krik said:
Crumas, I never "played" with a Cray machine but bellow is a "hello world" application for a CELL SPE. The big difference you will probably notice is the code is in C. Yes, IBM as C/C++ compilers for the PPE. They even support standard libraries. Here is the code: SPE code: --- #include int main( unsigned long spuid ) { printf("Hello, World! (From SPU:%d)n",spuid); return (0); } PPE code to load the spe code (this code actually blocks waiting for the SPE to finish but it could be loaded and run using a ppe_thread that would not block the PPE): #include #include int main() { unsigned int createflags = 0; unsigned int runflags = 0; unsigned int entry = SPE_DEFAULT_ENTRY; void* argp = NULL; void* envp = NULL; spe_program_handle_t* program = spe_image_open("spu_hello"); spe_context_ptr_t spe = spe_context_create(createflags, NULL); spe_stop_info_t stop_info; spe_program_load(spe, program); spe_context_run(spe, &entry, runflags, argp, envp, &stop_info); spe_image_close(program); spe_context_destroy(spe); return (0); } But I agree with you that the biggest problem of using the PPEs is to fragment your data structures in such a way you can send the PPE as much data as possible (up to the 256Kb cache limit) so that the PPE can run without having to access main memory or any other I/O. This prevents racing conditions and also lock wait slowdowns maximizing the crunch power of the PPE. This is what really makes it hard. But, for example, if you send a full mesh and it's skeleton animation data to a SPE and have it deform the mesh based on the skeleton animation per vertex weights then the PPE will beat any core doing the same task. This task is basically just a loop that performs matrix multiplications on each vertex based on vertex weights in relation to each bone. I'm actually working on some C code to do this right now (as I learn PPE/SPE coding on Linux/PS3). |
Having been a C programmer since 1980 (doing real-time telecomm and kernel work back then), I can appreciate the clean nature of the API. The race condition I was speaking of wasn't actually related to what goes on in the SPE (it's really a very advanced version of a bit-blitter, which is why the Amiga was so great 
... the challenges are actually due to co-ordinating the work among separate SPEs at the application layer. Having a bunch of powerful critters able to tear through matrix transforms isn't going to do you much good if you can't keep them fed or manage to mesh what they produce as output effectively.
You and I are probably not far off concerning what we believe is theoretically possible with the Cell. But just like a PC with a blazing fast CPU, if memory bandwidth limits or other factors holds it to 10 wait states, what good does it do? You have to be able to take advantage of it. Folding at Home is great for that machine because it can really use the Cell effectively without requiring a programmer to jump through big flaming hoops to get it done.
I hope the PS3 survives so that 3 or 4 years from now a killer game can be produced that uses the full potential of the Cell in a way that can't be duplicated elsewhere. Do I believe this can happen? Yes. Do I believe it will take years to produce just like Shadow of the Colossus or GT4 on the PS2? Yes. It wasn't impossible to produce those games on the PS2... just really, really hard.
