Incorrect peak FLOPS for AMD GPU

[davidpanderson]

The OpenCL GPU detection for a particular AMD GPU reports 43 ExaFLOPS - about 10,000X too big.
(this is 7.10.2, but same in 7.14.2)

22-Dec-2018 12:28:23 [---] cc_config.xml not found - using defaults
22-Dec-2018 12:28:23 [---] Starting BOINC client version 7.10.2 for windows_x86_64
22-Dec-2018 12:28:24 [---] OpenCL: AMD/ATI GPU 0: AMD Radeon(TM) Vega 8 Graphics (driver version 2580.6 (PAL,HSAIL), device version OpenCL 2.0 AMD-APP (2580.6), 6567MB, 6567MB available, 43980464 GFLOPS peak)

[adamradocz]

Maybe driver issue? The Intel OpenCL driver shows me different result too, then the Beignet driver,

[RichardHaselgrove]

Whatever the cause, we probably need to update lib/coproc.cpp for AMD/ATI, as we did last year for NVidia in #2707

[davidpanderson]

Addendum from user:

Update: On a hunch, I set the underclocked GPU back to auto (400-1100MHz).  The spurious peak GFLOPS issue went away. 02/02/2019 15:28:02 |  | OpenCL: AMD/ATI GPU 0: AMD Radeon(TM) Vega 8 Graphics (driver version 2766.5 (PAL,HSAIL), device version OpenCL 2.0 AMD-APP (2766.5), 6567MB, 6567MB available, 1127 GFLOPS peak)On underclocking it again, this time to a slightly different nominal speed, we get the old issue back:02/02/2019 15:57:59 |  | OpenCL: AMD/ATI GPU 0: AMD Radeon(TM) Vega 8 Graphics (driver version 2766.5 (PAL,HSAIL), device version OpenCL 2.0 AMD-APP (2766.5), 6567MB, 6567MB available, 43980464 GFLOPS peak)CPUZ seems to report the correct GPU speed in each case.

[RichardHaselgrove]

User Rob Smith reports:

AMD have a tool set that includes something that looks as if it could be of use (or at least a starter for ten) - AMD GPU Services (AGS)
One component of it appears to return a value for the FLOPS

float 	teraFlops
Teraflops of GPU. Zero if not GCN onwards. Calculated from iCoreClock * iNumCUs * 64 Pixels/clk * 2 instructions/MAD. 

https://github.com/GPUOpen-LibrariesAndSDKs/AGS_SDK

[JuhaSointusalo]

Workarounded in #3001.

[JuhaSointusalo]

Eeh, worked around...