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

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>

1 files changed, 14 insertions, 3 deletions

diff --git a/opencl/source/opencl_device.cxx b/opencl/source/opencl_device.cxx
index 0a0d28ee4200..3dc9a29413ad 100644
--- a/opencl/source/opencl_device.cxx
+++ b/opencl/source/opencl_device.cxx
@@ -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;
     }