| /external/XNNPACK/src/f32-vsqrt/gen/ |
| D | avx512f-nr1fma1adj-x128.c | 37 const __m512 vx7 = _mm512_loadu_ps(x + 112); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128() local
47 const __m512 vrsqrtx7 = _mm512_rsqrt14_ps(vx7); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
63 __m512 vsqrtx7 = _mm512_mul_ps(vrsqrtx7, vx7); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
99 const __m512 vadjustment7 = _mm512_fnmadd_ps(vsqrtx7, vsqrtx7, vx7); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
|
| D | fma3-nr1fma1adj-x64.c | 38 const __m256 vx7 = _mm256_loadu_ps(x + 56); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64() local
48 const __m256 vrsqrtx7 = _mm256_rsqrt_ps(vx7); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64()
64 __m256 vsqrtx7 = _mm256_mul_ps(vrsqrtx7, vx7); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64()
100 const __m256 vadjustment7 = _mm256_fnmadd_ps(vsqrtx7, vsqrtx7, vx7); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64()
|
| /external/XNNPACK/src/f32-velu/gen/ |
| D | velu-avx2-rr1-lut16-p3-gather-x64.c | 50 __m256 vx7 = _mm256_loadu_ps(x + 56); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x64() local
60 const __m256 vz7 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx7, vprescale)); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x64()
180 vx7 = _mm256_mul_ps(vx7, vbeta); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x64()
189 const __m256 vy7 = _mm256_blendv_ps(vx7, ve7, vx7); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x64()
|
| D | velu-avx2-rr1-p6-x64.c | 50 __m256 vx7 = _mm256_loadu_ps(x + 56); in xnn_f32_velu_ukernel__avx2_rr1_p6_x64() local
60 const __m256 vz7 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx7, vprescale)); in xnn_f32_velu_ukernel__avx2_rr1_p6_x64()
182 vx7 = _mm256_mul_ps(vx7, vbeta); in xnn_f32_velu_ukernel__avx2_rr1_p6_x64()
191 const __m256 vy7 = _mm256_blendv_ps(vx7, ve7, vx7); in xnn_f32_velu_ukernel__avx2_rr1_p6_x64()
|
| D | velu-avx2-rr1-lut8-p4-perm-x64.c | 50 __m256 vx7 = _mm256_loadu_ps(x + 56); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x64() local
60 const __m256 vz7 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx7, vprescale)); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x64()
180 vx7 = _mm256_mul_ps(vx7, vbeta); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x64()
189 const __m256 vy7 = _mm256_blendv_ps(vx7, ve7, vx7); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x64()
|
| D | velu-avx2-rr1-lut4-p4-perm-x64.c | 51 __m256 vx7 = _mm256_loadu_ps(x + 56); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x64() local
61 const __m256 vz7 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx7, vprescale)); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x64()
181 vx7 = _mm256_mul_ps(vx7, vbeta); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x64()
190 const __m256 vy7 = _mm256_blendv_ps(vx7, ve7, vx7); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x64()
|
| D | velu-avx2-rr1-lut8-p4-perm-x72.c | 50 __m256 vx7 = _mm256_loadu_ps(x + 56); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x72() local
61 const __m256 vz7 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx7, vprescale)); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x72()
194 vx7 = _mm256_mul_ps(vx7, vbeta); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x72()
205 const __m256 vy7 = _mm256_blendv_ps(vx7, ve7, vx7); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x72()
|
| D | velu-avx2-rr1-lut4-p4-perm-x72.c | 51 __m256 vx7 = _mm256_loadu_ps(x + 56); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x72() local
62 const __m256 vz7 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx7, vprescale)); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x72()
195 vx7 = _mm256_mul_ps(vx7, vbeta); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x72()
206 const __m256 vy7 = _mm256_blendv_ps(vx7, ve7, vx7); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x72()
|
| D | velu-avx2-rr1-p6-x72.c | 50 __m256 vx7 = _mm256_loadu_ps(x + 56); in xnn_f32_velu_ukernel__avx2_rr1_p6_x72() local
61 const __m256 vz7 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx7, vprescale)); in xnn_f32_velu_ukernel__avx2_rr1_p6_x72()
196 vx7 = _mm256_mul_ps(vx7, vbeta); in xnn_f32_velu_ukernel__avx2_rr1_p6_x72()
207 const __m256 vy7 = _mm256_blendv_ps(vx7, ve7, vx7); in xnn_f32_velu_ukernel__avx2_rr1_p6_x72()
|
| D | velu-avx2-rr1-lut16-p3-gather-x72.c | 50 __m256 vx7 = _mm256_loadu_ps(x + 56); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x72() local
61 const __m256 vz7 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx7, vprescale)); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x72()
194 vx7 = _mm256_mul_ps(vx7, vbeta); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x72()
205 const __m256 vy7 = _mm256_blendv_ps(vx7, ve7, vx7); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x72()
|
| D | velu-avx2-rr1-lut4-p4-perm-x80.c | 51 __m256 vx7 = _mm256_loadu_ps(x + 56); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x80() local
63 const __m256 vz7 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx7, vprescale)); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x80()
209 vx7 = _mm256_mul_ps(vx7, vbeta); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x80()
222 const __m256 vy7 = _mm256_blendv_ps(vx7, ve7, vx7); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x80()
|
| D | velu-avx2-rr1-lut16-p3-gather-x80.c | 50 __m256 vx7 = _mm256_loadu_ps(x + 56); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x80() local
62 const __m256 vz7 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx7, vprescale)); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x80()
208 vx7 = _mm256_mul_ps(vx7, vbeta); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x80()
221 const __m256 vy7 = _mm256_blendv_ps(vx7, ve7, vx7); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x80()
|
| D | velu-avx2-rr1-p6-x80.c | 50 __m256 vx7 = _mm256_loadu_ps(x + 56); in xnn_f32_velu_ukernel__avx2_rr1_p6_x80() local
62 const __m256 vz7 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx7, vprescale)); in xnn_f32_velu_ukernel__avx2_rr1_p6_x80()
210 vx7 = _mm256_mul_ps(vx7, vbeta); in xnn_f32_velu_ukernel__avx2_rr1_p6_x80()
223 const __m256 vy7 = _mm256_blendv_ps(vx7, ve7, vx7); in xnn_f32_velu_ukernel__avx2_rr1_p6_x80()
|
| D | velu-avx2-rr1-lut8-p4-perm-x80.c | 50 __m256 vx7 = _mm256_loadu_ps(x + 56); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x80() local
62 const __m256 vz7 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx7, vprescale)); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x80()
208 vx7 = _mm256_mul_ps(vx7, vbeta); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x80()
221 const __m256 vy7 = _mm256_blendv_ps(vx7, ve7, vx7); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x80()
|
| D | velu-avx512f-rr1-lut16-p3-perm-x128.c | 50 __m512 vx7 = _mm512_loadu_ps(x + 112); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128() local
60 const __m512 vz7 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx7, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
174 const __mmask16 vsign7 = _mm512_cmp_ps_mask(vx7, vzero, _CMP_NLT_US); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
183 vy7 = _mm512_mask_mul_ps(vy7, vsign7, vx7, vbeta); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | avx2-p5-x64-acc2.c | 65 const __m256 vx7 = _mm256_sub_ps(vi7, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc2() local
75 __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc2()
107 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc2()
186 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc2()
|
| D | avx2-p5-x64.c | 64 const __m256 vx7 = _mm256_sub_ps(vi7, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64() local
74 __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64()
106 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64()
185 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64()
|
| D | avx2-p5-x64-acc4.c | 67 const __m256 vx7 = _mm256_sub_ps(vi7, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc4() local
77 __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc4()
109 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc4()
188 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc4()
|
| /external/XNNPACK/src/f32-vscaleexpminusmax/gen/ |
| D | avx2-p5-x64.c | 65 const __m256 vx7 = _mm256_sub_ps(vi7, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64() local
75 __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64()
107 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64()
186 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64()
|
| D | avx2-p5-x72.c | 66 const __m256 vx7 = _mm256_sub_ps(vi7, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72() local
77 __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72()
112 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72()
199 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72()
|
| /external/XNNPACK/src/f32-raddexpminusmax/gen/ |
| D | avx2-p5-x64-acc4.c | 66 const __m256 vx7 = _mm256_sub_ps(vi7, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4() local
76 __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4()
108 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4()
187 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4()
|
| D | avx2-p5-x64.c | 63 const __m256 vx7 = _mm256_sub_ps(vi7, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64() local
73 __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64()
105 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64()
184 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64()
|
| D | avx2-p5-x64-acc2.c | 64 const __m256 vx7 = _mm256_sub_ps(vi7, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2() local
74 __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2()
106 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2()
185 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2()
|
| D | avx2-p5-x72-acc3.c | 66 const __m256 vx7 = _mm256_sub_ps(vi7, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3() local
77 __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3()
112 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3()
199 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3()
|
| D | avx2-p5-x72.c | 64 const __m256 vx7 = _mm256_sub_ps(vi7, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72() local
75 __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72()
110 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72()
197 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72()
|