Speed up cimg::fsize

[apocelipes]

In a POSIX environment, there is a faster way to retrieve a file's size that does not modify the file offset.

This can roughly improve fsize performance by 3x–10x

Benchmark code:

// A 1 MiB file. However size does not matter here.
static const char* filePath = "test.bin";

typedef int64_t cimg_int64;

inline cimg_int64 fsize(std::FILE *file) {
    if (!file) return (cimg_int64)-1;
    const long pos = std::ftell(file);
    std::fseek(file,0,SEEK_END);
    const cimg_int64 siz = (cimg_int64)std::ftell(file);
    std::fseek(file,pos,SEEK_SET);
    return siz;
}
inline cimg_int64 fsize(const char *const filename) {
    std::FILE *const file = std::fopen(filename,"rb");
    const cimg_int64 siz = fsize(file);
    std::fclose(file);
    return siz;
}

inline cimg_int64 fsizeOptimize(std::FILE *file) {
    if (!file) return (cimg_int64)-1;
    int fd = fileno(file);
    struct stat st;
    if (fstat(fd, &st) < 0) return -1;
    return (cimg_int64)st.st_size;
}
inline cimg_int64 fsizeOptimize(const char *const filename) {
    struct stat st;
    if (stat(filename, &st) < 0) return -1;
    return (cimg_int64)st.st_size;
}

static void BM_fsize_FILE(benchmark::State& state) {
    FILE* fp = std::fopen(filePath, "rb");
    for (auto _ : state) {
        benchmark::DoNotOptimize(fsize(fp));
        benchmark::ClobberMemory();
    }
    std::fclose(fp);
}
BENCHMARK(BM_fsize_FILE);
static void BM_fsize_filename(benchmark::State& state) {
    for (auto _ : state) {
        benchmark::DoNotOptimize(fsize(filePath));
        benchmark::ClobberMemory();
    }
}
BENCHMARK(BM_fsize_filename);

static void BM_fsizeOptimize_FILE(benchmark::State& state) {
    FILE* fp = std::fopen(filePath, "rb");
    for (auto _ : state) {
        benchmark::DoNotOptimize(fsizeOptimize(fp));
        benchmark::ClobberMemory();
    }
    std::fclose(fp);
}
BENCHMARK(BM_fsizeOptimize_FILE);
static void BM_fsizeOptimize_filename(benchmark::State& state) {
    for (auto _ : state) {
        benchmark::DoNotOptimize(fsizeOptimize(filePath));
        benchmark::ClobberMemory();
    }
}
BENCHMARK(BM_fsizeOptimize_filename);

Results:

# macOS
Run on (10 X 24 MHz CPU s)
CPU Caches:
  L1 Data 64 KiB
  L1 Instruction 128 KiB
  L2 Unified 4096 KiB (x10)
Load Average: 2.10, 2.08, 1.88
--------------------------------------------------------------------
Benchmark                          Time             CPU   Iterations
--------------------------------------------------------------------
BM_fsize_FILE                    604 ns          604 ns      1196254
BM_fsize_filename               6182 ns         6182 ns       111123
BM_fsizeOptimize_FILE            227 ns          227 ns      3107534
BM_fsizeOptimize_filename        405 ns          405 ns      1726758

# Linux
Run on (4 X 48 MHz CPU s)
Load Average: 0.13, 0.07, 0.02
--------------------------------------------------------------------
Benchmark                          Time             CPU   Iterations
--------------------------------------------------------------------
BM_fsize_FILE                    371 ns          370 ns      1871449
BM_fsize_filename               1213 ns         1208 ns       566134
BM_fsizeOptimize_FILE            141 ns          140 ns      4872801
BM_fsizeOptimize_filename        199 ns          198 ns      3489223

[apocelipes]

Conflicts are resolved.

[apocelipes]

The optimization path for Windows has also been added.

[dtschump]

I've included/rewritten your patch, with a few changes for better portability :

    //! Get file size.
    /**
       \param file Specified file to get size from.
       \return File size or '-1' if file does not exist.
    **/
    inline cimg_int64 fsize(std::FILE *const file) {
      if (!file) return (cimg_int64)-1;

#if cimg_OS==1 // Optimized for POSIX Environements (Linux, macOS, BSD, etc.)
      const int fd = fileno(file);
      struct stat st;
      if (fd>=0 && !fstat(fd,&st)) return (cimg_int64)st.st_size;
#elif cimg_OS==2 // Optimized for Windows Environements (MSVC, MinGW)
      const int fd = _fileno(file);
      struct _stati64 st;
      if (fd>=0 && !_fstati64(fd,&st)) return (cimg_int64)st.st_size;
#endif

      // Fallback, used for non-POSIX systems or if fstat failed (pipes or sockets).
      const cimg_long pos = cimg::ftell(file);
      cimg::fseek(file, 0, SEEK_END);
      const cimg_int64 siz = (cimg_int64)cimg::ftell(file);
      cimg::fseek(file, pos, SEEK_SET);
      return siz;
    }

    //! Get file size from filename.
    /**
       \param filename Specified filename to get size from.
       \return File size or '-1' if file does not exist.
    **/
    inline cimg_int64 fsize(const char *const filename) {
      if (!filename || !*filename) return (cimg_int64)-1;

#if cimg_OS==1 // Optimized for POSIX Environements (Linux, macOS, BSD, etc.)
      struct stat st;
      if (!stat(filename,&st)) return (cimg_int64)st.st_size;
#elif cimg_OS==2 // Optimized for Windows Environements (MSVC, MinGW)
      // Convert UTF-8 path to wchar_t.
      int len = MultiByteToWideChar(CP_UTF8,0,filename,-1,0,0);
      if (len>0) {
        CImg<wchar_t> wfilename((unsigned int)len);
        if (MultiByteToWideChar(CP_UTF8,0,filename,-1,wfilename,len)) {
          struct _stati64 st;
          if (!_wstati64(wfilename,&st)) return (cimg_int64)st.st_size;
        }
      }
#endif

      // Fallback.
      std::FILE *const file = cimg::std_fopen(filename, "rb");
      if (!file) return (cimg_int64)-1;
      const cimg_int64 siz = cimg::fsize(file);
      cimg::fclose(file);
      return siz;
    }

What do you think ?

[apocelipes]

I've included/rewritten your patch, with a few changes for better portability :

    //! Get file size.
    /**
       \param file Specified file to get size from.
       \return File size or '-1' if file does not exist.
    **/
    inline cimg_int64 fsize(std::FILE *const file) {
      if (!file) return (cimg_int64)-1;

#if cimg_OS==1 // Optimized for POSIX Environements (Linux, macOS, BSD, etc.)
      const int fd = fileno(file);
      struct stat st;
      if (fd>=0 && !fstat(fd,&st)) return (cimg_int64)st.st_size;
#elif cimg_OS==2 // Optimized for Windows Environements (MSVC, MinGW)
      const int fd = _fileno(file);
      struct _stati64 st;
      if (fd>=0 && !_fstati64(fd,&st)) return (cimg_int64)st.st_size;
#endif

      // Fallback, used for non-POSIX systems or if fstat failed (pipes or sockets).
      const cimg_long pos = cimg::ftell(file);
      cimg::fseek(file, 0, SEEK_END);
      const cimg_int64 siz = (cimg_int64)cimg::ftell(file);
      cimg::fseek(file, pos, SEEK_SET);
      return siz;
    }

    //! Get file size from filename.
    /**
       \param filename Specified filename to get size from.
       \return File size or '-1' if file does not exist.
    **/
    inline cimg_int64 fsize(const char *const filename) {
      if (!filename || !*filename) return (cimg_int64)-1;

#if cimg_OS==1 // Optimized for POSIX Environements (Linux, macOS, BSD, etc.)
      struct stat st;
      if (!stat(filename,&st)) return (cimg_int64)st.st_size;
#elif cimg_OS==2 // Optimized for Windows Environements (MSVC, MinGW)
      // Convert UTF-8 path to wchar_t.
      int len = MultiByteToWideChar(CP_UTF8,0,filename,-1,0,0);
      if (len>0) {
        CImg<wchar_t> wfilename((unsigned int)len);
        if (MultiByteToWideChar(CP_UTF8,0,filename,-1,wfilename,len)) {
          struct _stati64 st;
          if (!_wstati64(wfilename,&st)) return (cimg_int64)st.st_size;
        }
      }
#endif

      // Fallback.
      std::FILE *const file = cimg::std_fopen(filename, "rb");
      if (!file) return (cimg_int64)-1;
      const cimg_int64 siz = cimg::fsize(file);
      cimg::fclose(file);
      return siz;
    }

What do you think ?

@dtschump That's great! Additionally, fstat itself checks whether the file descriptor (fd) is valid, so this eliminates the need for a separate conditional branch. Given the typical use cases for CImg, it is highly unlikely that the obtained file descriptor (fd) would be invalid. So I removed that check in the new commit.

[dtschump]

OK, I've tested on Windows, and it does not work. I had to do other modifications to make it work.

[apocelipes]

I resolved the conflict. _filelengthi64 also no need to check fd so I removed the "if (fd>=0)"

[apocelipes]

@dtschump I fixed Windows code. Please take a look.

[apocelipes]

P.S. I don't really recommend using the POSIX-compatible functions provided by crt; they're slower than the native API and behave differently from POSIX.