xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/6x16-minmax-fp32-neonv8-mlal-lane-prfm.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 #include <xnnpack/intrinsics-polyfill.h>
17 
18 
xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     const int8_t* restrict a,
24     size_t a_stride,
25     const void* restrict w,
26     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31   assert(mr != 0);
32   assert(mr <= 6);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(kc % sizeof(int8_t) == 0);
36   assert(a != NULL);
37   assert(w != NULL);
38   assert(c != NULL);
39 
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44   if XNN_UNPREDICTABLE(mr < 2) {
45     a1 = a0;
46     c1 = c0;
47   }
48   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     a2 = a1;
52     c2 = c1;
53   }
54   const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
55   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56   if XNN_UNPREDICTABLE(mr < 4) {
57     a3 = a2;
58     c3 = c2;
59   }
60   const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
61   int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
62   if XNN_UNPREDICTABLE(mr <= 4) {
63     a4 = a3;
64     c4 = c3;
65   }
66   const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
67   int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
68   if XNN_UNPREDICTABLE(mr != 6) {
69     a5 = a4;
70     c5 = c4;
71   }
72 
73   do {
74     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
75     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
76     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
77     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
78     int32x4_t vacc1x0123 = vacc0x0123;
79     int32x4_t vacc1x4567 = vacc0x4567;
80     int32x4_t vacc1x89AB = vacc0x89AB;
81     int32x4_t vacc1xCDEF = vacc0xCDEF;
82     int32x4_t vacc2x0123 = vacc0x0123;
83     int32x4_t vacc2x4567 = vacc0x4567;
84     int32x4_t vacc2x89AB = vacc0x89AB;
85     int32x4_t vacc2xCDEF = vacc0xCDEF;
86     int32x4_t vacc3x0123 = vacc0x0123;
87     int32x4_t vacc3x4567 = vacc0x4567;
88     int32x4_t vacc3x89AB = vacc0x89AB;
89     int32x4_t vacc3xCDEF = vacc0xCDEF;
90     int32x4_t vacc4x0123 = vacc0x0123;
91     int32x4_t vacc4x4567 = vacc0x4567;
92     int32x4_t vacc4x89AB = vacc0x89AB;
93     int32x4_t vacc4xCDEF = vacc0xCDEF;
94     int32x4_t vacc5x0123 = vacc0x0123;
95     int32x4_t vacc5x4567 = vacc0x4567;
96     int32x4_t vacc5x89AB = vacc0x89AB;
97     int32x4_t vacc5xCDEF = vacc0xCDEF;
98 
99     size_t k = kc;
100     while (k >= 8 * sizeof(int8_t)) {
101       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
102       const int16x8_t vxa0 = vmovl_s8(va0);
103       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
104       const int16x8_t vxa1 = vmovl_s8(va1);
105       const int8x8_t va2 = vld1_s8(a2); a2 += 8;
106       const int16x8_t vxa2 = vmovl_s8(va2);
107       const int8x8_t va3 = vld1_s8(a3); a3 += 8;
108       const int16x8_t vxa3 = vmovl_s8(va3);
109       const int8x8_t va4 = vld1_s8(a4); a4 += 8;
110       const int16x8_t vxa4 = vmovl_s8(va4);
111       const int8x8_t va5 = vld1_s8(a5); a5 += 8;
112       const int16x8_t vxa5 = vmovl_s8(va5);
113 
114       const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
115       const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
116 
117       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
118       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
119       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
120       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
121       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
122       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
123       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
124       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
125       vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
126       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
127       vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
128       vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
129       const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
130       const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
131 
132       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
133       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
134       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
135       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
136       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
137       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
138       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
139       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
140       vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa4), 0);
141       vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa4), 0);
142       vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa5), 0);
143       vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa5), 0);
144       const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
145       const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
146 
147       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
148       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
149       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
150       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
151       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
152       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
153       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
154       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
155       vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
156       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
157       vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
158       vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
159       const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
160       const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
161 
162       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
163       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
164       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
165       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
166       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
167       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
168       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
169       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
170       vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa4), 1);
171       vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa4), 1);
172       vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa5), 1);
173       vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa5), 1);
174       const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
175       const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
176 
177       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
178       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
179       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
180       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
181       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
182       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
183       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
184       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
185       vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
186       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
187       vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
188       vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
189       const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
190       const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
191 
192       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
193       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
194       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
195       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
196       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
197       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
198       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
199       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
200       vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa4), 2);
201       vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa4), 2);
202       vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa5), 2);
203       vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa5), 2);
204       const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
205       const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
206 
207       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
208       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
209       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
210       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
211       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
212       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
213       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
214       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
215       vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
216       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
217       vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
218       vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
219       const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
220       const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
221 
222       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
223       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
224       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
225       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
226       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
227       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
228       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
229       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
230       vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa4), 3);
231       vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa4), 3);
232       vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa5), 3);
233       vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa5), 3);
234 
235       __builtin_prefetch((const int8_t*) w + 448);
236       __builtin_prefetch((const int8_t*) w + 512);
237 
238       const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
239       const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
240 
241       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
242       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
243       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
244       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
245       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
246       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
247       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
248       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
249       vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
250       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
251       vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
252       vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
253       const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
254       const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
255 
256       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
257       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
258       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
259       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
260       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
261       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
262       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
263       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
264       vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa4), 0);
265       vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa4), 0);
266       vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa5), 0);
267       vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa5), 0);
268       const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
269       const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
270 
271       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
272       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
273       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
274       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
275       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
276       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
277       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
278       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
279       vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
280       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
281       vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
282       vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
283       const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
284       const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
285 
286       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
287       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
288       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
289       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
290       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
291       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
292       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
293       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
294       vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa4), 1);
295       vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa4), 1);
296       vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa5), 1);
297       vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa5), 1);
298       const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
299       const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
300 
301       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
302       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
303       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
304       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
305       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
306       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
307       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
308       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
309       vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
310       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
311       vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
312       vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
313       const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
314       const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
315 
316       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
317       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
318       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
319       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
320       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
321       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
322       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
323       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
324       vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa4), 2);
325       vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa4), 2);
326       vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa5), 2);
327       vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa5), 2);
328       const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
329       const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
330 
331       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
332       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
333       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
334       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
335       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
336       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
337       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
338       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
339       vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa4), 3);
340       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa4), 3);
341       vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa5), 3);
342       vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa5), 3);
343       const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
344       const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
345 
346       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
347       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
348       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);
349       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);
350       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa2), 3);
351       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa2), 3);
352       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3);
353       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3);
354       vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa4), 3);
355       vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa4), 3);
356       vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa5), 3);
357       vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa5), 3);
358 
359       k -= 8 * sizeof(int8_t);
360     }
361     if XNN_UNLIKELY(k != 0) {
362       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
363       const int16x8_t vxa0 = vmovl_s8(va0);
364       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
365       const int16x8_t vxa1 = vmovl_s8(va1);
366       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
367       const int16x8_t vxa2 = vmovl_s8(va2);
368       const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
369       const int16x8_t vxa3 = vmovl_s8(va3);
370       const int8x8_t va4 = vld1_s8(a4); a4 = (const int8_t*) ((uintptr_t) a4 + k);
371       const int16x8_t vxa4 = vmovl_s8(va4);
372       const int8x8_t va5 = vld1_s8(a5); a5 = (const int8_t*) ((uintptr_t) a5 + k);
373       const int16x8_t vxa5 = vmovl_s8(va5);
374 
375       const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
376       const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
377       const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
378       const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
379 
380       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
381       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
382       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
383       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
384       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
385       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
386       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
387       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
388       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
389       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
390       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
391       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
392       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
393       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
394       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
395       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
396       vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
397       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
398       vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa4), 0);
399       vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa4), 0);
400       vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
401       vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
402       vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa5), 0);
403       vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa5), 0);
404 
405       if (k >= 2 * sizeof(int8_t)) {
406         const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
407         const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
408         const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
409         const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
410 
411         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
412         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
413         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
414         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
415         vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
416         vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
417         vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
418         vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
419         vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
420         vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
421         vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
422         vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
423         vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
424         vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
425         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
426         vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
427         vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
428         vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
429         vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa4), 1);
430         vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa4), 1);
431         vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
432         vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
433         vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa5), 1);
434         vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa5), 1);
435 
436         if (k > 2 * sizeof(int8_t)) {
437           const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
438           const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
439           const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
440           const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
441 
442           vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
443           vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
444           vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
445           vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
446           vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
447           vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
448           vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
449           vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
450           vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
451           vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
452           vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
453           vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
454           vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
455           vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
456           vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
457           vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
458           vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
459           vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
460           vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa4), 2);
461           vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa4), 2);
462           vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
463           vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
464           vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa5), 2);
465           vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa5), 2);
466 
467           if (k >= 4 * sizeof(int8_t)) {
468             const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
469             const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
470             const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
471             const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
472 
473             vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
474             vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
475             vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
476             vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
477             vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
478             vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
479             vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
480             vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
481             vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
482             vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
483             vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
484             vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
485             vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
486             vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
487             vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
488             vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
489             vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
490             vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
491             vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa4), 3);
492             vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa4), 3);
493             vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
494             vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
495             vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa5), 3);
496             vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa5), 3);
497 
498             if (k > 4 * sizeof(int8_t)) {
499               const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
500               const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
501               const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
502               const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
503 
504               vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
505               vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
506               vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
507               vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
508               vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
509               vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
510               vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
511               vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
512               vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
513               vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
514               vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
515               vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
516               vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
517               vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
518               vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
519               vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
520               vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
521               vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
522               vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa4), 0);
523               vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa4), 0);
524               vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
525               vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
526               vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa5), 0);
527               vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa5), 0);
528 
529               if (k >= 6 * sizeof(int8_t)) {
530                 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
531                 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
532                 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
533                 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
534 
535                 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
536                 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
537                 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
538                 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
539                 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
540                 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
541                 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
542                 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
543                 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
544                 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
545                 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
546                 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
547                 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
548                 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
549                 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
550                 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
551                 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
552                 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
553                 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa4), 1);
554                 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa4), 1);
555                 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
556                 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
557                 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa5), 1);
558                 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa5), 1);
559 
560                 if (k > 6 * sizeof(int8_t)) {
561                   const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
562                   const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
563                   const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
564                   const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
565 
566                   vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
567                   vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
568                   vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
569                   vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
570                   vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
571                   vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
572                   vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
573                   vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
574                   vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
575                   vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
576                   vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
577                   vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
578                   vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
579                   vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
580                   vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
581                   vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
582                   vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
583                   vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
584                   vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa4), 2);
585                   vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa4), 2);
586                   vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
587                   vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
588                   vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa5), 2);
589                   vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa5), 2);
590                 }
591               }
592             }
593           }
594         }
595       }
596     }
597 
598     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
599     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
600     float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
601     float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
602     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
603     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
604     float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB);
605     float32x4_t vfpacc1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
606     float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
607     float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
608     float32x4_t vfpacc2x89AB = vcvtq_f32_s32(vacc2x89AB);
609     float32x4_t vfpacc2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
610     float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
611     float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
612     float32x4_t vfpacc3x89AB = vcvtq_f32_s32(vacc3x89AB);
613     float32x4_t vfpacc3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
614     float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123);
615     float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567);
616     float32x4_t vfpacc4x89AB = vcvtq_f32_s32(vacc4x89AB);
617     float32x4_t vfpacc4xCDEF = vcvtq_f32_s32(vacc4xCDEF);
618     float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123);
619     float32x4_t vfpacc5x4567 = vcvtq_f32_s32(vacc5x4567);
620     float32x4_t vfpacc5x89AB = vcvtq_f32_s32(vacc5x89AB);
621     float32x4_t vfpacc5xCDEF = vcvtq_f32_s32(vacc5xCDEF);
622 
623     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
624     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
625     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
626     vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
627     vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale0123);
628     vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123);
629     vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123);
630     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
631     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
632     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
633     vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
634     vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale4567);
635     vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567);
636     vfpacc5x4567 = vmulq_f32(vfpacc5x4567, vscale4567);
637     const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
638     vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale89AB);
639     vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB);
640     vfpacc2x89AB = vmulq_f32(vfpacc2x89AB, vscale89AB);
641     vfpacc3x89AB = vmulq_f32(vfpacc3x89AB, vscale89AB);
642     vfpacc4x89AB = vmulq_f32(vfpacc4x89AB, vscale89AB);
643     vfpacc5x89AB = vmulq_f32(vfpacc5x89AB, vscale89AB);
644     const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
645     vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscaleCDEF);
646     vfpacc1xCDEF = vmulq_f32(vfpacc1xCDEF, vscaleCDEF);
647     vfpacc2xCDEF = vmulq_f32(vfpacc2xCDEF, vscaleCDEF);
648     vfpacc3xCDEF = vmulq_f32(vfpacc3xCDEF, vscaleCDEF);
649     vfpacc4xCDEF = vmulq_f32(vfpacc4xCDEF, vscaleCDEF);
650     vfpacc5xCDEF = vmulq_f32(vfpacc5xCDEF, vscaleCDEF);
651 
652     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
653     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
654     vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
655     vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
656     vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
657     vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
658     vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB);
659     vacc1xCDEF = vcvtnq_s32_f32(vfpacc1xCDEF);
660     vacc2x0123 = vcvtnq_s32_f32(vfpacc2x0123);
661     vacc2x4567 = vcvtnq_s32_f32(vfpacc2x4567);
662     vacc2x89AB = vcvtnq_s32_f32(vfpacc2x89AB);
663     vacc2xCDEF = vcvtnq_s32_f32(vfpacc2xCDEF);
664     vacc3x0123 = vcvtnq_s32_f32(vfpacc3x0123);
665     vacc3x4567 = vcvtnq_s32_f32(vfpacc3x4567);
666     vacc3x89AB = vcvtnq_s32_f32(vfpacc3x89AB);
667     vacc3xCDEF = vcvtnq_s32_f32(vfpacc3xCDEF);
668     vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123);
669     vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567);
670     vacc4x89AB = vcvtnq_s32_f32(vfpacc4x89AB);
671     vacc4xCDEF = vcvtnq_s32_f32(vfpacc4xCDEF);
672     vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123);
673     vacc5x4567 = vcvtnq_s32_f32(vfpacc5x4567);
674     vacc5x89AB = vcvtnq_s32_f32(vfpacc5x89AB);
675     vacc5xCDEF = vcvtnq_s32_f32(vfpacc5xCDEF);
676 
677     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
678 #if XNN_ARCH_ARM64
679     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
680     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
681     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
682     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
683     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
684     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
685     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
686     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
687     int16x8_t vacc4x01234567 = vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567);
688     int16x8_t vacc4x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF);
689     int16x8_t vacc5x01234567 = vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567);
690     int16x8_t vacc5x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc5x89AB), vacc5xCDEF);
691 
692     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
693     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
694     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
695     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
696     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
697     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
698     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
699     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
700     vacc4x01234567 = vqaddq_s16(vacc4x01234567, voutput_zero_point);
701     vacc4x89ABCDEF = vqaddq_s16(vacc4x89ABCDEF, voutput_zero_point);
702     vacc5x01234567 = vqaddq_s16(vacc5x01234567, voutput_zero_point);
703     vacc5x89ABCDEF = vqaddq_s16(vacc5x89ABCDEF, voutput_zero_point);
704 
705     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
706     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
707     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
708     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
709     int8x16_t vout4x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc4x89ABCDEF);
710     int8x16_t vout5x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc5x01234567), vacc5x89ABCDEF);
711 #else
712     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
713     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
714     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
715     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
716     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
717     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
718     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
719     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
720     int16x8_t vacc4x01234567 = vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567));
721     int16x8_t vacc4x89ABCDEF = vcombine_s16(vqmovn_s32(vacc4x89AB), vqmovn_s32(vacc4xCDEF));
722     int16x8_t vacc5x01234567 = vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567));
723     int16x8_t vacc5x89ABCDEF = vcombine_s16(vqmovn_s32(vacc5x89AB), vqmovn_s32(vacc5xCDEF));
724 
725     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
726     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
727     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
728     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
729     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
730     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
731     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
732     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
733     vacc4x01234567 = vqaddq_s16(vacc4x01234567, voutput_zero_point);
734     vacc4x89ABCDEF = vqaddq_s16(vacc4x89ABCDEF, voutput_zero_point);
735     vacc5x01234567 = vqaddq_s16(vacc5x01234567, voutput_zero_point);
736     vacc5x89ABCDEF = vqaddq_s16(vacc5x89ABCDEF, voutput_zero_point);
737 
738     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
739     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
740     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
741     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
742     int8x16_t vout4x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc4x89ABCDEF));
743     int8x16_t vout5x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc5x01234567), vqmovn_s16(vacc5x89ABCDEF));
744 #endif
745 
746     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
747     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
748     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
749     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
750     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
751     vout4x0123456789ABCDEF = vmaxq_s8(vout4x0123456789ABCDEF, voutput_min);
752     vout5x0123456789ABCDEF = vmaxq_s8(vout5x0123456789ABCDEF, voutput_min);
753 
754     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
755     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
756     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
757     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
758     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
759     vout4x0123456789ABCDEF = vminq_s8(vout4x0123456789ABCDEF, voutput_max);
760     vout5x0123456789ABCDEF = vminq_s8(vout5x0123456789ABCDEF, voutput_max);
761 
762     if (nc >= 16) {
763       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
764       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
765       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
766       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
767       vst1q_s8(c4 + 0, vout4x0123456789ABCDEF);
768       vst1q_s8(c5 + 0, vout5x0123456789ABCDEF);
769 
770       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
771       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
772       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
773       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
774       c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
775       c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
776 
777       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
778       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
779       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
780       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
781       a4 = (const int8_t*) ((uintptr_t) a4 - kc);
782       a5 = (const int8_t*) ((uintptr_t) a5 - kc);
783 
784       nc -= 16;
785     } else {
786       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
787       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
788       int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vget_low_s8(vout4x0123456789ABCDEF), vget_low_s8(vout5x0123456789ABCDEF));
789       if (nc & 8) {
790         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
791         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
792         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
793         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
794         vst1_s8(c4, vget_low_s8(vout4x01234567_5x01234567)); c4 += 8;
795         vst1_s8(c5, vget_high_s8(vout4x01234567_5x01234567)); c5 += 8;
796         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
797         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
798         vout4x01234567_5x01234567 = vcombine_s8(vget_high_s8(vout4x0123456789ABCDEF), vget_high_s8(vout5x0123456789ABCDEF));
799       }
800       if (nc & 4) {
801         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
802         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
803         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
804         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
805         vst1q_lane_u32((void*) c4, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4;
806         vst1q_lane_u32((void*) c5, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4;
807         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
808         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
809         vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
810       }
811       if (nc & 2) {
812         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
813         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
814         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
815         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
816         vst1q_lane_u16((void*) c4, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2;
817         vst1q_lane_u16((void*) c5, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2;
818         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
819         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
820         vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
821       }
822       if (nc & 1) {
823         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
824         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
825         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
826         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
827         vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0);
828         vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8);
829       }
830 
831       nc = 0;
832     }
833   } while (nc != 0);
834 }
835