Chris@10: /* Chris@10: * Copyright (c) 2003, 2007-11 Matteo Frigo Chris@10: * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology Chris@10: * Chris@10: * This program is free software; you can redistribute it and/or modify Chris@10: * it under the terms of the GNU General Public License as published by Chris@10: * the Free Software Foundation; either version 2 of the License, or Chris@10: * (at your option) any later version. Chris@10: * Chris@10: * This program is distributed in the hope that it will be useful, Chris@10: * but WITHOUT ANY WARRANTY; without even the implied warranty of Chris@10: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the Chris@10: * GNU General Public License for more details. Chris@10: * Chris@10: * You should have received a copy of the GNU General Public License Chris@10: * along with this program; if not, write to the Free Software Chris@10: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA Chris@10: * Chris@10: */ Chris@10: Chris@10: /* This file was automatically generated --- DO NOT EDIT */ Chris@10: /* Generated on Sun Nov 25 07:39:46 EST 2012 */ Chris@10: Chris@10: #include "codelet-rdft.h" Chris@10: Chris@10: #ifdef HAVE_FMA Chris@10: Chris@10: /* Generated by: ../../../genfft/gen_r2cf.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 13 -name r2cf_13 -include r2cf.h */ Chris@10: Chris@10: /* Chris@10: * This function contains 76 FP additions, 51 FP multiplications, Chris@10: * (or, 31 additions, 6 multiplications, 45 fused multiply/add), Chris@10: * 68 stack variables, 23 constants, and 26 memory accesses Chris@10: */ Chris@10: #include "r2cf.h" Chris@10: Chris@10: static void r2cf_13(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) Chris@10: { Chris@10: DK(KP516520780, +0.516520780623489722840901288569017135705033622); Chris@10: DK(KP300462606, +0.300462606288665774426601772289207995520941381); Chris@10: DK(KP581704778, +0.581704778510515730456870384989698884939833902); Chris@10: DK(KP859542535, +0.859542535098774820163672132761689612766401925); Chris@10: DK(KP769338817, +0.769338817572980603471413688209101117038278899); Chris@10: DK(KP686558370, +0.686558370781754340655719594850823015421401653); Chris@10: DK(KP514918778, +0.514918778086315755491789696138117261566051239); Chris@10: DK(KP251768516, +0.251768516431883313623436926934233488546674281); Chris@10: DK(KP503537032, +0.503537032863766627246873853868466977093348562); Chris@10: DK(KP904176221, +0.904176221990848204433795481776887926501523162); Chris@10: DK(KP575140729, +0.575140729474003121368385547455453388461001608); Chris@10: DK(KP957805992, +0.957805992594665126462521754605754580515587217); Chris@10: DK(KP600477271, +0.600477271932665282925769253334763009352012849); Chris@10: DK(KP522026385, +0.522026385161275033714027226654165028300441940); Chris@10: DK(KP301479260, +0.301479260047709873958013540496673347309208464); Chris@10: DK(KP226109445, +0.226109445035782405468510155372505010481906348); Chris@10: DK(KP853480001, +0.853480001859823990758994934970528322872359049); Chris@10: DK(KP083333333, +0.083333333333333333333333333333333333333333333); Chris@10: DK(KP612264650, +0.612264650376756543746494474777125408779395514); Chris@10: DK(KP038632954, +0.038632954644348171955506895830342264440241080); Chris@10: DK(KP302775637, +0.302775637731994646559610633735247973125648287); Chris@10: DK(KP866025403, +0.866025403784438646763723170752936183471402627); Chris@10: DK(KP500000000, +0.500000000000000000000000000000000000000000000); Chris@10: { Chris@10: INT i; Chris@10: for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(52, rs), MAKE_VOLATILE_STRIDE(52, csr), MAKE_VOLATILE_STRIDE(52, csi)) { Chris@10: E T15, T1a, T11, T17, T14, T1b; Chris@10: { Chris@10: E TN, TD, TV, TA, Tb, TZ, T12, TS, Tx, Tu, Ti, TU; Chris@10: TN = R0[0]; Chris@10: { Chris@10: E T3, TP, Th, TB, Tp, Te, Tm, TC, Tr, T6, T9, Ts; Chris@10: { Chris@10: E Tn, Tf, Tg, T1, T2; Chris@10: T1 = R0[WS(rs, 4)]; Chris@10: T2 = R1[WS(rs, 2)]; Chris@10: Tn = R0[WS(rs, 6)]; Chris@10: Tf = R0[WS(rs, 5)]; Chris@10: Tg = R0[WS(rs, 2)]; Chris@10: T3 = T1 - T2; Chris@10: TP = T1 + T2; Chris@10: { Chris@10: E Tk, To, Tc, Td; Chris@10: Tk = R1[0]; Chris@10: Th = Tf - Tg; Chris@10: To = Tf + Tg; Chris@10: Tc = R1[WS(rs, 4)]; Chris@10: Td = R1[WS(rs, 1)]; Chris@10: { Chris@10: E T4, Tl, T5, T7, T8; Chris@10: T4 = R1[WS(rs, 5)]; Chris@10: TB = Tn + To; Chris@10: Tp = FMS(KP500000000, To, Tn); Chris@10: Tl = Td + Tc; Chris@10: Te = Tc - Td; Chris@10: T5 = R0[WS(rs, 3)]; Chris@10: T7 = R1[WS(rs, 3)]; Chris@10: T8 = R0[WS(rs, 1)]; Chris@10: Tm = FNMS(KP500000000, Tl, Tk); Chris@10: TC = Tk + Tl; Chris@10: Tr = T4 + T5; Chris@10: T6 = T4 - T5; Chris@10: T9 = T7 - T8; Chris@10: Ts = T7 + T8; Chris@10: } Chris@10: } Chris@10: } Chris@10: { Chris@10: E TO, Ta, Tt, TQ; Chris@10: TD = TB - TC; Chris@10: TO = TC + TB; Chris@10: Ta = T6 + T9; Chris@10: TV = T6 - T9; Chris@10: Tt = Tr - Ts; Chris@10: TQ = Tr + Ts; Chris@10: { Chris@10: E TX, Tq, TR, TY; Chris@10: TX = Tm - Tp; Chris@10: Tq = Tm + Tp; Chris@10: TA = T3 + Ta; Chris@10: Tb = FNMS(KP500000000, Ta, T3); Chris@10: TR = TP + TQ; Chris@10: TY = FNMS(KP500000000, TQ, TP); Chris@10: TZ = TX + TY; Chris@10: T12 = TX - TY; Chris@10: T15 = TO - TR; Chris@10: TS = TO + TR; Chris@10: Tx = FNMS(KP866025403, Tt, Tq); Chris@10: Tu = FMA(KP866025403, Tt, Tq); Chris@10: Ti = Te + Th; Chris@10: TU = Th - Te; Chris@10: } Chris@10: } Chris@10: } Chris@10: Cr[0] = TN + TS; Chris@10: { Chris@10: E Tw, Tj, T13, TW; Chris@10: Tw = FNMS(KP866025403, Ti, Tb); Chris@10: Tj = FMA(KP866025403, Ti, Tb); Chris@10: T13 = TU - TV; Chris@10: TW = TU + TV; Chris@10: { Chris@10: E TE, TI, Tv, TF, TG, Ty; Chris@10: TE = FMA(KP302775637, TD, TA); Chris@10: TI = FNMS(KP302775637, TA, TD); Chris@10: Tv = FMA(KP038632954, Tu, Tj); Chris@10: TF = FNMS(KP038632954, Tj, Tu); Chris@10: TG = FNMS(KP612264650, Tw, Tx); Chris@10: Ty = FMA(KP612264650, Tx, Tw); Chris@10: { Chris@10: E TT, Tz, TK, TH, TM, T10, TL, TJ; Chris@10: TT = FNMS(KP083333333, TS, TN); Chris@10: Tz = FNMS(KP853480001, Ty, Tv); Chris@10: TK = FMA(KP853480001, Ty, Tv); Chris@10: TH = FNMS(KP853480001, TG, TF); Chris@10: TM = FMA(KP853480001, TG, TF); Chris@10: T1a = FNMS(KP226109445, TW, TZ); Chris@10: T10 = FMA(KP301479260, TZ, TW); Chris@10: TL = FNMS(KP522026385, Tz, TE); Chris@10: Ci[WS(csi, 1)] = KP600477271 * (FMA(KP957805992, TE, Tz)); Chris@10: TJ = FMA(KP522026385, TH, TI); Chris@10: Ci[WS(csi, 5)] = -(KP600477271 * (FNMS(KP957805992, TI, TH))); Chris@10: Ci[WS(csi, 4)] = -(KP575140729 * (FMA(KP904176221, TM, TL))); Chris@10: Ci[WS(csi, 3)] = KP575140729 * (FNMS(KP904176221, TM, TL)); Chris@10: Ci[WS(csi, 6)] = KP575140729 * (FMA(KP904176221, TK, TJ)); Chris@10: Ci[WS(csi, 2)] = KP575140729 * (FNMS(KP904176221, TK, TJ)); Chris@10: T11 = FMA(KP503537032, T10, TT); Chris@10: T17 = FNMS(KP251768516, T10, TT); Chris@10: } Chris@10: T14 = FNMS(KP514918778, T13, T12); Chris@10: T1b = FMA(KP686558370, T12, T13); Chris@10: } Chris@10: } Chris@10: } Chris@10: { Chris@10: E T1e, T1c, T18, T16, T1d, T19; Chris@10: T1e = FMA(KP769338817, T1b, T1a); Chris@10: T1c = FNMS(KP769338817, T1b, T1a); Chris@10: T18 = FNMS(KP859542535, T14, T15); Chris@10: T16 = FMA(KP581704778, T15, T14); Chris@10: T1d = FNMS(KP300462606, T18, T17); Chris@10: T19 = FMA(KP300462606, T18, T17); Chris@10: Cr[WS(csr, 1)] = FMA(KP516520780, T16, T11); Chris@10: Cr[WS(csr, 5)] = FNMS(KP516520780, T16, T11); Chris@10: Cr[WS(csr, 2)] = FMA(KP503537032, T1e, T1d); Chris@10: Cr[WS(csr, 6)] = FNMS(KP503537032, T1e, T1d); Chris@10: Cr[WS(csr, 3)] = FMA(KP503537032, T1c, T19); Chris@10: Cr[WS(csr, 4)] = FNMS(KP503537032, T1c, T19); Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: Chris@10: static const kr2c_desc desc = { 13, "r2cf_13", {31, 6, 45, 0}, &GENUS }; Chris@10: Chris@10: void X(codelet_r2cf_13) (planner *p) { Chris@10: X(kr2c_register) (p, r2cf_13, &desc); Chris@10: } Chris@10: Chris@10: #else /* HAVE_FMA */ Chris@10: Chris@10: /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 13 -name r2cf_13 -include r2cf.h */ Chris@10: Chris@10: /* Chris@10: * This function contains 76 FP additions, 34 FP multiplications, Chris@10: * (or, 57 additions, 15 multiplications, 19 fused multiply/add), Chris@10: * 55 stack variables, 20 constants, and 26 memory accesses Chris@10: */ Chris@10: #include "r2cf.h" Chris@10: Chris@10: static void r2cf_13(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) Chris@10: { Chris@10: DK(KP083333333, +0.083333333333333333333333333333333333333333333); Chris@10: DK(KP075902986, +0.075902986037193865983102897245103540356428373); Chris@10: DK(KP251768516, +0.251768516431883313623436926934233488546674281); Chris@10: DK(KP503537032, +0.503537032863766627246873853868466977093348562); Chris@10: DK(KP113854479, +0.113854479055790798974654345867655310534642560); Chris@10: DK(KP265966249, +0.265966249214837287587521063842185948798330267); Chris@10: DK(KP387390585, +0.387390585467617292130675966426762851778775217); Chris@10: DK(KP300462606, +0.300462606288665774426601772289207995520941381); Chris@10: DK(KP132983124, +0.132983124607418643793760531921092974399165133); Chris@10: DK(KP258260390, +0.258260390311744861420450644284508567852516811); Chris@10: DK(KP2_000000000, +2.000000000000000000000000000000000000000000000); Chris@10: DK(KP1_732050807, +1.732050807568877293527446341505872366942805254); Chris@10: DK(KP300238635, +0.300238635966332641462884626667381504676006424); Chris@10: DK(KP011599105, +0.011599105605768290721655456654083252189827041); Chris@10: DK(KP156891391, +0.156891391051584611046832726756003269660212636); Chris@10: DK(KP256247671, +0.256247671582936600958684654061725059144125175); Chris@10: DK(KP174138601, +0.174138601152135905005660794929264742616964676); Chris@10: DK(KP575140729, +0.575140729474003121368385547455453388461001608); Chris@10: DK(KP866025403, +0.866025403784438646763723170752936183471402627); Chris@10: DK(KP500000000, +0.500000000000000000000000000000000000000000000); Chris@10: { Chris@10: INT i; Chris@10: for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(52, rs), MAKE_VOLATILE_STRIDE(52, csr), MAKE_VOLATILE_STRIDE(52, csi)) { Chris@10: E T13, Tb, Tm, TW, TX, T14, TU, T10, Tz, TB, Tu, TC, TR, T11; Chris@10: T13 = R0[0]; Chris@10: { Chris@10: E Te, TO, Ta, Tv, To, T5, Tw, Tp, Th, Tr, Tk, Ts, Tl, TP, Tc; Chris@10: E Td; Chris@10: Tc = R0[WS(rs, 4)]; Chris@10: Td = R1[WS(rs, 2)]; Chris@10: Te = Tc - Td; Chris@10: TO = Tc + Td; Chris@10: { Chris@10: E T6, T7, T8, T9; Chris@10: T6 = R1[0]; Chris@10: T7 = R1[WS(rs, 1)]; Chris@10: T8 = R1[WS(rs, 4)]; Chris@10: T9 = T7 + T8; Chris@10: Ta = T6 + T9; Chris@10: Tv = T7 - T8; Chris@10: To = FNMS(KP500000000, T9, T6); Chris@10: } Chris@10: { Chris@10: E T1, T2, T3, T4; Chris@10: T1 = R0[WS(rs, 6)]; Chris@10: T2 = R0[WS(rs, 5)]; Chris@10: T3 = R0[WS(rs, 2)]; Chris@10: T4 = T2 + T3; Chris@10: T5 = T1 + T4; Chris@10: Tw = T2 - T3; Chris@10: Tp = FNMS(KP500000000, T4, T1); Chris@10: } Chris@10: { Chris@10: E Tf, Tg, Ti, Tj; Chris@10: Tf = R1[WS(rs, 5)]; Chris@10: Tg = R0[WS(rs, 3)]; Chris@10: Th = Tf - Tg; Chris@10: Tr = Tf + Tg; Chris@10: Ti = R1[WS(rs, 3)]; Chris@10: Tj = R0[WS(rs, 1)]; Chris@10: Tk = Ti - Tj; Chris@10: Ts = Ti + Tj; Chris@10: } Chris@10: Tl = Th + Tk; Chris@10: TP = Tr + Ts; Chris@10: Tb = T5 - Ta; Chris@10: Tm = Te + Tl; Chris@10: TW = Ta + T5; Chris@10: TX = TO + TP; Chris@10: T14 = TW + TX; Chris@10: { Chris@10: E TS, TT, Tx, Ty; Chris@10: TS = Tv + Tw; Chris@10: TT = Th - Tk; Chris@10: TU = TS - TT; Chris@10: T10 = TS + TT; Chris@10: Tx = KP866025403 * (Tv - Tw); Chris@10: Ty = FNMS(KP500000000, Tl, Te); Chris@10: Tz = Tx + Ty; Chris@10: TB = Ty - Tx; Chris@10: } Chris@10: { Chris@10: E Tq, Tt, TN, TQ; Chris@10: Tq = To - Tp; Chris@10: Tt = KP866025403 * (Tr - Ts); Chris@10: Tu = Tq - Tt; Chris@10: TC = Tq + Tt; Chris@10: TN = To + Tp; Chris@10: TQ = FNMS(KP500000000, TP, TO); Chris@10: TR = TN - TQ; Chris@10: T11 = TN + TQ; Chris@10: } Chris@10: } Chris@10: Cr[0] = T13 + T14; Chris@10: { Chris@10: E Tn, TG, TE, TF, TJ, TM, TK, TL; Chris@10: Tn = FNMS(KP174138601, Tm, KP575140729 * Tb); Chris@10: TG = FMA(KP174138601, Tb, KP575140729 * Tm); Chris@10: { Chris@10: E TA, TD, TH, TI; Chris@10: TA = FNMS(KP156891391, Tz, KP256247671 * Tu); Chris@10: TD = FNMS(KP300238635, TC, KP011599105 * TB); Chris@10: TE = TA + TD; Chris@10: TF = KP1_732050807 * (TD - TA); Chris@10: TH = FMA(KP300238635, TB, KP011599105 * TC); Chris@10: TI = FMA(KP256247671, Tz, KP156891391 * Tu); Chris@10: TJ = TH - TI; Chris@10: TM = KP1_732050807 * (TI + TH); Chris@10: } Chris@10: Ci[WS(csi, 5)] = FMA(KP2_000000000, TE, Tn); Chris@10: Ci[WS(csi, 1)] = FMA(KP2_000000000, TJ, TG); Chris@10: TK = TG - TJ; Chris@10: Ci[WS(csi, 4)] = TF - TK; Chris@10: Ci[WS(csi, 3)] = TF + TK; Chris@10: TL = Tn - TE; Chris@10: Ci[WS(csi, 2)] = TL - TM; Chris@10: Ci[WS(csi, 6)] = TL + TM; Chris@10: } Chris@10: { Chris@10: E TZ, T1b, T19, T1e, T16, T1a, TV, TY, T1c, T1d; Chris@10: TV = FNMS(KP132983124, TU, KP258260390 * TR); Chris@10: TY = KP300462606 * (TW - TX); Chris@10: TZ = FMA(KP2_000000000, TV, TY); Chris@10: T1b = TY - TV; Chris@10: { Chris@10: E T17, T18, T12, T15; Chris@10: T17 = FMA(KP387390585, TU, KP265966249 * TR); Chris@10: T18 = FNMS(KP503537032, T11, KP113854479 * T10); Chris@10: T19 = T17 - T18; Chris@10: T1e = T17 + T18; Chris@10: T12 = FMA(KP251768516, T10, KP075902986 * T11); Chris@10: T15 = FNMS(KP083333333, T14, T13); Chris@10: T16 = FMA(KP2_000000000, T12, T15); Chris@10: T1a = T15 - T12; Chris@10: } Chris@10: Cr[WS(csr, 1)] = TZ + T16; Chris@10: Cr[WS(csr, 5)] = T16 - TZ; Chris@10: T1c = T1a - T1b; Chris@10: Cr[WS(csr, 2)] = T19 + T1c; Chris@10: Cr[WS(csr, 6)] = T1c - T19; Chris@10: T1d = T1b + T1a; Chris@10: Cr[WS(csr, 3)] = T1d - T1e; Chris@10: Cr[WS(csr, 4)] = T1e + T1d; Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: Chris@10: static const kr2c_desc desc = { 13, "r2cf_13", {57, 15, 19, 0}, &GENUS }; Chris@10: Chris@10: void X(codelet_r2cf_13) (planner *p) { Chris@10: X(kr2c_register) (p, r2cf_13, &desc); Chris@10: } Chris@10: Chris@10: #endif /* HAVE_FMA */