13

我在边界寻址模式 ( cudaAddressModeBorder) 中使用 CUDA 纹理。我正在使用tex2D<float>(). 当纹理坐标落在纹理之外时,tex2D<float>()返回0.

如何将此返回的边框值更改0为其他值?我可以手动检查纹理坐标并自己设置边框值。我想知道是否有 CUDA API 可以设置这样的边界值。

4

2 回答 2

22

正如 sgarizvi 所提到的,CUDA 仅支持四种不可自定义的地址模式,即,clampborderwrapmirror,这些在 3.2.11.1 节中有描述。CUDA 编程指南。

前两者适用于非标准化和标准化坐标,而后两者仅适用于标准化坐标。

为了描述前两个,为了简单起见,让我们考虑非归一化坐标情况并考虑一维信号。在这种情况下,输入序列是c[k],其中k=0,...,M-1

cudaAddressModeClamp

信号c[k]在外部继续,k=0,...,M-1因此c[k] = c[0]fork < 0c[k] = c[M-1]for k >= M

cudaAddressModeBorder

信号c[k]在外部继续,k=0,...,M-1因此c[k] = 0fork < 0和 for k >= M

现在,为了描述最后两种地址模式,我们不得不考虑归一化坐标,因此假设 1D 输入信号样本为c[k / M],其中k=0,...,M-1

cudaAddressModeWrap

信号c[k / M]在外部继续k=0,...,M-1,因此它是周期性的,周期等于M。换句话说,c[(k + p * M) / M] = c[k / M]对于任何(正、负或消失的)整数p

cudaAddressModeMirror

信号c[k / M]在外部继续k=0,...,M-1,因此它是周期性的,周期等于2 * M - 2。换句话说,c[l / M] = c[k / M]对于任何lk这样的(l + k)mod(2 * M - 2) = 0

以下代码说明了所有四种可用的地址模式

#include <stdio.h>

texture<float, 1, cudaReadModeElementType> texture_clamp;
texture<float, 1, cudaReadModeElementType> texture_border;
texture<float, 1, cudaReadModeElementType> texture_wrap;
texture<float, 1, cudaReadModeElementType> texture_mirror;

/********************/
/* CUDA ERROR CHECK */
/********************/
#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, char *file, int line, bool abort=true)
{
    if (code != cudaSuccess) 
    {
        fprintf(stderr,"GPUassert: %s %s %d\n", cudaGetErrorString(code), file, line);
        if (abort) exit(code);
    }
}

/******************************/
/* CUDA ADDRESS MODE CLAMPING */
/******************************/
__global__ void Test_texture_clamping(const int M) {

    printf("Texture clamping - i = %i; value = %f\n", -threadIdx.x, tex1D(texture_clamp, -(float)threadIdx.x));
    printf("Texture clamping - i = %i; value = %f\n", M + threadIdx.x, tex1D(texture_clamp, (float)(M + threadIdx.x)));

}

/****************************/
/* CUDA ADDRESS MODE BORDER */
/****************************/
__global__ void Test_texture_border(const int M) {

    printf("Texture border - i = %i; value = %f\n", -threadIdx.x, tex1D(texture_border, -(float)threadIdx.x));
    printf("Texture border - i = %i; value = %f\n", M + threadIdx.x, tex1D(texture_border, (float)(M + threadIdx.x)));

}

/**************************/
/* CUDA ADDRESS MODE WRAP */
/**************************/
__global__ void Test_texture_wrap(const int M) {

    printf("Texture wrap - i = %i; value = %f\n", -threadIdx.x, tex1D(texture_wrap, -(float)threadIdx.x/(float)M));
    printf("Texture wrap - i = %i; value = %f\n", M + threadIdx.x, tex1D(texture_wrap, (float)(M + threadIdx.x)/(float)M));

}

/****************************/
/* CUDA ADDRESS MODE MIRROR */
/****************************/
__global__ void Test_texture_mirror(const int M) {

    printf("Texture mirror - i = %i; value = %f\n", -threadIdx.x, tex1D(texture_mirror, -(float)threadIdx.x/(float)M));
    printf("Texture mirror - i = %i; value = %f\n", M + threadIdx.x, tex1D(texture_mirror, (float)(M + threadIdx.x)/(float)M));

}

/********/
/* MAIN */
/********/
void main(){

    const int M = 4;

    // --- Host side memory allocation and initialization
    float *h_data = (float*)malloc(M * sizeof(float));

    for (int i=0; i<M; i++) h_data[i] = (float)i;

    // --- Texture clamping
    cudaArray* d_data_clamping = NULL; gpuErrchk(cudaMallocArray(&d_data_clamping, &texture_clamp.channelDesc, M, 1)); 
    gpuErrchk(cudaMemcpyToArray(d_data_clamping, 0, 0, h_data, M * sizeof(float), cudaMemcpyHostToDevice)); 
    cudaBindTextureToArray(texture_clamp, d_data_clamping); 
    texture_clamp.normalized = false; 
    texture_clamp.addressMode[0] = cudaAddressModeClamp;

    dim3 dimBlock(2 * M, 1); dim3 dimGrid(1, 1);
    Test_texture_clamping<<<dimGrid,dimBlock>>>(M);

    printf("\n\n\n");

    // --- Texture border
    cudaArray* d_data_border = NULL; gpuErrchk(cudaMallocArray(&d_data_border, &texture_border.channelDesc, M, 1)); 
    gpuErrchk(cudaMemcpyToArray(d_data_border, 0, 0, h_data, M * sizeof(float), cudaMemcpyHostToDevice)); 
    cudaBindTextureToArray(texture_border, d_data_border); 
    texture_border.normalized = false; 
    texture_border.addressMode[0] = cudaAddressModeBorder;

    Test_texture_border<<<dimGrid,dimBlock>>>(M);

    printf("\n\n\n");

    // --- Texture wrap
    cudaArray* d_data_wrap = NULL; gpuErrchk(cudaMallocArray(&d_data_wrap, &texture_wrap.channelDesc, M, 1)); 
    gpuErrchk(cudaMemcpyToArray(d_data_wrap, 0, 0, h_data, M * sizeof(float), cudaMemcpyHostToDevice)); 
    cudaBindTextureToArray(texture_wrap, d_data_wrap); 
    texture_wrap.normalized = true; 
    texture_wrap.addressMode[0] = cudaAddressModeWrap;

    Test_texture_wrap<<<dimGrid,dimBlock>>>(M);

    printf("\n\n\n");

    // --- Texture mirror
    cudaArray* d_data_mirror = NULL; gpuErrchk(cudaMallocArray(&d_data_mirror, &texture_mirror.channelDesc, M, 1)); 
    gpuErrchk(cudaMemcpyToArray(d_data_mirror, 0, 0, h_data, M * sizeof(float), cudaMemcpyHostToDevice)); 
    cudaBindTextureToArray(texture_mirror, d_data_mirror); 
    texture_mirror.normalized = true ; 
    texture_mirror.addressMode[0] = cudaAddressModeMirror;

    Test_texture_mirror<<<dimGrid,dimBlock>>>(M);

    printf("\n\n\n");
}

这些是输出

index                  -7  -6  -5  -4  -3  -2  -1  0  1  2  3  4  5  6  7  8  9  10  11
clamp                   0   0   0   0   0   0   0  0  1  2  3  3  3  3  3  3  3   3   3
border                  0   0   0   0   0   0   0  0  1  2  3  0  0  0  0  0  0   0   0
wrap                    1   2   3   0   1   2   3  0  1  2  3  0  1  2  3  0  1   2   3
mirror                  1   2   3   3   2   1   0  0  1  2  3  3  2  1  0  0  1   2   3
于 2014-12-10T22:02:42.880 回答
4

截至目前(CUDA 5.5),CUDA 纹理获取行为不可自定义。4 种自动内置模式(即BorderClampWrapMirror)中只有一种可用于超出范围的纹理提取。

于 2013-09-26T07:05:11.327 回答