GPU Timeout Error

Question

Created May ’22

Replies 2

Boosts 0

Views 1.1k

Participants 1

Hi - I've been developing code to determine star shapes using the GPU.

The code passes a patch of intensity values near the star and then uses a Levenberg-Marquardt algorithm to get a least squares fit.

I can get the code to work well for up to five stars at a time provided I pass an array of size 15x15. However, if I pass more than 5 stars, I get a GPU timeout error.

If I change the size of the array to 16x16, or 14x14 then even one star will cause a GPU timeout error.

The Levenberg-Marquardt algorithm does use lots of if statements - about 10 per loop, and the loop is executed multiple times. Not sure if there is a limit to the number of if statements allowed?

I use one threadgroup with a single thread per star.

Is there anyway to debug this to work out what causes the problem? In one run I did get a slightly different error:

Stack Overflow Exception. Please check the [MTLComputePipelineDescriptor maxCallStackDepth] setting.

The subroutine does call routines which call others, so that may be an issues - I think the call depth is max 2.

Any other thoughts gratefully received.

Colin

Boost

Answer 1

mcgillca OP

May ’22

OK - I've recoded this to take out the max call stack depth problem (no more than 1 call depth), but there is still an issue.

0

Answer 2

mcgillca OP

Jun ’22

I've finally figured this out - I was using a do while loop to control the iterations - I switched this to a for loop with a break statement - this solved all the issues.

Not sure exactly what is the reason, but there must be something that metal does with a do while loop that uses more resources.

0