CUDA / OpenGL互操作，用CUDA绘制到OpenGL纹理

问题描述

我在CUDA中编写一个渲染系统，希望通过OpenGL快速显示结果，而不需要触摸主存储器。我基本上做以下：

I am writing a rendering system in CUDA and want results to be quickly displayed via OpenGL, without touching main memory. I basically do the following:

创建和初始化OpenGL纹理，并在CUDA中注册为cudaGraphicsResource

Create and initialize OpenGL texture, and register it in CUDA as cudaGraphicsResource

GLuint viewGLTexture;
cudaGraphicsResource_t viewCudaResource;

void initialize() {
    glEnable(GL_TEXTURE_2D);
    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, view.getWidth(), view.getHeight(), 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
    }
    glBindTexture(GL_TEXTURE_2D, 0);

    cudaGraphicsGLRegisterImage(&viewCudaResource, viewGLTexture, GL_TEXTURE_2D, cudaGraphicsRegisterFlagsWriteDiscard)
}

$ b b

每当视图调整大小时，我会适当调整视口和纹理图像的大小：

Whenever view is resized I resize viewport and texture image appropriately:

void resize() {
    glViewport(0, 0, view.getWidth(), view.getHeight());

    glBindTexture(GL_TEXTURE_2D, viewGLTexture);
    {
        glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, view.getWidth(), view.getHeight(), 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
    }
    glBindTexture(GL_TEXTURE_2D, 0);
}

然后每个框架通过cudaArray将graphicsResource映射为cudaSurfaceObject，在其上，取消映射和同步，让OpenGL使用此纹理绘制一个全屏四边形：

And then each frame I map graphicsResource as a cudaSurfaceObject via cudaArray, call rendering kernel on it, unmap and synchronize to let OpenGL draw a fullscreen quad with this texture:

void renderFrame() {
    cudaGraphicsMapResources(1, &viewCudaResource);
    {
        cudaArray_t viewCudaArray;
        cudaGraphicsSubResourceGetMappedArray(&viewCudaArray, viewCudaResource, 0, 0);
        cudaResourceDesc viewCudaArrayResourceDesc;
        {
            viewCudaArrayResourceDesc.resType = cudaResourceTypeArray;
            viewCudaArrayResourceDesc.res.array.array = viewCudaArray;
        }
        cudaSurfaceObject_t viewCudaSurfaceObject;
        cudaCreateSurfaceObject(&viewCudaSurfaceObject, &viewCudaArrayResourceDesc);
        {
            invokeRenderingKernel(viewCudaSurfaceObject);
        }
        cudaDestroySurfaceObject(viewCudaSurfaceObject));
    }
    cudaGraphicsUnmapResources(1, &viewCudaResource);

    cudaStreamSynchronize(0);

    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();
}

问题是：每当视图调整大小时，所有CUDA调用都开始抛出未知错误和视觉上看起来像纹理实际上没有调整大小，只是伸展在整个视图。为什么会发生这种情况，如何解决？

The problem is: Whenever view is resized all CUDA calls start spewing out "unknown error"s and visually it looks like the texture is not in fact resized, just stretched across the whole view. Why is this happening and how do I fix it?

推荐答案

看起来interop需要在调整大小时重新注册纹理。以下工作：

It seems interop requires to re-register textures upon resize. The following works:

void resize() {
    glViewport(0, 0, view.getWidth(), view.getHeight());

        // unregister
    cudaGraphicsUnregisterResource(viewCudaResource);
        // resize
    glBindTexture(GL_TEXTURE_2D, viewGLTexture);
    {
        glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, view.getWidth(), view.getHeight(), 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
    }
    glBindTexture(GL_TEXTURE_2D, 0);
        // register back
    cudaGraphicsGLRegisterImage(&viewCudaResource, viewGLTexture, GL_TEXTURE_2D, cudaGraphicsRegisterFlagsWriteDiscard);
}

这篇关于CUDA / OpenGL互操作，用CUDA绘制到OpenGL纹理的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持！