/** * generated using metalibm 1.1 * sha1 git: b'"44931432e12dbbda32ae5520e55557ff367a96bb"'(clean) * * INFO: git status was clean when file was generated. * * command used for generation: * metalibm_functions/ml_cbrt.py --precision binary64 --target x86_sse2 \ * --output /root/s_cbrt_sse2.c * **/ #include #include #include #include #include static const __m128i my_cbrt_cst = {INT32_C(0)}/*sse*/; static const __m128i my_cbrt_cst1 = {INT32_C(0)}/*sse*/; static const __m128d my_cbrt_cst2 = {1.0}/*sse*/; static const __m128d my_cbrt_cst3 = {1.0}/*sse*/; static const __m128d my_cbrt_cst4 = {0.3333333333333333}/*sse*/; static const __m128d my_cbrt_cst5 = {1.0}/*sse*/; static const __m128d my_cbrt_cst6 = {0.3333333333333333}/*sse*/; static const __m128d my_cbrt_cst7 = {1.0}/*sse*/; static const __m128d my_cbrt_cst8 = {0.3333333333333333}/*sse*/; static const __m128d my_cbrt_cst9 = {1.0}/*sse*/; static const __m128d my_cbrt_cst10 = {0.3333333333333333}/*sse*/; static const __m128d my_cbrt_cst11 = {1.0}/*sse*/; static const __m128d my_cbrt_cst12 = {0.3333333333333333}/*sse*/; static const __m128d my_cbrt_cst13 = {1.0}/*sse*/; static const __m128d my_cbrt_cst14 = {0.3333333333333333}/*sse*/; static const __m128d my_cbrt_cst15 = {1.0}/*sse*/; static const __m128d my_cbrt_cst16 = {0.3333333333333333}/*sse*/; static const __m128d my_cbrt_cst17 = {1.0}/*sse*/; static const __m128d my_cbrt_cst18 = {0.3333333333333333}/*sse*/; static const double my_cbrt_cbrt_mod_table[6][1] = { { 0x1p-1}, { 0x1.428a2f98d728bp-1}, { 0x1.965fea53d6e3dp-1}, { 0x1p0}, { 0x1.428a2f98d728bp0}, { 0x1.965fea53d6e3dp0} }; static const double my_cbrt_cbrt_approx_table[64][1] = { { 0x1p0}, { 0x1.01539221d4c97p0}, { 0x1.02a3ad2ef6f48p0}, { 0x1.03f06771a2e33p0}, { 0x1.0539d6521256fp0}, { 0x1.06800e629d672p0}, { 0x1.07c3236b0a73ap0}, { 0x1.090328731deb2p0}, { 0x1.0a402fcc79298p0}, { 0x1.0b7a4b1bd64acp0}, { 0x1.0cb18b61ad8cfp0}, { 0x1.0de601024fb88p0}, { 0x1.0f17bbcd80046p0}, { 0x1.1046cb0597001p0}, { 0x1.11733d66373bdp0}, { 0x1.129d212a9ba9cp0}, { 0x1.13c484138704fp0}, { 0x1.14e9736cdaf39p0}, { 0x1.160bfc12dd091p0}, { 0x1.172c2a772f508p0}, { 0x1.184a0aa58191fp0}, { 0x1.1965a848001d3p0}, { 0x1.1a7f0eab8483dp0}, { 0x1.1b9648c38c55dp0}, { 0x1.1cab612df9a46p0}, { 0x1.1dbe6236a0c45p0}, { 0x1.1ecf55daa68a5p0}, { 0x1.1fde45cbb1f9fp0}, { 0x1.20eb3b72f42d5p0}, { 0x1.21f63ff409043p0}, { 0x1.22ff5c2fb2fdp0}, { 0x1.240698c6746e5p0}, { 0x1.250bfe1b082f5p0}, { 0x1.260f9454bb99bp0}, { 0x1.27116361abaeap0}, { 0x1.281172f8e7074p0}, { 0x1.290fca9c761f8p0}, { 0x1.2a0c719b4b6d1p0}, { 0x1.2b076f131c9d7p0}, { 0x1.2c00c9f2263edp0}, { 0x1.2cf888f8db02fp0}, { 0x1.2deeb2bb7fb79p0}, { 0x1.2ee34da3b4fe3p0}, { 0x1.2fd65ff1efbbcp0}, { 0x1.30c7efbee12adp0}, { 0x1.31b802fccf6a2p0}, { 0x1.32a69f78df567p0}, { 0x1.3393cadc50709p0}, { 0x1.347f8aadab855p0}, { 0x1.3569e451e4c2bp0}, { 0x1.3652dd0d71db1p0}, { 0x1.373a7a0554cdfp0}, { 0x1.3820c0401be52p0}, { 0x1.3905b4a6d76cep0}, { 0x1.39e95c0605a66p0}, { 0x1.3acbbb0e756b7p0}, { 0x1.3bacd6561ff5ep0}, { 0x1.3c8cb258fa341p0}, { 0x1.3d6b5379be10cp0}, { 0x1.3e48be02ac0cfp0}, { 0x1.3f24f62645865p0}, { 0x1.4p0}, { 0x1.40d9df94f1be1p0}, { 0x1.41b298d47800ep0} }; double my_cbrt(double); double my_cbrt(double x){ int64_t tmp; uint64_t _srl_cast; int64_t tmp1; int64_t tmp2; int64_t exp_vx; int64_t exp_vx_mod; int64_t tmp3; int32_t tmp4; double cbrt_mod; double tmp5; int64_t tmp6; uint64_t _srl_cast1; int64_t tmp7; int64_t table_index; int32_t tmp8; double seed; double init_mult; int64_t exp_vx_third; int64_t tmp9; int64_t tmp10; double exp_vx_third1; double init_approx; double tmp11; double tmp12; double tmp13; double tmp14; double tmp15; double carg; double tmp16; double tmp17; double tmp18; double approx_0; double tmp19; double tmp20; double tmp21; double carg1; double tmp22; double tmp23; double tmp24; double approx_1; double tmp25; double tmp26; double tmp27; double carg2; double tmp28; double tmp29; double tmp30; double approx_2; double tmp31; double tmp32; double tmp33; double carg3; double tmp34; double tmp35; double tmp36; double approx_3; double tmp37; double tmp38; double tmp39; double carg4; double tmp40; double tmp41; double tmp42; double approx_4; double tmp43; double tmp44; double tmp45; double carg5; double tmp46; double tmp47; double tmp48; double approx_5; double tmp49; double tmp50; double tmp51; double carg6; double tmp52; double tmp53; double tmp54; double approx_6; double tmp55; double tmp56; double tmp57; double carg7; double tmp58; double tmp59; double tmp60; double result; tmp = double_to_64b_encoding(x); _srl_cast = tmp; tmp1 = _srl_cast >> INT64_C(52); tmp2 = tmp1 & INT64_C(2047); exp_vx = tmp2 + INT64_C(-1023); exp_vx_mod = exp_vx % INT64_C(3); tmp3 = exp_vx_mod + INT64_C(3); tmp4 = _mm_cvtsi128_si32(my_cbrt_cst); cbrt_mod = my_cbrt_cbrt_mod_table[tmp3][tmp4]; tmp5 = ml_mantissa_extraction_fp64(x); tmp6 = double_to_64b_encoding(tmp5); _srl_cast1 = tmp6; tmp7 = _srl_cast1 >> INT64_C(46); table_index = tmp7 & INT64_C(63); tmp8 = _mm_cvtsi128_si32(my_cbrt_cst1); seed = my_cbrt_cbrt_approx_table[table_index][tmp8]; init_mult = seed * cbrt_mod; exp_vx_third = exp_vx / INT64_C(3); tmp9 = exp_vx_third + INT64_C(1023); tmp10 = tmp9 << INT64_C(52); exp_vx_third1 = double_from_64b_encoding(tmp10); init_approx = init_mult * exp_vx_third1; tmp11 = _mm_cvtsd_f64(my_cbrt_cst2); tmp12 = tmp11 / x; tmp13 = init_approx * init_approx; tmp14 = init_approx * tmp13; tmp15 = _mm_cvtsd_f64(my_cbrt_cst3); carg = - tmp14; tmp16 = fma(carg, tmp12, tmp15); tmp17 = _mm_cvtsd_f64(my_cbrt_cst4); tmp18 = init_approx * tmp17; approx_0 = fma(tmp18, tmp16, init_approx); tmp19 = approx_0 * approx_0; tmp20 = approx_0 * tmp19; tmp21 = _mm_cvtsd_f64(my_cbrt_cst5); carg1 = - tmp20; tmp22 = fma(carg1, tmp12, tmp21); tmp23 = _mm_cvtsd_f64(my_cbrt_cst6); tmp24 = approx_0 * tmp23; approx_1 = fma(tmp24, tmp22, approx_0); tmp25 = approx_1 * approx_1; tmp26 = approx_1 * tmp25; tmp27 = _mm_cvtsd_f64(my_cbrt_cst7); carg2 = - tmp26; tmp28 = fma(carg2, tmp12, tmp27); tmp29 = _mm_cvtsd_f64(my_cbrt_cst8); tmp30 = approx_1 * tmp29; approx_2 = fma(tmp30, tmp28, approx_1); tmp31 = approx_2 * approx_2; tmp32 = approx_2 * tmp31; tmp33 = _mm_cvtsd_f64(my_cbrt_cst9); carg3 = - tmp32; tmp34 = fma(carg3, tmp12, tmp33); tmp35 = _mm_cvtsd_f64(my_cbrt_cst10); tmp36 = approx_2 * tmp35; approx_3 = fma(tmp36, tmp34, approx_2); tmp37 = approx_3 * approx_3; tmp38 = approx_3 * tmp37; tmp39 = _mm_cvtsd_f64(my_cbrt_cst11); carg4 = - tmp38; tmp40 = fma(carg4, tmp12, tmp39); tmp41 = _mm_cvtsd_f64(my_cbrt_cst12); tmp42 = approx_3 * tmp41; approx_4 = fma(tmp42, tmp40, approx_3); tmp43 = approx_4 * approx_4; tmp44 = approx_4 * tmp43; tmp45 = _mm_cvtsd_f64(my_cbrt_cst13); carg5 = - tmp44; tmp46 = fma(carg5, tmp12, tmp45); tmp47 = _mm_cvtsd_f64(my_cbrt_cst14); tmp48 = approx_4 * tmp47; approx_5 = fma(tmp48, tmp46, approx_4); tmp49 = approx_5 * approx_5; tmp50 = approx_5 * tmp49; tmp51 = _mm_cvtsd_f64(my_cbrt_cst15); carg6 = - tmp50; tmp52 = fma(carg6, tmp12, tmp51); tmp53 = _mm_cvtsd_f64(my_cbrt_cst16); tmp54 = approx_5 * tmp53; approx_6 = fma(tmp54, tmp52, approx_5); tmp55 = approx_6 * approx_6; tmp56 = approx_6 * tmp55; tmp57 = _mm_cvtsd_f64(my_cbrt_cst17); carg7 = - tmp56; tmp58 = fma(carg7, tmp12, tmp57); tmp59 = _mm_cvtsd_f64(my_cbrt_cst18); tmp60 = approx_6 * tmp59; result = fma(tmp60, tmp58, approx_6); return result; }