5 Replies Latest reply on Jan 14, 2009 7:47 PM by MicahVillmow

    Two questions

    ryta1203
      1. Is there going to be more documentation regarding how to improve performance in the kernels, either CAL or Brook+?

      2. I'm having trouble getting any of the time.h structures to work in a .br file, anyone else have this problem, if so, how did you get it to work?? I'm interested in getting timing data without all the streamRead time included (since they are done once compared to 87000 times for each kernel). Being able to run it 1000 times and get good timing would be very helpful, thanks.
        • Two questions
          MicahVillmow
          1) This is currently being worked on by myself and our documentation person and we are targeting it for the next release.
            • Two questions
              ryta1203
              Originally posted by: MicahVillmow

              1) This is currently being worked on by myself and our documentation person and we are targeting it for the next release.


              Micah, thanks. I'm assuming since you are a self-named CAL person that this will be for CAL?

              Also, do you know anything about my second question?
            • Two questions
              MicahVillmow
              In the old runtime the way the samples were timing was understood as being correct, but I'll have to defer in the new runtime as I have not used it.

              Also, the doc is being targeted for the compute shader programming model, not specific for CAL.
                • Two questions
                  ryta1203
                  1. Good to hear about the docs.

                  2. I just want to use some old fashion time.h structures to get the timing, but unfortunately, it appears that the Brook+ compiler doesn't like it.

                  EDIT: Also, is it a good assumption to assume that the larger the ALU:Tex ratio the better and that if the Bottleneck is ALU Ops, that's a good thing too?
                • Two questions
                  MicahVillmow
                  Yes, the larger the ALU:Tex ratio can help, but the performance metrics are all based on heuristics for graphics and don't necessarily map to the compute world directly.

                  Another ratio that is important is ALU:CF where you want about a 10:1 ratio here.