OpenACC Julia 图形
▶ 书上的代码，逐步优化绘制 Julia 图形的代码
● 无并行优化（手动优化了变量等）
#include <stdio.h>
#include <stdlib.h>
#include <openacc.h> #define N (1024 * 8) int julia(const float cre, const float cim, float zre, float zim, const int maxIter)// 计算单点迭代次数
{
float zre2 = 0.0f, zim2 = 0.0f;
for (int iter = ; iter < maxIter; iter += ) // 一个迭代里计算两次
{
zre2 = zre * zre - zim * zim + cre, zim2 = * zre * zim + cim;
if (zre2 * zre2 + zim2 * zim2 > 4.0f)
return iter; zre = zre2 * zre2 - zim2 * zim2 + cre, zim = * zre2 * zim2 + cim;
if (zre * zre + zim * zim > 4.0f)
return iter;
}
return maxIter + + (maxIter % );
} int main()
{
const int maxIter = ; // 最大迭代次数
const float cre = -0.8350, cim = -0.2321, h = 4.0f / N; // 迭代常数和画幅步长
int *image = (int *)malloc(sizeof(int) * N * N);
FILE *pf = fopen("R:/output.txt", "w"); for (int i = ; i < N; i++)
{
for (int j = ; j < N; j++)
fprintf(pf, "%d ", julia(cre, cim, i * h - 2.0f, j * h - 2.0f, maxIter));
fprintf(pf, "\n");
} fclose(pf);
free(image);
//getchar();
return ;
}
● 输出结果（后面所有代码的输出都相同，不再写了）
● 改进 1，计算并行化
#include <stdio.h>
#include <stdlib.h>
#include <openacc.h> #define N (1024 * 8) #pragma acc routine seq
int julia(const float cre, const float cim, float zre, float zim, const int maxIter)
{
float zre2 = 0.0f, zim2 = 0.0f;
for (int iter = ; iter < maxIter; iter += )
{
zre2 = zre * zre - zim * zim + cre, zim2 = * zre * zim + cim;
if (zre2 * zre2 + zim2 * zim2 > 4.0f)
return iter; zre = zre2 * zre2 - zim2 * zim2 + cre, zim = * zre2 * zim2 + cim;
if (zre * zre + zim * zim > 4.0f)
return iter;
}
return maxIter + + (maxIter % );
} int main()
{
const int maxIter = ;
const float cre = -0.8350, cim = -0.2321, h = 4.0f / N;
int *image = (int *)malloc(sizeof(int) * N * N); #pragma acc data copyout(image[0:N * N]) // 数据域
{
#pragma acc kernels loop independent // loop 并行化，强制独立
for (int i = ; i < N; i++)
{
for (int j = ; j < N; j++)
image[i * N + j] = julia(cre, cim, i * h - 2.0f, j * h - 2.0f, maxIter);
}
}
/*// 注释掉写入文件的部分，防止 Nvvp 加入分析
FILE *pf = fopen("R:/output.txt", "w");
for (int i = 0; i < N; i++)
{
for (int j = 0; j < N; j++)
fprintf(pf, "%d ", image[i * N + j]);
fprintf(pf, "\n");
}
fclose(pf);
*/
free(image);
//getchar();
return ;
}
● 输出结果
D:\Code\OpenACC\OpenACCProject\OpenACCProject>pgcc -acc -Minfo main.c -o main_acc.exe
julia:
, Generating acc routine seq
Generating Tesla code
, FMA (fused multiply-add) instruction(s) generated
, FMA (fused multiply-add) instruction(s) generated
main:
, Generating copyout(image[:])
, Loop is parallelizable
FMA (fused multiply-add) instruction(s) generated
, Loop is parallelizable
Accelerator kernel generated
Generating Tesla code
, #pragma acc loop gang, vector(4) /* blockIdx.y threadIdx.y */
, #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
, FMA (fused multiply-add) instruction(s) generated D:\Code\OpenACC\OpenACCProject\OpenACCProject>main_acc.exe
launch CUDA kernel file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
line= device= threadid= num_gangs= num_workers= vector_length= grid=256x128 block=32x4
PGI: "acc_shutdown" not detected, performance results might be incomplete.
Please add the call "acc_shutdown(acc_device_nvidia)" to the end of your application to ensure that the performance results are complete. Accelerator Kernel Timing data
D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c
main NVIDIA devicenum=
time(us): ,
: data region reached times
: data copyout transfers:
device time(us): total=, max=, min= avg=,
: compute region reached time
: kernel launched time
grid: [256x128] block: [32x4]
device time(us): total= max= min= avg=
● 改进 2，分块计算，没有明显性能提升，为异步做准备
#include <stdio.h>
#include <stdlib.h>
#include <openacc.h> #define N (1024 * 8) #pragma acc routine seq
int julia(const float cre, const float cim, float zre, float zim, const int maxIter)
{
float zre2 = 0.0f, zim2 = 0.0f;
for (int iter = ; iter < maxIter; iter += )
{
zre2 = zre * zre - zim * zim + cre, zim2 = * zre * zim + cim;
if (zre2 * zre2 + zim2 * zim2 > 4.0f)
return iter; zre = zre2 * zre2 - zim2 * zim2 + cre, zim = * zre2 * zim2 + cim;
if (zre * zre + zim * zim > 4.0f)
return iter;
}
return maxIter + + (maxIter % );
} int main()
{
const int maxIter = ;
const float cre = -0.8350, cim = -0.2321, h = 4.0f / N;
int *image = (int *)malloc(sizeof(int) * N * N); #pragma acc data copyout(image[0:N * N])
{
const int numblock = ; // 指定分块数量
for (int block = ; block < numblock; block++) // 每次计算一块
{
const int start = block * (N / numblock), end = start + N / numblock; // 每块的始末下标
#pragma acc kernels loop independent
for (int i = start; i < end; i++)
{
for (int j = ; j < N; j++)
image[i * N + j] = julia(cre, cim, i * h - 2.0f, j * h - 2.0f, maxIter);
}
}
}
/*
FILE *pf = fopen("R:/output.txt", "w");
for (int i = 0; i < N; i++)
{
for (int j = 0; j < N; j++)
fprintf(pf, "%d ", image[i * N + j]);
fprintf(pf, "\n");
}
fclose(pf);
*/
free(image);
//getchar();
return ;
}
● 输出结果
D:\Code\OpenACC\OpenACCProject\OpenACCProject>pgcc -acc -Minfo main.c -o main_acc.exe
julia:
, Generating acc routine seq
Generating Tesla code
, FMA (fused multiply-add) instruction(s) generated
, FMA (fused multiply-add) instruction(s) generated
main:
, Generating copyout(image[:])
, Loop is parallelizable
FMA (fused multiply-add) instruction(s) generated
, Loop is parallelizable
Accelerator kernel generated
Generating Tesla code
, #pragma acc loop gang, vector(4) /* blockIdx.y threadIdx.y */
, #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
, FMA (fused multiply-add) instruction(s) generated D:\Code\OpenACC\OpenACCProject\OpenACCProject>main_acc.exe
launch CUDA kernel file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
line= device= threadid= num_gangs= num_workers= vector_length= grid=256x128 block=32x4
launch CUDA kernel file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
= device= threadid= num_gangs= num_workers= vector_length= grid=256x128 block=32x4
launch CUDA kernel file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
line= device= threadid= num_gangs= num_workers= vector_length= grid=256x128 block=32x4
launch CUDA kernel file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
line= device= threadid= num_gangs= num_workers= vector_length= grid=256x128 block=32x4
PGI: "acc_shutdown" not detected, performance results might be incomplete.
Please add the call "acc_shutdown(acc_device_nvidia)" to the end of your application to ensure that the performance results are complete. Accelerator Kernel Timing data
D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c
main NVIDIA devicenum=
time(us): ,
: data region reached times
: data copyout transfers:
device time(us): total=, max=, min= avg=,
: compute region reached times
: kernel launched times
grid: [256x128] block: [32x4]
device time(us): total= max= min= avg=
● 改进 3，分块传输，没有明显性能提升，为异步做准备
#include <stdio.h>
#include <stdlib.h>
#include <openacc.h> #define N (1024 * 8) #pragma acc routine seq
int julia(const float cre, const float cim, float zre, float zim, const int maxIter)
{
float zre2 = 0.0f, zim2 = 0.0f;
for (int iter = ; iter < maxIter; iter += )
{
zre2 = zre * zre - zim * zim + cre, zim2 = * zre * zim + cim;
if (zre2 * zre2 + zim2 * zim2 > 4.0f)
return iter; zre = zre2 * zre2 - zim2 * zim2 + cre, zim = * zre2 * zim2 + cim;
if (zre * zre + zim * zim > 4.0f)
return iter;
}
return maxIter + + (maxIter % );
} int main()
{
const int maxIter = ;
const float cre = -0.8350, cim = -0.2321, h = 4.0f / N;
int *image = (int *)malloc(sizeof(int) * N * N); #pragma acc data create(image[0:N * N]) // 改成 create，不需要从主机拷贝初始数据
{
const int numBlock = , blockSize = N * N / numBlock; // 仍然分块计算
for (int block = ; block < numBlock; block++)
{
const int start = block * (N / numBlock), end = start + N / numBlock;
#pragma acc kernels loop independent
for (int i = start; i < end; i++)
{
for (int j = ; j < N; j++)
image[i * N + j] = julia(cre, cim, i * h - 2.0f, j * h - 2.0f, maxIter);
}
#pragma acc update host(image[block * blockSize : blockSize]) // 每计算完一块就向主机回传数据
}
}
/*
FILE *pf = fopen("R:/output.txt", "w");
for (int i = 0; i < N; i++)
{
for (int j = 0; j < N; j++)
fprintf(pf, "%d ", image[i * N + j]);
fprintf(pf, "\n");
}
fclose(pf);
*/
free(image);
//getchar();
return ;
}
● 输出结果
D:\Code\OpenACC\OpenACCProject\OpenACCProject>pgcc -acc -Minfo main.c -o main_acc.exe
julia:
, Generating acc routine seq
Generating Tesla code
, FMA (fused multiply-add) instruction(s) generated
, FMA (fused multiply-add) instruction(s) generated
main:
, Generating create(image[:])
, Loop is parallelizable
FMA (fused multiply-add) instruction(s) generated
, Loop is parallelizable
Accelerator kernel generated
Generating Tesla code
, #pragma acc loop gang, vector(4) /* blockIdx.y threadIdx.y */
, #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
, FMA (fused multiply-add) instruction(s) generated
, Generating update self(image[block*blockSize:blockSize]) D:\Code\OpenACC\OpenACCProject\OpenACCProject>main_acc.exe
launch CUDA kernel file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
line= device= threadid= num_gangs= num_workers= vector_length= grid=256x128 block=32x4
launch CUDA kernel file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
line= device= threadid= num_gangs= num_workers= vector_length= grid=256x128 block=32x4
launch CUDA kernel file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
line= device= threadid= num_gangs= num_workers= vector_length= grid=256x128 block=32x4
launch CUDA kernel file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
line= device= threadid= num_gangs= num_workers= vector_length= grid=256x128 block=32x4
PGI: "acc_shutdown" not detected, performance results might be incomplete.
Please add the call "acc_shutdown(acc_device_nvidia)" to the end of your application to ensure that the performance results are complete. Accelerator Kernel Timing data
D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c
main NVIDIA devicenum=
time(us): ,
: data region reached times
: compute region reached times
: kernel launched times
grid: [256x128] block: [32x4]
elapsed time(us): total=, max=, min= avg=,
: update directive reached times
: data copyout transfers:
device time(us): total=, max=, min= avg=,
● 改进 4，异步计算 - 双向传输
#include <stdio.h>
#include <stdlib.h>
#include <openacc.h> #define N (1024 * 8) #pragma acc routine seq
int julia(const float cre, const float cim, float zre, float zim, const int maxIter)
{
float zre2 = 0.0f, zim2 = 0.0f;
for (int iter = ; iter < maxIter; iter += )
{
zre2 = zre * zre - zim * zim + cre, zim2 = * zre * zim + cim;
if (zre2 * zre2 + zim2 * zim2 > 4.0f)
return iter; zre = zre2 * zre2 - zim2 * zim2 + cre, zim = * zre2 * zim2 + cim;
if (zre * zre + zim * zim > 4.0f)
return iter;
}
return maxIter + + (maxIter % );
} int main()
{
const int maxIter = ;
const float cre = -0.8350, cim = -0.2321, h = 4.0f / N;
int *image = (int *)malloc(sizeof(int) * N * N); #pragma acc data create(image[0:N * N])
{
const int numBlock = , blockSize = N / numBlock * N;
for (int block = ; block < numBlock; block++)
{
const int start = block * (N / numBlock), end = start + N / numBlock;
#pragma acc kernels loop independent async(block + 1) // 异步计算，用块编号作标记
for (int i = start; i < end; i++)
{
for (int j = ; j < N; j++)
image[i * N + j] = julia(cre, cim, i * h - 2.0f, j * h - 2.0f, maxIter);
}
#pragma acc update host(image[block * blockSize : blockSize]) async(block + 1) // 计算完一块就异步传输
}
#pragma acc wait
}
/*
FILE *pf = fopen("R:/output.txt", "w");
for (int i = 0; i < N; i++)
{
for (int j = 0; j < N; j++)
fprintf(pf, "%d ", image[i * N + j]);
fprintf(pf, "\n");
}
fclose(pf);
*/
free(image);
//getchar();
return ;
}
● 输出结果
D:\Code\OpenACC\OpenACCProject\OpenACCProject>pgcc -acc -Minfo main.c -o main_acc.exe
julia:
, Generating acc routine seq
Generating Tesla code
, FMA (fused multiply-add) instruction(s) generated
, FMA (fused multiply-add) instruction(s) generated
main:
, Generating create(image[:])
, Loop is parallelizable
FMA (fused multiply-add) instruction(s) generated
, Loop is parallelizable
Accelerator kernel generated
Generating Tesla code
, #pragma acc loop gang, vector(4) /* blockIdx.y threadIdx.y */
, #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
, FMA (fused multiply-add) instruction(s) generated
, Generating update self(image[block*blockSize:blockSize]) D:\Code\OpenACC\OpenACCProject\OpenACCProject>main_acc.exe
launch CUDA kernel file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
line= device= threadid= queue= num_gangs= num_workers= vector_length= grid=256x128 block=32x4
launch CUDA kernel file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
line= device= threadid= queue= num_gangs= num_workers= vector_length= grid=256x128 block=32x4
launch CUDA kernel file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
line= device= threadid= queue= num_gangs= num_workers= vector_length= grid=256x128 block=32x4
launch CUDA kernel file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
line= device= threadid= queue= num_gangs= num_workers= vector_length= grid=256x128 block=32x4
PGI: "acc_shutdown" not detected, performance results might be incomplete.
Please add the call "acc_shutdown(acc_device_nvidia)" to the end of your application to ensure that the performance results are complete. Accelerator Kernel Timing data
Timing may be affected by asynchronous behavior
set PGI_ACC_SYNCHRONOUS to to disable async() clauses
D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c
main NVIDIA devicenum=
time(us): ,
: data region reached times
: compute region reached times
: kernel launched times
grid: [256x128] block: [32x4]
device time(us): total= max= min= avg=
: update directive reached times
: data copyout transfers:
device time(us): total=, max=, min= avg=,
● 使用统一内存访址（Ubuntu，win64 不支持）
● 改进 4，多设备版本 1，使用 OpenMP
#include <stdio.h>
#include <stdlib.h>
#include <omp.h>
#include <openacc.h> #define N (1024 * 8) #pragma acc routine seq
int julia(const float cre, const float cim, float zre, float zim, const int maxIter)
{
float zre2 = 0.0f, zim2 = 0.0f;
for (int iter = ; iter < maxIter; iter += )
{
zre2 = zre * zre - zim * zim + cre, zim2 = * zre * zim + cim;
if (zre2 * zre2 + zim2 * zim2 > 4.0f)
return iter; zre = zre2 * zre2 - zim2 * zim2 + cre, zim = * zre2 * zim2 + cim;
if (zre * zre + zim * zim > 4.0f)
return iter;
}
return maxIter + + (maxIter % );
} int main()
{
const int maxIter = ;
const int numBlock = acc_get_num_devices(acc_device_nvidia), blockSize = N / numBlock * N; // 使用 OpenMP 检测目标设备数量，以此作为分块数
const float cre = -0.8350, cim = -0.2321, h = 4.0f / N; int *image = (int *)malloc(sizeof(int) * N * N);
acc_init(acc_device_nvidia); // 一次性初始化全部目标设备 #pragma omp parallel num_threads(numBlock) // 使用多个线程，分别向目标设备发送任务
{
acc_set_device_num(omp_get_thread_num(), acc_device_nvidia);// 标记目标设备
#pragma omp for
for (int block = ; block < numBlock; block++)
{
const int start = block * (N / numBlock), end = start + N / numBlock;
#pragma acc data copyout(image[block * blockSize : blockSize])
{
#pragma acc kernels loop independent
for (int i = start; i < end; i++)
{
for (int j = ; j < N; j++)
image[i * N + j] = julia(cre, cim, i * h - 2.0f, j * h - 2.0f, maxIter);
}
}
}
}
/*
FILE *pf = fopen("R:/output.txt", "w");
for (int i = 0; i < N; i++)
{
for (int j = 0; j < N; j++)
fprintf(pf, "%d ", image[i * N + j]);
fprintf(pf, "\n");
}
fclose(pf);
*/
free(image);
//getchar();
return ;
}
● 输出结果
D:\Code\OpenACC\OpenACCProject\OpenACCProject>pgcc -acc -mp -Minfo main.c -o main_acc.exe
julia:
, Generating acc routine seq
Generating Tesla code
, FMA (fused multiply-add) instruction(s) generated
, FMA (fused multiply-add) instruction(s) generated
main:
, Parallel region activated
, Parallel loop activated with static block schedule
, Generating copyout(image[block*blockSize:blockSize])
, Loop is parallelizable
FMA (fused multiply-add) instruction(s) generated
, Loop is parallelizable
Accelerator kernel generated
Generating Tesla code
, #pragma acc loop gang, vector(4) /* blockIdx.y threadIdx.y */
, #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
, FMA (fused multiply-add) instruction(s) generated
, Barrier
, Parallel region terminated D:\Code\OpenACC\OpenACCProject\OpenACCProject>main_acc.exe
launch CUDA kernel file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
line= device= threadid= num_gangs= num_workers= vector_length= grid=256x128 block=32x4
PGI: "acc_shutdown" not detected, performance results might be incomplete.
Please add the call "acc_shutdown(acc_device_nvidia)" to the end of your application to ensure that the performance results are complete. Accelerator Kernel Timing data
D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c
main NVIDIA devicenum=
time(us): ,
: data region reached times
: data copyout transfers:
device time(us): total=, max=, min= avg=,
: compute region reached time
: kernel launched time
grid: [256x128] block: [32x4]
device time(us): total= max= min= avg=
● 改进 5，多设备版本 2，调整 OpenMP
#include <stdio.h>
#include <stdlib.h>
#include <omp.h>
#include <openacc.h> #define N (1024 * 8) #pragma acc routine seq
int julia(const float cre, const float cim, float zre, float zim, const int maxIter)
{
float zre2 = 0.0f, zim2 = 0.0f;
for (int iter = ; iter < maxIter; iter += )
{
zre2 = zre * zre - zim * zim + cre, zim2 = * zre * zim + cim;
if (zre2 * zre2 + zim2 * zim2 > 4.0f)
return iter; zre = zre2 * zre2 - zim2 * zim2 + cre, zim = * zre2 * zim2 + cim;
if (zre * zre + zim * zim > 4.0f)
return iter;
}
return maxIter + + (maxIter % );
} int main()
{
const int maxIter = ;
const int numBlock = acc_get_num_devices(acc_device_nvidia), blockSize = N / numBlock * N;
const float cre = -0.8350, cim = -0.2321, h = 4.0f / N; int *image = (int *)malloc(sizeof(int) * N * N);
acc_init(acc_device_nvidia); #pragma omp parallel for num_threads(numBlock) // 把函数 acc_set_device_num 单独放在一起
for(int block = ;block<numBlock;block++)
acc_set_device_num(block, acc_device_nvidia); #pragma omp for num_threads(numBlock)
for (int block = ; block < numBlock; block++)
{
const int start = block * (N / numBlock), end = start + N / numBlock;
#pragma acc data copyout(image[block * blockSize : blockSize])
{
#pragma acc kernels loop independent
for (int i = start; i < end; i++)
{
for (int j = ; j < N; j++)
image[i * N + j] = julia(cre, cim, i * h - 2.0f, j * h - 2.0f, maxIter);
}
}
}
/*
FILE *pf = fopen("R:/output.txt", "w");
for (int i = 0; i < N; i++)
{
for (int j = 0; j < N; j++)
fprintf(pf, "%d ", image[i * N + j]);
fprintf(pf, "\n");
}
fclose(pf);
*/
free(image);
//getchar();
return ;
}
● 输出结果
D:\Code\OpenACC\OpenACCProject\OpenACCProject>pgcc main.c -acc -Minfo -o main_acc.exe
julia:
, Generating acc routine seq
Generating Tesla code
, FMA (fused multiply-add) instruction(s) generated
, FMA (fused multiply-add) instruction(s) generated
main:
, Generating copyout(image[block*blockSize:blockSize])
, Loop is parallelizable
FMA (fused multiply-add) instruction(s) generated
, Loop is parallelizable
Accelerator kernel generated
Generating Tesla code
, #pragma acc loop gang, vector(4) /* blockIdx.y threadIdx.y */
, #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
, FMA (fused multiply-add) instruction(s) generated D:\Code\OpenACC\OpenACCProject\OpenACCProject>main_acc.exe
launch CUDA kernel file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
line= device= threadid= num_gangs= num_workers= vector_length= grid=256x128 block=32x4
PGI: "acc_shutdown" not detected, performance results might be incomplete.
Please add the call "acc_shutdown(acc_device_nvidia)" to the end of your application to ensure that the performance results are complete. Accelerator Kernel Timing data
D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c
main NVIDIA devicenum=
time(us): ,
: data region reached times
: data copyout transfers:
device time(us): total=, max=, min= avg=,
: compute region reached time
: kernel launched time
grid: [256x128] block: [32x4]
elapsed time(us): total=, max=, min=, avg=,
OpenACC Julia 图形的更多相关文章
- OpenCV绘制朱利亚(Julia)集合图形
朱利亚集合是一个在复平面上形成分形的点的集合.以法国数学家加斯顿·朱利亚(Gaston Julia)的名字命名. 朱利亚集合可以由下式进行反复迭代得到: 对于固定的复数c,取某一z值(如z = z0) ...
- CUDA+OpenCV 绘制朱利亚(Julia)集合图形
Julia集中的元素都是经过简单的迭代计算得到的,很适合用CUDA进行加速.对一个600*600的图像,需要进行360000次迭代计算,所以在CUDA中创建了600*600个线程块(block),每个 ...
- OpenACC 云水参数化方案
▶ 书上第十三章,用一系列步骤优化一个云水参数化方案.用于熟悉 Fortran 以及 OpenACC 在旗下的表现 ● 代码,文件较多,放在一起了 ! main.f90 PROGRAM main US ...
- 详解 CUDA By Example 中的 Julia Set 绘制GPU优化
笔者测试环境VS2019. 基本介绍 原书作者引入Julia Sets意在使用GPU加速图形的绘制.Julia Set 是指满足下式迭代收敛的复数集合 \[ Z_{n+1}=Z_{n}^2+C \] ...
- 现代3D图形编程学习-基础简介(3)-什么是opengl (译)
本书系列 现代3D图形编程学习 OpenGL是什么 在我们编写openGL程序之前,我们首先需要知道什么是OpenGL. 将OpenGL作为一个API OpenGL 通常被认为是应用程序接口(API) ...
- 超全面的.NET GDI+图形图像编程教程
本篇主题内容是.NET GDI+图形图像编程系列的教程,不要被这个滚动条吓到,为了查找方便,我没有分开写,上面加了目录了,而且很多都是源码和图片~ (*^_^*) 本人也为了学习深刻,另一方面也是为了 ...
- Ubuntu设置root用户登录图形界面
Ubuntu默认的是root用户不能登录图形界面的,只能以其他用户登录图形界面.这样就很麻烦,因为权限的问题,不能随意复制删除文件,用gedit编辑文件时经常不能保存,只能用vim去编辑. 解决的办法 ...
- 第六代智能英特尔® 酷睿™ 处理器图形 API 开发人员指南
欢迎查看第六代智能英特尔® 酷睿™ 处理器图形 API 开发人员指南,该处理器可为开发人员和最终用户提供领先的 CPU 和图形性能增强.各种新特性和功能以及显著提高的性能. 本指南旨在帮助软件开发人员 ...
- 通过Matrix进行二维图形仿射变换
Affine Transformation是一种二维坐标到二维坐标之间的线性变换,保持二维图形的"平直性"和"平行性".仿射变换可以通过一系列的原子变换的复合来 ...
随机推荐
- HDU 3746：Cyclic Nacklace（KMP循环节）
Cyclic Nacklace Time Limit: 2000/1000 MS (Java/Others) Memory Limit: 32768/32768 K (Java/Others) ...
- php empty详解
判断字符串是否为空,可以这么判断: if ($value=="") ... * 格式:bool empty ( mixed var ) * 功能:检查一个变量是否为空 ...
- RabbitMQ负载均衡方案之LVS
负载均衡的方案有很多,适合RabbitMQ使用的处理HAProxy之外还有LVS.LVS是Linux Virtual Server的简称,也就是Linux虚拟服务器, 是一个由章文嵩博士发起的自由软件 ...
- FP-growth算法发现频繁项集（二）——发现频繁项集
上篇介绍了如何构建FP树,FP树的每条路径都满足最小支持度,我们需要做的是在一条路径上寻找到更多的关联关系. 抽取条件模式基 首先从FP树头指针表中的单个频繁元素项开始.对于每一个元素项,获得其对应的 ...
- vue-router 知识点
vue-router配置scrollBehavior 第三个参数 savedPosition 当且仅当 popstate 导航 (通过浏览器的 前进/后退 按钮触发) 时才可用. 注意: 这个功能只在 ...
- Cropperjs 插件知识点
cropper.js 通过 canvas 实现图片裁剪,最后再通过 canvas 获取裁剪区域的图片base64串. 1. Container 容器 2. canvas 图片 3. crop 裁剪框
- button使用注意
<button type="button" class="btn btn-primary" onclick="ChangePassword(); ...
- linux SSH各配置项解释
Linux下SSH各配置项解释 关于ssh 设置的相关总结(ssh最大连接数.ssh连接时长.安全性配置等)以redhat6.3为例 ssh配置文件在: /etc/ssh/sshd_config可以打 ...
- Net分布式系统整体框架
Net分布式系统之一:系统整体框架介绍 一.设计目的 从事.Net平台开发系统已有8年多了,一直思考搭建.Net分布式系统架构.基于window平台搭建的大型分布式系统不多,之前了解过myspace. ...
- 转 WCF中同步和异步通讯总结
我这样分个类: WCF中, 以同步.异步角度考虑通讯的方式分为四种:跨进程同步.跨进程异步.发送队列端同步.发送队列端异步.之所以造成这样的结果源于两个因素,一个是传统概念上的同异步,一个是对于WCF ...