📄 hc2cfdftv_32.c
字号:
T29 = VZMULIJ(T28, VSUB(T1a, T18)); T1d = LD(&(Rp[WS(rs, 11)]), ms, &(Rp[WS(rs, 1)])); T1e = LD(&(Rm[WS(rs, 11)]), -ms, &(Rm[WS(rs, 1)])); T1f = VCONJ(T1e); T1c = LDW(&(W[TWVL * 42])); T1g = VZMULJ(T1c, VADD(T1d, T1f)); T2a = LDW(&(W[TWVL * 44])); T2b = VZMULIJ(T2a, VSUB(T1f, T1d)); TF = LD(&(Rp[WS(rs, 9)]), ms, &(Rp[WS(rs, 1)])); TG = LD(&(Rm[WS(rs, 9)]), -ms, &(Rm[WS(rs, 1)])); TH = VCONJ(TG); TE = LDW(&(W[TWVL * 34])); TI = VZMULJ(TE, VADD(TF, TH)); T1S = LDW(&(W[TWVL * 36])); T1T = VZMULIJ(T1S, VSUB(TH, TF)); TL = LD(&(Rp[WS(rs, 5)]), ms, &(Rp[WS(rs, 1)])); TM = LD(&(Rm[WS(rs, 5)]), -ms, &(Rm[WS(rs, 1)])); TN = VCONJ(TM); TK = LDW(&(W[TWVL * 18])); TO = VZMULJ(TK, VADD(TL, TN)); T1X = LDW(&(W[TWVL * 20])); T1Y = VZMULIJ(T1X, VSUB(TN, TL)); TX = LD(&(Rp[WS(rs, 15)]), ms, &(Rp[WS(rs, 1)])); TY = LD(&(Rm[WS(rs, 15)]), -ms, &(Rm[WS(rs, 1)])); TZ = VCONJ(TY); TW = LDW(&(W[TWVL * 58])); T10 = VZMULJ(TW, VADD(TX, TZ)); T21 = LDW(&(W[TWVL * 60])); T22 = VZMULIJ(T21, VSUB(TZ, TX)); T12 = LD(&(Rp[WS(rs, 7)]), ms, &(Rp[WS(rs, 1)])); T13 = LD(&(Rm[WS(rs, 7)]), -ms, &(Rm[WS(rs, 1)])); T14 = VCONJ(T13); T11 = LDW(&(W[TWVL * 26])); T15 = VZMULJ(T11, VADD(T12, T14)); T23 = LDW(&(W[TWVL * 28])); T24 = VZMULIJ(T23, VSUB(T14, T12)); TQ = LD(&(Rp[WS(rs, 13)]), ms, &(Rp[WS(rs, 1)])); TR = LD(&(Rm[WS(rs, 13)]), -ms, &(Rm[WS(rs, 1)])); TS = VCONJ(TR); TP = LDW(&(W[TWVL * 50])); TT = VZMULJ(TP, VADD(TQ, TS)); T1V = LDW(&(W[TWVL * 52])); T1W = VZMULIJ(T1V, VSUB(TS, TQ)); } TJ = VSUB(TD, TI); TU = VSUB(TO, TT); TV = VFNMS(LDK(KP382683432), TU, VMUL(LDK(KP923879532), TJ)); T2p = VFMA(LDK(KP382683432), TJ, VMUL(LDK(KP923879532), TU)); T16 = VSUB(T10, T15); T1h = VSUB(T1b, T1g); T1i = VFMA(LDK(KP923879532), T16, VMUL(LDK(KP382683432), T1h)); T2o = VFNMS(LDK(KP923879532), T1h, VMUL(LDK(KP382683432), T16)); T3J = VADD(T1Y, T1W); T3K = VADD(T1R, T1T); T3L = VSUB(T3J, T3K); T4q = VADD(T3K, T3J); T3G = VADD(T22, T24); T3H = VADD(T29, T2b); T3I = VSUB(T3G, T3H); T4r = VADD(T3G, T3H); T3u = VADD(T10, T15); T3v = VADD(T1b, T1g); T3w = VSUB(T3u, T3v); T4k = VADD(T3u, T3v); T3r = VADD(TD, TI); T3s = VADD(TO, TT); T3t = VSUB(T3r, T3s); T4j = VADD(T3r, T3s); T25 = VSUB(T22, T24); T2c = VSUB(T29, T2b); T1U = VSUB(T1R, T1T); T1Z = VSUB(T1W, T1Y); T20 = VMUL(LDK(KP707106781), VADD(T1U, T1Z)); T27 = VMUL(LDK(KP707106781), VSUB(T1Z, T1U)); T26 = VADD(T20, T25); T2V = VADD(T27, T2c); T2d = VSUB(T27, T2c); T2U = VSUB(T25, T20); } { V T4m, T4w, T4t, T4x, T4i, T4l, T4p, T4s, T4u, T4z, T4v, T4y, T4E, T4L, T4H; V T4K, T4A, T4F, T4D, T4G, T4B, T4C, T4I, T4N, T4J, T4M, T3O, T4c, T4d, T3X; V T40, T46, T49, T41, T3y, T47, T3T, T45, T3N, T44, T3W, T48, T3x, T3S, T3F; V T3M, T3U, T3V, T3Y, T4e, T4f, T3Z, T42, T4a, T4b, T43; T4i = VADD(T4g, T4h); T4l = VADD(T4j, T4k); T4m = VADD(T4i, T4l); T4w = VSUB(T4i, T4l); T4p = VADD(T4n, T4o); T4s = VADD(T4q, T4r); T4t = VADD(T4p, T4s); T4x = VBYI(VSUB(T4s, T4p)); T4u = VCONJ(VMUL(LDK(KP500000000), VSUB(T4m, T4t))); ST(&(Rm[WS(rs, 15)]), T4u, -ms, &(Rm[WS(rs, 1)])); T4z = VMUL(LDK(KP500000000), VADD(T4w, T4x)); ST(&(Rp[WS(rs, 8)]), T4z, ms, &(Rp[0])); T4v = VMUL(LDK(KP500000000), VADD(T4m, T4t)); ST(&(Rp[0]), T4v, ms, &(Rp[0])); T4y = VCONJ(VMUL(LDK(KP500000000), VSUB(T4w, T4x))); ST(&(Rm[WS(rs, 7)]), T4y, -ms, &(Rm[WS(rs, 1)])); T4A = VMUL(LDK(KP500000000), VSUB(T4g, T4h)); T4F = VSUB(T4k, T4j); T4B = VSUB(T4n, T4o); T4C = VSUB(T4r, T4q); T4D = VMUL(LDK(KP353553390), VADD(T4B, T4C)); T4G = VMUL(LDK(KP707106781), VSUB(T4C, T4B)); T4E = VADD(T4A, T4D); T4L = VMUL(LDK(KP500000000), VBYI(VSUB(T4G, T4F))); T4H = VMUL(LDK(KP500000000), VBYI(VADD(T4F, T4G))); T4K = VSUB(T4A, T4D); T4I = VCONJ(VSUB(T4E, T4H)); ST(&(Rm[WS(rs, 3)]), T4I, -ms, &(Rm[WS(rs, 1)])); T4N = VADD(T4K, T4L); ST(&(Rp[WS(rs, 12)]), T4N, ms, &(Rp[0])); T4J = VADD(T4E, T4H); ST(&(Rp[WS(rs, 4)]), T4J, ms, &(Rp[0])); T4M = VCONJ(VSUB(T4K, T4L)); ST(&(Rm[WS(rs, 11)]), T4M, -ms, &(Rm[WS(rs, 1)])); T3x = VMUL(LDK(KP353553390), VADD(T3t, T3w)); T3y = VADD(T3q, T3x); T47 = VSUB(T3q, T3x); T3S = VMUL(LDK(KP707106781), VSUB(T3w, T3t)); T3T = VADD(T3R, T3S); T45 = VSUB(T3S, T3R); T3F = VFMA(LDK(KP923879532), T3B, VMUL(LDK(KP382683432), T3E)); T3M = VFNMS(LDK(KP382683432), T3L, VMUL(LDK(KP923879532), T3I)); T3N = VMUL(LDK(KP500000000), VADD(T3F, T3M)); T44 = VSUB(T3M, T3F); T3U = VFNMS(LDK(KP382683432), T3B, VMUL(LDK(KP923879532), T3E)); T3V = VFMA(LDK(KP923879532), T3L, VMUL(LDK(KP382683432), T3I)); T3W = VADD(T3U, T3V); T48 = VMUL(LDK(KP500000000), VSUB(T3V, T3U)); T3O = VADD(T3y, T3N); T4c = VMUL(LDK(KP500000000), VBYI(VADD(T45, T44))); T4d = VADD(T47, T48); T3X = VMUL(LDK(KP500000000), VBYI(VADD(T3T, T3W))); T40 = VSUB(T3y, T3N); T46 = VMUL(LDK(KP500000000), VBYI(VSUB(T44, T45))); T49 = VSUB(T47, T48); T41 = VMUL(LDK(KP500000000), VBYI(VSUB(T3W, T3T))); T3Y = VCONJ(VSUB(T3O, T3X)); ST(&(Rm[WS(rs, 1)]), T3Y, -ms, &(Rm[WS(rs, 1)])); T4e = VADD(T4c, T4d); ST(&(Rp[WS(rs, 6)]), T4e, ms, &(Rp[0])); T4f = VCONJ(VSUB(T4d, T4c)); ST(&(Rm[WS(rs, 5)]), T4f, -ms, &(Rm[WS(rs, 1)])); T3Z = VADD(T3O, T3X); ST(&(Rp[WS(rs, 2)]), T3Z, ms, &(Rp[0])); T42 = VCONJ(VSUB(T40, T41)); ST(&(Rm[WS(rs, 13)]), T42, -ms, &(Rm[WS(rs, 1)])); T4a = VADD(T46, T49); ST(&(Rp[WS(rs, 10)]), T4a, ms, &(Rp[0])); T4b = VCONJ(VSUB(T49, T46)); ST(&(Rm[WS(rs, 9)]), T4b, -ms, &(Rm[WS(rs, 1)])); T43 = VADD(T40, T41); ST(&(Rp[WS(rs, 14)]), T43, ms, &(Rp[0])); { V T2g, T2K, T2L, T2v, T2y, T2E, T2H, T2z, T1k, T2F, T2u, T2G, T2f, T2C, T2r; V T2D, Ty, T1j, T2s, T2t, T1P, T2e, T2n, T2q, T2w, T2M, T2N, T2x, T2A, T2I; V T2J, T2B; Ty = VADD(Ta, Tx); T1j = VMUL(LDK(KP500000000), VADD(TV, T1i)); T1k = VADD(Ty, T1j); T2F = VSUB(Ty, T1j); T2s = VFNMS(LDK(KP195090322), T1B, VMUL(LDK(KP980785280), T1O)); T2t = VFMA(LDK(KP195090322), T26, VMUL(LDK(KP980785280), T2d)); T2u = VADD(T2s, T2t); T2G = VMUL(LDK(KP500000000), VSUB(T2t, T2s)); T1P = VFMA(LDK(KP980785280), T1B, VMUL(LDK(KP195090322), T1O)); T2e = VFNMS(LDK(KP195090322), T2d, VMUL(LDK(KP980785280), T26)); T2f = VMUL(LDK(KP500000000), VADD(T1P, T2e)); T2C = VSUB(T2e, T1P); T2n = VSUB(T2h, T2m); T2q = VSUB(T2o, T2p); T2r = VADD(T2n, T2q); T2D = VSUB(T2q, T2n); T2g = VADD(T1k, T2f); T2K = VMUL(LDK(KP500000000), VBYI(VADD(T2D, T2C))); T2L = VADD(T2F, T2G); T2v = VMUL(LDK(KP500000000), VBYI(VADD(T2r, T2u))); T2y = VSUB(T1k, T2f); T2E = VMUL(LDK(KP500000000), VBYI(VSUB(T2C, T2D))); T2H = VSUB(T2F, T2G); T2z = VMUL(LDK(KP500000000), VBYI(VSUB(T2u, T2r))); T2w = VCONJ(VSUB(T2g, T2v)); ST(&(Rm[0]), T2w, -ms, &(Rm[0])); T2M = VADD(T2K, T2L); ST(&(Rp[WS(rs, 7)]), T2M, ms, &(Rp[WS(rs, 1)])); T2N = VCONJ(VSUB(T2L, T2K)); ST(&(Rm[WS(rs, 6)]), T2N, -ms, &(Rm[0])); T2x = VADD(T2g, T2v); ST(&(Rp[WS(rs, 1)]), T2x, ms, &(Rp[WS(rs, 1)])); T2A = VCONJ(VSUB(T2y, T2z)); ST(&(Rm[WS(rs, 14)]), T2A, -ms, &(Rm[0])); T2I = VADD(T2E, T2H); ST(&(Rp[WS(rs, 9)]), T2I, ms, &(Rp[WS(rs, 1)])); T2J = VCONJ(VSUB(T2H, T2E)); ST(&(Rm[WS(rs, 8)]), T2J, -ms, &(Rm[0])); T2B = VADD(T2y, T2z); ST(&(Rp[WS(rs, 15)]), T2B, ms, &(Rp[WS(rs, 1)])); } { V T2Y, T3k, T3l, T35, T38, T3e, T3h, T39, T2Q, T3f, T34, T3g, T2X, T3c, T31; V T3d, T2O, T2P, T32, T33, T2T, T2W, T2Z, T30, T36, T3m, T3n, T37, T3a, T3i; V T3j, T3b; T2O = VSUB(Ta, Tx); T2P = VMUL(LDK(KP500000000), VADD(T2p, T2o)); T2Q = VADD(T2O, T2P); T3f = VSUB(T2O, T2P); T32 = VFNMS(LDK(KP555570233), T2R, VMUL(LDK(KP831469612), T2S)); T33 = VFMA(LDK(KP555570233), T2U, VMUL(LDK(KP831469612), T2V)); T34 = VADD(T32, T33); T3g = VMUL(LDK(KP500000000), VSUB(T33, T32)); T2T = VFMA(LDK(KP831469612), T2R, VMUL(LDK(KP555570233), T2S)); T2W = VFNMS(LDK(KP555570233), T2V, VMUL(LDK(KP831469612), T2U)); T2X = VMUL(LDK(KP500000000), VADD(T2T, T2W)); T3c = VSUB(T2W, T2T); T2Z = VADD(T2m, T2h); T30 = VSUB(T1i, TV); T31 = VADD(T2Z, T30); T3d = VSUB(T30, T2Z); T2Y = VADD(T2Q, T2X); T3k = VMUL(LDK(KP500000000), VBYI(VADD(T3d, T3c))); T3l = VADD(T3f, T3g); T35 = VMUL(LDK(KP500000000), VBYI(VADD(T31, T34))); T38 = VSUB(T2Q, T2X); T3e = VMUL(LDK(KP500000000), VBYI(VSUB(T3c, T3d))); T3h = VSUB(T3f, T3g); T39 = VMUL(LDK(KP500000000), VBYI(VSUB(T34, T31))); T36 = VCONJ(VSUB(T2Y, T35)); ST(&(Rm[WS(rs, 2)]), T36, -ms, &(Rm[0])); T3m = VADD(T3k, T3l); ST(&(Rp[WS(rs, 5)]), T3m, ms, &(Rp[WS(rs, 1)])); T3n = VCONJ(VSUB(T3l, T3k)); ST(&(Rm[WS(rs, 4)]), T3n, -ms, &(Rm[0])); T37 = VADD(T2Y, T35); ST(&(Rp[WS(rs, 3)]), T37, ms, &(Rp[WS(rs, 1)])); T3a = VCONJ(VSUB(T38, T39)); ST(&(Rm[WS(rs, 12)]), T3a, -ms, &(Rm[0])); T3i = VADD(T3e, T3h); ST(&(Rp[WS(rs, 11)]), T3i, ms, &(Rp[WS(rs, 1)])); T3j = VCONJ(VSUB(T3h, T3e)); ST(&(Rm[WS(rs, 10)]), T3j, -ms, &(Rm[0])); T3b = VADD(T38, T39); ST(&(Rp[WS(rs, 13)]), T3b, ms, &(Rp[WS(rs, 1)])); } } }}static const tw_instr twinstr[] = { VTW(1, 1), VTW(1, 2), VTW(1, 3), VTW(1, 4), VTW(1, 5), VTW(1, 6), VTW(1, 7), VTW(1, 8), VTW(1, 9), VTW(1, 10), VTW(1, 11), VTW(1, 12), VTW(1, 13), VTW(1, 14), VTW(1, 15), VTW(1, 16), VTW(1, 17), VTW(1, 18), VTW(1, 19), VTW(1, 20), VTW(1, 21), VTW(1, 22), VTW(1, 23), VTW(1, 24), VTW(1, 25), VTW(1, 26), VTW(1, 27), VTW(1, 28), VTW(1, 29), VTW(1, 30), VTW(1, 31), {TW_NEXT, VL, 0}};static const hc2c_desc desc = { 32, "hc2cfdftv_32", twinstr, &GENUS, {233, 117, 16, 0} };void X(codelet_hc2cfdftv_32) (planner *p) { X(khc2c_register) (p, hc2cfdftv_32, &desc, HC2C_VIA_DFT);}#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 + -