Home
last modified time | relevance | path

Searched refs:vsubq_f32 (Results 1 – 25 of 254) sorted by relevance

1234567891011

/external/XNNPACK/src/f32-ibilinear/gen/
Dneon-c8.c51 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 …]
Dneonfma-c8.c55 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 …]
Dneon-c4.c47 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/webrtc/common_audio/third_party/ooura/fft_size_128/
Dooura_fft_neon.cc46 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-raddstoreexpminusmax/gen/
Dneonfma-p5-x20.c52 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 …]
Dneon-p5-x20-acc5.c57 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 …]
Dneonfma-p5-x20-acc5.c56 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 …]
Dneon-p5-x20.c53 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 …]
Dneon-p5-x20-acc2.c54 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 …]
Dneonfma-p5-x20-acc2.c53 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 …]
Dneon-p5-x16.c52 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 …]
Dneonfma-p5-x16.c51 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 …]
Dneonfma-p5-x16-acc2.c52 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 …]
Dneon-p5-x16-acc2.c53 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 …]
Dneonfma-p5-x16-acc4.c54 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 …]
Dneon-p5-x16-acc4.c55 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 …]
/external/XNNPACK/src/f32-sigmoid/gen/
Dneonfma-rr1-p5-div-x24.c66 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
67 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
68 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
69 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
70 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
71 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
150 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
151 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
152 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
153 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
[all …]
Dneonfma-rr1-p5-div-x20.c62 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
63 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
64 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
65 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
66 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
134 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
135 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
136 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
137 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
138 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
[all …]
Dneonfma-rr1-p5-div-x16.c58 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
59 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
60 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
61 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
118 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
119 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
120 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
121 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
135 vn = vsubq_f32(vn, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
150 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
[all …]
Dneonfma-rr1-p5-nr2recps-x24.c66 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
67 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
68 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
69 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
70 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
71 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
171 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
172 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
173 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
174 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
[all …]
Dneonfma-rr1-p5-nr1recps1fma-x24.c66 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
67 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
68 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
69 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
70 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
71 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
171 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
172 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
173 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
174 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
[all …]
/external/XNNPACK/src/f32-velu/gen/
Dvelu-neon-rr2-p6-x24.c68 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
70 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
72 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
74 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
76 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
78 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
130 vs0123 = vsubq_f32(vs0123, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
132 vs4567 = vsubq_f32(vs4567, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
134 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
136 vsCDEF = vsubq_f32(vsCDEF, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
[all …]
Dvelu-neonfma-rr1-p6-x24.c67 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
69 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
71 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
73 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
75 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
77 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
122 vs0123 = vsubq_f32(vs0123, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
124 vs4567 = vsubq_f32(vs4567, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
126 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
128 vsCDEF = vsubq_f32(vsCDEF, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
[all …]
Dvelu-neonfma-rr1-p6-x20.c64 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
66 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
68 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
70 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
72 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
111 vs0123 = vsubq_f32(vs0123, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
113 vs4567 = vsubq_f32(vs4567, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
115 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
117 vsCDEF = vsubq_f32(vsCDEF, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
119 vsGHIJ = vsubq_f32(vsGHIJ, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
[all …]
Dvelu-neon-rr2-p6-x20.c65 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
67 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
69 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
71 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
73 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
118 vs0123 = vsubq_f32(vs0123, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
120 vs4567 = vsubq_f32(vs4567, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
122 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
124 vsCDEF = vsubq_f32(vsCDEF, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
126 vsGHIJ = vsubq_f32(vsGHIJ, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
[all …]

1234567891011