CUDA向量类型（float2、float3、float4）的效率

CUDA是一种用于并行计算的编程模型，可以有效地利用GPU的并行计算能力。在CUDA中，向量类型可以提高计算效率，其中包括float2、float3和float4等类型。这些向量类型可以在单个内存事务中处理多个数据元素，从而减少了内存访问的次数，提高了计算速度。

cuda
__global__ void vectorAdd(float2* a, float2* b, float2* c, int size) {
    int tid = blockIdx.x * blockDim.x + threadIdx.x;
    if (tid < size/2) {
        c[tid].x = a[tid].x + b[tid].x;
        c[tid].y = a[tid].y + b[tid].y;
    }
}
int main() {
    int size = 10000;
    float2 *h_a, *h_b, *h_c;
    float2 *d_a, *d_b, *d_c;
    
    // 分配内存并初始化数据
    
    // 在GPU上分配内存
    cudaMalloc((void**)&d_a, size * sizeof(float2));
    cudaMalloc((void**)&d_b, size * sizeof(float2));
    cudaMalloc((void**)&d_c, size * sizeof(float2));
    
    // 将数据从主机内存拷贝到GPU内存
    cudaMemcpy(d_a, h_a, size * sizeof(float2), cudaMemcpyHostToDevice);
    cudaMemcpy(d_b, h_b, size * sizeof(float2), cudaMemcpyHostToDevice);
    
    // 启动CUDA核函数
    int threadsPerBlock = 256;
    int blocksPerGrid = (size + threadsPerBlock - 1) / threadsPerBlock;
    vectorAdd<<>>(d_a, d_b, d_c, size);
    
    // 将结果从GPU内存拷贝到主机内存
    cudaMemcpy(h_c, d_c, size * sizeof(float2), cudaMemcpyDeviceToHost);
    
    // 清理内存
    
    return 0;
}

cuda
__global__ void vectorAdd(float3* a, float3* b, float3* c, int size) {
    int tid = blockIdx.x * blockDim.x + threadIdx.x;
    if (tid < size/3) {
        c[tid].x = a[tid].x + b[tid].x;
        c[tid].y = a[tid].y + b[tid].y;
        c[tid].z = a[tid].z + b[tid].z;
    }
}
int main() {
    int size = 10000;
    float3 *h_a, *h_b, *h_c;
    float3 *d_a, *d_b, *d_c;
    
    // 分配内存并初始化数据
    
    // 在GPU上分配内存
    cudaMalloc((void**)&d_a, size * sizeof(float3));
    cudaMalloc((void**)&d_b, size * sizeof(float3));
    cudaMalloc((void**)&d_c, size * sizeof(float3));
    
    // 将数据从主机内存拷贝到GPU内存
    cudaMemcpy(d_a, h_a, size * sizeof(float3), cudaMemcpyHostToDevice);
    cudaMemcpy(d_b, h_b, size * sizeof(float3), cudaMemcpyHostToDevice);
    
    // 启动CUDA核函数
    int threadsPerBlock = 256;
    int blocksPerGrid = (size + threadsPerBlock - 1) / threadsPerBlock;
    vectorAdd<<>>(d_a, d_b, d_c, size);
    
    // 将结果从GPU内存拷贝到主机内存
    cudaMemcpy(h_c, d_c, size * sizeof(float3), cudaMemcpyDeviceToHost);
    
    // 清理内存
    
    return 0;
}

cuda
__global__ void vectorAdd(float4* a, float4* b, float4* c, int size) {
    int tid = blockIdx.x * blockDim.x + threadIdx.x;
    if (tid < size/4) {
        c[tid].x = a[tid].x + b[tid].x;
        c[tid].y = a[tid].y + b[tid].y;
        c[tid].z = a[tid].z + b[tid].z;
        c[tid].w = a[tid].w + b[tid].w;
    }
}
int main() {
    int size = 10000;
    float4 *h_a, *h_b, *h_c;
    float4 *d_a, *d_b, *d_c;
    
    // 分配内存并初始化数据
    
    // 在GPU上分配内存
    cudaMalloc((void**)&d_a, size * sizeof(float4));
    cudaMalloc((void**)&d_b, size * sizeof(float4));
    cudaMalloc((void**)&d_c, size * sizeof(float4));
    
    // 将数据从主机内存拷贝到GPU内存
    cudaMemcpy(d_a, h_a, size * sizeof(float4), cudaMemcpyHostToDevice);
    cudaMemcpy(d_b, h_b, size * sizeof(float4), cudaMemcpyHostToDevice);
    
    // 启动CUDA核函数
    int threadsPerBlock = 256;
    int blocksPerGrid = (size + threadsPerBlock - 1) / threadsPerBlock;
    vectorAdd<<>>(d_a, d_b, d_c, size);
    
    // 将结果从GPU内存拷贝到主机内存
    cudaMemcpy(h_c, d_c, size * sizeof(float4), cudaMemcpyDeviceToHost);
    
    // 清理内存
    
    return 0;
}