Lines Matching refs:vxa3
104 const int16x8_t vxa3 = vmovl_s8(va3); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane() local
115 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
116 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
126 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
127 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
137 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
138 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
148 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
149 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
159 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
160 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
170 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
171 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
181 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
182 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
192 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
193 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
205 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
206 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
216 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
217 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
227 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
228 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
238 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
239 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
249 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
250 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
260 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
261 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
271 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
272 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
282 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
283 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
295 const int16x8_t vxa3 = vmovl_s8(va3); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane() local
314 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
315 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
316 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
317 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
337 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
338 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
339 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
340 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
360 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
361 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
362 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
363 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
383 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
384 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
385 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
386 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
406 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
407 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
408 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
409 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
429 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
430 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
431 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
432 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
452 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
453 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
454 … vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
455 … vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()