ncnn/src/platform.h.in

// Copyright 2017 Tencent
// SPDX-License-Identifier: BSD-3-Clause

#ifndef NCNN_PLATFORM_H
#define NCNN_PLATFORM_H

#cmakedefine01 NCNN_STDIO
#cmakedefine01 NCNN_STRING
#cmakedefine01 NCNN_SIMPLEOCV
#cmakedefine01 NCNN_SIMPLEOMP
#cmakedefine01 NCNN_SIMPLESTL
#cmakedefine01 NCNN_SIMPLEMATH
#cmakedefine01 NCNN_THREADS
#cmakedefine01 NCNN_BENCHMARK
#cmakedefine01 NCNN_C_API
#cmakedefine01 NCNN_PLATFORM_API
#cmakedefine01 NCNN_PIXEL
#cmakedefine01 NCNN_PIXEL_ROTATE
#cmakedefine01 NCNN_PIXEL_AFFINE
#cmakedefine01 NCNN_PIXEL_DRAWING
#cmakedefine01 NCNN_VULKAN
#cmakedefine01 NCNN_SIMPLEVK
#cmakedefine01 NCNN_SYSTEM_GLSLANG
#cmakedefine01 NCNN_RUNTIME_CPU
#cmakedefine01 NCNN_GNU_INLINE_ASM
#cmakedefine01 NCNN_AVX
#cmakedefine01 NCNN_XOP
#cmakedefine01 NCNN_FMA
#cmakedefine01 NCNN_F16C
#cmakedefine01 NCNN_AVX2
#cmakedefine01 NCNN_AVXVNNI
#cmakedefine01 NCNN_AVXVNNIINT8
#cmakedefine01 NCNN_AVXVNNIINT16
#cmakedefine01 NCNN_AVXNECONVERT
#cmakedefine01 NCNN_AVX512
#cmakedefine01 NCNN_AVX512VNNI
#cmakedefine01 NCNN_AVX512BF16
#cmakedefine01 NCNN_AVX512FP16
#cmakedefine01 NCNN_VFPV4
#cmakedefine01 NCNN_ARM82
#cmakedefine01 NCNN_ARM82DOT
#cmakedefine01 NCNN_ARM82FP16FML
#cmakedefine01 NCNN_ARM84BF16
#cmakedefine01 NCNN_ARM84I8MM
#cmakedefine01 NCNN_ARM86SVE
#cmakedefine01 NCNN_ARM86SVE2
#cmakedefine01 NCNN_ARM86SVEBF16
#cmakedefine01 NCNN_ARM86SVEI8MM
#cmakedefine01 NCNN_ARM86SVEF32MM
#cmakedefine01 NCNN_MSA
#cmakedefine01 NCNN_LSX
#cmakedefine01 NCNN_MMI
#cmakedefine01 NCNN_RVV
#cmakedefine01 NCNN_ZFH
#cmakedefine01 NCNN_ZVFH
#cmakedefine01 NCNN_XTHEADVECTOR
#cmakedefine01 NCNN_INT8
#cmakedefine01 NCNN_BF16
#cmakedefine01 NCNN_FORCE_INLINE

#cmakedefine NCNN_VERSION_STRING "@NCNN_VERSION_STRING@"

#include "ncnn_export.h"

#ifdef __cplusplus

#if NCNN_THREADS
#if defined _WIN32
#define WIN32_LEAN_AND_MEAN
#ifndef _WIN32_WINNT
#define _WIN32_WINNT 0x0501
#endif
#include <windows.h>
#include <process.h>
#else
#include <pthread.h>
#endif
#endif // NCNN_THREADS

#if __ANDROID_API__ >= 26
#ifndef VK_USE_PLATFORM_ANDROID_KHR
#define VK_USE_PLATFORM_ANDROID_KHR
#endif
#endif // __ANDROID_API__ >= 26

namespace ncnn {

#if NCNN_THREADS
#if defined _WIN32
#if _WIN32_WINNT > _WIN32_WINNT_WINXP // Windows Vista and later
class NCNN_EXPORT Mutex
{
public:
    Mutex() { InitializeSRWLock(&srwlock); }
    ~Mutex() {}
    void lock() { AcquireSRWLockExclusive(&srwlock); }
    void unlock() { ReleaseSRWLockExclusive(&srwlock); }
private:
    friend class ConditionVariable;
    SRWLOCK srwlock;
};

class NCNN_EXPORT ConditionVariable
{
public:
    ConditionVariable() { InitializeConditionVariable(&condvar); }
    ~ConditionVariable() {}
    void wait(Mutex& mutex) { SleepConditionVariableSRW(&condvar, &mutex.srwlock, INFINITE, 0); }
    void broadcast() { WakeAllConditionVariable(&condvar); }
    void signal() { WakeConditionVariable(&condvar); }
private:
    CONDITION_VARIABLE condvar;
};

#else // Windows XP compatibility

class NCNN_EXPORT Mutex
{
public:
    Mutex() { InitializeCriticalSection(&cs); }
    ~Mutex() { DeleteCriticalSection(&cs); }
    void lock() { EnterCriticalSection(&cs); }
    void unlock() { LeaveCriticalSection(&cs); }
private:
    friend class ConditionVariable;
    CRITICAL_SECTION cs;
};

class NCNN_EXPORT ConditionVariable
{
public:
    ConditionVariable() 
    { 
        signal_event = CreateEvent(0, FALSE, FALSE, 0); // Auto-reset event for signal()
        broadcast_event = CreateEvent(0, TRUE, FALSE, 0); // Manual-reset event for broadcast()
    }
    ~ConditionVariable() 
    { 
        CloseHandle(signal_event); 
        CloseHandle(broadcast_event); 
    }
    void wait(Mutex& mutex)
    {
        mutex.unlock();
        HANDLE events[2] = { signal_event, broadcast_event };
        WaitForMultipleObjects(2, events, FALSE, INFINITE); // Wait for either signal or broadcast
        mutex.lock();
    }
    void broadcast() 
    { 
        SetEvent(broadcast_event); // Wake all threads
        ResetEvent(broadcast_event); // Reset after waking all threads
    }
    void signal() 
    { 
        SetEvent(signal_event); // Wake one thread
    }
private:
    HANDLE signal_event;
    HANDLE broadcast_event;
};

#endif // _WIN32_WINNT > _WIN32_WINNT_WINXP

static unsigned __stdcall start_wrapper(void* args);
class NCNN_EXPORT Thread
{
public:
    Thread(void* (*start)(void*), void* args = 0) { _start = start; _args = args; handle = (HANDLE)_beginthreadex(0, 0, start_wrapper, this, 0, 0); }
    ~Thread() {}
    void join() { WaitForSingleObject(handle, INFINITE); CloseHandle(handle); }
private:
    friend unsigned __stdcall start_wrapper(void* args)
    {
        Thread* t = (Thread*)args;
        t->_start(t->_args);
        return 0;
    }
    HANDLE handle;
    void* (*_start)(void*);
    void* _args;
};

class NCNN_EXPORT ThreadLocalStorage
{
public:
    ThreadLocalStorage() { key = TlsAlloc(); }
    ~ThreadLocalStorage() { TlsFree(key); }
    void set(void* value) { TlsSetValue(key, (LPVOID)value); }
    void* get() { return (void*)TlsGetValue(key); }
private:
    DWORD key;
};
#else // defined _WIN32
class NCNN_EXPORT Mutex
{
public:
    Mutex() { pthread_mutex_init(&mutex, 0); }
    ~Mutex() { pthread_mutex_destroy(&mutex); }
    void lock() { pthread_mutex_lock(&mutex); }
    void unlock() { pthread_mutex_unlock(&mutex); }
private:
    friend class ConditionVariable;
    pthread_mutex_t mutex;
};

class NCNN_EXPORT ConditionVariable
{
public:
    ConditionVariable() { pthread_cond_init(&cond, 0); }
    ~ConditionVariable() { pthread_cond_destroy(&cond); }
    void wait(Mutex& mutex) { pthread_cond_wait(&cond, &mutex.mutex); }
    void broadcast() { pthread_cond_broadcast(&cond); }
    void signal() { pthread_cond_signal(&cond); }
private:
    pthread_cond_t cond;
};

class NCNN_EXPORT Thread
{
public:
    Thread(void* (*start)(void*), void* args = 0) { pthread_create(&t, 0, start, args); }
    ~Thread() {}
    void join() { pthread_join(t, 0); }
private:
    pthread_t t;
};

class NCNN_EXPORT ThreadLocalStorage
{
public:
    ThreadLocalStorage() { pthread_key_create(&key, 0); }
    ~ThreadLocalStorage() { pthread_key_delete(key); }
    void set(void* value) { pthread_setspecific(key, value); }
    void* get() { return pthread_getspecific(key); }
private:
    pthread_key_t key;
};
#endif // defined _WIN32
#else // NCNN_THREADS
class NCNN_EXPORT Mutex
{
public:
    Mutex() {}
    ~Mutex() {}
    void lock() {}
    void unlock() {}
};

class NCNN_EXPORT ConditionVariable
{
public:
    ConditionVariable() {}
    ~ConditionVariable() {}
    void wait(Mutex& /*mutex*/) {}
    void broadcast() {}
    void signal() {}
};

class NCNN_EXPORT Thread
{
public:
    Thread(void* (*/*start*/)(void*), void* /*args*/ = 0) {}
    ~Thread() {}
    void join() {}
};

class NCNN_EXPORT ThreadLocalStorage
{
public:
    ThreadLocalStorage() { data = 0; }
    ~ThreadLocalStorage() {}
    void set(void* value) { data = value; }
    void* get() { return data; }
private:
    void* data;
};
#endif // NCNN_THREADS

class NCNN_EXPORT MutexLockGuard
{
public:
    MutexLockGuard(Mutex& _mutex) : mutex(_mutex) { mutex.lock(); }
    ~MutexLockGuard() { mutex.unlock(); }
private:
    Mutex& mutex;
};

static inline void swap_endianness_16(void* x)
{
    unsigned char* xx = (unsigned char*)x;
    unsigned char x0 = xx[0];
    unsigned char x1 = xx[1];
    xx[0] = x1;
    xx[1] = x0;
}

static inline void swap_endianness_32(void* x)
{
    unsigned char* xx = (unsigned char*)x;
    unsigned char x0 = xx[0];
    unsigned char x1 = xx[1];
    unsigned char x2 = xx[2];
    unsigned char x3 = xx[3];
    xx[0] = x3;
    xx[1] = x2;
    xx[2] = x1;
    xx[3] = x0;
}

} // namespace ncnn

#if NCNN_SIMPLESTL
#include "simplestl.h"
#else
#include <algorithm>
#include <list>
#include <vector>
#include <stack>
#include <string>
#endif

// simplemath
#if NCNN_SIMPLEMATH
#include "simplemath.h"
#else
#include <math.h>
#include <fenv.h>
#endif

#if NCNN_VULKAN
#if NCNN_SIMPLEVK
#include "simplevk.h"
#else
#include <vulkan/vulkan.h>
#endif
#include "vulkan_header_fix.h"
#endif // NCNN_VULKAN

#endif // __cplusplus

#if NCNN_STDIO
#if NCNN_PLATFORM_API && __ANDROID_API__ >= 8
#include <android/log.h>
#define NCNN_LOGE(...) do { \
    fprintf(stderr, ##__VA_ARGS__); fprintf(stderr, "\n"); \
    __android_log_print(ANDROID_LOG_WARN, "ncnn", ##__VA_ARGS__); } while(0)
#else // NCNN_PLATFORM_API && __ANDROID_API__ >= 8
#include <stdio.h>
#define NCNN_LOGE(...) do { \
    fprintf(stderr, ##__VA_ARGS__); fprintf(stderr, "\n"); } while(0)
#endif // NCNN_PLATFORM_API && __ANDROID_API__ >= 8
#else
#define NCNN_LOGE(...)
#endif


#if NCNN_FORCE_INLINE
#ifdef _MSC_VER
    #define NCNN_FORCEINLINE __forceinline
#elif defined(__GNUC__)
    #define NCNN_FORCEINLINE inline __attribute__((__always_inline__))
#elif defined(__CLANG__)
    #if __has_attribute(__always_inline__)
        #define NCNN_FORCEINLINE inline __attribute__((__always_inline__))
    #else
        #define NCNN_FORCEINLINE inline
    #endif
#else
    #define NCNN_FORCEINLINE inline
#endif
#else
    #define NCNN_FORCEINLINE inline
#endif

#endif // NCNN_PLATFORM_H