55 …int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeo…  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
56 …int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeo…  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
57 …int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeo…  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
58 …int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeo…  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
59 …int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeo…  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
60 …int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeo…  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
61 …int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeo…  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
62 …int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeo…  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
63 …int32x4_t vacc0x8 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeo…  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
64 …int32x4_t vacc0x9 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeo…  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
65 …int32x4_t vacc0x10 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + size…  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
66 …int32x4_t vacc0x11 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + size…  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
67 …int32x4_t vacc0x12 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + size…  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
68 …int32x4_t vacc0x13 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + size…  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
69 …int32x4_t vacc0x14 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + size…  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
70 …int32x4_t vacc0x15 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + size…  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
71     int32x4_t vacc1x0 = vacc0x0;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
72     int32x4_t vacc1x1 = vacc0x1;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
73     int32x4_t vacc1x2 = vacc0x2;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
74     int32x4_t vacc1x3 = vacc0x3;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
75     int32x4_t vacc1x4 = vacc0x4;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
76     int32x4_t vacc1x5 = vacc0x5;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
77     int32x4_t vacc1x6 = vacc0x6;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
78     int32x4_t vacc1x7 = vacc0x7;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
79     int32x4_t vacc1x8 = vacc0x8;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
80     int32x4_t vacc1x9 = vacc0x9;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
81     int32x4_t vacc1x10 = vacc0x10;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
82     int32x4_t vacc1x11 = vacc0x11;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
83     int32x4_t vacc1x12 = vacc0x12;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
84     int32x4_t vacc1x13 = vacc0x13;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
85     int32x4_t vacc1x14 = vacc0x14;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
86     int32x4_t vacc1x15 = vacc0x15;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
87     int32x4_t vacc2x0 = vacc0x0;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
88     int32x4_t vacc2x1 = vacc0x1;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
89     int32x4_t vacc2x2 = vacc0x2;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
90     int32x4_t vacc2x3 = vacc0x3;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
91     int32x4_t vacc2x4 = vacc0x4;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
92     int32x4_t vacc2x5 = vacc0x5;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
93     int32x4_t vacc2x6 = vacc0x6;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
94     int32x4_t vacc2x7 = vacc0x7;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
95     int32x4_t vacc2x8 = vacc0x8;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
96     int32x4_t vacc2x9 = vacc0x9;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
97     int32x4_t vacc2x10 = vacc0x10;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
98     int32x4_t vacc2x11 = vacc0x11;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
99     int32x4_t vacc2x12 = vacc0x12;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
100     int32x4_t vacc2x13 = vacc0x13;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
101     int32x4_t vacc2x14 = vacc0x14;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
102     int32x4_t vacc2x15 = vacc0x15;  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
296     const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
297     const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
298     const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
299     const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
300     const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
301     const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
302     const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
303     const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
304     const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
305     const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
306     const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
307     const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
308     const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
309     const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
310     const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
311     const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, vacc1x15);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
312     const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
313     const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
314     const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
315     const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
316     const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
317     const int32x4_t vsum2xAB = vpaddq_s32(vacc2x10, vacc2x11);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
318     const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
319     const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
320     int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
321     int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
322     int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
323     int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
324     int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
325     int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
326     int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
327     int32x4_t vacc1xCDEF = vpaddq_s32(vsum1xCD, vsum1xEF);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
328     int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
329     int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
330     int32x4_t vacc2x89AB = vpaddq_s32(vsum2x89, vsum2xAB);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
331     int32x4_t vacc2xCDEF = vpaddq_s32(vsum2xCD, vsum2xEF);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
339     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
346     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
353     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB );  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
360     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF );  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
367     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
374     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
381     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB );  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
388     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF );  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
395     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23 );  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
402     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67 );  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
409     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB );  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
416     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF );  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
419     const int32x4_t vmultiplier = vld1q_dup_s32(&params->neon.multiplier);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
433     const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
434 … const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));  in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()