83     q31x4_t      vecB, vecA0, vecA1;  in arm_mat_mult_opt_q31_2x2_mve()  local
93     vecB = vldrwq_gather_shifted_offset_z_s32(pInB, vecColBOffs, p0);  in arm_mat_mult_opt_q31_2x2_mve()
99     acc0 = vrmlaldavhq(vecA0, vecB);  in arm_mat_mult_opt_q31_2x2_mve()
100     acc1 = vrmlaldavhq(vecA1, vecB);  in arm_mat_mult_opt_q31_2x2_mve()
112     vecB = vldrwq_gather_shifted_offset_z_s32(pInB, vecColBOffs, p0);  in arm_mat_mult_opt_q31_2x2_mve()
114     acc0 = vrmlaldavhq(vecA0, vecB);  in arm_mat_mult_opt_q31_2x2_mve()
115     acc1 = vrmlaldavhq(vecA1, vecB);  in arm_mat_mult_opt_q31_2x2_mve()
143     q31x4_t      vecB, vecA;  in arm_mat_mult_opt_q31_3x3_mve()  local
152     vecB = vldrwq_gather_shifted_offset_z_s32(pInB, vecColBOffs, p0);  in arm_mat_mult_opt_q31_3x3_mve()
155     acc0 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_3x3_mve()
157     acc1 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_3x3_mve()
159     acc2 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_3x3_mve()
173     vecB = vldrwq_gather_shifted_offset_z_s32(pInB, vecColBOffs, p0);  in arm_mat_mult_opt_q31_3x3_mve()
176     acc0 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_3x3_mve()
178     acc1 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_3x3_mve()
180     acc2 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_3x3_mve()
194     vecB = vldrwq_gather_shifted_offset_z_s32(pInB, vecColBOffs, p0);  in arm_mat_mult_opt_q31_3x3_mve()
197     acc0 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_3x3_mve()
199     acc1 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_3x3_mve()
201     acc2 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_3x3_mve()
230     q31x4_t      vecB, vecA;  in arm_mat_mult_opt_q31_4x4_mve()  local
236     vecB = vldrwq_gather_shifted_offset_s32(pInB, vecColBOffs);  in arm_mat_mult_opt_q31_4x4_mve()
239     acc0 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_4x4_mve()
241     acc1 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_4x4_mve()
243     acc2 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_4x4_mve()
245     acc3 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_4x4_mve()
261     vecB = vldrwq_gather_shifted_offset_s32(pInB, vecColBOffs);  in arm_mat_mult_opt_q31_4x4_mve()
264     acc0 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_4x4_mve()
266     acc1 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_4x4_mve()
268     acc2 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_4x4_mve()
270     acc3 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_4x4_mve()
287     vecB = vldrwq_gather_shifted_offset_s32(pInB, vecColBOffs);  in arm_mat_mult_opt_q31_4x4_mve()
290     acc0 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_4x4_mve()
292     acc1 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_4x4_mve()
294     acc2 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_4x4_mve()
296     acc3 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_4x4_mve()
313     vecB = vldrwq_gather_shifted_offset_s32(pInB, vecColBOffs);  in arm_mat_mult_opt_q31_4x4_mve()
316     acc0 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_4x4_mve()
318     acc1 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_4x4_mve()
320     acc2 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_4x4_mve()
322     acc3 = vrmlaldavhq(vecA, vecB);  in arm_mat_mult_opt_q31_4x4_mve()
425                 q31x4_t         vecA, vecA2, vecB, vecB2;  in arm_mat_mult_opt_q31()  local
455                     vecB = vld1q(pSrcBVec);  in arm_mat_mult_opt_q31()
457                     acc0 = vrmlaldavhaq(acc0, vecA, vecB);  in arm_mat_mult_opt_q31()
460                     acc1 = vrmlaldavhaq(acc1, vecA2, vecB);  in arm_mat_mult_opt_q31()
477                     vecB = vld1q(pSrcBVec);  in arm_mat_mult_opt_q31()
478                     acc0 = vrmlaldavhaq_p(acc0, vecA, vecB, p0);  in arm_mat_mult_opt_q31()
480                     acc1 = vrmlaldavhaq_p(acc1, vecA2, vecB, p0);  in arm_mat_mult_opt_q31()
529                 q31x4_t         vecA, vecB;  in arm_mat_mult_opt_q31()  local
547                     vecB = vld1q(pSrcBVec);  in arm_mat_mult_opt_q31()
549                     acc0 = vrmlaldavhaq(acc0, vecA, vecB);  in arm_mat_mult_opt_q31()
561                     vecB = vld1q(pSrcBVec);  in arm_mat_mult_opt_q31()
562                     acc0 = vrmlaldavhaq_p(acc0, vecA, vecB, p0);  in arm_mat_mult_opt_q31()
590                 q31x4_t         vecA, vecB;  in arm_mat_mult_opt_q31()  local
610                     vecB = vld1q(pSrcBVec);  in arm_mat_mult_opt_q31()
612                     acc0 = vrmlaldavhaq(acc0, vecA, vecB);  in arm_mat_mult_opt_q31()
624                     vecB = vld1q(pSrcBVec);  in arm_mat_mult_opt_q31()
625                     acc0 = vrmlaldavhaq_p(acc0, vecA, vecB, p0);  in arm_mat_mult_opt_q31()