */
#include <sched.h>
+#endif /* !defined(_WIN64) */
+
#include <thread>
#include <unordered_map>
#endif /* !defined(BARE_METAL) */
-#if !defined(BARE_METAL) && !defined(__APPLE__) && (defined(__arm__) || defined(__aarch64__))
+#if !defined(_WIN64)
+#if !defined(BARE_METAL) && !defined(__APPLE__) && !defined(__OpenBSD__) && (defined(__arm__) || defined(__aarch64__))
#include <asm/hwcap.h> /* Get HWCAP bits from asm/hwcap.h */
#include <sys/auxv.h>
-#endif /* !defined(BARE_METAL) && !defined(__APPLE__) && (defined(__arm__) || defined(__aarch64__)) */
-
-#define ARM_COMPUTE_CPU_FEATURE_HWCAP_CPUID (1 << 11)
-#define ARM_COMPUTE_GET_FEATURE_REG(var, freg) __asm __volatile("MRS %0, " #freg \
- : "=r"(var))
+#elif defined(__APPLE__) && defined(__aarch64__)
+#include <sys/sysctl.h>
+#include <sys/types.h>
+#endif /* defined(__APPLE__) && defined(__aarch64__)) */
+#endif /* !defined(BARE_METAL) && !defined(__APPLE__) && !defined(__OpenBSD__) && (defined(__arm__) || defined(__aarch64__)) */
+
+#define ARM_COMPUTE_CPU_FEATURE_HWCAP_CPUID (1 << 11)
+#define ARM_COMPUTE_GET_FEATURE_REG(var, freg) __asm __volatile("MRS %0, " #freg : "=r"(var))
namespace arm_compute
{
namespace cpuinfo
{
namespace
{
-#if !defined(BARE_METAL) && !defined(__APPLE__) && (defined(__arm__) || defined(__aarch64__))
+#if !defined(_WIN64) && !defined(BARE_METAL) && !defined(__APPLE__) && !defined(__OpenBSD__) && \
+ (defined(__arm__) || defined(__aarch64__))
/** Extract MIDR using CPUID information that are exposed to user-space
*
* @param[in] max_num_cpus Maximum number of possible CPUs
@@ -65,15 +73,15 @@ namespace
std::vector<uint32_t> midr_from_cpuid(uint32_t max_num_cpus)
{
std::vector<uint32_t> cpus;
- for(unsigned int i = 0; i < max_num_cpus; ++i)
+ for (unsigned int i = 0; i < max_num_cpus; ++i)
{
std::stringstream str;
str << "/sys/devices/system/cpu/cpu" << i << "/regs/identification/midr_el1";
std::ifstream file(str.str(), std::ios::in);
- if(file.is_open())
+ if (file.is_open())
{
std::string line;
- if(bool(getline(file, line)))
+ if (bool(getline(file, line)))
{
cpus.emplace_back(support::cpp11::stoul(line, nullptr, support::cpp11::NumericBase::BASE_16));
}
@@ -115,34 +123,35 @@ std::vector<uint32_t> midr_from_proc_cpuinfo(int max_num_cpus)
ARM_COMPUTE_ERROR_ON_MSG(ret_status != 0, "Regex compilation failed.");
std::ifstream file("/proc/cpuinfo", std::ios::in);
- if(file.is_open())
+ if (file.is_open())
{
std::string line;
int midr = 0;
int curcpu = -1;
- while(bool(getline(file, line)))
+ while (bool(getline(file, line)))
{
std::array<regmatch_t, 2> match;
ret_status = regexec(&proc_regex, line.c_str(), 2, match.data(), 0);
- if(ret_status == 0)
+ if (ret_status == 0)
{
std::string id = line.substr(match[1].rm_so, (match[1].rm_eo - match[1].rm_so));
int newcpu = support::cpp11::stoi(id, nullptr);
- if(curcpu >= 0 && midr == 0)
+ if (curcpu >= 0 && midr == 0)
{
// Matched a new CPU ID without any description of the previous one - looks like old format.
return {};
}
- if(curcpu >= 0 && curcpu < max_num_cpus)
+ if (curcpu >= 0 && curcpu < max_num_cpus)
{
cpus.emplace_back(midr);
}
else
{
- ARM_COMPUTE_LOG_INFO_MSG_CORE("Trying to populate a core id with id greater than the expected number of cores!");
+ ARM_COMPUTE_LOG_INFO_MSG_CORE(
+ "Trying to populate a core id with id greater than the expected number of cores!");
}
midr = 0;
@@ -152,7 +161,7 @@ std::vector<uint32_t> midr_from_proc_cpuinfo(int max_num_cpus)
}
ret_status = regexec(&imp_regex, line.c_str(), 2, match.data(), 0);
- if(ret_status == 0)
+ if (ret_status == 0)
{
std::string subexp = line.substr(match[1].rm_so, (match[1].rm_eo - match[1].rm_so));
int impv = support::cpp11::stoi(subexp, nullptr, support::cpp11::NumericBase::BASE_16);
@@ -162,7 +171,7 @@ std::vector<uint32_t> midr_from_proc_cpuinfo(int max_num_cpus)
}
ret_status = regexec(&var_regex, line.c_str(), 2, match.data(), 0);
- if(ret_status == 0)
+ if (ret_status == 0)
{
std::string subexp = line.substr(match[1].rm_so, (match[1].rm_eo - match[1].rm_so));
int varv = support::cpp11::stoi(subexp, nullptr, support::cpp11::NumericBase::BASE_16);
@@ -172,7 +181,7 @@ std::vector<uint32_t> midr_from_proc_cpuinfo(int max_num_cpus)
}
ret_status = regexec(&part_regex, line.c_str(), 2, match.data(), 0);
- if(ret_status == 0)
+ if (ret_status == 0)
{
std::string subexp = line.substr(match[1].rm_so, (match[1].rm_eo - match[1].rm_so));
int partv = support::cpp11::stoi(subexp, nullptr, support::cpp11::NumericBase::BASE_16);
@@ -182,7 +191,7 @@ std::vector<uint32_t> midr_from_proc_cpuinfo(int max_num_cpus)
}
ret_status = regexec(&rev_regex, line.c_str(), 2, match.data(), 0);
- if(ret_status == 0)
+ if (ret_status == 0)
{
std::string subexp = line.substr(match[1].rm_so, (match[1].rm_eo - match[1].rm_so));
int regv = support::cpp11::stoi(subexp, nullptr);
@@ -193,13 +202,14 @@ std::vector<uint32_t> midr_from_proc_cpuinfo(int max_num_cpus)
}
}
- if(curcpu >= 0 && curcpu < max_num_cpus)
+ if (curcpu >= 0 && curcpu < max_num_cpus)
{
cpus.emplace_back(midr);
}
else
{
- ARM_COMPUTE_LOG_INFO_MSG_CORE("Trying to populate a core id with id greater than the expected number of cores!");
+ ARM_COMPUTE_LOG_INFO_MSG_CORE(
+ "Trying to populate a core id with id greater than the expected number of cores!");
}
}
@@ -224,11 +234,11 @@ int get_max_cpus()
CPUspresent.open("/sys/devices/system/cpu/present", std::ios::in);
bool success = false;
- if(CPUspresent.is_open())
+ if (CPUspresent.is_open())
{
std::string line;
- if(bool(getline(CPUspresent, line)))
+ if (bool(getline(CPUspresent, line)))
{
/* The content of this file is a list of ranges or single values, e.g.
* 0-5, or 1-3,5,7 or similar. As we are interested in the
@@ -237,9 +247,9 @@ int get_max_cpus()
*/
auto startfrom = line.begin();
- for(auto i = line.begin(); i < line.end(); ++i)
+ for (auto i = line.begin(); i < line.end(); ++i)
{
- if(*i == '-' || *i == ',')
+ if (*i == '-' || *i == ',')
{
startfrom = i + 1;
}
@@ -253,38 +263,64 @@ int get_max_cpus()
}
// Return std::thread::hardware_concurrency() as a fallback.
- if(!success)
+ if (!success)
{
max_cpus = std::thread::hardware_concurrency();
}
return max_cpus;
}
-#endif /* !defined(BARE_METAL) && !defined(__APPLE__) && (defined(__arm__) || defined(__aarch64__)) */
+#elif defined(__aarch64__) && \
+ defined(__APPLE__) /* !defined(BARE_METAL) && !defined(__APPLE__) && (defined(__arm__) || defined(__aarch64__)) */
+/** Query features through sysctlbyname
+ *
+ * @return int value queried
+ */
+int get_hw_capability(const std::string &cap)
+{
+ int64_t result(0);
+ size_t size = sizeof(result);
+ sysctlbyname(cap.c_str(), &result, &size, NULL, 0);
+ return result;
+}
+#endif /* !defined(BARE_METAL) && !defined(__APPLE__) && !defined(__OpenBSD__) && (defined(__arm__) || defined(__aarch64__)) */
+
+#if defined(BARE_METAL) && defined(__aarch64__)
+uint64_t get_sve_feature_reg()
+{
+ uint64_t svefr0 = 0;
+ __asm __volatile(".inst 0xd5380483 // mrs x3, ID_AA64ZFR0_EL1\n"
+ "MOV %0, X3"
+ : "=r"(svefr0)
+ :
+ : "x3");
+ return svefr0;
+}
+#endif /* defined(BARE_METAL) && defined(__aarch64__) */
} // namespace
-CpuInfo::CpuInfo(CpuIsaInfo isa, std::vector<CpuModel> cpus)
- : _isa(std::move(isa)), _cpus(std::move(cpus))
+CpuInfo::CpuInfo(CpuIsaInfo isa, std::vector<CpuModel> cpus) : _isa(std::move(isa)), _cpus(std::move(cpus))
{
}
CpuInfo CpuInfo::build()
{
-#if !defined(BARE_METAL) && !defined(__APPLE__) && (defined(__arm__) || defined(__aarch64__))
+#if !defined(_WIN64) && !defined(BARE_METAL) && !defined(__APPLE__) && !defined(__OpenBSD__) && \
+ (defined(__arm__) || defined(__aarch64__))
const uint32_t hwcaps = getauxval(AT_HWCAP);
const uint32_t hwcaps2 = getauxval(AT_HWCAP2);
const uint32_t max_cpus = get_max_cpus();
// Populate midr values
std::vector<uint32_t> cpus_midr;
- if(hwcaps & ARM_COMPUTE_CPU_FEATURE_HWCAP_CPUID)
+ if (hwcaps & ARM_COMPUTE_CPU_FEATURE_HWCAP_CPUID)
{
cpus_midr = midr_from_cpuid(max_cpus);
}
- if(cpus_midr.empty())
+ if (cpus_midr.empty())
{
cpus_midr = midr_from_proc_cpuinfo(max_cpus);
}
- if(cpus_midr.empty())
+ if (cpus_midr.empty())
{
cpus_midr.resize(max_cpus, 0);
}
@@ -300,32 +336,49 @@ CpuInfo CpuInfo::build()
CpuInfo info(isa, cpus_model);
return info;
-#elif(BARE_METAL) && defined(__aarch64__) /* !defined(BARE_METAL) && !defined(__APPLE__) && (defined(__arm__) || defined(__aarch64__)) */
+#elif (BARE_METAL) && \
+ defined( \
+ __aarch64__) /* !defined(BARE_METAL) && !defined(__APPLE__) && !defined(__OpenBSD__) && (defined(__arm__) || defined(__aarch64__)) */
// Assume single CPU in bare metal mode. Just read the ID register and feature bits directly.
- uint64_t isar0 = 0, isar1 = 0, pfr0 = 0, svefr0 = 0, midr = 0;
+ uint64_t isar0 = 0, isar1 = 0, pfr0 = 0, pfr1 = 0, svefr0 = 0, midr = 0;
ARM_COMPUTE_GET_FEATURE_REG(isar0, ID_AA64ISAR0_EL1);
ARM_COMPUTE_GET_FEATURE_REG(isar1, ID_AA64ISAR1_EL1);
ARM_COMPUTE_GET_FEATURE_REG(pfr0, ID_AA64PFR0_EL1);
+ ARM_COMPUTE_GET_FEATURE_REG(pfr1, ID_AA64PFR1_EL1);
ARM_COMPUTE_GET_FEATURE_REG(midr, MIDR_EL1);
- if((pfr0 >> 32) & 0xf)
+ if ((pfr0 >> 32) & 0xf)
{
svefr0 = get_sve_feature_reg();
}
- CpuIsaInfo isa = init_cpu_isa_from_regs(isar0, isar1, pfr0, svefr0, midr);
+ CpuIsaInfo isa = init_cpu_isa_from_regs(isar0, isar1, pfr0, pfr1, svefr0, midr);
std::vector<CpuModel> cpus_model(1, midr_to_model(midr));
CpuInfo info(isa, cpus_model);
return info;
-#else /* #elif(BARE_METAL) && defined(__aarch64__) */
- CpuInfo info(CpuIsaInfo(), { CpuModel::GENERIC });
+#elif defined(__aarch64__) && defined(__APPLE__) /* #elif(BARE_METAL) && defined(__aarch64__) */
+ int ncpus = get_hw_capability("hw.perflevel0.logicalcpu");
+ CpuIsaInfo isainfo;
+ std::vector<CpuModel> cpus_model(ncpus);
+ isainfo.neon = get_hw_capability("hw.optional.neon");
+ isainfo.fp16 = get_hw_capability("hw.optional.neon_fp16");
+ isainfo.dot = get_hw_capability("hw.optional.arm.FEAT_DotProd");
+ CpuInfo info(isainfo, cpus_model);
+ return info;
+#elif defined(__aarch64__) && defined(_WIN64) /* #elif defined(__aarch64__) && defined(__APPLE__) */
+ CpuIsaInfo isainfo;
+ isainfo.neon = true;
+ CpuInfo info(isainfo, {CpuModel::GENERIC});
return info;
-#endif /* !defined(BARE_METAL) && !defined(__APPLE__) && (defined(__arm__) || defined(__aarch64__)) */
+#else /* #elif defined(__aarch64__) && defined(_WIN64) */
+ CpuInfo info(CpuIsaInfo(), {CpuModel::GENERIC});
+ return info;
+#endif /* !defined(BARE_METAL) && !defined(__APPLE__) && !defined(__OpenBSD__) && (defined(__arm__) || defined(__aarch64__)) */
}
CpuModel CpuInfo::cpu_model(uint32_t cpuid) const
{
- if(cpuid < _cpus.size())
+ if (cpuid < _cpus.size())
{
return _cpus[cpuid];
}
@@ -334,11 +387,12 @@ CpuModel CpuInfo::cpu_model(uint32_t cpuid) const
CpuModel CpuInfo::cpu_model() const
{
-#if defined(BARE_METAL) || defined(__APPLE__) || (!defined(__arm__) && !defined(__aarch64__))
+#if defined(_WIN64) || defined(BARE_METAL) || defined(__APPLE__) || defined(__OpenBSD__) || \
+ (!defined(__arm__) && !defined(__aarch64__))
return cpu_model(0);
-#else /* defined(BARE_METAL) || defined(__APPLE__) || (!defined(__arm__) && !defined(__aarch64__)) */
+#else /* defined(BARE_METAL) || defined(__APPLE__) || defined(__OpenBSD__) || (!defined(__arm__) && !defined(__aarch64__)) */
return cpu_model(sched_getcpu());
-#endif /* defined(BARE_METAL) || defined(__APPLE__) || (!defined(__arm__) && !defined(__aarch64__)) */
+#endif /* defined(BARE_METAL) || defined(__APPLE__) || defined(__OpenBSD__) || (!defined(__arm__) && !defined(__aarch64__)) */
}
uint32_t CpuInfo::num_cpus() const
@@ -350,7 +404,7 @@ uint32_t num_threads_hint()
{
unsigned int num_threads_hint = 1;
-#if !defined(BARE_METAL)
+#if !defined(BARE_METAL) && !defined(_WIN64) && !defined(ARM_COMPUTE_DISABLE_THREADS_HINT)
std::vector<std::string> cpus;
cpus.reserve(64);
@@ -363,13 +417,13 @@ uint32_t num_threads_hint()
// Read cpuinfo and get occurrence of each core
std::ifstream cpuinfo_file("/proc/cpuinfo", std::ios::in);
- if(cpuinfo_file.is_open())
+ if (cpuinfo_file.is_open())
{
std::string line;
- while(bool(getline(cpuinfo_file, line)))
+ while (bool(getline(cpuinfo_file, line)))
{
std::array<regmatch_t, 2> match;
- if(regexec(&cpu_part_rgx, line.c_str(), 2, match.data(), 0) == 0)
+ if (regexec(&cpu_part_rgx, line.c_str(), 2, match.data(), 0) == 0)
{
cpus.emplace_back(line.substr(match[1].rm_so, (match[1].rm_eo - match[1].rm_so)));
}
@@ -382,13 +436,13 @@ uint32_t num_threads_hint()
auto least_frequent_cpu_occurences = [](const std::vector<std::string> &cpus) -> uint32_t
{
std::unordered_map<std::string, uint32_t> cpus_freq;
- for(const auto &cpu : cpus)
+ for (const auto &cpu : cpus)
{
cpus_freq[cpu]++;
}
uint32_t vmin = cpus.size() + 1;
- for(const auto &cpu_freq : cpus_freq)
+ for (const auto &cpu_freq : cpus_freq)
{
vmin = std::min(vmin, cpu_freq.second);
}
@@ -397,9 +451,9 @@ uint32_t num_threads_hint()
// Set thread hint
num_threads_hint = cpus.empty() ? std::thread::hardware_concurrency() : least_frequent_cpu_occurences(cpus);
-#endif /* !defined(BARE_METAL) */
+#endif /* !defined(BARE_METAL) && !defined(_WIN64) && !defined(ARM_COMPUTE_DISABLE_THREADS_HINT) */
return num_threads_hint;
}
} // namespace cpuinfo
-} // namespace arm_compute \ No newline at end of file
+} // namespace arm_compute
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-04 01:00:53 +0000