ncnn/src/benchmark.cpp

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

#include "benchmark.h"

#if (__cplusplus >= 201103L || (defined(_MSVC_LANG) && _MSVC_LANG >= 201103L)) && !defined(__riscv) && !NCNN_SIMPLESTL
#define USE_CXX11_CLOCK 1
#else
#define USE_CXX11_CLOCK 0
#endif

#if USE_CXX11_CLOCK
#include <chrono>
#if NCNN_THREADS
#include <thread>
#endif
#include <numeric>
#include <algorithm>
#endif // USE_CXX11_CLOCK

#ifdef _WIN32
#define WIN32_LEAN_AND_MEAN
#include <windows.h>
#else                 // _WIN32
#include <sys/time.h> //gettimeofday()
#include <unistd.h>   // sleep()
#endif                // _WIN32

#if NCNN_BENCHMARK
#include "layer/convolution.h"
#include "layer/convolutiondepthwise.h"
#include "layer/deconvolution.h"
#include "layer/deconvolutiondepthwise.h"
#include "layer/convolution3d.h"
#include "layer/convolutiondepthwise3d.h"
#include "layer/deconvolution3d.h"
#include "layer/deconvolutiondepthwise3d.h"

#include <stdio.h>
#endif // NCNN_BENCHMARK

namespace ncnn {

double get_current_time()
{
#if USE_CXX11_CLOCK
    auto now = std::chrono::high_resolution_clock::now();
    auto usec = std::chrono::duration_cast<std::chrono::microseconds>(now.time_since_epoch());
    return usec.count() / 1000.0;
#else
#ifdef _WIN32
    LARGE_INTEGER freq;
    LARGE_INTEGER pc;
    QueryPerformanceFrequency(&freq);
    QueryPerformanceCounter(&pc);

    return pc.QuadPart * 1000.0 / freq.QuadPart;
#else  // _WIN32
    struct timeval tv;
    gettimeofday(&tv, NULL);

    return tv.tv_sec * 1000.0 + tv.tv_usec / 1000.0;
#endif // _WIN32
#endif
}

void sleep(unsigned long long int milliseconds)
{
#if USE_CXX11_CLOCK && NCNN_THREADS
    std::this_thread::sleep_for(std::chrono::milliseconds(milliseconds));
#else
#ifdef _WIN32
    Sleep(milliseconds);
#elif defined(__unix__) || defined(__APPLE__)
    usleep(milliseconds * 1000);
#elif _POSIX_TIMERS
    struct timespec ts;
    ts.tv_sec = milliseconds * 0.001;
    ts.tv_nsec = 0;
    nanosleep(&ts, &ts);
#else
    // TODO How to handle it ?
#endif
#endif
}

#if NCNN_BENCHMARK

void benchmark(const Layer* layer, double start, double end)
{
    fprintf(stderr, "%-24s %-30s %8.2lfms", layer->type.c_str(), layer->name.c_str(), end - start);
    fprintf(stderr, "    |");
    fprintf(stderr, "\n");
}

void benchmark(const Layer* layer, const Mat& bottom_blob, Mat& top_blob, double start, double end)
{
    fprintf(stderr, "%-24s %-30s %8.2lfms", layer->type.c_str(), layer->name.c_str(), end - start);

    char in_shape_str[64] = {'\0'};
    char out_shape_str[64] = {'\0'};

    if (bottom_blob.dims == 1)
    {
        sprintf(in_shape_str, "[%3d *%d]", bottom_blob.w, bottom_blob.elempack);
    }
    if (bottom_blob.dims == 2)
    {
        sprintf(in_shape_str, "[%3d, %3d *%d]", bottom_blob.w, bottom_blob.h, bottom_blob.elempack);
    }
    if (bottom_blob.dims == 3)
    {
        sprintf(in_shape_str, "[%3d, %3d, %3d *%d]", bottom_blob.w, bottom_blob.h, bottom_blob.c, bottom_blob.elempack);
    }
    if (bottom_blob.dims == 4)
    {
        sprintf(in_shape_str, "[%3d, %3d, %3d, %3d *%d]", bottom_blob.w, bottom_blob.h, bottom_blob.d, bottom_blob.c, bottom_blob.elempack);
    }

    if (top_blob.dims == 1)
    {
        sprintf(out_shape_str, "[%3d *%d]", top_blob.w, top_blob.elempack);
    }
    if (top_blob.dims == 2)
    {
        sprintf(out_shape_str, "[%3d, %3d *%d]", top_blob.w, top_blob.h, top_blob.elempack);
    }
    if (top_blob.dims == 3)
    {
        sprintf(out_shape_str, "[%3d, %3d, %3d *%d]", top_blob.w, top_blob.h, top_blob.c, top_blob.elempack);
    }
    if (top_blob.dims == 4)
    {
        sprintf(out_shape_str, "[%3d, %3d, %3d, %3d *%d]", top_blob.w, top_blob.h, top_blob.d, top_blob.c, top_blob.elempack);
    }

    fprintf(stderr, "    | %22s -> %-22s", in_shape_str, out_shape_str);

    if (layer->type == "Convolution")
    {
        fprintf(stderr, "     kernel: %1d x %1d     stride: %1d x %1d",
                ((Convolution*)layer)->kernel_w,
                ((Convolution*)layer)->kernel_h,
                ((Convolution*)layer)->stride_w,
                ((Convolution*)layer)->stride_h);
    }
    else if (layer->type == "ConvolutionDepthWise")
    {
        fprintf(stderr, "     kernel: %1d x %1d     stride: %1d x %1d",
                ((ConvolutionDepthWise*)layer)->kernel_w,
                ((ConvolutionDepthWise*)layer)->kernel_h,
                ((ConvolutionDepthWise*)layer)->stride_w,
                ((ConvolutionDepthWise*)layer)->stride_h);
    }
    else if (layer->type == "Deconvolution")
    {
        fprintf(stderr, "     kernel: %1d x %1d     stride: %1d x %1d",
                ((Deconvolution*)layer)->kernel_w,
                ((Deconvolution*)layer)->kernel_h,
                ((Deconvolution*)layer)->stride_w,
                ((Deconvolution*)layer)->stride_h);
    }
    else if (layer->type == "DeconvolutionDepthWise")
    {
        fprintf(stderr, "     kernel: %1d x %1d     stride: %1d x %1d",
                ((DeconvolutionDepthWise*)layer)->kernel_w,
                ((DeconvolutionDepthWise*)layer)->kernel_h,
                ((DeconvolutionDepthWise*)layer)->stride_w,
                ((DeconvolutionDepthWise*)layer)->stride_h);
    }
    else if (layer->type == "Convolution3D")
    {
        fprintf(stderr, "     kernel: %1d x %1d x %1d    stride: %1d x %1d x %1d",
                ((Convolution3D*)layer)->kernel_w,
                ((Convolution3D*)layer)->kernel_h,
                ((Convolution3D*)layer)->kernel_d,
                ((Convolution3D*)layer)->stride_w,
                ((Convolution3D*)layer)->stride_h,
                ((Convolution3D*)layer)->stride_d);
    }
    else if (layer->type == "ConvolutionDepthWise3D")
    {
        fprintf(stderr, "     kernel: %1d x %1d x %1d    stride: %1d x %1d x %1d",
                ((ConvolutionDepthWise3D*)layer)->kernel_w,
                ((ConvolutionDepthWise3D*)layer)->kernel_h,
                ((ConvolutionDepthWise3D*)layer)->kernel_d,
                ((ConvolutionDepthWise3D*)layer)->stride_w,
                ((ConvolutionDepthWise3D*)layer)->stride_h,
                ((ConvolutionDepthWise3D*)layer)->stride_d);
    }
    else if (layer->type == "Deconvolution3D")
    {
        fprintf(stderr, "     kernel: %1d x %1d x %1d    stride: %1d x %1d x %1d",
                ((Deconvolution3D*)layer)->kernel_w,
                ((Deconvolution3D*)layer)->kernel_h,
                ((Deconvolution3D*)layer)->kernel_d,
                ((Deconvolution3D*)layer)->stride_w,
                ((Deconvolution3D*)layer)->stride_h,
                ((Deconvolution3D*)layer)->stride_d);
    }
    else if (layer->type == "DeconvolutionDepthWise3D")
    {
        fprintf(stderr, "     kernel: %1d x %1d x %1d    stride: %1d x %1d x %1d",
                ((DeconvolutionDepthWise3D*)layer)->kernel_w,
                ((DeconvolutionDepthWise3D*)layer)->kernel_h,
                ((DeconvolutionDepthWise3D*)layer)->kernel_d,
                ((DeconvolutionDepthWise3D*)layer)->stride_w,
                ((DeconvolutionDepthWise3D*)layer)->stride_h,
                ((DeconvolutionDepthWise3D*)layer)->stride_d);
    }
    fprintf(stderr, "\n");
}

#endif // NCNN_BENCHMARK

} // namespace ncnn