📄 hc2cbdftv_32.c
字号:
T2h = VFMA(LDK(KP980785280), T2f, VMUL(LDK(KP195090322), T2g)); TV = LD(&(Rp[WS(rs, 3)]), ms, &(Rp[WS(rs, 1)])); TW = LD(&(Rm[WS(rs, 12)]), -ms, &(Rm[0])); TX = VCONJ(TW); TY = VSUB(TV, TX); T1I = VADD(TV, TX); T1b = LD(&(Rp[WS(rs, 15)]), ms, &(Rp[WS(rs, 1)])); T19 = LD(&(Rm[0]), -ms, &(Rm[0])); T1a = VCONJ(T19); T1c = VSUB(T1a, T1b); T1F = VADD(T1a, T1b); T11 = LD(&(Rp[WS(rs, 11)]), ms, &(Rp[WS(rs, 1)])); TZ = LD(&(Rm[WS(rs, 4)]), -ms, &(Rm[0])); T10 = VCONJ(TZ); T12 = VSUB(T10, T11); T1J = VADD(T10, T11); T14 = LD(&(Rp[WS(rs, 7)]), ms, &(Rp[WS(rs, 1)])); T15 = LD(&(Rm[WS(rs, 8)]), -ms, &(Rm[0])); T16 = VCONJ(T15); T17 = VSUB(T14, T16); T1G = VADD(T14, T16); T1H = VSUB(T1F, T1G); T1K = VSUB(T1I, T1J); T1L = VFMA(LDK(KP923879532), T1H, VMUL(LDK(KP382683432), T1K)); T25 = VFNMS(LDK(KP382683432), T1H, VMUL(LDK(KP923879532), T1K)); T3J = VADD(T1F, T1G); T3K = VADD(T1I, T1J); T3L = VSUB(T3J, T3K); T4q = VADD(T3J, T3K); T13 = VMUL(LDK(KP707106781), VSUB(TY, T12)); T18 = VSUB(T13, T17); T2i = VADD(T17, T13); T1d = VMUL(LDK(KP707106781), VADD(TY, T12)); T1e = VSUB(T1c, T1d); T2j = VADD(T1c, T1d); T1f = VFNMS(LDK(KP555570233), T1e, VMUL(LDK(KP831469612), T18)); T2o = VFMA(LDK(KP195090322), T2i, VMUL(LDK(KP980785280), T2j)); T1j = VFMA(LDK(KP555570233), T18, VMUL(LDK(KP831469612), T1e)); T2k = VFNMS(LDK(KP195090322), T2j, VMUL(LDK(KP980785280), T2i)); } { V T4L, T4G, T4s, T4y, T3W, T4g, T42, T4a, T3g, T4e, T3o, T3E, T1w, T46, T2M; V T40, T2u, T4w, T2C, T4k, T36, T3A, T3i, T3s, T28, T2O, T2w, T2G, T2Y, T4K; V T3y, T4C; { V T4E, T4F, T4D, T4o, T4r, T4l, T4x, T3Q, T48, T3V, T49, T3M, T3U, T3F, T4f; V T41, T47, T3c, T3n, T3f, T3m, T3a, T3b, T3d, T3e, T39, T4d, T3l, T3D, T1h; V T2K, T1v, T2L, Tz, T1g, T1k, T1u, T1, T45, T2J, T3Z, T2m, T2A, T2t, T2B; V T2e, T2l, T2p, T2s, T2b, T4v, T2z, T4j; T4E = VADD(T4m, T4n); T4F = VADD(T4p, T4q); T4L = VADD(T4E, T4F); T4D = LDW(&(W[TWVL * 30])); T4G = VZMUL(T4D, VSUB(T4E, T4F)); T4o = VSUB(T4m, T4n); T4r = VBYI(VSUB(T4p, T4q)); T4l = LDW(&(W[TWVL * 46])); T4s = VZMUL(T4l, VSUB(T4o, T4r)); T4x = LDW(&(W[TWVL * 14])); T4y = VZMUL(T4x, VADD(T4o, T4r)); T3M = VMUL(LDK(KP707106781), VSUB(T3I, T3L)); T3Q = VBYI(VSUB(T3M, T3P)); T48 = VBYI(VADD(T3P, T3M)); T3U = VMUL(LDK(KP707106781), VADD(T3I, T3L)); T3V = VSUB(T3T, T3U); T49 = VADD(T3T, T3U); T3F = LDW(&(W[TWVL * 22])); T3W = VZMUL(T3F, VADD(T3Q, T3V)); T4f = LDW(&(W[TWVL * 54])); T4g = VZMUL(T4f, VSUB(T49, T48)); T41 = LDW(&(W[TWVL * 38])); T42 = VZMUL(T41, VSUB(T3V, T3Q)); T47 = LDW(&(W[TWVL * 6])); T4a = VZMUL(T47, VADD(T48, T49)); T3a = VADD(T1t, T1n); T3b = VADD(TU, T1f); T3c = VBYI(VADD(T3a, T3b)); T3n = VBYI(VSUB(T3b, T3a)); T3d = VADD(Tf, Ty); T3e = VADD(T1i, T1j); T3f = VADD(T3d, T3e); T3m = VSUB(T3d, T3e); T39 = LDW(&(W[TWVL * 4])); T3g = VZMULI(T39, VADD(T3c, T3f)); T4d = LDW(&(W[TWVL * 56])); T4e = VZMULI(T4d, VSUB(T3f, T3c)); T3l = LDW(&(W[TWVL * 36])); T3o = VZMULI(T3l, VSUB(T3m, T3n)); T3D = LDW(&(W[TWVL * 24])); T3E = VZMULI(T3D, VADD(T3n, T3m)); Tz = VSUB(Tf, Ty); T1g = VSUB(TU, T1f); T1h = VSUB(Tz, T1g); T2K = VADD(Tz, T1g); T1k = VSUB(T1i, T1j); T1u = VSUB(T1n, T1t); T1v = VBYI(VSUB(T1k, T1u)); T2L = VBYI(VADD(T1u, T1k)); T1 = LDW(&(W[TWVL * 20])); T1w = VZMULI(T1, VADD(T1h, T1v)); T45 = LDW(&(W[TWVL * 8])); T46 = VZMULI(T45, VADD(T2K, T2L)); T2J = LDW(&(W[TWVL * 52])); T2M = VZMULI(T2J, VSUB(T2K, T2L)); T3Z = LDW(&(W[TWVL * 40])); T40 = VZMULI(T3Z, VSUB(T1h, T1v)); T2e = VSUB(T2c, T2d); T2l = VSUB(T2h, T2k); T2m = VSUB(T2e, T2l); T2A = VADD(T2e, T2l); T2p = VSUB(T2n, T2o); T2s = VSUB(T2q, T2r); T2t = VBYI(VSUB(T2p, T2s)); T2B = VBYI(VADD(T2s, T2p)); T2b = LDW(&(W[TWVL * 44])); T2u = VZMULI(T2b, VSUB(T2m, T2t)); T4v = LDW(&(W[TWVL * 16])); T4w = VZMULI(T4v, VADD(T2m, T2t)); T2z = LDW(&(W[TWVL * 12])); T2C = VZMULI(T2z, VADD(T2A, T2B)); T4j = LDW(&(W[TWVL * 48])); T4k = VZMULI(T4j, VSUB(T2A, T2B)); { V T32, T3q, T35, T3r, T30, T31, T33, T34, T2Z, T3z, T3h, T3p, T1Y, T2E, T27; V T2F, T1M, T1X, T23, T26, T1x, T2N, T2v, T2D, T2U, T3x, T2X, T3w, T2S, T2T; V T2V, T2W, T2R, T4J, T3v, T4B; T30 = VADD(T21, T22); T31 = VADD(T1E, T1L); T32 = VADD(T30, T31); T3q = VSUB(T30, T31); T33 = VADD(T1W, T1T); T34 = VADD(T24, T25); T35 = VBYI(VADD(T33, T34)); T3r = VBYI(VSUB(T34, T33)); T2Z = LDW(&(W[TWVL * 58])); T36 = VZMUL(T2Z, VSUB(T32, T35)); T3z = LDW(&(W[TWVL * 26])); T3A = VZMUL(T3z, VADD(T3q, T3r)); T3h = LDW(&(W[TWVL * 2])); T3i = VZMUL(T3h, VADD(T32, T35)); T3p = LDW(&(W[TWVL * 34])); T3s = VZMUL(T3p, VSUB(T3q, T3r)); T1M = VSUB(T1E, T1L); T1X = VSUB(T1T, T1W); T1Y = VBYI(VSUB(T1M, T1X)); T2E = VBYI(VADD(T1X, T1M)); T23 = VSUB(T21, T22); T26 = VSUB(T24, T25); T27 = VSUB(T23, T26); T2F = VADD(T23, T26); T1x = LDW(&(W[TWVL * 18])); T28 = VZMUL(T1x, VADD(T1Y, T27)); T2N = LDW(&(W[TWVL * 50])); T2O = VZMUL(T2N, VSUB(T2F, T2E)); T2v = LDW(&(W[TWVL * 42])); T2w = VZMUL(T2v, VSUB(T27, T1Y)); T2D = LDW(&(W[TWVL * 10])); T2G = VZMUL(T2D, VADD(T2E, T2F)); T2S = VADD(T2c, T2d); T2T = VADD(T2n, T2o); T2U = VADD(T2S, T2T); T3x = VSUB(T2S, T2T); T2V = VADD(T2r, T2q); T2W = VADD(T2h, T2k); T2X = VBYI(VADD(T2V, T2W)); T3w = VBYI(VSUB(T2W, T2V)); T2R = LDW(&(W[TWVL * 60])); T2Y = VZMULI(T2R, VSUB(T2U, T2X)); T4J = LDW(&(W[0])); T4K = VZMULI(T4J, VADD(T2X, T2U)); T3v = LDW(&(W[TWVL * 28])); T3y = VZMULI(T3v, VADD(T3w, T3x)); T4B = LDW(&(W[TWVL * 32])); T4C = VZMULI(T4B, VSUB(T3x, T3w)); } } { V T29, T4M, T2P, T4t, T4N, T2a, T4u, T2Q, T2x, T4H, T2H, T4z, T4I, T2y, T4A; V T2I, T37, T4h, T3B, T3X, T4i, T38, T3Y, T3C, T3j, T4b, T3t, T43, T4c, T3k; V T44, T3u; T29 = VADD(T1w, T28); ST(&(Rp[WS(rs, 5)]), T29, ms, &(Rp[WS(rs, 1)])); T4M = VADD(T4K, T4L); ST(&(Rp[0]), T4M, ms, &(Rp[0])); T2P = VADD(T2M, T2O); ST(&(Rp[WS(rs, 13)]), T2P, ms, &(Rp[WS(rs, 1)])); T4t = VADD(T4k, T4s); ST(&(Rp[WS(rs, 12)]), T4t, ms, &(Rp[0])); T4N = VCONJ(VSUB(T4L, T4K)); ST(&(Rm[0]), T4N, -ms, &(Rm[0])); T2a = VCONJ(VSUB(T28, T1w)); ST(&(Rm[WS(rs, 5)]), T2a, -ms, &(Rm[WS(rs, 1)])); T4u = VCONJ(VSUB(T4s, T4k)); ST(&(Rm[WS(rs, 12)]), T4u, -ms, &(Rm[0])); T2Q = VCONJ(VSUB(T2O, T2M)); ST(&(Rm[WS(rs, 13)]), T2Q, -ms, &(Rm[WS(rs, 1)])); T2x = VADD(T2u, T2w); ST(&(Rp[WS(rs, 11)]), T2x, ms, &(Rp[WS(rs, 1)])); T4H = VADD(T4C, T4G); ST(&(Rp[WS(rs, 8)]), T4H, ms, &(Rp[0])); T2H = VADD(T2C, T2G); ST(&(Rp[WS(rs, 3)]), T2H, ms, &(Rp[WS(rs, 1)])); T4z = VADD(T4w, T4y); ST(&(Rp[WS(rs, 4)]), T4z, ms, &(Rp[0])); T4I = VCONJ(VSUB(T4G, T4C)); ST(&(Rm[WS(rs, 8)]), T4I, -ms, &(Rm[0])); T2y = VCONJ(VSUB(T2w, T2u)); ST(&(Rm[WS(rs, 11)]), T2y, -ms, &(Rm[WS(rs, 1)])); T4A = VCONJ(VSUB(T4y, T4w)); ST(&(Rm[WS(rs, 4)]), T4A, -ms, &(Rm[0])); T2I = VCONJ(VSUB(T2G, T2C)); ST(&(Rm[WS(rs, 3)]), T2I, -ms, &(Rm[WS(rs, 1)])); T37 = VADD(T2Y, T36); ST(&(Rp[WS(rs, 15)]), T37, ms, &(Rp[WS(rs, 1)])); T4h = VADD(T4e, T4g); ST(&(Rp[WS(rs, 14)]), T4h, ms, &(Rp[0])); T3B = VADD(T3y, T3A); ST(&(Rp[WS(rs, 7)]), T3B, ms, &(Rp[WS(rs, 1)])); T3X = VADD(T3E, T3W); ST(&(Rp[WS(rs, 6)]), T3X, ms, &(Rp[0])); T4i = VCONJ(VSUB(T4g, T4e)); ST(&(Rm[WS(rs, 14)]), T4i, -ms, &(Rm[0])); T38 = VCONJ(VSUB(T36, T2Y)); ST(&(Rm[WS(rs, 15)]), T38, -ms, &(Rm[WS(rs, 1)])); T3Y = VCONJ(VSUB(T3W, T3E)); ST(&(Rm[WS(rs, 6)]), T3Y, -ms, &(Rm[0])); T3C = VCONJ(VSUB(T3A, T3y)); ST(&(Rm[WS(rs, 7)]), T3C, -ms, &(Rm[WS(rs, 1)])); T3j = VADD(T3g, T3i); ST(&(Rp[WS(rs, 1)]), T3j, ms, &(Rp[WS(rs, 1)])); T4b = VADD(T46, T4a); ST(&(Rp[WS(rs, 2)]), T4b, ms, &(Rp[0])); T3t = VADD(T3o, T3s); ST(&(Rp[WS(rs, 9)]), T3t, ms, &(Rp[WS(rs, 1)])); T43 = VADD(T40, T42); ST(&(Rp[WS(rs, 10)]), T43, ms, &(Rp[0])); T4c = VCONJ(VSUB(T4a, T46)); ST(&(Rm[WS(rs, 2)]), T4c, -ms, &(Rm[0])); T3k = VCONJ(VSUB(T3i, T3g)); ST(&(Rm[WS(rs, 1)]), T3k, -ms, &(Rm[WS(rs, 1)])); T44 = VCONJ(VSUB(T42, T40)); ST(&(Rm[WS(rs, 10)]), T44, -ms, &(Rm[0])); T3u = VCONJ(VSUB(T3s, T3o)); ST(&(Rm[WS(rs, 9)]), T3u, -ms, &(Rm[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, "hc2cbdftv_32", twinstr, &GENUS, {233, 88, 16, 0} };void X(codelet_hc2cbdftv_32) (planner *p) { X(khc2c_register) (p, hc2cbdftv_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 + -