12/03/2008: latency/ directory is updated to work with sdk3.0 and libspe2


note:	most code are outdated (pre-SDK 3.0). The ones I keep updating are /libtask and tasktest/
	If you find errors in the code, please send email to Guochun Shi <gshi@ncsa.uiuc.edu>.


bandwidth:  This directory contains code that measures the bandwidth between PPU and  SPUs, bandwidth can read ~24GB/s. For unknown reason
	    the bandwidth is not consistent, i.e. I may get 24 GB/s one day and get 15GB/s only the other day with the same code and the same machine.  

complex_matrix: 4x4 complex matrix implementation
	  
complex_matrix3x3: compute 3x3 complex matrix in the form of 4x4 complex matrix, i.e. there is bandwidth and compuation waste(9/16 = 56% efficiency)

complex_matrix3x3_native: compute 3x3 complex matrix in more efficient way, no bandwidth or computation power is wasted

dmalist: 	dmalist example

dotproduct: 	dotproduct example

dotproduct_v: 	another dot product example(see README in the directory)

eib_bandwidth:  measure the CELL bus bandwidth

event:   event(interrupt mailbox) sample code

latency: the latency is about 7~8 us roundtrip between ppu -- spu

libtask: a simple library which takes an task_t struct to do computation in spu, see tasktest/

tasktest:  sample code to use libtask library

matrix_mult:  This directory contains code that runs a big loop of 4x4 float value matrix multiplication
	       the speedup is ~150 with 8 SPUs compared with running the code in ppu only

specomm:	This directory contains code that demonstrate communication between SPUs including dma, mailbox and signal
	


FAQ:

1. What size of dma will work?

	dma requires size to be

	1, 2, 4, 8, 16, 32, ..., 16n

	 (n is an integer)