cannam@95: /* cannam@95: * Copyright (c) 2003, 2007-11 Matteo Frigo cannam@95: * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology cannam@95: * cannam@95: * This program is free software; you can redistribute it and/or modify cannam@95: * it under the terms of the GNU General Public License as published by cannam@95: * the Free Software Foundation; either version 2 of the License, or cannam@95: * (at your option) any later version. cannam@95: * cannam@95: * This program is distributed in the hope that it will be useful, cannam@95: * but WITHOUT ANY WARRANTY; without even the implied warranty of cannam@95: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the cannam@95: * GNU General Public License for more details. cannam@95: * cannam@95: * You should have received a copy of the GNU General Public License cannam@95: * along with this program; if not, write to the Free Software cannam@95: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA cannam@95: * cannam@95: */ cannam@95: cannam@95: /* This file was automatically generated --- DO NOT EDIT */ cannam@95: /* Generated on Sun Nov 25 07:37:36 EST 2012 */ cannam@95: cannam@95: #include "codelet-dft.h" cannam@95: cannam@95: #ifdef HAVE_FMA cannam@95: cannam@95: /* Generated by: ../../../genfft/gen_notw_c.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 64 -name n2bv_64 -with-ostride 2 -include n2b.h -store-multiple 2 */ cannam@95: cannam@95: /* cannam@95: * This function contains 456 FP additions, 258 FP multiplications, cannam@95: * (or, 198 additions, 0 multiplications, 258 fused multiply/add), cannam@95: * 178 stack variables, 15 constants, and 160 memory accesses cannam@95: */ cannam@95: #include "n2b.h" cannam@95: cannam@95: static void n2bv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) cannam@95: { cannam@95: DVK(KP773010453, +0.773010453362736960810906609758469800971041293); cannam@95: DVK(KP820678790, +0.820678790828660330972281985331011598767386482); cannam@95: DVK(KP881921264, +0.881921264348355029712756863660388349508442621); cannam@95: DVK(KP534511135, +0.534511135950791641089685961295362908582039528); cannam@95: DVK(KP995184726, +0.995184726672196886244836953109479921575474869); cannam@95: DVK(KP098491403, +0.098491403357164253077197521291327432293052451); cannam@95: DVK(KP980785280, +0.980785280403230449126182236134239036973933731); cannam@95: DVK(KP956940335, +0.956940335732208864935797886980269969482849206); cannam@95: DVK(KP303346683, +0.303346683607342391675883946941299872384187453); cannam@95: DVK(KP831469612, +0.831469612302545237078788377617905756738560812); cannam@95: DVK(KP923879532, +0.923879532511286756128183189396788286822416626); cannam@95: DVK(KP668178637, +0.668178637919298919997757686523080761552472251); cannam@95: DVK(KP198912367, +0.198912367379658006911597622644676228597850501); cannam@95: DVK(KP414213562, +0.414213562373095048801688724209698078569671875); cannam@95: DVK(KP707106781, +0.707106781186547524400844362104849039284835938); cannam@95: { cannam@95: INT i; cannam@95: const R *xi; cannam@95: R *xo; cannam@95: xi = ii; cannam@95: xo = io; cannam@95: for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(128, is), MAKE_VOLATILE_STRIDE(128, os)) { cannam@95: V T7z, T7A, T7B, T7C, T5T, T5S, T5X, T65, T8a, T8b, T8e, T8g, T5Z, T5R, T67; cannam@95: V T63, T5U, T64; cannam@95: { cannam@95: V T7, T26, T5k, T6A, T47, T69, T2V, T3z, T6B, T4e, T6a, T5n, T3M, T2Y, T27; cannam@95: V Tm, T3A, T3i, T29, TC, T5p, T4o, T6D, T6e, T3l, T3B, TR, T2a, T4x, T5q; cannam@95: V T6h, T6E, T39, T3H, T3I, T3c, T5N, T57, T72, T6w, T5O, T5e, T71, T6t, T2y; cannam@95: V T1W, T2x, T1N, T33, T34, T3E, T32, T1p, T2v, T1g, T2u, T4M, T5K, T6p, T6Z; cannam@95: V T6m, T6Y, T5L, T4T; cannam@95: { cannam@95: V T4g, T4l, T3g, Tu, Tx, T4h, TA, T4i; cannam@95: { cannam@95: V T1, T2, T23, T24, T4, T5, T20, T21; cannam@95: T1 = LD(&(xi[0]), ivs, &(xi[0])); cannam@95: T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0])); cannam@95: T23 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0])); cannam@95: T24 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0])); cannam@95: T4 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0])); cannam@95: T5 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0])); cannam@95: T20 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0])); cannam@95: T21 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0])); cannam@95: { cannam@95: V Ta, T48, Tk, T4c, T49, Td, Tf, Tg; cannam@95: { cannam@95: V T8, T43, T3, T45, T25, T5i, T6, T44, T22, T9, Ti, Tj, Tb, Tc; cannam@95: T8 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0])); cannam@95: T43 = VSUB(T1, T2); cannam@95: T3 = VADD(T1, T2); cannam@95: T45 = VSUB(T23, T24); cannam@95: T25 = VADD(T23, T24); cannam@95: T5i = VSUB(T4, T5); cannam@95: T6 = VADD(T4, T5); cannam@95: T44 = VSUB(T20, T21); cannam@95: T22 = VADD(T20, T21); cannam@95: T9 = LD(&(xi[WS(is, 36)]), ivs, &(xi[0])); cannam@95: Ti = LD(&(xi[WS(is, 12)]), ivs, &(xi[0])); cannam@95: Tj = LD(&(xi[WS(is, 44)]), ivs, &(xi[0])); cannam@95: Tb = LD(&(xi[WS(is, 20)]), ivs, &(xi[0])); cannam@95: Tc = LD(&(xi[WS(is, 52)]), ivs, &(xi[0])); cannam@95: { cannam@95: V T2T, T46, T5j, T2U; cannam@95: T7 = VSUB(T3, T6); cannam@95: T2T = VADD(T3, T6); cannam@95: T46 = VADD(T44, T45); cannam@95: T5j = VSUB(T44, T45); cannam@95: T26 = VSUB(T22, T25); cannam@95: T2U = VADD(T22, T25); cannam@95: Ta = VADD(T8, T9); cannam@95: T48 = VSUB(T8, T9); cannam@95: Tk = VADD(Ti, Tj); cannam@95: T4c = VSUB(Tj, Ti); cannam@95: T5k = VFMA(LDK(KP707106781), T5j, T5i); cannam@95: T6A = VFNMS(LDK(KP707106781), T5j, T5i); cannam@95: T47 = VFMA(LDK(KP707106781), T46, T43); cannam@95: T69 = VFNMS(LDK(KP707106781), T46, T43); cannam@95: T2V = VADD(T2T, T2U); cannam@95: T3z = VSUB(T2T, T2U); cannam@95: T49 = VSUB(Tb, Tc); cannam@95: Td = VADD(Tb, Tc); cannam@95: } cannam@95: Tf = LD(&(xi[WS(is, 60)]), ivs, &(xi[0])); cannam@95: Tg = LD(&(xi[WS(is, 28)]), ivs, &(xi[0])); cannam@95: } cannam@95: { cannam@95: V Te, T2W, T5l, T4a, Tq, Tt, Tv, Tw, T5m, T4d, Tl, T2X, Ty, Tz, To; cannam@95: V Tp; cannam@95: To = LD(&(xi[WS(is, 2)]), ivs, &(xi[0])); cannam@95: Tp = LD(&(xi[WS(is, 34)]), ivs, &(xi[0])); cannam@95: { cannam@95: V Th, T4b, Tr, Ts; cannam@95: Tr = LD(&(xi[WS(is, 18)]), ivs, &(xi[0])); cannam@95: Ts = LD(&(xi[WS(is, 50)]), ivs, &(xi[0])); cannam@95: Te = VSUB(Ta, Td); cannam@95: T2W = VADD(Ta, Td); cannam@95: T5l = VFMA(LDK(KP414213562), T48, T49); cannam@95: T4a = VFNMS(LDK(KP414213562), T49, T48); cannam@95: Th = VADD(Tf, Tg); cannam@95: T4b = VSUB(Tf, Tg); cannam@95: Tq = VADD(To, Tp); cannam@95: T4g = VSUB(To, Tp); cannam@95: T4l = VSUB(Tr, Ts); cannam@95: Tt = VADD(Tr, Ts); cannam@95: Tv = LD(&(xi[WS(is, 10)]), ivs, &(xi[0])); cannam@95: Tw = LD(&(xi[WS(is, 42)]), ivs, &(xi[0])); cannam@95: T5m = VFMA(LDK(KP414213562), T4b, T4c); cannam@95: T4d = VFNMS(LDK(KP414213562), T4c, T4b); cannam@95: Tl = VSUB(Th, Tk); cannam@95: T2X = VADD(Th, Tk); cannam@95: Ty = LD(&(xi[WS(is, 58)]), ivs, &(xi[0])); cannam@95: Tz = LD(&(xi[WS(is, 26)]), ivs, &(xi[0])); cannam@95: } cannam@95: T3g = VADD(Tq, Tt); cannam@95: Tu = VSUB(Tq, Tt); cannam@95: Tx = VADD(Tv, Tw); cannam@95: T4h = VSUB(Tv, Tw); cannam@95: T6B = VSUB(T4a, T4d); cannam@95: T4e = VADD(T4a, T4d); cannam@95: T6a = VADD(T5l, T5m); cannam@95: T5n = VSUB(T5l, T5m); cannam@95: T3M = VSUB(T2W, T2X); cannam@95: T2Y = VADD(T2W, T2X); cannam@95: T27 = VSUB(Te, Tl); cannam@95: Tm = VADD(Te, Tl); cannam@95: TA = VADD(Ty, Tz); cannam@95: T4i = VSUB(Ty, Tz); cannam@95: } cannam@95: } cannam@95: } cannam@95: { cannam@95: V TK, T4p, T4u, T4k, T6d, T4n, T6c, TL, TN, TO, T3j, TJ, TF, TI; cannam@95: { cannam@95: V TD, TE, TG, TH; cannam@95: TD = LD(&(xi[WS(is, 62)]), ivs, &(xi[0])); cannam@95: TE = LD(&(xi[WS(is, 30)]), ivs, &(xi[0])); cannam@95: TG = LD(&(xi[WS(is, 14)]), ivs, &(xi[0])); cannam@95: TH = LD(&(xi[WS(is, 46)]), ivs, &(xi[0])); cannam@95: TK = LD(&(xi[WS(is, 54)]), ivs, &(xi[0])); cannam@95: { cannam@95: V T3h, TB, T4j, T4m; cannam@95: T3h = VADD(Tx, TA); cannam@95: TB = VSUB(Tx, TA); cannam@95: T4j = VADD(T4h, T4i); cannam@95: T4m = VSUB(T4h, T4i); cannam@95: T4p = VSUB(TD, TE); cannam@95: TF = VADD(TD, TE); cannam@95: T4u = VSUB(TH, TG); cannam@95: TI = VADD(TG, TH); cannam@95: T3A = VSUB(T3g, T3h); cannam@95: T3i = VADD(T3g, T3h); cannam@95: T29 = VFMA(LDK(KP414213562), Tu, TB); cannam@95: TC = VFNMS(LDK(KP414213562), TB, Tu); cannam@95: T4k = VFMA(LDK(KP707106781), T4j, T4g); cannam@95: T6d = VFNMS(LDK(KP707106781), T4j, T4g); cannam@95: T4n = VFMA(LDK(KP707106781), T4m, T4l); cannam@95: T6c = VFNMS(LDK(KP707106781), T4m, T4l); cannam@95: TL = LD(&(xi[WS(is, 22)]), ivs, &(xi[0])); cannam@95: } cannam@95: TN = LD(&(xi[WS(is, 6)]), ivs, &(xi[0])); cannam@95: TO = LD(&(xi[WS(is, 38)]), ivs, &(xi[0])); cannam@95: } cannam@95: T3j = VADD(TF, TI); cannam@95: TJ = VSUB(TF, TI); cannam@95: { cannam@95: V T3a, T1E, T52, T5b, T1x, T4Z, T6r, T6u, T5a, T1U, T55, T5c, T1L, T3b; cannam@95: { cannam@95: V T4V, T1t, T58, T1w, T1Q, T1T, T1I, T4Y, T59, T1J, T53, T1H; cannam@95: { cannam@95: V T1r, TM, T4r, TP, T4q, T1s, T1u, T1v; cannam@95: T1r = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)])); cannam@95: T5p = VFMA(LDK(KP198912367), T4k, T4n); cannam@95: T4o = VFNMS(LDK(KP198912367), T4n, T4k); cannam@95: T6D = VFMA(LDK(KP668178637), T6c, T6d); cannam@95: T6e = VFNMS(LDK(KP668178637), T6d, T6c); cannam@95: TM = VADD(TK, TL); cannam@95: T4r = VSUB(TK, TL); cannam@95: TP = VADD(TN, TO); cannam@95: T4q = VSUB(TN, TO); cannam@95: T1s = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1u = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1v = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)])); cannam@95: { cannam@95: V T1R, T4X, T6g, T4t, T6f, T4w, T1S, T1O, T1P; cannam@95: T1O = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1P = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1R = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)])); cannam@95: { cannam@95: V T3k, TQ, T4s, T4v; cannam@95: T3k = VADD(TP, TM); cannam@95: TQ = VSUB(TM, TP); cannam@95: T4s = VADD(T4q, T4r); cannam@95: T4v = VSUB(T4r, T4q); cannam@95: T4V = VSUB(T1r, T1s); cannam@95: T1t = VADD(T1r, T1s); cannam@95: T58 = VSUB(T1v, T1u); cannam@95: T1w = VADD(T1u, T1v); cannam@95: T4X = VSUB(T1O, T1P); cannam@95: T1Q = VADD(T1O, T1P); cannam@95: T3l = VADD(T3j, T3k); cannam@95: T3B = VSUB(T3j, T3k); cannam@95: TR = VFNMS(LDK(KP414213562), TQ, TJ); cannam@95: T2a = VFMA(LDK(KP414213562), TJ, TQ); cannam@95: T6g = VFNMS(LDK(KP707106781), T4s, T4p); cannam@95: T4t = VFMA(LDK(KP707106781), T4s, T4p); cannam@95: T6f = VFNMS(LDK(KP707106781), T4v, T4u); cannam@95: T4w = VFMA(LDK(KP707106781), T4v, T4u); cannam@95: T1S = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)])); cannam@95: } cannam@95: { cannam@95: V T4W, T1A, T50, T51, T1D, T1F, T1G; cannam@95: { cannam@95: V T1y, T1z, T1B, T1C; cannam@95: T1y = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1z = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1B = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1C = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)])); cannam@95: T4x = VFNMS(LDK(KP198912367), T4w, T4t); cannam@95: T5q = VFMA(LDK(KP198912367), T4t, T4w); cannam@95: T6h = VFNMS(LDK(KP668178637), T6g, T6f); cannam@95: T6E = VFMA(LDK(KP668178637), T6f, T6g); cannam@95: T4W = VSUB(T1R, T1S); cannam@95: T1T = VADD(T1R, T1S); cannam@95: T1A = VADD(T1y, T1z); cannam@95: T50 = VSUB(T1y, T1z); cannam@95: T51 = VSUB(T1C, T1B); cannam@95: T1D = VADD(T1B, T1C); cannam@95: } cannam@95: T1F = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1G = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1I = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)])); cannam@95: T4Y = VADD(T4W, T4X); cannam@95: T59 = VSUB(T4X, T4W); cannam@95: T1J = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)])); cannam@95: T3a = VADD(T1A, T1D); cannam@95: T1E = VSUB(T1A, T1D); cannam@95: T52 = VFMA(LDK(KP414213562), T51, T50); cannam@95: T5b = VFNMS(LDK(KP414213562), T50, T51); cannam@95: T53 = VSUB(T1F, T1G); cannam@95: T1H = VADD(T1F, T1G); cannam@95: } cannam@95: } cannam@95: } cannam@95: { cannam@95: V T37, T54, T1K, T38; cannam@95: T1x = VSUB(T1t, T1w); cannam@95: T37 = VADD(T1t, T1w); cannam@95: T4Z = VFMA(LDK(KP707106781), T4Y, T4V); cannam@95: T6r = VFNMS(LDK(KP707106781), T4Y, T4V); cannam@95: T54 = VSUB(T1J, T1I); cannam@95: T1K = VADD(T1I, T1J); cannam@95: T6u = VFNMS(LDK(KP707106781), T59, T58); cannam@95: T5a = VFMA(LDK(KP707106781), T59, T58); cannam@95: T38 = VADD(T1T, T1Q); cannam@95: T1U = VSUB(T1Q, T1T); cannam@95: T55 = VFNMS(LDK(KP414213562), T54, T53); cannam@95: T5c = VFMA(LDK(KP414213562), T53, T54); cannam@95: T1L = VSUB(T1H, T1K); cannam@95: T3b = VADD(T1H, T1K); cannam@95: T39 = VADD(T37, T38); cannam@95: T3H = VSUB(T37, T38); cannam@95: } cannam@95: } cannam@95: { cannam@95: V T4A, TW, T4N, TZ, T1j, T1m, T4O, T4D, T13, T4F, T16, T4G, T1a, T4I, T4J; cannam@95: V T1d; cannam@95: { cannam@95: V TU, TV, TX, TY, T56, T6v; cannam@95: TU = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)])); cannam@95: T56 = VADD(T52, T55); cannam@95: T6v = VSUB(T55, T52); cannam@95: { cannam@95: V T5d, T6s, T1V, T1M; cannam@95: T5d = VADD(T5b, T5c); cannam@95: T6s = VSUB(T5c, T5b); cannam@95: T1V = VSUB(T1L, T1E); cannam@95: T1M = VADD(T1E, T1L); cannam@95: T3I = VSUB(T3b, T3a); cannam@95: T3c = VADD(T3a, T3b); cannam@95: T5N = VFNMS(LDK(KP923879532), T56, T4Z); cannam@95: T57 = VFMA(LDK(KP923879532), T56, T4Z); cannam@95: T72 = VFNMS(LDK(KP923879532), T6v, T6u); cannam@95: T6w = VFMA(LDK(KP923879532), T6v, T6u); cannam@95: T5O = VFNMS(LDK(KP923879532), T5d, T5a); cannam@95: T5e = VFMA(LDK(KP923879532), T5d, T5a); cannam@95: T71 = VFMA(LDK(KP923879532), T6s, T6r); cannam@95: T6t = VFNMS(LDK(KP923879532), T6s, T6r); cannam@95: T2y = VFNMS(LDK(KP707106781), T1V, T1U); cannam@95: T1W = VFMA(LDK(KP707106781), T1V, T1U); cannam@95: T2x = VFNMS(LDK(KP707106781), T1M, T1x); cannam@95: T1N = VFMA(LDK(KP707106781), T1M, T1x); cannam@95: TV = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)])); cannam@95: } cannam@95: TX = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)])); cannam@95: TY = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)])); cannam@95: { cannam@95: V T1h, T1i, T1k, T1l; cannam@95: T1h = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1i = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1k = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1l = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)])); cannam@95: { cannam@95: V T11, T4B, T4C, T12, T14, T15; cannam@95: T11 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)])); cannam@95: T4A = VSUB(TU, TV); cannam@95: TW = VADD(TU, TV); cannam@95: T4N = VSUB(TX, TY); cannam@95: TZ = VADD(TX, TY); cannam@95: T1j = VADD(T1h, T1i); cannam@95: T4B = VSUB(T1h, T1i); cannam@95: T1m = VADD(T1k, T1l); cannam@95: T4C = VSUB(T1k, T1l); cannam@95: T12 = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)])); cannam@95: T14 = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)])); cannam@95: T15 = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)])); cannam@95: { cannam@95: V T18, T19, T1b, T1c; cannam@95: T18 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)])); cannam@95: T19 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1b = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1c = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)])); cannam@95: T4O = VSUB(T4B, T4C); cannam@95: T4D = VADD(T4B, T4C); cannam@95: T13 = VADD(T11, T12); cannam@95: T4F = VSUB(T11, T12); cannam@95: T16 = VADD(T14, T15); cannam@95: T4G = VSUB(T14, T15); cannam@95: T1a = VADD(T18, T19); cannam@95: T4I = VSUB(T18, T19); cannam@95: T4J = VSUB(T1b, T1c); cannam@95: T1d = VADD(T1b, T1c); cannam@95: } cannam@95: } cannam@95: } cannam@95: } cannam@95: { cannam@95: V T30, T10, T6k, T4E, T4Q, T4H, T17, T6n, T4P, T1e, T4K, T4R, T1n, T31; cannam@95: T30 = VADD(TW, TZ); cannam@95: T10 = VSUB(TW, TZ); cannam@95: T6k = VFNMS(LDK(KP707106781), T4D, T4A); cannam@95: T4E = VFMA(LDK(KP707106781), T4D, T4A); cannam@95: T4Q = VFMA(LDK(KP414213562), T4F, T4G); cannam@95: T4H = VFNMS(LDK(KP414213562), T4G, T4F); cannam@95: T33 = VADD(T13, T16); cannam@95: T17 = VSUB(T13, T16); cannam@95: T6n = VFNMS(LDK(KP707106781), T4O, T4N); cannam@95: T4P = VFMA(LDK(KP707106781), T4O, T4N); cannam@95: T34 = VADD(T1a, T1d); cannam@95: T1e = VSUB(T1a, T1d); cannam@95: T4K = VFMA(LDK(KP414213562), T4J, T4I); cannam@95: T4R = VFNMS(LDK(KP414213562), T4I, T4J); cannam@95: T1n = VSUB(T1j, T1m); cannam@95: T31 = VADD(T1j, T1m); cannam@95: { cannam@95: V T1f, T1o, T6o, T4L, T4S, T6l; cannam@95: T1f = VADD(T17, T1e); cannam@95: T1o = VSUB(T17, T1e); cannam@95: T6o = VSUB(T4H, T4K); cannam@95: T4L = VADD(T4H, T4K); cannam@95: T4S = VADD(T4Q, T4R); cannam@95: T6l = VSUB(T4Q, T4R); cannam@95: T3E = VSUB(T30, T31); cannam@95: T32 = VADD(T30, T31); cannam@95: T1p = VFMA(LDK(KP707106781), T1o, T1n); cannam@95: T2v = VFNMS(LDK(KP707106781), T1o, T1n); cannam@95: T1g = VFMA(LDK(KP707106781), T1f, T10); cannam@95: T2u = VFNMS(LDK(KP707106781), T1f, T10); cannam@95: T4M = VFMA(LDK(KP923879532), T4L, T4E); cannam@95: T5K = VFNMS(LDK(KP923879532), T4L, T4E); cannam@95: T6p = VFMA(LDK(KP923879532), T6o, T6n); cannam@95: T6Z = VFNMS(LDK(KP923879532), T6o, T6n); cannam@95: T6m = VFNMS(LDK(KP923879532), T6l, T6k); cannam@95: T6Y = VFMA(LDK(KP923879532), T6l, T6k); cannam@95: T5L = VFNMS(LDK(KP923879532), T4S, T4P); cannam@95: T4T = VFMA(LDK(KP923879532), T4S, T4P); cannam@95: } cannam@95: } cannam@95: } cannam@95: } cannam@95: } cannam@95: } cannam@95: { cannam@95: V T6b, T6F, T7n, T7o, T7p, T7q, T7r, T7s, T7t, T7u, T7v, T7w, T7x, T7y, T7f; cannam@95: V T6X, T70, T79, T7a, T73, T6C, T76, T77, T6i; cannam@95: { cannam@95: V T2Z, T3r, T3s, T3m, T3d, T3v; cannam@95: T2Z = VSUB(T2V, T2Y); cannam@95: T3r = VADD(T2V, T2Y); cannam@95: T3s = VADD(T3i, T3l); cannam@95: T3m = VSUB(T3i, T3l); cannam@95: T3d = VSUB(T39, T3c); cannam@95: T3v = VADD(T39, T3c); cannam@95: { cannam@95: V T3x, T3t, T3Q, T3J, T3D, T3V, T3G, T3P, T3u, T36, T3O, T3Y, T6V, T6W; cannam@95: { cannam@95: V T3N, T3C, T3F, T35; cannam@95: T3N = VSUB(T3A, T3B); cannam@95: T3C = VADD(T3A, T3B); cannam@95: T3F = VSUB(T33, T34); cannam@95: T35 = VADD(T33, T34); cannam@95: T3x = VADD(T3r, T3s); cannam@95: T3t = VSUB(T3r, T3s); cannam@95: T3Q = VFMA(LDK(KP414213562), T3H, T3I); cannam@95: T3J = VFNMS(LDK(KP414213562), T3I, T3H); cannam@95: T3D = VFMA(LDK(KP707106781), T3C, T3z); cannam@95: T3V = VFNMS(LDK(KP707106781), T3C, T3z); cannam@95: T3G = VFNMS(LDK(KP414213562), T3F, T3E); cannam@95: T3P = VFMA(LDK(KP414213562), T3E, T3F); cannam@95: T3u = VADD(T32, T35); cannam@95: T36 = VSUB(T32, T35); cannam@95: T3O = VFMA(LDK(KP707106781), T3N, T3M); cannam@95: T3Y = VFNMS(LDK(KP707106781), T3N, T3M); cannam@95: } cannam@95: T6b = VFNMS(LDK(KP923879532), T6a, T69); cannam@95: T6V = VFMA(LDK(KP923879532), T6a, T69); cannam@95: T6W = VADD(T6D, T6E); cannam@95: T6F = VSUB(T6D, T6E); cannam@95: { cannam@95: V T3R, T3W, T3K, T3Z; cannam@95: T3R = VSUB(T3P, T3Q); cannam@95: T3W = VADD(T3P, T3Q); cannam@95: T3K = VADD(T3G, T3J); cannam@95: T3Z = VSUB(T3G, T3J); cannam@95: { cannam@95: V T3e, T3n, T3w, T3y; cannam@95: T3e = VADD(T36, T3d); cannam@95: T3n = VSUB(T36, T3d); cannam@95: T3w = VSUB(T3u, T3v); cannam@95: T3y = VADD(T3u, T3v); cannam@95: { cannam@95: V T41, T3X, T3S, T3U; cannam@95: T41 = VFMA(LDK(KP923879532), T3W, T3V); cannam@95: T3X = VFNMS(LDK(KP923879532), T3W, T3V); cannam@95: T3S = VFNMS(LDK(KP923879532), T3R, T3O); cannam@95: T3U = VFMA(LDK(KP923879532), T3R, T3O); cannam@95: { cannam@95: V T42, T40, T3L, T3T; cannam@95: T42 = VFNMS(LDK(KP923879532), T3Z, T3Y); cannam@95: T40 = VFMA(LDK(KP923879532), T3Z, T3Y); cannam@95: T3L = VFNMS(LDK(KP923879532), T3K, T3D); cannam@95: T3T = VFMA(LDK(KP923879532), T3K, T3D); cannam@95: { cannam@95: V T3o, T3q, T3f, T3p; cannam@95: T3o = VFNMS(LDK(KP707106781), T3n, T3m); cannam@95: T3q = VFMA(LDK(KP707106781), T3n, T3m); cannam@95: T3f = VFNMS(LDK(KP707106781), T3e, T2Z); cannam@95: T3p = VFMA(LDK(KP707106781), T3e, T2Z); cannam@95: T7n = VSUB(T3x, T3y); cannam@95: STM2(&(xo[64]), T7n, ovs, &(xo[0])); cannam@95: T7o = VADD(T3x, T3y); cannam@95: STM2(&(xo[0]), T7o, ovs, &(xo[0])); cannam@95: T7p = VFMAI(T3w, T3t); cannam@95: STM2(&(xo[32]), T7p, ovs, &(xo[0])); cannam@95: T7q = VFNMSI(T3w, T3t); cannam@95: STM2(&(xo[96]), T7q, ovs, &(xo[0])); cannam@95: T7r = VFNMSI(T40, T3X); cannam@95: STM2(&(xo[88]), T7r, ovs, &(xo[0])); cannam@95: T7s = VFMAI(T40, T3X); cannam@95: STM2(&(xo[40]), T7s, ovs, &(xo[0])); cannam@95: T7t = VFMAI(T42, T41); cannam@95: STM2(&(xo[104]), T7t, ovs, &(xo[0])); cannam@95: T7u = VFNMSI(T42, T41); cannam@95: STM2(&(xo[24]), T7u, ovs, &(xo[0])); cannam@95: T7v = VFMAI(T3U, T3T); cannam@95: STM2(&(xo[8]), T7v, ovs, &(xo[0])); cannam@95: T7w = VFNMSI(T3U, T3T); cannam@95: STM2(&(xo[120]), T7w, ovs, &(xo[0])); cannam@95: T7x = VFMAI(T3S, T3L); cannam@95: STM2(&(xo[72]), T7x, ovs, &(xo[0])); cannam@95: T7y = VFNMSI(T3S, T3L); cannam@95: STM2(&(xo[56]), T7y, ovs, &(xo[0])); cannam@95: T7z = VFNMSI(T3q, T3p); cannam@95: STM2(&(xo[112]), T7z, ovs, &(xo[0])); cannam@95: T7A = VFMAI(T3q, T3p); cannam@95: STM2(&(xo[16]), T7A, ovs, &(xo[0])); cannam@95: T7B = VFMAI(T3o, T3f); cannam@95: STM2(&(xo[80]), T7B, ovs, &(xo[0])); cannam@95: T7C = VFNMSI(T3o, T3f); cannam@95: STM2(&(xo[48]), T7C, ovs, &(xo[0])); cannam@95: T7f = VFNMS(LDK(KP831469612), T6W, T6V); cannam@95: T6X = VFMA(LDK(KP831469612), T6W, T6V); cannam@95: } cannam@95: } cannam@95: } cannam@95: } cannam@95: } cannam@95: T70 = VFMA(LDK(KP303346683), T6Z, T6Y); cannam@95: T79 = VFNMS(LDK(KP303346683), T6Y, T6Z); cannam@95: T7a = VFNMS(LDK(KP303346683), T71, T72); cannam@95: T73 = VFMA(LDK(KP303346683), T72, T71); cannam@95: T6C = VFMA(LDK(KP923879532), T6B, T6A); cannam@95: T76 = VFNMS(LDK(KP923879532), T6B, T6A); cannam@95: T77 = VSUB(T6e, T6h); cannam@95: T6i = VADD(T6e, T6h); cannam@95: } cannam@95: } cannam@95: { cannam@95: V T2r, T2D, T2C, T2s, T5H, T5o, T5v, T5D, T7L, T7O, T7Q, T7S, T5r, T5I, T5x; cannam@95: V T5h, T5F, T5B; cannam@95: { cannam@95: V TT, T2f, T7E, T7F, T7I, T7K, T2n, T1Y, T28, T2b, T2l, T2p, T2j, T2k; cannam@95: { cannam@95: V T1q, T2d, T7h, T7l, T2e, T1X, T75, T7d, T7m, T7k, T7c, T7e, Tn, TS; cannam@95: T2r = VFNMS(LDK(KP707106781), Tm, T7); cannam@95: Tn = VFMA(LDK(KP707106781), Tm, T7); cannam@95: TS = VADD(TC, TR); cannam@95: T2D = VSUB(TC, TR); cannam@95: { cannam@95: V T7b, T7j, T74, T7i, T78, T7g; cannam@95: T1q = VFNMS(LDK(KP198912367), T1p, T1g); cannam@95: T2d = VFMA(LDK(KP198912367), T1g, T1p); cannam@95: T7g = VADD(T79, T7a); cannam@95: T7b = VSUB(T79, T7a); cannam@95: T7j = VSUB(T70, T73); cannam@95: T74 = VADD(T70, T73); cannam@95: T7i = VFNMS(LDK(KP831469612), T77, T76); cannam@95: T78 = VFMA(LDK(KP831469612), T77, T76); cannam@95: T2j = VFNMS(LDK(KP923879532), TS, Tn); cannam@95: TT = VFMA(LDK(KP923879532), TS, Tn); cannam@95: T7h = VFMA(LDK(KP956940335), T7g, T7f); cannam@95: T7l = VFNMS(LDK(KP956940335), T7g, T7f); cannam@95: T2e = VFMA(LDK(KP198912367), T1N, T1W); cannam@95: T1X = VFNMS(LDK(KP198912367), T1W, T1N); cannam@95: T75 = VFNMS(LDK(KP956940335), T74, T6X); cannam@95: T7d = VFMA(LDK(KP956940335), T74, T6X); cannam@95: T7m = VFMA(LDK(KP956940335), T7j, T7i); cannam@95: T7k = VFNMS(LDK(KP956940335), T7j, T7i); cannam@95: T7c = VFNMS(LDK(KP956940335), T7b, T78); cannam@95: T7e = VFMA(LDK(KP956940335), T7b, T78); cannam@95: } cannam@95: T2k = VADD(T2d, T2e); cannam@95: T2f = VSUB(T2d, T2e); cannam@95: { cannam@95: V T7D, T7G, T7H, T7J; cannam@95: T7D = VFMAI(T7k, T7h); cannam@95: STM2(&(xo[90]), T7D, ovs, &(xo[2])); cannam@95: STN2(&(xo[88]), T7r, T7D, ovs); cannam@95: T7E = VFNMSI(T7k, T7h); cannam@95: STM2(&(xo[38]), T7E, ovs, &(xo[2])); cannam@95: T7F = VFNMSI(T7m, T7l); cannam@95: STM2(&(xo[102]), T7F, ovs, &(xo[2])); cannam@95: T7G = VFMAI(T7m, T7l); cannam@95: STM2(&(xo[26]), T7G, ovs, &(xo[2])); cannam@95: STN2(&(xo[24]), T7u, T7G, ovs); cannam@95: T7H = VFMAI(T7e, T7d); cannam@95: STM2(&(xo[122]), T7H, ovs, &(xo[2])); cannam@95: STN2(&(xo[120]), T7w, T7H, ovs); cannam@95: T7I = VFNMSI(T7e, T7d); cannam@95: STM2(&(xo[6]), T7I, ovs, &(xo[2])); cannam@95: T7J = VFMAI(T7c, T75); cannam@95: STM2(&(xo[58]), T7J, ovs, &(xo[2])); cannam@95: STN2(&(xo[56]), T7y, T7J, ovs); cannam@95: T7K = VFNMSI(T7c, T75); cannam@95: STM2(&(xo[70]), T7K, ovs, &(xo[2])); cannam@95: T2n = VSUB(T1q, T1X); cannam@95: T1Y = VADD(T1q, T1X); cannam@95: } cannam@95: T2C = VFNMS(LDK(KP707106781), T27, T26); cannam@95: T28 = VFMA(LDK(KP707106781), T27, T26); cannam@95: T2b = VSUB(T29, T2a); cannam@95: T2s = VADD(T29, T2a); cannam@95: } cannam@95: T2l = VFNMS(LDK(KP980785280), T2k, T2j); cannam@95: T2p = VFMA(LDK(KP980785280), T2k, T2j); cannam@95: { cannam@95: V T5z, T4z, T5A, T5g; cannam@95: { cannam@95: V T4f, T4y, T1Z, T2h, T4U, T5t, T2m, T2c, T5u, T5f; cannam@95: T5H = VFNMS(LDK(KP923879532), T4e, T47); cannam@95: T4f = VFMA(LDK(KP923879532), T4e, T47); cannam@95: T4y = VADD(T4o, T4x); cannam@95: T5T = VSUB(T4o, T4x); cannam@95: T1Z = VFNMS(LDK(KP980785280), T1Y, TT); cannam@95: T2h = VFMA(LDK(KP980785280), T1Y, TT); cannam@95: T4U = VFNMS(LDK(KP098491403), T4T, T4M); cannam@95: T5t = VFMA(LDK(KP098491403), T4M, T4T); cannam@95: T2m = VFNMS(LDK(KP923879532), T2b, T28); cannam@95: T2c = VFMA(LDK(KP923879532), T2b, T28); cannam@95: T5u = VFMA(LDK(KP098491403), T57, T5e); cannam@95: T5f = VFNMS(LDK(KP098491403), T5e, T57); cannam@95: T5z = VFNMS(LDK(KP980785280), T4y, T4f); cannam@95: T4z = VFMA(LDK(KP980785280), T4y, T4f); cannam@95: T5S = VFNMS(LDK(KP923879532), T5n, T5k); cannam@95: T5o = VFMA(LDK(KP923879532), T5n, T5k); cannam@95: { cannam@95: V T2o, T2q, T2i, T2g; cannam@95: T2o = VFMA(LDK(KP980785280), T2n, T2m); cannam@95: T2q = VFNMS(LDK(KP980785280), T2n, T2m); cannam@95: T2i = VFMA(LDK(KP980785280), T2f, T2c); cannam@95: T2g = VFNMS(LDK(KP980785280), T2f, T2c); cannam@95: T5A = VADD(T5t, T5u); cannam@95: T5v = VSUB(T5t, T5u); cannam@95: T5D = VSUB(T4U, T5f); cannam@95: T5g = VADD(T4U, T5f); cannam@95: T7L = VFNMSI(T2o, T2l); cannam@95: STM2(&(xo[92]), T7L, ovs, &(xo[0])); cannam@95: { cannam@95: V T7M, T7N, T7P, T7R; cannam@95: T7M = VFMAI(T2o, T2l); cannam@95: STM2(&(xo[36]), T7M, ovs, &(xo[0])); cannam@95: STN2(&(xo[36]), T7M, T7E, ovs); cannam@95: T7N = VFMAI(T2q, T2p); cannam@95: STM2(&(xo[100]), T7N, ovs, &(xo[0])); cannam@95: STN2(&(xo[100]), T7N, T7F, ovs); cannam@95: T7O = VFNMSI(T2q, T2p); cannam@95: STM2(&(xo[28]), T7O, ovs, &(xo[0])); cannam@95: T7P = VFMAI(T2i, T2h); cannam@95: STM2(&(xo[4]), T7P, ovs, &(xo[0])); cannam@95: STN2(&(xo[4]), T7P, T7I, ovs); cannam@95: T7Q = VFNMSI(T2i, T2h); cannam@95: STM2(&(xo[124]), T7Q, ovs, &(xo[0])); cannam@95: T7R = VFMAI(T2g, T1Z); cannam@95: STM2(&(xo[68]), T7R, ovs, &(xo[0])); cannam@95: STN2(&(xo[68]), T7R, T7K, ovs); cannam@95: T7S = VFNMSI(T2g, T1Z); cannam@95: STM2(&(xo[60]), T7S, ovs, &(xo[0])); cannam@95: T5r = VSUB(T5p, T5q); cannam@95: T5I = VADD(T5p, T5q); cannam@95: } cannam@95: } cannam@95: } cannam@95: T5x = VFMA(LDK(KP995184726), T5g, T4z); cannam@95: T5h = VFNMS(LDK(KP995184726), T5g, T4z); cannam@95: T5F = VFMA(LDK(KP995184726), T5A, T5z); cannam@95: T5B = VFNMS(LDK(KP995184726), T5A, T5z); cannam@95: } cannam@95: } cannam@95: { cannam@95: V T6J, T6R, T6L, T6z, T6T, T6P; cannam@95: { cannam@95: V T6N, T6j, T6O, T6y; cannam@95: { cannam@95: V T6q, T6H, T5C, T5s, T6I, T6x; cannam@95: T6q = VFNMS(LDK(KP534511135), T6p, T6m); cannam@95: T6H = VFMA(LDK(KP534511135), T6m, T6p); cannam@95: T5C = VFNMS(LDK(KP980785280), T5r, T5o); cannam@95: T5s = VFMA(LDK(KP980785280), T5r, T5o); cannam@95: T6I = VFMA(LDK(KP534511135), T6t, T6w); cannam@95: T6x = VFNMS(LDK(KP534511135), T6w, T6t); cannam@95: T6N = VFMA(LDK(KP831469612), T6i, T6b); cannam@95: T6j = VFNMS(LDK(KP831469612), T6i, T6b); cannam@95: { cannam@95: V T5E, T5G, T5y, T5w; cannam@95: T5E = VFMA(LDK(KP995184726), T5D, T5C); cannam@95: T5G = VFNMS(LDK(KP995184726), T5D, T5C); cannam@95: T5y = VFMA(LDK(KP995184726), T5v, T5s); cannam@95: T5w = VFNMS(LDK(KP995184726), T5v, T5s); cannam@95: T6O = VADD(T6H, T6I); cannam@95: T6J = VSUB(T6H, T6I); cannam@95: T6R = VSUB(T6q, T6x); cannam@95: T6y = VADD(T6q, T6x); cannam@95: { cannam@95: V T7T, T7U, T7V, T7W; cannam@95: T7T = VFNMSI(T5E, T5B); cannam@95: STM2(&(xo[94]), T7T, ovs, &(xo[2])); cannam@95: STN2(&(xo[92]), T7L, T7T, ovs); cannam@95: T7U = VFMAI(T5E, T5B); cannam@95: STM2(&(xo[34]), T7U, ovs, &(xo[2])); cannam@95: STN2(&(xo[32]), T7p, T7U, ovs); cannam@95: T7V = VFMAI(T5G, T5F); cannam@95: STM2(&(xo[98]), T7V, ovs, &(xo[2])); cannam@95: STN2(&(xo[96]), T7q, T7V, ovs); cannam@95: T7W = VFNMSI(T5G, T5F); cannam@95: STM2(&(xo[30]), T7W, ovs, &(xo[2])); cannam@95: STN2(&(xo[28]), T7O, T7W, ovs); cannam@95: { cannam@95: V T7X, T7Y, T7Z, T80; cannam@95: T7X = VFMAI(T5y, T5x); cannam@95: STM2(&(xo[2]), T7X, ovs, &(xo[2])); cannam@95: STN2(&(xo[0]), T7o, T7X, ovs); cannam@95: T7Y = VFNMSI(T5y, T5x); cannam@95: STM2(&(xo[126]), T7Y, ovs, &(xo[2])); cannam@95: STN2(&(xo[124]), T7Q, T7Y, ovs); cannam@95: T7Z = VFMAI(T5w, T5h); cannam@95: STM2(&(xo[66]), T7Z, ovs, &(xo[2])); cannam@95: STN2(&(xo[64]), T7n, T7Z, ovs); cannam@95: T80 = VFNMSI(T5w, T5h); cannam@95: STM2(&(xo[62]), T80, ovs, &(xo[2])); cannam@95: STN2(&(xo[60]), T7S, T80, ovs); cannam@95: } cannam@95: } cannam@95: } cannam@95: } cannam@95: T6L = VFMA(LDK(KP881921264), T6y, T6j); cannam@95: T6z = VFNMS(LDK(KP881921264), T6y, T6j); cannam@95: T6T = VFMA(LDK(KP881921264), T6O, T6N); cannam@95: T6P = VFNMS(LDK(KP881921264), T6O, T6N); cannam@95: } cannam@95: { cannam@95: V T2H, T2P, T81, T84, T86, T88, T2J, T2B, T2R, T2N; cannam@95: { cannam@95: V T2L, T2t, T2M, T2A; cannam@95: { cannam@95: V T2w, T2F, T6Q, T6G, T2G, T2z; cannam@95: T2w = VFMA(LDK(KP668178637), T2v, T2u); cannam@95: T2F = VFNMS(LDK(KP668178637), T2u, T2v); cannam@95: T6Q = VFNMS(LDK(KP831469612), T6F, T6C); cannam@95: T6G = VFMA(LDK(KP831469612), T6F, T6C); cannam@95: T2G = VFNMS(LDK(KP668178637), T2x, T2y); cannam@95: T2z = VFMA(LDK(KP668178637), T2y, T2x); cannam@95: T2L = VFNMS(LDK(KP923879532), T2s, T2r); cannam@95: T2t = VFMA(LDK(KP923879532), T2s, T2r); cannam@95: { cannam@95: V T6S, T6U, T6M, T6K; cannam@95: T6S = VFMA(LDK(KP881921264), T6R, T6Q); cannam@95: T6U = VFNMS(LDK(KP881921264), T6R, T6Q); cannam@95: T6M = VFMA(LDK(KP881921264), T6J, T6G); cannam@95: T6K = VFNMS(LDK(KP881921264), T6J, T6G); cannam@95: T2M = VADD(T2F, T2G); cannam@95: T2H = VSUB(T2F, T2G); cannam@95: T2P = VSUB(T2w, T2z); cannam@95: T2A = VADD(T2w, T2z); cannam@95: T81 = VFNMSI(T6S, T6P); cannam@95: STM2(&(xo[86]), T81, ovs, &(xo[2])); cannam@95: { cannam@95: V T82, T83, T85, T87; cannam@95: T82 = VFMAI(T6S, T6P); cannam@95: STM2(&(xo[42]), T82, ovs, &(xo[2])); cannam@95: STN2(&(xo[40]), T7s, T82, ovs); cannam@95: T83 = VFMAI(T6U, T6T); cannam@95: STM2(&(xo[106]), T83, ovs, &(xo[2])); cannam@95: STN2(&(xo[104]), T7t, T83, ovs); cannam@95: T84 = VFNMSI(T6U, T6T); cannam@95: STM2(&(xo[22]), T84, ovs, &(xo[2])); cannam@95: T85 = VFMAI(T6M, T6L); cannam@95: STM2(&(xo[10]), T85, ovs, &(xo[2])); cannam@95: STN2(&(xo[8]), T7v, T85, ovs); cannam@95: T86 = VFNMSI(T6M, T6L); cannam@95: STM2(&(xo[118]), T86, ovs, &(xo[2])); cannam@95: T87 = VFMAI(T6K, T6z); cannam@95: STM2(&(xo[74]), T87, ovs, &(xo[2])); cannam@95: STN2(&(xo[72]), T7x, T87, ovs); cannam@95: T88 = VFNMSI(T6K, T6z); cannam@95: STM2(&(xo[54]), T88, ovs, &(xo[2])); cannam@95: } cannam@95: } cannam@95: } cannam@95: T2J = VFMA(LDK(KP831469612), T2A, T2t); cannam@95: T2B = VFNMS(LDK(KP831469612), T2A, T2t); cannam@95: T2R = VFNMS(LDK(KP831469612), T2M, T2L); cannam@95: T2N = VFMA(LDK(KP831469612), T2M, T2L); cannam@95: } cannam@95: { cannam@95: V T61, T5J, T62, T5Q; cannam@95: { cannam@95: V T5M, T5V, T2O, T2E, T5W, T5P; cannam@95: T5M = VFMA(LDK(KP820678790), T5L, T5K); cannam@95: T5V = VFNMS(LDK(KP820678790), T5K, T5L); cannam@95: T2O = VFMA(LDK(KP923879532), T2D, T2C); cannam@95: T2E = VFNMS(LDK(KP923879532), T2D, T2C); cannam@95: T5W = VFNMS(LDK(KP820678790), T5N, T5O); cannam@95: T5P = VFMA(LDK(KP820678790), T5O, T5N); cannam@95: T61 = VFNMS(LDK(KP980785280), T5I, T5H); cannam@95: T5J = VFMA(LDK(KP980785280), T5I, T5H); cannam@95: { cannam@95: V T2Q, T2S, T2K, T2I; cannam@95: T2Q = VFNMS(LDK(KP831469612), T2P, T2O); cannam@95: T2S = VFMA(LDK(KP831469612), T2P, T2O); cannam@95: T2K = VFMA(LDK(KP831469612), T2H, T2E); cannam@95: T2I = VFNMS(LDK(KP831469612), T2H, T2E); cannam@95: T62 = VADD(T5V, T5W); cannam@95: T5X = VSUB(T5V, T5W); cannam@95: T65 = VSUB(T5M, T5P); cannam@95: T5Q = VADD(T5M, T5P); cannam@95: { cannam@95: V T89, T8c, T8d, T8f; cannam@95: T89 = VFMAI(T2Q, T2N); cannam@95: STM2(&(xo[84]), T89, ovs, &(xo[0])); cannam@95: STN2(&(xo[84]), T89, T81, ovs); cannam@95: T8a = VFNMSI(T2Q, T2N); cannam@95: STM2(&(xo[44]), T8a, ovs, &(xo[0])); cannam@95: T8b = VFNMSI(T2S, T2R); cannam@95: STM2(&(xo[108]), T8b, ovs, &(xo[0])); cannam@95: T8c = VFMAI(T2S, T2R); cannam@95: STM2(&(xo[20]), T8c, ovs, &(xo[0])); cannam@95: STN2(&(xo[20]), T8c, T84, ovs); cannam@95: T8d = VFMAI(T2K, T2J); cannam@95: STM2(&(xo[116]), T8d, ovs, &(xo[0])); cannam@95: STN2(&(xo[116]), T8d, T86, ovs); cannam@95: T8e = VFNMSI(T2K, T2J); cannam@95: STM2(&(xo[12]), T8e, ovs, &(xo[0])); cannam@95: T8f = VFMAI(T2I, T2B); cannam@95: STM2(&(xo[52]), T8f, ovs, &(xo[0])); cannam@95: STN2(&(xo[52]), T8f, T88, ovs); cannam@95: T8g = VFNMSI(T2I, T2B); cannam@95: STM2(&(xo[76]), T8g, ovs, &(xo[0])); cannam@95: } cannam@95: } cannam@95: } cannam@95: T5Z = VFMA(LDK(KP773010453), T5Q, T5J); cannam@95: T5R = VFNMS(LDK(KP773010453), T5Q, T5J); cannam@95: T67 = VFNMS(LDK(KP773010453), T62, T61); cannam@95: T63 = VFMA(LDK(KP773010453), T62, T61); cannam@95: } cannam@95: } cannam@95: } cannam@95: } cannam@95: } cannam@95: } cannam@95: T5U = VFNMS(LDK(KP980785280), T5T, T5S); cannam@95: T64 = VFMA(LDK(KP980785280), T5T, T5S); cannam@95: { cannam@95: V T68, T66, T5Y, T60; cannam@95: T68 = VFMA(LDK(KP773010453), T65, T64); cannam@95: T66 = VFNMS(LDK(KP773010453), T65, T64); cannam@95: T5Y = VFNMS(LDK(KP773010453), T5X, T5U); cannam@95: T60 = VFMA(LDK(KP773010453), T5X, T5U); cannam@95: { cannam@95: V T8h, T8i, T8j, T8k; cannam@95: T8h = VFMAI(T66, T63); cannam@95: STM2(&(xo[82]), T8h, ovs, &(xo[2])); cannam@95: STN2(&(xo[80]), T7B, T8h, ovs); cannam@95: T8i = VFNMSI(T66, T63); cannam@95: STM2(&(xo[46]), T8i, ovs, &(xo[2])); cannam@95: STN2(&(xo[44]), T8a, T8i, ovs); cannam@95: T8j = VFNMSI(T68, T67); cannam@95: STM2(&(xo[110]), T8j, ovs, &(xo[2])); cannam@95: STN2(&(xo[108]), T8b, T8j, ovs); cannam@95: T8k = VFMAI(T68, T67); cannam@95: STM2(&(xo[18]), T8k, ovs, &(xo[2])); cannam@95: STN2(&(xo[16]), T7A, T8k, ovs); cannam@95: { cannam@95: V T8l, T8m, T8n, T8o; cannam@95: T8l = VFMAI(T60, T5Z); cannam@95: STM2(&(xo[114]), T8l, ovs, &(xo[2])); cannam@95: STN2(&(xo[112]), T7z, T8l, ovs); cannam@95: T8m = VFNMSI(T60, T5Z); cannam@95: STM2(&(xo[14]), T8m, ovs, &(xo[2])); cannam@95: STN2(&(xo[12]), T8e, T8m, ovs); cannam@95: T8n = VFMAI(T5Y, T5R); cannam@95: STM2(&(xo[50]), T8n, ovs, &(xo[2])); cannam@95: STN2(&(xo[48]), T7C, T8n, ovs); cannam@95: T8o = VFNMSI(T5Y, T5R); cannam@95: STM2(&(xo[78]), T8o, ovs, &(xo[2])); cannam@95: STN2(&(xo[76]), T8g, T8o, ovs); cannam@95: } cannam@95: } cannam@95: } cannam@95: } cannam@95: } cannam@95: VLEAVE(); cannam@95: } cannam@95: cannam@95: static const kdft_desc desc = { 64, XSIMD_STRING("n2bv_64"), {198, 0, 258, 0}, &GENUS, 0, 2, 0, 0 }; cannam@95: cannam@95: void XSIMD(codelet_n2bv_64) (planner *p) { cannam@95: X(kdft_register) (p, n2bv_64, &desc); cannam@95: } cannam@95: cannam@95: #else /* HAVE_FMA */ cannam@95: cannam@95: /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 64 -name n2bv_64 -with-ostride 2 -include n2b.h -store-multiple 2 */ cannam@95: cannam@95: /* cannam@95: * This function contains 456 FP additions, 124 FP multiplications, cannam@95: * (or, 404 additions, 72 multiplications, 52 fused multiply/add), cannam@95: * 128 stack variables, 15 constants, and 160 memory accesses cannam@95: */ cannam@95: #include "n2b.h" cannam@95: cannam@95: static void n2bv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) cannam@95: { cannam@95: DVK(KP634393284, +0.634393284163645498215171613225493370675687095); cannam@95: DVK(KP773010453, +0.773010453362736960810906609758469800971041293); cannam@95: DVK(KP290284677, +0.290284677254462367636192375817395274691476278); cannam@95: DVK(KP956940335, +0.956940335732208864935797886980269969482849206); cannam@95: DVK(KP098017140, +0.098017140329560601994195563888641845861136673); cannam@95: DVK(KP995184726, +0.995184726672196886244836953109479921575474869); cannam@95: DVK(KP471396736, +0.471396736825997648556387625905254377657460319); cannam@95: DVK(KP881921264, +0.881921264348355029712756863660388349508442621); cannam@95: DVK(KP195090322, +0.195090322016128267848284868477022240927691618); cannam@95: DVK(KP980785280, +0.980785280403230449126182236134239036973933731); cannam@95: DVK(KP555570233, +0.555570233019602224742830813948532874374937191); cannam@95: DVK(KP831469612, +0.831469612302545237078788377617905756738560812); cannam@95: DVK(KP923879532, +0.923879532511286756128183189396788286822416626); cannam@95: DVK(KP382683432, +0.382683432365089771728459984030398866761344562); cannam@95: DVK(KP707106781, +0.707106781186547524400844362104849039284835938); cannam@95: { cannam@95: INT i; cannam@95: const R *xi; cannam@95: R *xo; cannam@95: xi = ii; cannam@95: xo = io; cannam@95: for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(128, is), MAKE_VOLATILE_STRIDE(128, os)) { cannam@95: V T4p, T5u, Tb, T3A, T2q, T3v, T6G, T78, Tq, T3w, T6B, T79, T2l, T3B, T4w; cannam@95: V T5r, TI, T2g, T6u, T74, T3q, T3D, T4E, T5o, TZ, T2h, T6x, T75, T3t, T3E; cannam@95: V T4L, T5p, T23, T2N, T6m, T70, T6p, T71, T2c, T2O, T3i, T3Y, T5f, T5R, T5k; cannam@95: V T5S, T3l, T3Z, T1s, T2K, T6f, T6X, T6i, T6Y, T1B, T2L, T3b, T3V, T4Y, T5O; cannam@95: V T53, T5P, T3e, T3W; cannam@95: { cannam@95: V T3, T4n, T2p, T4o, T6, T5s, T9, T5t; cannam@95: { cannam@95: V T1, T2, T2n, T2o; cannam@95: T1 = LD(&(xi[0]), ivs, &(xi[0])); cannam@95: T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0])); cannam@95: T3 = VSUB(T1, T2); cannam@95: T4n = VADD(T1, T2); cannam@95: T2n = LD(&(xi[WS(is, 16)]), ivs, &(xi[0])); cannam@95: T2o = LD(&(xi[WS(is, 48)]), ivs, &(xi[0])); cannam@95: T2p = VSUB(T2n, T2o); cannam@95: T4o = VADD(T2n, T2o); cannam@95: } cannam@95: { cannam@95: V T4, T5, T7, T8; cannam@95: T4 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0])); cannam@95: T5 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0])); cannam@95: T6 = VSUB(T4, T5); cannam@95: T5s = VADD(T4, T5); cannam@95: T7 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0])); cannam@95: T8 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0])); cannam@95: T9 = VSUB(T7, T8); cannam@95: T5t = VADD(T7, T8); cannam@95: } cannam@95: T4p = VSUB(T4n, T4o); cannam@95: T5u = VSUB(T5s, T5t); cannam@95: { cannam@95: V Ta, T2m, T6E, T6F; cannam@95: Ta = VMUL(LDK(KP707106781), VADD(T6, T9)); cannam@95: Tb = VSUB(T3, Ta); cannam@95: T3A = VADD(T3, Ta); cannam@95: T2m = VMUL(LDK(KP707106781), VSUB(T6, T9)); cannam@95: T2q = VSUB(T2m, T2p); cannam@95: T3v = VADD(T2p, T2m); cannam@95: T6E = VADD(T4n, T4o); cannam@95: T6F = VADD(T5s, T5t); cannam@95: T6G = VSUB(T6E, T6F); cannam@95: T78 = VADD(T6E, T6F); cannam@95: } cannam@95: } cannam@95: { cannam@95: V Te, T4q, To, T4t, Th, T4r, Tl, T4u; cannam@95: { cannam@95: V Tc, Td, Tm, Tn; cannam@95: Tc = LD(&(xi[WS(is, 4)]), ivs, &(xi[0])); cannam@95: Td = LD(&(xi[WS(is, 36)]), ivs, &(xi[0])); cannam@95: Te = VSUB(Tc, Td); cannam@95: T4q = VADD(Tc, Td); cannam@95: Tm = LD(&(xi[WS(is, 60)]), ivs, &(xi[0])); cannam@95: Tn = LD(&(xi[WS(is, 28)]), ivs, &(xi[0])); cannam@95: To = VSUB(Tm, Tn); cannam@95: T4t = VADD(Tm, Tn); cannam@95: } cannam@95: { cannam@95: V Tf, Tg, Tj, Tk; cannam@95: Tf = LD(&(xi[WS(is, 20)]), ivs, &(xi[0])); cannam@95: Tg = LD(&(xi[WS(is, 52)]), ivs, &(xi[0])); cannam@95: Th = VSUB(Tf, Tg); cannam@95: T4r = VADD(Tf, Tg); cannam@95: Tj = LD(&(xi[WS(is, 12)]), ivs, &(xi[0])); cannam@95: Tk = LD(&(xi[WS(is, 44)]), ivs, &(xi[0])); cannam@95: Tl = VSUB(Tj, Tk); cannam@95: T4u = VADD(Tj, Tk); cannam@95: } cannam@95: { cannam@95: V Ti, Tp, T6z, T6A; cannam@95: Ti = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th)); cannam@95: Tp = VFNMS(LDK(KP382683432), To, VMUL(LDK(KP923879532), Tl)); cannam@95: Tq = VSUB(Ti, Tp); cannam@95: T3w = VADD(Ti, Tp); cannam@95: T6z = VADD(T4q, T4r); cannam@95: T6A = VADD(T4t, T4u); cannam@95: T6B = VSUB(T6z, T6A); cannam@95: T79 = VADD(T6z, T6A); cannam@95: } cannam@95: { cannam@95: V T2j, T2k, T4s, T4v; cannam@95: T2j = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te)); cannam@95: T2k = VFMA(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl)); cannam@95: T2l = VSUB(T2j, T2k); cannam@95: T3B = VADD(T2j, T2k); cannam@95: T4s = VSUB(T4q, T4r); cannam@95: T4v = VSUB(T4t, T4u); cannam@95: T4w = VMUL(LDK(KP707106781), VADD(T4s, T4v)); cannam@95: T5r = VMUL(LDK(KP707106781), VSUB(T4s, T4v)); cannam@95: } cannam@95: } cannam@95: { cannam@95: V TB, T4z, TF, T4y, Ty, T4C, TG, T4B; cannam@95: { cannam@95: V Tz, TA, TD, TE; cannam@95: Tz = LD(&(xi[WS(is, 18)]), ivs, &(xi[0])); cannam@95: TA = LD(&(xi[WS(is, 50)]), ivs, &(xi[0])); cannam@95: TB = VSUB(Tz, TA); cannam@95: T4z = VADD(Tz, TA); cannam@95: TD = LD(&(xi[WS(is, 2)]), ivs, &(xi[0])); cannam@95: TE = LD(&(xi[WS(is, 34)]), ivs, &(xi[0])); cannam@95: TF = VSUB(TD, TE); cannam@95: T4y = VADD(TD, TE); cannam@95: { cannam@95: V Ts, Tt, Tu, Tv, Tw, Tx; cannam@95: Ts = LD(&(xi[WS(is, 10)]), ivs, &(xi[0])); cannam@95: Tt = LD(&(xi[WS(is, 42)]), ivs, &(xi[0])); cannam@95: Tu = VSUB(Ts, Tt); cannam@95: Tv = LD(&(xi[WS(is, 58)]), ivs, &(xi[0])); cannam@95: Tw = LD(&(xi[WS(is, 26)]), ivs, &(xi[0])); cannam@95: Tx = VSUB(Tv, Tw); cannam@95: Ty = VMUL(LDK(KP707106781), VSUB(Tu, Tx)); cannam@95: T4C = VADD(Tv, Tw); cannam@95: TG = VMUL(LDK(KP707106781), VADD(Tu, Tx)); cannam@95: T4B = VADD(Ts, Tt); cannam@95: } cannam@95: } cannam@95: { cannam@95: V TC, TH, T6s, T6t; cannam@95: TC = VSUB(Ty, TB); cannam@95: TH = VSUB(TF, TG); cannam@95: TI = VFMA(LDK(KP831469612), TC, VMUL(LDK(KP555570233), TH)); cannam@95: T2g = VFNMS(LDK(KP555570233), TC, VMUL(LDK(KP831469612), TH)); cannam@95: T6s = VADD(T4y, T4z); cannam@95: T6t = VADD(T4B, T4C); cannam@95: T6u = VSUB(T6s, T6t); cannam@95: T74 = VADD(T6s, T6t); cannam@95: } cannam@95: { cannam@95: V T3o, T3p, T4A, T4D; cannam@95: T3o = VADD(TB, Ty); cannam@95: T3p = VADD(TF, TG); cannam@95: T3q = VFMA(LDK(KP980785280), T3o, VMUL(LDK(KP195090322), T3p)); cannam@95: T3D = VFNMS(LDK(KP195090322), T3o, VMUL(LDK(KP980785280), T3p)); cannam@95: T4A = VSUB(T4y, T4z); cannam@95: T4D = VSUB(T4B, T4C); cannam@95: T4E = VFMA(LDK(KP382683432), T4A, VMUL(LDK(KP923879532), T4D)); cannam@95: T5o = VFNMS(LDK(KP382683432), T4D, VMUL(LDK(KP923879532), T4A)); cannam@95: } cannam@95: } cannam@95: { cannam@95: V TS, T4J, TW, T4I, TP, T4G, TX, T4F; cannam@95: { cannam@95: V TQ, TR, TU, TV; cannam@95: TQ = LD(&(xi[WS(is, 14)]), ivs, &(xi[0])); cannam@95: TR = LD(&(xi[WS(is, 46)]), ivs, &(xi[0])); cannam@95: TS = VSUB(TQ, TR); cannam@95: T4J = VADD(TQ, TR); cannam@95: TU = LD(&(xi[WS(is, 62)]), ivs, &(xi[0])); cannam@95: TV = LD(&(xi[WS(is, 30)]), ivs, &(xi[0])); cannam@95: TW = VSUB(TU, TV); cannam@95: T4I = VADD(TU, TV); cannam@95: { cannam@95: V TJ, TK, TL, TM, TN, TO; cannam@95: TJ = LD(&(xi[WS(is, 6)]), ivs, &(xi[0])); cannam@95: TK = LD(&(xi[WS(is, 38)]), ivs, &(xi[0])); cannam@95: TL = VSUB(TJ, TK); cannam@95: TM = LD(&(xi[WS(is, 54)]), ivs, &(xi[0])); cannam@95: TN = LD(&(xi[WS(is, 22)]), ivs, &(xi[0])); cannam@95: TO = VSUB(TM, TN); cannam@95: TP = VMUL(LDK(KP707106781), VSUB(TL, TO)); cannam@95: T4G = VADD(TM, TN); cannam@95: TX = VMUL(LDK(KP707106781), VADD(TL, TO)); cannam@95: T4F = VADD(TJ, TK); cannam@95: } cannam@95: } cannam@95: { cannam@95: V TT, TY, T6v, T6w; cannam@95: TT = VSUB(TP, TS); cannam@95: TY = VSUB(TW, TX); cannam@95: TZ = VFNMS(LDK(KP555570233), TY, VMUL(LDK(KP831469612), TT)); cannam@95: T2h = VFMA(LDK(KP555570233), TT, VMUL(LDK(KP831469612), TY)); cannam@95: T6v = VADD(T4I, T4J); cannam@95: T6w = VADD(T4F, T4G); cannam@95: T6x = VSUB(T6v, T6w); cannam@95: T75 = VADD(T6v, T6w); cannam@95: } cannam@95: { cannam@95: V T3r, T3s, T4H, T4K; cannam@95: T3r = VADD(TS, TP); cannam@95: T3s = VADD(TW, TX); cannam@95: T3t = VFNMS(LDK(KP195090322), T3s, VMUL(LDK(KP980785280), T3r)); cannam@95: T3E = VFMA(LDK(KP195090322), T3r, VMUL(LDK(KP980785280), T3s)); cannam@95: T4H = VSUB(T4F, T4G); cannam@95: T4K = VSUB(T4I, T4J); cannam@95: T4L = VFNMS(LDK(KP382683432), T4K, VMUL(LDK(KP923879532), T4H)); cannam@95: T5p = VFMA(LDK(KP923879532), T4K, VMUL(LDK(KP382683432), T4H)); cannam@95: } cannam@95: } cannam@95: { cannam@95: V T21, T5h, T26, T5g, T1Y, T5d, T27, T5c, T55, T56, T1J, T57, T29, T58, T59; cannam@95: V T1Q, T5a, T2a; cannam@95: { cannam@95: V T1Z, T20, T24, T25; cannam@95: T1Z = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)])); cannam@95: T20 = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)])); cannam@95: T21 = VSUB(T1Z, T20); cannam@95: T5h = VADD(T1Z, T20); cannam@95: T24 = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)])); cannam@95: T25 = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)])); cannam@95: T26 = VSUB(T24, T25); cannam@95: T5g = VADD(T24, T25); cannam@95: } cannam@95: { cannam@95: V T1S, T1T, T1U, T1V, T1W, T1X; cannam@95: T1S = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1T = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1U = VSUB(T1S, T1T); cannam@95: T1V = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1W = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1X = VSUB(T1V, T1W); cannam@95: T1Y = VMUL(LDK(KP707106781), VSUB(T1U, T1X)); cannam@95: T5d = VADD(T1V, T1W); cannam@95: T27 = VMUL(LDK(KP707106781), VADD(T1U, T1X)); cannam@95: T5c = VADD(T1S, T1T); cannam@95: } cannam@95: { cannam@95: V T1F, T1I, T1M, T1P; cannam@95: { cannam@95: V T1D, T1E, T1G, T1H; cannam@95: T1D = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1E = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1F = VSUB(T1D, T1E); cannam@95: T55 = VADD(T1D, T1E); cannam@95: T1G = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1H = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1I = VSUB(T1G, T1H); cannam@95: T56 = VADD(T1G, T1H); cannam@95: } cannam@95: T1J = VFNMS(LDK(KP382683432), T1I, VMUL(LDK(KP923879532), T1F)); cannam@95: T57 = VSUB(T55, T56); cannam@95: T29 = VFMA(LDK(KP382683432), T1F, VMUL(LDK(KP923879532), T1I)); cannam@95: { cannam@95: V T1K, T1L, T1N, T1O; cannam@95: T1K = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1L = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1M = VSUB(T1K, T1L); cannam@95: T58 = VADD(T1K, T1L); cannam@95: T1N = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1O = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1P = VSUB(T1N, T1O); cannam@95: T59 = VADD(T1N, T1O); cannam@95: } cannam@95: T1Q = VFMA(LDK(KP923879532), T1M, VMUL(LDK(KP382683432), T1P)); cannam@95: T5a = VSUB(T58, T59); cannam@95: T2a = VFNMS(LDK(KP382683432), T1M, VMUL(LDK(KP923879532), T1P)); cannam@95: } cannam@95: { cannam@95: V T1R, T22, T6k, T6l; cannam@95: T1R = VSUB(T1J, T1Q); cannam@95: T22 = VSUB(T1Y, T21); cannam@95: T23 = VSUB(T1R, T22); cannam@95: T2N = VADD(T22, T1R); cannam@95: T6k = VADD(T5g, T5h); cannam@95: T6l = VADD(T5c, T5d); cannam@95: T6m = VSUB(T6k, T6l); cannam@95: T70 = VADD(T6k, T6l); cannam@95: } cannam@95: { cannam@95: V T6n, T6o, T28, T2b; cannam@95: T6n = VADD(T55, T56); cannam@95: T6o = VADD(T58, T59); cannam@95: T6p = VSUB(T6n, T6o); cannam@95: T71 = VADD(T6n, T6o); cannam@95: T28 = VSUB(T26, T27); cannam@95: T2b = VSUB(T29, T2a); cannam@95: T2c = VSUB(T28, T2b); cannam@95: T2O = VADD(T28, T2b); cannam@95: } cannam@95: { cannam@95: V T3g, T3h, T5b, T5e; cannam@95: T3g = VADD(T26, T27); cannam@95: T3h = VADD(T1J, T1Q); cannam@95: T3i = VADD(T3g, T3h); cannam@95: T3Y = VSUB(T3g, T3h); cannam@95: T5b = VMUL(LDK(KP707106781), VSUB(T57, T5a)); cannam@95: T5e = VSUB(T5c, T5d); cannam@95: T5f = VSUB(T5b, T5e); cannam@95: T5R = VADD(T5e, T5b); cannam@95: } cannam@95: { cannam@95: V T5i, T5j, T3j, T3k; cannam@95: T5i = VSUB(T5g, T5h); cannam@95: T5j = VMUL(LDK(KP707106781), VADD(T57, T5a)); cannam@95: T5k = VSUB(T5i, T5j); cannam@95: T5S = VADD(T5i, T5j); cannam@95: T3j = VADD(T21, T1Y); cannam@95: T3k = VADD(T29, T2a); cannam@95: T3l = VADD(T3j, T3k); cannam@95: T3Z = VSUB(T3k, T3j); cannam@95: } cannam@95: } cannam@95: { cannam@95: V T1q, T50, T1v, T4Z, T1n, T4W, T1w, T4V, T4O, T4P, T18, T4Q, T1y, T4R, T4S; cannam@95: V T1f, T4T, T1z; cannam@95: { cannam@95: V T1o, T1p, T1t, T1u; cannam@95: T1o = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1p = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1q = VSUB(T1o, T1p); cannam@95: T50 = VADD(T1o, T1p); cannam@95: T1t = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1u = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1v = VSUB(T1t, T1u); cannam@95: T4Z = VADD(T1t, T1u); cannam@95: } cannam@95: { cannam@95: V T1h, T1i, T1j, T1k, T1l, T1m; cannam@95: T1h = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1i = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1j = VSUB(T1h, T1i); cannam@95: T1k = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1l = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1m = VSUB(T1k, T1l); cannam@95: T1n = VMUL(LDK(KP707106781), VSUB(T1j, T1m)); cannam@95: T4W = VADD(T1k, T1l); cannam@95: T1w = VMUL(LDK(KP707106781), VADD(T1j, T1m)); cannam@95: T4V = VADD(T1h, T1i); cannam@95: } cannam@95: { cannam@95: V T14, T17, T1b, T1e; cannam@95: { cannam@95: V T12, T13, T15, T16; cannam@95: T12 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)])); cannam@95: T13 = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)])); cannam@95: T14 = VSUB(T12, T13); cannam@95: T4O = VADD(T12, T13); cannam@95: T15 = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)])); cannam@95: T16 = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)])); cannam@95: T17 = VSUB(T15, T16); cannam@95: T4P = VADD(T15, T16); cannam@95: } cannam@95: T18 = VFNMS(LDK(KP382683432), T17, VMUL(LDK(KP923879532), T14)); cannam@95: T4Q = VSUB(T4O, T4P); cannam@95: T1y = VFMA(LDK(KP382683432), T14, VMUL(LDK(KP923879532), T17)); cannam@95: { cannam@95: V T19, T1a, T1c, T1d; cannam@95: T19 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1a = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1b = VSUB(T19, T1a); cannam@95: T4R = VADD(T19, T1a); cannam@95: T1c = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1d = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)])); cannam@95: T1e = VSUB(T1c, T1d); cannam@95: T4S = VADD(T1c, T1d); cannam@95: } cannam@95: T1f = VFMA(LDK(KP923879532), T1b, VMUL(LDK(KP382683432), T1e)); cannam@95: T4T = VSUB(T4R, T4S); cannam@95: T1z = VFNMS(LDK(KP382683432), T1b, VMUL(LDK(KP923879532), T1e)); cannam@95: } cannam@95: { cannam@95: V T1g, T1r, T6d, T6e; cannam@95: T1g = VSUB(T18, T1f); cannam@95: T1r = VSUB(T1n, T1q); cannam@95: T1s = VSUB(T1g, T1r); cannam@95: T2K = VADD(T1r, T1g); cannam@95: T6d = VADD(T4Z, T50); cannam@95: T6e = VADD(T4V, T4W); cannam@95: T6f = VSUB(T6d, T6e); cannam@95: T6X = VADD(T6d, T6e); cannam@95: } cannam@95: { cannam@95: V T6g, T6h, T1x, T1A; cannam@95: T6g = VADD(T4O, T4P); cannam@95: T6h = VADD(T4R, T4S); cannam@95: T6i = VSUB(T6g, T6h); cannam@95: T6Y = VADD(T6g, T6h); cannam@95: T1x = VSUB(T1v, T1w); cannam@95: T1A = VSUB(T1y, T1z); cannam@95: T1B = VSUB(T1x, T1A); cannam@95: T2L = VADD(T1x, T1A); cannam@95: } cannam@95: { cannam@95: V T39, T3a, T4U, T4X; cannam@95: T39 = VADD(T1v, T1w); cannam@95: T3a = VADD(T18, T1f); cannam@95: T3b = VADD(T39, T3a); cannam@95: T3V = VSUB(T39, T3a); cannam@95: T4U = VMUL(LDK(KP707106781), VSUB(T4Q, T4T)); cannam@95: T4X = VSUB(T4V, T4W); cannam@95: T4Y = VSUB(T4U, T4X); cannam@95: T5O = VADD(T4X, T4U); cannam@95: } cannam@95: { cannam@95: V T51, T52, T3c, T3d; cannam@95: T51 = VSUB(T4Z, T50); cannam@95: T52 = VMUL(LDK(KP707106781), VADD(T4Q, T4T)); cannam@95: T53 = VSUB(T51, T52); cannam@95: T5P = VADD(T51, T52); cannam@95: T3c = VADD(T1q, T1n); cannam@95: T3d = VADD(T1y, T1z); cannam@95: T3e = VADD(T3c, T3d); cannam@95: T3W = VSUB(T3d, T3c); cannam@95: } cannam@95: } cannam@95: { cannam@95: V T7n, T7o, T7p, T7q, T7r, T7s, T7t, T7u, T7v, T7w, T7x, T7y, T7z, T7A, T7B; cannam@95: V T7C, T7D, T7E, T7F, T7G, T7H, T7I, T7J, T7K; cannam@95: { cannam@95: V T7h, T7l, T7k, T7m; cannam@95: { cannam@95: V T7f, T7g, T7i, T7j; cannam@95: T7f = VADD(T78, T79); cannam@95: T7g = VADD(T74, T75); cannam@95: T7h = VSUB(T7f, T7g); cannam@95: T7l = VADD(T7f, T7g); cannam@95: T7i = VADD(T6X, T6Y); cannam@95: T7j = VADD(T70, T71); cannam@95: T7k = VBYI(VSUB(T7i, T7j)); cannam@95: T7m = VADD(T7i, T7j); cannam@95: } cannam@95: T7n = VSUB(T7h, T7k); cannam@95: STM2(&(xo[96]), T7n, ovs, &(xo[0])); cannam@95: T7o = VADD(T7l, T7m); cannam@95: STM2(&(xo[0]), T7o, ovs, &(xo[0])); cannam@95: T7p = VADD(T7h, T7k); cannam@95: STM2(&(xo[32]), T7p, ovs, &(xo[0])); cannam@95: T7q = VSUB(T7l, T7m); cannam@95: STM2(&(xo[64]), T7q, ovs, &(xo[0])); cannam@95: } cannam@95: { cannam@95: V T76, T7a, T73, T7b, T6Z, T72; cannam@95: T76 = VSUB(T74, T75); cannam@95: T7a = VSUB(T78, T79); cannam@95: T6Z = VSUB(T6X, T6Y); cannam@95: T72 = VSUB(T70, T71); cannam@95: T73 = VMUL(LDK(KP707106781), VSUB(T6Z, T72)); cannam@95: T7b = VMUL(LDK(KP707106781), VADD(T6Z, T72)); cannam@95: { cannam@95: V T77, T7c, T7d, T7e; cannam@95: T77 = VBYI(VSUB(T73, T76)); cannam@95: T7c = VSUB(T7a, T7b); cannam@95: T7r = VADD(T77, T7c); cannam@95: STM2(&(xo[48]), T7r, ovs, &(xo[0])); cannam@95: T7s = VSUB(T7c, T77); cannam@95: STM2(&(xo[80]), T7s, ovs, &(xo[0])); cannam@95: T7d = VBYI(VADD(T76, T73)); cannam@95: T7e = VADD(T7a, T7b); cannam@95: T7t = VADD(T7d, T7e); cannam@95: STM2(&(xo[16]), T7t, ovs, &(xo[0])); cannam@95: T7u = VSUB(T7e, T7d); cannam@95: STM2(&(xo[112]), T7u, ovs, &(xo[0])); cannam@95: } cannam@95: } cannam@95: { cannam@95: V T6C, T6S, T6I, T6P, T6r, T6Q, T6L, T6T, T6y, T6H; cannam@95: T6y = VMUL(LDK(KP707106781), VSUB(T6u, T6x)); cannam@95: T6C = VSUB(T6y, T6B); cannam@95: T6S = VADD(T6B, T6y); cannam@95: T6H = VMUL(LDK(KP707106781), VADD(T6u, T6x)); cannam@95: T6I = VSUB(T6G, T6H); cannam@95: T6P = VADD(T6G, T6H); cannam@95: { cannam@95: V T6j, T6q, T6J, T6K; cannam@95: T6j = VFNMS(LDK(KP382683432), T6i, VMUL(LDK(KP923879532), T6f)); cannam@95: T6q = VFMA(LDK(KP923879532), T6m, VMUL(LDK(KP382683432), T6p)); cannam@95: T6r = VSUB(T6j, T6q); cannam@95: T6Q = VADD(T6j, T6q); cannam@95: T6J = VFMA(LDK(KP382683432), T6f, VMUL(LDK(KP923879532), T6i)); cannam@95: T6K = VFNMS(LDK(KP382683432), T6m, VMUL(LDK(KP923879532), T6p)); cannam@95: T6L = VSUB(T6J, T6K); cannam@95: T6T = VADD(T6J, T6K); cannam@95: } cannam@95: { cannam@95: V T6D, T6M, T6V, T6W; cannam@95: T6D = VBYI(VSUB(T6r, T6C)); cannam@95: T6M = VSUB(T6I, T6L); cannam@95: T7v = VADD(T6D, T6M); cannam@95: STM2(&(xo[40]), T7v, ovs, &(xo[0])); cannam@95: T7w = VSUB(T6M, T6D); cannam@95: STM2(&(xo[88]), T7w, ovs, &(xo[0])); cannam@95: T6V = VSUB(T6P, T6Q); cannam@95: T6W = VBYI(VSUB(T6T, T6S)); cannam@95: T7x = VSUB(T6V, T6W); cannam@95: STM2(&(xo[72]), T7x, ovs, &(xo[0])); cannam@95: T7y = VADD(T6V, T6W); cannam@95: STM2(&(xo[56]), T7y, ovs, &(xo[0])); cannam@95: } cannam@95: { cannam@95: V T6N, T6O, T6R, T6U; cannam@95: T6N = VBYI(VADD(T6C, T6r)); cannam@95: T6O = VADD(T6I, T6L); cannam@95: T7z = VADD(T6N, T6O); cannam@95: STM2(&(xo[24]), T7z, ovs, &(xo[0])); cannam@95: T7A = VSUB(T6O, T6N); cannam@95: STM2(&(xo[104]), T7A, ovs, &(xo[0])); cannam@95: T6R = VADD(T6P, T6Q); cannam@95: T6U = VBYI(VADD(T6S, T6T)); cannam@95: T7B = VSUB(T6R, T6U); cannam@95: STM2(&(xo[120]), T7B, ovs, &(xo[0])); cannam@95: T7C = VADD(T6R, T6U); cannam@95: STM2(&(xo[8]), T7C, ovs, &(xo[0])); cannam@95: } cannam@95: } cannam@95: { cannam@95: V T5N, T68, T61, T69, T5U, T65, T5Y, T66; cannam@95: { cannam@95: V T5L, T5M, T5Z, T60; cannam@95: T5L = VADD(T4p, T4w); cannam@95: T5M = VADD(T5o, T5p); cannam@95: T5N = VSUB(T5L, T5M); cannam@95: T68 = VADD(T5L, T5M); cannam@95: T5Z = VFNMS(LDK(KP195090322), T5O, VMUL(LDK(KP980785280), T5P)); cannam@95: T60 = VFMA(LDK(KP195090322), T5R, VMUL(LDK(KP980785280), T5S)); cannam@95: T61 = VSUB(T5Z, T60); cannam@95: T69 = VADD(T5Z, T60); cannam@95: } cannam@95: { cannam@95: V T5Q, T5T, T5W, T5X; cannam@95: T5Q = VFMA(LDK(KP980785280), T5O, VMUL(LDK(KP195090322), T5P)); cannam@95: T5T = VFNMS(LDK(KP195090322), T5S, VMUL(LDK(KP980785280), T5R)); cannam@95: T5U = VSUB(T5Q, T5T); cannam@95: T65 = VADD(T5Q, T5T); cannam@95: T5W = VADD(T4E, T4L); cannam@95: T5X = VADD(T5u, T5r); cannam@95: T5Y = VSUB(T5W, T5X); cannam@95: T66 = VADD(T5X, T5W); cannam@95: } cannam@95: { cannam@95: V T5V, T62, T6b, T6c; cannam@95: T5V = VADD(T5N, T5U); cannam@95: T62 = VBYI(VADD(T5Y, T61)); cannam@95: T7D = VSUB(T5V, T62); cannam@95: STM2(&(xo[100]), T7D, ovs, &(xo[0])); cannam@95: T7E = VADD(T5V, T62); cannam@95: STM2(&(xo[28]), T7E, ovs, &(xo[0])); cannam@95: T6b = VBYI(VADD(T66, T65)); cannam@95: T6c = VADD(T68, T69); cannam@95: T7F = VADD(T6b, T6c); cannam@95: STM2(&(xo[4]), T7F, ovs, &(xo[0])); cannam@95: T7G = VSUB(T6c, T6b); cannam@95: STM2(&(xo[124]), T7G, ovs, &(xo[0])); cannam@95: } cannam@95: { cannam@95: V T63, T64, T67, T6a; cannam@95: T63 = VSUB(T5N, T5U); cannam@95: T64 = VBYI(VSUB(T61, T5Y)); cannam@95: T7H = VSUB(T63, T64); cannam@95: STM2(&(xo[92]), T7H, ovs, &(xo[0])); cannam@95: T7I = VADD(T63, T64); cannam@95: STM2(&(xo[36]), T7I, ovs, &(xo[0])); cannam@95: T67 = VBYI(VSUB(T65, T66)); cannam@95: T6a = VSUB(T68, T69); cannam@95: T7J = VADD(T67, T6a); cannam@95: STM2(&(xo[60]), T7J, ovs, &(xo[0])); cannam@95: T7K = VSUB(T6a, T67); cannam@95: STM2(&(xo[68]), T7K, ovs, &(xo[0])); cannam@95: } cannam@95: } cannam@95: { cannam@95: V T7M, T7O, T7P, T7R; cannam@95: { cannam@95: V T11, T2C, T2v, T2D, T2e, T2z, T2s, T2A; cannam@95: { cannam@95: V Tr, T10, T2t, T2u; cannam@95: Tr = VSUB(Tb, Tq); cannam@95: T10 = VSUB(TI, TZ); cannam@95: T11 = VSUB(Tr, T10); cannam@95: T2C = VADD(Tr, T10); cannam@95: T2t = VFNMS(LDK(KP471396736), T1s, VMUL(LDK(KP881921264), T1B)); cannam@95: T2u = VFMA(LDK(KP471396736), T23, VMUL(LDK(KP881921264), T2c)); cannam@95: T2v = VSUB(T2t, T2u); cannam@95: T2D = VADD(T2t, T2u); cannam@95: } cannam@95: { cannam@95: V T1C, T2d, T2i, T2r; cannam@95: T1C = VFMA(LDK(KP881921264), T1s, VMUL(LDK(KP471396736), T1B)); cannam@95: T2d = VFNMS(LDK(KP471396736), T2c, VMUL(LDK(KP881921264), T23)); cannam@95: T2e = VSUB(T1C, T2d); cannam@95: T2z = VADD(T1C, T2d); cannam@95: T2i = VSUB(T2g, T2h); cannam@95: T2r = VSUB(T2l, T2q); cannam@95: T2s = VSUB(T2i, T2r); cannam@95: T2A = VADD(T2r, T2i); cannam@95: } cannam@95: { cannam@95: V T2f, T2w, T7L, T2F, T2G, T7N; cannam@95: T2f = VADD(T11, T2e); cannam@95: T2w = VBYI(VADD(T2s, T2v)); cannam@95: T7L = VSUB(T2f, T2w); cannam@95: STM2(&(xo[106]), T7L, ovs, &(xo[2])); cannam@95: STN2(&(xo[104]), T7A, T7L, ovs); cannam@95: T7M = VADD(T2f, T2w); cannam@95: STM2(&(xo[22]), T7M, ovs, &(xo[2])); cannam@95: T2F = VBYI(VADD(T2A, T2z)); cannam@95: T2G = VADD(T2C, T2D); cannam@95: T7N = VADD(T2F, T2G); cannam@95: STM2(&(xo[10]), T7N, ovs, &(xo[2])); cannam@95: STN2(&(xo[8]), T7C, T7N, ovs); cannam@95: T7O = VSUB(T2G, T2F); cannam@95: STM2(&(xo[118]), T7O, ovs, &(xo[2])); cannam@95: } cannam@95: { cannam@95: V T2x, T2y, T7Q, T2B, T2E, T7S; cannam@95: T2x = VSUB(T11, T2e); cannam@95: T2y = VBYI(VSUB(T2v, T2s)); cannam@95: T7P = VSUB(T2x, T2y); cannam@95: STM2(&(xo[86]), T7P, ovs, &(xo[2])); cannam@95: T7Q = VADD(T2x, T2y); cannam@95: STM2(&(xo[42]), T7Q, ovs, &(xo[2])); cannam@95: STN2(&(xo[40]), T7v, T7Q, ovs); cannam@95: T2B = VBYI(VSUB(T2z, T2A)); cannam@95: T2E = VSUB(T2C, T2D); cannam@95: T7R = VADD(T2B, T2E); cannam@95: STM2(&(xo[54]), T7R, ovs, &(xo[2])); cannam@95: T7S = VSUB(T2E, T2B); cannam@95: STM2(&(xo[74]), T7S, ovs, &(xo[2])); cannam@95: STN2(&(xo[72]), T7x, T7S, ovs); cannam@95: } cannam@95: } cannam@95: { cannam@95: V T3n, T3O, T3J, T3R, T3y, T3Q, T3G, T3N; cannam@95: { cannam@95: V T3f, T3m, T3H, T3I; cannam@95: T3f = VFNMS(LDK(KP098017140), T3e, VMUL(LDK(KP995184726), T3b)); cannam@95: T3m = VFMA(LDK(KP995184726), T3i, VMUL(LDK(KP098017140), T3l)); cannam@95: T3n = VSUB(T3f, T3m); cannam@95: T3O = VADD(T3f, T3m); cannam@95: T3H = VFMA(LDK(KP098017140), T3b, VMUL(LDK(KP995184726), T3e)); cannam@95: T3I = VFNMS(LDK(KP098017140), T3i, VMUL(LDK(KP995184726), T3l)); cannam@95: T3J = VSUB(T3H, T3I); cannam@95: T3R = VADD(T3H, T3I); cannam@95: } cannam@95: { cannam@95: V T3u, T3x, T3C, T3F; cannam@95: T3u = VADD(T3q, T3t); cannam@95: T3x = VADD(T3v, T3w); cannam@95: T3y = VSUB(T3u, T3x); cannam@95: T3Q = VADD(T3x, T3u); cannam@95: T3C = VADD(T3A, T3B); cannam@95: T3F = VADD(T3D, T3E); cannam@95: T3G = VSUB(T3C, T3F); cannam@95: T3N = VADD(T3C, T3F); cannam@95: } cannam@95: { cannam@95: V T3z, T3K, T7T, T7U; cannam@95: T3z = VBYI(VSUB(T3n, T3y)); cannam@95: T3K = VSUB(T3G, T3J); cannam@95: T7T = VADD(T3z, T3K); cannam@95: STM2(&(xo[34]), T7T, ovs, &(xo[2])); cannam@95: STN2(&(xo[32]), T7p, T7T, ovs); cannam@95: T7U = VSUB(T3K, T3z); cannam@95: STM2(&(xo[94]), T7U, ovs, &(xo[2])); cannam@95: STN2(&(xo[92]), T7H, T7U, ovs); cannam@95: } cannam@95: { cannam@95: V T3T, T3U, T7V, T7W; cannam@95: T3T = VSUB(T3N, T3O); cannam@95: T3U = VBYI(VSUB(T3R, T3Q)); cannam@95: T7V = VSUB(T3T, T3U); cannam@95: STM2(&(xo[66]), T7V, ovs, &(xo[2])); cannam@95: STN2(&(xo[64]), T7q, T7V, ovs); cannam@95: T7W = VADD(T3T, T3U); cannam@95: STM2(&(xo[62]), T7W, ovs, &(xo[2])); cannam@95: STN2(&(xo[60]), T7J, T7W, ovs); cannam@95: } cannam@95: { cannam@95: V T3L, T3M, T7X, T7Y; cannam@95: T3L = VBYI(VADD(T3y, T3n)); cannam@95: T3M = VADD(T3G, T3J); cannam@95: T7X = VADD(T3L, T3M); cannam@95: STM2(&(xo[30]), T7X, ovs, &(xo[2])); cannam@95: STN2(&(xo[28]), T7E, T7X, ovs); cannam@95: T7Y = VSUB(T3M, T3L); cannam@95: STM2(&(xo[98]), T7Y, ovs, &(xo[2])); cannam@95: STN2(&(xo[96]), T7n, T7Y, ovs); cannam@95: } cannam@95: { cannam@95: V T3P, T3S, T7Z, T80; cannam@95: T3P = VADD(T3N, T3O); cannam@95: T3S = VBYI(VADD(T3Q, T3R)); cannam@95: T7Z = VSUB(T3P, T3S); cannam@95: STM2(&(xo[126]), T7Z, ovs, &(xo[2])); cannam@95: STN2(&(xo[124]), T7G, T7Z, ovs); cannam@95: T80 = VADD(T3P, T3S); cannam@95: STM2(&(xo[2]), T80, ovs, &(xo[2])); cannam@95: STN2(&(xo[0]), T7o, T80, ovs); cannam@95: } cannam@95: } cannam@95: { cannam@95: V T81, T83, T86, T88; cannam@95: { cannam@95: V T4N, T5G, T5z, T5H, T5m, T5D, T5w, T5E; cannam@95: { cannam@95: V T4x, T4M, T5x, T5y; cannam@95: T4x = VSUB(T4p, T4w); cannam@95: T4M = VSUB(T4E, T4L); cannam@95: T4N = VSUB(T4x, T4M); cannam@95: T5G = VADD(T4x, T4M); cannam@95: T5x = VFNMS(LDK(KP555570233), T4Y, VMUL(LDK(KP831469612), T53)); cannam@95: T5y = VFMA(LDK(KP555570233), T5f, VMUL(LDK(KP831469612), T5k)); cannam@95: T5z = VSUB(T5x, T5y); cannam@95: T5H = VADD(T5x, T5y); cannam@95: } cannam@95: { cannam@95: V T54, T5l, T5q, T5v; cannam@95: T54 = VFMA(LDK(KP831469612), T4Y, VMUL(LDK(KP555570233), T53)); cannam@95: T5l = VFNMS(LDK(KP555570233), T5k, VMUL(LDK(KP831469612), T5f)); cannam@95: T5m = VSUB(T54, T5l); cannam@95: T5D = VADD(T54, T5l); cannam@95: T5q = VSUB(T5o, T5p); cannam@95: T5v = VSUB(T5r, T5u); cannam@95: T5w = VSUB(T5q, T5v); cannam@95: T5E = VADD(T5v, T5q); cannam@95: } cannam@95: { cannam@95: V T5n, T5A, T82, T5J, T5K, T84; cannam@95: T5n = VADD(T4N, T5m); cannam@95: T5A = VBYI(VADD(T5w, T5z)); cannam@95: T81 = VSUB(T5n, T5A); cannam@95: STM2(&(xo[108]), T81, ovs, &(xo[0])); cannam@95: T82 = VADD(T5n, T5A); cannam@95: STM2(&(xo[20]), T82, ovs, &(xo[0])); cannam@95: STN2(&(xo[20]), T82, T7M, ovs); cannam@95: T5J = VBYI(VADD(T5E, T5D)); cannam@95: T5K = VADD(T5G, T5H); cannam@95: T83 = VADD(T5J, T5K); cannam@95: STM2(&(xo[12]), T83, ovs, &(xo[0])); cannam@95: T84 = VSUB(T5K, T5J); cannam@95: STM2(&(xo[116]), T84, ovs, &(xo[0])); cannam@95: STN2(&(xo[116]), T84, T7O, ovs); cannam@95: } cannam@95: { cannam@95: V T5B, T5C, T85, T5F, T5I, T87; cannam@95: T5B = VSUB(T4N, T5m); cannam@95: T5C = VBYI(VSUB(T5z, T5w)); cannam@95: T85 = VSUB(T5B, T5C); cannam@95: STM2(&(xo[84]), T85, ovs, &(xo[0])); cannam@95: STN2(&(xo[84]), T85, T7P, ovs); cannam@95: T86 = VADD(T5B, T5C); cannam@95: STM2(&(xo[44]), T86, ovs, &(xo[0])); cannam@95: T5F = VBYI(VSUB(T5D, T5E)); cannam@95: T5I = VSUB(T5G, T5H); cannam@95: T87 = VADD(T5F, T5I); cannam@95: STM2(&(xo[52]), T87, ovs, &(xo[0])); cannam@95: STN2(&(xo[52]), T87, T7R, ovs); cannam@95: T88 = VSUB(T5I, T5F); cannam@95: STM2(&(xo[76]), T88, ovs, &(xo[0])); cannam@95: } cannam@95: } cannam@95: { cannam@95: V T2J, T34, T2X, T35, T2Q, T31, T2U, T32; cannam@95: { cannam@95: V T2H, T2I, T2V, T2W; cannam@95: T2H = VADD(Tb, Tq); cannam@95: T2I = VADD(T2g, T2h); cannam@95: T2J = VSUB(T2H, T2I); cannam@95: T34 = VADD(T2H, T2I); cannam@95: T2V = VFNMS(LDK(KP290284677), T2K, VMUL(LDK(KP956940335), T2L)); cannam@95: T2W = VFMA(LDK(KP290284677), T2N, VMUL(LDK(KP956940335), T2O)); cannam@95: T2X = VSUB(T2V, T2W); cannam@95: T35 = VADD(T2V, T2W); cannam@95: } cannam@95: { cannam@95: V T2M, T2P, T2S, T2T; cannam@95: T2M = VFMA(LDK(KP956940335), T2K, VMUL(LDK(KP290284677), T2L)); cannam@95: T2P = VFNMS(LDK(KP290284677), T2O, VMUL(LDK(KP956940335), T2N)); cannam@95: T2Q = VSUB(T2M, T2P); cannam@95: T31 = VADD(T2M, T2P); cannam@95: T2S = VADD(TI, TZ); cannam@95: T2T = VADD(T2q, T2l); cannam@95: T2U = VSUB(T2S, T2T); cannam@95: T32 = VADD(T2T, T2S); cannam@95: } cannam@95: { cannam@95: V T2R, T2Y, T89, T8a; cannam@95: T2R = VADD(T2J, T2Q); cannam@95: T2Y = VBYI(VADD(T2U, T2X)); cannam@95: T89 = VSUB(T2R, T2Y); cannam@95: STM2(&(xo[102]), T89, ovs, &(xo[2])); cannam@95: STN2(&(xo[100]), T7D, T89, ovs); cannam@95: T8a = VADD(T2R, T2Y); cannam@95: STM2(&(xo[26]), T8a, ovs, &(xo[2])); cannam@95: STN2(&(xo[24]), T7z, T8a, ovs); cannam@95: } cannam@95: { cannam@95: V T37, T38, T8b, T8c; cannam@95: T37 = VBYI(VADD(T32, T31)); cannam@95: T38 = VADD(T34, T35); cannam@95: T8b = VADD(T37, T38); cannam@95: STM2(&(xo[6]), T8b, ovs, &(xo[2])); cannam@95: STN2(&(xo[4]), T7F, T8b, ovs); cannam@95: T8c = VSUB(T38, T37); cannam@95: STM2(&(xo[122]), T8c, ovs, &(xo[2])); cannam@95: STN2(&(xo[120]), T7B, T8c, ovs); cannam@95: } cannam@95: { cannam@95: V T2Z, T30, T8d, T8e; cannam@95: T2Z = VSUB(T2J, T2Q); cannam@95: T30 = VBYI(VSUB(T2X, T2U)); cannam@95: T8d = VSUB(T2Z, T30); cannam@95: STM2(&(xo[90]), T8d, ovs, &(xo[2])); cannam@95: STN2(&(xo[88]), T7w, T8d, ovs); cannam@95: T8e = VADD(T2Z, T30); cannam@95: STM2(&(xo[38]), T8e, ovs, &(xo[2])); cannam@95: STN2(&(xo[36]), T7I, T8e, ovs); cannam@95: } cannam@95: { cannam@95: V T33, T36, T8f, T8g; cannam@95: T33 = VBYI(VSUB(T31, T32)); cannam@95: T36 = VSUB(T34, T35); cannam@95: T8f = VADD(T33, T36); cannam@95: STM2(&(xo[58]), T8f, ovs, &(xo[2])); cannam@95: STN2(&(xo[56]), T7y, T8f, ovs); cannam@95: T8g = VSUB(T36, T33); cannam@95: STM2(&(xo[70]), T8g, ovs, &(xo[2])); cannam@95: STN2(&(xo[68]), T7K, T8g, ovs); cannam@95: } cannam@95: } cannam@95: { cannam@95: V T41, T4g, T4b, T4j, T44, T4i, T48, T4f; cannam@95: { cannam@95: V T3X, T40, T49, T4a; cannam@95: T3X = VFNMS(LDK(KP634393284), T3W, VMUL(LDK(KP773010453), T3V)); cannam@95: T40 = VFMA(LDK(KP773010453), T3Y, VMUL(LDK(KP634393284), T3Z)); cannam@95: T41 = VSUB(T3X, T40); cannam@95: T4g = VADD(T3X, T40); cannam@95: T49 = VFMA(LDK(KP634393284), T3V, VMUL(LDK(KP773010453), T3W)); cannam@95: T4a = VFNMS(LDK(KP634393284), T3Y, VMUL(LDK(KP773010453), T3Z)); cannam@95: T4b = VSUB(T49, T4a); cannam@95: T4j = VADD(T49, T4a); cannam@95: } cannam@95: { cannam@95: V T42, T43, T46, T47; cannam@95: T42 = VSUB(T3D, T3E); cannam@95: T43 = VSUB(T3w, T3v); cannam@95: T44 = VSUB(T42, T43); cannam@95: T4i = VADD(T43, T42); cannam@95: T46 = VSUB(T3A, T3B); cannam@95: T47 = VSUB(T3q, T3t); cannam@95: T48 = VSUB(T46, T47); cannam@95: T4f = VADD(T46, T47); cannam@95: } cannam@95: { cannam@95: V T45, T4c, T8h, T8i; cannam@95: T45 = VBYI(VSUB(T41, T44)); cannam@95: T4c = VSUB(T48, T4b); cannam@95: T8h = VADD(T45, T4c); cannam@95: STM2(&(xo[46]), T8h, ovs, &(xo[2])); cannam@95: STN2(&(xo[44]), T86, T8h, ovs); cannam@95: T8i = VSUB(T4c, T45); cannam@95: STM2(&(xo[82]), T8i, ovs, &(xo[2])); cannam@95: STN2(&(xo[80]), T7s, T8i, ovs); cannam@95: } cannam@95: { cannam@95: V T4l, T4m, T8j, T8k; cannam@95: T4l = VSUB(T4f, T4g); cannam@95: T4m = VBYI(VSUB(T4j, T4i)); cannam@95: T8j = VSUB(T4l, T4m); cannam@95: STM2(&(xo[78]), T8j, ovs, &(xo[2])); cannam@95: STN2(&(xo[76]), T88, T8j, ovs); cannam@95: T8k = VADD(T4l, T4m); cannam@95: STM2(&(xo[50]), T8k, ovs, &(xo[2])); cannam@95: STN2(&(xo[48]), T7r, T8k, ovs); cannam@95: } cannam@95: { cannam@95: V T4d, T4e, T8l, T8m; cannam@95: T4d = VBYI(VADD(T44, T41)); cannam@95: T4e = VADD(T48, T4b); cannam@95: T8l = VADD(T4d, T4e); cannam@95: STM2(&(xo[18]), T8l, ovs, &(xo[2])); cannam@95: STN2(&(xo[16]), T7t, T8l, ovs); cannam@95: T8m = VSUB(T4e, T4d); cannam@95: STM2(&(xo[110]), T8m, ovs, &(xo[2])); cannam@95: STN2(&(xo[108]), T81, T8m, ovs); cannam@95: } cannam@95: { cannam@95: V T4h, T4k, T8n, T8o; cannam@95: T4h = VADD(T4f, T4g); cannam@95: T4k = VBYI(VADD(T4i, T4j)); cannam@95: T8n = VSUB(T4h, T4k); cannam@95: STM2(&(xo[114]), T8n, ovs, &(xo[2])); cannam@95: STN2(&(xo[112]), T7u, T8n, ovs); cannam@95: T8o = VADD(T4h, T4k); cannam@95: STM2(&(xo[14]), T8o, ovs, &(xo[2])); cannam@95: STN2(&(xo[12]), T83, T8o, ovs); cannam@95: } cannam@95: } cannam@95: } cannam@95: } cannam@95: } cannam@95: } cannam@95: } cannam@95: VLEAVE(); cannam@95: } cannam@95: cannam@95: static const kdft_desc desc = { 64, XSIMD_STRING("n2bv_64"), {404, 72, 52, 0}, &GENUS, 0, 2, 0, 0 }; cannam@95: cannam@95: void XSIMD(codelet_n2bv_64) (planner *p) { cannam@95: X(kdft_register) (p, n2bv_64, &desc); cannam@95: } cannam@95: cannam@95: #endif /* HAVE_FMA */