| /external/XNNPACK/src/f32-ibilinear/gen/ |
| D | neon-c8.c | 51 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); in xnn_f32_ibilinear_ukernel__neon_c8()
52 const float32x4_t vbd0123 = vsubq_f32(vbr0123, vbl0123); in xnn_f32_ibilinear_ukernel__neon_c8()
53 const float32x4_t vtd4567 = vsubq_f32(vtr4567, vtl4567); in xnn_f32_ibilinear_ukernel__neon_c8()
54 const float32x4_t vbd4567 = vsubq_f32(vbr4567, vbl4567); in xnn_f32_ibilinear_ukernel__neon_c8()
61 const float32x4_t vd0123 = vsubq_f32(vb0123, vt0123); in xnn_f32_ibilinear_ukernel__neon_c8()
62 const float32x4_t vd4567 = vsubq_f32(vb4567, vt4567); in xnn_f32_ibilinear_ukernel__neon_c8()
76 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); in xnn_f32_ibilinear_ukernel__neon_c8()
77 const float32x4_t vbd0123 = vsubq_f32(vbr0123, vbl0123); in xnn_f32_ibilinear_ukernel__neon_c8()
82 const float32x4_t vd0123 = vsubq_f32(vb0123, vt0123); in xnn_f32_ibilinear_ukernel__neon_c8()
95 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); in xnn_f32_ibilinear_ukernel__neon_c8()
[all …]
|
| D | neonfma-c8.c | 55 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); in xnn_f32_ibilinear_ukernel__neonfma_c8()
56 const float32x4_t vbd0123 = vsubq_f32(vbr0123, vbl0123); in xnn_f32_ibilinear_ukernel__neonfma_c8()
57 const float32x4_t vtd4567 = vsubq_f32(vtr4567, vtl4567); in xnn_f32_ibilinear_ukernel__neonfma_c8()
58 const float32x4_t vbd4567 = vsubq_f32(vbr4567, vbl4567); in xnn_f32_ibilinear_ukernel__neonfma_c8()
72 const float32x4_t vd0123 = vsubq_f32(vb0123, vt0123); in xnn_f32_ibilinear_ukernel__neonfma_c8()
73 const float32x4_t vd4567 = vsubq_f32(vb4567, vt4567); in xnn_f32_ibilinear_ukernel__neonfma_c8()
92 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); in xnn_f32_ibilinear_ukernel__neonfma_c8()
93 const float32x4_t vbd0123 = vsubq_f32(vbr0123, vbl0123); in xnn_f32_ibilinear_ukernel__neonfma_c8()
103 const float32x4_t vd0123 = vsubq_f32(vb0123, vt0123); in xnn_f32_ibilinear_ukernel__neonfma_c8()
120 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); in xnn_f32_ibilinear_ukernel__neonfma_c8()
[all …]
|
| D | neon-c4.c | 47 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); in xnn_f32_ibilinear_ukernel__neon_c4()
48 const float32x4_t vbd0123 = vsubq_f32(vbr0123, vbl0123); in xnn_f32_ibilinear_ukernel__neon_c4()
53 const float32x4_t vd0123 = vsubq_f32(vb0123, vt0123); in xnn_f32_ibilinear_ukernel__neon_c4()
65 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); in xnn_f32_ibilinear_ukernel__neon_c4()
66 const float32x4_t vbd0123 = vsubq_f32(vbr0123, vbl0123); in xnn_f32_ibilinear_ukernel__neon_c4()
71 const float32x4_t vd0123 = vsubq_f32(vb0123, vt0123); in xnn_f32_ibilinear_ukernel__neon_c4()
|
| /external/ComputeLibrary/src/cpu/kernels/sub/neon/ |
| D | qsymm16.cpp | 111 …vcvtnq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[0], af.val[0]) : vsubq_f32(af.val… in sub_qsymm16_neon()
112 …vcvtnq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[1], af.val[1]) : vsubq_f32(af.val… in sub_qsymm16_neon()
114 …vcvtq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[0], af.val[0]) : vsubq_f32(af.val[… in sub_qsymm16_neon()
115 …vcvtq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[1], af.val[1]) : vsubq_f32(af.val[… in sub_qsymm16_neon()
176 vcvtnq_s32_f32(vmulq_f32(vsubq_f32(af.val[0], bf.val[0]), invvscaleo)), in sub_qsymm16_neon()
177 vcvtnq_s32_f32(vmulq_f32(vsubq_f32(af.val[1], bf.val[1]), invvscaleo)), in sub_qsymm16_neon()
179 vcvtq_s32_f32(vmulq_f32(vsubq_f32(af.val[0], bf.val[0]), invvscaleo)), in sub_qsymm16_neon()
180 vcvtq_s32_f32(vmulq_f32(vsubq_f32(af.val[1], bf.val[1]), invvscaleo)), in sub_qsymm16_neon()
|
| /external/webrtc/common_audio/third_party/ooura/fft_size_128/ |
| D | ooura_fft_neon.cc | 46 const float32x4_t x1v = vsubq_f32(a01v, a23v); in cft1st_128_neon()
48 const float32x4_t x3v = vsubq_f32(a45v, a67v); in cft1st_128_neon()
52 x0v = vsubq_f32(x0v, x2v); in cft1st_128_neon()
89 const float32x4_t x1r0_1i0_1r1_x1i1 = vsubq_f32(a_00_32, a_08_40); in cftmdl_128_neon()
97 const float32x4_t x3r0_3i0_3r1_x3i1 = vsubq_f32(a_16_48, a_24_56); in cftmdl_128_neon()
99 const float32x4_t xx1 = vsubq_f32(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1); in cftmdl_128_neon()
147 const float32x4_t x1r0_1i0_1r1_x1i1 = vsubq_f32(a_00_32, a_08_40); in cftmdl_128_neon()
155 const float32x4_t x3r0_3i0_3r1_x3i1 = vsubq_f32(a_16_48, a_24_56); in cftmdl_128_neon()
157 const float32x4_t xx1 = vsubq_f32(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1); in cftmdl_128_neon()
200 const float32x4_t wkrt = vsubq_f32(mm_half, c_k1); // 28, 29, 30, 31, in rftfsub_128_neon()
[all …]
|
| /external/XNNPACK/src/f32-vsigmoid/gen/ |
| D | vsigmoid-neonfma-rr1-p5-div-x24.c | 66 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x24()
67 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x24()
68 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x24()
69 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x24()
70 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x24()
71 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x24()
150 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x24()
151 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x24()
152 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x24()
153 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x24()
[all …]
|
| D | vsigmoid-neonfma-rr1-p5-div-x20.c | 62 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()
63 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()
64 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()
65 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()
66 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()
134 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()
135 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()
136 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()
137 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()
138 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()
[all …]
|
| D | vsigmoid-neonfma-rr1-p5-div-x16.c | 58 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x16()
59 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x16()
60 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x16()
61 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x16()
118 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x16()
119 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x16()
120 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x16()
121 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x16()
135 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x16()
150 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x16()
[all …]
|
| D | vsigmoid-neonfma-rr1-p5-nr2recps-x24.c | 66 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
67 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
68 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
69 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
70 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
71 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
171 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
172 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
173 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
174 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
[all …]
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | neonfma-rr1-p5-x20-acc5.c | 51 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc5()
52 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc5()
53 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc5()
54 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc5()
55 const float32x4_t vxGHIJ = vsubq_f32(viGHIJ, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc5()
69 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc5()
70 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc5()
71 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc5()
72 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc5()
73 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc5()
[all …]
|
| D | neonfma-rr1-p5-x20.c | 47 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20()
48 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20()
49 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20()
50 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20()
51 const float32x4_t vxGHIJ = vsubq_f32(viGHIJ, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20()
65 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20()
66 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20()
67 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20()
68 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20()
69 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20()
[all …]
|
| D | neon-rr2-p5-x20.c | 48 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20()
49 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20()
50 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20()
51 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20()
52 const float32x4_t vxGHIJ = vsubq_f32(viGHIJ, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20()
66 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20()
67 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20()
68 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20()
69 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20()
70 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20()
[all …]
|
| D | neonfma-rr1-p5-x20-acc2.c | 48 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc2()
49 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc2()
50 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc2()
51 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc2()
52 const float32x4_t vxGHIJ = vsubq_f32(viGHIJ, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc2()
66 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc2()
67 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc2()
68 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc2()
69 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc2()
70 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc2()
[all …]
|
| D | neon-rr2-p5-x20-acc5.c | 52 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc5()
53 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc5()
54 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc5()
55 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc5()
56 const float32x4_t vxGHIJ = vsubq_f32(viGHIJ, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc5()
70 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc5()
71 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc5()
72 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc5()
73 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc5()
74 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc5()
[all …]
|
| D | neon-rr2-p5-x20-acc2.c | 49 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc2()
50 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc2()
51 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc2()
52 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc2()
53 const float32x4_t vxGHIJ = vsubq_f32(viGHIJ, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc2()
67 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc2()
68 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc2()
69 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc2()
70 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc2()
71 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc2()
[all …]
|
| D | neon-rr2-p5-x16.c | 47 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16()
48 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16()
49 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16()
50 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16()
62 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16()
63 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16()
64 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16()
65 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16()
127 const float32x4_t vx = vsubq_f32(vi, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16()
133 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16()
[all …]
|
| D | neonfma-rr1-p5-x16-acc4.c | 49 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc4()
50 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc4()
51 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc4()
52 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc4()
64 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc4()
65 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc4()
66 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc4()
67 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc4()
127 const float32x4_t vx = vsubq_f32(vi, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc4()
133 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc4()
[all …]
|
| D | neon-rr2-p5-x16-acc4.c | 50 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc4()
51 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc4()
52 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc4()
53 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc4()
65 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc4()
66 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc4()
67 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc4()
68 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc4()
133 const float32x4_t vx = vsubq_f32(vi, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc4()
139 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc4()
[all …]
|
| D | neonfma-rr1-p5-x16.c | 46 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16()
47 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16()
48 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16()
49 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16()
61 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16()
62 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16()
63 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16()
64 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16()
121 const float32x4_t vx = vsubq_f32(vi, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16()
127 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16()
[all …]
|
| D | neonfma-rr1-p5-x16-acc2.c | 47 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc2()
48 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc2()
49 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc2()
50 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc2()
62 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc2()
63 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc2()
64 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc2()
65 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc2()
123 const float32x4_t vx = vsubq_f32(vi, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc2()
129 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc2()
[all …]
|
| D | neon-rr2-p5-x16-acc2.c | 48 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc2()
49 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc2()
50 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc2()
51 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc2()
63 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc2()
64 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc2()
65 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc2()
66 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc2()
129 const float32x4_t vx = vsubq_f32(vi, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc2()
135 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc2()
[all …]
|
| /external/XNNPACK/src/f32-velu/gen/ |
| D | velu-neonfma-rr1-p6-x24.c | 66 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
68 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
70 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
72 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
74 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
76 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
121 vs0123 = vsubq_f32(vs0123, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
123 vs4567 = vsubq_f32(vs4567, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
125 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
127 vsCDEF = vsubq_f32(vsCDEF, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
[all …]
|
| D | velu-neon-rr2-p6-x24.c | 67 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
69 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
71 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
73 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
75 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
77 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
129 vs0123 = vsubq_f32(vs0123, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
131 vs4567 = vsubq_f32(vs4567, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
133 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
135 vsCDEF = vsubq_f32(vsCDEF, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
[all …]
|
| D | velu-neonfma-rr1-p6-x20.c | 63 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
65 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
67 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
69 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
71 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
110 vs0123 = vsubq_f32(vs0123, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
112 vs4567 = vsubq_f32(vs4567, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
114 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
116 vsCDEF = vsubq_f32(vsCDEF, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
118 vsGHIJ = vsubq_f32(vsGHIJ, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
[all …]
|
| D | velu-neon-rr2-p6-x20.c | 64 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
66 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
68 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
70 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
72 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
117 vs0123 = vsubq_f32(vs0123, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
119 vs4567 = vsubq_f32(vs4567, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
121 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
123 vsCDEF = vsubq_f32(vsCDEF, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
125 vsGHIJ = vsubq_f32(vsGHIJ, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
[all …]
|