I have an array of 8192 floats. Threads in the same warp will read the same float. In CUDA I would have used constant memory but its not clear if the same thing is possible using constant buffers in DirectCompute - there seems to be some strangeness concerning four-vectors and a maximum number of variables per constant buffer. Should I even be using constant buffers or would a raw or structured buffer be just as fast? I’m targetting Fermi hardware.
I have an array of 8192 floats. Threads in the same warp will read the same float. In CUDA I would have used constant memory but its not clear if the same thing is possible using constant buffers in DirectCompute - there seems to be some strangeness concerning four-vectors and a maximum number of variables per constant buffer. Should I even be using constant buffers or would a raw or structured buffer be just as fast? I’m targetting Fermi hardware.