| /external/XNNPACK/src/f32-bilinear/gen/ |
| D | neon-c8.c | 51 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); in xnn_f32_bilinear_ukernel__neon_c8()
52 const float32x4_t vbd0123 = vsubq_f32(vbr0123, vbl0123); in xnn_f32_bilinear_ukernel__neon_c8()
53 const float32x4_t vtd4567 = vsubq_f32(vtr4567, vtl4567); in xnn_f32_bilinear_ukernel__neon_c8()
54 const float32x4_t vbd4567 = vsubq_f32(vbr4567, vbl4567); in xnn_f32_bilinear_ukernel__neon_c8()
61 const float32x4_t vd0123 = vsubq_f32(vb0123, vt0123); in xnn_f32_bilinear_ukernel__neon_c8()
62 const float32x4_t vd4567 = vsubq_f32(vb4567, vt4567); in xnn_f32_bilinear_ukernel__neon_c8()
76 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); in xnn_f32_bilinear_ukernel__neon_c8()
77 const float32x4_t vbd0123 = vsubq_f32(vbr0123, vbl0123); in xnn_f32_bilinear_ukernel__neon_c8()
82 const float32x4_t vd0123 = vsubq_f32(vb0123, vt0123); in xnn_f32_bilinear_ukernel__neon_c8()
95 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); in xnn_f32_bilinear_ukernel__neon_c8()
[all …]
|
| D | neonfma-c8.c | 55 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); in xnn_f32_bilinear_ukernel__neonfma_c8()
56 const float32x4_t vbd0123 = vsubq_f32(vbr0123, vbl0123); in xnn_f32_bilinear_ukernel__neonfma_c8()
57 const float32x4_t vtd4567 = vsubq_f32(vtr4567, vtl4567); in xnn_f32_bilinear_ukernel__neonfma_c8()
58 const float32x4_t vbd4567 = vsubq_f32(vbr4567, vbl4567); in xnn_f32_bilinear_ukernel__neonfma_c8()
72 const float32x4_t vd0123 = vsubq_f32(vb0123, vt0123); in xnn_f32_bilinear_ukernel__neonfma_c8()
73 const float32x4_t vd4567 = vsubq_f32(vb4567, vt4567); in xnn_f32_bilinear_ukernel__neonfma_c8()
92 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); in xnn_f32_bilinear_ukernel__neonfma_c8()
93 const float32x4_t vbd0123 = vsubq_f32(vbr0123, vbl0123); in xnn_f32_bilinear_ukernel__neonfma_c8()
103 const float32x4_t vd0123 = vsubq_f32(vb0123, vt0123); in xnn_f32_bilinear_ukernel__neonfma_c8()
120 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); in xnn_f32_bilinear_ukernel__neonfma_c8()
[all …]
|
| D | neon-c4.c | 47 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); in xnn_f32_bilinear_ukernel__neon_c4()
48 const float32x4_t vbd0123 = vsubq_f32(vbr0123, vbl0123); in xnn_f32_bilinear_ukernel__neon_c4()
53 const float32x4_t vd0123 = vsubq_f32(vb0123, vt0123); in xnn_f32_bilinear_ukernel__neon_c4()
65 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); in xnn_f32_bilinear_ukernel__neon_c4()
66 const float32x4_t vbd0123 = vsubq_f32(vbr0123, vbl0123); in xnn_f32_bilinear_ukernel__neon_c4()
71 const float32x4_t vd0123 = vsubq_f32(vb0123, vt0123); in xnn_f32_bilinear_ukernel__neon_c4()
|
| /external/webrtc/webrtc/modules/audio_processing/aec/ |
| D | aec_rdft_neon.c | 45 const float32x4_t x1v = vsubq_f32(a01v, a23v); in cft1st_128_neon()
47 const float32x4_t x3v = vsubq_f32(a45v, a67v); in cft1st_128_neon()
51 x0v = vsubq_f32(x0v, x2v); in cft1st_128_neon()
88 const float32x4_t x1r0_1i0_1r1_x1i1 = vsubq_f32(a_00_32, a_08_40); in cftmdl_128_neon()
96 const float32x4_t x3r0_3i0_3r1_x3i1 = vsubq_f32(a_16_48, a_24_56); in cftmdl_128_neon()
98 const float32x4_t xx1 = vsubq_f32(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1); in cftmdl_128_neon()
146 const float32x4_t x1r0_1i0_1r1_x1i1 = vsubq_f32(a_00_32, a_08_40); in cftmdl_128_neon()
154 const float32x4_t x3r0_3i0_3r1_x3i1 = vsubq_f32(a_16_48, a_24_56); in cftmdl_128_neon()
156 const float32x4_t xx1 = vsubq_f32(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1); in cftmdl_128_neon()
199 const float32x4_t wkrt = vsubq_f32(mm_half, c_k1); // 28, 29, 30, 31, in rftfsub_128_neon()
[all …]
|
| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | neonfma-rr1-p5-div-x24.c | 85 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
86 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
87 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
88 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
89 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
90 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
180 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
181 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
182 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
183 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
[all …]
|
| D | neonfma-rr1-p5-div-x20.c | 81 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
82 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
83 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
84 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
85 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
164 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
165 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
166 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
167 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
168 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
[all …]
|
| D | neonfma-rr1-p5-div-x16.c | 77 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
78 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
79 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
80 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
148 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
149 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
150 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
151 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
183 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
213 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
[all …]
|
| D | neonfma-rr1-p5-nr1recps1fma-x24.c | 85 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
86 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
87 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
88 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
89 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
90 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
204 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
205 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
206 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
207 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
[all …]
|
| D | neonfma-rr1-p5-nr2recps-x24.c | 85 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
86 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
87 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
88 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
89 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
90 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
204 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
205 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
206 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
207 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
[all …]
|
| D | neonfma-rr1-p5-nr2fma-x24.c | 85 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
86 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
87 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
88 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
89 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
90 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
204 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
205 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
206 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
207 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
[all …]
|
| D | neonfma-rr1-p5-nr1recps1fma-x20.c | 81 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
82 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
83 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
84 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
85 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
185 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
186 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
187 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
188 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
189 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
[all …]
|
| D | neonfma-rr1-p5-nr2recps-x20.c | 81 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
82 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
83 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
84 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
85 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
185 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
186 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
187 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
188 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
189 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
[all …]
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | neonfma-p5-x20-acc2.c | 53 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
54 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
55 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
56 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
57 const float32x4_t vxGHIJ = vsubq_f32(viGHIJ, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
80 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
81 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
82 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
83 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
84 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
[all …]
|
| D | neonfma-p5-x20-acc5.c | 56 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
57 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
58 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
59 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
60 const float32x4_t vxGHIJ = vsubq_f32(viGHIJ, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
83 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
84 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
85 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
86 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
87 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
[all …]
|
| D | neon-p5-x20-acc5.c | 57 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
58 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
59 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
60 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
61 const float32x4_t vxGHIJ = vsubq_f32(viGHIJ, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
84 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
85 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
86 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
87 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
88 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
[all …]
|
| D | neonfma-p5-x20.c | 52 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
53 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
54 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
55 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
56 const float32x4_t vxGHIJ = vsubq_f32(viGHIJ, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
79 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
80 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
81 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
82 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
83 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
[all …]
|
| D | neon-p5-x20.c | 53 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
54 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
55 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
56 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
57 const float32x4_t vxGHIJ = vsubq_f32(viGHIJ, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
80 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
81 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
82 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
83 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
84 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
[all …]
|
| D | neon-p5-x20-acc2.c | 54 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
55 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
56 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
57 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
58 const float32x4_t vxGHIJ = vsubq_f32(viGHIJ, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
81 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
82 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
83 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
84 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
85 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
[all …]
|
| D | neonfma-p5-x16-acc4.c | 54 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
55 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
56 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
57 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
78 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
79 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
80 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
81 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
160 const float32x4_t vx = vsubq_f32(vi, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
175 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
[all …]
|
| D | neon-p5-x16-acc2.c | 53 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
54 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
55 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
56 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
77 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
78 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
79 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
80 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
157 const float32x4_t vx = vsubq_f32(vi, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
172 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
[all …]
|
| D | neon-p5-x16.c | 52 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
53 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
54 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
55 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
76 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
77 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
78 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
79 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
154 const float32x4_t vx = vsubq_f32(vi, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
169 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
[all …]
|
| D | neonfma-p5-x16.c | 51 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
52 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
53 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
54 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
75 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
76 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
77 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
78 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
153 const float32x4_t vx = vsubq_f32(vi, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
168 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
[all …]
|
| D | neonfma-p5-x16-acc2.c | 52 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
53 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
54 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
55 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
76 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
77 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
78 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
79 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
156 const float32x4_t vx = vsubq_f32(vi, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
171 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
[all …]
|
| D | neon-p5-x16-acc4.c | 55 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
56 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
57 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
58 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
79 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
80 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
81 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
82 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
161 const float32x4_t vx = vsubq_f32(vi, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
176 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
[all …]
|
| D | neon-p5-x12.c | 51 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
52 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
53 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
72 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
73 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
74 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
138 const float32x4_t vx = vsubq_f32(vi, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
153 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
195 const float32x4_t vx = vsubq_f32(vi, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
210 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
|