Lines Matching refs:vxa3
104 const int16x8_t vxa3 = vmovl_s8(va3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm() local
115 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
116 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
126 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
127 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
137 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
138 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
148 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
149 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
159 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
160 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
170 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
171 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
181 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
182 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
192 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
193 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
207 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
208 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
218 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
219 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
229 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
230 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
240 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
241 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
251 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
252 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
262 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
263 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
273 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
274 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
284 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
285 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
297 const int16x8_t vxa3 = vmovl_s8(va3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm() local
316 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
317 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
318 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
319 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
339 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
340 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
341 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
342 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
362 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
363 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
364 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
365 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
385 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
386 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
387 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
388 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
408 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
409 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
410 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
411 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
431 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
432 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
433 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
434 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
454 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
455 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
456 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()
457 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm()