1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker // Template: src/qs8-gemm/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/gemm.h>
16*4bdc9457SAndroid Build Coastguard Worker
17*4bdc9457SAndroid Build Coastguard Worker
xnn_qu8_gemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18*4bdc9457SAndroid Build Coastguard Worker void xnn_qu8_gemm_minmax_rndnu_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 const uint8_t* restrict a,
23*4bdc9457SAndroid Build Coastguard Worker size_t a_stride,
24*4bdc9457SAndroid Build Coastguard Worker const void* restrict w,
25*4bdc9457SAndroid Build Coastguard Worker uint8_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 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29*4bdc9457SAndroid Build Coastguard Worker {
30*4bdc9457SAndroid Build Coastguard Worker assert(mr != 0);
31*4bdc9457SAndroid Build Coastguard Worker assert(mr <= 6);
32*4bdc9457SAndroid Build Coastguard Worker assert(nc != 0);
33*4bdc9457SAndroid Build Coastguard Worker assert(kc != 0);
34*4bdc9457SAndroid Build Coastguard Worker assert(kc % sizeof(uint8_t) == 0);
35*4bdc9457SAndroid Build Coastguard Worker assert(a != NULL);
36*4bdc9457SAndroid Build Coastguard Worker assert(w != NULL);
37*4bdc9457SAndroid Build Coastguard Worker assert(c != NULL);
38*4bdc9457SAndroid Build Coastguard Worker
39*4bdc9457SAndroid Build Coastguard Worker const uint8_t* a0 = a;
40*4bdc9457SAndroid Build Coastguard Worker uint8_t* c0 = c;
41*4bdc9457SAndroid Build Coastguard Worker const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
42*4bdc9457SAndroid Build Coastguard Worker uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
43*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(mr < 2) {
44*4bdc9457SAndroid Build Coastguard Worker a1 = a0;
45*4bdc9457SAndroid Build Coastguard Worker c1 = c0;
46*4bdc9457SAndroid Build Coastguard Worker }
47*4bdc9457SAndroid Build Coastguard Worker const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
48*4bdc9457SAndroid Build Coastguard Worker uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
49*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(mr <= 2) {
50*4bdc9457SAndroid Build Coastguard Worker a2 = a1;
51*4bdc9457SAndroid Build Coastguard Worker c2 = c1;
52*4bdc9457SAndroid Build Coastguard Worker }
53*4bdc9457SAndroid Build Coastguard Worker const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
54*4bdc9457SAndroid Build Coastguard Worker uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
55*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(mr < 4) {
56*4bdc9457SAndroid Build Coastguard Worker a3 = a2;
57*4bdc9457SAndroid Build Coastguard Worker c3 = c2;
58*4bdc9457SAndroid Build Coastguard Worker }
59*4bdc9457SAndroid Build Coastguard Worker const uint8_t* a4 = (const uint8_t*) ((uintptr_t) a3 + a_stride);
60*4bdc9457SAndroid Build Coastguard Worker uint8_t* c4 = (uint8_t*) ((uintptr_t) c3 + cm_stride);
61*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(mr <= 4) {
62*4bdc9457SAndroid Build Coastguard Worker a4 = a3;
63*4bdc9457SAndroid Build Coastguard Worker c4 = c3;
64*4bdc9457SAndroid Build Coastguard Worker }
65*4bdc9457SAndroid Build Coastguard Worker const uint8_t* a5 = (const uint8_t*) ((uintptr_t) a4 + a_stride);
66*4bdc9457SAndroid Build Coastguard Worker uint8_t* c5 = (uint8_t*) ((uintptr_t) c4 + cm_stride);
67*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(mr != 6) {
68*4bdc9457SAndroid Build Coastguard Worker a5 = a4;
69*4bdc9457SAndroid Build Coastguard Worker c5 = c4;
70*4bdc9457SAndroid Build Coastguard Worker }
71*4bdc9457SAndroid Build Coastguard Worker
72*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb_zero_point = vld1_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
73*4bdc9457SAndroid Build Coastguard Worker do {
74*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
75*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
76*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
77*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
78*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc1x0123 = vacc0x0123;
79*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc1x4567 = vacc0x4567;
80*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc1x89AB = vacc0x89AB;
81*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc1xCDEF = vacc0xCDEF;
82*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc2x0123 = vacc0x0123;
83*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc2x4567 = vacc0x4567;
84*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc2x89AB = vacc0x89AB;
85*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc2xCDEF = vacc0xCDEF;
86*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc3x0123 = vacc0x0123;
87*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc3x4567 = vacc0x4567;
88*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc3x89AB = vacc0x89AB;
89*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc3xCDEF = vacc0xCDEF;
90*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc4x0123 = vacc0x0123;
91*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc4x4567 = vacc0x4567;
92*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc4x89AB = vacc0x89AB;
93*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc4xCDEF = vacc0xCDEF;
94*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc5x0123 = vacc0x0123;
95*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc5x4567 = vacc0x4567;
96*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc5x89AB = vacc0x89AB;
97*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc5xCDEF = vacc0xCDEF;
98*4bdc9457SAndroid Build Coastguard Worker
99*4bdc9457SAndroid Build Coastguard Worker size_t k = kc;
100*4bdc9457SAndroid Build Coastguard Worker while (k >= 8 * sizeof(uint8_t)) {
101*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t va0 = vld1_u8(a0); a0 += 8;
102*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
103*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t va1 = vld1_u8(a1); a1 += 8;
104*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxa1 = vreinterpretq_s16_u16(vmovl_u8(va1));
105*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t va2 = vld1_u8(a2); a2 += 8;
106*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxa2 = vreinterpretq_s16_u16(vmovl_u8(va2));
107*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t va3 = vld1_u8(a3); a3 += 8;
108*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxa3 = vreinterpretq_s16_u16(vmovl_u8(va3));
109*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t va4 = vld1_u8(a4); a4 += 8;
110*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxa4 = vreinterpretq_s16_u16(vmovl_u8(va4));
111*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t va5 = vld1_u8(a5); a5 += 8;
112*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxa5 = vreinterpretq_s16_u16(vmovl_u8(va5));
113*4bdc9457SAndroid Build Coastguard Worker
114*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
115*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb01234567c0 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c0, vb_zero_point));
116*4bdc9457SAndroid Build Coastguard Worker
117*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
118*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
119*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
120*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
121*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
122*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
123*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
124*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
125*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
126*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
127*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
128*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
129*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb89ABCDEFc0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
130*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb89ABCDEFc0 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc0, vb_zero_point));
131*4bdc9457SAndroid Build Coastguard Worker
132*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
133*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
134*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
135*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
136*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
137*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
138*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
139*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
140*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa4), 0);
141*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa4), 0);
142*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa5), 0);
143*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa5), 0);
144*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb01234567c1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
145*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb01234567c1 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c1, vb_zero_point));
146*4bdc9457SAndroid Build Coastguard Worker
147*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
148*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
149*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
150*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
151*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
152*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
153*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
154*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
155*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
156*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
157*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
158*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
159*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
160*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb89ABCDEFc1 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc1, vb_zero_point));
161*4bdc9457SAndroid Build Coastguard Worker
162*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
163*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
164*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
165*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
166*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
167*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
168*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
169*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
170*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa4), 1);
171*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa4), 1);
172*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa5), 1);
173*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa5), 1);
174*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb01234567c2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
175*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point));
176*4bdc9457SAndroid Build Coastguard Worker
177*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
178*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
179*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
180*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
181*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
182*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
183*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
184*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
185*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
186*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
187*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
188*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
189*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb89ABCDEFc2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
190*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb89ABCDEFc2 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc2, vb_zero_point));
191*4bdc9457SAndroid Build Coastguard Worker
192*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
193*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
194*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
195*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
196*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
197*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
198*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
199*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
200*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa4), 2);
201*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa4), 2);
202*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa5), 2);
203*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa5), 2);
204*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
205*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb01234567c3 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c3, vb_zero_point));
206*4bdc9457SAndroid Build Coastguard Worker
207*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
208*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
209*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
210*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
211*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
212*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
213*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
214*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
215*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
216*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
217*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
218*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
219*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb89ABCDEFc3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
220*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb89ABCDEFc3 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc3, vb_zero_point));
221*4bdc9457SAndroid Build Coastguard Worker
222*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
223*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
224*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
225*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
226*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
227*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
228*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
229*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
230*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa4), 3);
231*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa4), 3);
232*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa5), 3);
233*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa5), 3);
234*4bdc9457SAndroid Build Coastguard Worker
235*4bdc9457SAndroid Build Coastguard Worker
236*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb01234567c4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
237*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point));
238*4bdc9457SAndroid Build Coastguard Worker
239*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
240*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
241*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
242*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
243*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
244*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
245*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
246*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
247*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
248*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
249*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
250*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
251*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb89ABCDEFc4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
252*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb89ABCDEFc4 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc4, vb_zero_point));
253*4bdc9457SAndroid Build Coastguard Worker
254*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
255*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
256*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
257*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
258*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
259*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
260*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
261*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
262*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa4), 0);
263*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa4), 0);
264*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa5), 0);
265*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa5), 0);
266*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb01234567c5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
267*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb01234567c5 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c5, vb_zero_point));
268*4bdc9457SAndroid Build Coastguard Worker
269*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
270*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
271*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
272*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
273*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
274*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
275*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
276*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
277*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
278*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
279*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
280*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
281*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb89ABCDEFc5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
282*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb89ABCDEFc5 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc5, vb_zero_point));
283*4bdc9457SAndroid Build Coastguard Worker
284*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
285*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
286*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
287*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
288*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
289*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
290*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
291*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
292*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa4), 1);
293*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa4), 1);
294*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa5), 1);
295*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa5), 1);
296*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb01234567c6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
297*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));
298*4bdc9457SAndroid Build Coastguard Worker
299*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
300*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
301*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
302*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
303*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
304*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
305*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
306*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
307*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
308*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
309*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
310*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
311*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb89ABCDEFc6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
312*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb89ABCDEFc6 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc6, vb_zero_point));
313*4bdc9457SAndroid Build Coastguard Worker
314*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
315*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
316*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
317*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
318*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
319*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
320*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
321*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
322*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa4), 2);
323*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa4), 2);
324*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa5), 2);
325*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa5), 2);
326*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb01234567c7 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
327*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb01234567c7 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c7, vb_zero_point));
328*4bdc9457SAndroid Build Coastguard Worker
329*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
330*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
331*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
332*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
333*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
334*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
335*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
336*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
337*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa4), 3);
338*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa4), 3);
339*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa5), 3);
340*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa5), 3);
341*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb89ABCDEFc7 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
342*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb89ABCDEFc7 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc7, vb_zero_point));
343*4bdc9457SAndroid Build Coastguard Worker
344*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
345*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
346*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);
347*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);
348*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa2), 3);
349*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa2), 3);
350*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3);
351*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3);
352*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa4), 3);
353*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa4), 3);
354*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa5), 3);
355*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa5), 3);
356*4bdc9457SAndroid Build Coastguard Worker
357*4bdc9457SAndroid Build Coastguard Worker k -= 8 * sizeof(uint8_t);
358*4bdc9457SAndroid Build Coastguard Worker }
359*4bdc9457SAndroid Build Coastguard Worker if XNN_UNLIKELY(k != 0) {
360*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t va0 = vld1_u8(a0); a0 = (const uint8_t*) ((uintptr_t) a0 + k);
361*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
362*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t va1 = vld1_u8(a1); a1 = (const uint8_t*) ((uintptr_t) a1 + k);
363*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxa1 = vreinterpretq_s16_u16(vmovl_u8(va1));
364*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t va2 = vld1_u8(a2); a2 = (const uint8_t*) ((uintptr_t) a2 + k);
365*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxa2 = vreinterpretq_s16_u16(vmovl_u8(va2));
366*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t va3 = vld1_u8(a3); a3 = (const uint8_t*) ((uintptr_t) a3 + k);
367*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxa3 = vreinterpretq_s16_u16(vmovl_u8(va3));
368*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t va4 = vld1_u8(a4); a4 = (const uint8_t*) ((uintptr_t) a4 + k);
369*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxa4 = vreinterpretq_s16_u16(vmovl_u8(va4));
370*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t va5 = vld1_u8(a5); a5 = (const uint8_t*) ((uintptr_t) a5 + k);
371*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxa5 = vreinterpretq_s16_u16(vmovl_u8(va5));
372*4bdc9457SAndroid Build Coastguard Worker
373*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
374*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb01234567c0 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c0, vb_zero_point));
375*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb89ABCDEFc0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
376*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb89ABCDEFc0 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc0, vb_zero_point));
377*4bdc9457SAndroid Build Coastguard Worker
378*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
379*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
380*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
381*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
382*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
383*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
384*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
385*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
386*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
387*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
388*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
389*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
390*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
391*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
392*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
393*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
394*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
395*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
396*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa4), 0);
397*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa4), 0);
398*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
399*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
400*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa5), 0);
401*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa5), 0);
402*4bdc9457SAndroid Build Coastguard Worker
403*4bdc9457SAndroid Build Coastguard Worker if (k >= 2 * sizeof(uint8_t)) {
404*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb01234567c1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
405*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb01234567c1 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c1, vb_zero_point));
406*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
407*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb89ABCDEFc1 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc1, vb_zero_point));
408*4bdc9457SAndroid Build Coastguard Worker
409*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
410*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
411*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
412*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
413*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
414*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
415*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
416*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
417*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
418*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
419*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
420*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
421*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
422*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
423*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
424*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
425*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
426*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
427*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa4), 1);
428*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa4), 1);
429*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
430*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
431*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa5), 1);
432*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa5), 1);
433*4bdc9457SAndroid Build Coastguard Worker
434*4bdc9457SAndroid Build Coastguard Worker if (k > 2 * sizeof(uint8_t)) {
435*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb01234567c2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
436*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point));
437*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb89ABCDEFc2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
438*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb89ABCDEFc2 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc2, vb_zero_point));
439*4bdc9457SAndroid Build Coastguard Worker
440*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
441*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
442*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
443*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
444*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
445*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
446*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
447*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
448*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
449*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
450*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
451*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
452*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
453*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
454*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
455*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
456*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
457*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
458*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa4), 2);
459*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa4), 2);
460*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
461*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
462*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa5), 2);
463*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa5), 2);
464*4bdc9457SAndroid Build Coastguard Worker
465*4bdc9457SAndroid Build Coastguard Worker if (k >= 4 * sizeof(uint8_t)) {
466*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
467*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb01234567c3 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c3, vb_zero_point));
468*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb89ABCDEFc3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
469*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb89ABCDEFc3 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc3, vb_zero_point));
470*4bdc9457SAndroid Build Coastguard Worker
471*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
472*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
473*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
474*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
475*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
476*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
477*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
478*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
479*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
480*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
481*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
482*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
483*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
484*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
485*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
486*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
487*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
488*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
489*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa4), 3);
490*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa4), 3);
491*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
492*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
493*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa5), 3);
494*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa5), 3);
495*4bdc9457SAndroid Build Coastguard Worker
496*4bdc9457SAndroid Build Coastguard Worker if (k > 4 * sizeof(uint8_t)) {
497*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb01234567c4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
498*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point));
499*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb89ABCDEFc4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
500*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb89ABCDEFc4 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc4, vb_zero_point));
501*4bdc9457SAndroid Build Coastguard Worker
502*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
503*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
504*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
505*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
506*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
507*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
508*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
509*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
510*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
511*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
512*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
513*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
514*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
515*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
516*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
517*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
518*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
519*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
520*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa4), 0);
521*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa4), 0);
522*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
523*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
524*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa5), 0);
525*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa5), 0);
526*4bdc9457SAndroid Build Coastguard Worker
527*4bdc9457SAndroid Build Coastguard Worker if (k >= 6 * sizeof(uint8_t)) {
528*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb01234567c5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
529*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb01234567c5 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c5, vb_zero_point));
530*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb89ABCDEFc5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
531*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb89ABCDEFc5 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc5, vb_zero_point));
532*4bdc9457SAndroid Build Coastguard Worker
533*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
534*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
535*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
536*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
537*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
538*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
539*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
540*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
541*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
542*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
543*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
544*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
545*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
546*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
547*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
548*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
549*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
550*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
551*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa4), 1);
552*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa4), 1);
553*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
554*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
555*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa5), 1);
556*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa5), 1);
557*4bdc9457SAndroid Build Coastguard Worker
558*4bdc9457SAndroid Build Coastguard Worker if (k > 6 * sizeof(uint8_t)) {
559*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb01234567c6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
560*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));
561*4bdc9457SAndroid Build Coastguard Worker const uint8x8_t vb89ABCDEFc6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
562*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vxb89ABCDEFc6 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc6, vb_zero_point));
563*4bdc9457SAndroid Build Coastguard Worker
564*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
565*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
566*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
567*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
568*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
569*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
570*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
571*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
572*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
573*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
574*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
575*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
576*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
577*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
578*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
579*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
580*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
581*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
582*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa4), 2);
583*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa4), 2);
584*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
585*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
586*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa5), 2);
587*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa5), 2);
588*4bdc9457SAndroid Build Coastguard Worker }
589*4bdc9457SAndroid Build Coastguard Worker }
590*4bdc9457SAndroid Build Coastguard Worker }
591*4bdc9457SAndroid Build Coastguard Worker }
592*4bdc9457SAndroid Build Coastguard Worker }
593*4bdc9457SAndroid Build Coastguard Worker }
594*4bdc9457SAndroid Build Coastguard Worker }
595*4bdc9457SAndroid Build Coastguard Worker
596*4bdc9457SAndroid Build Coastguard Worker const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
597*4bdc9457SAndroid Build Coastguard Worker const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
598*4bdc9457SAndroid Build Coastguard Worker const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
599*4bdc9457SAndroid Build Coastguard Worker
600*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
601*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
602*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
603*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
604*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
605*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
606*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
607*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
608*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
609*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
610*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
611*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
612*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
613*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
614*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
615*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
616*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vqshlq_s32(vacc4x0123, vright_pre_shift);
617*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vqshlq_s32(vacc4x4567, vright_pre_shift);
618*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vqshlq_s32(vacc4x89AB, vright_pre_shift);
619*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vqshlq_s32(vacc4xCDEF, vright_pre_shift);
620*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vqshlq_s32(vacc5x0123, vright_pre_shift);
621*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vqshlq_s32(vacc5x4567, vright_pre_shift);
622*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vqshlq_s32(vacc5x89AB, vright_pre_shift);
623*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vqshlq_s32(vacc5xCDEF, vright_pre_shift);
624*4bdc9457SAndroid Build Coastguard Worker
625*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
626*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
627*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
628*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
629*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
630*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
631*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
632*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
633*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
634*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
635*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
636*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
637*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
638*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
639*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
640*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
641*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vqdmulhq_s32(vacc4x0123, vmultiplier);
642*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vqdmulhq_s32(vacc4x4567, vmultiplier);
643*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vqdmulhq_s32(vacc4x89AB, vmultiplier);
644*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vqdmulhq_s32(vacc4xCDEF, vmultiplier);
645*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vqdmulhq_s32(vacc5x0123, vmultiplier);
646*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vqdmulhq_s32(vacc5x4567, vmultiplier);
647*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vqdmulhq_s32(vacc5x89AB, vmultiplier);
648*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vqdmulhq_s32(vacc5xCDEF, vmultiplier);
649*4bdc9457SAndroid Build Coastguard Worker
650*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
651*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
652*4bdc9457SAndroid Build Coastguard Worker vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
653*4bdc9457SAndroid Build Coastguard Worker vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
654*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
655*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
656*4bdc9457SAndroid Build Coastguard Worker vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
657*4bdc9457SAndroid Build Coastguard Worker vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
658*4bdc9457SAndroid Build Coastguard Worker vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
659*4bdc9457SAndroid Build Coastguard Worker vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
660*4bdc9457SAndroid Build Coastguard Worker vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
661*4bdc9457SAndroid Build Coastguard Worker vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
662*4bdc9457SAndroid Build Coastguard Worker vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
663*4bdc9457SAndroid Build Coastguard Worker vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
664*4bdc9457SAndroid Build Coastguard Worker vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
665*4bdc9457SAndroid Build Coastguard Worker vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
666*4bdc9457SAndroid Build Coastguard Worker vacc4x0123 = vrshlq_s32(vacc4x0123, vright_post_shift);
667*4bdc9457SAndroid Build Coastguard Worker vacc4x4567 = vrshlq_s32(vacc4x4567, vright_post_shift);
668*4bdc9457SAndroid Build Coastguard Worker vacc4x89AB = vrshlq_s32(vacc4x89AB, vright_post_shift);
669*4bdc9457SAndroid Build Coastguard Worker vacc4xCDEF = vrshlq_s32(vacc4xCDEF, vright_post_shift);
670*4bdc9457SAndroid Build Coastguard Worker vacc5x0123 = vrshlq_s32(vacc5x0123, vright_post_shift);
671*4bdc9457SAndroid Build Coastguard Worker vacc5x4567 = vrshlq_s32(vacc5x4567, vright_post_shift);
672*4bdc9457SAndroid Build Coastguard Worker vacc5x89AB = vrshlq_s32(vacc5x89AB, vright_post_shift);
673*4bdc9457SAndroid Build Coastguard Worker vacc5xCDEF = vrshlq_s32(vacc5xCDEF, vright_post_shift);
674*4bdc9457SAndroid Build Coastguard Worker
675*4bdc9457SAndroid Build Coastguard Worker const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
676*4bdc9457SAndroid Build Coastguard Worker #if XNN_ARCH_ARM64
677*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
678*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
679*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
680*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
681*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
682*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
683*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
684*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
685*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc4x01234567 = vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567);
686*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc4x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF);
687*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc5x01234567 = vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567);
688*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc5x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc5x89AB), vacc5xCDEF);
689*4bdc9457SAndroid Build Coastguard Worker
690*4bdc9457SAndroid Build Coastguard Worker vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
691*4bdc9457SAndroid Build Coastguard Worker vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
692*4bdc9457SAndroid Build Coastguard Worker vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
693*4bdc9457SAndroid Build Coastguard Worker vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
694*4bdc9457SAndroid Build Coastguard Worker vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
695*4bdc9457SAndroid Build Coastguard Worker vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
696*4bdc9457SAndroid Build Coastguard Worker vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
697*4bdc9457SAndroid Build Coastguard Worker vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
698*4bdc9457SAndroid Build Coastguard Worker vacc4x01234567 = vqaddq_s16(vacc4x01234567, voutput_zero_point);
699*4bdc9457SAndroid Build Coastguard Worker vacc4x89ABCDEF = vqaddq_s16(vacc4x89ABCDEF, voutput_zero_point);
700*4bdc9457SAndroid Build Coastguard Worker vacc5x01234567 = vqaddq_s16(vacc5x01234567, voutput_zero_point);
701*4bdc9457SAndroid Build Coastguard Worker vacc5x89ABCDEF = vqaddq_s16(vacc5x89ABCDEF, voutput_zero_point);
702*4bdc9457SAndroid Build Coastguard Worker
703*4bdc9457SAndroid Build Coastguard Worker uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
704*4bdc9457SAndroid Build Coastguard Worker uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
705*4bdc9457SAndroid Build Coastguard Worker uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
706*4bdc9457SAndroid Build Coastguard Worker uint8x16_t vout3x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc3x01234567), vacc3x89ABCDEF);
707*4bdc9457SAndroid Build Coastguard Worker uint8x16_t vout4x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc4x01234567), vacc4x89ABCDEF);
708*4bdc9457SAndroid Build Coastguard Worker uint8x16_t vout5x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc5x01234567), vacc5x89ABCDEF);
709*4bdc9457SAndroid Build Coastguard Worker #else
710*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
711*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
712*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
713*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
714*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
715*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
716*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
717*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
718*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc4x01234567 = vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567));
719*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc4x89ABCDEF = vcombine_s16(vqmovn_s32(vacc4x89AB), vqmovn_s32(vacc4xCDEF));
720*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc5x01234567 = vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567));
721*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc5x89ABCDEF = vcombine_s16(vqmovn_s32(vacc5x89AB), vqmovn_s32(vacc5xCDEF));
722*4bdc9457SAndroid Build Coastguard Worker
723*4bdc9457SAndroid Build Coastguard Worker vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
724*4bdc9457SAndroid Build Coastguard Worker vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
725*4bdc9457SAndroid Build Coastguard Worker vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
726*4bdc9457SAndroid Build Coastguard Worker vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
727*4bdc9457SAndroid Build Coastguard Worker vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
728*4bdc9457SAndroid Build Coastguard Worker vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
729*4bdc9457SAndroid Build Coastguard Worker vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
730*4bdc9457SAndroid Build Coastguard Worker vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
731*4bdc9457SAndroid Build Coastguard Worker vacc4x01234567 = vqaddq_s16(vacc4x01234567, voutput_zero_point);
732*4bdc9457SAndroid Build Coastguard Worker vacc4x89ABCDEF = vqaddq_s16(vacc4x89ABCDEF, voutput_zero_point);
733*4bdc9457SAndroid Build Coastguard Worker vacc5x01234567 = vqaddq_s16(vacc5x01234567, voutput_zero_point);
734*4bdc9457SAndroid Build Coastguard Worker vacc5x89ABCDEF = vqaddq_s16(vacc5x89ABCDEF, voutput_zero_point);
735*4bdc9457SAndroid Build Coastguard Worker
736*4bdc9457SAndroid Build Coastguard Worker uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
737*4bdc9457SAndroid Build Coastguard Worker uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
738*4bdc9457SAndroid Build Coastguard Worker uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
739*4bdc9457SAndroid Build Coastguard Worker uint8x16_t vout3x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc3x01234567), vqmovun_s16(vacc3x89ABCDEF));
740*4bdc9457SAndroid Build Coastguard Worker uint8x16_t vout4x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc4x01234567), vqmovun_s16(vacc4x89ABCDEF));
741*4bdc9457SAndroid Build Coastguard Worker uint8x16_t vout5x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc5x01234567), vqmovun_s16(vacc5x89ABCDEF));
742*4bdc9457SAndroid Build Coastguard Worker #endif
743*4bdc9457SAndroid Build Coastguard Worker
744*4bdc9457SAndroid Build Coastguard Worker const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
745*4bdc9457SAndroid Build Coastguard Worker vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
746*4bdc9457SAndroid Build Coastguard Worker vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
747*4bdc9457SAndroid Build Coastguard Worker vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
748*4bdc9457SAndroid Build Coastguard Worker vout3x0123456789ABCDEF = vmaxq_u8(vout3x0123456789ABCDEF, voutput_min);
749*4bdc9457SAndroid Build Coastguard Worker vout4x0123456789ABCDEF = vmaxq_u8(vout4x0123456789ABCDEF, voutput_min);
750*4bdc9457SAndroid Build Coastguard Worker vout5x0123456789ABCDEF = vmaxq_u8(vout5x0123456789ABCDEF, voutput_min);
751*4bdc9457SAndroid Build Coastguard Worker
752*4bdc9457SAndroid Build Coastguard Worker const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
753*4bdc9457SAndroid Build Coastguard Worker vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
754*4bdc9457SAndroid Build Coastguard Worker vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
755*4bdc9457SAndroid Build Coastguard Worker vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
756*4bdc9457SAndroid Build Coastguard Worker vout3x0123456789ABCDEF = vminq_u8(vout3x0123456789ABCDEF, voutput_max);
757*4bdc9457SAndroid Build Coastguard Worker vout4x0123456789ABCDEF = vminq_u8(vout4x0123456789ABCDEF, voutput_max);
758*4bdc9457SAndroid Build Coastguard Worker vout5x0123456789ABCDEF = vminq_u8(vout5x0123456789ABCDEF, voutput_max);
759*4bdc9457SAndroid Build Coastguard Worker
760*4bdc9457SAndroid Build Coastguard Worker if (nc >= 16) {
761*4bdc9457SAndroid Build Coastguard Worker vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
762*4bdc9457SAndroid Build Coastguard Worker vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
763*4bdc9457SAndroid Build Coastguard Worker vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
764*4bdc9457SAndroid Build Coastguard Worker vst1q_u8(c3 + 0, vout3x0123456789ABCDEF);
765*4bdc9457SAndroid Build Coastguard Worker vst1q_u8(c4 + 0, vout4x0123456789ABCDEF);
766*4bdc9457SAndroid Build Coastguard Worker vst1q_u8(c5 + 0, vout5x0123456789ABCDEF);
767*4bdc9457SAndroid Build Coastguard Worker
768*4bdc9457SAndroid Build Coastguard Worker c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
769*4bdc9457SAndroid Build Coastguard Worker c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
770*4bdc9457SAndroid Build Coastguard Worker c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
771*4bdc9457SAndroid Build Coastguard Worker c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
772*4bdc9457SAndroid Build Coastguard Worker c4 = (uint8_t*) ((uintptr_t) c4 + cn_stride);
773*4bdc9457SAndroid Build Coastguard Worker c5 = (uint8_t*) ((uintptr_t) c5 + cn_stride);
774*4bdc9457SAndroid Build Coastguard Worker
775*4bdc9457SAndroid Build Coastguard Worker a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
776*4bdc9457SAndroid Build Coastguard Worker a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
777*4bdc9457SAndroid Build Coastguard Worker a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
778*4bdc9457SAndroid Build Coastguard Worker a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
779*4bdc9457SAndroid Build Coastguard Worker a4 = (const uint8_t*) ((uintptr_t) a4 - kc);
780*4bdc9457SAndroid Build Coastguard Worker a5 = (const uint8_t*) ((uintptr_t) a5 - kc);
781*4bdc9457SAndroid Build Coastguard Worker
782*4bdc9457SAndroid Build Coastguard Worker nc -= 16;
783*4bdc9457SAndroid Build Coastguard Worker } else {
784*4bdc9457SAndroid Build Coastguard Worker uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
785*4bdc9457SAndroid Build Coastguard Worker uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vget_low_u8(vout2x0123456789ABCDEF), vget_low_u8(vout3x0123456789ABCDEF));
786*4bdc9457SAndroid Build Coastguard Worker uint8x16_t vout4x01234567_5x01234567 = vcombine_u8(vget_low_u8(vout4x0123456789ABCDEF), vget_low_u8(vout5x0123456789ABCDEF));
787*4bdc9457SAndroid Build Coastguard Worker if (nc & 8) {
788*4bdc9457SAndroid Build Coastguard Worker vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
789*4bdc9457SAndroid Build Coastguard Worker vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
790*4bdc9457SAndroid Build Coastguard Worker vst1_u8(c2, vget_low_u8(vout2x01234567_3x01234567)); c2 += 8;
791*4bdc9457SAndroid Build Coastguard Worker vst1_u8(c3, vget_high_u8(vout2x01234567_3x01234567)); c3 += 8;
792*4bdc9457SAndroid Build Coastguard Worker vst1_u8(c4, vget_low_u8(vout4x01234567_5x01234567)); c4 += 8;
793*4bdc9457SAndroid Build Coastguard Worker vst1_u8(c5, vget_high_u8(vout4x01234567_5x01234567)); c5 += 8;
794*4bdc9457SAndroid Build Coastguard Worker vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
795*4bdc9457SAndroid Build Coastguard Worker vout2x01234567_3x01234567 = vcombine_u8(vget_high_u8(vout2x0123456789ABCDEF), vget_high_u8(vout3x0123456789ABCDEF));
796*4bdc9457SAndroid Build Coastguard Worker vout4x01234567_5x01234567 = vcombine_u8(vget_high_u8(vout4x0123456789ABCDEF), vget_high_u8(vout5x0123456789ABCDEF));
797*4bdc9457SAndroid Build Coastguard Worker }
798*4bdc9457SAndroid Build Coastguard Worker if (nc & 4) {
799*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
800*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
801*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u32((void*) c2, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
802*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u32((void*) c3, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
803*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u32((void*) c4, vreinterpretq_u32_u8(vout4x01234567_5x01234567), 0); c4 += 4;
804*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u32((void*) c5, vreinterpretq_u32_u8(vout4x01234567_5x01234567), 2); c5 += 4;
805*4bdc9457SAndroid Build Coastguard Worker vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
806*4bdc9457SAndroid Build Coastguard Worker vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
807*4bdc9457SAndroid Build Coastguard Worker vout4x01234567_5x01234567 = vextq_u8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
808*4bdc9457SAndroid Build Coastguard Worker }
809*4bdc9457SAndroid Build Coastguard Worker if (nc & 2) {
810*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
811*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
812*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u16((void*) c2, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
813*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u16((void*) c3, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
814*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u16((void*) c4, vreinterpretq_u16_u8(vout4x01234567_5x01234567), 0); c4 += 2;
815*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u16((void*) c5, vreinterpretq_u16_u8(vout4x01234567_5x01234567), 4); c5 += 2;
816*4bdc9457SAndroid Build Coastguard Worker vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
817*4bdc9457SAndroid Build Coastguard Worker vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
818*4bdc9457SAndroid Build Coastguard Worker vout4x01234567_5x01234567 = vextq_u8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
819*4bdc9457SAndroid Build Coastguard Worker }
820*4bdc9457SAndroid Build Coastguard Worker if (nc & 1) {
821*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
822*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
823*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
824*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
825*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u8(c4, vout4x01234567_5x01234567, 0);
826*4bdc9457SAndroid Build Coastguard Worker vst1q_lane_u8(c5, vout4x01234567_5x01234567, 8);
827*4bdc9457SAndroid Build Coastguard Worker }
828*4bdc9457SAndroid Build Coastguard Worker
829*4bdc9457SAndroid Build Coastguard Worker nc = 0;
830*4bdc9457SAndroid Build Coastguard Worker }
831*4bdc9457SAndroid Build Coastguard Worker } while (nc != 0);
832*4bdc9457SAndroid Build Coastguard Worker }
833