Multi thread calculation with Hyper-Threading Technology sometimes returns erroneous answer
kanekaka
Have anyone experienced similar problem?
I wish 'ThreadProcessorCount' to return 'core number' instead of 'thread number'.
ThreadSafe Static Function TSPeakDetect(wSou, x1, x2, xWidth, wPkDt) Wave wSou, wPkDt Variable x1, x2, xWidth Variable vDeltaX = deltax(wSou), i for(i = 0; x1 < x2; i += 1) if (x1 + xWidth > x2) xWidth = x2 - x1 + vDeltaX endif WaveStats /M=1 /Q/Z/R = (x1, x1 + xWidth - vDeltaX) wSou wPkDt[i * 2] = V_min wPkDt[i * 2 + 1] = V_max x1 += xWidth endfor return i end Function/S MTPeakDetect(wSou, xWidth) Wave wSou Variable xWidth Variable vDeltaX = deltax(wSou), vNumpnt = numpnts(wSou) Variable vSize = 2 * ceil(vDeltaX * vNumpnt / xWidth), x1 = leftx(wSou), x2 = pnt2x(wSou, vNumpnt - 1) String sDest = UniqueName(NameOfWave(wSou) + "PkDt", 1, 0) Duplicate wSou, $sDest Redimension /N=(vSize) $sDest SetScale/P x x1, xWidth / 2, WaveUnits(wSou, 0), $sDest Wave wPkDt = $sDest Variable vPC = ThreadProcessorCount, i if (vPC >= 2) if ((vPC >= 4) && !mod(vPC, 2)) vPC /= 2 // By Core i5(2 core) with HTT(4 thread), 4 thread calcuration answers wrong result. endif Variable vMT = ceil(vSize / 2 / vPC), vTGS = 0, vTGW = 100 String sDests = UniqueName("Tmp", 1, 0), sX1 = UniqueName("sX1_", 1, 0), sX2 = UniqueName("sX2_", 1, 0) Make /D/N=(vPC) $sX1, $sX2 Wave wX1 = $sX1 Wave wX2 = $sX2 wX1[0] = x1 for(i = 1; i < vPC; i += 1) Make /N=(vMT * 2) $StringFromList(i - 1, sDests) sDest = UniqueName("Tmp", 1, 0) sDests += ";" + sDest wX2[i - 1] = wX1[i - 1] + xWidth * vMT wX1[i] = wX2[i - 1] endfor Make /N=(vMT * 2) $StringFromList(i - 1, sDests) wX2[i - 1] = x2 vMT = ThreadGroupCreate(vPC) for(i = 0; i < vPC; i += 1) ThreadStart vMT, i, TSPeakDetect(wSou, wX1[i], wX2[i], xWidth, $StringFromList(i, sDests)) endfor do vTGS = ThreadGroupWait(vMT, vTGW) while (vTGS != 0) x1 = 0 for(i = 0; i < vPC; i += 1) x2 = ThreadReturnValue(vMT, i) * 2 Wave wTmp = $StringFromList(i, sDests) wPkDt[x1, x1 + x2 - 1] = wTmp[p - x1] x1 = x2 KillWaves /Z wTmp endfor KillWaves /Z $sX1, $sX2 vTGS = ThreadGroupRelease(vMT) else for(i = 0; x1 < x2; i += 1) if (x1 + xWidth > x2) xWidth = x2 - x1 + vDeltaX endif WaveStats /M=1 /Q/Z/R = (x1, x1 + xWidth - vDeltaX) wSou wPkDt[i * 2] = V_min wPkDt[i * 2 + 1] = V_max x1 += xWidth endfor endif return NameOfWave(wPkDt) end
ThreadProcessorCount always returns the number of processors including hyperthreads. Even if that was an option, I don't think there is any guarantee that threads created with ThreadGroupCreate and executed with ThreadStart will actually run on separate cores/processors.
I opened your experiment in Igor 8 and executed MTPeakDetect(tsTest1, 2e-5) and Igor gave an index out of range error. If you are ignoring this, that might explain your problem. It's hard to track down the cause of these types of errors in a multiple thread situation, but I think the problem is here:
wPkDt[x1, x1 + x2 - 1] = wTmp[p - x1]If I add a print statement just before that line, like this:
then I get this output:
Note that the next to the last line has a value for x2 that will cause the wave assignment statement to give the index out of range error.
I suspect that your problem ultimately comes down to floating point rounding error, since you're passing around x values and telling WaveStats to use an X range instead of a point range (/R=() vs /R=[]).
I don't know whether the issues I pointed out above are responsible for your calculation getting the wrong answer, but I have a feeling they are.
October 22, 2018 at 10:52 am - Permalink
In reply to ThreadProcessorCount always… by aclight
Thank you so much aclight.
Your suggestion helps me.
One mistake caused the trouble.
Modified procedure is bellow.
October 22, 2018 at 08:11 pm - Permalink