xref: /aosp_15_r20/external/XNNPACK/src/qu8-gemm/gen/4x16-minmax-fp32-neon-mlal-lane.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/neon-mlal-lane.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/gemm.h>
16 
17 
xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const uint8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     uint8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(mr != 0);
31   assert(mr <= 4);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(uint8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   const uint8_t* a0 = a;
40   uint8_t* c0 = c;
41   const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
42   uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
43   if XNN_UNPREDICTABLE(mr < 2) {
44     a1 = a0;
45     c1 = c0;
46   }
47   const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
48   uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     a2 = a1;
51     c2 = c1;
52   }
53   const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
54   uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
55   if XNN_UNPREDICTABLE(mr != 4) {
56     a3 = a2;
57     c3 = c2;
58   }
59 
60   const uint8x8_t vb_zero_point = vld1_dup_u8(&params->fp32_neon.kernel_zero_point[0]);
61   do {
62     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
63     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
64     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
65     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
66     int32x4_t vacc1x0123 = vacc0x0123;
67     int32x4_t vacc1x4567 = vacc0x4567;
68     int32x4_t vacc1x89AB = vacc0x89AB;
69     int32x4_t vacc1xCDEF = vacc0xCDEF;
70     int32x4_t vacc2x0123 = vacc0x0123;
71     int32x4_t vacc2x4567 = vacc0x4567;
72     int32x4_t vacc2x89AB = vacc0x89AB;
73     int32x4_t vacc2xCDEF = vacc0xCDEF;
74     int32x4_t vacc3x0123 = vacc0x0123;
75     int32x4_t vacc3x4567 = vacc0x4567;
76     int32x4_t vacc3x89AB = vacc0x89AB;
77     int32x4_t vacc3xCDEF = vacc0xCDEF;
78 
79     size_t k = kc;
80     while (k >= 8 * sizeof(uint8_t)) {
81       const uint8x8_t va0 = vld1_u8(a0); a0 += 8;
82       const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
83       const uint8x8_t va1 = vld1_u8(a1); a1 += 8;
84       const int16x8_t vxa1 = vreinterpretq_s16_u16(vmovl_u8(va1));
85       const uint8x8_t va2 = vld1_u8(a2); a2 += 8;
86       const int16x8_t vxa2 = vreinterpretq_s16_u16(vmovl_u8(va2));
87       const uint8x8_t va3 = vld1_u8(a3); a3 += 8;
88       const int16x8_t vxa3 = vreinterpretq_s16_u16(vmovl_u8(va3));
89 
90       const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
91       const int16x8_t vxb01234567c0 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c0, vb_zero_point));
92 
93       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
94       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
95       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
96       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
97       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
98       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
99       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
100       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
101       const uint8x8_t vb89ABCDEFc0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
102       const int16x8_t vxb89ABCDEFc0 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc0, vb_zero_point));
103 
104       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
105       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
106       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
107       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
108       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
109       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
110       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
111       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
112       const uint8x8_t vb01234567c1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
113       const int16x8_t vxb01234567c1 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c1, vb_zero_point));
114 
115       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
116       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
117       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
118       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
119       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
120       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
121       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
122       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
123       const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
124       const int16x8_t vxb89ABCDEFc1 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc1, vb_zero_point));
125 
126       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
127       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
128       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
129       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
130       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
131       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
132       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
133       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
134       const uint8x8_t vb01234567c2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
135       const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point));
136 
137       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
138       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
139       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
140       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
141       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
142       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
143       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
144       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
145       const uint8x8_t vb89ABCDEFc2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
146       const int16x8_t vxb89ABCDEFc2 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc2, vb_zero_point));
147 
148       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
149       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
150       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
151       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
152       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
153       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
154       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
155       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
156       const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
157       const int16x8_t vxb01234567c3 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c3, vb_zero_point));
158 
159       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
160       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
161       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
162       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
163       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
164       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
165       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
166       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
167       const uint8x8_t vb89ABCDEFc3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
168       const int16x8_t vxb89ABCDEFc3 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc3, vb_zero_point));
169 
170       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
171       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
172       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
173       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
174       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
175       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
176       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
177       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
178 
179 
180       const uint8x8_t vb01234567c4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
181       const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point));
182 
183       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
184       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
185       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
186       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
187       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
188       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
189       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
190       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
191       const uint8x8_t vb89ABCDEFc4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
192       const int16x8_t vxb89ABCDEFc4 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc4, vb_zero_point));
193 
194       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
195       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
196       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
197       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
198       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
199       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
200       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
201       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
202       const uint8x8_t vb01234567c5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
203       const int16x8_t vxb01234567c5 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c5, vb_zero_point));
204 
205       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
206       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
207       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
208       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
209       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
210       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
211       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
212       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
213       const uint8x8_t vb89ABCDEFc5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
214       const int16x8_t vxb89ABCDEFc5 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc5, vb_zero_point));
215 
216       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
217       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
218       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
219       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
220       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
221       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
222       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
223       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
224       const uint8x8_t vb01234567c6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
225       const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));
226 
227       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
228       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
229       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
230       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
231       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
232       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
233       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
234       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
235       const uint8x8_t vb89ABCDEFc6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
236       const int16x8_t vxb89ABCDEFc6 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc6, vb_zero_point));
237 
238       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
239       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
240       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
241       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
242       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
243       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
244       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
245       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
246       const uint8x8_t vb01234567c7 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
247       const int16x8_t vxb01234567c7 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c7, vb_zero_point));
248 
249       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
250       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
251       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
252       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
253       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
254       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
255       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
256       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
257       const uint8x8_t vb89ABCDEFc7 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
258       const int16x8_t vxb89ABCDEFc7 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc7, vb_zero_point));
259 
260       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
261       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
262       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);
263       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);
264       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa2), 3);
265       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa2), 3);
266       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3);
267       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3);
268 
269       k -= 8 * sizeof(uint8_t);
270     }
271     if XNN_UNLIKELY(k != 0) {
272       const uint8x8_t va0 = vld1_u8(a0); a0 = (const uint8_t*) ((uintptr_t) a0 + k);
273       const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
274       const uint8x8_t va1 = vld1_u8(a1); a1 = (const uint8_t*) ((uintptr_t) a1 + k);
275       const int16x8_t vxa1 = vreinterpretq_s16_u16(vmovl_u8(va1));
276       const uint8x8_t va2 = vld1_u8(a2); a2 = (const uint8_t*) ((uintptr_t) a2 + k);
277       const int16x8_t vxa2 = vreinterpretq_s16_u16(vmovl_u8(va2));
278       const uint8x8_t va3 = vld1_u8(a3); a3 = (const uint8_t*) ((uintptr_t) a3 + k);
279       const int16x8_t vxa3 = vreinterpretq_s16_u16(vmovl_u8(va3));
280 
281       const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
282       const int16x8_t vxb01234567c0 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c0, vb_zero_point));
283       const uint8x8_t vb89ABCDEFc0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
284       const int16x8_t vxb89ABCDEFc0 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc0, vb_zero_point));
285 
286       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
287       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
288       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
289       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
290       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
291       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
292       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
293       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
294       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
295       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
296       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
297       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
298       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
299       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
300       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
301       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
302 
303       if (k >= 2 * sizeof(uint8_t)) {
304         const uint8x8_t vb01234567c1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
305         const int16x8_t vxb01234567c1 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c1, vb_zero_point));
306         const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
307         const int16x8_t vxb89ABCDEFc1 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc1, vb_zero_point));
308 
309         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
310         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
311         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
312         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
313         vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
314         vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
315         vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
316         vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
317         vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
318         vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
319         vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
320         vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
321         vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
322         vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
323         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
324         vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
325 
326         if (k > 2 * sizeof(uint8_t)) {
327           const uint8x8_t vb01234567c2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
328           const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point));
329           const uint8x8_t vb89ABCDEFc2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
330           const int16x8_t vxb89ABCDEFc2 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc2, vb_zero_point));
331 
332           vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
333           vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
334           vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
335           vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
336           vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
337           vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
338           vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
339           vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
340           vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
341           vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
342           vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
343           vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
344           vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
345           vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
346           vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
347           vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
348 
349           if (k >= 4 * sizeof(uint8_t)) {
350             const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
351             const int16x8_t vxb01234567c3 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c3, vb_zero_point));
352             const uint8x8_t vb89ABCDEFc3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
353             const int16x8_t vxb89ABCDEFc3 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc3, vb_zero_point));
354 
355             vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
356             vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
357             vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
358             vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
359             vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
360             vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
361             vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
362             vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
363             vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
364             vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
365             vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
366             vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
367             vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
368             vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
369             vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
370             vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
371 
372             if (k > 4 * sizeof(uint8_t)) {
373               const uint8x8_t vb01234567c4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
374               const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point));
375               const uint8x8_t vb89ABCDEFc4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
376               const int16x8_t vxb89ABCDEFc4 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc4, vb_zero_point));
377 
378               vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
379               vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
380               vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
381               vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
382               vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
383               vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
384               vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
385               vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
386               vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
387               vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
388               vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
389               vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
390               vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
391               vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
392               vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
393               vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
394 
395               if (k >= 6 * sizeof(uint8_t)) {
396                 const uint8x8_t vb01234567c5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
397                 const int16x8_t vxb01234567c5 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c5, vb_zero_point));
398                 const uint8x8_t vb89ABCDEFc5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
399                 const int16x8_t vxb89ABCDEFc5 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc5, vb_zero_point));
400 
401                 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
402                 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
403                 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
404                 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
405                 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
406                 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
407                 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
408                 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
409                 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
410                 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
411                 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
412                 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
413                 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
414                 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
415                 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
416                 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
417 
418                 if (k > 6 * sizeof(uint8_t)) {
419                   const uint8x8_t vb01234567c6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
420                   const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));
421                   const uint8x8_t vb89ABCDEFc6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
422                   const int16x8_t vxb89ABCDEFc6 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc6, vb_zero_point));
423 
424                   vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
425                   vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
426                   vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
427                   vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
428                   vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
429                   vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
430                   vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
431                   vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
432                   vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
433                   vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
434                   vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
435                   vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
436                   vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
437                   vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
438                   vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
439                   vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
440                 }
441               }
442             }
443           }
444         }
445       }
446     }
447 
448     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
449     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
450     float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
451     float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
452     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
453     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
454     float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB);
455     float32x4_t vfpacc1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
456     float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
457     float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
458     float32x4_t vfpacc2x89AB = vcvtq_f32_s32(vacc2x89AB);
459     float32x4_t vfpacc2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
460     float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
461     float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
462     float32x4_t vfpacc3x89AB = vcvtq_f32_s32(vacc3x89AB);
463     float32x4_t vfpacc3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
464 
465     const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neon.scale);
466     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
467     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
468     vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale);
469     vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscale);
470     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale);
471     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale);
472     vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale);
473     vfpacc1xCDEF = vmulq_f32(vfpacc1xCDEF, vscale);
474     vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale);
475     vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale);
476     vfpacc2x89AB = vmulq_f32(vfpacc2x89AB, vscale);
477     vfpacc2xCDEF = vmulq_f32(vfpacc2xCDEF, vscale);
478     vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale);
479     vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale);
480     vfpacc3x89AB = vmulq_f32(vfpacc3x89AB, vscale);
481     vfpacc3xCDEF = vmulq_f32(vfpacc3xCDEF, vscale);
482 
483     const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
484     vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
485     vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
486     vacc0x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x89AB, vmagic_bias));
487     vacc0xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc0xCDEF, vmagic_bias));
488     vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
489     vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
490     vacc1x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x89AB, vmagic_bias));
491     vacc1xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc1xCDEF, vmagic_bias));
492     vacc2x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc2x0123, vmagic_bias));
493     vacc2x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc2x4567, vmagic_bias));
494     vacc2x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc2x89AB, vmagic_bias));
495     vacc2xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc2xCDEF, vmagic_bias));
496     vacc3x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc3x0123, vmagic_bias));
497     vacc3x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc3x4567, vmagic_bias));
498     vacc3x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc3x89AB, vmagic_bias));
499     vacc3xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc3xCDEF, vmagic_bias));
500 
501     const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
502     vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
503     vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
504     vacc0x89AB = vqsubq_s32(vacc0x89AB, vmagic_bias_less_output_zero_point);
505     vacc0xCDEF = vqsubq_s32(vacc0xCDEF, vmagic_bias_less_output_zero_point);
506     vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
507     vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
508     vacc1x89AB = vqsubq_s32(vacc1x89AB, vmagic_bias_less_output_zero_point);
509     vacc1xCDEF = vqsubq_s32(vacc1xCDEF, vmagic_bias_less_output_zero_point);
510     vacc2x0123 = vqsubq_s32(vacc2x0123, vmagic_bias_less_output_zero_point);
511     vacc2x4567 = vqsubq_s32(vacc2x4567, vmagic_bias_less_output_zero_point);
512     vacc2x89AB = vqsubq_s32(vacc2x89AB, vmagic_bias_less_output_zero_point);
513     vacc2xCDEF = vqsubq_s32(vacc2xCDEF, vmagic_bias_less_output_zero_point);
514     vacc3x0123 = vqsubq_s32(vacc3x0123, vmagic_bias_less_output_zero_point);
515     vacc3x4567 = vqsubq_s32(vacc3x4567, vmagic_bias_less_output_zero_point);
516     vacc3x89AB = vqsubq_s32(vacc3x89AB, vmagic_bias_less_output_zero_point);
517     vacc3xCDEF = vqsubq_s32(vacc3xCDEF, vmagic_bias_less_output_zero_point);
518 
519 #if XNN_ARCH_ARM64
520     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
521     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
522     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
523     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
524     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
525     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
526     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
527     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
528 
529 
530     uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
531     uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
532     uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
533     uint8x16_t vout3x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc3x01234567), vacc3x89ABCDEF);
534 #else
535     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
536     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
537     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
538     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
539     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
540     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
541     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
542     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
543 
544 
545     uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
546     uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
547     uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
548     uint8x16_t vout3x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc3x01234567), vqmovun_s16(vacc3x89ABCDEF));
549 #endif
550 
551     const uint8x16_t voutput_min = vld1q_dup_u8(&params->fp32_neon.output_min);
552     vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
553     vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
554     vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
555     vout3x0123456789ABCDEF = vmaxq_u8(vout3x0123456789ABCDEF, voutput_min);
556 
557     const uint8x16_t voutput_max = vld1q_dup_u8(&params->fp32_neon.output_max);
558     vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
559     vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
560     vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
561     vout3x0123456789ABCDEF = vminq_u8(vout3x0123456789ABCDEF, voutput_max);
562 
563     if (nc >= 16) {
564       vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
565       vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
566       vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
567       vst1q_u8(c3 + 0, vout3x0123456789ABCDEF);
568 
569       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
570       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
571       c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
572       c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
573 
574       a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
575       a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
576       a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
577       a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
578 
579       nc -= 16;
580     } else {
581       uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
582       uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vget_low_u8(vout2x0123456789ABCDEF), vget_low_u8(vout3x0123456789ABCDEF));
583       if (nc & 8) {
584         vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
585         vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
586         vst1_u8(c2, vget_low_u8(vout2x01234567_3x01234567)); c2 += 8;
587         vst1_u8(c3, vget_high_u8(vout2x01234567_3x01234567)); c3 += 8;
588         vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
589         vout2x01234567_3x01234567 = vcombine_u8(vget_high_u8(vout2x0123456789ABCDEF), vget_high_u8(vout3x0123456789ABCDEF));
590       }
591       if (nc & 4) {
592         vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
593         vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
594         vst1q_lane_u32((void*) c2, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
595         vst1q_lane_u32((void*) c3, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
596         vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
597         vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
598       }
599       if (nc & 2) {
600         vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
601         vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
602         vst1q_lane_u16((void*) c2, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
603         vst1q_lane_u16((void*) c3, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
604         vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
605         vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
606       }
607       if (nc & 1) {
608         vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
609         vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
610         vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
611         vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
612       }
613 
614       nc = 0;
615     }
616   } while (nc != 0);
617 }
618