From 1a3d12deacb7e2dbff45927a0d2fffa9da912a9e Mon Sep 17 00:00:00 2001 From: babrodtk Date: Wed, 5 Aug 2015 11:02:45 +0200 Subject: [PATCH] implemented struct to keep track of derivatives during interpolation --- opm/autodiff/VFPProperties.cpp | 100 ++++++++++++++++++--------------- 1 file changed, 54 insertions(+), 46 deletions(-) diff --git a/opm/autodiff/VFPProperties.cpp b/opm/autodiff/VFPProperties.cpp index d9d2cd26a..bb47bd80a 100644 --- a/opm/autodiff/VFPProperties.cpp +++ b/opm/autodiff/VFPProperties.cpp @@ -361,21 +361,53 @@ VFPProdProperties::InterpData VFPProdProperties::find_interp_data(const double& #pragma GCC optimize ("unroll-loops") #endif +namespace detail { + + //An "ADB-like" structure with a value and a set of derivatives + //Just defined to make sure that operator+ and operator* do the + //correct thing for the use in this function + //Wastes some space (AoS versus SoA), but resulting code is easier to read and maintain + struct adb_like { + adb_like() : value(0.0), dthp(0.0), dwfr(0.0), dgfr(0.0), dalq(0.0), dflo(0.0) {}; + double value; + double dthp; + double dwfr; + double dgfr; + double dalq; + double dflo; + }; + + adb_like operator+(adb_like lhs, const adb_like& rhs) { + lhs.value += rhs.value; + lhs.dthp += rhs.dthp; + lhs.dwfr += rhs.dwfr; + lhs.dgfr += rhs.dgfr; + lhs.dalq += rhs.dalq; + lhs.dflo += rhs.dflo; + return lhs; + } + + adb_like operator*(double lhs, adb_like rhs) { + rhs.value *= lhs; + rhs.dthp *= lhs; + rhs.dwfr *= lhs; + rhs.dgfr *= lhs; + rhs.dalq *= lhs; + rhs.dflo *= lhs; + return rhs; + } +} + double VFPProdProperties::interpolate(const VFPProdTable::array_type& array, const InterpData& flo_i, const InterpData& thp_i, const InterpData& wfr_i, const InterpData& gfr_i, const InterpData& alq_i) { - //Values in our hypercube - double nn[2][2][2][2][2]; - //Derivatives in our hypercube - double dthp[2][2][2][2][2]; - double dwfr[2][2][2][2][2]; - double dgfr[2][2][2][2][2]; - double dalq[2][2][2][2][2]; - double dflo[2][2][2][2][2]; + //Values and derivatives in a 5D hypercube + detail::adb_like nn[2][2][2][2][2]; + //Pick out nearest neighbors (nn) to our evaluation point //This is not really required, but performance-wise it may pay off, since the 32-elements @@ -395,7 +427,7 @@ double VFPProdProperties::interpolate(const VFPProdTable::array_type& array, const int fi = flo_i.ind_[f]; //Copy element - nn[t][w][g][a][f] = array[ti][wi][gi][ai][fi]; + nn[t][w][g][a][f].value = array[ti][wi][gi][ai][fi]; } } } @@ -403,23 +435,23 @@ double VFPProdProperties::interpolate(const VFPProdTable::array_type& array, } //Calculate derivatives + //Note that the derivative of the two end points of a line aligned with the + //"axis of the derivative" are equal for (int i=0; i<=1; ++i) { for (int j=0; j<=1; ++j) { for (int k=0; k<=1; ++k) { for (int l=0; l<=1; ++l) { - dthp[0][i][j][k][l] = nn[1][i][j][k][l] - nn[0][i][j][k][l]; - dwfr[i][0][j][k][l] = nn[i][1][j][k][l] - nn[i][0][j][k][l]; - dgfr[i][j][0][k][l] = nn[i][j][1][k][l] - nn[i][j][0][k][l]; - dalq[i][j][k][0][l] = nn[i][j][k][1][l] - nn[i][j][k][0][l]; - dflo[i][j][k][l][0] = nn[i][j][k][l][1] - nn[i][j][k][l][0]; + nn[0][i][j][k][l].dthp = nn[1][i][j][k][l].value - nn[0][i][j][k][l].value; + nn[i][0][j][k][l].dwfr = nn[i][1][j][k][l].value - nn[i][0][j][k][l].value; + nn[i][j][0][k][l].dgfr = nn[i][j][1][k][l].value - nn[i][j][0][k][l].value; + nn[i][j][k][0][l].dalq = nn[i][j][k][1][l].value - nn[i][j][k][0][l].value; + nn[i][j][k][l][0].dflo = nn[i][j][k][l][1].value - nn[i][j][k][l][0].value; - //For simplicity of the rest of the interpolation code, - //we copy the derivatives in the full hypercube - dthp[1][i][j][k][l] = dthp[0][i][j][k][l]; - dwfr[i][1][j][k][l] = dwfr[i][0][j][k][l]; - dgfr[i][j][1][k][l] = dgfr[i][j][0][k][l]; - dalq[i][j][k][1][l] = dalq[i][j][k][0][l]; - dflo[i][j][k][l][1] = dflo[i][j][k][l][0]; + nn[1][i][j][k][l].dthp = nn[0][i][j][k][l].dthp; + nn[i][1][j][k][l].dwfr = nn[i][0][j][k][l].dwfr; + nn[i][j][1][k][l].dgfr = nn[i][j][0][k][l].dgfr; + nn[i][j][k][1][l].dalq = nn[i][j][k][0][l].dalq; + nn[i][j][k][l][1].dflo = nn[i][j][k][l][0].dflo; } } } @@ -438,12 +470,6 @@ double VFPProdProperties::interpolate(const VFPProdTable::array_type& array, for (int g=0; g<=1; ++g) { for (int a=0; a<=1; ++a) { nn[t][w][g][a][0] = a*nn[t][w][g][a][0] + b*nn[t][w][g][a][1]; - - dthp[t][w][g][a][0] = a*dthp[t][w][g][a][0] + b*dthp[t][w][g][a][1]; - dwfr[t][w][g][a][0] = a*dwfr[t][w][g][a][0] + b*dwfr[t][w][g][a][1]; - dgfr[t][w][g][a][0] = a*dgfr[t][w][g][a][0] + b*dgfr[t][w][g][a][1]; - dalq[t][w][g][a][0] = a*dalq[t][w][g][a][0] + b*dalq[t][w][g][a][1]; - dflo[t][w][g][a][0] = a*dflo[t][w][g][a][0] + b*dflo[t][w][g][a][1]; } } } @@ -455,12 +481,6 @@ double VFPProdProperties::interpolate(const VFPProdTable::array_type& array, for (int w=0; w<=1; ++w) { for (int g=0; g<=1; ++g) { nn[t][w][g][0][0] = a*nn[t][w][g][0][0] + b*nn[t][w][g][1][0]; - - dthp[t][w][g][0][0] = a*dthp[t][w][g][0][0] + b*dthp[t][w][g][1][0]; - dwfr[t][w][g][0][0] = a*dwfr[t][w][g][0][0] + b*dwfr[t][w][g][1][0]; - dgfr[t][w][g][0][0] = a*dgfr[t][w][g][0][0] + b*dgfr[t][w][g][1][0]; - dalq[t][w][g][0][0] = a*dalq[t][w][g][0][0] + b*dalq[t][w][g][1][0]; - dflo[t][w][g][0][0] = a*dflo[t][w][g][0][0] + b*dflo[t][w][g][1][0]; } } } @@ -470,12 +490,6 @@ double VFPProdProperties::interpolate(const VFPProdTable::array_type& array, for (int t=0; t<=1; ++t) { for (int w=0; w<=1; ++w) { nn[t][w][0][0][0] = a*nn[t][w][0][0][0] + b*nn[t][w][1][0][0]; - - dthp[t][w][0][0][0] = a*dthp[t][w][0][0][0] + b*dthp[t][w][1][0][0]; - dwfr[t][w][0][0][0] = a*dwfr[t][w][0][0][0] + b*dwfr[t][w][1][0][0]; - dgfr[t][w][0][0][0] = a*dgfr[t][w][0][0][0] + b*dgfr[t][w][1][0][0]; - dalq[t][w][0][0][0] = a*dalq[t][w][0][0][0] + b*dalq[t][w][1][0][0]; - dflo[t][w][0][0][0] = a*dflo[t][w][0][0][0] + b*dflo[t][w][1][0][0]; } } @@ -483,17 +497,11 @@ double VFPProdProperties::interpolate(const VFPProdTable::array_type& array, a = (1.0-b); for (int t=0; t<=1; ++t) { nn[t][0][0][0][0] = a*nn[t][0][0][0][0] + b*nn[t][1][0][0][0]; - - dthp[t][0][0][0][0] = a*dthp[t][0][0][0][0] + b*dthp[t][1][0][0][0]; - dwfr[t][0][0][0][0] = a*dwfr[t][0][0][0][0] + b*dwfr[t][1][0][0][0]; - dgfr[t][0][0][0][0] = a*dgfr[t][0][0][0][0] + b*dgfr[t][1][0][0][0]; - dalq[t][0][0][0][0] = a*dalq[t][0][0][0][0] + b*dalq[t][1][0][0][0]; - dflo[t][0][0][0][0] = a*dflo[t][0][0][0][0] + b*dflo[t][1][0][0][0]; } b = thp_i.factor_; a = (1.0-b); - return a*nn[0][0][0][0][0] + b*nn[1][0][0][0][0]; + return a*nn[0][0][0][0][0].value + b*nn[1][0][0][0][0].value; } #ifdef __GNUC__