/external/XNNPACK/src/f32-rmax/ |
D | avx512f.c | 33 vmax0 = _mm512_max_ps(vmax0, vx0); in xnn_f32_rmax_ukernel__avx512f() 34 vmax1 = _mm512_max_ps(vmax1, vx1); in xnn_f32_rmax_ukernel__avx512f() 35 vmax2 = _mm512_max_ps(vmax2, vx2); in xnn_f32_rmax_ukernel__avx512f() 36 vmax3 = _mm512_max_ps(vmax3, vx3); in xnn_f32_rmax_ukernel__avx512f() 38 __m512 vmax = _mm512_max_ps(_mm512_max_ps(vmax0, vmax1), _mm512_max_ps(vmax2, vmax3)); in xnn_f32_rmax_ukernel__avx512f() 41 vmax = _mm512_max_ps(vmax, vx); in xnn_f32_rmax_ukernel__avx512f()
|
/external/XNNPACK/src/f32-raddextexp/gen/ |
D | avx512f-p5-scalef-x192-acc6.c | 186 __m512 vmax_e0 = _mm512_max_ps(vacce0, vn0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() 187 __m512 vmax_e1 = _mm512_max_ps(vacce1, vn1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() 188 __m512 vmax_e2 = _mm512_max_ps(vacce2, vn2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() 189 __m512 vmax_e3 = _mm512_max_ps(vacce3, vn3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() 190 __m512 vmax_e4 = _mm512_max_ps(vacce4, vn4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() 191 __m512 vmax_e5 = _mm512_max_ps(vacce5, vn5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() 192 vmax_e0 = _mm512_max_ps(vmax_e0, vn6); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() 193 vmax_e1 = _mm512_max_ps(vmax_e1, vn7); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() 194 vmax_e2 = _mm512_max_ps(vmax_e2, vn8); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() 195 vmax_e3 = _mm512_max_ps(vmax_e3, vn9); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() [all …]
|
D | avx512f-p5-scalef-x160-acc5.c | 166 __m512 vmax_e0 = _mm512_max_ps(vacce0, vn0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() 167 __m512 vmax_e1 = _mm512_max_ps(vacce1, vn1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() 168 __m512 vmax_e2 = _mm512_max_ps(vacce2, vn2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() 169 __m512 vmax_e3 = _mm512_max_ps(vacce3, vn3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() 170 __m512 vmax_e4 = _mm512_max_ps(vacce4, vn4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() 171 vmax_e0 = _mm512_max_ps(vmax_e0, vn5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() 172 vmax_e1 = _mm512_max_ps(vmax_e1, vn6); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() 173 vmax_e2 = _mm512_max_ps(vmax_e2, vn7); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() 174 vmax_e3 = _mm512_max_ps(vmax_e3, vn8); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() 175 vmax_e4 = _mm512_max_ps(vmax_e4, vn9); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() [all …]
|
D | avx512f-p5-scalef-x192-acc3.c | 180 __m512 vmax_e0 = _mm512_max_ps(vacce0, vn0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() 181 __m512 vmax_e1 = _mm512_max_ps(vacce1, vn1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() 182 __m512 vmax_e2 = _mm512_max_ps(vacce2, vn2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() 183 vmax_e0 = _mm512_max_ps(vmax_e0, vn3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() 184 vmax_e1 = _mm512_max_ps(vmax_e1, vn4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() 185 vmax_e2 = _mm512_max_ps(vmax_e2, vn5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() 186 vmax_e0 = _mm512_max_ps(vmax_e0, vn6); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() 187 vmax_e1 = _mm512_max_ps(vmax_e1, vn7); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() 188 vmax_e2 = _mm512_max_ps(vmax_e2, vn8); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() 189 vmax_e0 = _mm512_max_ps(vmax_e0, vn9); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() [all …]
|
D | avx512f-p5-scalef-x192-acc2.c | 178 __m512 vmax_e0 = _mm512_max_ps(vacce0, vn0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() 179 __m512 vmax_e1 = _mm512_max_ps(vacce1, vn1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() 180 vmax_e0 = _mm512_max_ps(vmax_e0, vn2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() 181 vmax_e1 = _mm512_max_ps(vmax_e1, vn3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() 182 vmax_e0 = _mm512_max_ps(vmax_e0, vn4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() 183 vmax_e1 = _mm512_max_ps(vmax_e1, vn5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() 184 vmax_e0 = _mm512_max_ps(vmax_e0, vn6); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() 185 vmax_e1 = _mm512_max_ps(vmax_e1, vn7); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() 186 vmax_e0 = _mm512_max_ps(vmax_e0, vn8); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() 187 vmax_e1 = _mm512_max_ps(vmax_e1, vn9); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() [all …]
|
D | avx512f-p5-scalef-x160-acc2.c | 160 __m512 vmax_e0 = _mm512_max_ps(vacce0, vn0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() 161 __m512 vmax_e1 = _mm512_max_ps(vacce1, vn1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() 162 vmax_e0 = _mm512_max_ps(vmax_e0, vn2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() 163 vmax_e1 = _mm512_max_ps(vmax_e1, vn3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() 164 vmax_e0 = _mm512_max_ps(vmax_e0, vn4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() 165 vmax_e1 = _mm512_max_ps(vmax_e1, vn5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() 166 vmax_e0 = _mm512_max_ps(vmax_e0, vn6); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() 167 vmax_e1 = _mm512_max_ps(vmax_e1, vn7); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() 168 vmax_e0 = _mm512_max_ps(vmax_e0, vn8); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() 169 vmax_e1 = _mm512_max_ps(vmax_e1, vn9); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() [all …]
|
D | avx512f-p5-scalef-x144-acc3.c | 153 __m512 vmax_e0 = _mm512_max_ps(vacce0, vn0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() 154 __m512 vmax_e1 = _mm512_max_ps(vacce1, vn1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() 155 __m512 vmax_e2 = _mm512_max_ps(vacce2, vn2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() 156 vmax_e0 = _mm512_max_ps(vmax_e0, vn3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() 157 vmax_e1 = _mm512_max_ps(vmax_e1, vn4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() 158 vmax_e2 = _mm512_max_ps(vmax_e2, vn5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() 159 vmax_e0 = _mm512_max_ps(vmax_e0, vn6); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() 160 vmax_e1 = _mm512_max_ps(vmax_e1, vn7); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() 161 vmax_e2 = _mm512_max_ps(vmax_e2, vn8); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() 196 const __m512 vmax_acce01 = _mm512_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() [all …]
|
D | avx512f-p5-scalef-x128-acc4.c | 146 __m512 vmax_e0 = _mm512_max_ps(vacce0, vn0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() 147 __m512 vmax_e1 = _mm512_max_ps(vacce1, vn1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() 148 __m512 vmax_e2 = _mm512_max_ps(vacce2, vn2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() 149 __m512 vmax_e3 = _mm512_max_ps(vacce3, vn3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() 150 vmax_e0 = _mm512_max_ps(vmax_e0, vn4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() 151 vmax_e1 = _mm512_max_ps(vmax_e1, vn5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() 152 vmax_e2 = _mm512_max_ps(vmax_e2, vn6); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() 153 vmax_e3 = _mm512_max_ps(vmax_e3, vn7); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() 189 const __m512 vmax_acce01 = _mm512_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() 190 const __m512 vmax_acce23 = _mm512_max_ps(vacce2, vacce3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() [all …]
|
D | avx512f-p5-scalef-x192.c | 176 __m512 vmax_e0 = _mm512_max_ps(vacce0, vn0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() 177 vmax_e0 = _mm512_max_ps(vmax_e0, vn1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() 178 vmax_e0 = _mm512_max_ps(vmax_e0, vn2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() 179 vmax_e0 = _mm512_max_ps(vmax_e0, vn3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() 180 vmax_e0 = _mm512_max_ps(vmax_e0, vn4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() 181 vmax_e0 = _mm512_max_ps(vmax_e0, vn5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() 182 vmax_e0 = _mm512_max_ps(vmax_e0, vn6); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() 183 vmax_e0 = _mm512_max_ps(vmax_e0, vn7); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() 184 vmax_e0 = _mm512_max_ps(vmax_e0, vn8); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() 185 vmax_e0 = _mm512_max_ps(vmax_e0, vn9); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() [all …]
|
/external/XNNPACK/src/f32-relu/gen/ |
D | avx512f-x32.c | 37 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEF, vzero); in xnn_f32_relu_ukernel__avx512f_x32() 38 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUV, vzero); in xnn_f32_relu_ukernel__avx512f_x32() 48 vacc = _mm512_max_ps(vacc, vzero); in xnn_f32_relu_ukernel__avx512f_x32() 61 vacc = _mm512_max_ps(vacc, vzero); in xnn_f32_relu_ukernel__avx512f_x32()
|
/external/XNNPACK/src/f32-vbinary/gen/ |
D | vmaxc-avx512f-x32.c | 39 __m512 vy0123456789ABCDEF = _mm512_max_ps(va0123456789ABCDEF, vb); in xnn_f32_vmaxc_ukernel__avx512f_x32() 40 __m512 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vaGHIJKLMNOPQRSTUV, vb); in xnn_f32_vmaxc_ukernel__avx512f_x32() 52 __m512 vy = _mm512_max_ps(va, vb); in xnn_f32_vmaxc_ukernel__avx512f_x32() 65 __m512 vy = _mm512_max_ps(va, vb); in xnn_f32_vmaxc_ukernel__avx512f_x32()
|
D | vmax-avx512f-x32.c | 42 __m512 vy0123456789ABCDEF = _mm512_max_ps(va0123456789ABCDEF, vb0123456789ABCDEF); in xnn_f32_vmax_ukernel__avx512f_x32() 43 __m512 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV); in xnn_f32_vmax_ukernel__avx512f_x32() 58 __m512 vy = _mm512_max_ps(va, vb); in xnn_f32_vmax_ukernel__avx512f_x32() 72 __m512 vy = _mm512_max_ps(va, vb); in xnn_f32_vmax_ukernel__avx512f_x32()
|
D | vrdivc-minmax-avx512f-x32.c | 45 vy0123456789ABCDEF = _mm512_max_ps(vy0123456789ABCDEF, vy_min); in xnn_f32_vrdivc_minmax_ukernel__avx512f_x32() 46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vrdivc_minmax_ukernel__avx512f_x32() 60 vy = _mm512_max_ps(vy, vy_min); in xnn_f32_vrdivc_minmax_ukernel__avx512f_x32() 75 vy = _mm512_max_ps(vy, vy_min); in xnn_f32_vrdivc_minmax_ukernel__avx512f_x32()
|
D | vrsubc-minmax-avx512f-x32.c | 45 vy0123456789ABCDEF = _mm512_max_ps(vy0123456789ABCDEF, vy_min); in xnn_f32_vrsubc_minmax_ukernel__avx512f_x32() 46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vrsubc_minmax_ukernel__avx512f_x32() 60 vy = _mm512_max_ps(vy, vy_min); in xnn_f32_vrsubc_minmax_ukernel__avx512f_x32() 75 vy = _mm512_max_ps(vy, vy_min); in xnn_f32_vrsubc_minmax_ukernel__avx512f_x32()
|
D | vaddc-minmax-avx512f-x32.c | 45 vy0123456789ABCDEF = _mm512_max_ps(vy0123456789ABCDEF, vy_min); in xnn_f32_vaddc_minmax_ukernel__avx512f_x32() 46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vaddc_minmax_ukernel__avx512f_x32() 60 vy = _mm512_max_ps(vy, vy_min); in xnn_f32_vaddc_minmax_ukernel__avx512f_x32() 75 vy = _mm512_max_ps(vy, vy_min); in xnn_f32_vaddc_minmax_ukernel__avx512f_x32()
|
D | vdivc-minmax-avx512f-x32.c | 45 vy0123456789ABCDEF = _mm512_max_ps(vy0123456789ABCDEF, vy_min); in xnn_f32_vdivc_minmax_ukernel__avx512f_x32() 46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vdivc_minmax_ukernel__avx512f_x32() 60 vy = _mm512_max_ps(vy, vy_min); in xnn_f32_vdivc_minmax_ukernel__avx512f_x32() 75 vy = _mm512_max_ps(vy, vy_min); in xnn_f32_vdivc_minmax_ukernel__avx512f_x32()
|
D | vsubc-minmax-avx512f-x32.c | 45 vy0123456789ABCDEF = _mm512_max_ps(vy0123456789ABCDEF, vy_min); in xnn_f32_vsubc_minmax_ukernel__avx512f_x32() 46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vsubc_minmax_ukernel__avx512f_x32() 60 vy = _mm512_max_ps(vy, vy_min); in xnn_f32_vsubc_minmax_ukernel__avx512f_x32() 75 vy = _mm512_max_ps(vy, vy_min); in xnn_f32_vsubc_minmax_ukernel__avx512f_x32()
|
D | vmulc-minmax-avx512f-x32.c | 45 vy0123456789ABCDEF = _mm512_max_ps(vy0123456789ABCDEF, vy_min); in xnn_f32_vmulc_minmax_ukernel__avx512f_x32() 46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vmulc_minmax_ukernel__avx512f_x32() 60 vy = _mm512_max_ps(vy, vy_min); in xnn_f32_vmulc_minmax_ukernel__avx512f_x32() 75 vy = _mm512_max_ps(vy, vy_min); in xnn_f32_vmulc_minmax_ukernel__avx512f_x32()
|
/external/XNNPACK/src/f32-gemm/gen-inc/ |
D | 8x16inc-minmax-avx512f-broadcast.c | 134 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast() 135 vacc1x0123456789ABCDEF = _mm512_max_ps(vacc1x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast() 136 vacc2x0123456789ABCDEF = _mm512_max_ps(vacc2x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast() 137 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast() 138 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast() 139 vacc5x0123456789ABCDEF = _mm512_max_ps(vacc5x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast() 140 vacc6x0123456789ABCDEF = _mm512_max_ps(vacc6x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast() 141 vacc7x0123456789ABCDEF = _mm512_max_ps(vacc7x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast()
|
D | 7x16inc-minmax-avx512f-broadcast.c | 124 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast() 125 vacc1x0123456789ABCDEF = _mm512_max_ps(vacc1x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast() 126 vacc2x0123456789ABCDEF = _mm512_max_ps(vacc2x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast() 127 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast() 128 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast() 129 vacc5x0123456789ABCDEF = _mm512_max_ps(vacc5x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast() 130 vacc6x0123456789ABCDEF = _mm512_max_ps(vacc6x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast()
|
/external/XNNPACK/src/f32-gemm/gen/ |
D | 8x16-minmax-avx512f-broadcast.c | 132 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast() 133 vacc1x0123456789ABCDEF = _mm512_max_ps(vacc1x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast() 134 vacc2x0123456789ABCDEF = _mm512_max_ps(vacc2x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast() 135 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast() 136 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast() 137 vacc5x0123456789ABCDEF = _mm512_max_ps(vacc5x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast() 138 vacc6x0123456789ABCDEF = _mm512_max_ps(vacc6x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast() 139 vacc7x0123456789ABCDEF = _mm512_max_ps(vacc7x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast()
|
D | 7x16-minmax-avx512f-broadcast.c | 122 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast() 123 vacc1x0123456789ABCDEF = _mm512_max_ps(vacc1x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast() 124 vacc2x0123456789ABCDEF = _mm512_max_ps(vacc2x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast() 125 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast() 126 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast() 127 vacc5x0123456789ABCDEF = _mm512_max_ps(vacc5x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast() 128 vacc6x0123456789ABCDEF = _mm512_max_ps(vacc6x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast()
|
/external/XNNPACK/src/f32-clamp/gen/ |
D | avx512f-x32.c | 38 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEF, vy_min); in xnn_f32_clamp_ukernel__avx512f_x32() 39 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_clamp_ukernel__avx512f_x32() 52 vacc = _mm512_max_ps(vacc, vy_min); in xnn_f32_clamp_ukernel__avx512f_x32() 66 vacc = _mm512_max_ps(vacc, vy_min); in xnn_f32_clamp_ukernel__avx512f_x32()
|
/external/XNNPACK/src/f32-igemm/gen/ |
D | 8x16-minmax-avx512f-broadcast.c | 168 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast() 169 vacc1x0123456789ABCDEF = _mm512_max_ps(vacc1x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast() 170 vacc2x0123456789ABCDEF = _mm512_max_ps(vacc2x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast() 171 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast() 172 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast() 173 vacc5x0123456789ABCDEF = _mm512_max_ps(vacc5x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast() 174 vacc6x0123456789ABCDEF = _mm512_max_ps(vacc6x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast() 175 vacc7x0123456789ABCDEF = _mm512_max_ps(vacc7x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast()
|
D | 7x16-minmax-avx512f-broadcast.c | 155 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast() 156 vacc1x0123456789ABCDEF = _mm512_max_ps(vacc1x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast() 157 vacc2x0123456789ABCDEF = _mm512_max_ps(vacc2x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast() 158 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast() 159 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast() 160 vacc5x0123456789ABCDEF = _mm512_max_ps(vacc5x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast() 161 vacc6x0123456789ABCDEF = _mm512_max_ps(vacc6x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast()
|