经过上一回的努力，我们终于将 GPU 计算的时间缩减到同 CPU 一个数量级，但是发现内存和显存之间的 memcpy 成了最主要的性能损耗。
一、固定内存
cudaMallocHostconst size_t NTB = 256;
const size_t EXT = 8;
#define divCeil(a, b) (((a) + (b) - 1) / (b))
struct Ctx {
float *x, *y;
float *xd, *yd, *rd;
size_t n;
};
extern "C" __declspec(dllexport) void init(Ctx **p, size_t n) {
Ctx *ctx = (Ctx *)malloc(sizeof(Ctx));
ctx->n = n;
size_t sz = sizeof(float) * n;
cudaMallocHost(&(ctx->x), sz);
cudaMallocHost(&(ctx->y), sz);
cudaMalloc(&(ctx->xd), sz);
cudaMalloc(&(ctx->yd), sz);
cudaMallocManaged(&(ctx->rd), sizeof(float) * divCeil(n, NTB) / EXT);
*p = ctx;
}
extern "C" __declspec(dllexport) void getInputs(Ctx *ctx, float **px, float **py) {
*px = ctx->x;
*py = ctx->y;
}
extern "C" __declspec(dllexport) void deinit(Ctx *ctx) {
cudaFreeHost(ctx->x);
cudaFreeHost(ctx->y);
cudaFree(ctx->xd);
cudaFree(ctx->yd);
cudaFree(ctx->rd);
free(ctx);
}
__global__ void devDot(float *x, float *y, size_t n, float *r) {
__shared__ float rb[NTB];
size_t itb = threadIdx.x;
size_t i = blockIdx.x * blockDim.x * EXT + itb;
float s = 0.0;
for (size_t j = 0; j < EXT && i < n; j++, i += blockDim.x) {
s += x[i] * y[i];
}
rb[itb] = s;
__syncthreads();
for (size_t i = NTB >> 1; i != 0; i >>= 1) {
if (itb < i) rb[itb] += rb[itb + i];
__syncthreads();
}
if (0 == itb) r[blockIdx.x] = rb[0];
}
extern "C" __declspec(dllexport) void dot(Ctx *ctx, float *r) {
size_t sz = sizeof(float) * ctx->n;
cudaMemcpy(ctx->xd, ctx->x, sz, cudaMemcpyHostToDevice);
cudaMemcpy(ctx->yd, ctx->y, sz, cudaMemcpyHostToDevice);
size_t nb = divCeil(ctx->n, NTB) / EXT;
float *rd = ctx->rd;
devDot<<<nb, NTB>>>(ctx->xd, ctx->yd, ctx->n, rd);
cudaDeviceSynchronize();
float s = 0.0;
for (size_t i = 0; i < nb; i++) s += rd[i];
*r = s;
}
package main
import (
"math/rand"
"reflect"
"syscall"
"time"
"unsafe"
)
const N = 1 << 20
type Lib struct {
dll *syscall.DLL
deinitProc *syscall.Proc
dotProc *syscall.Proc
handler uintptr
X, Y []float32
}
func LoadLib() (*Lib, error) {
l := &Lib{}
var err error
defer func() {
if nil != err {
l.Release()
}
}()
if l.dll, err = syscall.LoadDLL("cuda.dll"); nil != err {
return nil, err
}
if l.deinitProc, err = l.dll.FindProc("deinit"); nil != err {
return nil, err
}
if l.dotProc, err = l.dll.FindProc("dot"); nil != err {
return nil, err
}
proc, err := l.dll.FindProc("init")
if nil != err {
return nil, err
}
proc.Call(uintptr(unsafe.Pointer(&l.handler)), uintptr(N))
proc, err = l.dll.FindProc("getInputs")
if nil != err {
return nil, err
}
xh := (*reflect.SliceHeader)(unsafe.Pointer(&l.X))
yh := (*reflect.SliceHeader)(unsafe.Pointer(&l.Y))
xh.Len, xh.Cap, yh.Len, yh.Cap = N, N, N, N
proc.Call(l.handler,
uintptr(unsafe.Pointer(&xh.Data)), uintptr(unsafe.Pointer(&yh.Data)))
return l, nil
}
func (l *Lib) Release() {
if nil != l.deinitProc && 0 != l.handler {
l.deinitProc.Call(l.handler)
}
if nil != l.dll {
l.dll.Release()
}
}
func (l *Lib) Dot() float32 {
var r float32
l.dotProc.Call(l.handler, uintptr(unsafe.Pointer(&r)))
return r
}
func main() {
lib, err := LoadLib()
if nil != err {
println(err.Error())
return
}
defer lib.Release()
rand.Seed(time.Now().Unix())
x, y := lib.X, lib.Y
for i := 0; i < N; i++ {
x[i], y[i] = rand.Float32(), rand.Float32()
}
t := time.Now()
var r float32
for i := 0; i < 100; i++ {
r = 0
for i := 0; i < N; i++ {
r += x[i] * y[i]
}
}
println(time.Now().Sub(t).Microseconds())
println(r)
t = time.Now()
for i := 0; i < 100; i++ {
r = lib.Dot()
}
println(time.Now().Sub(t).Microseconds())
println(r)
}
cudaMemcpy二、内存映射
cudaHostAllocc 端有改动的代码如下：
struct Ctx {
float *x, *y, *r;
size_t n;
};
extern "C" __declspec(dllexport) void init(Ctx **p, size_t n) {
Ctx *ctx = (Ctx *)malloc(sizeof(Ctx));
ctx->n = n;
size_t sz = sizeof(float) * n;
cudaHostAlloc(&(ctx->x), sz, cudaHostAllocMapped);
cudaHostAlloc(&(ctx->y), sz, cudaHostAllocMapped);
cudaMallocManaged(&(ctx->r), sizeof(float) * divCeil(n, NTB) / EXT);
*p = ctx;
}
extern "C" __declspec(dllexport) void deinit(Ctx *ctx) {
cudaFreeHost(ctx->x);
cudaFreeHost(ctx->y);
cudaFree(ctx->r);
free(ctx);
}
extern "C" __declspec(dllexport) void dot(Ctx *ctx, float *r) {
size_t nb = divCeil(ctx->n, NTB) / EXT;
float *rd = ctx->r;
devDot<<<nb, NTB>>>(ctx->x, ctx->y, ctx->n, rd);
cudaDeviceSynchronize();
float s = 0.0;
for (size_t i = 0; i < nb; i++) s += rd[i];
*r = s;
}
cudaMemcpydevDot三、统一寻址
cudaMallocManagedc 端有改动的代码如下：
extern "C" __declspec(dllexport) void init(Ctx **p, size_t n) {
Ctx *ctx = (Ctx *)malloc(sizeof(Ctx));
ctx->n = n;
size_t sz = sizeof(float) * n;
cudaMallocManaged(&(ctx->x), sz);
cudaMallocManaged(&(ctx->y), sz);
cudaMallocManaged(&(ctx->r), sizeof(float) * divCeil(n, NTB) / EXT);
*p = ctx;
}
extern "C" __declspec(dllexport) void deinit(Ctx *ctx) {
cudaFree(ctx->x);
cudaFree(ctx->y);
cudaFree(ctx->r);
free(ctx);
}
devDotcudaMemcpycudaLaunchKernel