xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/6x16-minmax-fp32-neon-mlal-lane.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/qs8-igemm/neon-mlal-lane.c.in
3*4bdc9457SAndroid Build Coastguard Worker //   Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2020 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker 
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker 
12*4bdc9457SAndroid Build Coastguard Worker #include <arm_neon.h>
13*4bdc9457SAndroid Build Coastguard Worker 
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/igemm.h>
16*4bdc9457SAndroid Build Coastguard Worker 
17*4bdc9457SAndroid Build Coastguard Worker 
xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18*4bdc9457SAndroid Build Coastguard Worker void xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane(
19*4bdc9457SAndroid Build Coastguard Worker     size_t mr,
20*4bdc9457SAndroid Build Coastguard Worker     size_t nc,
21*4bdc9457SAndroid Build Coastguard Worker     size_t kc,
22*4bdc9457SAndroid Build Coastguard Worker     size_t ks,
23*4bdc9457SAndroid Build Coastguard Worker     const int8_t** restrict a,
24*4bdc9457SAndroid Build Coastguard Worker     const void* restrict w,
25*4bdc9457SAndroid Build Coastguard Worker     int8_t* restrict c,
26*4bdc9457SAndroid Build Coastguard Worker     size_t cm_stride,
27*4bdc9457SAndroid Build Coastguard Worker     size_t cn_stride,
28*4bdc9457SAndroid Build Coastguard Worker     size_t a_offset,
29*4bdc9457SAndroid Build Coastguard Worker     const int8_t* zero,
30*4bdc9457SAndroid Build Coastguard Worker     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31*4bdc9457SAndroid Build Coastguard Worker {
32*4bdc9457SAndroid Build Coastguard Worker   assert(mr != 0);
33*4bdc9457SAndroid Build Coastguard Worker   assert(mr <= 6);
34*4bdc9457SAndroid Build Coastguard Worker   assert(nc != 0);
35*4bdc9457SAndroid Build Coastguard Worker   assert(kc != 0);
36*4bdc9457SAndroid Build Coastguard Worker   assert(ks != 0);
37*4bdc9457SAndroid Build Coastguard Worker   assert(ks % (6 * sizeof(void*)) == 0);
38*4bdc9457SAndroid Build Coastguard Worker   assert(a_offset % sizeof(int8_t) == 0);
39*4bdc9457SAndroid Build Coastguard Worker   assert(a != NULL);
40*4bdc9457SAndroid Build Coastguard Worker   assert(w != NULL);
41*4bdc9457SAndroid Build Coastguard Worker   assert(c != NULL);
42*4bdc9457SAndroid Build Coastguard Worker 
43*4bdc9457SAndroid Build Coastguard Worker   int8_t* c0 = c;
44*4bdc9457SAndroid Build Coastguard Worker   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNPREDICTABLE(mr < 2) {
46*4bdc9457SAndroid Build Coastguard Worker     c1 = c0;
47*4bdc9457SAndroid Build Coastguard Worker   }
48*4bdc9457SAndroid Build Coastguard Worker   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNPREDICTABLE(mr <= 2) {
50*4bdc9457SAndroid Build Coastguard Worker     c2 = c1;
51*4bdc9457SAndroid Build Coastguard Worker   }
52*4bdc9457SAndroid Build Coastguard Worker   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
53*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNPREDICTABLE(mr < 4) {
54*4bdc9457SAndroid Build Coastguard Worker     c3 = c2;
55*4bdc9457SAndroid Build Coastguard Worker   }
56*4bdc9457SAndroid Build Coastguard Worker   int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
57*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNPREDICTABLE(mr <= 4) {
58*4bdc9457SAndroid Build Coastguard Worker     c4 = c3;
59*4bdc9457SAndroid Build Coastguard Worker   }
60*4bdc9457SAndroid Build Coastguard Worker   int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
61*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNPREDICTABLE(mr != 6) {
62*4bdc9457SAndroid Build Coastguard Worker     c5 = c4;
63*4bdc9457SAndroid Build Coastguard Worker   }
64*4bdc9457SAndroid Build Coastguard Worker 
65*4bdc9457SAndroid Build Coastguard Worker   do {
66*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
67*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
68*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
69*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
70*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc1x0123 = vacc0x0123;
71*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc1x4567 = vacc0x4567;
72*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc1x89AB = vacc0x89AB;
73*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc1xCDEF = vacc0xCDEF;
74*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc2x0123 = vacc0x0123;
75*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc2x4567 = vacc0x4567;
76*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc2x89AB = vacc0x89AB;
77*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc2xCDEF = vacc0xCDEF;
78*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc3x0123 = vacc0x0123;
79*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc3x4567 = vacc0x4567;
80*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc3x89AB = vacc0x89AB;
81*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc3xCDEF = vacc0xCDEF;
82*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc4x0123 = vacc0x0123;
83*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc4x4567 = vacc0x4567;
84*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc4x89AB = vacc0x89AB;
85*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc4xCDEF = vacc0xCDEF;
86*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc5x0123 = vacc0x0123;
87*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc5x4567 = vacc0x4567;
88*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc5x89AB = vacc0x89AB;
89*4bdc9457SAndroid Build Coastguard Worker     int32x4_t vacc5xCDEF = vacc0xCDEF;
90*4bdc9457SAndroid Build Coastguard Worker 
91*4bdc9457SAndroid Build Coastguard Worker     size_t p = ks;
92*4bdc9457SAndroid Build Coastguard Worker     do {
93*4bdc9457SAndroid Build Coastguard Worker       const int8_t* restrict a0 = a[0];
94*4bdc9457SAndroid Build Coastguard Worker       if XNN_UNPREDICTABLE(a0 != zero) {
95*4bdc9457SAndroid Build Coastguard Worker         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
96*4bdc9457SAndroid Build Coastguard Worker       }
97*4bdc9457SAndroid Build Coastguard Worker       const int8_t* restrict a1 = a[1];
98*4bdc9457SAndroid Build Coastguard Worker       if XNN_UNPREDICTABLE(a1 != zero) {
99*4bdc9457SAndroid Build Coastguard Worker         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
100*4bdc9457SAndroid Build Coastguard Worker       }
101*4bdc9457SAndroid Build Coastguard Worker       const int8_t* restrict a2 = a[2];
102*4bdc9457SAndroid Build Coastguard Worker       if XNN_UNPREDICTABLE(a2 != zero) {
103*4bdc9457SAndroid Build Coastguard Worker         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
104*4bdc9457SAndroid Build Coastguard Worker       }
105*4bdc9457SAndroid Build Coastguard Worker       const int8_t* restrict a3 = a[3];
106*4bdc9457SAndroid Build Coastguard Worker       if XNN_UNPREDICTABLE(a3 != zero) {
107*4bdc9457SAndroid Build Coastguard Worker         a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
108*4bdc9457SAndroid Build Coastguard Worker       }
109*4bdc9457SAndroid Build Coastguard Worker       const int8_t* restrict a4 = a[4];
110*4bdc9457SAndroid Build Coastguard Worker       if XNN_UNPREDICTABLE(a4 != zero) {
111*4bdc9457SAndroid Build Coastguard Worker         a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
112*4bdc9457SAndroid Build Coastguard Worker       }
113*4bdc9457SAndroid Build Coastguard Worker       const int8_t* restrict a5 = a[5];
114*4bdc9457SAndroid Build Coastguard Worker       if XNN_UNPREDICTABLE(a5 != zero) {
115*4bdc9457SAndroid Build Coastguard Worker         a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
116*4bdc9457SAndroid Build Coastguard Worker       }
117*4bdc9457SAndroid Build Coastguard Worker       a += 6;
118*4bdc9457SAndroid Build Coastguard Worker 
119*4bdc9457SAndroid Build Coastguard Worker       size_t k = kc;
120*4bdc9457SAndroid Build Coastguard Worker       while (k >= 8 * sizeof(int8_t)) {
121*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
122*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxa0 = vmovl_s8(va0);
123*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t va1 = vld1_s8(a1); a1 += 8;
124*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxa1 = vmovl_s8(va1);
125*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t va2 = vld1_s8(a2); a2 += 8;
126*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxa2 = vmovl_s8(va2);
127*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t va3 = vld1_s8(a3); a3 += 8;
128*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxa3 = vmovl_s8(va3);
129*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t va4 = vld1_s8(a4); a4 += 8;
130*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxa4 = vmovl_s8(va4);
131*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t va5 = vld1_s8(a5); a5 += 8;
132*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxa5 = vmovl_s8(va5);
133*4bdc9457SAndroid Build Coastguard Worker 
134*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
135*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
136*4bdc9457SAndroid Build Coastguard Worker 
137*4bdc9457SAndroid Build Coastguard Worker         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
138*4bdc9457SAndroid Build Coastguard Worker         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
139*4bdc9457SAndroid Build Coastguard Worker         vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
140*4bdc9457SAndroid Build Coastguard Worker         vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
141*4bdc9457SAndroid Build Coastguard Worker         vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
142*4bdc9457SAndroid Build Coastguard Worker         vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
143*4bdc9457SAndroid Build Coastguard Worker         vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
144*4bdc9457SAndroid Build Coastguard Worker         vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
145*4bdc9457SAndroid Build Coastguard Worker         vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
146*4bdc9457SAndroid Build Coastguard Worker         vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
147*4bdc9457SAndroid Build Coastguard Worker         vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
148*4bdc9457SAndroid Build Coastguard Worker         vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
149*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
150*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
151*4bdc9457SAndroid Build Coastguard Worker 
152*4bdc9457SAndroid Build Coastguard Worker         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
153*4bdc9457SAndroid Build Coastguard Worker         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
154*4bdc9457SAndroid Build Coastguard Worker         vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
155*4bdc9457SAndroid Build Coastguard Worker         vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
156*4bdc9457SAndroid Build Coastguard Worker         vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
157*4bdc9457SAndroid Build Coastguard Worker         vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
158*4bdc9457SAndroid Build Coastguard Worker         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
159*4bdc9457SAndroid Build Coastguard Worker         vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
160*4bdc9457SAndroid Build Coastguard Worker         vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa4), 0);
161*4bdc9457SAndroid Build Coastguard Worker         vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa4), 0);
162*4bdc9457SAndroid Build Coastguard Worker         vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa5), 0);
163*4bdc9457SAndroid Build Coastguard Worker         vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa5), 0);
164*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
165*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
166*4bdc9457SAndroid Build Coastguard Worker 
167*4bdc9457SAndroid Build Coastguard Worker         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
168*4bdc9457SAndroid Build Coastguard Worker         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
169*4bdc9457SAndroid Build Coastguard Worker         vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
170*4bdc9457SAndroid Build Coastguard Worker         vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
171*4bdc9457SAndroid Build Coastguard Worker         vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
172*4bdc9457SAndroid Build Coastguard Worker         vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
173*4bdc9457SAndroid Build Coastguard Worker         vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
174*4bdc9457SAndroid Build Coastguard Worker         vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
175*4bdc9457SAndroid Build Coastguard Worker         vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
176*4bdc9457SAndroid Build Coastguard Worker         vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
177*4bdc9457SAndroid Build Coastguard Worker         vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
178*4bdc9457SAndroid Build Coastguard Worker         vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
179*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
180*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
181*4bdc9457SAndroid Build Coastguard Worker 
182*4bdc9457SAndroid Build Coastguard Worker         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
183*4bdc9457SAndroid Build Coastguard Worker         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
184*4bdc9457SAndroid Build Coastguard Worker         vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
185*4bdc9457SAndroid Build Coastguard Worker         vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
186*4bdc9457SAndroid Build Coastguard Worker         vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
187*4bdc9457SAndroid Build Coastguard Worker         vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
188*4bdc9457SAndroid Build Coastguard Worker         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
189*4bdc9457SAndroid Build Coastguard Worker         vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
190*4bdc9457SAndroid Build Coastguard Worker         vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa4), 1);
191*4bdc9457SAndroid Build Coastguard Worker         vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa4), 1);
192*4bdc9457SAndroid Build Coastguard Worker         vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa5), 1);
193*4bdc9457SAndroid Build Coastguard Worker         vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa5), 1);
194*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
195*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
196*4bdc9457SAndroid Build Coastguard Worker 
197*4bdc9457SAndroid Build Coastguard Worker         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
198*4bdc9457SAndroid Build Coastguard Worker         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
199*4bdc9457SAndroid Build Coastguard Worker         vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
200*4bdc9457SAndroid Build Coastguard Worker         vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
201*4bdc9457SAndroid Build Coastguard Worker         vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
202*4bdc9457SAndroid Build Coastguard Worker         vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
203*4bdc9457SAndroid Build Coastguard Worker         vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
204*4bdc9457SAndroid Build Coastguard Worker         vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
205*4bdc9457SAndroid Build Coastguard Worker         vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
206*4bdc9457SAndroid Build Coastguard Worker         vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
207*4bdc9457SAndroid Build Coastguard Worker         vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
208*4bdc9457SAndroid Build Coastguard Worker         vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
209*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
210*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
211*4bdc9457SAndroid Build Coastguard Worker 
212*4bdc9457SAndroid Build Coastguard Worker         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
213*4bdc9457SAndroid Build Coastguard Worker         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
214*4bdc9457SAndroid Build Coastguard Worker         vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
215*4bdc9457SAndroid Build Coastguard Worker         vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
216*4bdc9457SAndroid Build Coastguard Worker         vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
217*4bdc9457SAndroid Build Coastguard Worker         vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
218*4bdc9457SAndroid Build Coastguard Worker         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
219*4bdc9457SAndroid Build Coastguard Worker         vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
220*4bdc9457SAndroid Build Coastguard Worker         vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa4), 2);
221*4bdc9457SAndroid Build Coastguard Worker         vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa4), 2);
222*4bdc9457SAndroid Build Coastguard Worker         vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa5), 2);
223*4bdc9457SAndroid Build Coastguard Worker         vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa5), 2);
224*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
225*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
226*4bdc9457SAndroid Build Coastguard Worker 
227*4bdc9457SAndroid Build Coastguard Worker         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
228*4bdc9457SAndroid Build Coastguard Worker         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
229*4bdc9457SAndroid Build Coastguard Worker         vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
230*4bdc9457SAndroid Build Coastguard Worker         vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
231*4bdc9457SAndroid Build Coastguard Worker         vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
232*4bdc9457SAndroid Build Coastguard Worker         vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
233*4bdc9457SAndroid Build Coastguard Worker         vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
234*4bdc9457SAndroid Build Coastguard Worker         vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
235*4bdc9457SAndroid Build Coastguard Worker         vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
236*4bdc9457SAndroid Build Coastguard Worker         vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
237*4bdc9457SAndroid Build Coastguard Worker         vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
238*4bdc9457SAndroid Build Coastguard Worker         vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
239*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
240*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
241*4bdc9457SAndroid Build Coastguard Worker 
242*4bdc9457SAndroid Build Coastguard Worker         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
243*4bdc9457SAndroid Build Coastguard Worker         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
244*4bdc9457SAndroid Build Coastguard Worker         vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
245*4bdc9457SAndroid Build Coastguard Worker         vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
246*4bdc9457SAndroid Build Coastguard Worker         vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
247*4bdc9457SAndroid Build Coastguard Worker         vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
248*4bdc9457SAndroid Build Coastguard Worker         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
249*4bdc9457SAndroid Build Coastguard Worker         vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
250*4bdc9457SAndroid Build Coastguard Worker         vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa4), 3);
251*4bdc9457SAndroid Build Coastguard Worker         vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa4), 3);
252*4bdc9457SAndroid Build Coastguard Worker         vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa5), 3);
253*4bdc9457SAndroid Build Coastguard Worker         vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa5), 3);
254*4bdc9457SAndroid Build Coastguard Worker 
255*4bdc9457SAndroid Build Coastguard Worker 
256*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
257*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
258*4bdc9457SAndroid Build Coastguard Worker 
259*4bdc9457SAndroid Build Coastguard Worker         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
260*4bdc9457SAndroid Build Coastguard Worker         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
261*4bdc9457SAndroid Build Coastguard Worker         vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
262*4bdc9457SAndroid Build Coastguard Worker         vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
263*4bdc9457SAndroid Build Coastguard Worker         vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
264*4bdc9457SAndroid Build Coastguard Worker         vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
265*4bdc9457SAndroid Build Coastguard Worker         vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
266*4bdc9457SAndroid Build Coastguard Worker         vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
267*4bdc9457SAndroid Build Coastguard Worker         vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
268*4bdc9457SAndroid Build Coastguard Worker         vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
269*4bdc9457SAndroid Build Coastguard Worker         vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
270*4bdc9457SAndroid Build Coastguard Worker         vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
271*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
272*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
273*4bdc9457SAndroid Build Coastguard Worker 
274*4bdc9457SAndroid Build Coastguard Worker         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
275*4bdc9457SAndroid Build Coastguard Worker         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
276*4bdc9457SAndroid Build Coastguard Worker         vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
277*4bdc9457SAndroid Build Coastguard Worker         vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
278*4bdc9457SAndroid Build Coastguard Worker         vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
279*4bdc9457SAndroid Build Coastguard Worker         vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
280*4bdc9457SAndroid Build Coastguard Worker         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
281*4bdc9457SAndroid Build Coastguard Worker         vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
282*4bdc9457SAndroid Build Coastguard Worker         vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa4), 0);
283*4bdc9457SAndroid Build Coastguard Worker         vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa4), 0);
284*4bdc9457SAndroid Build Coastguard Worker         vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa5), 0);
285*4bdc9457SAndroid Build Coastguard Worker         vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa5), 0);
286*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
287*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
288*4bdc9457SAndroid Build Coastguard Worker 
289*4bdc9457SAndroid Build Coastguard Worker         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
290*4bdc9457SAndroid Build Coastguard Worker         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
291*4bdc9457SAndroid Build Coastguard Worker         vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
292*4bdc9457SAndroid Build Coastguard Worker         vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
293*4bdc9457SAndroid Build Coastguard Worker         vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
294*4bdc9457SAndroid Build Coastguard Worker         vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
295*4bdc9457SAndroid Build Coastguard Worker         vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
296*4bdc9457SAndroid Build Coastguard Worker         vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
297*4bdc9457SAndroid Build Coastguard Worker         vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
298*4bdc9457SAndroid Build Coastguard Worker         vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
299*4bdc9457SAndroid Build Coastguard Worker         vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
300*4bdc9457SAndroid Build Coastguard Worker         vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
301*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
302*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
303*4bdc9457SAndroid Build Coastguard Worker 
304*4bdc9457SAndroid Build Coastguard Worker         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
305*4bdc9457SAndroid Build Coastguard Worker         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
306*4bdc9457SAndroid Build Coastguard Worker         vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
307*4bdc9457SAndroid Build Coastguard Worker         vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
308*4bdc9457SAndroid Build Coastguard Worker         vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
309*4bdc9457SAndroid Build Coastguard Worker         vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
310*4bdc9457SAndroid Build Coastguard Worker         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
311*4bdc9457SAndroid Build Coastguard Worker         vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
312*4bdc9457SAndroid Build Coastguard Worker         vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa4), 1);
313*4bdc9457SAndroid Build Coastguard Worker         vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa4), 1);
314*4bdc9457SAndroid Build Coastguard Worker         vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa5), 1);
315*4bdc9457SAndroid Build Coastguard Worker         vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa5), 1);
316*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
317*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
318*4bdc9457SAndroid Build Coastguard Worker 
319*4bdc9457SAndroid Build Coastguard Worker         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
320*4bdc9457SAndroid Build Coastguard Worker         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
321*4bdc9457SAndroid Build Coastguard Worker         vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
322*4bdc9457SAndroid Build Coastguard Worker         vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
323*4bdc9457SAndroid Build Coastguard Worker         vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
324*4bdc9457SAndroid Build Coastguard Worker         vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
325*4bdc9457SAndroid Build Coastguard Worker         vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
326*4bdc9457SAndroid Build Coastguard Worker         vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
327*4bdc9457SAndroid Build Coastguard Worker         vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
328*4bdc9457SAndroid Build Coastguard Worker         vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
329*4bdc9457SAndroid Build Coastguard Worker         vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
330*4bdc9457SAndroid Build Coastguard Worker         vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
331*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
332*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
333*4bdc9457SAndroid Build Coastguard Worker 
334*4bdc9457SAndroid Build Coastguard Worker         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
335*4bdc9457SAndroid Build Coastguard Worker         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
336*4bdc9457SAndroid Build Coastguard Worker         vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
337*4bdc9457SAndroid Build Coastguard Worker         vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
338*4bdc9457SAndroid Build Coastguard Worker         vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
339*4bdc9457SAndroid Build Coastguard Worker         vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
340*4bdc9457SAndroid Build Coastguard Worker         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
341*4bdc9457SAndroid Build Coastguard Worker         vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
342*4bdc9457SAndroid Build Coastguard Worker         vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa4), 2);
343*4bdc9457SAndroid Build Coastguard Worker         vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa4), 2);
344*4bdc9457SAndroid Build Coastguard Worker         vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa5), 2);
345*4bdc9457SAndroid Build Coastguard Worker         vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa5), 2);
346*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
347*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
348*4bdc9457SAndroid Build Coastguard Worker 
349*4bdc9457SAndroid Build Coastguard Worker         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
350*4bdc9457SAndroid Build Coastguard Worker         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
351*4bdc9457SAndroid Build Coastguard Worker         vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
352*4bdc9457SAndroid Build Coastguard Worker         vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
353*4bdc9457SAndroid Build Coastguard Worker         vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
354*4bdc9457SAndroid Build Coastguard Worker         vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
355*4bdc9457SAndroid Build Coastguard Worker         vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
356*4bdc9457SAndroid Build Coastguard Worker         vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
357*4bdc9457SAndroid Build Coastguard Worker         vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa4), 3);
358*4bdc9457SAndroid Build Coastguard Worker         vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa4), 3);
359*4bdc9457SAndroid Build Coastguard Worker         vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa5), 3);
360*4bdc9457SAndroid Build Coastguard Worker         vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa5), 3);
361*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
362*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
363*4bdc9457SAndroid Build Coastguard Worker 
364*4bdc9457SAndroid Build Coastguard Worker         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
365*4bdc9457SAndroid Build Coastguard Worker         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
366*4bdc9457SAndroid Build Coastguard Worker         vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);
367*4bdc9457SAndroid Build Coastguard Worker         vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);
368*4bdc9457SAndroid Build Coastguard Worker         vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa2), 3);
369*4bdc9457SAndroid Build Coastguard Worker         vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa2), 3);
370*4bdc9457SAndroid Build Coastguard Worker         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3);
371*4bdc9457SAndroid Build Coastguard Worker         vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3);
372*4bdc9457SAndroid Build Coastguard Worker         vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa4), 3);
373*4bdc9457SAndroid Build Coastguard Worker         vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa4), 3);
374*4bdc9457SAndroid Build Coastguard Worker         vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa5), 3);
375*4bdc9457SAndroid Build Coastguard Worker         vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa5), 3);
376*4bdc9457SAndroid Build Coastguard Worker 
377*4bdc9457SAndroid Build Coastguard Worker         k -= 8 * sizeof(int8_t);
378*4bdc9457SAndroid Build Coastguard Worker       }
379*4bdc9457SAndroid Build Coastguard Worker       if XNN_UNLIKELY(k != 0) {
380*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
381*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxa0 = vmovl_s8(va0);
382*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
383*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxa1 = vmovl_s8(va1);
384*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
385*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxa2 = vmovl_s8(va2);
386*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
387*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxa3 = vmovl_s8(va3);
388*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t va4 = vld1_s8(a4); a4 = (const int8_t*) ((uintptr_t) a4 + k);
389*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxa4 = vmovl_s8(va4);
390*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t va5 = vld1_s8(a5); a5 = (const int8_t*) ((uintptr_t) a5 + k);
391*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxa5 = vmovl_s8(va5);
392*4bdc9457SAndroid Build Coastguard Worker 
393*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
394*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
395*4bdc9457SAndroid Build Coastguard Worker         const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
396*4bdc9457SAndroid Build Coastguard Worker         const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
397*4bdc9457SAndroid Build Coastguard Worker 
398*4bdc9457SAndroid Build Coastguard Worker         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
399*4bdc9457SAndroid Build Coastguard Worker         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
400*4bdc9457SAndroid Build Coastguard Worker         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
401*4bdc9457SAndroid Build Coastguard Worker         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
402*4bdc9457SAndroid Build Coastguard Worker         vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
403*4bdc9457SAndroid Build Coastguard Worker         vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
404*4bdc9457SAndroid Build Coastguard Worker         vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
405*4bdc9457SAndroid Build Coastguard Worker         vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
406*4bdc9457SAndroid Build Coastguard Worker         vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
407*4bdc9457SAndroid Build Coastguard Worker         vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
408*4bdc9457SAndroid Build Coastguard Worker         vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
409*4bdc9457SAndroid Build Coastguard Worker         vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
410*4bdc9457SAndroid Build Coastguard Worker         vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
411*4bdc9457SAndroid Build Coastguard Worker         vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
412*4bdc9457SAndroid Build Coastguard Worker         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
413*4bdc9457SAndroid Build Coastguard Worker         vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
414*4bdc9457SAndroid Build Coastguard Worker         vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
415*4bdc9457SAndroid Build Coastguard Worker         vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
416*4bdc9457SAndroid Build Coastguard Worker         vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa4), 0);
417*4bdc9457SAndroid Build Coastguard Worker         vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa4), 0);
418*4bdc9457SAndroid Build Coastguard Worker         vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
419*4bdc9457SAndroid Build Coastguard Worker         vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
420*4bdc9457SAndroid Build Coastguard Worker         vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa5), 0);
421*4bdc9457SAndroid Build Coastguard Worker         vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa5), 0);
422*4bdc9457SAndroid Build Coastguard Worker 
423*4bdc9457SAndroid Build Coastguard Worker         if (k >= 2 * sizeof(int8_t)) {
424*4bdc9457SAndroid Build Coastguard Worker           const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
425*4bdc9457SAndroid Build Coastguard Worker           const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
426*4bdc9457SAndroid Build Coastguard Worker           const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
427*4bdc9457SAndroid Build Coastguard Worker           const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
428*4bdc9457SAndroid Build Coastguard Worker 
429*4bdc9457SAndroid Build Coastguard Worker           vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
430*4bdc9457SAndroid Build Coastguard Worker           vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
431*4bdc9457SAndroid Build Coastguard Worker           vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
432*4bdc9457SAndroid Build Coastguard Worker           vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
433*4bdc9457SAndroid Build Coastguard Worker           vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
434*4bdc9457SAndroid Build Coastguard Worker           vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
435*4bdc9457SAndroid Build Coastguard Worker           vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
436*4bdc9457SAndroid Build Coastguard Worker           vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
437*4bdc9457SAndroid Build Coastguard Worker           vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
438*4bdc9457SAndroid Build Coastguard Worker           vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
439*4bdc9457SAndroid Build Coastguard Worker           vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
440*4bdc9457SAndroid Build Coastguard Worker           vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
441*4bdc9457SAndroid Build Coastguard Worker           vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
442*4bdc9457SAndroid Build Coastguard Worker           vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
443*4bdc9457SAndroid Build Coastguard Worker           vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
444*4bdc9457SAndroid Build Coastguard Worker           vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
445*4bdc9457SAndroid Build Coastguard Worker           vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
446*4bdc9457SAndroid Build Coastguard Worker           vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
447*4bdc9457SAndroid Build Coastguard Worker           vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa4), 1);
448*4bdc9457SAndroid Build Coastguard Worker           vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa4), 1);
449*4bdc9457SAndroid Build Coastguard Worker           vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
450*4bdc9457SAndroid Build Coastguard Worker           vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
451*4bdc9457SAndroid Build Coastguard Worker           vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa5), 1);
452*4bdc9457SAndroid Build Coastguard Worker           vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa5), 1);
453*4bdc9457SAndroid Build Coastguard Worker 
454*4bdc9457SAndroid Build Coastguard Worker           if (k > 2 * sizeof(int8_t)) {
455*4bdc9457SAndroid Build Coastguard Worker             const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
456*4bdc9457SAndroid Build Coastguard Worker             const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
457*4bdc9457SAndroid Build Coastguard Worker             const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
458*4bdc9457SAndroid Build Coastguard Worker             const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
459*4bdc9457SAndroid Build Coastguard Worker 
460*4bdc9457SAndroid Build Coastguard Worker             vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
461*4bdc9457SAndroid Build Coastguard Worker             vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
462*4bdc9457SAndroid Build Coastguard Worker             vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
463*4bdc9457SAndroid Build Coastguard Worker             vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
464*4bdc9457SAndroid Build Coastguard Worker             vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
465*4bdc9457SAndroid Build Coastguard Worker             vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
466*4bdc9457SAndroid Build Coastguard Worker             vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
467*4bdc9457SAndroid Build Coastguard Worker             vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
468*4bdc9457SAndroid Build Coastguard Worker             vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
469*4bdc9457SAndroid Build Coastguard Worker             vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
470*4bdc9457SAndroid Build Coastguard Worker             vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
471*4bdc9457SAndroid Build Coastguard Worker             vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
472*4bdc9457SAndroid Build Coastguard Worker             vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
473*4bdc9457SAndroid Build Coastguard Worker             vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
474*4bdc9457SAndroid Build Coastguard Worker             vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
475*4bdc9457SAndroid Build Coastguard Worker             vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
476*4bdc9457SAndroid Build Coastguard Worker             vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
477*4bdc9457SAndroid Build Coastguard Worker             vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
478*4bdc9457SAndroid Build Coastguard Worker             vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa4), 2);
479*4bdc9457SAndroid Build Coastguard Worker             vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa4), 2);
480*4bdc9457SAndroid Build Coastguard Worker             vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
481*4bdc9457SAndroid Build Coastguard Worker             vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
482*4bdc9457SAndroid Build Coastguard Worker             vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa5), 2);
483*4bdc9457SAndroid Build Coastguard Worker             vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa5), 2);
484*4bdc9457SAndroid Build Coastguard Worker 
485*4bdc9457SAndroid Build Coastguard Worker             if (k >= 4 * sizeof(int8_t)) {
486*4bdc9457SAndroid Build Coastguard Worker               const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
487*4bdc9457SAndroid Build Coastguard Worker               const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
488*4bdc9457SAndroid Build Coastguard Worker               const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
489*4bdc9457SAndroid Build Coastguard Worker               const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
490*4bdc9457SAndroid Build Coastguard Worker 
491*4bdc9457SAndroid Build Coastguard Worker               vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
492*4bdc9457SAndroid Build Coastguard Worker               vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
493*4bdc9457SAndroid Build Coastguard Worker               vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
494*4bdc9457SAndroid Build Coastguard Worker               vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
495*4bdc9457SAndroid Build Coastguard Worker               vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
496*4bdc9457SAndroid Build Coastguard Worker               vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
497*4bdc9457SAndroid Build Coastguard Worker               vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
498*4bdc9457SAndroid Build Coastguard Worker               vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
499*4bdc9457SAndroid Build Coastguard Worker               vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
500*4bdc9457SAndroid Build Coastguard Worker               vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
501*4bdc9457SAndroid Build Coastguard Worker               vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
502*4bdc9457SAndroid Build Coastguard Worker               vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
503*4bdc9457SAndroid Build Coastguard Worker               vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
504*4bdc9457SAndroid Build Coastguard Worker               vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
505*4bdc9457SAndroid Build Coastguard Worker               vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
506*4bdc9457SAndroid Build Coastguard Worker               vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
507*4bdc9457SAndroid Build Coastguard Worker               vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
508*4bdc9457SAndroid Build Coastguard Worker               vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
509*4bdc9457SAndroid Build Coastguard Worker               vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa4), 3);
510*4bdc9457SAndroid Build Coastguard Worker               vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa4), 3);
511*4bdc9457SAndroid Build Coastguard Worker               vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
512*4bdc9457SAndroid Build Coastguard Worker               vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
513*4bdc9457SAndroid Build Coastguard Worker               vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa5), 3);
514*4bdc9457SAndroid Build Coastguard Worker               vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa5), 3);
515*4bdc9457SAndroid Build Coastguard Worker 
516*4bdc9457SAndroid Build Coastguard Worker               if (k > 4 * sizeof(int8_t)) {
517*4bdc9457SAndroid Build Coastguard Worker                 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
518*4bdc9457SAndroid Build Coastguard Worker                 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
519*4bdc9457SAndroid Build Coastguard Worker                 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
520*4bdc9457SAndroid Build Coastguard Worker                 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
521*4bdc9457SAndroid Build Coastguard Worker 
522*4bdc9457SAndroid Build Coastguard Worker                 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
523*4bdc9457SAndroid Build Coastguard Worker                 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
524*4bdc9457SAndroid Build Coastguard Worker                 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
525*4bdc9457SAndroid Build Coastguard Worker                 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
526*4bdc9457SAndroid Build Coastguard Worker                 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
527*4bdc9457SAndroid Build Coastguard Worker                 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
528*4bdc9457SAndroid Build Coastguard Worker                 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
529*4bdc9457SAndroid Build Coastguard Worker                 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
530*4bdc9457SAndroid Build Coastguard Worker                 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
531*4bdc9457SAndroid Build Coastguard Worker                 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
532*4bdc9457SAndroid Build Coastguard Worker                 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
533*4bdc9457SAndroid Build Coastguard Worker                 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
534*4bdc9457SAndroid Build Coastguard Worker                 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
535*4bdc9457SAndroid Build Coastguard Worker                 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
536*4bdc9457SAndroid Build Coastguard Worker                 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
537*4bdc9457SAndroid Build Coastguard Worker                 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
538*4bdc9457SAndroid Build Coastguard Worker                 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
539*4bdc9457SAndroid Build Coastguard Worker                 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
540*4bdc9457SAndroid Build Coastguard Worker                 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa4), 0);
541*4bdc9457SAndroid Build Coastguard Worker                 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa4), 0);
542*4bdc9457SAndroid Build Coastguard Worker                 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
543*4bdc9457SAndroid Build Coastguard Worker                 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
544*4bdc9457SAndroid Build Coastguard Worker                 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa5), 0);
545*4bdc9457SAndroid Build Coastguard Worker                 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa5), 0);
546*4bdc9457SAndroid Build Coastguard Worker 
547*4bdc9457SAndroid Build Coastguard Worker                 if (k >= 6 * sizeof(int8_t)) {
548*4bdc9457SAndroid Build Coastguard Worker                   const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
549*4bdc9457SAndroid Build Coastguard Worker                   const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
550*4bdc9457SAndroid Build Coastguard Worker                   const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
551*4bdc9457SAndroid Build Coastguard Worker                   const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
552*4bdc9457SAndroid Build Coastguard Worker 
553*4bdc9457SAndroid Build Coastguard Worker                   vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
554*4bdc9457SAndroid Build Coastguard Worker                   vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
555*4bdc9457SAndroid Build Coastguard Worker                   vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
556*4bdc9457SAndroid Build Coastguard Worker                   vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
557*4bdc9457SAndroid Build Coastguard Worker                   vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
558*4bdc9457SAndroid Build Coastguard Worker                   vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
559*4bdc9457SAndroid Build Coastguard Worker                   vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
560*4bdc9457SAndroid Build Coastguard Worker                   vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
561*4bdc9457SAndroid Build Coastguard Worker                   vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
562*4bdc9457SAndroid Build Coastguard Worker                   vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
563*4bdc9457SAndroid Build Coastguard Worker                   vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
564*4bdc9457SAndroid Build Coastguard Worker                   vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
565*4bdc9457SAndroid Build Coastguard Worker                   vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
566*4bdc9457SAndroid Build Coastguard Worker                   vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
567*4bdc9457SAndroid Build Coastguard Worker                   vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
568*4bdc9457SAndroid Build Coastguard Worker                   vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
569*4bdc9457SAndroid Build Coastguard Worker                   vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
570*4bdc9457SAndroid Build Coastguard Worker                   vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
571*4bdc9457SAndroid Build Coastguard Worker                   vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa4), 1);
572*4bdc9457SAndroid Build Coastguard Worker                   vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa4), 1);
573*4bdc9457SAndroid Build Coastguard Worker                   vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
574*4bdc9457SAndroid Build Coastguard Worker                   vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
575*4bdc9457SAndroid Build Coastguard Worker                   vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa5), 1);
576*4bdc9457SAndroid Build Coastguard Worker                   vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa5), 1);
577*4bdc9457SAndroid Build Coastguard Worker 
578*4bdc9457SAndroid Build Coastguard Worker                   if (k > 6 * sizeof(int8_t)) {
579*4bdc9457SAndroid Build Coastguard Worker                     const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
580*4bdc9457SAndroid Build Coastguard Worker                     const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
581*4bdc9457SAndroid Build Coastguard Worker                     const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
582*4bdc9457SAndroid Build Coastguard Worker                     const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
583*4bdc9457SAndroid Build Coastguard Worker 
584*4bdc9457SAndroid Build Coastguard Worker                     vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
585*4bdc9457SAndroid Build Coastguard Worker                     vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
586*4bdc9457SAndroid Build Coastguard Worker                     vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
587*4bdc9457SAndroid Build Coastguard Worker                     vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
588*4bdc9457SAndroid Build Coastguard Worker                     vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
589*4bdc9457SAndroid Build Coastguard Worker                     vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
590*4bdc9457SAndroid Build Coastguard Worker                     vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
591*4bdc9457SAndroid Build Coastguard Worker                     vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
592*4bdc9457SAndroid Build Coastguard Worker                     vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
593*4bdc9457SAndroid Build Coastguard Worker                     vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
594*4bdc9457SAndroid Build Coastguard Worker                     vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
595*4bdc9457SAndroid Build Coastguard Worker                     vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
596*4bdc9457SAndroid Build Coastguard Worker                     vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
597*4bdc9457SAndroid Build Coastguard Worker                     vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
598*4bdc9457SAndroid Build Coastguard Worker                     vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
599*4bdc9457SAndroid Build Coastguard Worker                     vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
600*4bdc9457SAndroid Build Coastguard Worker                     vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
601*4bdc9457SAndroid Build Coastguard Worker                     vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
602*4bdc9457SAndroid Build Coastguard Worker                     vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa4), 2);
603*4bdc9457SAndroid Build Coastguard Worker                     vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa4), 2);
604*4bdc9457SAndroid Build Coastguard Worker                     vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
605*4bdc9457SAndroid Build Coastguard Worker                     vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
606*4bdc9457SAndroid Build Coastguard Worker                     vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa5), 2);
607*4bdc9457SAndroid Build Coastguard Worker                     vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa5), 2);
608*4bdc9457SAndroid Build Coastguard Worker                   }
609*4bdc9457SAndroid Build Coastguard Worker                 }
610*4bdc9457SAndroid Build Coastguard Worker               }
611*4bdc9457SAndroid Build Coastguard Worker             }
612*4bdc9457SAndroid Build Coastguard Worker           }
613*4bdc9457SAndroid Build Coastguard Worker         }
614*4bdc9457SAndroid Build Coastguard Worker       }
615*4bdc9457SAndroid Build Coastguard Worker       p -= 6 * sizeof(void*);
616*4bdc9457SAndroid Build Coastguard Worker     } while (p != 0);
617*4bdc9457SAndroid Build Coastguard Worker 
618*4bdc9457SAndroid Build Coastguard Worker     // Post-accumulation work
619*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
620*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
621*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
622*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
623*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
624*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
625*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB);
626*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
627*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
628*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
629*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc2x89AB = vcvtq_f32_s32(vacc2x89AB);
630*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
631*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
632*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
633*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc3x89AB = vcvtq_f32_s32(vacc3x89AB);
634*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
635*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123);
636*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567);
637*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc4x89AB = vcvtq_f32_s32(vacc4x89AB);
638*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc4xCDEF = vcvtq_f32_s32(vacc4xCDEF);
639*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123);
640*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc5x4567 = vcvtq_f32_s32(vacc5x4567);
641*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc5x89AB = vcvtq_f32_s32(vacc5x89AB);
642*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfpacc5xCDEF = vcvtq_f32_s32(vacc5xCDEF);
643*4bdc9457SAndroid Build Coastguard Worker 
644*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
645*4bdc9457SAndroid Build Coastguard Worker     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
646*4bdc9457SAndroid Build Coastguard Worker     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
647*4bdc9457SAndroid Build Coastguard Worker     vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
648*4bdc9457SAndroid Build Coastguard Worker     vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale0123);
649*4bdc9457SAndroid Build Coastguard Worker     vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123);
650*4bdc9457SAndroid Build Coastguard Worker     vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123);
651*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
652*4bdc9457SAndroid Build Coastguard Worker     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
653*4bdc9457SAndroid Build Coastguard Worker     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
654*4bdc9457SAndroid Build Coastguard Worker     vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
655*4bdc9457SAndroid Build Coastguard Worker     vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale4567);
656*4bdc9457SAndroid Build Coastguard Worker     vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567);
657*4bdc9457SAndroid Build Coastguard Worker     vfpacc5x4567 = vmulq_f32(vfpacc5x4567, vscale4567);
658*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
659*4bdc9457SAndroid Build Coastguard Worker     vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale89AB);
660*4bdc9457SAndroid Build Coastguard Worker     vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB);
661*4bdc9457SAndroid Build Coastguard Worker     vfpacc2x89AB = vmulq_f32(vfpacc2x89AB, vscale89AB);
662*4bdc9457SAndroid Build Coastguard Worker     vfpacc3x89AB = vmulq_f32(vfpacc3x89AB, vscale89AB);
663*4bdc9457SAndroid Build Coastguard Worker     vfpacc4x89AB = vmulq_f32(vfpacc4x89AB, vscale89AB);
664*4bdc9457SAndroid Build Coastguard Worker     vfpacc5x89AB = vmulq_f32(vfpacc5x89AB, vscale89AB);
665*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
666*4bdc9457SAndroid Build Coastguard Worker     vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscaleCDEF);
667*4bdc9457SAndroid Build Coastguard Worker     vfpacc1xCDEF = vmulq_f32(vfpacc1xCDEF, vscaleCDEF);
668*4bdc9457SAndroid Build Coastguard Worker     vfpacc2xCDEF = vmulq_f32(vfpacc2xCDEF, vscaleCDEF);
669*4bdc9457SAndroid Build Coastguard Worker     vfpacc3xCDEF = vmulq_f32(vfpacc3xCDEF, vscaleCDEF);
670*4bdc9457SAndroid Build Coastguard Worker     vfpacc4xCDEF = vmulq_f32(vfpacc4xCDEF, vscaleCDEF);
671*4bdc9457SAndroid Build Coastguard Worker     vfpacc5xCDEF = vmulq_f32(vfpacc5xCDEF, vscaleCDEF);
672*4bdc9457SAndroid Build Coastguard Worker 
673*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
674*4bdc9457SAndroid Build Coastguard Worker     vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
675*4bdc9457SAndroid Build Coastguard Worker     vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
676*4bdc9457SAndroid Build Coastguard Worker     vacc0x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x89AB, vmagic_bias));
677*4bdc9457SAndroid Build Coastguard Worker     vacc0xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc0xCDEF, vmagic_bias));
678*4bdc9457SAndroid Build Coastguard Worker     vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
679*4bdc9457SAndroid Build Coastguard Worker     vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
680*4bdc9457SAndroid Build Coastguard Worker     vacc1x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x89AB, vmagic_bias));
681*4bdc9457SAndroid Build Coastguard Worker     vacc1xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc1xCDEF, vmagic_bias));
682*4bdc9457SAndroid Build Coastguard Worker     vacc2x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc2x0123, vmagic_bias));
683*4bdc9457SAndroid Build Coastguard Worker     vacc2x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc2x4567, vmagic_bias));
684*4bdc9457SAndroid Build Coastguard Worker     vacc2x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc2x89AB, vmagic_bias));
685*4bdc9457SAndroid Build Coastguard Worker     vacc2xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc2xCDEF, vmagic_bias));
686*4bdc9457SAndroid Build Coastguard Worker     vacc3x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc3x0123, vmagic_bias));
687*4bdc9457SAndroid Build Coastguard Worker     vacc3x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc3x4567, vmagic_bias));
688*4bdc9457SAndroid Build Coastguard Worker     vacc3x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc3x89AB, vmagic_bias));
689*4bdc9457SAndroid Build Coastguard Worker     vacc3xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc3xCDEF, vmagic_bias));
690*4bdc9457SAndroid Build Coastguard Worker     vacc4x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x0123, vmagic_bias));
691*4bdc9457SAndroid Build Coastguard Worker     vacc4x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x4567, vmagic_bias));
692*4bdc9457SAndroid Build Coastguard Worker     vacc4x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x89AB, vmagic_bias));
693*4bdc9457SAndroid Build Coastguard Worker     vacc4xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc4xCDEF, vmagic_bias));
694*4bdc9457SAndroid Build Coastguard Worker     vacc5x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc5x0123, vmagic_bias));
695*4bdc9457SAndroid Build Coastguard Worker     vacc5x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc5x4567, vmagic_bias));
696*4bdc9457SAndroid Build Coastguard Worker     vacc5x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc5x89AB, vmagic_bias));
697*4bdc9457SAndroid Build Coastguard Worker     vacc5xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc5xCDEF, vmagic_bias));
698*4bdc9457SAndroid Build Coastguard Worker 
699*4bdc9457SAndroid Build Coastguard Worker     const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
700*4bdc9457SAndroid Build Coastguard Worker     vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
701*4bdc9457SAndroid Build Coastguard Worker     vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
702*4bdc9457SAndroid Build Coastguard Worker     vacc0x89AB = vqsubq_s32(vacc0x89AB, vmagic_bias_less_output_zero_point);
703*4bdc9457SAndroid Build Coastguard Worker     vacc0xCDEF = vqsubq_s32(vacc0xCDEF, vmagic_bias_less_output_zero_point);
704*4bdc9457SAndroid Build Coastguard Worker     vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
705*4bdc9457SAndroid Build Coastguard Worker     vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
706*4bdc9457SAndroid Build Coastguard Worker     vacc1x89AB = vqsubq_s32(vacc1x89AB, vmagic_bias_less_output_zero_point);
707*4bdc9457SAndroid Build Coastguard Worker     vacc1xCDEF = vqsubq_s32(vacc1xCDEF, vmagic_bias_less_output_zero_point);
708*4bdc9457SAndroid Build Coastguard Worker     vacc2x0123 = vqsubq_s32(vacc2x0123, vmagic_bias_less_output_zero_point);
709*4bdc9457SAndroid Build Coastguard Worker     vacc2x4567 = vqsubq_s32(vacc2x4567, vmagic_bias_less_output_zero_point);
710*4bdc9457SAndroid Build Coastguard Worker     vacc2x89AB = vqsubq_s32(vacc2x89AB, vmagic_bias_less_output_zero_point);
711*4bdc9457SAndroid Build Coastguard Worker     vacc2xCDEF = vqsubq_s32(vacc2xCDEF, vmagic_bias_less_output_zero_point);
712*4bdc9457SAndroid Build Coastguard Worker     vacc3x0123 = vqsubq_s32(vacc3x0123, vmagic_bias_less_output_zero_point);
713*4bdc9457SAndroid Build Coastguard Worker     vacc3x4567 = vqsubq_s32(vacc3x4567, vmagic_bias_less_output_zero_point);
714*4bdc9457SAndroid Build Coastguard Worker     vacc3x89AB = vqsubq_s32(vacc3x89AB, vmagic_bias_less_output_zero_point);
715*4bdc9457SAndroid Build Coastguard Worker     vacc3xCDEF = vqsubq_s32(vacc3xCDEF, vmagic_bias_less_output_zero_point);
716*4bdc9457SAndroid Build Coastguard Worker     vacc4x0123 = vqsubq_s32(vacc4x0123, vmagic_bias_less_output_zero_point);
717*4bdc9457SAndroid Build Coastguard Worker     vacc4x4567 = vqsubq_s32(vacc4x4567, vmagic_bias_less_output_zero_point);
718*4bdc9457SAndroid Build Coastguard Worker     vacc4x89AB = vqsubq_s32(vacc4x89AB, vmagic_bias_less_output_zero_point);
719*4bdc9457SAndroid Build Coastguard Worker     vacc4xCDEF = vqsubq_s32(vacc4xCDEF, vmagic_bias_less_output_zero_point);
720*4bdc9457SAndroid Build Coastguard Worker     vacc5x0123 = vqsubq_s32(vacc5x0123, vmagic_bias_less_output_zero_point);
721*4bdc9457SAndroid Build Coastguard Worker     vacc5x4567 = vqsubq_s32(vacc5x4567, vmagic_bias_less_output_zero_point);
722*4bdc9457SAndroid Build Coastguard Worker     vacc5x89AB = vqsubq_s32(vacc5x89AB, vmagic_bias_less_output_zero_point);
723*4bdc9457SAndroid Build Coastguard Worker     vacc5xCDEF = vqsubq_s32(vacc5xCDEF, vmagic_bias_less_output_zero_point);
724*4bdc9457SAndroid Build Coastguard Worker 
725*4bdc9457SAndroid Build Coastguard Worker #if XNN_ARCH_ARM64
726*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
727*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
728*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
729*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
730*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
731*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
732*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
733*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
734*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc4x01234567 = vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567);
735*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc4x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF);
736*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc5x01234567 = vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567);
737*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc5x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc5x89AB), vacc5xCDEF);
738*4bdc9457SAndroid Build Coastguard Worker 
739*4bdc9457SAndroid Build Coastguard Worker 
740*4bdc9457SAndroid Build Coastguard Worker     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
741*4bdc9457SAndroid Build Coastguard Worker     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
742*4bdc9457SAndroid Build Coastguard Worker     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
743*4bdc9457SAndroid Build Coastguard Worker     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
744*4bdc9457SAndroid Build Coastguard Worker     int8x16_t vout4x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc4x89ABCDEF);
745*4bdc9457SAndroid Build Coastguard Worker     int8x16_t vout5x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc5x01234567), vacc5x89ABCDEF);
746*4bdc9457SAndroid Build Coastguard Worker #else
747*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
748*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
749*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
750*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
751*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
752*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
753*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
754*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
755*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc4x01234567 = vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567));
756*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc4x89ABCDEF = vcombine_s16(vqmovn_s32(vacc4x89AB), vqmovn_s32(vacc4xCDEF));
757*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc5x01234567 = vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567));
758*4bdc9457SAndroid Build Coastguard Worker     int16x8_t vacc5x89ABCDEF = vcombine_s16(vqmovn_s32(vacc5x89AB), vqmovn_s32(vacc5xCDEF));
759*4bdc9457SAndroid Build Coastguard Worker 
760*4bdc9457SAndroid Build Coastguard Worker 
761*4bdc9457SAndroid Build Coastguard Worker     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
762*4bdc9457SAndroid Build Coastguard Worker     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
763*4bdc9457SAndroid Build Coastguard Worker     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
764*4bdc9457SAndroid Build Coastguard Worker     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
765*4bdc9457SAndroid Build Coastguard Worker     int8x16_t vout4x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc4x89ABCDEF));
766*4bdc9457SAndroid Build Coastguard Worker     int8x16_t vout5x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc5x01234567), vqmovn_s16(vacc5x89ABCDEF));
767*4bdc9457SAndroid Build Coastguard Worker #endif
768*4bdc9457SAndroid Build Coastguard Worker 
769*4bdc9457SAndroid Build Coastguard Worker     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neon.output_min);
770*4bdc9457SAndroid Build Coastguard Worker     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
771*4bdc9457SAndroid Build Coastguard Worker     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
772*4bdc9457SAndroid Build Coastguard Worker     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
773*4bdc9457SAndroid Build Coastguard Worker     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
774*4bdc9457SAndroid Build Coastguard Worker     vout4x0123456789ABCDEF = vmaxq_s8(vout4x0123456789ABCDEF, voutput_min);
775*4bdc9457SAndroid Build Coastguard Worker     vout5x0123456789ABCDEF = vmaxq_s8(vout5x0123456789ABCDEF, voutput_min);
776*4bdc9457SAndroid Build Coastguard Worker 
777*4bdc9457SAndroid Build Coastguard Worker     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neon.output_max);
778*4bdc9457SAndroid Build Coastguard Worker     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
779*4bdc9457SAndroid Build Coastguard Worker     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
780*4bdc9457SAndroid Build Coastguard Worker     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
781*4bdc9457SAndroid Build Coastguard Worker     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
782*4bdc9457SAndroid Build Coastguard Worker     vout4x0123456789ABCDEF = vminq_s8(vout4x0123456789ABCDEF, voutput_max);
783*4bdc9457SAndroid Build Coastguard Worker     vout5x0123456789ABCDEF = vminq_s8(vout5x0123456789ABCDEF, voutput_max);
784*4bdc9457SAndroid Build Coastguard Worker 
785*4bdc9457SAndroid Build Coastguard Worker     if (nc >= 16) {
786*4bdc9457SAndroid Build Coastguard Worker       vst1q_s8(c5 + 0, vout5x0123456789ABCDEF);
787*4bdc9457SAndroid Build Coastguard Worker       vst1q_s8(c4 + 0, vout4x0123456789ABCDEF);
788*4bdc9457SAndroid Build Coastguard Worker       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
789*4bdc9457SAndroid Build Coastguard Worker       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
790*4bdc9457SAndroid Build Coastguard Worker       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
791*4bdc9457SAndroid Build Coastguard Worker       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
792*4bdc9457SAndroid Build Coastguard Worker 
793*4bdc9457SAndroid Build Coastguard Worker       c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
794*4bdc9457SAndroid Build Coastguard Worker       c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
795*4bdc9457SAndroid Build Coastguard Worker       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
796*4bdc9457SAndroid Build Coastguard Worker       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
797*4bdc9457SAndroid Build Coastguard Worker       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
798*4bdc9457SAndroid Build Coastguard Worker       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
799*4bdc9457SAndroid Build Coastguard Worker 
800*4bdc9457SAndroid Build Coastguard Worker       a = (const int8_t**restrict) ((uintptr_t) a - ks);
801*4bdc9457SAndroid Build Coastguard Worker 
802*4bdc9457SAndroid Build Coastguard Worker       nc -= 16;
803*4bdc9457SAndroid Build Coastguard Worker     } else {
804*4bdc9457SAndroid Build Coastguard Worker       int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vget_low_s8(vout4x0123456789ABCDEF), vget_low_s8(vout5x0123456789ABCDEF));
805*4bdc9457SAndroid Build Coastguard Worker       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
806*4bdc9457SAndroid Build Coastguard Worker       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
807*4bdc9457SAndroid Build Coastguard Worker       if (nc & 8) {
808*4bdc9457SAndroid Build Coastguard Worker         vst1_s8(c5, vget_high_s8(vout4x01234567_5x01234567)); c5 += 8;
809*4bdc9457SAndroid Build Coastguard Worker         vst1_s8(c4, vget_low_s8(vout4x01234567_5x01234567)); c4 += 8;
810*4bdc9457SAndroid Build Coastguard Worker         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
811*4bdc9457SAndroid Build Coastguard Worker         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
812*4bdc9457SAndroid Build Coastguard Worker         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
813*4bdc9457SAndroid Build Coastguard Worker         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
814*4bdc9457SAndroid Build Coastguard Worker         vout4x01234567_5x01234567 = vcombine_s8(vget_high_s8(vout4x0123456789ABCDEF), vget_high_s8(vout5x0123456789ABCDEF));
815*4bdc9457SAndroid Build Coastguard Worker         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
816*4bdc9457SAndroid Build Coastguard Worker         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
817*4bdc9457SAndroid Build Coastguard Worker       }
818*4bdc9457SAndroid Build Coastguard Worker       if (nc & 4) {
819*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_u32((void*) c5, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4;
820*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_u32((void*) c4, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4;
821*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
822*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
823*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
824*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
825*4bdc9457SAndroid Build Coastguard Worker         vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
826*4bdc9457SAndroid Build Coastguard Worker         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
827*4bdc9457SAndroid Build Coastguard Worker         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
828*4bdc9457SAndroid Build Coastguard Worker       }
829*4bdc9457SAndroid Build Coastguard Worker       if (nc & 2) {
830*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_u16((void*) c5, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2;
831*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_u16((void*) c4, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2;
832*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
833*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
834*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
835*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
836*4bdc9457SAndroid Build Coastguard Worker         vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
837*4bdc9457SAndroid Build Coastguard Worker         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
838*4bdc9457SAndroid Build Coastguard Worker         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
839*4bdc9457SAndroid Build Coastguard Worker       }
840*4bdc9457SAndroid Build Coastguard Worker       if (nc & 1) {
841*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8);
842*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0);
843*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
844*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
845*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
846*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
847*4bdc9457SAndroid Build Coastguard Worker       }
848*4bdc9457SAndroid Build Coastguard Worker 
849*4bdc9457SAndroid Build Coastguard Worker       nc = 0;
850*4bdc9457SAndroid Build Coastguard Worker     }
851*4bdc9457SAndroid Build Coastguard Worker   } while (nc != 0);
852*4bdc9457SAndroid Build Coastguard Worker }
853