📄 r2cf_64.c
字号:
T2B = FNMS(KP707106781, T1I, T1H); T1J = FMA(KP707106781, T1I, T1H); T27 = FMA(KP707106781, T26, T1R); T2D = FNMS(KP707106781, T26, T1R); T2L = FNMS(KP923879532, T2I, T2H); T2J = FMA(KP923879532, T2I, T2H); T2M = FMA(KP668178637, T2A, T2B); T2C = FNMS(KP668178637, T2B, T2A); T2E = FNMS(KP707106781, T2f, T2e); T2g = FMA(KP707106781, T2f, T2e); } T2N = FNMS(KP668178637, T2D, T2E); T2F = FMA(KP668178637, T2E, T2D); T2v = FNMS(KP923879532, T1c, T15); T1d = FMA(KP923879532, T1c, T15); { E T2Q, T2O, T2K, T2G; T2Q = T2M - T2N; T2O = T2M + T2N; T2K = T2F - T2C; T2G = T2C + T2F; Cr[WS(csr, 10)] = FMA(KP831469612, T2Q, T2P); Cr[WS(csr, 22)] = FNMS(KP831469612, T2Q, T2P); Ci[WS(csi, 26)] = FNMS(KP831469612, T2O, T2L); Ci[WS(csi, 6)] = -(FMA(KP831469612, T2O, T2L)); Ci[WS(csi, 22)] = FMS(KP831469612, T2K, T2J); Ci[WS(csi, 10)] = FMA(KP831469612, T2K, T2J); Cr[WS(csr, 6)] = FMA(KP831469612, T2G, T2z); Cr[WS(csr, 26)] = FNMS(KP831469612, T2G, T2z); } T2r = FMA(KP923879532, T2o, T2l); T2p = FNMS(KP923879532, T2o, T2l); T2s = FNMS(KP198912367, T1A, T1J); T1K = FMA(KP198912367, T1J, T1A); } { E T63, T5r, T5R, T6d, T5J, T5M, T6e, T5y, T6j, T6b, T66, T67, T64, T5U, T5Z; E T5G; { E T5S, T5u, T5x, T5T, T2t, T2h; T63 = FMA(KP923879532, T5q, T5p); T5r = FNMS(KP923879532, T5q, T5p); T5R = FNMS(KP923879532, T5Q, T5P); T6d = FMA(KP923879532, T5Q, T5P); T2t = FMA(KP198912367, T27, T2g); T2h = FNMS(KP198912367, T2g, T27); T5S = FNMS(KP668178637, T5s, T5t); T5u = FMA(KP668178637, T5t, T5s); { E T2w, T2u, T2q, T2i; T2w = T2t - T2s; T2u = T2s + T2t; T2q = T2h - T1K; T2i = T1K + T2h; Cr[WS(csr, 14)] = FMA(KP980785280, T2w, T2v); Cr[WS(csr, 18)] = FNMS(KP980785280, T2w, T2v); Ci[WS(csi, 30)] = FMS(KP980785280, T2u, T2r); Ci[WS(csi, 2)] = FMA(KP980785280, T2u, T2r); Ci[WS(csi, 18)] = FMA(KP980785280, T2q, T2p); Ci[WS(csi, 14)] = FMS(KP980785280, T2q, T2p); Cr[WS(csr, 2)] = FMA(KP980785280, T2i, T1d); Cr[WS(csr, 30)] = FNMS(KP980785280, T2i, T1d); T5x = FNMS(KP668178637, T5w, T5v); T5T = FMA(KP668178637, T5v, T5w); } { E T69, T6a, T5C, T5F; T5J = FNMS(KP923879532, T5I, T5H); T69 = FMA(KP923879532, T5I, T5H); T6a = FNMS(KP923879532, T5L, T5K); T5M = FMA(KP923879532, T5L, T5K); T6e = T5x + T5u; T5y = T5u - T5x; T6j = FNMS(KP303346683, T69, T6a); T6b = FMA(KP303346683, T6a, T69); T66 = FMA(KP923879532, T5B, T5A); T5C = FNMS(KP923879532, T5B, T5A); T5F = FNMS(KP923879532, T5E, T5D); T67 = FMA(KP923879532, T5E, T5D); T64 = T5T + T5S; T5U = T5S - T5T; T5Z = FMA(KP534511135, T5C, T5F); T5G = FNMS(KP534511135, T5F, T5C); } } { E T61, T6i, T68, T62; { E T5z, T5Y, T5N, T5X, T5V, T60, T5W, T5O; T61 = FNMS(KP831469612, T5y, T5r); T5z = FMA(KP831469612, T5y, T5r); T6i = FNMS(KP303346683, T66, T67); T68 = FMA(KP303346683, T67, T66); T5Y = FMA(KP534511135, T5J, T5M); T5N = FNMS(KP534511135, T5M, T5J); T5X = FNMS(KP831469612, T5U, T5R); T5V = FMA(KP831469612, T5U, T5R); T60 = T5Y - T5Z; T62 = T5Z + T5Y; T5W = T5N - T5G; T5O = T5G + T5N; Ci[WS(csi, 27)] = FMA(KP881921264, T60, T5X); Ci[WS(csi, 5)] = FMS(KP881921264, T60, T5X); Cr[WS(csr, 5)] = FMA(KP881921264, T5O, T5z); Cr[WS(csr, 27)] = FNMS(KP881921264, T5O, T5z); Ci[WS(csi, 21)] = FMS(KP881921264, T5W, T5V); Ci[WS(csi, 11)] = FMA(KP881921264, T5W, T5V); } { E T6g, T6f, T6h, T6k, T65, T6c; T6l = FNMS(KP831469612, T64, T63); T65 = FMA(KP831469612, T64, T63); T6c = T68 + T6b; T6g = T6b - T68; T6f = FNMS(KP831469612, T6e, T6d); T6h = FMA(KP831469612, T6e, T6d); Cr[WS(csr, 11)] = FMA(KP881921264, T62, T61); Cr[WS(csr, 21)] = FNMS(KP881921264, T62, T61); Cr[WS(csr, 3)] = FMA(KP956940335, T6c, T65); Cr[WS(csr, 29)] = FNMS(KP956940335, T6c, T65); T6k = T6i - T6j; T6m = T6i + T6j; Ci[WS(csi, 29)] = FMS(KP956940335, T6k, T6h); Ci[WS(csi, 3)] = FMA(KP956940335, T6k, T6h); Ci[WS(csi, 19)] = FMA(KP956940335, T6g, T6f); Ci[WS(csi, 13)] = FMS(KP956940335, T6g, T6f); } } } { E T55, T3L, T4T, T5f, T4D, T4K, T5g, T44, T5l, T5d, T58, T59, T56, T4W, T51; E T4q; { E T4U, T3U, T43, T4V; T55 = FNMS(KP923879532, T3K, T3D); T3L = FMA(KP923879532, T3K, T3D); T4T = FMA(KP923879532, T4S, T4P); T5f = FNMS(KP923879532, T4S, T4P); Cr[WS(csr, 13)] = FNMS(KP956940335, T6m, T6l); Cr[WS(csr, 19)] = FMA(KP956940335, T6m, T6l); T4U = FMA(KP198912367, T3Q, T3T); T3U = FNMS(KP198912367, T3T, T3Q); T43 = FMA(KP198912367, T42, T3Z); T4V = FNMS(KP198912367, T3Z, T42); { E T5b, T5c, T4i, T4p; T4D = FMA(KP923879532, T4C, T4v); T5b = FNMS(KP923879532, T4C, T4v); T5c = FNMS(KP923879532, T4J, T4G); T4K = FMA(KP923879532, T4J, T4G); T5g = T43 - T3U; T44 = T3U + T43; T5l = FNMS(KP820678790, T5b, T5c); T5d = FMA(KP820678790, T5c, T5b); T58 = FNMS(KP923879532, T4h, T4a); T4i = FMA(KP923879532, T4h, T4a); T4p = FMA(KP923879532, T4o, T4l); T59 = FNMS(KP923879532, T4o, T4l); T56 = T4U - T4V; T4W = T4U + T4V; T51 = FMA(KP098491403, T4i, T4p); T4q = FNMS(KP098491403, T4p, T4i); } } { E T53, T5k, T5a, T54; { E T45, T50, T4L, T4Z, T4X, T52, T4Y, T4M; T53 = FNMS(KP980785280, T44, T3L); T45 = FMA(KP980785280, T44, T3L); T5k = FNMS(KP820678790, T58, T59); T5a = FMA(KP820678790, T59, T58); T50 = FMA(KP098491403, T4D, T4K); T4L = FNMS(KP098491403, T4K, T4D); T4Z = FMA(KP980785280, T4W, T4T); T4X = FNMS(KP980785280, T4W, T4T); T52 = T50 - T51; T54 = T51 + T50; T4Y = T4L - T4q; T4M = T4q + T4L; Ci[WS(csi, 31)] = FMA(KP995184726, T52, T4Z); Ci[WS(csi, 1)] = FMS(KP995184726, T52, T4Z); Cr[WS(csr, 1)] = FMA(KP995184726, T4M, T45); Cr[WS(csr, 31)] = FNMS(KP995184726, T4M, T45); Ci[WS(csi, 17)] = FMS(KP995184726, T4Y, T4X); Ci[WS(csi, 15)] = FMA(KP995184726, T4Y, T4X); } { E T5i, T5h, T5j, T5m, T57, T5e; T5n = FNMS(KP980785280, T56, T55); T57 = FMA(KP980785280, T56, T55); T5e = T5a + T5d; T5i = T5d - T5a; T5h = FNMS(KP980785280, T5g, T5f); T5j = FMA(KP980785280, T5g, T5f); Cr[WS(csr, 15)] = FMA(KP995184726, T54, T53); Cr[WS(csr, 17)] = FNMS(KP995184726, T54, T53); Cr[WS(csr, 7)] = FMA(KP773010453, T5e, T57); Cr[WS(csr, 25)] = FNMS(KP773010453, T5e, T57); T5m = T5k - T5l; T5o = T5k + T5l; Ci[WS(csi, 25)] = FMS(KP773010453, T5m, T5j); Ci[WS(csi, 7)] = FMA(KP773010453, T5m, T5j); Ci[WS(csi, 23)] = FMA(KP773010453, T5i, T5h); Ci[WS(csi, 9)] = FMS(KP773010453, T5i, T5h); } } } } } } Cr[WS(csr, 9)] = FNMS(KP773010453, T5o, T5n); Cr[WS(csr, 23)] = FMA(KP773010453, T5o, T5n); }}static const kr2c_desc desc = { 64, "r2cf_64", {198, 0, 196, 0}, &GENUS };void X(codelet_r2cf_64) (planner *p) { X(kr2c_register) (p, r2cf_64, &desc);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_r2cf -compact -variables 4 -pipeline-latency 4 -n 64 -name r2cf_64 -include r2cf.h *//* * This function contains 394 FP additions, 124 FP multiplications, * (or, 342 additions, 72 multiplications, 52 fused multiply/add), * 106 stack variables, 15 constants, and 128 memory accesses */#include "r2cf.h"static void r2cf_64(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs){ DK(KP773010453, +0.773010453362736960810906609758469800971041293); DK(KP634393284, +0.634393284163645498215171613225493370675687095); DK(KP098017140, +0.098017140329560601994195563888641845861136673); DK(KP995184726, +0.995184726672196886244836953109479921575474869); DK(KP290284677, +0.290284677254462367636192375817395274691476278); DK(KP956940335, +0.956940335732208864935797886980269969482849206); DK(KP471396736, +0.471396736825997648556387625905254377657460319); DK(KP881921264, +0.881921264348355029712756863660388349508442621); DK(KP195090322, +0.195090322016128267848284868477022240927691618); DK(KP980785280, +0.980785280403230449126182236134239036973933731); DK(KP555570233, +0.555570233019602224742830813948532874374937191); DK(KP831469612, +0.831469612302545237078788377617905756738560812); DK(KP382683432, +0.382683432365089771728459984030398866761344562); DK(KP923879532, +0.923879532511286756128183189396788286822416626); DK(KP707106781, +0.707106781186547524400844362104849039284835938); INT i; for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) { E T4l, T5a, T15, T3n, T2T, T3Q, T7, Te, Tf, T4A, T4L, T1X, T3B, T23, T3y; E T5I, T66, T4R, T52, T2j, T3F, T2H, T3I, T5P, T69, T1i, T3t, T1l, T3u, TZ; E T63, T4v, T58, T1r, T3r, T1u, T3q, TK, T62, T4s, T57, Tm, Tt, Tu, T4o; E T5b, T1c, T3R, T2Q, T3o, T1M, T3z, T5L, T67, T26, T3C, T4H, T4M, T2y, T3J; E T5S, T6a, T2C, T3G, T4Y, T53; { E T3, T11, Td, T13, T6, T2S, Ta, T12, T14, T2R; { E T1, T2, Tb, Tc; T1 = R0[0]; T2 = R0[WS(rs, 16)]; T3 = T1 + T2; T11 = T1 - T2; Tb = R0[WS(rs, 28)]; Tc = R0[WS(rs, 12)]; Td = Tb + Tc; T13 = Tb - Tc; } { E T4, T5, T8, T9; T4 = R0[WS(rs, 8)]; T5 = R0[WS(rs, 24)]; T6 = T4 + T5; T2S = T4 - T5; T8 = R0[WS(rs, 4)]; T9 = R0[WS(rs, 20)]; Ta = T8 + T9; T12 = T8 - T9; } T4l = T3 - T6; T5a = Td - Ta; T14 = KP707106781 * (T12 + T13); T15 = T11 + T14; T3n = T11 - T14; T2R = KP707106781 * (T13 - T12); T2T = T2R - T2S; T3Q = T2S + T2R; T7 = T3 + T6; Te = Ta + Td; Tf = T7 + Te; } { E T1P, T4J, T21, T4y, T1S, T4K, T1W, T4z; { E T1N, T1O, T1Z, T20; T1N = R1[WS(rs, 28)]; T1O = R1[WS(rs, 12)]; T1P = T1N - T1O; T4J = T1N + T1O; T1Z = R1[0]; T20 = R1[WS(rs, 16)]; T21 = T1Z - T20; T4y = T1Z + T20; } { E T1Q, T1R, T1U, T1V; T1Q = R1[WS(rs, 4)]; T1R = R1[WS(rs, 20)]; T1S = T1Q - T1R; T4K = T1Q + T1R; T1U = R1[WS(rs, 8)]; T1V = R1[WS(rs, 24)]; T1W = T1U - T1V; T4z = T1U + T1V; } T4A = T4y - T4z; T4L = T4J - T4K; { E T1T, T22, T5G, T5H; T1T = KP707106781 * (T1P - T1S); T1X = T1T - T1W; T3B = T1W + T1T; T22 = KP707106781 * (T1S + T1P); T23 = T21 + T22; T3y = T21 - T22; T5G = T4y + T4z; T5H = T4K + T4J; T5I = T5G + T5H; T66 = T5G - T5H; } } { E T2b, T4P, T2G, T4Q, T2e, T51, T2h, T50; { E T29, T2a, T2E, T2F; T29 = R1[WS(rs, 31)]; T2a = R1[WS(rs, 15)]; T2b = T29 - T2a; T4P = T29 + T2a; T2E = R1[WS(rs, 7)]; T2F = R1[WS(rs, 23)]; T2G = T2E - T2F; T4Q = T2E + T2F; } { E T2c, T2d, T2f, T2g; T2c = R1[WS(rs, 3)]; T2d = R1[WS(rs, 19)]; T2e = T2c - T2d; T51 = T2c + T2d; T2f = R1[WS(rs, 27)]; T2g = R1[WS(rs, 11)]; T2h = T2f - T2g; T50 = T2f + T2g; } T4R = T4P - T4Q; T52 = T50 - T51; { E T2i, T2D, T5N, T5O; T2i = KP707106781 * (T2e + T2h); T2j = T2b + T2i; T3F = T2b - T2i; T2D = KP707106781 * (T2h - T2e); T2H = T2D - T2G; T3I = T2G + T2D; T5N = T4P + T4Q; T5O = T51 + T50; T5P = T5N + T5O; T69 = T5N - T5O; } } { E TN, T1e, TX, T1g, TQ, T1k, TU, T1f, T1h, T1j; { E TL, TM, TV, TW; TL = R0[WS(rs, 31)]; TM = R0[WS(rs, 15)]; TN = TL + TM; T1e = TL - TM; TV = R0[WS(rs, 27)]; TW = R0[WS(rs, 11)]; TX = TV + TW; T1g = TV - TW; } { E TO, TP, TS, TT; TO = R0[WS(rs, 7)]; TP = R0[WS(rs, 23)]; TQ = TO + TP; T1k = TO - TP; TS = R0[WS(rs, 3)]; TT = R0[WS(rs, 19)]; TU = TS + TT; T1f = TS - TT; } T1h = KP707106781 * (T1f + T1g); T1i = T1e + T1h; T3t = T1e - T1h; T1j = KP707106781 * (T1g - T1f); T1l = T1j - T1k; T3u = T1k + T1j; { E TR, TY, T4t, T4u; TR = TN + TQ; TY = TU + TX; TZ = TR + TY; T63 = TR - TY; T4t = TN - TQ; T4u = TX - TU; T4v = FNMS(KP382683432, T4u, KP923879532 * T4t); T58 = FMA(KP382683432, T4t, KP923879532 * T4u); } } { E Ty, T1s, TI, T1n, TB, T1q, TF, T1o, T1p, T1t; { E Tw, Tx, TG, TH; Tw = R0[WS(rs, 1)]; Tx = R0[WS(rs, 17)]; Ty = Tw + Tx; T1s = Tw - Tx; TG = R0[WS(rs, 29)]; TH = R0[WS(rs, 13)]; TI = TG + TH; T1n = TG - TH; } { E Tz, TA, TD, TE; Tz = R0[WS(rs, 9)]; TA = R0[WS(rs, 25)]; TB = Tz + TA; T1q = Tz - TA; TD = R0[WS(rs, 5)]; TE = R0[WS(rs, 21)]; TF = TD + TE; T1o = TD - TE; } T1p = KP707106781 * (T1n - T1o); T1r = T1p - T1q; T3r = T1q + T1p; T1t = KP707106781 * (T1o + T1n); T1u = T1s + T1t; T3q = T1s - T1t; { E TC, TJ, T4q, T4r; TC = Ty + TB; TJ = TF + TI; TK = TC + TJ; T62 = TC - TJ; T4q = Ty - TB; T4r = TI - TF; T4s = FMA(KP923879532, T4q, KP382683432 * T4r); T57 = FNMS(KP382683432, T4q, KP923879532 * T4r); } } {
⌨️ 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 + -