C++ GPU编程(英伟达CUDA)

安装编译环境https://developer.download.nvidia.cn/compute/cuda/12.5.0/local_installers/cuda_12.5.0_555.85_windows.exe

CMakeLists.txt

cmake_minimum_required(VERSION 3.10)

set(CMAKE_CXX_STANDARD 17)
set(CMAKE_BUILD_TYPE Release)
#set(CMAKE_CUDA_ARCHITECTURES 52;70;75;86)
set(CMAKE_CUDA_SEPARABLE_COMPILATION ON)

project(hellocuda LANGUAGES CXX CUDA)

add_executable(main main.cu hello.cu)

target_include_directories(main PUBLIC ../../include)
target_compile_options(main PUBLIC $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
target_compile_options(main PUBLIC $<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr>)
# --use_fast_math   sinf 替换成 __sinf
# --ftz=true 会把极小数(denormal)退化为0
# --prec-div=false 降低除法的精度换取速度。
# --prec-sqrt=false 降低开方的精度换取速度。
# --fmad 因为非常重要，所以默认就是开启的，会自动把 a * b + c 优化成乘加(FMA)指令。
# 开启 --use_fast_math 后会自动开启上述所有

 CudaAllocator.h

template <class T>
struct CudaAllocator {
    using value_type = T;

    T* allocate(size_t size) {
        T* ptr = nullptr;
        checkCudaErrors(cudaMallocManaged(&ptr, size * sizeof(T)));
        return ptr;
    }

    void deallocate(T* ptr, size_t size = 0) {
        checkCudaErrors(cudaFree(ptr));
    }

    template <class ...Args>
    void construct(T* p, Args &&...args) {
        if constexpr (!(sizeof...(Args) == 0 && std::is_pod_v<T>))
            ::new((void*)p) T(std::forward<Args>(args)...);
    }
};

 hello.cu

#include <cstdio>
#include <cuda_runtime.h>

__device__ void say_hello3() {  // 定义
    printf("Hello, world!\n");
}

 main.cu

#include <cstdio>
#include <cuda_runtime.h>
#include "helper_cuda.h"
#include <vector>
#include "CudaAllocator.h"
#include <thrust/universal_vector.h>
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>
#include <thrust/generate.h>
#include <thrust/for_each.h>

__device__ void say_hello3();  // 声明

__device__ __inline__ void say_hello() {
    printf("Hello, world, world from GPU!\n");
}

__host__ void say_hello_host() {
    printf("Hello, world from CPU!\n");
}

__host__ __device__ void say_hello2() {
#ifdef __CUDA_ARCH__
    printf("Hello, world from GPU architecture %d!\n", __CUDA_ARCH__);
#else
    printf("Hello, world from CPU!\n");
#endif
}

__global__ void another() {
    int localVal = 1;
    printf("another: Thread %d of %d\n", threadIdx.x, blockDim.x);
}

__global__ void kernel() {
    unsigned int tid = blockDim.x * blockIdx.x + threadIdx.x;
    unsigned int tnum = blockDim.x * gridDim.x;
    printf("Flattened Thread %d of %d\n", tid, tnum);
    say_hello2();
    say_hello3();
    another << <1, tnum >> > ();  // 核函数调用核函数
    printf("kernel: called another with %d threads\n", tnum);
}

__global__ void kernel2() {
    printf("Block (%d,%d,%d) of (%d,%d,%d), Thread (%d,%d,%d) of (%d,%d,%d)\n",
        blockIdx.x, blockIdx.y, blockIdx.z,
        gridDim.x, gridDim.y, gridDim.z,
        threadIdx.x, threadIdx.y, threadIdx.z,
        blockDim.x, blockDim.y, blockDim.z);
}

// 默认host
constexpr const char* cuthead(const char* p) {
    return p + 1;
}

// 内存访问
__global__ void kernel_memory(int* pret) {
    *pret = 42;
}

// 数组并行处理
__global__ void kernel_array(int* arr, int n) {
    for (int i = blockDim.x * blockIdx.x + threadIdx.x;
        i < n; i += blockDim.x * gridDim.x) {
        arr[i] = i;
    }
}

// 并行处理模板
template <class Func>
__global__ void parallel_for(int n, Func func) {
    for (int i = blockDim.x * blockIdx.x + threadIdx.x;
        i < n; i += blockDim.x * gridDim.x) {
        func(i);
    }
}

// 自定义加减乘除
__device__ __inline__ int float_atomic_add(float* dst, float src) {
    int old = __float_as_int(*dst), expect;
    do {
        expect = old;
        old = atomicCAS((int*)dst, expect,
            __float_as_int(__int_as_float(expect) + src));
    } while (expect != old);
    return old;
}


// map
template <int blockSize, class T>
__global__ void parallel_sum_kernel(T* sum, T const* arr, int n) {
    __shared__ volatile int local_sum[blockSize];
    int j = threadIdx.x;
    int i = blockIdx.x;
    T temp_sum = 0;
    for (int t = i * blockSize + j; t < n; t += blockSize * gridDim.x) {
        temp_sum += arr[t];
    }
    local_sum[j] = temp_sum;
    __syncthreads();
    if constexpr (blockSize >= 1024) {
        if (j < 512)
            local_sum[j] += local_sum[j + 512];
        __syncthreads();
    }
    if constexpr (blockSize >= 512) {
        if (j < 256)
            local_sum[j] += local_sum[j + 256];
        __syncthreads();
    }
    if constexpr (blockSize >= 256) {
        if (j < 128)
            local_sum[j] += local_sum[j + 128];
        __syncthreads();
    }
    if constexpr (blockSize >= 128) {
        if (j < 64)
            local_sum[j] += local_sum[j + 64];
        __syncthreads();
    }
    if (j < 32) {
        if constexpr (blockSize >= 64)
            local_sum[j] += local_sum[j + 32];
        if constexpr (blockSize >= 32)
            local_sum[j] += local_sum[j + 16];
        if constexpr (blockSize >= 16)
            local_sum[j] += local_sum[j + 8];
        if constexpr (blockSize >= 8)
            local_sum[j] += local_sum[j + 4];
        if constexpr (blockSize >= 4)
            local_sum[j] += local_sum[j + 2];
        if (j == 0) {
            sum[i] = local_sum[0] + local_sum[1];
        }
    }
}

// reduce
template <int reduceScale = 4096, int blockSize = 256, int cutoffSize = reduceScale * 2, class T>
int parallel_sum(T const* arr, int n) {
    if (n > cutoffSize) {
        std::vector<int, CudaAllocator<int>> sum(n / reduceScale);
        parallel_sum_kernel<blockSize> << <n / reduceScale, blockSize >> > (sum.data(), arr, n);
        return parallel_sum(sum.data(), n / reduceScale);
    }
    else {
        checkCudaErrors(cudaDeviceSynchronize());
        T final_sum = 0;
        for (int i = 0; i < n; i++) {
            final_sum += arr[i];
        }
        return final_sum;
    }
}

// 共享内存
template <int blockSize, class T>
__global__ void parallel_transpose(T* out, T const* in, int nx, int ny) {
    int x = blockIdx.x * blockSize + threadIdx.x;
    int y = blockIdx.y * blockSize + threadIdx.y;
    if (x >= nx || y >= ny) return;
    __shared__ T tmp[blockSize * blockSize];
    int rx = blockIdx.y * blockSize + threadIdx.x;
    int ry = blockIdx.x * blockSize + threadIdx.y;
    tmp[threadIdx.y * blockSize + threadIdx.x] = in[ry * nx + rx];
    __syncthreads();
    out[y * nx + x] = tmp[threadIdx.x * blockSize + threadIdx.y];
}

void testArray() {
    int n = 65535;
    int* arr;
    checkCudaErrors(cudaMallocManaged(&arr, n * sizeof(int)));

    int nthreads = 128;
    int nblocks = (n + nthreads + 1) / nthreads;
    kernel_array << <nblocks, nthreads >> > (arr, n);

    checkCudaErrors(cudaDeviceSynchronize());
    for (int i = 0; i < n; i++) {
        printf("arr[%d]: %d\n", i, arr[i]);
    }

    cudaFree(arr);
}

void testVector() {
    int n = 65536;
    std::vector<int, CudaAllocator<int>> arr(n);

    parallel_for << <32, 128 >> > (n, [arr = arr.data()] __device__(int i) {
        arr[i] = i;
    });

    checkCudaErrors(cudaDeviceSynchronize());
    for (int i = 0; i < n; i++) {
        printf("arr[%d] = %d\n", i, arr[i]);
    }

}

void testMath() {
    int n = 1 << 25;
    std::vector<float, CudaAllocator<float>> gpu(n);
    std::vector<float> cpu(n);

    for (int i = 0; i < n; i++) {
        cpu[i] = sinf(i);
    }

    parallel_for << <n / 512, 128 >> > (n, [gpu = gpu.data()] __device__(int i) {
        gpu[i] = __sinf(i);
    });
    checkCudaErrors(cudaDeviceSynchronize());

    //for (int i = 0; i < n; i++) {
        //printf("diff %d = %f\n", i, gpu[i] - cpu[i]);
    //}
}

void testTrust() {
    int n = 65536;
    thrust::universal_vector<float> x(n);   // CPU或GPU
    thrust::host_vector<float> x_host(n);   // CPU
    thrust::device_vector<float> x_dev = x_host;    // GPU

    thrust::for_each(x_dev.begin(), x_dev.end(), [] __device__(float& x) {
        x += 100.f;
    });
}

void testAtomic() {
    // atomicAdd(dst, src) 
    // atomicSub
    // atomicOr
    // atomicAnd
    // atomicXor
    // atomicMax
    // atomicMin
    // atomicExch   old = *dst; *dst = src
    // atomicCAS    automic::exchange
}

void test_share() {
    int nx = 1 << 14, ny = 1 << 14;
    std::vector<int, CudaAllocator<int>> in(nx * ny);
    std::vector<int, CudaAllocator<int>> out(nx * ny);
    parallel_transpose<32> << <dim3(nx / 32, ny / 32, 1), dim3(32, 32, 1) >> >
        (out.data(), in.data(), nx, ny);
}

// host 可以调用 global；global 可以调用 device；device 可以调用 device
// 实际上 GPU 的板块相当于 CPU 的线程，GPU 的线程相当于 CPU 的SIMD
// 数学函数sqrtf，rsqrtf，cbrtf，rcbrtf，powf，sinf，cosf，sinpif，cospif，sincosf，sincospif，logf，log2f，log10f，expf，exp2f，exp10f，tanf，atanf，asinf，acosf，fmodf，fabsf，fminf，fmaxf
// 低精度，高性能 __sinf,__expf、__logf、__cosf、__powf, __fdividef(x, y) 
// old = atomicAdd(dst, src) 相当于 old = *dst; *dst += src 
int main() {
    // 三重尖括号里的，第一个参数表示板块数量，第二个参数决定着启动 kernel 时所用 GPU 的线程数量
    kernel<<<1, 1>>>(); // CPU调用GPU函数异步执行
    kernel2<<<dim3(2, 1, 1), dim3(2, 2, 2)>>>();
    cudaDeviceSynchronize(); // CPU等待GPU完成
    say_hello_host();

    int* pret;
    checkCudaErrors(cudaMallocManaged(&pret, sizeof(int)));
    kernel_memory << <1, 1 >> > (pret);
    checkCudaErrors(cudaDeviceSynchronize());
    printf("result: %d\n", *pret);
    cudaFree(pret);

    testArray();

    float sin10 = sinf(10.f);

    /// map-reduce
    int n = 1 << 24;
    std::vector<int, CudaAllocator<int>> arr(n);
    int final_sum = parallel_sum(arr.data(), n);
    return 0;
}

参考

GitHub - rapidsai/cudf: cuDF - GPU DataFrame Library

GitHub - stotko/stdgpu: stdgpu: Efficient STL-like Data Structures on the GPU

cuda-samples/Common/helper_string.h at 5f97d7d0dff880bc6567faa4c5e62e389a6d6999 · NVIDIA/cuda-samples · GitHub

 https://github.com/NVIDIA/cuda-samples/blob/5f97d7d0dff880bc6567faa4c5e62e389a6d6999/Common/helper_cuda.h#L31

https://www.nvidia.cn/docs/IO/51635/NVIDIA_CUDA_Programming_Guide_1.1_chs.pdf

https://developer.download.nvidia.cn/CUDA/training/StreamsAndConcurrencyWebinar.pdf

CUDA Pro Tip: Write Flexible Kernels with Grid-Stride Loops | NVIDIA Technical Blog

https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#texture-and-surface-memory

---

创作不易，小小的支持一下吧！