| /external/XNNPACK/src/math/ |
| D | exp-avx512f-rr2-p5-scalef.c | 20 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_math_f32_exp__avx512f_rr2_p5_scalef()
23 const __m512 vzero_cutoff = _mm512_set1_ps(-0x1.9FE368p+6f); in xnn_math_f32_exp__avx512f_rr2_p5_scalef()
25 const __m512 vinf_cutoff = _mm512_set1_ps(0x1.62E42Ep+6f); in xnn_math_f32_exp__avx512f_rr2_p5_scalef()
27 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_math_f32_exp__avx512f_rr2_p5_scalef()
28 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_math_f32_exp__avx512f_rr2_p5_scalef()
30 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_math_f32_exp__avx512f_rr2_p5_scalef()
31 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); in xnn_math_f32_exp__avx512f_rr2_p5_scalef()
32 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); in xnn_math_f32_exp__avx512f_rr2_p5_scalef()
33 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); in xnn_math_f32_exp__avx512f_rr2_p5_scalef()
34 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); in xnn_math_f32_exp__avx512f_rr2_p5_scalef()
[all …]
|
| D | exp-avx512f-rr2-p5.c | 21 const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p+23f); in xnn_math_f32_exp__avx512f_rr2_p5()
23 const __m512 vzero_cutoff = _mm512_set1_ps(-0x1.9FE368p+6f); in xnn_math_f32_exp__avx512f_rr2_p5()
25 const __m512 vinf_cutoff = _mm512_set1_ps(0x1.62E42Ep+6f); in xnn_math_f32_exp__avx512f_rr2_p5()
26 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_math_f32_exp__avx512f_rr2_p5()
27 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_math_f32_exp__avx512f_rr2_p5()
28 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_math_f32_exp__avx512f_rr2_p5()
29 const __m512 vplus_inf = _mm512_set1_ps(INFINITY); in xnn_math_f32_exp__avx512f_rr2_p5()
31 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); in xnn_math_f32_exp__avx512f_rr2_p5()
32 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); in xnn_math_f32_exp__avx512f_rr2_p5()
33 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); in xnn_math_f32_exp__avx512f_rr2_p5()
[all …]
|
| D | expm1minus-avx512f-rr1-p6.c | 22 const __m512 vsat_cutoff = _mm512_set1_ps(-0x1.154246p+4f); in xnn_math_f32_expm1minus__avx512f_rr1_p6()
24 const __m512 vmagic_bias = _mm512_set1_ps(0x1.8000FEp23f); in xnn_math_f32_expm1minus__avx512f_rr1_p6()
25 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_math_f32_expm1minus__avx512f_rr1_p6()
26 const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_math_f32_expm1minus__avx512f_rr1_p6()
30 const __m512 vc6 = _mm512_set1_ps(0x1.6b7338p-10f); in xnn_math_f32_expm1minus__avx512f_rr1_p6()
31 const __m512 vc5 = _mm512_set1_ps(0x1.12278Ep-7f); in xnn_math_f32_expm1minus__avx512f_rr1_p6()
32 const __m512 vc4 = _mm512_set1_ps(0x1.555716p-5f); in xnn_math_f32_expm1minus__avx512f_rr1_p6()
33 const __m512 vc3 = _mm512_set1_ps(0x1.5554B0p-3f); in xnn_math_f32_expm1minus__avx512f_rr1_p6()
34 const __m512 vc2 = _mm512_set1_ps(0x1.FFFFFEp-2f); in xnn_math_f32_expm1minus__avx512f_rr1_p6()
35 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_math_f32_expm1minus__avx512f_rr1_p6()
|
| D | extexp-avx512f-p5.c | 22 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_math_f32_extexp__avx512f_p5()
23 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_math_f32_extexp__avx512f_p5()
24 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_math_f32_extexp__avx512f_p5()
26 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_math_f32_extexp__avx512f_p5()
27 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); in xnn_math_f32_extexp__avx512f_p5()
28 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); in xnn_math_f32_extexp__avx512f_p5()
29 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); in xnn_math_f32_extexp__avx512f_p5()
30 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); in xnn_math_f32_extexp__avx512f_p5()
31 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_math_f32_extexp__avx512f_p5()
|
| D | sigmoid-avx512f-rr2-p5-scalef-div.c | 23 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_div()
24 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_div()
25 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_div()
28 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_div()
29 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_div()
30 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_div()
31 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_div()
32 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_div()
33 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_div()
|
| D | sigmoid-avx512f-rr2-p5-scalef-nr1fma.c | 23 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma()
24 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma()
25 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma()
28 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma()
29 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma()
30 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma()
31 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma()
32 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma()
33 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma()
|
| D | sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c | 23 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma1adj()
24 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma1adj()
25 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma1adj()
28 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma1adj()
29 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma1adj()
30 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma1adj()
31 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma1adj()
32 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma1adj()
33 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma1adj()
|
| D | exp-avx512f-rr2-lut32-p2-perm2.c | 21 const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p23f); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2()
22 const __m512 vlog2e_x32 = _mm512_set1_ps(0x1.715476p5f); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2()
24 const __m512 vzero_cutoff = _mm512_set1_ps(-0x1.9FE368p6f); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2()
26 const __m512 vinf_cutoff = _mm512_set1_ps(0x1.62E42Ep6f); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2()
27 const __m512 vminus_ln2_o32_hi = _mm512_set1_ps(-0x1.62e43p-6f); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2()
28 const __m512 vminus_ln2_o32_lo = _mm512_set1_ps(0x1.05c61p-34f); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2()
29 const __m512 vplus_inf = _mm512_set1_ps(INFINITY); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2()
31 const __m512 vc1 = _mm512_set1_ps(0x1.0000F6p-0f); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2()
32 const __m512 vc2 = _mm512_set1_ps(0x1.000000p-1f); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2()
|
| D | exp-avx512f-rr2-lut16-p3-perm.c | 21 const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p23f); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm()
22 const __m512 vlog2e_x16 = _mm512_set1_ps(0x1.715476p4f); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm()
24 const __m512 vzero_cutoff = _mm512_set1_ps(-0x1.9FE368p6f); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm()
26 const __m512 vinf_cutoff = _mm512_set1_ps(0x1.62E42Ep6f); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm()
27 const __m512 vminus_ln2_o16_hi = _mm512_set1_ps(-0x1.62e43p-5f); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm()
28 const __m512 vminus_ln2_o16_lo = _mm512_set1_ps(0x1.05c61p-33f); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm()
29 const __m512 vplus_inf = _mm512_set1_ps(INFINITY); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm()
31 const __m512 vc2 = _mm512_set1_ps(0x1.00021Ep-1f); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm()
32 const __m512 vc3 = _mm512_set1_ps(0x1.55559Ap-3f); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm()
|
| D | sigmoid-avx512f-rr1-p5-scalef-div.c | 23 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_div()
24 const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_div()
27 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_div()
28 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_div()
29 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_div()
30 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_div()
31 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_div()
32 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_div()
|
| D | sigmoid-avx512f-rr1-p5-scalef-nr1fma.c | 23 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma()
24 const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma()
27 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma()
28 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma()
29 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma()
30 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma()
31 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma()
32 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma()
|
| D | sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c | 23 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma1adj()
24 const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma1adj()
27 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma1adj()
28 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma1adj()
29 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma1adj()
30 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma1adj()
31 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma1adj()
32 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma1adj()
|
| D | expm1minus-avx512f-rr1-lut16-p3-perm.c | 24 const __m512 vsat_cutoff = _mm512_set1_ps(-0x1.154246p+4f); in xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm()
26 const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p19f); in xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm()
27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm()
32 const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm()
36 const __m512 vc3 = _mm512_set1_ps(0x1.55561Cp-3f); in xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm()
37 const __m512 vc2 = _mm512_set1_ps(0x1.0001ECp-1f); in xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm()
38 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm()
|
| D | sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c | 24 const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p18f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_div()
25 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_div()
37 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62e43p-1f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_div()
38 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05c61p-29f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_div()
41 const __m512 vc2 = _mm512_set1_ps(0x1.000000p-1f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_div()
42 const __m512 vc1 = _mm512_set1_ps(0x1.0000F6p-0f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_div()
43 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_div()
|
| D | sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c | 24 const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p19f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_div()
25 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_div()
32 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62e43p-1f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_div()
33 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05c61p-29f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_div()
36 const __m512 vc3 = _mm512_set1_ps(0x1.55559Ap-3f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_div()
37 const __m512 vc2 = _mm512_set1_ps(0x1.00021Ep-1f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_div()
38 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_div()
|
| D | sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c | 24 const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p19f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma()
25 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma()
32 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62e43p-1f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma()
33 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05c61p-29f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma()
36 const __m512 vc3 = _mm512_set1_ps(0x1.55559Ap-3f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma()
37 const __m512 vc2 = _mm512_set1_ps(0x1.00021Ep-1f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma()
38 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma()
|
| D | sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c | 24 const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p18f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma()
25 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma()
37 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62e43p-1f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma()
38 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05c61p-29f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma()
41 const __m512 vc2 = _mm512_set1_ps(0x1.000000p-1f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma()
42 const __m512 vc1 = _mm512_set1_ps(0x1.0000F6p-0f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma()
43 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma()
|
| D | sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c | 24 const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p19f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma1adj()
25 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma1adj()
32 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62e43p-1f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma1adj()
33 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05c61p-29f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma1adj()
36 const __m512 vc3 = _mm512_set1_ps(0x1.55559Ap-3f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma1adj()
37 const __m512 vc2 = _mm512_set1_ps(0x1.00021Ep-1f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma1adj()
38 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma1adj()
|
| D | sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c | 24 const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p18f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma1adj()
25 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma1adj()
37 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62e43p-1f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma1adj()
38 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05c61p-29f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma1adj()
41 const __m512 vc2 = _mm512_set1_ps(0x1.000000p-1f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma1adj()
42 const __m512 vc1 = _mm512_set1_ps(0x1.0000F6p-0f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma1adj()
43 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma1adj()
|
| D | exp-avx512f-rr2-lut16-p3-perm-scalef.c | 21 const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p19f); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm_scalef()
22 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm_scalef()
23 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62e43p-1f); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm_scalef()
24 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05c61p-29f); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm_scalef()
26 const __m512 vc2 = _mm512_set1_ps(0x1.00021Ep-1f); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm_scalef()
27 const __m512 vc3 = _mm512_set1_ps(0x1.55559Ap-3f); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm_scalef()
|
| D | exp-avx512f-rr2-lut32-p2-perm2-scalef.c | 21 const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p18f); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2_scalef()
22 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2_scalef()
23 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62e43p-1f); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2_scalef()
24 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05c61p-29f); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2_scalef()
26 const __m512 vc1 = _mm512_set1_ps(0x1.0000F6p-0f); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2_scalef()
27 const __m512 vc2 = _mm512_set1_ps(0x1.000000p-1f); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2_scalef()
|
| /external/XNNPACK/src/f32-vscaleexpminusmax/gen/ |
| D | avx512f-p5-scalef-x16.c | 27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16()
28 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16()
29 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16()
31 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16()
32 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16()
33 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16()
34 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16()
35 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16()
36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16()
38 const __m512 vscale = _mm512_set1_ps(scale); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16()
[all …]
|
| D | avx512f-p5-scalef-x32.c | 27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32()
28 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32()
29 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32()
31 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32()
32 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32()
33 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32()
34 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32()
35 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32()
36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32()
38 const __m512 vscale = _mm512_set1_ps(scale); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32()
[all …]
|
| /external/XNNPACK/src/f32-vscaleextexp/gen/ |
| D | avx512f-p5-scalef-x16.c | 28 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x16()
29 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x16()
30 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x16()
32 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x16()
33 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x16()
34 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x16()
35 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x16()
36 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x16()
37 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x16()
39 const __m512 vscalev = _mm512_set1_ps(scale_value); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x16()
[all …]
|
| D | avx512f-p5-scalef-x32.c | 28 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x32()
29 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x32()
30 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x32()
32 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x32()
33 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x32()
34 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x32()
35 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x32()
36 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x32()
37 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x32()
39 const __m512 vscalev = _mm512_set1_ps(scale_value); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x32()
[all …]
|