| /external/XNNPACK/src/f32-vscaleexpminusmax/gen/ |
| D | avx2-p5-x56.c | 101 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56() local
109 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56()
118 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56()
126 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56()
134 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56()
142 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56()
154 vt6 = _mm256_mul_ps(vt6, vs6); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56()
162 __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56()
|
| D | avx512f-p5-scalef-x112.c | 78 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112() local
86 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112()
95 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112()
103 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112()
111 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112()
119 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112()
127 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112()
|
| D | avx2-p5-x64.c | 106 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64() local
115 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64()
125 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64()
134 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64()
143 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64()
152 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64()
165 vt6 = _mm256_mul_ps(vt6, vs6); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64()
174 __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64()
|
| D | avx512f-p5-scalef-x128.c | 81 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128() local
90 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128()
100 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128()
109 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128()
118 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128()
127 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128()
136 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128()
|
| D | avx2-p5-x72.c | 111 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72() local
121 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72()
132 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72()
142 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72()
152 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72()
162 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72()
176 vt6 = _mm256_mul_ps(vt6, vs6); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72()
186 __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72()
|
| D | avx512f-p5-scalef-x144.c | 84 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144() local
94 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
105 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
115 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
125 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
135 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
145 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
|
| /external/XNNPACK/src/f32-vscaleextexp/gen/ |
| D | avx512f-p5-scalef-x112.c | 70 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112() local
78 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
87 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
95 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
103 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
111 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
119 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
|
| D | avx512f-p5-scalef-x128.c | 72 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128() local
81 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128()
91 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128()
100 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128()
109 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128()
118 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128()
127 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128()
|
| D | avx512f-p5-scalef-x144.c | 74 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144() local
84 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
95 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
105 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
115 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
125 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
135 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
|
| D | avx2-p5-x56.c | 76 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x56() local
84 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x56()
93 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x56()
101 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x56()
109 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x56()
117 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x56()
125 vp6 = _mm256_fmadd_ps(vp6, vt6, vc0); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x56()
|
| /external/XNNPACK/src/f32-raddexpminusmax/gen/ |
| D | avx2-p5-x64-acc2.c | 105 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2() local
114 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2()
124 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2()
133 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2()
142 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2()
151 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2()
164 vt6 = _mm256_mul_ps(vt6, vs6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2()
173 __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2()
|
| D | avx2-p5-x64.c | 104 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64() local
113 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64()
123 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64()
132 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64()
141 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64()
150 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64()
163 vt6 = _mm256_mul_ps(vt6, vs6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64()
172 __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64()
|
| D | avx2-p5-x64-acc4.c | 107 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4() local
116 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4()
126 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4()
135 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4()
144 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4()
153 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4()
166 vt6 = _mm256_mul_ps(vt6, vs6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4()
175 __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4()
|
| D | avx512f-p5-scalef-x128-acc2.c | 81 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2() local
90 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2()
100 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2()
109 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2()
118 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2()
127 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2()
136 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2()
|
| D | avx512f-p5-scalef-x128-acc4.c | 83 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc4() local
92 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc4()
102 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc4()
111 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc4()
120 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc4()
129 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc4()
138 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc4()
|
| D | avx512f-p5-scalef-x128.c | 80 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128() local
89 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128()
99 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128()
108 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128()
117 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128()
126 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128()
135 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128()
|
| D | avx2-p5-x72-acc3.c | 111 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3() local
121 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3()
132 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3()
142 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3()
152 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3()
162 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3()
176 vt6 = _mm256_mul_ps(vt6, vs6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3()
186 __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72_acc3()
|
| D | avx2-p5-x72.c | 109 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72() local
119 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72()
130 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72()
140 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72()
150 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72()
160 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72()
174 vt6 = _mm256_mul_ps(vt6, vs6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72()
184 __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72()
|
| D | avx512f-p5-scalef-x144-acc3.c | 85 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3() local
95 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
106 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
116 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
126 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
136 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
146 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
|
| D | avx512f-p5-scalef-x144.c | 83 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144() local
93 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
104 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
114 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
124 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
134 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
144 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
|
| D | avx2-p5-x80-acc2.c | 115 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2() local
126 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
138 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
149 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
160 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
171 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
186 vt6 = _mm256_mul_ps(vt6, vs6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
197 __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
|
| /external/XNNPACK/src/f32-velu/gen/ |
| D | velu-avx2-rr1-p6-x56.c | 86 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_velu_ukernel__avx2_rr1_p6_x56() local
94 __m256 vp6 = _mm256_fmadd_ps(vc6, vt6, vc5); in xnn_f32_velu_ukernel__avx2_rr1_p6_x56()
102 vp6 = _mm256_fmadd_ps(vp6, vt6, vc4); in xnn_f32_velu_ukernel__avx2_rr1_p6_x56()
110 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3); in xnn_f32_velu_ukernel__avx2_rr1_p6_x56()
118 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2); in xnn_f32_velu_ukernel__avx2_rr1_p6_x56()
132 vp6 = _mm256_mul_ps(vp6, vt6); in xnn_f32_velu_ukernel__avx2_rr1_p6_x56()
133 vt6 = _mm256_mul_ps(vt6, vs6); in xnn_f32_velu_ukernel__avx2_rr1_p6_x56()
148 vp6 = _mm256_fmadd_ps(vp6, vt6, vt6); in xnn_f32_velu_ukernel__avx2_rr1_p6_x56()
|
| D | velu-avx512f-rr1-p6-x112.c | 88 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112() local
96 __m512 vp6 = _mm512_fmadd_ps(vc6, vt6, vc5); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
104 vp6 = _mm512_fmadd_ps(vp6, vt6, vc4); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
112 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
120 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
134 vp6 = _mm512_mul_ps(vp6, vt6); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
135 vt6 = _mm512_mul_ps(vt6, vs6); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
151 vp6 = _mm512_fmadd_ps(vp6, vt6, vt6); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
|
| /external/XNNPACK/src/f32-raddextexp/gen/ |
| D | avx512f-p5-scalef-x128.c | 72 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128() local
81 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
91 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
100 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
109 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
118 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
127 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
|
| D | avx512f-p5-scalef-x128-acc2.c | 74 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2() local
83 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
93 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
102 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
111 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
120 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
129 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
|