GPU Calc: Optimized GAUSS

AMLOEXT-276

Change-Id: I29517305ffda2673961c2fa4dc46503690933645
Signed-off-by: haochen <haochen@multicorewareinc.com>
Signed-off-by: Wei Wei <weiwei@multicorewareinc.com>
Signed-off-by: I-Jui (Ray) Sung <ray@multicorewareinc.com>

1 files changed, 70 insertions, 1 deletions

diff --git a/sc/source/core/opencl/opinlinefun_statistical.cxx b/sc/source/core/opencl/opinlinefun_statistical.cxx
index 81e8150b06bf..ab034794dc15 100644
--- a/sc/source/core/opencl/opinlinefun_statistical.cxx
+++ b/sc/source/core/opencl/opinlinefun_statistical.cxx
@@ -1272,7 +1272,76 @@ std::string lcl_IterateInverse =
 "    }\n"
 "    return fRx;\n"
 "}\n";
-
+std::string phiDecl=
+"double phi(double x);\n";
+std::string phi =
+"double phi(double x)\n"
+"{\n"
+"    return  0.39894228040143268 * exp(-(x * x) / 2.0);\n"
+"}\n";
+std::string taylorDecl =
+"double taylor(double* pPolynom, uint nMax, double x);\n";
+std::string taylor =
+"double taylor(double* pPolynom, uint nMax, double x)\n"
+"{\n"
+"    double nVal = pPolynom[nMax];\n"
+"    for (short i = nMax-1; i >= 0; i--)\n"
+"    {\n"
+"        nVal = pPolynom[i] + (nVal * x);\n"
+"    }\n"
+"    return nVal;\n"
+"}";
+std::string gaussDecl = "double gauss(double x);\n";
+std::string gauss =
+"double gauss(double x)\n"
+"{\n"
+"    double xAbs = fabs(x);\n"
+"    uint xShort = (uint)(floor(xAbs));\n"
+"    double nVal = 0.0;\n"
+"    if (xShort == 0)\n"
+"    {\n"
+"        double t0[] =\n"
+"        { 0.39894228040143268, -0.06649038006690545,  0.00997355701003582,\n"
+"         -0.00118732821548045,  0.00011543468761616, -0.00000944465625950,\n"
+"          0.00000066596935163, -0.00000004122667415,  0.00000000227352982,\n"
+"          0.00000000011301172,  0.00000000000511243, -0.00000000000021218 };\n"
+"        nVal = taylor(t0, 11, (xAbs * xAbs)) * xAbs;\n"
+"    }\n"
+"    else if ((xShort >= 1) && (xShort <= 2))\n"
+"    {\n"
+"        double t2[] =\n"
+"        { 0.47724986805182079,  0.05399096651318805, -0.05399096651318805,\n"
+"          0.02699548325659403, -0.00449924720943234, -0.00224962360471617,\n"
+"          0.00134977416282970, -0.00011783742691370, -0.00011515930357476,\n"
+"          0.00003704737285544,  0.00000282690796889, -0.00000354513195524,\n"
+"          0.00000037669563126,  0.00000019202407921, -0.00000005226908590,\n"
+"         -0.00000000491799345,  0.00000000366377919, -0.00000000015981997,\n"
+"         -0.00000000017381238,  0.00000000002624031,  0.00000000000560919,\n"
+"         -0.00000000000172127, -0.00000000000008634, 0.00000000000007894 };\n"
+"        nVal = taylor(t2, 23, (xAbs - 2.0));\n"
+"    }\n"
+"    else if ((xShort >= 3) && (xShort <= 4))\n"
+"    {\n"
+"       double t4[] =\n"
+"       { 0.49996832875816688,  0.00013383022576489, -0.00026766045152977,\n"
+"         0.00033457556441221, -0.00028996548915725,  0.00018178605666397,\n"
+"        -0.00008252863922168,  0.00002551802519049, -0.00000391665839292,\n"
+"        -0.00000074018205222,  0.00000064422023359, -0.00000017370155340,\n"
+"         0.00000000909595465,  0.00000000944943118, -0.00000000329957075,\n"
+"         0.00000000029492075,  0.00000000011874477, -0.00000000004420396,\n"
+"         0.00000000000361422,  0.00000000000143638, -0.00000000000045848 };\n"
+"        nVal = taylor(t4, 20, (xAbs - 4.0));\n"
+"    }\n"
+"    else\n"
+"    {\n"
+"        double asympt[] = { -1.0, 1.0, -3.0, 15.0, -105.0 };\n"
+"        nVal = 0.5 + phi(xAbs) * taylor(asympt, 4, 1.0/(xAbs * xAbs))/xAbs;\n"
+"    }\n"
+"    if (x < 0.0)\n"
+"        return -nVal;\n"
+"    else\n"
+"        return nVal;\n"
+"}\n";
 #endif
 
 /* vim:set shiftwidth=4 softtabstop=4 expandtab: */