I was able to get a AmiBroker dll to work with Nvidia CUDA drivers. These drivers allow C code to run on the graphics shares of a modern video card. These are the same processors that allow high speed 3d graphics. Several math intensive applications report a 50-100 fold performance improvement over running on the host cpu.
The mid range card that came on my system has 64 cores, each able to perform one floating point operation per clock. As a simple test, I wrote a AmiBroker plug in, called by AFL. It calculated the average price (H+L+C)/3 for 60464 bars in 21us. This works out to about 8.5GF (billion floating point operations per second) and 46GB/s memory transfer speed. (read 3 floats and write one per bar), (2 floating point adds and 1 multiply per bar) The 46GB/s transfer rate is not far from the available memory bandwidth on the card, but the simple test calculation is not very "dense" so, I should be able to get a much higher calculation rate once I move more of my code to the graphics cores. Several of the CUDA demos report > 150GF/s. Memory is the bottleneck of this simple test. I used one thread per bar. High end graphics cards are available now that would improve performance by another factor of 2 to 4. A few problems: * The above numbers do not include the time needed to copy the data from ami to the graphics card or copy the results back. This time is much greater than the calculation time in this simple test. * This is not a general AFL accelerator. My goal is to reduce my current 25s backtest time down to < 1s per pass. To do this, I will need to move the data set for all symbols to the graphics card once and make many passes over the data with different optimization values. Each CUDA thread will work on one symbol, rather than a thread per bar as in my first test. There is not much point in writing a CUDA routine to just execute directly from AFL code. There is too much overhead. In my application, the AFL code is a very small part of the total time for each backtest. Even if I reduced the time to zero, it would not reduce the time per pass very much. Also, the time needed to copy the price data on each pass would greatly reduce the benefit. As far as I can tell, the current Ami API does not allow injecting a externally generated trade list into the backtest, so I will need to perform the full backtest and fitness function calculation externally. I had no compatibility problems getting the CUDA api to run as a Ami plug in. Why go to the trouble? Using Fred's IO program would get much of the same benefit for less trouble, or I could wait until Ami finally supports multi cores, or finds other clever ways to reduce the per pass overhead. The real answer is that I just had to try it....
