baskett@SGI.COM (11/01/88)
In response to several posting about pixel transfer rates on Iris workstations, I ran some experiments on our latest products. Pixel access on the latest series of IRIS products has been improved via new data paths to pixel memory and new library calls to do block mode transfers on those new data paths. The products involved are the GT, GTX, and the Personal Iris. Below is a table of results I obtained last week by running a simple C program on some of the machines in my building. The numbers below are not the best or the worst you can get and they are not the official specifications but they do give some idea of the kind of results you can achieve and they do show moderatly high pixel transfer rates. Rates are all to the nearest .1 megapixels per second. The block size I used was moderate: 100 x 100 pixels. GL routine GT GTX Personal Iris rectwrite 1.3 8.0 0.9 lrectwrite 1.0 5.7 4.4 rectread 0.6 3.0 0.6 lrectread 0.6 2.0 2.5 rectwrite transfers a vector of 16 bit numbers from main memory to a rectangle of 12 bit color index mode pixels in frame buffer memory. lrectwrite transfers a vector of 32 bit numbers from main memory to a rectangle of 24 bit RGB mode pixels in frame buffer memory. rectread and lrectread are the corresponding inverse transfers. There is also rectcopy which will move a rectangle of pixels from one area of the screen to another at high speed. Users of the Personal Iris might want to note that color index pixels can be transfered by lrectwrite and lrectread for the best transfer rates. Users of the GTX should note that the transfer rates above are with the host memory pages locked in main memory with the mpin system function to avoid per page overhead that reduces the transfer rates. The transfer rates on the Personal Iris will be even better on future software and microcode releases of this very new product. Hope this helps. Forest Baskett