The most efficient way to modify CMSampleBuffer contents

I want开发者_StackOverflow社区 to modify the contents of CMSampleBuffer and then write it to a file with AVAssetWriter/AVAssetWriterInput.

The way I do this is by creating a Core Graphics bitmap context and then drawing into it, but it's way too slow. Specifically, I need to draw an image into the buffer.

So can one provide some sort of a hint or suggestion on how to do it more efficiently?

I thought about using OpenGL to accomplish that, that is, first create a texture A from CMSampleBuffer. Then render texture B, created from image I want to draw, into texture A, then retrieve the data backing the texture A from OpenGL and finally hand that data over to AVAssetWriter/AVAssetWriterInput. But the docs say that transferring texture data from GPU back to CPU is kinda expensive.

So, any suggestions on how to approach that?

Thanks in advance

OpenGL is probably the way to go. It might be slightly more efficient to render to an offscreen framebuffer rather than a texture, though.

To extract a texture from a sample buffer:

// Note the caller is responsible for calling glDeleteTextures on the return value.
- (GLuint)textureFromSampleBuffer:(CMSampleBufferRef)sampleBuffer {
    GLuint texture = 0;

    glGenTextures(1, &texture);
    glBindTexture(GL_TEXTURE_2D, texture);
    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);

    CVImageBufferRef pixelBuffer = CMSampleBufferGetImageBuffer(sampleBuffer);
    CVPixelBufferLockBaseAddress(pixelBuffer, 0);
    int width = CVPixelBufferGetWidth(pixelBuffer);
    int height = CVPixelBufferGetHeight(pixelBuffer);
    glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, width, height, 0, GL_BGRA, GL_UNSIGNED_BYTE, CVPixelBufferGetBaseAddress(pixelBuffer));
    CVPixelBufferUnlockBaseAddress(pixelBuffer, 0);

    return texture;
}

To process the texture through OpenGL, you could then do something like this:

// This function exists to free the malloced data when the CGDataProviderRef is
// eventually freed.
void dataProviderFreeData(void *info, const void *data, size_t size){
    free((void *)data);
}

// Returns an autoreleased CGImageRef.
- (CGImageRef)processTexture:(GLuint)texture width:(int)width height:(int)height {
    CGImageRef newImage = NULL;

    // Set up framebuffer and renderbuffer.
    GLuint framebuffer;
    glGenFramebuffers(1, &framebuffer);
    glBindFramebuffer(GL_FRAMEBUFFER, framebuffer);

    GLuint colorRenderbuffer;
    glGenRenderbuffers(1, &colorRenderbuffer);
    glBindRenderbuffer(GL_RENDERBUFFER, colorRenderbuffer);
    glRenderbufferStorage(GL_RENDERBUFFER, GL_RGBA8_OES, width, height);
    glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, colorRenderbuffer);

    GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER);
    if (status != GL_FRAMEBUFFER_COMPLETE) {
        NSLog(@"Failed to create OpenGL frame buffer: %x", status);
    } else {
        glViewport(0, 0, width, height);
        glClearColor(0.0,0.0,0.0,1.0);
        glClear(GL_COLOR_BUFFER_BIT);

        // Do whatever is necessary to actually draw the texture to the framebuffer
        [self renderTextureToCurrentFrameBuffer:texture];

        // Read the pixels out of the framebuffer
        void *data = malloc(width * height * 4);
        glReadPixels(0, 0, width, height, GL_RGBA, GL_UNSIGNED_BYTE, data);

        // Convert the data to a CGImageRef. Note that CGDataProviderRef takes
        // ownership of our malloced data buffer, and the CGImageRef internally
        // retains the CGDataProviderRef. Hence the callback above, to free the data
        // buffer when the provider is finally released.
        CGDataProviderRef dataProvider = CGDataProviderCreateWithData(NULL, data, width * height * 4, dataProviderFreeData);
        CGColorSpaceRef colorspace = CGColorSpaceCreateDeviceRGB();
        newImage = CGImageCreate(width, height, 8, 32, width*4, colorspace, kCGBitmapByteOrder32Big | kCGImageAlphaPremultipliedLast, dataProvider, NULL, true, kCGRenderingIntentDefault);
        CFRelease(dataProvider);
        CGColorSpaceRelease(colorspace);

        // Autorelease the CGImageRef
        newImage = (CGImageRef)[NSMakeCollectable(newImage) autorelease];
    }

    // Clean up the framebuffer and renderbuffer.
    glDeleteRenderbuffers(1, &colorRenderbuffer);
    glDeleteFramebuffers(1, &framebuffer);

    return newImage;
}