diapolo

profiling for read and write buffers

Discussion created by diapolo on May 3, 2010

Was it broken before 2.1?

Look at this results:

2.10:
up: 0.000000 s / exec: 0.054278 s / down: 0.000000s / 309095084
56 / up: 0.000022s / exec: 3.003121 s / down: 0.000022s / 313044173

2.0.1:
up: 0.087022 s / exec: 0.025523 s / down: 0.000044s / 149012759
56 / up: 4.966646s / exec: 1.429298 s / down: 0.002827s / 147135768

 

56 is the number of passes

up is the time it takes to load 256MB data into a kernel

down is the time it takes to read an int4 into my host program

exec is the kernel execution time (16777216 work-items)

 

It seems that kernel execution was speed up 50% for me, but because I take up- and download time into account for the messurement of results/s, the result is 50% worse. But yeah it seems, that 256MB of data and 0.000000s for upload with 2.0.1 were pretty wrong .

Dia

Outcomes