platform.cpp

/*******************************************************************************
* Copyright 2020-2025 Intel Corporation
* Copyright 2020-2024 FUJITSU LIMITED
* Copyright 2022-2024 Arm Ltd. and affiliates
*
* 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.
*******************************************************************************/

#include <thread>

#include "cpu/platform.hpp"

#if DNNL_CPU_THREADING_RUNTIME == DNNL_RUNTIME_THREADPOOL
#include <algorithm>

#if defined(_WIN32)
#include <windows.h>
#elif defined(__GLIBC__)
#include <sched.h>
#endif
#endif

#if DNNL_X64
#include "cpu/x64/cpu_isa_traits.hpp"
#elif DNNL_AARCH64
#include "cpu/aarch64/cpu_isa_traits.hpp"
#if DNNL_AARCH64_USE_ACL
// For checking if fp16 isa is supported on the platform
#include "arm_compute/core/CPP/CPPTypes.h"
#endif
#endif

// For DNNL_X64 build we compute the timestamp using rdtsc. Use std::chrono for
// other builds.
#if !DNNL_X64
#include <chrono>
#endif

namespace dnnl {
namespace impl {
namespace cpu {
namespace platform {

const char *get_isa_info() {
#if DNNL_X64
    return x64::get_isa_info();
#elif DNNL_AARCH64
    return aarch64::get_isa_info();
#else
    return "Generic";
#endif
}

dnnl_cpu_isa_t get_effective_cpu_isa() {
#if DNNL_X64
    return x64::get_effective_cpu_isa();
#elif DNNL_AARCH64
    return aarch64::get_effective_cpu_isa();
#else
    return dnnl_cpu_isa_default;
#endif
}

status_t set_max_cpu_isa(dnnl_cpu_isa_t isa) {
#if DNNL_X64
    return x64::set_max_cpu_isa(isa);
#else
    return status::unimplemented;
#endif
}

status_t set_cpu_isa_hints(dnnl_cpu_isa_hints_t isa_hints) {
#if DNNL_X64
    return x64::set_cpu_isa_hints(isa_hints);
#elif DNNL_AARCH64
    return status::success;
#else
    return status::unimplemented;
#endif
}

dnnl_cpu_isa_hints_t get_cpu_isa_hints() {
#if DNNL_X64
    return x64::get_cpu_isa_hints();
#else
    return dnnl_cpu_isa_no_hints;
#endif
}

bool prefer_ymm_requested() {
#if DNNL_X64
    const bool prefer_ymm = x64::get_cpu_isa_hints() == dnnl_cpu_isa_prefer_ymm;
    return prefer_ymm;
#else
    return false;
#endif
}

bool has_data_type_support(data_type_t data_type) {
    // Notice: see notes in header
    switch (data_type) {
        case data_type::bf16:
#if DNNL_X64
            return x64::mayiuse(x64::avx512_core)
                    || x64::mayiuse(x64::avx2_vnni_2);
#elif DNNL_PPC64
#if defined(USE_CBLAS) && defined(BLAS_HAS_SBGEMM) && defined(__MMA__)
            return true;
#endif
#elif DNNL_AARCH64
            return aarch64::mayiuse_bf16();
#else
            return false;
#endif
        case data_type::f16:
#if DNNL_X64
            return x64::mayiuse(x64::avx512_core_fp16)
                    || x64::mayiuse(x64::avx2_vnni_2);
#elif DNNL_AARCH64_USE_ACL
            return arm_compute::CPUInfo::get().has_fp16();
#else
            return false;
#endif
        case data_type::f8_e5m2:
        case data_type::f8_e4m3:
#if DNNL_X64
            return x64::mayiuse(x64::avx512_core_fp16);
#else
            return false;
#endif
        default: return true;
    }
}

bool has_training_support(data_type_t data_type) {
    // TODO: maybe return false for int8, but some primitives like prelu
    // have training support
    switch (data_type) {
        case data_type::bf16:
#if DNNL_X64
            return x64::mayiuse(x64::avx512_core);
#elif DNNL_PPC64
#if defined(USE_CBLAS) && defined(BLAS_HAS_SBGEMM) && defined(__MMA__)
            return true;
#endif
#elif DNNL_AARCH64_USE_ACL
            return arm_compute::CPUInfo::get().has_bf16();
#else
            return false;
#endif
        case data_type::f16:
#if DNNL_X64
            return x64::mayiuse(x64::avx512_core_fp16);
#elif DNNL_AARCH64_USE_ACL
            return arm_compute::CPUInfo::get().has_fp16();
#else
            return false;
#endif
        default: return true;
    }
}

float s8s8_weights_scale_factor() {
#if DNNL_X64
    return x64::mayiuse(x64::avx512_core_vnni) || x64::mayiuse(x64::avx2_vnni)
            ? 1.0f
            : 0.5f;
#else
    return 1.0f;
#endif
}

unsigned get_per_core_cache_size(int level) {
    auto guess = [](int level) {
        switch (level) {
            case 1: return 32U * 1024;
            case 2: return 512U * 1024;
            case 3: return 1024U * 1024;
            default: return 0U;
        }
    };

#if DNNL_X64
    using namespace x64;
    if (cpu().getDataCacheLevels() == 0) return guess(level);

    if (level > 0 && (unsigned)level <= cpu().getDataCacheLevels()) {
        unsigned l = level - 1;
        return cpu().getDataCacheSize(l) / cpu().getCoresSharingDataCache(l);
    } else
        return 0;
#else
    return guess(level);
#endif
}

unsigned get_num_cores() {
#if DNNL_X64
    return x64::cpu().getNumCores(Xbyak::util::CoreLevel);
#elif DNNL_AARCH64_USE_ACL
    return aarch64::cpu().getNumCores(Xbyak_aarch64::util::CoreLevel);
#else
    return 1;
#endif
}

#if DNNL_CPU_THREADING_RUNTIME == DNNL_RUNTIME_THREADPOOL
// The purpose of this function is to return the potential maximum number of
// threads in user's threadpool. It is assumed that the number of threads in an
// actual threadpool will not exceed the number cores in a socket reported by
// the OS, which may or may not be equal to the number of total physical cores
// in a socket depending on the OS configuration (read -- VM environment). In
// order to simulate the number of cores available in such environment, this
// function supports process affinity.
unsigned get_max_threads_to_use() {
    // TODO: the logic below should involve number of sockets to provide exact
    // number of cores on 2+ socket systems.
    int num_cores_per_socket = (int)dnnl::impl::cpu::platform::get_num_cores();
    // It may happen that XByak doesn't get num of threads identified, e.g. for
    // AMD. In order to make threadpool working, we supply an additional
    // condition to have some reasonable number of threads available at
    // primitive descriptor creation time.
    if (num_cores_per_socket == 0)
        num_cores_per_socket = std::thread::hardware_concurrency();

#if defined(_WIN32)
    DWORD_PTR proc_affinity_mask;
    DWORD_PTR sys_affinity_mask;
    if (GetProcessAffinityMask(
                GetCurrentProcess(), &proc_affinity_mask, &sys_affinity_mask)) {
        int masked_nthr = 0;
        for (int i = 0; i < CHAR_BIT * sizeof(proc_affinity_mask);
                i++, proc_affinity_mask >>= 1)
            masked_nthr += proc_affinity_mask & 1;
        return std::min(masked_nthr, num_cores_per_socket);
    }
#elif defined(__GLIBC__)
    cpu_set_t cpu_set;
    // Check if the affinity of the process has been set using, e.g.,
    // numactl.
    if (::sched_getaffinity(0, sizeof(cpu_set_t), &cpu_set) == 0)
        return std::min(CPU_COUNT(&cpu_set), num_cores_per_socket);
#endif
    return num_cores_per_socket;
}
#endif

int get_vector_register_size() {
#if DNNL_X64
    using namespace x64;
    if (mayiuse(avx512_core)) return cpu_isa_traits_t<avx512_core>::vlen;
    if (mayiuse(avx)) return cpu_isa_traits_t<avx>::vlen;
    if (mayiuse(sse41)) return cpu_isa_traits_t<sse41>::vlen;
#elif DNNL_AARCH64
    using namespace aarch64;
    if (mayiuse(asimd)) return cpu_isa_traits<asimd>::vlen;
    if (mayiuse(sve_512)) return cpu_isa_traits<sve_512>::vlen;
    if (mayiuse(sve_256)) return cpu_isa_traits<sve_256>::vlen;
#endif
    return 0;
}

/* The purpose of this function is to provide a very efficient timestamp
 * calculation (used primarily for primitive cache). For DNNL_X64, this can be
 * accomplished using *rdtsc* since it provides a timestamp value that (i) is
 * independent for each core, and (ii) is synchronized across cores in multiple
 * sockets.
 * TODO: For now, use std::chrono::steady_clock for other builds, however
 * another more optimized function may be called here.
 */
size_t get_timestamp() {
#if DNNL_X64
    return static_cast<size_t>(Xbyak::util::Clock::getRdtsc());
#else
    return static_cast<size_t>(
            std::chrono::steady_clock::now().time_since_epoch().count());
#endif
}

} // namespace platform
} // namespace cpu
} // namespace impl
} // namespace dnnl