📄 q1_6.c
字号:
T5R = FMA(KP866025403, T5Q, T5P); TY = TL * TX; TR = TL * TQ; T5T = W[8]; T5W = W[9]; iio[WS(vs, 2)] = FNMS(TS, TQ, TY); rio[WS(vs, 2)] = FMA(TS, TX, TR); T5Y = T5T * T5X; T5V = T5T * T5U; iio[WS(vs, 5) + WS(rs, 5)] = FNMS(T5W, T5U, T5Y); rio[WS(vs, 5) + WS(rs, 5)] = FMA(T5W, T5X, T5V); T65 = W[2]; T6c = W[3]; } } } } { E T5g, T5j, T5f, T5i; { E T1E, T1H, T3M, T3P, T56, T5d, T58, T5e, T57; { E T1s, T1B, T1f, T1u; { E T5K, T5S, T5J, T6i, T6b, T5v; T6i = T65 * T6h; T6b = T65 * T6a; T5v = W[0]; T5K = W[1]; iio[WS(vs, 2) + WS(rs, 5)] = FNMS(T6c, T6a, T6i); rio[WS(vs, 2) + WS(rs, 5)] = FMA(T6c, T6h, T6b); T5S = T5v * T5R; T5J = T5v * T5I; T1E = FNMS(KP866025403, T1r, T1k); T1s = FMA(KP866025403, T1r, T1k); T1B = FMA(KP866025403, T1A, T1z); T1H = FNMS(KP866025403, T1A, T1z); iio[WS(vs, 1) + WS(rs, 5)] = FNMS(T5K, T5I, T5S); rio[WS(vs, 1) + WS(rs, 5)] = FMA(T5K, T5R, T5J); T1f = W[0]; T1u = W[1]; } { E T3U, T3W, T3T, T1C, T1t, T3R; T1C = T1f * T1B; T1t = T1f * T1s; T3R = W[4]; T3U = W[5]; iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T1u, T1s, T1C); rio[WS(vs, 1) + WS(rs, 1)] = FMA(T1u, T1B, T1t); T3W = T3R * T3V; T3T = T3R * T3S; { E T3A, T3J, T3n, T3C, T3K, T3B, T51; T3M = FNMS(KP866025403, T3z, T3s); T3A = FMA(KP866025403, T3z, T3s); T3J = FMA(KP866025403, T3I, T3H); T3P = FNMS(KP866025403, T3I, T3H); iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T3U, T3S, T3W); rio[WS(vs, 3) + WS(rs, 3)] = FMA(T3U, T3V, T3T); T3n = W[0]; T3C = W[1]; T5g = FMA(KP866025403, T55, T52); T56 = FNMS(KP866025403, T55, T52); T5d = FNMS(KP866025403, T5c, T5b); T5j = FMA(KP866025403, T5c, T5b); T3K = T3n * T3J; T3B = T3n * T3A; T51 = W[2]; T58 = W[3]; iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T3C, T3A, T3K); rio[WS(vs, 1) + WS(rs, 3)] = FMA(T3C, T3J, T3B); T5e = T51 * T5d; T57 = T51 * T56; } } } { E T38, T3b, T3O, T3Q, T3N, T37, T3a; { E T2Y, T35, T2T, T30, T36, T2Z, T3L; T38 = FMA(KP866025403, T2X, T2U); T2Y = FNMS(KP866025403, T2X, T2U); T35 = FNMS(KP866025403, T34, T33); T3b = FMA(KP866025403, T34, T33); iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T58, T56, T5e); rio[WS(vs, 2) + WS(rs, 4)] = FMA(T58, T5d, T57); T2T = W[2]; T30 = W[3]; T36 = T2T * T35; T2Z = T2T * T2Y; T3L = W[8]; T3O = W[9]; iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T30, T2Y, T36); rio[WS(vs, 2) + WS(rs, 2)] = FMA(T30, T35, T2Z); T3Q = T3L * T3P; T3N = T3L * T3M; } iio[WS(vs, 5) + WS(rs, 3)] = FNMS(T3O, T3M, T3Q); rio[WS(vs, 5) + WS(rs, 3)] = FMA(T3O, T3P, T3N); T37 = W[6]; T3a = W[7]; { E T1G, T1I, T1F, T3c, T39, T1D; T3c = T37 * T3b; T39 = T37 * T38; T1D = W[8]; T1G = W[9]; iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T3a, T38, T3c); rio[WS(vs, 4) + WS(rs, 2)] = FMA(T3a, T3b, T39); T1I = T1D * T1H; T1F = T1D * T1E; iio[WS(vs, 5) + WS(rs, 1)] = FNMS(T1G, T1E, T1I); rio[WS(vs, 5) + WS(rs, 1)] = FMA(T1G, T1H, T1F); T5f = W[6]; T5i = W[7]; } } } { E T4Q, T4T, T2I, T2w, T2F, T2L, T2y, T2G, T2x, T4V, T4Y; { E T1M, T1O, T1L, T5k, T5h, T1J; T5k = T5f * T5j; T5h = T5f * T5g; T1J = W[4]; T1M = W[5]; iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T5i, T5g, T5k); rio[WS(vs, 4) + WS(rs, 4)] = FMA(T5i, T5j, T5h); T1O = T1J * T1N; T1L = T1J * T1K; iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T1M, T1K, T1O); rio[WS(vs, 3) + WS(rs, 1)] = FMA(T1M, T1N, T1L); T4V = W[4]; T4Y = W[5]; } { E T4E, T4N, T4G, T4O, T4F, T50, T4X, T4r; T4Q = FNMS(KP866025403, T4D, T4w); T4E = FMA(KP866025403, T4D, T4w); T4N = FMA(KP866025403, T4M, T4L); T4T = FNMS(KP866025403, T4M, T4L); T50 = T4V * T4Z; T4X = T4V * T4W; T4r = W[0]; T4G = W[1]; iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T4Y, T4W, T50); rio[WS(vs, 3) + WS(rs, 4)] = FMA(T4Y, T4Z, T4X); T4O = T4r * T4N; T4F = T4r * T4E; { E T2N, T2Q, T2S, T2P, T2j; iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T4G, T4E, T4O); rio[WS(vs, 1) + WS(rs, 4)] = FMA(T4G, T4N, T4F); T2N = W[4]; T2Q = W[5]; T2I = FNMS(KP866025403, T2v, T2o); T2w = FMA(KP866025403, T2v, T2o); T2F = FMA(KP866025403, T2E, T2D); T2L = FNMS(KP866025403, T2E, T2D); T2S = T2N * T2R; T2P = T2N * T2O; T2j = W[0]; T2y = W[1]; iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T2Q, T2O, T2S); rio[WS(vs, 3) + WS(rs, 2)] = FMA(T2Q, T2R, T2P); T2G = T2j * T2F; T2x = T2j * T2w; } } { E T1U, T21, T2H, T2K; { E T24, T27, T23, T26; T1U = FNMS(KP866025403, T1T, T1Q); T24 = FMA(KP866025403, T1T, T1Q); T27 = FMA(KP866025403, T20, T1Z); T21 = FNMS(KP866025403, T20, T1Z); iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T2y, T2w, T2G); rio[WS(vs, 1) + WS(rs, 2)] = FMA(T2y, T2F, T2x); T23 = W[6]; T26 = W[7]; { E T42, T49, T44, T4a, T43, T28, T25, T3X; T4c = FMA(KP866025403, T41, T3Y); T42 = FNMS(KP866025403, T41, T3Y); T49 = FNMS(KP866025403, T48, T47); T4f = FMA(KP866025403, T48, T47); T28 = T23 * T27; T25 = T23 * T24; T3X = W[2]; T44 = W[3]; iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T26, T24, T28); rio[WS(vs, 4) + WS(rs, 1)] = FMA(T26, T27, T25); T4a = T3X * T49; T43 = T3X * T42; iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T44, T42, T4a); rio[WS(vs, 2) + WS(rs, 3)] = FMA(T44, T49, T43); T2H = W[8]; T2K = W[9]; } } { E T4S, T4U, T4R, T2M, T2J, T4P; T2M = T2H * T2L; T2J = T2H * T2I; T4P = W[8]; T4S = W[9]; iio[WS(vs, 5) + WS(rs, 2)] = FNMS(T2K, T2I, T2M); rio[WS(vs, 5) + WS(rs, 2)] = FMA(T2K, T2L, T2J); T4U = T4P * T4T; T4R = T4P * T4Q; { E T1P, T1W, T22, T1V, T4b; iio[WS(vs, 5) + WS(rs, 4)] = FNMS(T4S, T4Q, T4U); rio[WS(vs, 5) + WS(rs, 4)] = FMA(T4S, T4T, T4R); T1P = W[2]; T1W = W[3]; T22 = T1P * T21; T1V = T1P * T1U; T4b = W[6]; T4e = W[7]; iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T1W, T1U, T22); rio[WS(vs, 2) + WS(rs, 1)] = FMA(T1W, T21, T1V); T4g = T4b * T4f; T4d = T4b * T4c; } } } } } } } iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T4e, T4c, T4g); rio[WS(vs, 4) + WS(rs, 3)] = FMA(T4e, T4f, T4d); }}static const tw_instr twinstr[] = { {TW_FULL, 0, 6}, {TW_NEXT, 1, 0}};static const ct_desc desc = { 6, "q1_6", twinstr, &GENUS, {144, 60, 132, 0}, 0, 0, 0 };void X(codelet_q1_6) (planner *p) { X(kdft_difsq_register) (p, q1_6, &desc);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_twidsq -compact -variables 4 -pipeline-latency 4 -reload-twiddle -dif -n 6 -name q1_6 -include q.h *//* * This function contains 276 FP additions, 168 FP multiplications, * (or, 192 additions, 84 multiplications, 84 fused multiply/add), * 85 stack variables, 2 constants, and 144 memory accesses */#include "q.h"static void q1_6(R *rio, R *iio, const R *W, stride rs, stride vs, INT mb, INT me, INT ms){ DK(KP500000000, +0.500000000000000000000000000000000000000000000); DK(KP866025403, +0.866025403784438646763723170752936183471402627); INT m; for (m = mb, W = W + (mb * 10); m < me; m = m + 1, rio = rio + ms, iio = iio + ms, W = W + 10, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(vs)) { E T3, Tc, Tt, TM, TX, T16, T1n, T1G, T2h, T2A, T1R, T20, T2L, T2U, T3b; E T3u, T3F, T3O, T45, T4o, T4Z, T5i, T4z, T4I, Ta, TP, Tf, Tq, Tn, TN; E Tu, TJ, T14, T1J, T19, T1k, T1h, T1H, T1o, T1D, T2b, T2B, T2i, T2x, T1Y; E T2D, T23, T2e, T2S, T3x, T2X, T38, T35, T3v, T3c, T3r, T3M, T4r, T3R, T42; E T3Z, T4p, T46, T4l, T4T, T5j, T50, T5f, T4G, T5l, T4L, T4W; { E T1, T2, T1l, T1m; T1 = rio[0]; T2 = rio[WS(rs, 3)]; T3 = T1 + T2; Tc = T1 - T2; { E Tr, Ts, TV, TW; Tr = iio[0]; Ts = iio[WS(rs, 3)]; Tt = Tr - Ts; TM = Tr + Ts; TV = rio[WS(vs, 1)]; TW = rio[WS(vs, 1) + WS(rs, 3)]; TX = TV + TW; T16 = TV - TW; } T1l = iio[WS(vs, 1)]; T1m = iio[WS(vs, 1) + WS(rs, 3)]; T1n = T1l - T1m; T1G = T1l + T1m; { E T2f, T2g, T1P, T1Q; T2f = iio[WS(vs, 2)]; T2g = iio[WS(vs, 2) + WS(rs, 3)]; T2h = T2f - T2g; T2A = T2f + T2g; T1P = rio[WS(vs, 2)]; T1Q = rio[WS(vs, 2) + WS(rs, 3)]; T1R = T1P + T1Q; T20 = T1P - T1Q; } } { E T2J, T2K, T43, T44; T2J = rio[WS(vs, 3)]; T2K = rio[WS(vs, 3) + WS(rs, 3)]; T2L = T2J + T2K; T2U = T2J - T2K; { E T39, T3a, T3D, T3E; T39 = iio[WS(vs, 3)]; T3a = iio[WS(vs, 3) + WS(rs, 3)]; T3b = T39 - T3a; T3u = T39 + T3a; T3D = rio[WS(vs, 4)]; T3E = rio[WS(vs, 4) + WS(rs, 3)]; T3F = T3D + T3E; T3O = T3D - T3E; } T43 = iio[WS(vs, 4)]; T44 = iio[WS(vs, 4) + WS(rs, 3)]; T45 = T43 - T44; T4o = T43 + T44; { E T4X, T4Y, T4x, T4y; T4X = iio[WS(vs, 5)]; T4Y = iio[WS(vs, 5) + WS(rs, 3)]; T4Z = T4X - T4Y; T5i = T4X + T4Y; T4x = rio[WS(vs, 5)]; T4y = rio[WS(vs, 5) + WS(rs, 3)]; T4z = T4x + T4y; T4I = T4x - T4y; } } { E T6, Td, T9, Te; { E T4, T5, T7, T8; T4 = rio[WS(rs, 2)]; T5 = rio[WS(rs, 5)]; T6 = T4 + T5; Td = T4 - T5; T7 = rio[WS(rs, 4)]; T8 = rio[WS(rs, 1)]; T9 = T7 + T8; Te = T7 - T8; } Ta = T6 + T9; TP = KP866025403 * (T9 - T6); Tf = Td + Te; Tq = KP866025403 * (Te - Td); } { E Tj, TH, Tm, TI; { E Th, Ti, Tk, Tl; Th = iio[WS(rs, 2)]; Ti = iio[WS(rs, 5)]; Tj = Th - Ti; TH = Th + Ti; Tk = iio[WS(rs, 4)]; Tl = iio[WS(rs, 1)]; Tm = Tk - Tl; TI = Tk + Tl; } Tn = KP866025403 * (Tj - Tm); TN = TH + TI; Tu = Tj + Tm; TJ = KP866025403 * (TH - TI); } { E T10, T17, T13, T18; { E TY, TZ, T11, T12; TY = rio[WS(vs, 1) + WS(rs, 2)]; TZ = rio[WS(vs, 1) + WS(rs, 5)]; T10 = TY + TZ; T17 = TY - TZ; T11 = rio[WS(vs, 1) + WS(rs, 4)]; T12 = rio[WS(vs, 1) + WS(rs, 1)]; T13 = T11 + T12; T18 = T11 - T12; } T14 = T10 + T13; T1J = KP866025403 * (T13 - T10); T19 = T17 + T18; T1k = KP866025403 * (T18 - T17); } { E T1d, T1B, T1g, T1C; { E T1b, T1c, T1e, T1f; T1b = iio[WS(vs, 1) + WS(rs, 2)]; T1c = iio[WS(vs, 1) + WS(rs, 5)]; T1d = T1b - T1c; T1B = T1b + T1c; T1e = iio[WS(vs, 1) + WS(rs, 4)]; T1f = iio[WS(vs, 1) + WS(rs, 1)]; T1g = T1e - T1f; T1C = T1e + T1f; } T1h = KP866025403 * (T1d - T1g); T1H = T1B + T1C; T1o = T1d + T1g; T1D = KP866025403 * (T1B - T1C); } { E T27, T2v, T2a, T2w; { E T25, T26, T28, T29; T25 = iio[WS(vs, 2) + WS(rs, 2)]; T26 = iio[WS(vs, 2) + WS(rs, 5)]; T27 = T25 - T26; T2v = T25 + T26; T28 = iio[WS(vs, 2) + WS(rs, 4)]; T29 = iio[WS(vs, 2) + WS(rs, 1)]; T2a = T28 - T29; T2w = T28 + T29; } T2b = KP866025403 * (T27 - T2a); T2B = T2v + T2w; T2i = T27 + T2a; T2x = KP866025403 * (T2v - T2w); } { E T1U, T21, T1X, T22; { E T1S, T1T, T1V, T1W; T1S = rio[WS(vs, 2) + WS(rs, 2)]; T1T = rio[WS(vs, 2) + WS(rs, 5)]; T1U = T1S + T1T; T21 = T1S - T1T; T1V = rio[WS(vs, 2) + WS(rs, 4)]; T1W = rio[WS(vs, 2) + WS(rs, 1)]; T1X = T1V + T1W;
⌨️ 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 + -