jarrod said:
Pana:
Okay... but what is that 33M Vert/sec reflective of in PSP then? Is it similarly peak performance, limited screnario or reflective of real world applictation results? Is it fair to say PSP is roughly half as powerful as PS2 when it comes to 3D goemetry potential?
The PSP GPU's T&L is hardwired: going by the clock-speed and the quoted number...
166 MHz, 33.2 MVertices/s gives you 5 cycles per Vertex which is about the same number of cycles Flipper takes for each Vertex.
That figure should include basic transform with perspective projection and lighting (4 parallel lights processed by the Lighting unit).
Calculation of ST coordinates (the set-up the T&L units does) consist of only a simple MUL: you already got Q ( which is 1/w ) for perspective projection and you only really need to take the ST coordinates (stored ina register) and multiply them by Q.
In games people want multiple textures, point-lights, etc... high-resolution textures, multiple rendering-passes to do certain effects, you have to help the GPU in doing clipping (the GPU onlu does front-plane clipping, the other plane are not really taken care of... very brute-force) and the T&L unit will not always run full-speed (sometimes, making it work with the VFPU will waste a bit of time moving data around... time in which you lose efficiency).
You have quite a lot of processing performance in the GPU, but you have lost the flexibility you had in VU1 when doing T&L (bye-bye branching in T&L code, bye-bye Vertex creation and deletion in the T&L code, etc...) on the GPU (you have the VFPU on the CPU side to help to do custom T&L... just like you do on the GCN with Gekko and Flipper).
Look at the Elan T&L chip: only 100 MHz and 10 MTransistors and it pushes (it sustains) 10 MVertices/s with 6 point-lights (very computationally expensive) plus doing set-up work for DOT3 calculations, etc...