# Set export http_proxy="http://111.11.11.11:1111" export https_proxy="http://111.11.11.11:1111" # Unset unset http_proxy unset https_proxy
# Set import os os.environ['http_proxy']="http://111.11.11.11:1111" os.environ['https_proxy']="http://111.11.11.11:1111" # Unset import os del os.environ['http_proxy'] del os.environ['https_proxy']
参考python包: import distutils.util
import os
import re
import sys
def get_platform():
"""Return a string that identifies the current platform. This is used mainly to
distinguish platform-specific build directories and platform-specific built
distributions. Typically includes the OS name and version and the
architecture (as supplied by 'os.uname()'), although the exact information
included depends on the OS; eg. on Linux, the kernel version isn't
particularly important.
Examples of returned values:
linux-i586
linux-alpha (?)
solaris-2.6-sun4u
Windows will return one of:
win-amd64 (64bit Windows on AMD64 (aka x86_64, Intel64, EM64T, etc)
win32 (all others - specifically, sys.platform is returned)
For other non-POSIX platforms, currently just returns 'sys.platform'.
"""
if os.name == 'nt':
if 'amd64' in sys.version.lower():
return 'win-amd64'
return sys.platform
# Set for cross builds explicitly
if "_PYTHON_HOST_PLATFORM" in os.environ:
return os.environ["_PYTHON_HOST_PLATFORM"]
if os.name != "posix" or not hasattr(os, 'uname'):
# XXX what about the architecture? NT is Intel or Alpha,
# Mac OS is M68k or PPC, etc.
return sys.platform
# Try to distinguish various flavours of Unix
(osname, host, release, version, machine) = os.uname()
# Convert the OS name to lowercase, remove '/' characters, and translate
# spaces (for "Power Macintosh")
osname = osname.lower().replace('/', '')
machine = machine.replace(' ', '_')
machine = machine.replace('/', '-')
if osname[:5] == "linux":
# At least on Linux/Intel, 'machine' is the processor --
# i386, etc.
# XXX what about Alpha, SPARC, etc?
return "%s-%s" % (osname, machine)
elif osname[:5] == "sunos":
if release[0] >= "5": # SunOS 5 == Solaris 2
osname = "solaris"
release = "%d.%s" % (int(release[0]) - 3, release[2:])
# We can't use "platform.architecture()[0]" because a
# bootstrap problem. We use a dict to get an error
# if some suspicious happens.
bitness = {2147483647: "32bit", 9223372036854775807: "64bit"}
machine += ".%s" % bitness[sys.maxsize]
# fall through to standard osname-release-machine representation
elif osname[:3] == "aix":
return "%s-%s.%s" % (osname, version, release)
elif osname[:6] == "cygwin":
osname = "cygwin"
rel_re = re.compile(r'[\d.]+', re.ASCII)
m = rel_re.match(release)
if m:
release = m.group()
elif osname[:6] == "darwin":
import _osx_support, distutils.sysconfig
osname, release, machine = _osx_support.get_platform_osx(
distutils.sysconfig.get_config_vars(),
osname, release, machine)
return "%s-%s-%s" % (osname, release, machine)
cuda 通过<<< >>>符号来分配索引线程的方式，下面展示15种索引方式：
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <stdio.h>
#include <stdlib.h>
#include <iostream>
using namespace std;
//thread 1D
__global__ void testThread1(int *c, const int *a, const int *b)
{
int i = threadIdx.x;
c[i] = b[i] - a[i];
}
//thread 2D
__global__ void testThread2(int *c, const int *a, const int *b)
{
int i = threadIdx.x + threadIdx.y*blockDim.x;
c[i] = b[i] - a[i];
}
//thread 3D
__global__ void testThread3(int *c, const int *a, const int *b)
{
int i = threadIdx.x + threadIdx.y*blockDim.x + threadIdx.z*blockDim.x*blockDim.y;
c[i] = b[i] - a[i];
}
//block 1D
__global__ void testBlock1(int *c, const int *a, const int *b)
{
int i = blockIdx.x;
c[i] = b[i] - a[i];
}
//block 2D
__global__ void testBlock2(int *c, const int *a, const int *b)
{
int i = blockIdx.x + blockIdx.y*gridDim.x;
c[i] = b[i] - a[i];
}
//block 3D
__global__ void testBlock3(int *c, const int *a, const int *b)
{
int i = blockIdx.x + blockIdx.y*gridDim.x + blockIdx.z*gridDim.x*gridDim.y;
c[i] = b[i] - a[i];
}
//block-thread 1D-1D
__global__ void testBlockThread1(int *c, const int *a, const int *b)
{
int i = threadIdx.x + blockDim.x*blockIdx.x;
c[i] = b[i] - a[i];
}
//block-thread 1D-2D
__global__ void testBlockThread2(int *c, const int *a, const int *b)
{
int threadId_2D = threadIdx.x + threadIdx.y*blockDim.x;
int i = threadId_2D+ (blockDim.x*blockDim.y)*blockIdx.x;
c[i] = b[i] - a[i];
}
//block-thread 1D-3D
__global__ void testBlockThread3(int *c, const int *a, const int *b)
{
int threadId_3D = threadIdx.x + threadIdx.y*blockDim.x + threadIdx.z*blockDim.x*blockDim.y;
int i = threadId_3D + (blockDim.x*blockDim.y*blockDim.z)*blockIdx.x;
c[i] = b[i] - a[i];
}
//block-thread 2D-1D
__global__ void testBlockThread4(int *c, const int *a, const int *b)
{
int blockId_2D = blockIdx.x + blockIdx.y*gridDim.x;
int i = threadIdx.x + blockDim.x*blockId_2D;
c[i] = b[i] - a[i];
}
//block-thread 3D-1D
__global__ void testBlockThread5(int *c, const int *a, const int *b)
{
int blockId_3D = blockIdx.x + blockIdx.y*gridDim.x + blockIdx.z*gridDim.x*gridDim.y;
int i = threadIdx.x + blockDim.x*blockId_3D;
c[i] = b[i] - a[i];
}
//block-thread 2D-2D
__global__ void testBlockThread6(int *c, const int *a, const int *b)
{
int threadId_2D = threadIdx.x + threadIdx.y*blockDim.x;
int blockId_2D = blockIdx.x + blockIdx.y*gridDim.x;
int i = threadId_2D + (blockDim.x*blockDim.y)*blockId_2D;
c[i] = b[i] - a[i];
}
//block-thread 2D-3D
__global__ void testBlockThread7(int *c, const int *a, const int *b)
{
int threadId_3D = threadIdx.x + threadIdx.y*blockDim.x + threadIdx.z*blockDim.x*blockDim.y;
int blockId_2D = blockIdx.x + blockIdx.y*gridDim.x;
int i = threadId_3D + (blockDim.x*blockDim.y*blockDim.z)*blockId_2D;
c[i] = b[i] - a[i];
}
//block-thread 3D-2D
__global__ void testBlockThread8(int *c, const int *a, const int *b)
{
int threadId_2D = threadIdx.x + threadIdx.y*blockDim.x;
int blockId_3D = blockIdx.x + blockIdx.y*gridDim.x + blockIdx.z*gridDim.x*gridDim.y;
int i = threadId_2D + (blockDim.x*blockDim.y)*blockId_3D;
c[i] = b[i] - a[i];
}
//block-thread 3D-3D
__global__ void testBlockThread9(int *c, const int *a, const int *b)
{
int threadId_3D = threadIdx.x + threadIdx.y*blockDim.x + threadIdx.z*blockDim.x*blockDim.y;
int blockId_3D = blockIdx.x + blockIdx.y*gridDim.x + blockIdx.z*gridDim.x*gridDim.y;
int i = threadId_3D + (blockDim.x*blockDim.y*blockDim.z)*blockId_3D;
c[i] = b[i] - a[i];
}
void addWithCuda(int *c, const int *a, const int *b, unsigned int size)
{
int *dev_a = 0;
int *dev_b = 0;
int *dev_c = 0;
cudaSetDevice(0);
cudaMalloc((void**)&dev_c, size * sizeof(int));
cudaMalloc((void**)&dev_a, size * sizeof(int));
cudaMalloc((void**)&dev_b, size * sizeof(int));
cudaMemcpy(dev_a, a, size * sizeof(int), cudaMemcpyHostToDevice);
cudaMemcpy(dev_b, b, size * sizeof(int), cudaMemcpyHostToDevice);
//testThread1<<<1, size>>>(dev_c, dev_a, dev_b);
//uint3 s;s.x = size/5;s.y = 5;s.z = 1;
//testThread2 <<<1,s>>>(dev_c, dev_a, dev_b);
//uint3 s; s.x = size / 10; s.y = 5; s.z = 2;
//testThread3<<<1, s >>>(dev_c, dev_a, dev_b);
//testBlock1<<<size,1 >>>(dev_c, dev_a, dev_b);
//uint3 s; s.x = size / 5; s.y = 5; s.z = 1;
//testBlock2<<<s, 1 >>>(dev_c, dev_a, dev_b);
//uint3 s; s.x = size / 10; s.y = 5; s.z = 2;
//testBlock3<<<s, 1 >>>(dev_c, dev_a, dev_b);
//testBlockThread1<<<size/10, 10>>>(dev_c, dev_a, dev_b);
//uint3 s1; s1.x = size / 100; s1.y = 1; s1.z = 1;
//uint3 s2; s2.x = 10; s2.y = 10; s2.z = 1;
//testBlockThread2 << <s1, s2 >> >(dev_c, dev_a, dev_b);
//uint3 s1; s1.x = size / 100; s1.y = 1; s1.z = 1;
//uint3 s2; s2.x = 10; s2.y = 5; s2.z = 2;
//testBlockThread3 << <s1, s2 >> >(dev_c, dev_a, dev_b);
//uint3 s1; s1.x = 10; s1.y = 10; s1.z = 1;
//uint3 s2; s2.x = size / 100; s2.y = 1; s2.z = 1;
//testBlockThread4 << <s1, s2 >> >(dev_c, dev_a, dev_b);
//uint3 s1; s1.x = 10; s1.y = 5; s1.z = 2;
//uint3 s2; s2.x = size / 100; s2.y = 1; s2.z = 1;
//testBlockThread5 << <s1, s2 >> >(dev_c, dev_a, dev_b);
//uint3 s1; s1.x = size / 100; s1.y = 10; s1.z = 1;
//uint3 s2; s2.x = 5; s2.y = 2; s2.z = 1;
//testBlockThread6 << <s1, s2 >> >(dev_c, dev_a, dev_b);
//uint3 s1; s1.x = size / 100; s1.y = 5; s1.z = 1;
//uint3 s2; s2.x = 5; s2.y = 2; s2.z = 2;
//testBlockThread7 << <s1, s2 >> >(dev_c, dev_a, dev_b);
//uint3 s1; s1.x = 5; s1.y = 2; s1.z = 2;
//uint3 s2; s2.x = size / 100; s2.y = 5; s2.z = 1;
//testBlockThread8 <<<s1, s2 >>>(dev_c, dev_a, dev_b);
uint3 s1; s1.x = 5; s1.y = 2; s1.z = 2;
uint3 s2; s2.x = size / 200; s2.y = 5; s2.z = 2;
testBlockThread9<<<s1, s2 >>>(dev_c, dev_a, dev_b);
cudaMemcpy(c, dev_c, size*sizeof(int), cudaMemcpyDeviceToHost);
cudaFree(dev_a);
cudaFree(dev_b);
cudaFree(dev_c);
cudaGetLastError();
}
int main()
{
const int n = 1000;
int *a = new int[n];
int *b = new int[n];
int *c = new int[n];
int *cc = new int[n];
for (int i = 0; i < n; i++)
{
a[i] = rand() % 100;
b[i] = rand() % 100;
c[i] = b[i] - a[i];
}
addWithCuda(cc, a, b, n);
FILE *fp = fopen("out.txt", "w");
for (int i = 0; i < n; i++)
fprintf(fp, "%d %d\n", c[i], cc[i]);
fclose(fp);
bool flag = true;
for (int i = 0; i < n; i++)
{
if (c[i] != cc[i])
{
flag = false;
break;
}
}
if (flag == false)
printf("no pass");
else
printf("pass");
cudaDeviceReset();
delete[] a;
delete[] b;
delete[] c;
delete[] cc;
getchar();
return 0;
}
Full global thread ID in x and y dimensions can be computed by:
