get_physical_cpu_count api family

[nihui]

• test on linux
• test on android
• test on ios
• test on macos
• test on windows
• test on big.little cpu
• test on smp cpu (eg most x86)
• test on smp + big.little cpu (eg intel 12th/13th)
• test on 3-cluster cpu (eg qcom 8gen1)
• resolve interaction with thread affinity
• set for option default
• set for benchncnn default

[codecov-commenter]

Codecov Report

Merging #4302 (4356346) into master (9c6f110) will decrease coverage by 0.01%.
The diff coverage is 49.18%.

@@            Coverage Diff             @@
##           master    #4302      +/-   ##
==========================================
- Coverage   91.52%   91.51%   -0.02%     
==========================================
  Files         783      783              
  Lines      183445   183505      +60     
==========================================
+ Hits       167903   167938      +35     
- Misses      15542    15567      +25     

Impacted Files	Coverage Δ
src/cpu.cpp	61.64% <48.33%> (-0.47%)	⬇️
src/option.cpp	100.00% <100.00%> (ø)

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

[kagurazakakotori]

Tested on i9-12900K (E-cores and HT enabled) with Ubuntu 22.10 + kernel 5.19, num_threads is 8 by default in benchncnn. On my setup, it does not run on E-cores.

$ ./benchncnn 
loop_count = 4
num_threads = 8
powersave = 2
gpu_device = -1
cooling_down = 1
          squeezenet  min =    1.94  max =    1.94  avg =    1.94
     squeezenet_int8  min =    3.37  max =    3.38  avg =    3.38
           mobilenet  min =    2.41  max =    2.49  avg =    2.44
      mobilenet_int8  min =    3.69  max =    3.80  avg =    3.75
        mobilenet_v2  min =    2.39  max =    2.85  avg =    2.57
        mobilenet_v3  min =    1.56  max =    1.65  avg =    1.60
          shufflenet  min =    2.04  max =    2.09  avg =    2.06
       shufflenet_v2  min =    2.01  max =    2.15  avg =    2.06
             mnasnet  min =    1.74  max =    1.78  avg =    1.76
     proxylessnasnet  min =    2.80  max =    3.15  avg =    2.98
     efficientnet_b0  min =    3.87  max =    3.98  avg =    3.94
   efficientnetv2_b0  min =    4.20  max =    4.61  avg =    4.31
        regnety_400m  min =    4.86  max =    4.96  avg =    4.90
           blazeface  min =    0.70  max =    0.72  avg =    0.71
           googlenet  min =    6.47  max =    6.58  avg =    6.52
      googlenet_int8  min =    8.18  max =    8.25  avg =    8.22
            resnet18  min =    5.19  max =    5.41  avg =    5.33
       resnet18_int8  min =    6.85  max =    6.97  avg =    6.90
             alexnet  min =    4.24  max =    4.59  avg =    4.35
               vgg16  min =   22.17  max =   23.28  avg =   22.94
          vgg16_int8  min =   19.29  max =   20.24  avg =   19.91
            resnet50  min =   10.80  max =   12.17  avg =   11.47
       resnet50_int8  min =   11.71  max =   19.37  avg =   15.53
      squeezenet_ssd  min =    7.87  max =    8.12  avg =    8.01
 squeezenet_ssd_int8  min =    5.40  max =    5.64  avg =    5.57
       mobilenet_ssd  min =    4.34  max =    4.72  avg =    4.48
  mobilenet_ssd_int8  min =    4.57  max =    4.70  avg =    4.64
      mobilenet_yolo  min =   19.91  max =   20.95  avg =   20.30
  mobilenetv2_yolov3  min =    7.23  max =    7.80  avg =    7.61
         yolov4-tiny  min =   11.43  max =   12.44  avg =   12.16
           nanodet_m  min =    4.48  max =    4.57  avg =    4.52
    yolo-fastest-1.1  min =    2.53  max =    2.55  avg =    2.54
      yolo-fastestv2  min =    2.52  max =    2.54  avg =    2.53
  vision_transformer  min =  205.26  max =  212.66  avg =  208.73
          FastestDet  min =    2.66  max =    2.80  avg =    2.73

Without this PR it's 24 by default.

loop_count = 4
num_threads = 24
powersave = 0
gpu_device = -1
cooling_down = 1
          squeezenet  min =    2.41  max =    4.56  avg =    3.22
     squeezenet_int8  min =    3.40  max =    3.51  avg =    3.43
           mobilenet  min =    2.87  max =    4.48  avg =    3.45
      mobilenet_int8  min =    3.70  max =    4.16  avg =    3.86
        mobilenet_v2  min =    3.25  max =    5.36  avg =    3.80
        mobilenet_v3  min =    2.40  max =    8.69  avg =    4.12
          shufflenet  min =    3.34  max =    3.35  avg =    3.34
       shufflenet_v2  min =    3.24  max =    3.30  avg =    3.26
             mnasnet  min =    2.89  max =    4.28  avg =    3.32
     proxylessnasnet  min =    3.21  max =   16.49  avg =    8.53
     efficientnet_b0  min =    4.43  max =    4.57  avg =    4.47
   efficientnetv2_b0  min =    6.16  max =   42.60  avg =   21.47
        regnety_400m  min =    9.73  max =   11.94  avg =   10.37
           blazeface  min =    1.17  max =    1.18  avg =    1.17
           googlenet  min =    7.08  max =    7.19  avg =    7.14
      googlenet_int8  min =   10.02  max =   10.30  avg =   10.17
            resnet18  min =    6.72  max =   16.32  avg =   10.69
       resnet18_int8  min =    8.16  max =    8.24  avg =    8.20
             alexnet  min =    4.71  max =    7.42  avg =    5.41
               vgg16  min =   23.82  max =   33.07  avg =   26.71
          vgg16_int8  min =   21.11  max =   30.79  avg =   24.42
            resnet50  min =   11.59  max =   15.28  avg =   12.69
       resnet50_int8  min =   15.11  max =   25.67  avg =   18.08
      squeezenet_ssd  min =    9.06  max =   18.40  avg =   12.14
 squeezenet_ssd_int8  min =    7.10  max =    7.33  avg =    7.19
       mobilenet_ssd  min =    5.14  max =    5.23  avg =    5.17
  mobilenet_ssd_int8  min =    6.92  max =    9.76  avg =    7.75
      mobilenet_yolo  min =   21.21  max =   37.19  avg =   25.40
  mobilenetv2_yolov3  min =    7.64  max =    7.72  avg =    7.68
         yolov4-tiny  min =   14.61  max =   22.32  avg =   17.40
           nanodet_m  min =    7.02  max =    7.12  avg =    7.07
    yolo-fastest-1.1  min =    4.36  max =    6.55  avg =    4.99
      yolo-fastestv2  min =    4.65  max =    4.73  avg =    4.68
  vision_transformer  min =  157.35  max =  691.71  avg =  406.97
          FastestDet  min =    3.11  max =    3.14  avg =    3.13

[nihui]

#include <stdio.h>
#include <stdlib.h>

#include <vector>

#define WIN32_LEAN_AND_MEAN
#include <windows.h>
#include <powerbase.h>

#pragma comment(lib, "PowrProf.lib")
// Helper function to count set bits in the processor mask.
DWORD CountSetBits(ULONG_PTR bitMask)
{
    DWORD LSHIFT = sizeof(ULONG_PTR) * 8 - 1;
    DWORD bitSetCount = 0;
    ULONG_PTR bitTest = (ULONG_PTR)1 << LSHIFT;
    DWORD i;

    for (i = 0; i <= LSHIFT; ++i)
    {
        bitSetCount += ((bitMask & bitTest) ? 1 : 0);
        bitTest /= 2;
    }

    return bitSetCount;
}
static int get_physical_cpucount()
{
    int count = 0;
    typedef BOOL(WINAPI* LPFN_GLPI)(PSYSTEM_LOGICAL_PROCESSOR_INFORMATION, PDWORD);
    LPFN_GLPI glpi = (LPFN_GLPI)GetProcAddress(GetModuleHandle(TEXT("kernel32")), "GetLogicalProcessorInformation");
    if (glpi == NULL)
    {
        printf("GetLogicalProcessorInformation is not supported\n");
        return -1;
    }

    DWORD return_length = 0;
    glpi(NULL, &return_length);

    PSYSTEM_LOGICAL_PROCESSOR_INFORMATION buffer = (PSYSTEM_LOGICAL_PROCESSOR_INFORMATION)malloc(return_length);
    glpi(buffer, &return_length);

    PSYSTEM_LOGICAL_PROCESSOR_INFORMATION ptr = buffer;
    DWORD byte_offset = 0;
    while (byte_offset + sizeof(SYSTEM_LOGICAL_PROCESSOR_INFORMATION) <= return_length)
    {
        if (ptr->Relationship == RelationProcessorCore)
        {
            count++;

            int cc = CountSetBits(ptr->ProcessorMask);
            printf("smt count = %d   %x\n", cc, ptr->ProcessorMask);
        }

        byte_offset += sizeof(SYSTEM_LOGICAL_PROCESSOR_INFORMATION);
        ptr++;
    }

    free(buffer);

    return count;
}

static int get_cpu_count()
{
    SYSTEM_INFO system_info;
    GetSystemInfo(&system_info);
    return system_info.dwNumberOfProcessors;
}

static std::vector<int> get_cpu_max_mhz()
{
    typedef struct _PROCESSOR_POWER_INFORMATION {
        ULONG Number;
        ULONG MaxMhz;
        ULONG CurrentMhz;
        ULONG MhzLimit;
        ULONG MaxIdleState;
        ULONG CurrentIdleState;
    } PROCESSOR_POWER_INFORMATION, * PPROCESSOR_POWER_INFORMATION;

    int count = get_cpu_count();

    DWORD return_length = sizeof(PROCESSOR_POWER_INFORMATION) * count;
    PPROCESSOR_POWER_INFORMATION buffer = (PPROCESSOR_POWER_INFORMATION)malloc(return_length);

    CallNtPowerInformation(ProcessorInformation, NULL, 0, buffer, return_length);

    std::vector<int> ret;
    for (int i = 0; i < count; i++)
    {
        ULONG max_mhz = buffer[i].MaxMhz;
        printf("%d max_mhz = %d\n", i, max_mhz);

        ret.push_back(max_mhz);
    }

    free(buffer);
    return ret;
}

int main()
{
    printf("get_cpu_count = %d\n", get_cpu_count());
    printf("------------------------------------\n");
    printf("get_physical_cpucount = %d\n", get_physical_cpucount());
    printf("------------------------------------\n");
    get_cpu_max_mhz();
    return 0;
}

[MouriNaruto]

@nihui

Here are results.

i7-11800H

get_cpu_count = 16
------------------------------------
smt count = 2   3
smt count = 2   c
smt count = 2   30
smt count = 2   c0
smt count = 2   300
smt count = 2   c00
smt count = 2   3000
smt count = 2   c000
get_physical_cpucount = 8
------------------------------------
0 max_mhz = 2304
1 max_mhz = 2304
2 max_mhz = 2304
3 max_mhz = 2304
4 max_mhz = 2304
5 max_mhz = 2304
6 max_mhz = 2304
7 max_mhz = 2304
8 max_mhz = 2304
9 max_mhz = 2304
10 max_mhz = 2304
11 max_mhz = 2304
12 max_mhz = 2304
13 max_mhz = 2304
14 max_mhz = 2304
15 max_mhz = 2304

Snapdragon 8cx Gen 3

get_cpu_count = 8
------------------------------------
smt count = 1   1
smt count = 1   2
smt count = 1   4
smt count = 1   8
smt count = 1   10
smt count = 1   20
smt count = 1   40
smt count = 1   80
get_physical_cpucount = 8
------------------------------------
0 max_mhz = 2438
1 max_mhz = 2438
2 max_mhz = 2438
3 max_mhz = 2438
4 max_mhz = 2995
5 max_mhz = 2995
6 max_mhz = 2995
7 max_mhz = 2995

Kenji Mouri