|
| 1 | +/* |
| 2 | + * Copyright (c) Meta Platforms, Inc. and affiliates. |
| 3 | + * All rights reserved. |
| 4 | + * |
| 5 | + * This source code is licensed under the BSD-style license found in the |
| 6 | + * LICENSE file in the root directory of this source tree. |
| 7 | + */ |
| 8 | + |
| 9 | +#include <executorch/runtime/platform/assert.h> |
| 10 | +#include <fstream> |
| 11 | +#include <mutex> |
| 12 | +#include <string> |
| 13 | +#include <vector> |
| 14 | + |
| 15 | +#include "cpuinfo_utils.h" |
| 16 | + |
| 17 | +namespace torch { |
| 18 | +namespace executorch { |
| 19 | +namespace cpuinfo { |
| 20 | + |
| 21 | +// Ignore revisions (last digit (4 LSBs)) |
| 22 | +#define CPUINFO_ARM_MIDR_CORTEX_A520 UINT32_C(0x410FD800) |
| 23 | +#define CPUINFO_ARM_MIDR_CORTEX_A53 UINT32_C(0x410FD030) |
| 24 | +#define CPUINFO_ARM_MIDR_CORTEX_A55 UINT32_C(0x410FD050) |
| 25 | +#define CPUINFO_ARM_MIDR_CORTEX_A57 UINT32_C(0x410FD070) |
| 26 | + |
| 27 | +#define RIVISION_MASK UINT32_C(0xFFFFFFF0) |
| 28 | + |
| 29 | +namespace { |
| 30 | +bool is_non_performant_core(const struct cpuinfo_uarch_info* uarch_info) { |
| 31 | + switch (uarch_info->uarch) { |
| 32 | + case cpuinfo_uarch_cortex_a55: |
| 33 | + case cpuinfo_uarch_cortex_a53: |
| 34 | + case cpuinfo_uarch_cortex_a510: |
| 35 | + return true; |
| 36 | + // This can be so many other cores. |
| 37 | + // Need to update this to better account for slow cores |
| 38 | + // Also does not account Apple's A/M series cores |
| 39 | + // And not yet qcomm's |
| 40 | + default: |
| 41 | + break; |
| 42 | + } |
| 43 | + // A520 is not yet updated in cpuinfo |
| 44 | + // Hence decode it separately. |
| 45 | +#if CPUINFO_ARCH_ARM || CPUINFO_ARCH_ARM64 |
| 46 | + if ((uarch_info->midr & RIVISION_MASK) == CPUINFO_ARM_MIDR_CORTEX_A520) { |
| 47 | + return true; |
| 48 | + } |
| 49 | +#endif |
| 50 | + return false; |
| 51 | +} |
| 52 | + |
| 53 | +std::vector<uint32_t>* get_static_cpu_midr_vector() { |
| 54 | + static std::vector<uint32_t> cpu_midrs; |
| 55 | + return &cpu_midrs; |
| 56 | +} |
| 57 | + |
| 58 | +uint32_t _get_model_specific_num_cores() { |
| 59 | + // Not sure how reliable this is but going with it for now. |
| 60 | + const std::string kImageVersionPath = "/sys/devices/soc0/image_version"; |
| 61 | + ET_LOG(Info, "Reading file %s", kImageVersionPath.c_str()); |
| 62 | + std::fstream image_version_file(kImageVersionPath, std::ios_base::in); |
| 63 | + if (image_version_file.is_open()) { |
| 64 | + std::string x; |
| 65 | + std::getline(image_version_file, x); |
| 66 | + // Hardcoding some rules for now |
| 67 | + if (x.find("S911") != std::string::npos) { |
| 68 | + // Samsung S23 has: |
| 69 | + // 1x3.36 GHz Cortex-X3 |
| 70 | + // 2x2.8 GHz Cortex-A715 |
| 71 | + // 2x2.8 GHz Cortex-A710 |
| 72 | + // 3x2.0 GHz Cortex-A510 |
| 73 | + // And we have balanced execution with 4 cores. |
| 74 | + return 4; |
| 75 | + } |
| 76 | + } |
| 77 | + ET_LOG(Info, "Failed to open midr file %s", kImageVersionPath.c_str()); |
| 78 | + return 0; |
| 79 | +} |
| 80 | + |
| 81 | +bool populate_available_cpu_mids() { |
| 82 | + std::vector<uint32_t>* cpu_midrs = get_static_cpu_midr_vector(); |
| 83 | + uint32_t num_possible_cores = cpuinfo_get_processors_count(); |
| 84 | + cpu_midrs->resize(num_possible_cores); |
| 85 | + const std::string kMidrFilePathPrefix = "/sys/devices/system/cpu/cpu"; |
| 86 | + const std::string kMidrFilePathSuffix = "/regs/identification/midr_el1"; |
| 87 | + for (int32_t i = 0; i < num_possible_cores; ++i) { |
| 88 | + std::string midr_file_path = |
| 89 | + kMidrFilePathPrefix + std::to_string(i) + kMidrFilePathSuffix; |
| 90 | + ET_LOG(Info, "Reading file %s", midr_file_path.c_str()); |
| 91 | + std::fstream midr_file(midr_file_path, std::ios_base::in); |
| 92 | + uint32_t tmp{0}; |
| 93 | + if (midr_file.is_open()) { |
| 94 | + std::string x; |
| 95 | + std::getline(midr_file, x); |
| 96 | + tmp = std::stoi(x, nullptr, 16); |
| 97 | + (*cpu_midrs)[i] = tmp; |
| 98 | + } else { |
| 99 | + ET_LOG(Info, "Failed to open midr file %s", midr_file_path.c_str()); |
| 100 | + cpu_midrs->clear(); |
| 101 | + return false; |
| 102 | + } |
| 103 | + } |
| 104 | + return true; |
| 105 | +} |
| 106 | + |
| 107 | +uint32_t _get_num_performant_cores() { |
| 108 | + // @lint-ignore CLANGTIDY facebook-hte-std::once_flag |
| 109 | + static std::once_flag flag; |
| 110 | + // @lint-ignore CLANGTIDY facebook-hte-std::call_once |
| 111 | + std::call_once(flag, []() { populate_available_cpu_mids(); }); |
| 112 | + std::vector<uint32_t>* cpu_midrs = get_static_cpu_midr_vector(); |
| 113 | + uint32_t num_possible_cores = cpuinfo_get_processors_count(); |
| 114 | + if (num_possible_cores != cpu_midrs->size()) { |
| 115 | + ET_LOG(Info, "CPU info and manual query on # of cpus dont match."); |
| 116 | + return 0; |
| 117 | + } |
| 118 | + for (int32_t i = 0; i < cpu_midrs->size(); ++i) { |
| 119 | + uint32_t masked_midr = (*cpu_midrs)[i] & RIVISION_MASK; |
| 120 | + switch (masked_midr) { |
| 121 | + case CPUINFO_ARM_MIDR_CORTEX_A520: |
| 122 | + case CPUINFO_ARM_MIDR_CORTEX_A53: |
| 123 | + case CPUINFO_ARM_MIDR_CORTEX_A55: |
| 124 | + case CPUINFO_ARM_MIDR_CORTEX_A57: |
| 125 | + num_possible_cores--; |
| 126 | + break; |
| 127 | + default: |
| 128 | + break; |
| 129 | + } |
| 130 | + } |
| 131 | + return num_possible_cores; |
| 132 | +} |
| 133 | + |
| 134 | +} // namespace |
| 135 | + |
| 136 | +uint32_t get_num_performant_cores() { |
| 137 | + ET_CHECK_MSG(cpuinfo_initialize(), "cpuinfo cannot be initialized."); |
| 138 | + // First try and see if we have number of cores profiled for this specific |
| 139 | + // device |
| 140 | + uint32_t model_specific_num_cores = _get_model_specific_num_cores(); |
| 141 | + if (model_specific_num_cores > 0) { |
| 142 | + return model_specific_num_cores; |
| 143 | + } |
| 144 | + |
| 145 | + // Else looks at either the # of litte cores if found |
| 146 | + // Or parse the midr in "Something seems wrong" section. |
| 147 | + const uint32_t uarch_count = cpuinfo_get_uarchs_count(); |
| 148 | + uint32_t num_possible_cores = cpuinfo_get_processors_count(); |
| 149 | + uint32_t num_non_performant_core = 0; |
| 150 | + if (uarch_count > 1) { |
| 151 | + for (int32_t i = 0; i < uarch_count; ++i) { |
| 152 | + const struct cpuinfo_uarch_info* uarch_info = cpuinfo_get_uarch(i); |
| 153 | + if (is_non_performant_core(uarch_info)) { |
| 154 | + num_non_performant_core += uarch_info->processor_count; |
| 155 | + } |
| 156 | + } |
| 157 | + ET_LOG(Info, "Number of efficient cores %d", num_non_performant_core); |
| 158 | + if (num_possible_cores <= num_non_performant_core) { |
| 159 | + ET_LOG( |
| 160 | + Info, "Total number of cores must be larger than efficient cores."); |
| 161 | + return 0; |
| 162 | + } |
| 163 | + return (num_possible_cores - num_non_performant_core); |
| 164 | + } else { |
| 165 | + // Something seems wrong. Lets check each processor's midr |
| 166 | + // In one plua 12 while it has 2 little cores, the topology |
| 167 | + // reported in /sys/devices/system/cpu/cpu* /topology/core_siblings_list |
| 168 | + // report wrong topology which results in wront configratuon |
| 169 | + return _get_num_performant_cores(); |
| 170 | + } |
| 171 | +} |
| 172 | + |
| 173 | +} // namespace cpuinfo |
| 174 | +} // namespace executorch |
| 175 | +} // namespace torch |
0 commit comments