📄 t3fv_32.c
字号:
* (or, 228 additions, 142 multiplications, 16 fused multiply/add), * 90 stack variables, and 64 memory accesses *//* * Generator Id's : * $Id: algsimp.ml,v 1.9 2006-02-12 23:34:12 athena Exp $ * $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $ * $Id: gen_twiddle_c.ml,v 1.14 2006-02-12 23:34:12 athena Exp $ */#include "t3f.h"static const R *t3fv_32(R *ri, R *ii, const R *W, stride ios, INT m, INT dist){ 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; R *x; x = ri; for (i = m; i > 0; i = i - VL, x = x + (VL * dist), W = W + (TWVL * 8), MAKE_VOLATILE_STRIDE(ios)) { V T2, T5, T3, T4, Tc, T1C, TP, Tz, Tn, T6, TS, Tf, TK, T7, T8; V Tv, T1w, T22, Tg, Tk, T1D, T1R, TC, T18, T12, T1t, TH, TL, TT, T1n; V T1c; T2 = LDW(&(W[0])); T5 = LDW(&(W[TWVL * 4])); T3 = LDW(&(W[TWVL * 2])); T4 = VZMULJ(T2, T3); Tc = VZMUL(T2, T3); T1C = VZMULJ(T2, T5); TP = VZMULJ(T3, T5); Tz = VZMUL(T2, T5); Tn = VZMUL(T3, T5); T6 = VZMUL(T4, T5); TS = VZMULJ(Tc, T5); Tf = VZMULJ(T4, T5); TK = VZMUL(Tc, T5); T7 = LDW(&(W[TWVL * 6])); T8 = VZMULJ(T6, T7); Tv = VZMULJ(T5, T7); T1w = VZMULJ(Tn, T7); T22 = VZMULJ(T3, T7); Tg = VZMULJ(Tf, T7); Tk = VZMUL(T2, T7); T1D = VZMULJ(T1C, T7); T1R = VZMULJ(Tc, T7); TC = VZMULJ(T2, T7); T18 = VZMULJ(TP, T7); T12 = VZMULJ(Tz, T7); T1t = VZMUL(Tc, T7); TH = VZMUL(T3, T7); TL = VZMULJ(TK, T7); TT = VZMULJ(TS, T7); T1n = VZMULJ(T4, T7); T1c = VZMUL(T4, T7); { V Tb, T25, T2T, T3x, Tr, T1Z, T2W, T3K, TX, T27, T3g, T3z, TG, T28, T3j; V T3y, T1N, T2v, T3a, T3G, T1V, T2w, T37, T3F, T1j, T2s, T33, T3D, T1r, T2t; V T30, T3C; { V T1, T24, Ta, T21, T23, T9, T20, T2R, T2S; T1 = LD(&(x[0]), dist, &(x[0])); T23 = LD(&(x[WS(ios, 24)]), dist, &(x[0])); T24 = VZMULJ(T22, T23); T9 = LD(&(x[WS(ios, 16)]), dist, &(x[0])); Ta = VZMULJ(T8, T9); T20 = LD(&(x[WS(ios, 8)]), dist, &(x[0])); T21 = VZMULJ(T1C, T20); Tb = VSUB(T1, Ta); T25 = VSUB(T21, T24); T2R = VADD(T1, Ta); T2S = VADD(T21, T24); T2T = VADD(T2R, T2S); T3x = VSUB(T2R, T2S); } { V Te, Tp, Ti, Tm; { V Td, To, Th, Tl; Td = LD(&(x[WS(ios, 4)]), dist, &(x[0])); Te = VZMULJ(Tc, Td); To = LD(&(x[WS(ios, 12)]), dist, &(x[0])); Tp = VZMULJ(Tn, To); Th = LD(&(x[WS(ios, 20)]), dist, &(x[0])); Ti = VZMULJ(Tg, Th); Tl = LD(&(x[WS(ios, 28)]), dist, &(x[0])); Tm = VZMULJ(Tk, Tl); } { V Tj, Tq, T2U, T2V; Tj = VSUB(Te, Ti); Tq = VSUB(Tm, Tp); Tr = VMUL(LDK(KP707106781), VADD(Tj, Tq)); T1Z = VMUL(LDK(KP707106781), VSUB(Tq, Tj)); T2U = VADD(Te, Ti); T2V = VADD(Tm, Tp); T2W = VADD(T2U, T2V); T3K = VSUB(T2V, T2U); } } { V TJ, TV, TN, TR; { V TI, TU, TM, TQ; TI = LD(&(x[WS(ios, 30)]), dist, &(x[0])); TJ = VZMULJ(TH, TI); TU = LD(&(x[WS(ios, 22)]), dist, &(x[0])); TV = VZMULJ(TT, TU); TM = LD(&(x[WS(ios, 14)]), dist, &(x[0])); TN = VZMULJ(TL, TM); TQ = LD(&(x[WS(ios, 6)]), dist, &(x[0])); TR = VZMULJ(TP, TQ); } { V TO, TW, T3e, T3f; TO = VSUB(TJ, TN); TW = VSUB(TR, TV); TX = VFMA(LDK(KP923879532), TO, VMUL(LDK(KP382683432), TW)); T27 = VFNMS(LDK(KP923879532), TW, VMUL(LDK(KP382683432), TO)); T3e = VADD(TJ, TN); T3f = VADD(TR, TV); T3g = VADD(T3e, T3f); T3z = VSUB(T3e, T3f); } } { V Tu, TE, Tx, TB; { V Tt, TD, Tw, TA; Tt = LD(&(x[WS(ios, 2)]), dist, &(x[0])); Tu = VZMULJ(T4, Tt); TD = LD(&(x[WS(ios, 26)]), dist, &(x[0])); TE = VZMULJ(TC, TD); Tw = LD(&(x[WS(ios, 18)]), dist, &(x[0])); Tx = VZMULJ(Tv, Tw); TA = LD(&(x[WS(ios, 10)]), dist, &(x[0])); TB = VZMULJ(Tz, TA); } { V Ty, TF, T3h, T3i; Ty = VSUB(Tu, Tx); TF = VSUB(TB, TE); TG = VFNMS(LDK(KP382683432), TF, VMUL(LDK(KP923879532), Ty)); T28 = VFMA(LDK(KP382683432), Ty, VMUL(LDK(KP923879532), TF)); T3h = VADD(Tu, Tx); T3i = VADD(TB, TE); T3j = VADD(T3h, T3i); T3y = VSUB(T3h, T3i); } } { V T1v, T1y, T1T, T1Q, T1I, T1K, T1L, T1B, T1F, T1G; { V T1u, T1x, T1S, T1P; T1u = LD(&(x[WS(ios, 31)]), dist, &(x[WS(ios, 1)])); T1v = VZMULJ(T1t, T1u); T1x = LD(&(x[WS(ios, 15)]), dist, &(x[WS(ios, 1)])); T1y = VZMULJ(T1w, T1x); T1S = LD(&(x[WS(ios, 23)]), dist, &(x[WS(ios, 1)])); T1T = VZMULJ(T1R, T1S); T1P = LD(&(x[WS(ios, 7)]), dist, &(x[WS(ios, 1)])); T1Q = VZMULJ(Tf, T1P); { V T1H, T1J, T1A, T1E; T1H = LD(&(x[WS(ios, 27)]), dist, &(x[WS(ios, 1)])); T1I = VZMULJ(T7, T1H); T1J = LD(&(x[WS(ios, 11)]), dist, &(x[WS(ios, 1)])); T1K = VZMULJ(T6, T1J); T1L = VSUB(T1I, T1K); T1A = LD(&(x[WS(ios, 3)]), dist, &(x[WS(ios, 1)])); T1B = VZMULJ(T3, T1A); T1E = LD(&(x[WS(ios, 19)]), dist, &(x[WS(ios, 1)])); T1F = VZMULJ(T1D, T1E); T1G = VSUB(T1B, T1F); } } { V T1z, T1M, T38, T39; T1z = VSUB(T1v, T1y); T1M = VMUL(LDK(KP707106781), VADD(T1G, T1L)); T1N = VADD(T1z, T1M); T2v = VSUB(T1z, T1M); T38 = VADD(T1B, T1F); T39 = VADD(T1I, T1K); T3a = VADD(T38, T39); T3G = VSUB(T39, T38); } { V T1O, T1U, T35, T36; T1O = VMUL(LDK(KP707106781), VSUB(T1L, T1G)); T1U = VSUB(T1Q, T1T); T1V = VSUB(T1O, T1U); T2w = VADD(T1U, T1O); T35 = VADD(T1v, T1y); T36 = VADD(T1Q, T1T); T37 = VADD(T35, T36); T3F = VSUB(T35, T36); } } { V T11, T14, T1p, T1m, T1e, T1g, T1h, T17, T1a, T1b; { V T10, T13, T1o, T1l; T10 = LD(&(x[WS(ios, 1)]), dist, &(x[WS(ios, 1)])); T11 = VZMULJ(T2, T10); T13 = LD(&(x[WS(ios, 17)]), dist, &(x[WS(ios, 1)])); T14 = VZMULJ(T12, T13); T1o = LD(&(x[WS(ios, 25)]), dist, &(x[WS(ios, 1)])); T1p = VZMULJ(T1n, T1o); T1l = LD(&(x[WS(ios, 9)]), dist, &(x[WS(ios, 1)])); T1m = VZMULJ(T5, T1l); { V T1d, T1f, T16, T19; T1d = LD(&(x[WS(ios, 29)]), dist, &(x[WS(ios, 1)])); T1e = VZMULJ(T1c, T1d); T1f = LD(&(x[WS(ios, 13)]), dist, &(x[WS(ios, 1)])); T1g = VZMULJ(TK, T1f); T1h = VSUB(T1e, T1g); T16 = LD(&(x[WS(ios, 5)]), dist, &(x[WS(ios, 1)])); T17 = VZMULJ(TS, T16); T19 = LD(&(x[WS(ios, 21)]), dist, &(x[WS(ios, 1)])); T1a = VZMULJ(T18, T19); T1b = VSUB(T17, T1a); } } { V T15, T1i, T31, T32; T15 = VSUB(T11, T14); T1i = VMUL(LDK(KP707106781), VADD(T1b, T1h)); T1j = VADD(T15, T1i); T2s = VSUB(T15, T1i); T31 = VADD(T17, T1a); T32 = VADD(T1e, T1g); T33 = VADD(T31, T32); T3D = VSUB(T32, T31); } { V T1k, T1q, T2Y, T2Z; T1k = VMUL(LDK(KP707106781), VSUB(T1h, T1b)); T1q = VSUB(T1m, T1p); T1r = VSUB(T1k, T1q); T2t = VADD(T1q, T1k); T2Y = VADD(T11, T14); T2Z = VADD(T1m, T1p); T30 = VADD(T2Y, T2Z); T3C = VSUB(T2Y, T2Z); } } { V T3r, T3v, T3u, T3w; { V T3p, T3q, T3s, T3t; T3p = VADD(T2T, T2W); T3q = VADD(T3j, T3g); T3r = VADD(T3p, T3q); T3v = VSUB(T3p, T3q); T3s = VADD(T30, T33); T3t = VADD(T37, T3a); T3u = VADD(T3s, T3t); T3w = VBYI(VSUB(T3t, T3s)); } ST(&(x[WS(ios, 16)]), VSUB(T3r, T3u), dist, &(x[0])); ST(&(x[WS(ios, 8)]), VADD(T3v, T3w), dist, &(x[0])); ST(&(x[0]), VADD(T3r, T3u), dist, &(x[0])); ST(&(x[WS(ios, 24)]), VSUB(T3v, T3w), dist, &(x[0])); } { V T2X, T3k, T3c, T3l, T34, T3b; T2X = VSUB(T2T, T2W); T3k = VSUB(T3g, T3j); T34 = VSUB(T30, T33); T3b = VSUB(T37, T3a); T3c = VMUL(LDK(KP707106781), VADD(T34, T3b)); T3l = VMUL(LDK(KP707106781), VSUB(T3b, T34)); { V T3d, T3m, T3n, T3o; T3d = VADD(T2X, T3c); T3m = VBYI(VADD(T3k, T3l)); ST(&(x[WS(ios, 28)]), VSUB(T3d, T3m), dist, &(x[0])); ST(&(x[WS(ios, 4)]), VADD(T3d, T3m), dist, &(x[0])); T3n = VSUB(T2X, T3c); T3o = VBYI(VSUB(T3l, T3k)); ST(&(x[WS(ios, 20)]), VSUB(T3n, T3o), dist, &(x[0])); ST(&(x[WS(ios, 12)]), VADD(T3n, T3o), dist, &(x[0])); } } { V T3B, T3W, T3M, T3U, T3I, T3T, T3P, T3X, T3A, T3L; T3A = VMUL(LDK(KP707106781), VADD(T3y, T3z)); T3B = VADD(T3x, T3A); T3W = VSUB(T3x, T3A); T3L = VMUL(LDK(KP707106781), VSUB(T3z, T3y)); T3M = VADD(T3K, T3L); T3U = VSUB(T3L, T3K); { V T3E, T3H, T3N, T3O; T3E = VFMA(LDK(KP923879532), T3C, VMUL(LDK(KP382683432), T3D)); T3H = VFNMS(LDK(KP382683432), T3G, VMUL(LDK(KP923879532), T3F)); T3I = VADD(T3E, T3H); T3T = VSUB(T3H, T3E); T3N = VFNMS(LDK(KP382683432), T3C, VMUL(LDK(KP923879532), T3D)); T3O = VFMA(LDK(KP382683432), T3F, VMUL(LDK(KP923879532), T3G)); T3P = VADD(T3N, T3O); T3X = VSUB(T3O, T3N); } { V T3J, T3Q, T3Z, T40; T3J = VADD(T3B, T3I); T3Q = VBYI(VADD(T3M, T3P)); ST(&(x[WS(ios, 30)]), VSUB(T3J, T3Q), dist, &(x[0])); ST(&(x[WS(ios, 2)]), VADD(T3J, T3Q), dist, &(x[0])); T3Z = VBYI(VADD(T3U, T3T)); T40 = VADD(T3W, T3X); ST(&(x[WS(ios, 6)]), VADD(T3Z, T40), dist, &(x[0])); ST(&(x[WS(ios, 26)]), VSUB(T40, T3Z), dist, &(x[0])); } { V T3R, T3S, T3V, T3Y; T3R = VSUB(T3B, T3I); T3S = VBYI(VSUB(T3P, T3M)); ST(&(x[WS(ios, 18)]), VSUB(T3R, T3S), dist, &(x[0])); ST(&(x[WS(ios, 14)]), VADD(T3R, T3S), dist, &(x[0])); T3V = VBYI(VSUB(T3T, T3U)); T3Y = VSUB(T3W, T3X); ST(&(x[WS(ios, 10)]), VADD(T3V, T3Y), dist, &(x[0])); ST(&(x[WS(ios, 22)]), VSUB(T3Y, T3V), dist, &(x[0])); } } { V TZ, T2k, T2d, T2l, T1X, T2h, T2a, T2i; { V Ts, TY, T2b, T2c; Ts = VADD(Tb, Tr); TY = VADD(TG, TX); TZ = VADD(Ts, TY); T2k = VSUB(Ts, TY); T2b = VFNMS(LDK(KP195090322), T1j, VMUL(LDK(KP980785280), T1r)); T2c = VFMA(LDK(KP195090322), T1N, VMUL(LDK(KP980785280), T1V)); T2d = VADD(T2b, T2c); T2l = VSUB(T2c, T2b); } { V T1s, T1W, T26, T29; T1s = VFMA(LDK(KP980785280), T1j, VMUL(LDK(KP195090322), T1r)); T1W = VFNMS(LDK(KP195090322), T1V, VMUL(LDK(KP980785280), T1N)); T1X = VADD(T1s, T1W); T2h = VSUB(T1W, T1s); T26 = VSUB(T1Z, T25); T29 = VSUB(T27, T28); T2a = VADD(T26, T29); T2i = VSUB(T29, T26); } { V T1Y, T2e, T2n, T2o; T1Y = VADD(TZ, T1X); T2e = VBYI(VADD(T2a, T2d)); ST(&(x[WS(ios, 31)]), VSUB(T1Y, T2e), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 1)]), VADD(T1Y, T2e), dist, &(x[WS(ios, 1)])); T2n = VBYI(VADD(T2i, T2h)); T2o = VADD(T2k, T2l); ST(&(x[WS(ios, 7)]), VADD(T2n, T2o), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 25)]), VSUB(T2o, T2n), dist, &(x[WS(ios, 1)])); } { V T2f, T2g, T2j, T2m; T2f = VSUB(TZ, T1X); T2g = VBYI(VSUB(T2d, T2a)); ST(&(x[WS(ios, 17)]), VSUB(T2f, T2g), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 15)]), VADD(T2f, T2g), dist, &(x[WS(ios, 1)])); T2j = VBYI(VSUB(T2h, T2i)); T2m = VSUB(T2k, T2l); ST(&(x[WS(ios, 9)]), VADD(T2j, T2m), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 23)]), VSUB(T2m, T2j), dist, &(x[WS(ios, 1)])); } } { V T2r, T2M, T2F, T2N, T2y, T2J, T2C, T2K; { V T2p, T2q, T2D, T2E; T2p = VSUB(Tb, Tr); T2q = VADD(T28, T27); T2r = VADD(T2p, T2q); T2M = VSUB(T2p, T2q); T2D = VFNMS(LDK(KP555570233), T2s, VMUL(LDK(KP831469612), T2t)); T2E = VFMA(LDK(KP555570233), T2v, VMUL(LDK(KP831469612), T2w)); T2F = VADD(T2D, T2E); T2N = VSUB(T2E, T2D); } { V T2u, T2x, T2A, T2B; T2u = VFMA(LDK(KP831469612), T2s, VMUL(LDK(KP555570233), T2t)); T2x = VFNMS(LDK(KP555570233), T2w, VMUL(LDK(KP831469612), T2v)); T2y = VADD(T2u, T2x); T2J = VSUB(T2x, T2u); T2A = VADD(T25, T1Z); T2B = VSUB(TX, TG); T2C = VADD(T2A, T2B); T2K = VSUB(T2B, T2A); } { V T2z, T2G, T2P, T2Q; T2z = VADD(T2r, T2y); T2G = VBYI(VADD(T2C, T2F)); ST(&(x[WS(ios, 29)]), VSUB(T2z, T2G), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 3)]), VADD(T2z, T2G), dist, &(x[WS(ios, 1)])); T2P = VBYI(VADD(T2K, T2J)); T2Q = VADD(T2M, T2N); ST(&(x[WS(ios, 5)]), VADD(T2P, T2Q), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 27)]), VSUB(T2Q, T2P), dist, &(x[WS(ios, 1)])); } { V T2H, T2I, T2L, T2O; T2H = VSUB(T2r, T2y); T2I = VBYI(VSUB(T2F, T2C)); ST(&(x[WS(ios, 19)]), VSUB(T2H, T2I), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 13)]), VADD(T2H, T2I), dist, &(x[WS(ios, 1)])); T2L = VBYI(VSUB(T2J, T2K)); T2O = VSUB(T2M, T2N); ST(&(x[WS(ios, 11)]), VADD(T2L, T2O), dist, &(x[WS(ios, 1)])); ST(&(x[WS(ios, 21)]), VSUB(T2O, T2L), dist, &(x[WS(ios, 1)])); } } } } return W;}static const tw_instr twinstr[] = { VTW(1), VTW(3), VTW(9), VTW(27), {TW_NEXT, VL, 0}};static const ct_desc desc = { 32, "t3fv_32", twinstr, &GENUS, {228, 142, 16, 0}, 0, 0, 0 };void X(codelet_t3fv_32) (planner *p) { X(kdft_dit_register) (p, t3fv_32, &desc);}#endif /* HAVE_FMA */
⌨️ 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 + -