📄 n2sv_32.c
字号:
for (i = v; i > 0; i = i - (2 * VL), ri = ri + ((2 * VL) * ivs), ii = ii + ((2 * VL) * ivs), ro = ro + ((2 * VL) * ovs), io = io + ((2 * VL) * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) { V T7, T4r, T4Z, T18, T1z, T3t, T3T, T2T, Te, T1f, T50, T4s, T2W, T3u, T1G; V T3U, Tm, T1n, T1O, T2Z, T3y, T3X, T4w, T53, Tt, T1u, T1V, T2Y, T3B, T3W; V T4z, T52, T2t, T3L, T3O, T2K, TR, TY, T5F, T5G, T5H, T5I, T4R, T5j, T2E; V T3P, T4W, T5k, T2N, T3M, T22, T3E, T3H, T2j, TC, TJ, T5A, T5B, T5C, T5D; V T4G, T5g, T2d, T3F, T4L, T5h, T2m, T3I; { V T3, T1x, T14, T2S, T6, T2R, T17, T1y; { V T1, T2, T12, T13; T1 = LD(&(ri[0]), ivs, &(ri[0])); T2 = LD(&(ri[WS(is, 16)]), ivs, &(ri[0])); T3 = VADD(T1, T2); T1x = VSUB(T1, T2); T12 = LD(&(ii[0]), ivs, &(ii[0])); T13 = LD(&(ii[WS(is, 16)]), ivs, &(ii[0])); T14 = VADD(T12, T13); T2S = VSUB(T12, T13); } { V T4, T5, T15, T16; T4 = LD(&(ri[WS(is, 8)]), ivs, &(ri[0])); T5 = LD(&(ri[WS(is, 24)]), ivs, &(ri[0])); T6 = VADD(T4, T5); T2R = VSUB(T4, T5); T15 = LD(&(ii[WS(is, 8)]), ivs, &(ii[0])); T16 = LD(&(ii[WS(is, 24)]), ivs, &(ii[0])); T17 = VADD(T15, T16); T1y = VSUB(T15, T16); } T7 = VADD(T3, T6); T4r = VSUB(T3, T6); T4Z = VSUB(T14, T17); T18 = VADD(T14, T17); T1z = VSUB(T1x, T1y); T3t = VADD(T1x, T1y); T3T = VSUB(T2S, T2R); T2T = VADD(T2R, T2S); } { V Ta, T1B, T1b, T1A, Td, T1D, T1e, T1E; { V T8, T9, T19, T1a; T8 = LD(&(ri[WS(is, 4)]), ivs, &(ri[0])); T9 = LD(&(ri[WS(is, 20)]), ivs, &(ri[0])); Ta = VADD(T8, T9); T1B = VSUB(T8, T9); T19 = LD(&(ii[WS(is, 4)]), ivs, &(ii[0])); T1a = LD(&(ii[WS(is, 20)]), ivs, &(ii[0])); T1b = VADD(T19, T1a); T1A = VSUB(T19, T1a); } { V Tb, Tc, T1c, T1d; Tb = LD(&(ri[WS(is, 28)]), ivs, &(ri[0])); Tc = LD(&(ri[WS(is, 12)]), ivs, &(ri[0])); Td = VADD(Tb, Tc); T1D = VSUB(Tb, Tc); T1c = LD(&(ii[WS(is, 28)]), ivs, &(ii[0])); T1d = LD(&(ii[WS(is, 12)]), ivs, &(ii[0])); T1e = VADD(T1c, T1d); T1E = VSUB(T1c, T1d); } Te = VADD(Ta, Td); T1f = VADD(T1b, T1e); T50 = VSUB(Td, Ta); T4s = VSUB(T1b, T1e); { V T2U, T2V, T1C, T1F; T2U = VSUB(T1D, T1E); T2V = VADD(T1B, T1A); T2W = VMUL(LDK(KP707106781), VSUB(T2U, T2V)); T3u = VMUL(LDK(KP707106781), VADD(T2V, T2U)); T1C = VSUB(T1A, T1B); T1F = VADD(T1D, T1E); T1G = VMUL(LDK(KP707106781), VSUB(T1C, T1F)); T3U = VMUL(LDK(KP707106781), VADD(T1C, T1F)); } } { V Ti, T1L, T1j, T1J, Tl, T1I, T1m, T1M, T1K, T1N; { V Tg, Th, T1h, T1i; Tg = LD(&(ri[WS(is, 2)]), ivs, &(ri[0])); Th = LD(&(ri[WS(is, 18)]), ivs, &(ri[0])); Ti = VADD(Tg, Th); T1L = VSUB(Tg, Th); T1h = LD(&(ii[WS(is, 2)]), ivs, &(ii[0])); T1i = LD(&(ii[WS(is, 18)]), ivs, &(ii[0])); T1j = VADD(T1h, T1i); T1J = VSUB(T1h, T1i); } { V Tj, Tk, T1k, T1l; Tj = LD(&(ri[WS(is, 10)]), ivs, &(ri[0])); Tk = LD(&(ri[WS(is, 26)]), ivs, &(ri[0])); Tl = VADD(Tj, Tk); T1I = VSUB(Tj, Tk); T1k = LD(&(ii[WS(is, 10)]), ivs, &(ii[0])); T1l = LD(&(ii[WS(is, 26)]), ivs, &(ii[0])); T1m = VADD(T1k, T1l); T1M = VSUB(T1k, T1l); } Tm = VADD(Ti, Tl); T1n = VADD(T1j, T1m); T1K = VADD(T1I, T1J); T1N = VSUB(T1L, T1M); T1O = VFNMS(LDK(KP923879532), T1N, VMUL(LDK(KP382683432), T1K)); T2Z = VFMA(LDK(KP923879532), T1K, VMUL(LDK(KP382683432), T1N)); { V T3w, T3x, T4u, T4v; T3w = VSUB(T1J, T1I); T3x = VADD(T1L, T1M); T3y = VFNMS(LDK(KP382683432), T3x, VMUL(LDK(KP923879532), T3w)); T3X = VFMA(LDK(KP382683432), T3w, VMUL(LDK(KP923879532), T3x)); T4u = VSUB(T1j, T1m); T4v = VSUB(Ti, Tl); T4w = VSUB(T4u, T4v); T53 = VADD(T4v, T4u); } } { V Tp, T1S, T1q, T1Q, Ts, T1P, T1t, T1T, T1R, T1U; { V Tn, To, T1o, T1p; Tn = LD(&(ri[WS(is, 30)]), ivs, &(ri[0])); To = LD(&(ri[WS(is, 14)]), ivs, &(ri[0])); Tp = VADD(Tn, To); T1S = VSUB(Tn, To); T1o = LD(&(ii[WS(is, 30)]), ivs, &(ii[0])); T1p = LD(&(ii[WS(is, 14)]), ivs, &(ii[0])); T1q = VADD(T1o, T1p); T1Q = VSUB(T1o, T1p); } { V Tq, Tr, T1r, T1s; Tq = LD(&(ri[WS(is, 6)]), ivs, &(ri[0])); Tr = LD(&(ri[WS(is, 22)]), ivs, &(ri[0])); Ts = VADD(Tq, Tr); T1P = VSUB(Tq, Tr); T1r = LD(&(ii[WS(is, 6)]), ivs, &(ii[0])); T1s = LD(&(ii[WS(is, 22)]), ivs, &(ii[0])); T1t = VADD(T1r, T1s); T1T = VSUB(T1r, T1s); } Tt = VADD(Tp, Ts); T1u = VADD(T1q, T1t); T1R = VADD(T1P, T1Q); T1U = VSUB(T1S, T1T); T1V = VFMA(LDK(KP382683432), T1R, VMUL(LDK(KP923879532), T1U)); T2Y = VFNMS(LDK(KP923879532), T1R, VMUL(LDK(KP382683432), T1U)); { V T3z, T3A, T4x, T4y; T3z = VSUB(T1Q, T1P); T3A = VADD(T1S, T1T); T3B = VFMA(LDK(KP923879532), T3z, VMUL(LDK(KP382683432), T3A)); T3W = VFNMS(LDK(KP382683432), T3z, VMUL(LDK(KP923879532), T3A)); T4x = VSUB(Tp, Ts); T4y = VSUB(T1q, T1t); T4z = VADD(T4x, T4y); T52 = VSUB(T4x, T4y); } } { V TN, T2p, T2J, T4S, TQ, T2G, T2s, T4T, TU, T2x, T2w, T4O, TX, T2z, T2C; V T4P; { V TL, TM, T2H, T2I; TL = LD(&(ri[WS(is, 31)]), ivs, &(ri[WS(is, 1)])); TM = LD(&(ri[WS(is, 15)]), ivs, &(ri[WS(is, 1)])); TN = VADD(TL, TM); T2p = VSUB(TL, TM); T2H = LD(&(ii[WS(is, 31)]), ivs, &(ii[WS(is, 1)])); T2I = LD(&(ii[WS(is, 15)]), ivs, &(ii[WS(is, 1)])); T2J = VSUB(T2H, T2I); T4S = VADD(T2H, T2I); } { V TO, TP, T2q, T2r; TO = LD(&(ri[WS(is, 7)]), ivs, &(ri[WS(is, 1)])); TP = LD(&(ri[WS(is, 23)]), ivs, &(ri[WS(is, 1)])); TQ = VADD(TO, TP); T2G = VSUB(TO, TP); T2q = LD(&(ii[WS(is, 7)]), ivs, &(ii[WS(is, 1)])); T2r = LD(&(ii[WS(is, 23)]), ivs, &(ii[WS(is, 1)])); T2s = VSUB(T2q, T2r); T4T = VADD(T2q, T2r); } { V TS, TT, T2u, T2v; TS = LD(&(ri[WS(is, 3)]), ivs, &(ri[WS(is, 1)])); TT = LD(&(ri[WS(is, 19)]), ivs, &(ri[WS(is, 1)])); TU = VADD(TS, TT); T2x = VSUB(TS, TT); T2u = LD(&(ii[WS(is, 3)]), ivs, &(ii[WS(is, 1)])); T2v = LD(&(ii[WS(is, 19)]), ivs, &(ii[WS(is, 1)])); T2w = VSUB(T2u, T2v); T4O = VADD(T2u, T2v); } { V TV, TW, T2A, T2B; TV = LD(&(ri[WS(is, 27)]), ivs, &(ri[WS(is, 1)])); TW = LD(&(ri[WS(is, 11)]), ivs, &(ri[WS(is, 1)])); TX = VADD(TV, TW); T2z = VSUB(TV, TW); T2A = LD(&(ii[WS(is, 27)]), ivs, &(ii[WS(is, 1)])); T2B = LD(&(ii[WS(is, 11)]), ivs, &(ii[WS(is, 1)])); T2C = VSUB(T2A, T2B); T4P = VADD(T2A, T2B); } T2t = VSUB(T2p, T2s); T3L = VADD(T2p, T2s); T3O = VSUB(T2J, T2G); T2K = VADD(T2G, T2J); TR = VADD(TN, TQ); TY = VADD(TU, TX); T5F = VSUB(TR, TY); { V T4N, T4Q, T2y, T2D; T5G = VADD(T4S, T4T); T5H = VADD(T4O, T4P); T5I = VSUB(T5G, T5H); T4N = VSUB(TN, TQ); T4Q = VSUB(T4O, T4P); T4R = VSUB(T4N, T4Q); T5j = VADD(T4N, T4Q); T2y = VSUB(T2w, T2x); T2D = VADD(T2z, T2C); T2E = VMUL(LDK(KP707106781), VSUB(T2y, T2D)); T3P = VMUL(LDK(KP707106781), VADD(T2y, T2D)); { V T4U, T4V, T2L, T2M; T4U = VSUB(T4S, T4T); T4V = VSUB(TX, TU); T4W = VSUB(T4U, T4V); T5k = VADD(T4V, T4U); T2L = VSUB(T2z, T2C); T2M = VADD(T2x, T2w); T2N = VMUL(LDK(KP707106781), VSUB(T2L, T2M)); T3M = VMUL(LDK(KP707106781), VADD(T2M, T2L)); } } } { V Ty, T2f, T21, T4C, TB, T1Y, T2i, T4D, TF, T28, T2b, T4I, TI, T23, T26; V T4J; { V Tw, Tx, T1Z, T20; Tw = LD(&(ri[WS(is, 1)]), ivs, &(ri[WS(is, 1)])); Tx = LD(&(ri[WS(is, 17)]), ivs, &(ri[WS(is, 1)])); Ty = VADD(Tw, Tx); T2f = VSUB(Tw, Tx); T1Z = LD(&(ii[WS(is, 1)]), ivs, &(ii[WS(is, 1)])); T20 = LD(&(ii[WS(is, 17)]), ivs, &(ii[WS(is, 1)])); T21 = VSUB(T1Z, T20); T4C = VADD(T1Z, T20); } { V Tz, TA, T2g, T2h; Tz = LD(&(ri[WS(is, 9)]), ivs, &(ri[WS(is, 1)])); TA = LD(&(ri[WS(is, 25)]), ivs, &(ri[WS(is, 1)])); TB = VADD(Tz, TA); T1Y = VSUB(Tz, TA); T2g = LD(&(ii[WS(is, 9)]), ivs, &(ii[WS(is, 1)])); T2h = LD(&(ii[WS(is, 25)]), ivs, &(ii[WS(is, 1)])); T2i = VSUB(T2g, T2h); T4D = VADD(T2g, T2h); } { V TD, TE, T29, T2a; TD = LD(&(ri[WS(is, 5)]), ivs, &(ri[WS(is, 1)])); TE = LD(&(ri[WS(is, 21)]), ivs, &(ri[WS(is, 1)])); TF = VADD(TD, TE); T28 = VSUB(TD, TE); T29 = LD(&(ii[WS(is, 5)]), ivs, &(ii[WS(is, 1)])); T2a = LD(&(ii[WS(is, 21)]), ivs, &(ii[WS(is, 1)])); T2b = VSUB(T29, T2a); T4I = VADD(T29, T2a); } { V TG, TH, T24, T25; TG = LD(&(ri[WS(is, 29)]), ivs, &(ri[WS(is, 1)])); TH = LD(&(ri[WS(is, 13)]), ivs, &(ri[WS(is, 1)])); TI = VADD(TG, TH); T23 = VSUB(TG, TH); T24 = LD(&(ii[WS(is, 29)]), ivs, &(ii[WS(is, 1)])); T25 = LD(&(ii[WS(is, 13)]), ivs, &(ii[WS(is, 1)])); T26 = VSUB(T24, T25); T4J = VADD(T24, T25); } T22 = VADD(T1Y, T21); T3E = VADD(T2f, T2i); T3H = VSUB(T21, T1Y); T2j = VSUB(T2f, T2i); TC = VADD(Ty, TB); TJ = VADD(TF, TI); T5A = VSUB(TC, TJ); { V T4E, T4F, T27, T2c; T5B = VADD(T4C, T4D); T5C = VADD(T4I, T4J); T5D = VSUB(T5B, T5C); T4E = VSUB(T4C, T4D); T4F = VSUB(TI, TF); T4G = VSUB(T4E, T4F); T5g = VADD(T4F, T4E); T27 = VSUB(T23, T26); T2c = VADD(T28, T2b); T2d = VMUL(LDK(KP707106781), VSUB(T27, T2c)); T3F = VMUL(LDK(KP707106781), VADD(T2c, T27)); { V T4H, T4K, T2k, T2l; T4H = VSUB(Ty, TB); T4K = VSUB(T4I, T4J); T4L = VSUB(T4H, T4K); T5h = VADD(T4H, T4K); T2k = VSUB(T2b, T28); T2l = VADD(T23, T26); T2m = VMUL(LDK(KP707106781), VSUB(T2k, T2l)); T3I = VMUL(LDK(KP707106781), VADD(T2k, T2l)); } } } { V T61, T62, T63, T64, T65, T66, T67, T68, T69, T6a, T6b, T6c, T6d, T6e, T6f; V T6g, T6h, T6i, T6j, T6k, T6l, T6m, T6n, T6o, T6p, T6q, T6r, T6s, T6t, T6u; V T6v, T6w; { V T4B, T57, T5a, T5c, T4Y, T56, T55, T5b; { V T4t, T4A, T58, T59; T4t = VSUB(T4r, T4s); T4A = VMUL(LDK(KP707106781), VSUB(T4w, T4z)); T4B = VADD(T4t, T4A); T57 = VSUB(T4t, T4A); T58 = VFNMS(LDK(KP923879532), T4L, VMUL(LDK(KP382683432), T4G)); T59 = VFMA(LDK(KP382683432), T4W, VMUL(LDK(KP923879532), T4R)); T5a = VSUB(T58, T59); T5c = VADD(T58, T59); } { V T4M, T4X, T51, T54; T4M = VFMA(LDK(KP923879532), T4G, VMUL(LDK(KP382683432), T4L)); T4X = VFNMS(LDK(KP923879532), T4W, VMUL(LDK(KP382683432), T4R)); T4Y = VADD(T4M, T4X); T56 = VSUB(T4X, T4M); T51 = VSUB(T4Z, T50); T54 = VMUL(LDK(KP707106781), VSUB(T52, T53)); T55 = VSUB(T51, T54); T5b = VADD(T51, T54); } T61 = VSUB(T4B, T4Y); STM4(&(ro[22]), T61, ovs, &(ro[0])); T62 = VSUB(T5b, T5c); STM4(&(io[22]), T62, ovs, &(io[0])); T63 = VADD(T4B, T4Y); STM4(&(ro[6]), T63, ovs, &(ro[0])); T64 = VADD(T5b, T5c); STM4(&(io[6]), T64, ovs, &(io[0])); T65 = VSUB(T55, T56); STM4(&(io[30]), T65, ovs, &(io[0])); T66 = VSUB(T57, T5a);
⌨️ 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 + -