Script that compiles a CUDA application to report the attributes of a kernel

cuda_kernel_query.py

# Compiles a CUDA application to report the attributes of a kernel
# 
# usage: cuda_kernel_query.py <file> <entry name> [<define list>]
# 
# file        - name of the CUDA source file contain the kernel of interest
# entry name  - name of the kernel to examine
# define list - comma delimited list of proprocessor defines for compilation 
#
# If BLOCK_SIZE is defined in the file, this value is used in the occupancy
# calculation, otherwise the blockSize returned by
# cudaOccupancyMaxPotentialBlockSize() is instead.

import sys
import os
from pathlib import Path

cuda_template = """
#include <stdio.h>
#include <cuda_runtime.h>

#define CUDA_CHECK(err) if( (err) != cudaSuccess ) { fprintf(stderr, "ERROR: %s in %s at line %d\n", cudaGetErrorString(err), __FILE__, __LINE__); exit(EXIT_FAILURE); }

#include "FILE_NAME"

int main()
{ 
    int minGridSize;
    int maxPotentialBlockSize;
    int blockSize;
    int numBlocks;
    cudaDeviceProp deviceProp;
    cudaFuncAttributes attrib;

    CUDA_CHECK(cudaFuncGetAttributes(&attrib, KERNEL_NAME))
    CUDA_CHECK(cudaGetDeviceProperties(&deviceProp, 0))

    CUDA_CHECK(cudaOccupancyMaxPotentialBlockSize(&minGridSize,&maxPotentialBlockSize,KERNEL_NAME,0,deviceProp.maxThreadsPerBlock))
  
#ifdef BLOCK_SIZE
    blockSize = BLOCK_SIZE;
#else
    blockSize = maxPotentialBlockSize;
#endif

    CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&numBlocks,KERNEL_NAME,blockSize,0))

    int maxWarps = deviceProp.maxThreadsPerMultiProcessor / deviceProp.warpSize;

    printf("Device Info\n");
    printf("-----------\n");
    printf("deviceProp.maxThreadsPerBlock:          %d\n", deviceProp.maxThreadsPerBlock);
    printf("deviceProp.maxThreadsPerMultiProcessor: %d\n", deviceProp.maxThreadsPerMultiProcessor );
    printf("deviceProp.warpSize:                    %d\n", deviceProp.warpSize);
    printf("maxWarps:                               %d\n\n", maxWarps);

    printf("Kernal KERNEL_NAME in file FILE_NAME\n\n");

    printf("minGridSize:                      %d\n", minGridSize);
    printf("MaxPotentialBlockSize:            %d\n", maxPotentialBlockSize);
    printf("MaxActiveBlocksPerMultiprocessor: %d\n", numBlocks);
#ifdef BLOCK_SIZE    
    printf("BLOCK_SIZE:                       %d\n", blockSize);
#endif
    printf("Occupancy:                        %f%%\n\n", (double)(numBlocks * blockSize / deviceProp.warpSize) / maxWarps * 100.0);

    printf("cudaFuncAttributes\n");
    printf("------------------\n");
    printf(" binaryVersion:             %.1f\n", attrib.binaryVersion / 10.0f);
    printf(" cacheModeCA:               %d\n", attrib.cacheModeCA);
    printf(" constSizeBytes:            %zu\n", attrib.constSizeBytes);
    printf(" localSizeBytes:            %zu\n", attrib.localSizeBytes);
    printf(" maxDynamicSharedSizeBytes: %d\n", attrib.maxDynamicSharedSizeBytes);
    printf(" maxThreadsPerBlock:        %d\n", attrib.maxThreadsPerBlock);
    printf(" numRegs:                   %d\n", attrib.numRegs);
    printf(" preferredShmemCarveout:    %d\n", attrib.preferredShmemCarveout);
    printf(" ptxVersion:                %d\n", attrib.ptxVersion);
    printf(" sharedSizeBytes:           %zu\n", attrib.sharedSizeBytes);

    return 0;
}
"""

if len(sys.argv) != 3 and len(sys.argv) != 4:
    print("    usage: %s <file> <entry name> [<define list>]" % sys.argv[0])
    sys.exit(0)

p = Path(sys.argv[1])

define_list = ''

if not p.exists():
    print("Error: %s does not exist" % sys.argv[1])
    sys.exit(-1)

if not p.is_file():
    print("Error: %s is not a file" % sys.argv[1])
    sys.exit(-1)

if len(sys.argv) == 4:
    define_list = ' '.join(['-D%s' % d for d in sys.argv[3].split(',')])

p = Path("./kquery")

# remove old executable if it exists
if p.exists():
    p.unlink()

cuda_template = cuda_template.replace('FILE_NAME', sys.argv[1])
cuda_template = cuda_template.replace('KERNEL_NAME', sys.argv[2])

with open('kernel_query.cu','w') as f:
    f.write(cuda_template)

cmd_str = "nvcc -ccbin g++ %s -m64 -gencode arch=compute_35,code=sm_35 -o kquery kernel_query.cu" % define_list
print(cmd_str)
os.system(cmd_str)

p = Path("./kquery")

if p.exists():
    os.system("./kquery")