cuda：用纹理内存实现插值操作

#include <cuda_runtime.h>
#include <iostream>

// 定义纹理引用
texture<float, 1, cudaReadModeElementType> texRef;

__global__ void linearInterpolation(float* output, const int width, float scale) {
    int x = blockIdx.x * blockDim.x + threadIdx.x;
    float texX = x * scale;

    if (x < width) {
        output[x] = tex1Dfetch(texRef, texX);
    }
}

int main() {
    const int width = 100;
    float h_input[width];
    float h_output[width];

    // 初始化输入数据
    for (int i = 0; i < width; i++) {
        h_input[i] = i*i;
    }

    float* d_input;
    float* d_output;

    cudaMalloc(&d_input, width * sizeof(float));
    cudaMalloc(&d_output, width * sizeof(float));

    cudaMemcpy(d_input, h_input, width * sizeof(float), cudaMemcpyHostToDevice);

    // 将纹理与输入数据绑定
    cudaBindTexture(NULL, texRef, d_input, width * sizeof(float));

    // 定义线程块和网格尺寸
    int threadsPerBlock = 256;
    int blocksPerGrid = (width + threadsPerBlock - 1) / threadsPerBlock;

    // 运行核函数
    float scale = 0.2f; // some scale factor for interpolation
    linearInterpolation << <blocksPerGrid, threadsPerBlock >> > (d_output, width, scale);

    cudaMemcpy(h_output, d_output, width * sizeof(float), cudaMemcpyDeviceToHost);

    // 打印结果
    for (int i = 0; i < width; i++) {
        std::cout << h_output[i] << " ";
    }
    std::cout << std::endl;

    cudaUnbindTexture(texRef);

    cudaFree(d_input);
    cudaFree(d_output);

    return 0;
}