Lines Matching refs:vmlaq_f32
66 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vx0123, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
67 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
68 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vx89AB, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
69 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vxCDEF, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
70 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vxGHIJ, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
89 float32x4_t vt0123 = vmlaq_f32(vx0123, vn0123, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
90 float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
91 float32x4_t vt89AB = vmlaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
92 float32x4_t vtCDEF = vmlaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
93 float32x4_t vtGHIJ = vmlaq_f32(vxGHIJ, vnGHIJ, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
95 vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
96 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
97 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
98 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
99 vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
102 float32x4_t vp0123 = vmlaq_f32(vc4, vc5, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
103 float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
104 float32x4_t vp89AB = vmlaq_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
105 float32x4_t vpCDEF = vmlaq_f32(vc4, vc5, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
106 float32x4_t vpGHIJ = vmlaq_f32(vc4, vc5, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
108 vp0123 = vmlaq_f32(vc3, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
109 vp4567 = vmlaq_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
110 vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
111 vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
112 vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
114 vp0123 = vmlaq_f32(vc2, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
115 vp4567 = vmlaq_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
116 vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
117 vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
118 vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
120 vp0123 = vmlaq_f32(vc1, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
121 vp4567 = vmlaq_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
122 vp89AB = vmlaq_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
123 vpCDEF = vmlaq_f32(vc1, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
124 vpGHIJ = vmlaq_f32(vc1, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
136 float32x4_t vf0123 = vmlaq_f32(vs0123, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
137 float32x4_t vf4567 = vmlaq_f32(vs4567, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
138 float32x4_t vf89AB = vmlaq_f32(vs89AB, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
139 float32x4_t vfCDEF = vmlaq_f32(vsCDEF, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
140 float32x4_t vfGHIJ = vmlaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
181 float32x4_t vn = vmlaq_f32(vmagic_bias, vx, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
192 float32x4_t vt = vmlaq_f32(vx, vn, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
193 vt = vmlaq_f32(vt, vn, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
196 float32x4_t vp = vmlaq_f32(vc4, vc5, vt); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
197 vp = vmlaq_f32(vc3, vp, vt); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
198 vp = vmlaq_f32(vc2, vp, vt); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
199 vp = vmlaq_f32(vc1, vp, vt); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
206 float32x4_t vf = vmlaq_f32(vs, vp, vt); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
238 float32x4_t vn = vmlaq_f32(vmagic_bias, vx, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
249 float32x4_t vt = vmlaq_f32(vx, vn, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
250 vt = vmlaq_f32(vt, vn, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
253 float32x4_t vp = vmlaq_f32(vc4, vc5, vt); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
254 vp = vmlaq_f32(vc3, vp, vt); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
255 vp = vmlaq_f32(vc2, vp, vt); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
256 vp = vmlaq_f32(vc1, vp, vt); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
263 float32x4_t vf = vmlaq_f32(vs, vp, vt); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()