CUDA Tensor Stacking Optimization Guide

void stack_naive(float** inputs, float* output, int n, int size) {
    for (int i = 0; i < n; i++) {
        cudaMemcpy(output + i * size, inputs[i],
                   size * sizeof(float), cudaMemcpyDeviceToDevice);
    }
}

void stack_opt(float** inputs, float* output, int n, int size) {
    // Copy input pointers to device
    float** d_inputs;
    cudaMalloc(&d_inputs, n * sizeof(float*));
    cudaMemcpy(d_inputs, inputs, n * sizeof(float*), cudaMemcpyHostToDevice);

    int total = n * size;
    int blocks = (total + 255) / 256;
    stack_kernel<<<blocks, 256>>>(d_inputs, output, n, size);

    cudaFree(d_inputs);
}