mir/libreoffice
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Don't waste too much time evaluating the "score" for the "native CPU"

Browse Source 
This code took tens of second typically, which is utterly pointless.
Instead run the loop for max a second, after which estimate how long
it would have taken to process all of the data.

Don't confuse the "native CPU" term here with a CPU-based device
of an actual OpenCL implementation. That is a completely different
thing. This "native CPU" thing here is just our way to estimate how
much time it takes to do calculations using normal C++ code on the
CPU.

Change-Id: I92f5eedc06bbaaef6a9b5322fefec9d41f0db505
Reviewed-on: https://gerrit.libreoffice.org/26774
Reviewed-by: Michael Meeks <michael.meeks@collabora.com>
Tested-by: Jenkins <ci@libreoffice.org>
This commit is contained in:
Tor Lillqvist
2016-06-29 18:25:26 +03:00
parent f4f580f4a4
commit 68106cbac5
1 changed files with 14 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
opencl/source/opencl_device.cxx
View File

@@ -327,7 +327,8 @@ ds_status evaluateScoreForDevice(ds_device& rDevice, std::unique_ptr<LibreOffice
timer kernelTime;
timerStart(&kernelTime);
for (unsigned long j = 0; j < testData->outputSize; j++)
unsigned long j;
for (j = 0; j < testData->outputSize; j++)
{
double fAverage = 0.0f;
double fMin = DBL_MAX;
@@ -340,15 +341,25 @@ ds_status evaluateScoreForDevice(ds_device& rDevice, std::unique_ptr<LibreOffice
}
fAverage /= testData->inputSize;
testData->output[j] = fAverage + (fMin * fSoP);
// Don't run for much longer than one second
if (j > 0 && j % 100 == 0)
{
rDevice.fTime = timerCurrent(&kernelTime);
if (rDevice.fTime >= 1)
break;
}
}
rDevice.fTime = timerCurrent(&kernelTime);
// Scale time to how long it would have taken to go all the way to outputSize
rDevice.fTime /= ((double) j / testData->outputSize);
// InterpretTail - the S/W fallback is nothing like as efficient
// as any good openCL implementation: no SIMD, tons of branching
// in the inner loops etc. Generously characterise it as only 10x
// slower than the above.
float fInterpretTailFactor = 10.0;
rDevice.fTime = timerCurrent(&kernelTime);
rDevice.fTime *= fInterpretTailFactor;
rDevice.bErrors = false;
}