📄 t2bv_64.c
字号:
} } } } } } { V T5N, T5U, T68, T5z, T6U, T6f; T5N = VFMA(LDK(KP923879532), T5M, T5F); T6L = VFNMS(LDK(KP923879532), T5M, T5F); T6M = VFNMS(LDK(KP923879532), T5T, T5Q); T5U = VFMA(LDK(KP923879532), T5T, T5Q); T68 = VFMA(LDK(KP923879532), T67, T60); T6O = VFNMS(LDK(KP923879532), T67, T60); T5z = VADD(T5p, T5y); T6U = VSUB(T5p, T5y); T6P = VFNMS(LDK(KP923879532), T6e, T6b); T6f = VFMA(LDK(KP923879532), T6e, T6b); { V T5V, T6u, T6g, T6v, T6s, T6J; T6s = VSUB(T6q, T6r); T6J = VADD(T6q, T6r); T5V = VFNMS(LDK(KP098491403), T5U, T5N); T6u = VFMA(LDK(KP098491403), T5N, T5U); T75 = VFMA(LDK(KP980785280), T6U, T6T); T6V = VFNMS(LDK(KP980785280), T6U, T6T); T5A = VFMA(LDK(KP980785280), T5z, T5g); T6A = VFNMS(LDK(KP980785280), T5z, T5g); T6g = VFNMS(LDK(KP098491403), T6f, T68); T6v = VFMA(LDK(KP098491403), T68, T6f); T72 = VFNMS(LDK(KP980785280), T6J, T6I); T6K = VFMA(LDK(KP980785280), T6J, T6I); T6t = VFMA(LDK(KP980785280), T6s, T6p); T6D = VFNMS(LDK(KP980785280), T6s, T6p); T6w = VSUB(T6u, T6v); T6B = VADD(T6u, T6v); T6h = VADD(T5V, T6g); T6E = VSUB(T5V, T6g); } } } } } } } { V T6W, T6N, T6G, T6C, T6z, T6x, T6H, T6F, T6y, T6i, T6X, T6Q; T6W = VFNMS(LDK(KP820678790), T6L, T6M); T6N = VFMA(LDK(KP820678790), T6M, T6L); T6G = VFMA(LDK(KP995184726), T6B, T6A); T6C = VFNMS(LDK(KP995184726), T6B, T6A); T6z = VFMA(LDK(KP995184726), T6w, T6t); T6x = VFNMS(LDK(KP995184726), T6w, T6t); T6H = VFNMS(LDK(KP995184726), T6E, T6D); T6F = VFMA(LDK(KP995184726), T6E, T6D); T6y = VFMA(LDK(KP995184726), T6h, T5A); T6i = VFNMS(LDK(KP995184726), T6h, T5A); T6X = VFNMS(LDK(KP820678790), T6O, T6P); T6Q = VFMA(LDK(KP820678790), T6P, T6O); { V T73, T6Y, T76, T6R; ST(&(x[WS(ios, 49)]), VFMAI(T6H, T6G), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 15)]), VFNMSI(T6H, T6G), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 47)]), VFNMSI(T6F, T6C), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 17)]), VFMAI(T6F, T6C), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 1)]), VFMAI(T6z, T6y), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 63)]), VFNMSI(T6z, T6y), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 33)]), VFMAI(T6x, T6i), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 31)]), VFNMSI(T6x, T6i), dist, &(x[WS(ios, 1)])); T73 = VADD(T6W, T6X); T6Y = VSUB(T6W, T6X); T76 = VSUB(T6N, T6Q); T6R = VADD(T6N, T6Q); { V T78, T74, T71, T6Z, T79, T77, T70, T6S; T78 = VFNMS(LDK(KP773010453), T73, T72); T74 = VFMA(LDK(KP773010453), T73, T72); T71 = VFMA(LDK(KP773010453), T6Y, T6V); T6Z = VFNMS(LDK(KP773010453), T6Y, T6V); T79 = VFMA(LDK(KP773010453), T76, T75); T77 = VFNMS(LDK(KP773010453), T76, T75); T70 = VFMA(LDK(KP773010453), T6R, T6K); T6S = VFNMS(LDK(KP773010453), T6R, T6K); ST(&(x[WS(ios, 55)]), VFNMSI(T79, T78), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 9)]), VFMAI(T79, T78), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 41)]), VFMAI(T77, T74), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 23)]), VFNMSI(T77, T74), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 57)]), VFMAI(T71, T70), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 7)]), VFNMSI(T71, T70), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 25)]), VFMAI(T6Z, T6S), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 39)]), VFNMSI(T6Z, T6S), dist, &(x[WS(ios, 1)])); } } } } return W;}static const tw_instr twinstr[] = { VTW(1), VTW(2), VTW(3), VTW(4), VTW(5), VTW(6), VTW(7), VTW(8), VTW(9), VTW(10), VTW(11), VTW(12), VTW(13), VTW(14), VTW(15), VTW(16), VTW(17), VTW(18), VTW(19), VTW(20), VTW(21), VTW(22), VTW(23), VTW(24), VTW(25), VTW(26), VTW(27), VTW(28), VTW(29), VTW(30), VTW(31), VTW(32), VTW(33), VTW(34), VTW(35), VTW(36), VTW(37), VTW(38), VTW(39), VTW(40), VTW(41), VTW(42), VTW(43), VTW(44), VTW(45), VTW(46), VTW(47), VTW(48), VTW(49), VTW(50), VTW(51), VTW(52), VTW(53), VTW(54), VTW(55), VTW(56), VTW(57), VTW(58), VTW(59), VTW(60), VTW(61), VTW(62), VTW(63), {TW_NEXT, VL, 0}};static const ct_desc desc = { 64, "t2bv_64", twinstr, &GENUS, {261, 126, 258, 0}, 0, 0, 0 };void X(codelet_t2bv_64) (planner *p) { X(kdft_dit_register) (p, t2bv_64, &desc);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_twiddle_c -simd -compact -variables 4 -pipeline-latency 8 -n 64 -name t2bv_64 -include t2b.h -sign 1 *//* * This function contains 519 FP additions, 250 FP multiplications, * (or, 467 additions, 198 multiplications, 52 fused multiply/add), * 107 stack variables, and 128 memory accesses *//* * Generator Id's : * $Id: algsimp.ml,v 1.8 2006-01-05 03:04:27 stevenj Exp $ * $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $ * $Id: gen_twiddle_c.ml,v 1.13 2006-01-05 03:04:27 stevenj Exp $ */#include "t2b.h"static const R *t2bv_64(R *ri, R *ii, const R *W, stride ios, INT m, INT dist){ DVK(KP290284677, +0.290284677254462367636192375817395274691476278); DVK(KP956940335, +0.956940335732208864935797886980269969482849206); DVK(KP471396736, +0.471396736825997648556387625905254377657460319); DVK(KP881921264, +0.881921264348355029712756863660388349508442621); DVK(KP634393284, +0.634393284163645498215171613225493370675687095); DVK(KP773010453, +0.773010453362736960810906609758469800971041293); DVK(KP098017140, +0.098017140329560601994195563888641845861136673); DVK(KP995184726, +0.995184726672196886244836953109479921575474869); DVK(KP195090322, +0.195090322016128267848284868477022240927691618); DVK(KP980785280, +0.980785280403230449126182236134239036973933731); DVK(KP555570233, +0.555570233019602224742830813948532874374937191); DVK(KP831469612, +0.831469612302545237078788377617905756738560812); DVK(KP382683432, +0.382683432365089771728459984030398866761344562); DVK(KP923879532, +0.923879532511286756128183189396788286822416626); DVK(KP707106781, +0.707106781186547524400844362104849039284835938); INT i; R *x; x = ii; for (i = m; i > 0; i = i - VL, x = x + (VL * dist), W = W + (TWVL * 126), MAKE_VOLATILE_STRIDE(ios)) { V Tg, T4B, T6v, T7G, T3r, T4w, T5q, T7F, T5Y, T62, T28, T4d, T2g, T4a, T7g; V T7Y, T6f, T6j, T2Z, T4k, T37, T4h, T7n, T81, T7w, T7x, T7y, T5M, T6q, T1k; V T4s, T1r, T4t, T7t, T7u, T7v, T5F, T6p, TV, T4p, T12, T4q, T7A, T7B, TD; V T4x, T3k, T4C, T5x, T6s, T1R, T4b, T7j, T7Z, T2j, T4e, T5V, T63, T2I, T4i; V T7q, T82, T3a, T4l, T6c, T6k; { V T1, T3, T3p, T3n, Tb, Td, Te, T6, T8, T9, T2, T3o, T3m; T1 = LD(&(x[0]), dist, &(x[0])); T2 = LD(&(x[WS(ios, 32)]), dist, &(x[0])); T3 = BYTW(&(W[TWVL * 62]), T2); T3o = LD(&(x[WS(ios, 48)]), dist, &(x[0])); T3p = BYTW(&(W[TWVL * 94]), T3o); T3m = LD(&(x[WS(ios, 16)]), dist, &(x[0])); T3n = BYTW(&(W[TWVL * 30]), T3m); { V Ta, Tc, T5, T7; Ta = LD(&(x[WS(ios, 56)]), dist, &(x[0])); Tb = BYTW(&(W[TWVL * 110]), Ta); Tc = LD(&(x[WS(ios, 24)]), dist, &(x[0])); Td = BYTW(&(W[TWVL * 46]), Tc); Te = VSUB(Tb, Td); T5 = LD(&(x[WS(ios, 8)]), dist, &(x[0])); T6 = BYTW(&(W[TWVL * 14]), T5); T7 = LD(&(x[WS(ios, 40)]), dist, &(x[0])); T8 = BYTW(&(W[TWVL * 78]), T7); T9 = VSUB(T6, T8); } { V T4, Tf, T6t, T6u; T4 = VSUB(T1, T3); Tf = VMUL(LDK(KP707106781), VADD(T9, Te)); Tg = VSUB(T4, Tf); T4B = VADD(T4, Tf); T6t = VADD(T6, T8); T6u = VADD(Tb, Td); T6v = VSUB(T6t, T6u); T7G = VADD(T6t, T6u); } { V T3l, T3q, T5o, T5p; T3l = VMUL(LDK(KP707106781), VSUB(T9, Te)); T3q = VSUB(T3n, T3p); T3r = VSUB(T3l, T3q); T4w = VADD(T3q, T3l); T5o = VADD(T1, T3); T5p = VADD(T3n, T3p); T5q = VSUB(T5o, T5p); T7F = VADD(T5o, T5p); } } { V T24, T26, T61, T2b, T2d, T60, T1W, T5W, T21, T5X, T22, T27; { V T23, T25, T2a, T2c; T23 = LD(&(x[WS(ios, 17)]), dist, &(x[WS(ios, 1)])); T24 = BYTW(&(W[TWVL * 32]), T23); T25 = LD(&(x[WS(ios, 49)]), dist, &(x[WS(ios, 1)])); T26 = BYTW(&(W[TWVL * 96]), T25); T61 = VADD(T24, T26); T2a = LD(&(x[WS(ios, 1)]), dist, &(x[WS(ios, 1)])); T2b = BYTW(&(W[0]), T2a); T2c = LD(&(x[WS(ios, 33)]), dist, &(x[WS(ios, 1)])); T2d = BYTW(&(W[TWVL * 64]), T2c); T60 = VADD(T2b, T2d); } { V T1T, T1V, T1S, T1U; T1S = LD(&(x[WS(ios, 9)]), dist, &(x[WS(ios, 1)])); T1T = BYTW(&(W[TWVL * 16]), T1S); T1U = LD(&(x[WS(ios, 41)]), dist, &(x[WS(ios, 1)])); T1V = BYTW(&(W[TWVL * 80]), T1U); T1W = VSUB(T1T, T1V); T5W = VADD(T1T, T1V); } { V T1Y, T20, T1X, T1Z; T1X = LD(&(x[WS(ios, 57)]), dist, &(x[WS(ios, 1)])); T1Y = BYTW(&(W[TWVL * 112]), T1X); T1Z = LD(&(x[WS(ios, 25)]), dist, &(x[WS(ios, 1)])); T20 = BYTW(&(W[TWVL * 48]), T1Z); T21 = VSUB(T1Y, T20); T5X = VADD(T1Y, T20); } T5Y = VSUB(T5W, T5X); T62 = VSUB(T60, T61); T22 = VMUL(LDK(KP707106781), VSUB(T1W, T21)); T27 = VSUB(T24, T26); T28 = VSUB(T22, T27); T4d = VADD(T27, T22); { V T2e, T2f, T7e, T7f; T2e = VSUB(T2b, T2d); T2f = VMUL(LDK(KP707106781), VADD(T1W, T21)); T2g = VSUB(T2e, T2f); T4a = VADD(T2e, T2f); T7e = VADD(T60, T61); T7f = VADD(T5W, T5X); T7g = VSUB(T7e, T7f); T7Y = VADD(T7e, T7f); } } { V T2V, T2X, T6i, T32, T34, T6h, T2N, T6d, T2S, T6e, T2T, T2Y; { V T2U, T2W, T31, T33; T2U = LD(&(x[WS(ios, 15)]), dist, &(x[WS(ios, 1)])); T2V = BYTW(&(W[TWVL * 28]), T2U); T2W = LD(&(x[WS(ios, 47)]), dist, &(x[WS(ios, 1)])); T2X = BYTW(&(W[TWVL * 92]), T2W); T6i = VADD(T2V, T2X); T31 = LD(&(x[WS(ios, 63)]), dist, &(x[WS(ios, 1)])); T32 = BYTW(&(W[TWVL * 124]), T31); T33 = LD(&(x[WS(ios, 31)]), dist, &(x[WS(ios, 1)])); T34 = BYTW(&(W[TWVL * 60]), T33); T6h = VADD(T32, T34); } { V T2K, T2M, T2J, T2L; T2J = LD(&(x[WS(ios, 7)]), dist, &(x[WS(ios, 1)])); T2K = BYTW(&(W[TWVL * 12]), T2J); T2L = LD(&(x[WS(ios, 39)]), dist, &(x[WS(ios, 1)])); T2M = BYTW(&(W[TWVL * 76]), T2L); T2N = VSUB(T2K, T2M); T6d = VADD(T2K, T2M); } { V T2P, T2R, T2O, T2Q; T2O = LD(&(x[WS(ios, 55)]), dist, &(x[WS(ios, 1)])); T2P = BYTW(&(W[TWVL * 108]), T2O); T2Q = LD(&(x[WS(ios, 23)]), dist, &(x[WS(ios, 1)])); T2R = BYTW(&(W[TWVL * 44]), T2Q); T2S = VSUB(T2P, T2R); T6e = VADD(T2P, T2R); } T6f = VSUB(T6d, T6e); T6j = VSUB(T6h, T6i); T2T = VMUL(LDK(KP707106781), VSUB(T2N, T2S)); T2Y = VSUB(T2V, T2X); T2Z = VSUB(T2T, T2Y); T4k = VADD(T2Y, T2T); { V T35, T36, T7l, T7m; T35 = VSUB(T32, T34); T36 = VMUL(LDK(KP707106781), VADD(T2N, T2S)); T37 = VSUB(T35, T36); T4h = VADD(T35, T36); T7l = VADD(T6h, T6i); T7m = VADD(T6d, T6e); T7n = VSUB(T7l, T7m); T81 = VADD(T7l, T7m); } } { V T1g, T1i, T5K, T1m, T1o, T5J, T18, T5G, T1d, T5H, T5I, T5L; { V T1f, T1h, T1l, T1n; T1f = LD(&(x[WS(ios, 14)]), dist, &(x[0])); T1g = BYTW(&(W[TWVL * 26]), T1f); T1h = LD(&(x[WS(ios, 46)]), dist, &(x[0])); T1i = BYTW(&(W[TWVL * 90]), T1h); T5K = VADD(T1g, T1i); T1l = LD(&(x[WS(ios, 62)]), dist, &(x[0])); T1m = BYTW(&(W[TWVL * 122]), T1l); T1n = LD(&(x[WS(ios, 30)]), dist, &(x[0])); T1o = BYTW(&(W[TWVL * 58]), T1n); T5J = VADD(T1m, T1o);
⌨️ 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 + -