环境：

  • NVIDIA GeForce GTX 1050
  • cuda 10.2.89 windows
  • visual studio 2017
  • windows SDK 10.0.14393.0
  • go 1.13.4 windows/amd64
lib.cu
__global__ void devDot(double *x, double *y, int n, double *r) {    double res = 0.0;    for (int i = 0; i < n; i++) {        res += x[i] * y[i];    }    *r = res;}extern "C" __declspec(dllexport) void dot(double *x, double *y, int n, double *r) {    double *xd, *yd, *rd;    int sz = sizeof(double) * n;    cudaMalloc(&xd, sz);    cudaMalloc(&yd, sz);    cudaMalloc(&rd, sizeof(double));    cudaMemcpy(xd, x, sz, cudaMemcpyHostToDevice);    cudaMemcpy(yd, y, sz, cudaMemcpyHostToDevice);    devDot<<<1, 1>>>(xd, yd, n, rd);    cudaDeviceSynchronize();    cudaMemcpy(r, rd, sizeof(double), cudaMemcpyDeviceToHost);    cudaFree(xd);    cudaFree(yd);    cudaFree(rd);}
cu__global__cudaMalloccudaMalloccudaMemcpydevDot<<<1, 1>>>
PATH
nvcc lib.cu -o libcuda.dll --shared
main.gomain.go
package mainimport (    "math/rand"    "syscall"    "time"    "unsafe")type Lib struct {    dll     *syscall.DLL    dotProc *syscall.Proc}func LoadLib() (*Lib, error) {    l := &Lib{}    var err error    if l.dll, err = syscall.LoadDLL("libcuda.dll"); nil != err {        return nil, err    }    if l.dotProc, err = l.dll.FindProc("dot"); nil != err {        l.dll.Release()        return nil, err    }    return l, nil}func (l *Lib) Release() {    l.dll.Release()}func (l *Lib) Dot(x, y []float64) float64 {    var r float64    l.dotProc.Call(        uintptr(unsafe.Pointer(&x[0])),        uintptr(unsafe.Pointer(&y[0])),        uintptr(len(x)),        uintptr(unsafe.Pointer(&r)),    )    return r}const N = 1 << 20func main() {    lib, err := LoadLib()    if nil != err {        println(err.Error())        return    }    defer lib.Release()    rand.Seed(time.Now().Unix())    x, y := make([]float64, N), make([]float64, N)    for i := 0; i < N; i++ {        x[i], y[i] = rand.Float64(), rand.Float64()    }    var r float64    for i := 0; i < N; i++ {        r += x[i] * y[i]    }    println(r)    println(lib.Dot(x, y))    t := time.Now()    for i := 0; i < 100; i++ {        var r float64        for i := 0; i < N; i++ {            r += x[i] * y[i]        }    }    println(time.Now().Sub(t).Microseconds())    t = time.Now()    for i := 0; i < 100; i++ {        lib.Dot(x, y)    }    println(time.Now().Sub(t).Microseconds())}

在 golang 中使用动态加载，比较计算结果和运行时间，运行的结果基本是同样的画风，计算结果正确，但是在性能上，这种哈喽级别的 CUDA 尝试终究惨败被虐出 N 个数量级，浮点数计算的性能优势在其他性能损耗面前完全入不敷出。这就是所谓的从入门到放弃……吗？

给这篇文章的作者打赏

微信扫一扫打赏 支付宝扫一扫打赏
×
打赏