如何在dll中使用CUDA

Question

我的目的是通过存根库使用CUDA，存根库封装了CUDA特定的函数、方法和数据。CUDA内容将驻留在dll文件中，并将通过LoadLibrary和GetProcAddress Windows API函数动态使用。我使用Visual Studio 2010 C++编译器创建CUDA dll，但其余部分由另一个编译器完成。这意味着，我不能在CUDA dll中使用堆内存(即malloc、new或任何位于堆栈内存之外的东西。即使是全局变量也会导致内存损坏)，所以它实际上是一个存根库。

然而，首先我想编写一个测试程序：CUDAhost.exe和CUDAdevice.dll，它们都是由Visual Studio 2010编译的，这两个项目在一个解决方案中。该程序在屏幕上绘制OpenGL纹理，但首先通过CUDA将图像数据从readGLTexture复制到viewGLTexture。注意，为了避免使用堆内存，我使用了void指针引用void*&cReadCudaResource和void*&cViewCudaResource。我的问题是我无法使程序工作，窗口是黑色的。我找不到错误。我不确定这是否可能，或者我应该选择完全不同的解决方案。我希望你能帮助我。任何建议都将不胜感激。以下是源代码：

CUDAhost.cpp:

#include "stdafx.h"
const unsigned int window_width  = 512;
const unsigned int window_height = 512;
GLuint viewGLTexture;
GLuint readGLTexture;
void* cViewCudaResource;
void* cReadCudaResource;
HINSTANCE dll;
typedef void (*SETCUDA)(unsigned int& readGLTexture, void*   &cReadCudaResource, unsigned int& viewGLTexture, void* &cViewCudaResource);
SETCUDA setCuda;
typedef void (*DRAWPICTURE)(void* &cReadCudaResource, void* &cViewCudaResource);
DRAWPICTURE drawPicture;
bool loadTexture(const wchar_t* name, GLuint& number) {
FILE* file;
BITMAPFILEHEADER bitmapFileHeader;
BITMAPINFOHEADER bitmapInfoHeader;
unsigned char *bitmap;
unsigned char temp;
wchar_t path[45]={0};
int width;
int height;
//prepare file path
wcsncat_s(path, L"Textures\", 45);
wcsncat_s(path, name, 45);
wcsncat_s(path, L".bmp", 45);
//open BMP file
file=_wfopen(path, L"rb");
if (file==NULL) {
return false;
}
//read bmp file header and sequre it is bmp file
fread(&bitmapFileHeader, sizeof(BITMAPFILEHEADER), 1, file);
if (bitmapFileHeader.bfType != 0x4D42) {
fclose(file);
return false;
}
//read bmp info header and move to the beginning of image data
fread(&bitmapInfoHeader, sizeof(BITMAPINFOHEADER), 1, file);
fseek(file, bitmapFileHeader.bfOffBits, SEEK_SET);
//allocate memory space
bitmap=(unsigned char*)malloc(bitmapInfoHeader.biSizeImage);
if (!bitmap) {
free(bitmap);
bitmap=NULL;
fclose(file);
return false;
}
//read image
fread(bitmap, 1, bitmapInfoHeader.biSizeImage, file);
if (file==NULL) {
free(bitmap);
bitmap=NULL;
fclose(file);
return false;
}
//rearrange bgr to rgb
for (int i=0; i<bitmapInfoHeader.biSizeImage; i+=3) {
temp=bitmap[i];
bitmap[i]=bitmap[i+2];
bitmap[i+2]=temp;
}
//query image width and height
width=bitmapInfoHeader.biWidth;
height=abs(bitmapInfoHeader.biHeight);
//close bmp file
fclose(file);
glGetError();
//create OpenGL texture
glGenTextures(1, &number);
glBindTexture(GL_TEXTURE_2D, number);
glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_RGB, GL_UNSIGNED_BYTE, bitmap);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
//free temporary buffer
free(bitmap);
bitmap=NULL;
//if success, return true
if (0==glGetError()) {
return true;
} else {
return false;
}
}
void initGLandCUDA(int argc, char* argv[]) {
glutInit(&argc, argv);
glutInitDisplayMode(GLUT_RGBA);
glutInitWindowSize(window_width, window_height);
glutCreateWindow("CUDA GL Interop");
glewInit();
glEnable(GL_TEXTURE_2D);
bool success=loadTexture(L"Tex", readGLTexture);
glGenTextures(1, &viewGLTexture);
glBindTexture(GL_TEXTURE_2D, viewGLTexture);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, window_width, window_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
glBindTexture(GL_TEXTURE_2D, 0);
dll=LoadLibraryW(L"CUDAdevice.dll");
if (dll) {
setCuda=(SETCUDA)GetProcAddress(dll, "setCuda");
setCuda(readGLTexture, cReadCudaResource, viewGLTexture, cViewCudaResource);
}
}    
void renderFrame() {
if (dll) {
drawPicture=(DRAWPICTURE)GetProcAddress(dll, "drawPicture");
drawPicture(cReadCudaResource, cViewCudaResource);
}
glBindTexture(GL_TEXTURE_2D, viewGLTexture);
{
glBegin(GL_QUADS);
{
glTexCoord2f(0.0f, 0.0f); glVertex2f(-1.0f, -1.0f);
glTexCoord2f(1.0f, 0.0f); glVertex2f(+1.0f, -1.0f);
glTexCoord2f(1.0f, 1.0f); glVertex2f(+1.0f, +1.0f);
glTexCoord2f(0.0f, 1.0f); glVertex2f(-1.0f, +1.0f);
}
glEnd();
}
glBindTexture(GL_TEXTURE_2D, 0);
glFinish();
}
int _tmain(int argc, _TCHAR* argv[])
{
initGLandCUDA(argc, reinterpret_cast<char**>(argv));
glutDisplayFunc(renderFrame);
glutMainLoop();
return 0;
}

dllmain.cpp:

BOOL APIENTRY DllMain( HMODULE hModule,
DWORD  ul_reason_for_call,
LPVOID lpReserved
)
{
switch (ul_reason_for_call)
{
case DLL_PROCESS_ATTACH:
case DLL_THREAD_ATTACH:
case DLL_THREAD_DETACH:
case DLL_PROCESS_DETACH:
break;
}
return TRUE;
}
//this function is used to setup CUDA
void setCuda(unsigned int& readGLTexture, void* &cReadCudaResource, unsigned int& viewGLTexture, void* &cViewCudaResource) {
struct cudaGraphicsResource* viewCudaResource;
struct cudaGraphicsResource* readCudaResource;
cudaError cError;
cudaGLSetGLDevice(0);
cError=cudaGraphicsGLRegisterImage(&viewCudaResource, viewGLTexture, GL_TEXTURE_2D, cudaGraphicsRegisterFlagsReadOnly);
cError=cudaGraphicsGLRegisterImage(&readCudaResource, readGLTexture, GL_TEXTURE_2D, cudaGraphicsRegisterFlagsSurfaceLoadStore);
cReadCudaResource=reinterpret_cast<void*>(readCudaResource);
cViewCudaResource=reinterpret_cast<void*>(viewCudaResource);
}    
//this function is used to draw texture image via CUDA
void drawPicture(void* &cReadCudaResource, void* &cViewCudaResource) {
cudaError cError;
struct cudaGraphicsResource* viewCudaResource=reinterpret_cast<cudaGraphicsResource*>(cReadCudaResource);
struct cudaGraphicsResource* readCudaResource=reinterpret_cast<cudaGraphicsResource*>(cViewCudaResource);
cudaArray *readCudaArray;
cudaArray *viewCudaArray;
cError=cudaGraphicsMapResources(1, &readCudaResource);
cError=cudaGraphicsMapResources(1, &viewCudaResource);
cError=cudaGraphicsSubResourceGetMappedArray(&readCudaArray, readCudaResource, 0, 0);
cError=cudaGraphicsSubResourceGetMappedArray(&viewCudaArray, viewCudaResource, 0, 0);
callCUDAKernel(readCudaArray, viewCudaArray);
cudaGraphicsUnmapResources(1, &viewCudaResource);
cudaStreamSynchronize(0);
}

kernels.cu:

#include "stdafx.h"
texture<uchar4, cudaTextureType2D, cudaReadModeElementType> readCudaTextureObject;
surface<void, cudaSurfaceType2D> viewCudaSurfaceObject;

__global__ void renderingKernel() {
unsigned int x = blockIdx.x * blockDim.x + threadIdx.x;
unsigned int y = blockIdx.y * blockDim.y + threadIdx.y;
uchar4 dd=tex2D(readCudaTextureObject, x, y);
surf2Dwrite(dd, viewCudaSurfaceObject, x*sizeof(dd), y, cudaBoundaryModeZero);
}

void callCUDAKernel(cudaArray *readCudaArray, cudaArray *viewCudaArray) {
cudaError cError;
cError=cudaBindTextureToArray(readCudaTextureObject, readCudaArray);
cError=cudaBindSurfaceToArray(viewCudaSurfaceObject, viewCudaArray);
dim3 block(256, 1, 1);
dim3 grid(2, 512, 1);
renderingKernel<<<grid, block>>>();
cudaPeekAtLastError();
cudaDeviceSynchronize();
}

CUDA设备的stdafx.h:

#pragma once
#include "targetver.h"
#define WIN32_LEAN_AND_MEAN             // Exclude rarely-used stuff from Windows headers
// Windows Header Files:
#include <windows.h>

// TODO: reference additional headers your program requires here
#include <cuda_runtime_api.h>
#include <cuda_gl_interop.h>
#include "kernels.h"
#if defined (__cplusplus)
extern "C"
{
#endif
__declspec(dllexport) void setCuda(unsigned int& readGLTexture, void* &cReadCudaResource, unsigned int& viewGLTexture, void* &cViewCudaResource);
__declspec(dllexport) void drawPicture(void* &cReadCudaResource, void* &cViewCudaResource);
#if defined (__cplusplus)
}
#endif

kernels.h:

#ifndef __kernels_H
#define __kernels_H
void callCUDAKernel(cudaArray *readCudaArray, cudaArray *viewCudaArray);
#endif

PS。我还设置了CUDA库、头文件、源文件和二进制文件的路径，向其他依赖项添加了cudart.lib，设置了CUDA4.2目标和compute_20，sm_21。测试程序使用GLEW和GLUT库。

Answer 1

如果有人读到这篇文章，最好在这里更新解决方案。。。

以上代码无法工作。整个想法是为了避免在DLL模块中使用堆内存。但是，CUDA需要纹理和表面是全局的：

texture<uchar4, cudaTextureType2D, cudaReadModeElementType>readCudaTextureObject;
surface<void, cudaSurfaceType2D> viewCudaSurfaceObject;

然而，为了避免使用堆内存，需要这些内存是本地的，所以这是不可能的。

此外，根据我的理解，在CUDA DLL模块中使用堆内存应该没有任何问题。我已经制作了可工作的EXE/DLL测试程序，这些程序使用DLL模块内的堆内存。我在以前的尝试中没有发现这个错误，但它可能是由于GLEW库造成的。我的工作程序没有使用GLEW。