📄 n2bv_64.c
字号:
{ V T2H, T2P, T81, T84, T86, T88, T2J, T2B, T2R, T2N; { V T2L, T2t, T2M, T2A; { V T2w, T2F, T6Q, T6G, T2G, T2z; T2w = VFMA(LDK(KP668178637), T2v, T2u); T2F = VFNMS(LDK(KP668178637), T2u, T2v); T6Q = VFNMS(LDK(KP831469612), T6F, T6C); T6G = VFMA(LDK(KP831469612), T6F, T6C); T2G = VFNMS(LDK(KP668178637), T2x, T2y); T2z = VFMA(LDK(KP668178637), T2y, T2x); T2L = VFNMS(LDK(KP923879532), T2s, T2r); T2t = VFMA(LDK(KP923879532), T2s, T2r); { V T6S, T6U, T6M, T6K; T6S = VFMA(LDK(KP881921264), T6R, T6Q); T6U = VFNMS(LDK(KP881921264), T6R, T6Q); T6M = VFMA(LDK(KP881921264), T6J, T6G); T6K = VFNMS(LDK(KP881921264), T6J, T6G); T2M = VADD(T2F, T2G); T2H = VSUB(T2F, T2G); T2P = VSUB(T2w, T2z); T2A = VADD(T2w, T2z); T81 = VFNMSI(T6S, T6P); STM2(&(xo[86]), T81, ovs, &(xo[2])); { V T82, T83, T85, T87; T82 = VFMAI(T6S, T6P); STM2(&(xo[42]), T82, ovs, &(xo[2])); STN2(&(xo[40]), T7s, T82, ovs); T83 = VFMAI(T6U, T6T); STM2(&(xo[106]), T83, ovs, &(xo[2])); STN2(&(xo[104]), T7t, T83, ovs); T84 = VFNMSI(T6U, T6T); STM2(&(xo[22]), T84, ovs, &(xo[2])); T85 = VFMAI(T6M, T6L); STM2(&(xo[10]), T85, ovs, &(xo[2])); STN2(&(xo[8]), T7v, T85, ovs); T86 = VFNMSI(T6M, T6L); STM2(&(xo[118]), T86, ovs, &(xo[2])); T87 = VFMAI(T6K, T6z); STM2(&(xo[74]), T87, ovs, &(xo[2])); STN2(&(xo[72]), T7x, T87, ovs); T88 = VFNMSI(T6K, T6z); STM2(&(xo[54]), T88, ovs, &(xo[2])); } } } T2J = VFMA(LDK(KP831469612), T2A, T2t); T2B = VFNMS(LDK(KP831469612), T2A, T2t); T2R = VFNMS(LDK(KP831469612), T2M, T2L); T2N = VFMA(LDK(KP831469612), T2M, T2L); } { V T61, T5J, T62, T5Q; { V T5M, T5V, T2O, T2E, T5W, T5P; T5M = VFMA(LDK(KP820678790), T5L, T5K); T5V = VFNMS(LDK(KP820678790), T5K, T5L); T2O = VFMA(LDK(KP923879532), T2D, T2C); T2E = VFNMS(LDK(KP923879532), T2D, T2C); T5W = VFNMS(LDK(KP820678790), T5N, T5O); T5P = VFMA(LDK(KP820678790), T5O, T5N); T61 = VFNMS(LDK(KP980785280), T5I, T5H); T5J = VFMA(LDK(KP980785280), T5I, T5H); { V T2Q, T2S, T2K, T2I; T2Q = VFNMS(LDK(KP831469612), T2P, T2O); T2S = VFMA(LDK(KP831469612), T2P, T2O); T2K = VFMA(LDK(KP831469612), T2H, T2E); T2I = VFNMS(LDK(KP831469612), T2H, T2E); T62 = VADD(T5V, T5W); T5X = VSUB(T5V, T5W); T65 = VSUB(T5M, T5P); T5Q = VADD(T5M, T5P); { V T89, T8c, T8d, T8f; T89 = VFMAI(T2Q, T2N); STM2(&(xo[84]), T89, ovs, &(xo[0])); STN2(&(xo[84]), T89, T81, ovs); T8a = VFNMSI(T2Q, T2N); STM2(&(xo[44]), T8a, ovs, &(xo[0])); T8b = VFNMSI(T2S, T2R); STM2(&(xo[108]), T8b, ovs, &(xo[0])); T8c = VFMAI(T2S, T2R); STM2(&(xo[20]), T8c, ovs, &(xo[0])); STN2(&(xo[20]), T8c, T84, ovs); T8d = VFMAI(T2K, T2J); STM2(&(xo[116]), T8d, ovs, &(xo[0])); STN2(&(xo[116]), T8d, T86, ovs); T8e = VFNMSI(T2K, T2J); STM2(&(xo[12]), T8e, ovs, &(xo[0])); T8f = VFMAI(T2I, T2B); STM2(&(xo[52]), T8f, ovs, &(xo[0])); STN2(&(xo[52]), T8f, T88, ovs); T8g = VFNMSI(T2I, T2B); STM2(&(xo[76]), T8g, ovs, &(xo[0])); } } } T5Z = VFMA(LDK(KP773010453), T5Q, T5J); T5R = VFNMS(LDK(KP773010453), T5Q, T5J); T67 = VFNMS(LDK(KP773010453), T62, T61); T63 = VFMA(LDK(KP773010453), T62, T61); } } } } } } T5U = VFNMS(LDK(KP980785280), T5T, T5S); T64 = VFMA(LDK(KP980785280), T5T, T5S); { V T68, T66, T5Y, T60; T68 = VFMA(LDK(KP773010453), T65, T64); T66 = VFNMS(LDK(KP773010453), T65, T64); T5Y = VFNMS(LDK(KP773010453), T5X, T5U); T60 = VFMA(LDK(KP773010453), T5X, T5U); { V T8h, T8i, T8j, T8k; T8h = VFMAI(T66, T63); STM2(&(xo[82]), T8h, ovs, &(xo[2])); STN2(&(xo[80]), T7B, T8h, ovs); T8i = VFNMSI(T66, T63); STM2(&(xo[46]), T8i, ovs, &(xo[2])); STN2(&(xo[44]), T8a, T8i, ovs); T8j = VFNMSI(T68, T67); STM2(&(xo[110]), T8j, ovs, &(xo[2])); STN2(&(xo[108]), T8b, T8j, ovs); T8k = VFMAI(T68, T67); STM2(&(xo[18]), T8k, ovs, &(xo[2])); STN2(&(xo[16]), T7A, T8k, ovs); { V T8l, T8m, T8n, T8o; T8l = VFMAI(T60, T5Z); STM2(&(xo[114]), T8l, ovs, &(xo[2])); STN2(&(xo[112]), T7z, T8l, ovs); T8m = VFNMSI(T60, T5Z); STM2(&(xo[14]), T8m, ovs, &(xo[2])); STN2(&(xo[12]), T8e, T8m, ovs); T8n = VFMAI(T5Y, T5R); STM2(&(xo[50]), T8n, ovs, &(xo[2])); STN2(&(xo[48]), T7C, T8n, ovs); T8o = VFNMSI(T5Y, T5R); STM2(&(xo[78]), T8o, ovs, &(xo[2])); STN2(&(xo[76]), T8g, T8o, ovs); } } } }}static const kdft_desc desc = { 64, "n2bv_64", {198, 0, 258, 0}, &GENUS, 0, 2, 0, 0 };void X(codelet_n2bv_64) (planner *p) { X(kdft_register) (p, n2bv_64, &desc);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_notw_c -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 64 -name n2bv_64 -with-ostride 2 -include n2b.h -store-multiple 2 *//* * This function contains 456 FP additions, 124 FP multiplications, * (or, 404 additions, 72 multiplications, 52 fused multiply/add), * 128 stack variables, 15 constants, and 160 memory accesses */#include "n2b.h"static void n2bv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs){ DVK(KP634393284, +0.634393284163645498215171613225493370675687095); DVK(KP773010453, +0.773010453362736960810906609758469800971041293); DVK(KP290284677, +0.290284677254462367636192375817395274691476278); DVK(KP956940335, +0.956940335732208864935797886980269969482849206); DVK(KP098017140, +0.098017140329560601994195563888641845861136673); DVK(KP995184726, +0.995184726672196886244836953109479921575474869); DVK(KP471396736, +0.471396736825997648556387625905254377657460319); DVK(KP881921264, +0.881921264348355029712756863660388349508442621); DVK(KP195090322, +0.195090322016128267848284868477022240927691618); DVK(KP980785280, +0.980785280403230449126182236134239036973933731); DVK(KP555570233, +0.555570233019602224742830813948532874374937191); DVK(KP831469612, +0.831469612302545237078788377617905756738560812); DVK(KP923879532, +0.923879532511286756128183189396788286822416626); DVK(KP382683432, +0.382683432365089771728459984030398866761344562); 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 T4p, T5u, Tb, T3A, T2q, T3v, T6G, T78, Tq, T3w, T6B, T79, T2l, T3B, T4w; V T5r, TI, T2g, T6u, T74, T3q, T3D, T4E, T5o, TZ, T2h, T6x, T75, T3t, T3E; V T4L, T5p, T23, T2N, T6m, T70, T6p, T71, T2c, T2O, T3i, T3Y, T5f, T5R, T5k; V T5S, T3l, T3Z, T1s, T2K, T6f, T6X, T6i, T6Y, T1B, T2L, T3b, T3V, T4Y, T5O; V T53, T5P, T3e, T3W; { V T3, T4n, T2p, T4o, T6, T5s, T9, T5t; { V T1, T2, T2n, T2o; T1 = LD(&(xi[0]), ivs, &(xi[0])); T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0])); T3 = VSUB(T1, T2); T4n = VADD(T1, T2); T2n = LD(&(xi[WS(is, 16)]), ivs, &(xi[0])); T2o = LD(&(xi[WS(is, 48)]), ivs, &(xi[0])); T2p = VSUB(T2n, T2o); T4o = VADD(T2n, T2o); } { V T4, T5, T7, T8; T4 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0])); T5 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0])); T6 = VSUB(T4, T5); T5s = VADD(T4, T5); T7 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0])); T8 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0])); T9 = VSUB(T7, T8); T5t = VADD(T7, T8); } T4p = VSUB(T4n, T4o); T5u = VSUB(T5s, T5t); { V Ta, T2m, T6E, T6F; Ta = VMUL(LDK(KP707106781), VADD(T6, T9)); Tb = VSUB(T3, Ta); T3A = VADD(T3, Ta); T2m = VMUL(LDK(KP707106781), VSUB(T6, T9)); T2q = VSUB(T2m, T2p); T3v = VADD(T2p, T2m); T6E = VADD(T4n, T4o); T6F = VADD(T5s, T5t); T6G = VSUB(T6E, T6F); T78 = VADD(T6E, T6F); } } { V Te, T4q, To, T4t, Th, T4r, Tl, T4u; { V Tc, Td, Tm, Tn; Tc = LD(&(xi[WS(is, 4)]), ivs, &(xi[0])); Td = LD(&(xi[WS(is, 36)]), ivs, &(xi[0])); Te = VSUB(Tc, Td); T4q = VADD(Tc, Td); Tm = LD(&(xi[WS(is, 60)]), ivs, &(xi[0])); Tn = LD(&(xi[WS(is, 28)]), ivs, &(xi[0])); To = VSUB(Tm, Tn); T4t = VADD(Tm, Tn); } { V Tf, Tg, Tj, Tk; Tf = LD(&(xi[WS(is, 20)]), ivs, &(xi[0])); Tg = LD(&(xi[WS(is, 52)]), ivs, &(xi[0])); Th = VSUB(Tf, Tg); T4r = VADD(Tf, Tg); Tj = LD(&(xi[WS(is, 12)]), ivs, &(xi[0])); Tk = LD(&(xi[WS(is, 44)]), ivs, &(xi[0])); Tl = VSUB(Tj, Tk); T4u = VADD(Tj, Tk); } { V Ti, Tp, T6z, T6A; Ti = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th)); Tp = VFNMS(LDK(KP382683432), To, VMUL(LDK(KP923879532), Tl)); Tq = VSUB(Ti, Tp); T3w = VADD(Ti, Tp); T6z = VADD(T4q, T4r); T6A = VADD(T4t, T4u); T6B = VSUB(T6z, T6A); T79 = VADD(T6z, T6A); } { V T2j, T2k, T4s, T4v; T2j = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te)); T2k = VFMA(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl)); T2l = VSUB(T2j, T2k); T3B = VADD(T2j, T2k); T4s = VSUB(T4q, T4r); T4v = VSUB(T4t, T4u); T4w = VMUL(LDK(KP707106781), VADD(T4s, T4v)); T5r = VMUL(LDK(KP707106781), VSUB(T4s, T4v)); } } { V TB, T4z, TF, T4y, Ty, T4C, TG, T4B; { V Tz, TA, TD, TE; Tz = LD(&(xi[WS(is, 18)]), ivs, &(xi[0])); TA = LD(&(xi[WS(is, 50)]), ivs, &(xi[0])); TB = VSUB(Tz, TA); T4z = VADD(Tz, TA); TD = LD(&(xi[WS(is, 2)]), ivs, &(xi[0])); TE = LD(&(xi[WS(is, 34)]), ivs, &(xi[0])); TF = VSUB(TD, TE); T4y = VADD(TD, TE); { V Ts, Tt, Tu, Tv, Tw, Tx; Ts = LD(&(xi[WS(is, 10)]), ivs, &(xi[0])); Tt = LD(&(xi[WS(is, 42)]), ivs, &(xi[0])); Tu = VSUB(Ts, Tt); Tv = LD(&(xi[WS(is, 58)]), ivs, &(xi[0])); Tw = LD(&(xi[WS(is, 26)]), ivs, &(xi[0])); Tx = VSUB(Tv, Tw); Ty = VMUL(LDK(KP707106781), VSUB(Tu, Tx)); T4C = VADD(Tv, Tw); TG = VMUL(LDK(KP707106781), VADD(Tu, Tx)); T4B = VADD(Ts, Tt); } } { V TC, TH, T6s, T6t; TC = VSUB(Ty, TB); TH = VSUB(TF, TG); TI = VFMA(LDK(KP831469612), TC, VMUL(LDK(KP555570233), TH)); T2g = VFNMS(LDK(KP555570233), TC, VMUL(LDK(KP831469612), TH)); T6s = VADD(T4y, T4z); T6t = VADD(T4B, T4C); T6u = VSUB(T6s, T6t); T74 = VADD(T6s, T6t); } { V T3o, T3p, T4A, T4D; T3o = VADD(TB, Ty); T3p = VADD(TF, TG); T3q = VFMA(LDK(KP980785280), T3o, VMUL(LDK(KP195090322), T3p)); T3D = VFNMS(LDK(KP195090322), T3o, VMUL(LDK(KP980785280), T3p)); T4A = VSUB(T4y, T4z); T4D = VSUB(T4B, T4C); T4E = VFMA(LDK(KP382683432), T4A, VMUL(LDK(KP923879532), T4D)); T5o = VFNMS(LDK(KP382683432), T4D, VMUL(LDK(KP923879532), T4A)); } } { V TS, T4J, TW, T4I, TP, T4G, TX, T4F; { V TQ, TR, TU, TV; TQ = LD(&(xi[WS(is, 14)]), ivs, &(xi[0])); TR = LD(&(xi[WS(is, 46)]), ivs, &(xi[0])); TS = VSUB(TQ, TR); T4J = VADD(TQ, TR); TU = LD(&(xi[WS(is, 62)]), ivs, &(xi[0])); TV = LD(&(xi[WS(is, 30)]), ivs, &(xi[0])); TW = VSUB(TU, TV); T4I = VADD(TU, TV); { V TJ, TK, TL, TM, TN, TO; TJ = LD(&(xi[WS(is, 6)]), ivs, &(xi[0])); TK = LD(&(xi[WS(is, 38)]), ivs, &(xi[0])); TL = VSUB(TJ, TK); TM = LD(&(xi[WS(is, 54)]), ivs, &(xi[0])); TN = LD(&(xi[WS(is, 22)]), ivs, &(xi[0])); TO = VSUB(TM, TN); TP = VMUL(LDK(KP707106781), VSUB(TL, TO)); T4G = VADD(TM, TN); TX = VMUL(LDK(KP707106781), VADD(TL, TO)); T4F = VADD(TJ, TK); } } { V TT, TY, T6v, T6w; TT = VSUB(TP, TS);
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -