cudpp_segmented_scan_cta.cu "Removed dead synchronization intrinsic" advisories about cudpp

[GoogleCodeExporter]

What do you know - CUDA 2.2 seems to have fixed this bug - so I will comment 
the offending syncthreads for 
now (only tested on linux though - I will test on windows as well)

Original comment by [email protected] on 25 Jun 2009 at 5:18

[GoogleCodeExporter]

Verified on Linux with CUDA 2.2

Original comment by [email protected] on 25 Jun 2009 at 5:22

• Changed state: Fixed

[GoogleCodeExporter]

Verified on Windows XP with CUDA 2.2

Original comment by [email protected] on 25 Jun 2009 at 3:21

[GoogleCodeExporter]

Reverted since behavior much more stable with the syncthreads

Original comment by [email protected] on 25 Jun 2009 at 7:00

• Changed state: WontFix

[GoogleCodeExporter]

FIne, I'll fix it.  I explained how to fix it in the description, I'll try it 
myself.  But we aren't marking this "wont fix".

Original comment by [email protected] on 25 Jun 2009 at 9:38

• Changed state: Started

[GoogleCodeExporter]

I tried your approach and it didn't work (that is putting the syncthreads 
before calls to warpSegScan)

Original comment by [email protected] on 25 Jun 2009 at 9:41

[GoogleCodeExporter]

This advisory is fixed in CUDA 2.3.  There is a bug in the compiler where 
moving the
__syncthreads() from the beginning of warpSegScan() to right before each call 
of it
causes the code to fail.  I filed an NVIDIA bug for this.

Closing this for now.

Original comment by [email protected] on 26 Jun 2009 at 8:03

• Changed state: WontFix

[GoogleCodeExporter]

Original comment by [email protected] on 26 Jun 2009 at 8:04

• Changed state: Verified

[GoogleCodeExporter]

I fixed this for CUDA 2.2.

Original comment by [email protected] on 30 Jun 2009 at 5:07