It still blows my mind that we can stream video at 60FPS. I was making an animation app that did frame-by-frame playback and 16.6ms goes by fast! Just unpacking a frame into memory and copying it to the GPU seemed like it took a while.
> 16.6ms
That's sixteen million nanoseconds, you should be able to issue thirty million instructions in that time from an ordinary 2GHz CPU. A GPU will give you several billion. Just don't waste them.
Instead create a GPU texture that’s backed by a fixed buffer in main memory. Decode into that buffer, unlock it, and draw using the texture. The GPU will do direct memory access over PCIe, avoiding the copy.
The CPU can’t be writing into the buffer while the GPU may be reading from it, so you can either use locks or double buffering to synchronize access.
With a dedicated GPU with its own memory there still is usually a memory to memory copy, it just doesn’t have to involve the CPU.
It definitely used to be that GPUs did real DMA texture reads though, at least in the early days of dedicated GPUs with fairly little local RAM. I'm thinking back to when the Mac OS X accelerated window compositor was introduced — the graphics RAM simply wouldn't have been enough to hold more than a handful of window buffers.
Frame A is copied to the GPU. Frame B is being decoded into
Frame B is copied to the GPU. Frame C is being decoded into
Frame C is copied to the GPU. Frame A is being decoded into