0_Simple__simpleLayeredTexture

二维分层纹理

▶ 源代码。用纹理方法把元素按原顺序从 CUDA3D 数组中取出来，求个相反数再加上层数放入全局内存，输出。

 #include <stdio.h>

 #include "cuda_runtime.h"

 #include "device_launch_parameters.h"

 #include <helper_functions.h>

 #include <helper_cuda.h>

 #define MIN_EPSILON_ERROR 5e-3f

 #define OUTPUT 5

 texture<float, cudaTextureType2DLayered> tex;

 __global__ void transformKernel(float *g_odata, int width, int height, int layer)

 {

     unsigned int x = blockIdx.x*blockDim.x + threadIdx.x;

     unsigned int y = blockIdx.y*blockDim.y + threadIdx.y;

     float u = (x + 0.5f) / (float)width;

     float v = (y + 0.5f) / (float)height;

     g_odata[layer*width*height + y*width + x] = - tex2DLayered(tex, u, v, layer) + layer;

 }

 int main(int argc, char **argv)

 {

     unsigned int width = , height = , num_layers = ;

     unsigned int size = width * height * num_layers * sizeof(float);

     float *h_data = (float *)malloc(size);

     float *h_data_ref = (float *)malloc(size);

     float *d_data = NULL;

     cudaMalloc((void **)&d_data, size);

     for (unsigned int layer = ; layer < num_layers; layer++)

     {

         for (int i = ; i < (int)(width * height); i++)

             h_data[layer*width*height + i] = (float)i;

     }

     for (unsigned int layer = ; layer < num_layers; layer++)

     {

         for (int i = ; i < (int)(width * height); i++)

             h_data_ref[layer*width*height + i] = - h_data[layer*width*height + i] + layer;

     }

     printf("\n\t\Input data\n\t");

     for (int i = ; i < num_layers; i++)

     {

         for (int j = ; j < OUTPUT; j++)

         {

             for(int k=;k<OUTPUT;k++)

                 printf("%2.1f ", h_data[i*width*height+j*width+k]);

             printf("\n\t");

         }

         printf("\n\t");

     }

     printf("\n\t\Ideal output data\n\t");

     for (int i = ; i < num_layers; i++)

     {

         for (int j = ; j < OUTPUT; j++)

         {

             for (int k = ; k<OUTPUT; k++)

                 printf("%2.1f ", h_data_ref[i*width*height + j*width + k]);

             printf("\n\t");

         }

         printf("\n\t");

     }

     // 设置 CUDA 3D 数组参数和数据拷贝

     cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc(, , , , cudaChannelFormatKindFloat);

     cudaArray *cu_3darray;

     cudaMalloc3DArray(&cu_3darray, &channelDesc, make_cudaExtent(width, height, num_layers), cudaArrayLayered);

     cudaMemcpy3DParms myparms = {  };

     myparms.srcPos = make_cudaPos(, , );

     myparms.dstPos = make_cudaPos(, , );

     myparms.srcPtr = make_cudaPitchedPtr(h_data, width * sizeof(float), width, height);

     myparms.dstArray = cu_3darray;

     myparms.extent = make_cudaExtent(width, height, num_layers);

     myparms.kind = cudaMemcpyHostToDevice;

     cudaMemcpy3D(&myparms);

     // 设置纹理参数并绑定

     tex.addressMode[] = cudaAddressModeWrap;

     tex.addressMode[] = cudaAddressModeWrap;

     tex.filterMode = cudaFilterModeLinear;

     tex.normalized = true;

     cudaBindTextureToArray(tex, cu_3darray, channelDesc);

     dim3 dimBlock(, , );

     dim3 dimGrid(width / dimBlock.x, height / dimBlock.y, );

     printf("Covering 2D data of %d * %d * %d: Grid size is %d x %d, each block has 8 x 8 threads\n", width, height, num_layers, dimGrid.x, dimGrid.y);

     transformKernel << < dimGrid, dimBlock >> >(d_data, width, height, );// 预跑

     cudaDeviceSynchronize();

     StopWatchInterface *timer = NULL;

     sdkCreateTimer(&timer);

     sdkStartTimer(&timer);

     for (unsigned int layer = ; layer < num_layers; layer++)// 启用多个核，每个核完成一层

         transformKernel << < dimGrid, dimBlock,  >> >(d_data, width, height, layer);

     cudaDeviceSynchronize();

     sdkStopTimer(&timer);

     printf("\n\Time: %.3f msec, %.2f Mtexlookups/sec\n", sdkGetTimerValue(&timer), (width *height *num_layers / (sdkGetTimerValue(&timer) / 1000.0f) / 1e6));

     sdkDeleteTimer(&timer);

     // 返回计算结果并检验

     memset(h_data, , size);

     cudaMemcpy(h_data, d_data, size, cudaMemcpyDeviceToHost);

     if (checkCmdLineFlag(argc, (const char **)argv, "regression"))

         sdkWriteFile<float>("./data/regression.dat", h_data, width * width, 0.0f, false);

     else

         printf("Comparing kernel output to expected data return %d\n", compareData(h_data, h_data_ref, width * height * num_layers, MIN_EPSILON_ERROR, 0.0f));

     printf("\n\tActual output data\n\t");

     for (int i = ; i < num_layers; i++)

     {

         for (int j = ; j < OUTPUT; j++)

         {

             for (int k = ; k<OUTPUT; k++)

                 printf("%2.1f ", h_data[i*width*height + j*width + k]);

             printf("\n\t");

         }

         printf("\n\t");

     }

     free(h_data);

     free(h_data_ref);

     cudaFree(d_data);

     cudaFreeArray(cu_3darray);

     getchar();

     return ;

 }

▶ 输出结果

    Input data

    0.0 1.0 2.0 3.0 4.0

    512.0 513.0 514.0 515.0 516.0

    1024.0 1025.0 1026.0 1027.0 1028.0

    1536.0 1537.0 1538.0 1539.0 1540.0

    2048.0 2049.0 2050.0 2051.0 2052.0

    0.0 1.0 2.0 3.0 4.0

    512.0 513.0 514.0 515.0 516.0

    1024.0 1025.0 1026.0 1027.0 1028.0

    1536.0 1537.0 1538.0 1539.0 1540.0

    2048.0 2049.0 2050.0 2051.0 2052.0

    0.0 1.0 2.0 3.0 4.0

    512.0 513.0 514.0 515.0 516.0

    1024.0 1025.0 1026.0 1027.0 1028.0

    1536.0 1537.0 1538.0 1539.0 1540.0

    2048.0 2049.0 2050.0 2051.0 2052.0

    0.0 1.0 2.0 3.0 4.0

    512.0 513.0 514.0 515.0 516.0

    1024.0 1025.0 1026.0 1027.0 1028.0

    1536.0 1537.0 1538.0 1539.0 1540.0

    2048.0 2049.0 2050.0 2051.0 2052.0

    0.0 1.0 2.0 3.0 4.0

    512.0 513.0 514.0 515.0 516.0

    1024.0 1025.0 1026.0 1027.0 1028.0

    1536.0 1537.0 1538.0 1539.0 1540.0

    2048.0 2049.0 2050.0 2051.0 2052.0

    Ideal output data

    0.0 -1.0 -2.0 -3.0 -4.0

    -512.0 -513.0 -514.0 -515.0 -516.0

    -1024.0 -1025.0 -1026.0 -1027.0 -1028.0

    -1536.0 -1537.0 -1538.0 -1539.0 -1540.0

    -2048.0 -2049.0 -2050.0 -2051.0 -2052.0

    1.0 0.0 -1.0 -2.0 -3.0

    -511.0 -512.0 -513.0 -514.0 -515.0

    -1023.0 -1024.0 -1025.0 -1026.0 -1027.0

    -1535.0 -1536.0 -1537.0 -1538.0 -1539.0

    -2047.0 -2048.0 -2049.0 -2050.0 -2051.0

    2.0 1.0 0.0 -1.0 -2.0

    -510.0 -511.0 -512.0 -513.0 -514.0

    -1022.0 -1023.0 -1024.0 -1025.0 -1026.0

    -1534.0 -1535.0 -1536.0 -1537.0 -1538.0

    -2046.0 -2047.0 -2048.0 -2049.0 -2050.0

    3.0 2.0 1.0 0.0 -1.0

    -509.0 -510.0 -511.0 -512.0 -513.0

    -1021.0 -1022.0 -1023.0 -1024.0 -1025.0

    -1533.0 -1534.0 -1535.0 -1536.0 -1537.0

    -2045.0 -2046.0 -2047.0 -2048.0 -2049.0

    4.0 3.0 2.0 1.0 0.0

    -508.0 -509.0 -510.0 -511.0 -512.0

    -1020.0 -1021.0 -1022.0 -1023.0 -1024.0

    -1532.0 -1533.0 -1534.0 -1535.0 -1536.0

    -2044.0 -2045.0 -2046.0 -2047.0 -2048.0

    Covering 2D data of  *  * : Grid size is  x , each block has  x  threads

Time: 0.995 msec, 1317.00 Mtexlookups/sec

Comparing kernel output to expected data return 

    Actual output data

    0.0 -1.0 -2.0 -3.0 -4.0

    -512.0 -513.0 -514.0 -515.0 -516.0

    -1024.0 -1025.0 -1026.0 -1027.0 -1028.0

    -1536.0 -1537.0 -1538.0 -1539.0 -1540.0

    -2048.0 -2049.0 -2050.0 -2051.0 -2052.0

    1.0 0.0 -1.0 -2.0 -3.0

    -511.0 -512.0 -513.0 -514.0 -515.0

    -1023.0 -1024.0 -1025.0 -1026.0 -1027.0

    -1535.0 -1536.0 -1537.0 -1538.0 -1539.0

    -2047.0 -2048.0 -2049.0 -2050.0 -2051.0

    2.0 1.0 0.0 -1.0 -2.0

    -510.0 -511.0 -512.0 -513.0 -514.0

    -1022.0 -1023.0 -1024.0 -1025.0 -1026.0

    -1534.0 -1535.0 -1536.0 -1537.0 -1538.0

    -2046.0 -2047.0 -2048.0 -2049.0 -2050.0

    3.0 2.0 1.0 0.0 -1.0

    -509.0 -510.0 -511.0 -512.0 -513.0

    -1021.0 -1022.0 -1023.0 -1024.0 -1025.0

    -1533.0 -1534.0 -1535.0 -1536.0 -1537.0

    -2045.0 -2046.0 -2047.0 -2048.0 -2049.0

    4.0 3.0 2.0 1.0 0.0

    -508.0 -509.0 -510.0 -511.0 -512.0

    -1020.0 -1021.0 -1022.0 -1023.0 -1024.0

    -1532.0 -1533.0 -1534.0 -1535.0 -1536.0

    -2044.0 -2045.0 -2046.0 -2047.0 -2048.0

▶ 涨姿势

● 与前面立方体贴图纹理不同的地方：申请 CUDA3D 数组的时候使用标志 cudaArrayLayered 而不是 cudaArrayCubemap，并注意调整相关的维度参数。

0_Simple__simpleLayeredTexture的更多相关文章

随机推荐

tarjan强联通分量（模板）
#include<cstdio> #include<cstring> #include<cstdlib> #include<iostream> #inc ...
chrom 自带截屏用法
1 . F12调出控制台 2 . Ctrl+Shift+p 3 . 输入capture
LeetCode Majority Element Python
Given an array of size n, find the majority element. The majority element is the element that appear ...
使用webpack搭建vue项目中遇到的问题
1:data数据文件经试验,需要放在生成的build文件夹中才能生效,但是应该把data文件夹先放在src中,然后如何定义config文件,让其复制过去? new CopyWebpackPlugin( ...
hasura 的3factor 架构结论
hasura 是一个很不错的开发团队,开发了好几款,不错的产品,最近的graphql engine 是很热的一款 graphql 引擎,同时团队提出了3factor 系统架构理论参考网站 https ...
Tensoflw.js - 02 - 模型与内存管理（易懂）
Tensoflw.js - 02 - 模型与内存管理(易懂) 参考 W3Cschool 文档:https://www.w3cschool.cn/tensorflowjs/ 本文主要翻译一些英文注释,添 ...
JS页面打印预览功能
点击按钮后就可以直接打印预览,并且隐藏了按钮,如果你不希望按钮显示,可以把那部分代码删除就可以了. <html><head><meta http-equiv=" ...
Java性能分析之线程栈详解与性能分析
Java性能分析之线程栈详解 Java性能分析迈不过去的一个关键点是线程栈,新的性能班级也讲到了JVM这一块,所以本篇文章对线程栈进行基础知识普及以及如何对线程栈进行性能分析. 基本概念线程堆栈也称 ...
【转】每天一个linux命令（20）：find命令之exec
原文网址:http://www.cnblogs.com/peida/archive/2012/11/14/2769248.html find是我们很常用的一个Linux命令,但是我们一般查找出来的并不 ...
SQL SERVER 2008 彻底卸载干净方法（转）
最近安装SQL SERVER 2008失败后,再重新安装时老是报错,东搞西搞的很难卸干净.但又不方便重装系统,经按下面方法终于搞定并成功安装上2008 1.停掉SQL SERVER 2008所有相关服 ...

0_Simple__simpleLayeredTexture

0_Simple__simpleLayeredTexture的更多相关文章

随机推荐

热门专题