| /external/XNNPACK/src/f32-vscaleexpminusmax/gen/ |
| D | avx2-p5-x48.c | 69 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x48() local
78 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x48()
86 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x48()
95 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x48()
102 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x48()
|
| D | avx2-p5-x56.c | 71 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56() local
81 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56()
90 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56()
100 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56()
108 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56()
|
| D | avx2-p5-x64.c | 73 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64() local
84 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64()
94 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64()
105 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64()
114 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64()
|
| D | avx2-p5-x72.c | 75 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72() local
87 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72()
98 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72()
110 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72()
120 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72()
|
| D | avx512f-p5-scalef-x96.c | 65 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x96() local
74 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x96()
81 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x96()
127 __m512 vf5 = _mm512_scalef_ps(vp5, vn5); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x96()
|
| D | avx2-p5-x80.c | 77 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80() local
90 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80()
102 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80()
115 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80()
126 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80()
|
| D | avx512f-p5-scalef-x112.c | 67 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112() local
77 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112()
85 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112()
137 __m512 vf5 = _mm512_scalef_ps(vp5, vn5); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112()
|
| /external/XNNPACK/src/f32-raddexpminusmax/gen/ |
| D | avx2-p5-x64.c | 71 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64() local
82 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64()
92 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64()
103 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64()
112 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64()
|
| D | avx2-p5-x64-acc2.c | 72 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2() local
83 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2()
93 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2()
104 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2()
113 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2()
|
| D | avx2-p5-x64-acc4.c | 74 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4() local
85 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4()
95 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4()
106 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4()
115 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4()
|
| D | avx2-p5-x72.c | 73 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72() local
85 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72()
96 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72()
108 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72()
118 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72()
|
| D | avx2-p5-x72-acc3.c | 75 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3() local
87 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3()
98 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3()
110 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3()
120 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3()
|
| D | avx2-p5-x80.c | 75 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80() local
88 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80()
100 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80()
113 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80()
124 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80()
|
| D | avx2-p5-x80-acc2.c | 76 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2() local
89 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
101 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
114 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
125 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | avx2-p5-x64-acc2.c | 73 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc2() local
84 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc2()
94 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc2()
105 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc2()
114 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc2()
|
| D | avx2-p5-x64.c | 72 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64() local
83 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64()
93 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64()
104 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64()
113 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64()
|
| D | avx2-p5-x64-acc4.c | 75 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc4() local
86 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc4()
96 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc4()
107 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc4()
116 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc4()
|
| D | avx2-p5-x72.c | 74 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x72() local
86 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x72()
97 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x72()
109 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x72()
119 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x72()
|
| D | avx2-p5-x72-acc3.c | 76 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x72_acc3() local
88 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x72_acc3()
99 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x72_acc3()
111 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x72_acc3()
121 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x72_acc3()
|
| /external/XNNPACK/src/f32-raddextexp/gen/ |
| D | avx512f-p5-scalef-x128.c | 60 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128() local
71 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
80 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
145 vmax_e0 = _mm512_max_ps(vmax_e0, vn5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
155 const __m512 vdelta_e5 = _mm512_sub_ps(vn5, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
|
| D | avx512f-p5-scalef-x128-acc2.c | 62 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2() local
73 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
82 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
147 vmax_e1 = _mm512_max_ps(vmax_e1, vn5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
158 const __m512 vdelta_e5 = _mm512_sub_ps(vn5, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
|
| D | avx512f-p5-scalef-x144.c | 61 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144() local
73 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144()
83 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144()
154 vmax_e0 = _mm512_max_ps(vmax_e0, vn5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144()
165 const __m512 vdelta_e5 = _mm512_sub_ps(vn5, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144()
|
| /external/XNNPACK/src/f32-vscaleextexp/gen/ |
| D | avx512f-p5-scalef-x96.c | 58 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x96() local
67 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x96()
74 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x96()
130 const __m512 ve5 = _mm512_add_ps(vn5, vscalee); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x96()
|
| D | avx512f-p5-scalef-x112.c | 59 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112() local
69 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
77 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
140 const __m512 ve5 = _mm512_add_ps(vn5, vscalee); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
|
| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | avx2-rr1-p5-div-x48.c | 77 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x48() local
86 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x48()
94 vn5 = _mm256_sub_ps(vn5, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x48()
102 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x48()
|