?? n2bv_32.c
字號:
*/#include "n2b.h"static void n2bv_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs){ DVK(KP555570233, +0.555570233019602224742830813948532874374937191); DVK(KP831469612, +0.831469612302545237078788377617905756738560812); DVK(KP195090322, +0.195090322016128267848284868477022240927691618); DVK(KP980785280, +0.980785280403230449126182236134239036973933731); DVK(KP382683432, +0.382683432365089771728459984030398866761344562); DVK(KP923879532, +0.923879532511286756128183189396788286822416626); DVK(KP707106781, +0.707106781186547524400844362104849039284835938); INT i; const R *xi; R *xo; xi = ii; xo = io; for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) { V T2f, T2k, T2N, T2M, T19, T1B, Tb, T1p, TT, T1v, TY, T1w, T2E, T2F, T2G; V T24, T2o, TC, T1s, TH, T1t, T2B, T2C, T2D, T1X, T2n, T2I, T2J, Tq, T1A; V T14, T1q, T2c, T2l; { V T3, T2i, T18, T2j, T6, T2d, T9, T2e, T15, Ta; { V T1, T2, T16, T17; T1 = LD(&(xi[0]), ivs, &(xi[0])); T2 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0])); T3 = VSUB(T1, T2); T2i = VADD(T1, T2); T16 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0])); T17 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0])); T18 = VSUB(T16, T17); T2j = VADD(T16, T17); } { V T4, T5, T7, T8; T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0])); T5 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0])); T6 = VSUB(T4, T5); T2d = VADD(T4, T5); T7 = LD(&(xi[WS(is, 28)]), ivs, &(xi[0])); T8 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0])); T9 = VSUB(T7, T8); T2e = VADD(T7, T8); } T2f = VSUB(T2d, T2e); T2k = VSUB(T2i, T2j); T2N = VADD(T2d, T2e); T2M = VADD(T2i, T2j); T15 = VMUL(LDK(KP707106781), VSUB(T6, T9)); T19 = VSUB(T15, T18); T1B = VADD(T18, T15); Ta = VMUL(LDK(KP707106781), VADD(T6, T9)); Tb = VSUB(T3, Ta); T1p = VADD(T3, Ta); } { V TL, T21, TW, T1Y, TO, T22, TS, T1Z; { V TJ, TK, TU, TV; TJ = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)])); TK = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)])); TL = VSUB(TJ, TK); T21 = VADD(TJ, TK); TU = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)])); TV = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)])); TW = VSUB(TU, TV); T1Y = VADD(TU, TV); } { V TM, TN, TQ, TR; TM = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)])); TN = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)])); TO = VSUB(TM, TN); T22 = VADD(TM, TN); TQ = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)])); TR = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)])); TS = VSUB(TQ, TR); T1Z = VADD(TQ, TR); } { V TP, TX, T20, T23; TP = VMUL(LDK(KP707106781), VSUB(TL, TO)); TT = VSUB(TP, TS); T1v = VADD(TS, TP); TX = VMUL(LDK(KP707106781), VADD(TL, TO)); TY = VSUB(TW, TX); T1w = VADD(TW, TX); T2E = VADD(T1Y, T1Z); T2F = VADD(T21, T22); T2G = VSUB(T2E, T2F); T20 = VSUB(T1Y, T1Z); T23 = VSUB(T21, T22); T24 = VFMA(LDK(KP923879532), T20, VMUL(LDK(KP382683432), T23)); T2o = VFNMS(LDK(KP382683432), T20, VMUL(LDK(KP923879532), T23)); } } { V Tu, T1U, TF, T1R, Tx, T1V, TB, T1S; { V Ts, Tt, TD, TE; Ts = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)])); Tt = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)])); Tu = VSUB(Ts, Tt); T1U = VADD(Ts, Tt); TD = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)])); TE = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)])); TF = VSUB(TD, TE); T1R = VADD(TD, TE); } { V Tv, Tw, Tz, TA; Tv = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)])); Tw = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)])); Tx = VSUB(Tv, Tw); T1V = VADD(Tv, Tw); Tz = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)])); TA = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)])); TB = VSUB(Tz, TA); T1S = VADD(Tz, TA); } { V Ty, TG, T1T, T1W; Ty = VMUL(LDK(KP707106781), VSUB(Tu, Tx)); TC = VSUB(Ty, TB); T1s = VADD(TB, Ty); TG = VMUL(LDK(KP707106781), VADD(Tu, Tx)); TH = VSUB(TF, TG); T1t = VADD(TF, TG); T2B = VADD(T1R, T1S); T2C = VADD(T1U, T1V); T2D = VSUB(T2B, T2C); T1T = VSUB(T1R, T1S); T1W = VSUB(T1U, T1V); T1X = VFNMS(LDK(KP382683432), T1W, VMUL(LDK(KP923879532), T1T)); T2n = VFMA(LDK(KP382683432), T1T, VMUL(LDK(KP923879532), T1W)); } } { V Te, T26, To, T29, Th, T27, Tl, T2a, Ti, Tp; { V Tc, Td, Tm, Tn; Tc = LD(&(xi[WS(is, 2)]), ivs, &(xi[0])); Td = LD(&(xi[WS(is, 18)]), ivs, &(xi[0])); Te = VSUB(Tc, Td); T26 = VADD(Tc, Td); Tm = LD(&(xi[WS(is, 30)]), ivs, &(xi[0])); Tn = LD(&(xi[WS(is, 14)]), ivs, &(xi[0])); To = VSUB(Tm, Tn); T29 = VADD(Tm, Tn); } { V Tf, Tg, Tj, Tk; Tf = LD(&(xi[WS(is, 10)]), ivs, &(xi[0])); Tg = LD(&(xi[WS(is, 26)]), ivs, &(xi[0])); Th = VSUB(Tf, Tg); T27 = VADD(Tf, Tg); Tj = LD(&(xi[WS(is, 6)]), ivs, &(xi[0])); Tk = LD(&(xi[WS(is, 22)]), ivs, &(xi[0])); Tl = VSUB(Tj, Tk); T2a = VADD(Tj, Tk); } T2I = VADD(T26, T27); T2J = VADD(T29, T2a); Ti = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th)); Tp = VFNMS(LDK(KP382683432), To, VMUL(LDK(KP923879532), Tl)); Tq = VSUB(Ti, Tp); T1A = VADD(Ti, Tp); { V T12, T13, T28, T2b; T12 = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te)); T13 = VFMA(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl)); T14 = VSUB(T12, T13); T1q = VADD(T12, T13); T28 = VSUB(T26, T27); T2b = VSUB(T29, T2a); T2c = VMUL(LDK(KP707106781), VSUB(T28, T2b)); T2l = VMUL(LDK(KP707106781), VADD(T28, T2b)); } } { V T31, T32, T33, T34, T35, T36, T37, T38, T39, T3a, T3b, T3c; { V T2L, T2R, T2Q, T2S; { V T2H, T2K, T2O, T2P; T2H = VMUL(LDK(KP707106781), VSUB(T2D, T2G)); T2K = VSUB(T2I, T2J); T2L = VBYI(VSUB(T2H, T2K)); T2R = VBYI(VADD(T2K, T2H)); T2O = VSUB(T2M, T2N); T2P = VMUL(LDK(KP707106781), VADD(T2D, T2G)); T2Q = VSUB(T2O, T2P); T2S = VADD(T2O, T2P); } T31 = VADD(T2L, T2Q); STM2(&(xo[24]), T31, ovs, &(xo[0])); T32 = VSUB(T2S, T2R); STM2(&(xo[56]), T32, ovs, &(xo[0])); T33 = VSUB(T2Q, T2L); STM2(&(xo[40]), T33, ovs, &(xo[0])); T34 = VADD(T2R, T2S); STM2(&(xo[8]), T34, ovs, &(xo[0])); } { V T2h, T2r, T2q, T2s; { V T25, T2g, T2m, T2p; T25 = VSUB(T1X, T24); T2g = VSUB(T2c, T2f); T2h = VBYI(VSUB(T25, T2g)); T2r = VBYI(VADD(T2g, T25)); T2m = VSUB(T2k, T2l); T2p = VSUB(T2n, T2o); T2q = VSUB(T2m, T2p); T2s = VADD(T2m, T2p); } T35 = VADD(T2h, T2q); STM2(&(xo[20]), T35, ovs, &(xo[0])); T36 = VSUB(T2s, T2r); STM2(&(xo[52]), T36, ovs, &(xo[0])); T37 = VSUB(T2q, T2h); STM2(&(xo[44]), T37, ovs, &(xo[0])); T38 = VADD(T2r, T2s); STM2(&(xo[12]), T38, ovs, &(xo[0])); } { V T2V, T2Z, T2Y, T30; { V T2T, T2U, T2W, T2X; T2T = VADD(T2M, T2N); T2U = VADD(T2I, T2J); T2V = VSUB(T2T, T2U); T2Z = VADD(T2T, T2U); T2W = VADD(T2B, T2C); T2X = VADD(T2E, T2F); T2Y = VBYI(VSUB(T2W, T2X)); T30 = VADD(T2W, T2X); } T39 = VSUB(T2V, T2Y); STM2(&(xo[48]), T39, ovs, &(xo[0])); T3a = VADD(T2Z, T30); STM2(&(xo[0]), T3a, ovs, &(xo[0])); T3b = VADD(T2V, T2Y); STM2(&(xo[16]), T3b, ovs, &(xo[0])); T3c = VSUB(T2Z, T30); STM2(&(xo[32]), T3c, ovs, &(xo[0])); } { V T3d, T3e, T3f, T3g; { V T2v, T2z, T2y, T2A; { V T2t, T2u, T2w, T2x; T2t = VADD(T2k, T2l); T2u = VADD(T1X, T24); T2v = VADD(T2t, T2u); T2z = VSUB(T2t, T2u); T2w = VADD(T2f, T2c); T2x = VADD(T2n, T2o); T2y = VBYI(VADD(T2w, T2x)); T2A = VBYI(VSUB(T2x, T2w)); } T3d = VSUB(T2v, T2y); STM2(&(xo[60]), T3d, ovs, &(xo[0])); T3e = VADD(T2z, T2A); STM2(&(xo[28]), T3e, ovs, &(xo[0])); T3f = VADD(T2v, T2y); STM2(&(xo[4]), T3f, ovs, &(xo[0])); T3g = VSUB(T2z, T2A); STM2(&(xo[36]), T3g, ovs, &(xo[0])); } { V T1r, T1C, T1M, T1K, T1F, T1N, T1y, T1J; T1r = VSUB(T1p, T1q); T1C = VSUB(T1A, T1B); T1M = VADD(T1p, T1q); T1K = VADD(T1B, T1A); { V T1D, T1E, T1u, T1x; T1D = VFNMS(LDK(KP195090322), T1s, VMUL(LDK(KP980785280), T1t)); T1E = VFMA(LDK(KP195090322), T1v, VMUL(LDK(KP980785280), T1w)); T1F = VSUB(T1D, T1E); T1N = VADD(T1D, T1E); T1u = VFMA(LDK(KP980785280), T1s, VMUL(LDK(KP195090322), T1t)); T1x = VFNMS(LDK(KP195090322), T1w, VMUL(LDK(KP980785280), T1v)); T1y = VSUB(T1u, T1x); T1J = VADD(T1u, T1x); } { V T1z, T1G, T3h, T3i; T1z = VADD(T1r, T1y); T1G = VBYI(VADD(T1C, T1F)); T3h = VSUB(T1z, T1G); STM2(&(xo[50]), T3h, ovs, &(xo[2])); STN2(&(xo[48]), T39, T3h, ovs); T3i = VADD(T1z, T1G); STM2(&(xo[14]), T3i, ovs, &(xo[2])); STN2(&(xo[12]), T38, T3i, ovs); } { V T1P, T1Q, T3j, T3k; T1P = VBYI(VADD(T1K, T1J)); T1Q = VADD(T1M, T1N); T3j = VADD(T1P, T1Q); STM2(&(xo[2]), T3j, ovs, &(xo[2])); STN2(&(xo[0]), T3a, T3j, ovs); T3k = VSUB(T1Q, T1P); STM2(&(xo[62]), T3k, ovs, &(xo[2])); STN2(&(xo[60]), T3d, T3k, ovs); } { V T1H, T1I, T3l, T3m; T1H = VSUB(T1r, T1y); T1I = VBYI(VSUB(T1F, T1C)); T3l = VSUB(T1H, T1I); STM2(&(xo[46]), T3l, ovs, &(xo[2])); STN2(&(xo[44]), T37, T3l, ovs); T3m = VADD(T1H, T1I); STM2(&(xo[18]), T3m, ovs, &(xo[2])); STN2(&(xo[16]), T3b, T3m, ovs); } { V T1L, T1O, T3n, T3o; T1L = VBYI(VSUB(T1J, T1K)); T1O = VSUB(T1M, T1N); T3n = VADD(T1L, T1O); STM2(&(xo[30]), T3n, ovs, &(xo[2])); STN2(&(xo[28]), T3e, T3n, ovs); T3o = VSUB(T1O, T1L); STM2(&(xo[34]), T3o, ovs, &(xo[2])); STN2(&(xo[32]), T3c, T3o, ovs); } } { V Tr, T1a, T1k, T1i, T1d, T1l, T10, T1h; Tr = VSUB(Tb, Tq); T1a = VSUB(T14, T19); T1k = VADD(Tb, Tq); T1i = VADD(T19, T14); { V T1b, T1c, TI, TZ; T1b = VFNMS(LDK(KP555570233), TC, VMUL(LDK(KP831469612), TH)); T1c = VFMA(LDK(KP555570233), TT, VMUL(LDK(KP831469612), TY)); T1d = VSUB(T1b, T1c); T1l = VADD(T1b, T1c); TI = VFMA(LDK(KP831469612), TC, VMUL(LDK(KP555570233), TH)); TZ = VFNMS(LDK(KP555570233), TY, VMUL(LDK(KP831469612), TT)); T10 = VSUB(TI, TZ); T1h = VADD(TI, TZ); } { V T11, T1e, T3p, T3q; T11 = VADD(Tr, T10); T1e = VBYI(VADD(T1a, T1d)); T3p = VSUB(T11, T1e); STM2(&(xo[54]), T3p, ovs, &(xo[2])); STN2(&(xo[52]), T36, T3p, ovs); T3q = VADD(T11, T1e); STM2(&(xo[10]), T3q, ovs, &(xo[2])); STN2(&(xo[8]), T34, T3q, ovs); } { V T1n, T1o, T3r, T3s; T1n = VBYI(VADD(T1i, T1h)); T1o = VADD(T1k, T1l); T3r = VADD(T1n, T1o); STM2(&(xo[6]), T3r, ovs, &(xo[2])); STN2(&(xo[4]), T3f, T3r, ovs); T3s = VSUB(T1o, T1n); STM2(&(xo[58]), T3s, ovs, &(xo[2])); STN2(&(xo[56]), T32, T3s, ovs); } { V T1f, T1g, T3t, T3u; T1f = VSUB(Tr, T10); T1g = VBYI(VSUB(T1d, T1a)); T3t = VSUB(T1f, T1g); STM2(&(xo[42]), T3t, ovs, &(xo[2])); STN2(&(xo[40]), T33, T3t, ovs); T3u = VADD(T1f, T1g); STM2(&(xo[22]), T3u, ovs, &(xo[2])); STN2(&(xo[20]), T35, T3u, ovs); } { V T1j, T1m, T3v, T3w; T1j = VBYI(VSUB(T1h, T1i)); T1m = VSUB(T1k, T1l); T3v = VADD(T1j, T1m); STM2(&(xo[26]), T3v, ovs, &(xo[2])); STN2(&(xo[24]), T31, T3v, ovs); T3w = VSUB(T1m, T1j); STM2(&(xo[38]), T3w, ovs, &(xo[2])); STN2(&(xo[36]), T3g, T3w, ovs); } } } } }}static const kdft_desc desc = { 32, "n2bv_32", {170, 26, 16, 0}, &GENUS, 0, 2, 0, 0 };void X(codelet_n2bv_32) (planner *p) { X(kdft_register) (p, n2bv_32, &desc);}#endif /* HAVE_FMA */
?? 快捷鍵說明
復制代碼
Ctrl + C
搜索代碼
Ctrl + F
全屏模式
F11
切換主題
Ctrl + Shift + D
顯示快捷鍵
?
增大字號
Ctrl + =
減小字號
Ctrl + -