OpenCL and OpenCL1K apps - what the difference?
27 Feb 2021 11:56:49 UTC
Topic 224928
(moderation:
I have hsgamma_FGRPB1G_1.22_windows_x86_64__FGRPopencl-nvidia.exe and hsgamma_FGRPB1G_1.22_windows_x86_64__FGRPopencl1K-nvidia.exe
apps on same host.
The file size just the same.
Also there are 2 tasks downloaded one for first one for second.
But on web page both tasks are just the same type "Gamma-ray pulsar binary search #1 on GPUs v1.22 () windows_x86_64"
So what the difference?
The K is a beta version but I
)
The K is a beta version but I don't know what is the actual difference.
I believe the "K" versions
)
I believe the "K" versions were just a test compile run in figuring out and optimizing the application for the TV (Turing-Volta) applications which were needed for those new gpus to be able to process tasks without erroring out.
FWIW --- Maybe this is the
)
2021-02-28 07:11:08.0557 [PID=15912] [version] Checking plan class 'FGRPopencl-ati'
2021-02-28 07:11:08.0581 [PID=15912] [version] reading plan classes from file '/BOINC/projects/EinsteinAtHome/plan_class_spec.xml'
2021-02-28 07:11:08.0581 [PID=15912] [version] parsed project prefs setting 'gpu_util_fgrp': 1.000000
2021-02-28 07:11:08.0581 [PID=15912] [version] No ATI devices found
2021-02-28 07:11:08.0581 [PID=15912] [version] Checking plan class 'FGRPopencl-intel_gpu'
2021-02-28 07:11:08.0582 [PID=15912] [version] parsed project prefs setting 'gpu_util_fgrp': 1.000000
2021-02-28 07:11:08.0582 [PID=15912] [version] No Intel GPU devices found
2021-02-28 07:11:08.0582 [PID=15912] [version] Checking plan class 'FGRPopencl-nvidia'
2021-02-28 07:11:08.0582 [PID=15912] [version] parsed project prefs setting 'gpu_util_fgrp': 1.000000
2021-02-28 07:11:08.0582 [PID=15912] [version] GPU RAM calculated: min: 766 MB, use: 750 MB, WU#528059745 CPU: 429 MB
2021-02-28 07:11:08.0582 [PID=15912] [version] NVidia compute capability: 502
2021-02-28 07:11:08.0582 [PID=15912] [version] Peak flops supplied: 1.5072e+11
2021-02-28 07:11:08.0582 [PID=15912] [version] plan class ok
2021-02-28 07:11:08.0582 [PID=15912] [version] Checking plan class 'FGRPopencl1K-ati'
2021-02-28 07:11:08.0582 [PID=15912] [version] parsed project prefs setting 'gpu_util_fgrp': 1.000000
2021-02-28 07:11:08.0582 [PID=15912] [version] No ATI devices found
2021-02-28 07:11:08.0582 [PID=15912] [version] Checking plan class 'FGRPopencl1K-nvidia'
2021-02-28 07:11:08.0582 [PID=15912] [version] parsed project prefs setting 'gpu_util_fgrp': 1.000000
2021-02-28 07:11:08.0582 [PID=15912] [version] GPU RAM calculated: min: 1000 MB, use: 750 MB, WU#528059745 CPU: 429 MB
2021-02-28 07:11:08.0582 [PID=15912] [version] NVidia compute capability: 502
2021-02-28 07:11:08.0582 [PID=15912] [version] Peak flops supplied: 1.5072e+11
2021-02-28 07:11:08.0582 [PID=15912] [version] plan class ok
2021-02-28 07:11:08.0583 [PID=15912] [version] beta test app versions not allowed in project prefs.
2021-02-28 07:11:08.0583 [PID=15912] [version] Checking plan class 'FGRPopenclTV-nvidia'
2021-02-28 07:11:08.0583 [PID=15912] [version] parsed project prefs setting 'gpu_util_fgrp': 1.000000
2021-02-28 07:11:08.0583 [PID=15912] [version] GPU RAM calculated: min: 1000 MB, use: 750 MB, WU#528059745 CPU: 429 MB
2021-02-28 07:11:08.0583 [PID=15912] [version] NVidia compute capability: 502
2021-02-28 07:11:08.0583 [PID=15912] [version] CUDA compute capability required min: 700, supplied: 502
2021-02-28 07:11:08.0583 [PID=15912] [version] Best version of app hsgamma_FGRPB1G is 1.22 ID 1137 FGRPopencl-nvidia (151.14 GFLOPS)
The notice that 700 compute
)
The notice that 700 compute capability being required while the gpu only has 502 is interesting.
The 700 Compute Capability was what the FGRPopenclTV-nvidia apps were developed for. Volta was the first 700 level CC cards followed up by all the cards of the Turing release with CC of 750. Seems the FGRPopencl1K-nvidia application also has a minimum of CC 700.
Looks like if the card is older and below CC 700, then it falls back to the basic FGRPopencl-nvidia application.
From cited log it seems
)
From cited log it seems biggest difference is GPU RAM. "1K" requires 1GB while another executable requires less.
Seems only plan class differs.
I did MD5 for another issue so:
D4A60788AAA42F236E319475895A86D4 *hsgamma_FGRPB1G_1.22_windows_x86_64__FGRPopencl-nvidia.exe
D4A60788AAA42F236E319475895A86D4 *hsgamma_FGRPB1G_1.22_windows_x86_64__FGRPopencl1K-nvidia.exe
Binaries are identical. Just different plan classes. So, If I see differencies in run times they are not from this...