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 = 512, height = 512, num_layers = 5;
    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 = 0; layer < num_layers; layer++)
    {
        for (int i = 0; i < (int)(width * height); i++)
            h_data[layer*width*height + i] = (float)i;
    }
   
    for (unsigned int layer = 0; layer < num_layers; layer++)
    {
        for (int i = 0; i < (int)(width * height); i++)
            h_data_ref[layer*width*height + i] = - h_data[layer*width*height + i] + layer;
    }

    printf("
	Input data
	");
    for (int i = 0; i < num_layers; i++)
    {
        for (int j = 0; j < OUTPUT; j++)
        {
            for(int k=0;k<OUTPUT;k++)
                printf("%2.1f ", h_data[i*width*height+j*width+k]);
            printf("
	");
        }
        printf("
	");
    }
    printf("
	Ideal output data
	");
    for (int i = 0; i < num_layers; i++)
    {
        for (int j = 0; j < OUTPUT; j++)
        {
            for (int k = 0; k<OUTPUT; k++)
                printf("%2.1f ", h_data_ref[i*width*height + j*width + k]);
            printf("
	");
        }
        printf("
	");
    }

    // 设置 CUDA 3D 数组参数和数据拷贝
    cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc(32, 0, 0, 0, cudaChannelFormatKindFloat);
    cudaArray *cu_3darray;
    cudaMalloc3DArray(&cu_3darray, &channelDesc, make_cudaExtent(width, height, num_layers), cudaArrayLayered);
    cudaMemcpy3DParms myparms = { 0 };
    myparms.srcPos = make_cudaPos(0, 0, 0);
    myparms.dstPos = make_cudaPos(0, 0, 0);
    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[0] = cudaAddressModeWrap;
    tex.addressMode[1] = cudaAddressModeWrap;
    tex.filterMode = cudaFilterModeLinear;
    tex.normalized = true;
    cudaBindTextureToArray(tex, cu_3darray, channelDesc);

    dim3 dimBlock(8, 8, 1);
    dim3 dimGrid(width / dimBlock.x, height / dimBlock.y, 1);
    printf("Covering 2D data of %d * %d * %d: Grid size is %d x %d, each block has 8 x 8 threads
", width, height, num_layers, dimGrid.x, dimGrid.y);
    transformKernel << < dimGrid, dimBlock >> >(d_data, width, height, 0);// 预跑
    cudaDeviceSynchronize();

    StopWatchInterface *timer = NULL;
    sdkCreateTimer(&timer);
    sdkStartTimer(&timer);

    for (unsigned int layer = 0; layer < num_layers; layer++)// 启用多个核，每个核完成一层
        transformKernel << < dimGrid, dimBlock, 0 >> >(d_data, width, height, layer);
    cudaDeviceSynchronize();
    
    sdkStopTimer(&timer);
    printf("
Time: %.3f msec, %.2f Mtexlookups/sec
", 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, 0, 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
", compareData(h_data, h_data_ref, width * height * num_layers, MIN_EPSILON_ERROR, 0.0f));

    printf("
	Actual output data
	");
    for (int i = 0; i < num_layers; i++)
    {
        for (int j = 0; j < OUTPUT; j++)
        {
            for (int k = 0; k<OUTPUT; k++)
                printf("%2.1f ", h_data[i*width*height + j*width + k]);
            printf("
	");
        }
        printf("
	");
    }

    free(h_data);
    free(h_data_ref);
    cudaFree(d_data);
    cudaFreeArray(cu_3darray);

    getchar();
    return 0;
}

▶ 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 512 * 512 * 5: Grid size is 64 x 64, each block has 8 x 8 threads

Time: 0.995 msec, 1317.00 Mtexlookups/sec
Comparing kernel output to expected data return 1

    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。