AMDave
07-04-2014, 12:54 AM
Astropulse 7,00 released for Linux 32&64, Win 32&64, Win32+AMD/NVIDIA/Intel GPU
For now in offline testing it survived -ffa_block 512 -ffa_ffa_block_fetch 256 and crashed driver on -ffa_block 512 -ffa_ffa_block_fetch 512
So i think runtime error reporting is misplaced and real cause is fetch kernel (longest kernel in whole app, btw).
Hence, quite possible that higher -ffa_block values will go to, only -ffa_block_fetch should be severely limited (to 256 in my case).
Will check this in next few tests.
Also it seems time to implement some default values scaling based on device capabilities for 7.01
It's not uncommon to have 2 same vendor devices with vastly different capabilities (almost any upgrade with keeping old old card plugged will lead to such situation).
EDIT: it survived -ffa_block 1024 -ffa_ffa_block_fetch 256 too so I think recommended workaround would be:
add -ffa_block 1024 -ffa_block_fetch 256 to ap_cmdline_7.00_windows_intelx86__opencl_nvidia_10 0.txt file
[-ffa_block required to allow desired block_fetch modification]
Will test live on loaded host now and TDR solution later.
EDIT2: can confirm it works and make progress with recommended above settings.
Next will be to check TDR solution (as more appropriate for fast GPU/slow GPU mix).
More... (http://setiweb.ssl.berkeley.edu/beta/forum_thread.php?id=2182)
For now in offline testing it survived -ffa_block 512 -ffa_ffa_block_fetch 256 and crashed driver on -ffa_block 512 -ffa_ffa_block_fetch 512
So i think runtime error reporting is misplaced and real cause is fetch kernel (longest kernel in whole app, btw).
Hence, quite possible that higher -ffa_block values will go to, only -ffa_block_fetch should be severely limited (to 256 in my case).
Will check this in next few tests.
Also it seems time to implement some default values scaling based on device capabilities for 7.01
It's not uncommon to have 2 same vendor devices with vastly different capabilities (almost any upgrade with keeping old old card plugged will lead to such situation).
EDIT: it survived -ffa_block 1024 -ffa_ffa_block_fetch 256 too so I think recommended workaround would be:
add -ffa_block 1024 -ffa_block_fetch 256 to ap_cmdline_7.00_windows_intelx86__opencl_nvidia_10 0.txt file
[-ffa_block required to allow desired block_fetch modification]
Will test live on loaded host now and TDR solution later.
EDIT2: can confirm it works and make progress with recommended above settings.
Next will be to check TDR solution (as more appropriate for fast GPU/slow GPU mix).
More... (http://setiweb.ssl.berkeley.edu/beta/forum_thread.php?id=2182)