search

taijusti / sleep_apnea

0 stars 0 forks source link

remove bottleneck on host #30

Closed taijusti closed 9 years ago

taijusti commented 9 years ago

adding additional instances of device and boosting the number of points still shows approximately a 2x slowdown over matlab. this suggests the bottle neck is host. i think the issue is the way we're collecting values for KKT violators. we sequentially add them into a BRAM. this, multiplied by the many iterations that the main SMO loop runs is slowing things down (i think). we need to somehow get rid of this. i think i'm going to change the SMO algorithm some more

taijusti commented 9 years ago

system now is approximately on par with matlab on an i7. at 2048 points, with 2 devices, it finishes in approx 1 second. trying with 8192 points, 4 instances of device. closing