1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker // Template: src/qs8-dwconv/unipass-neon-mul16.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/dwconv.h>
15*4bdc9457SAndroid Build Coastguard Worker
16*4bdc9457SAndroid Build Coastguard Worker
xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17*4bdc9457SAndroid Build Coastguard Worker void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16(
18*4bdc9457SAndroid Build Coastguard Worker size_t channels,
19*4bdc9457SAndroid Build Coastguard Worker size_t output_width,
20*4bdc9457SAndroid Build Coastguard Worker const int8_t** input,
21*4bdc9457SAndroid Build Coastguard Worker const void* weights,
22*4bdc9457SAndroid Build Coastguard Worker int8_t* output,
23*4bdc9457SAndroid Build Coastguard Worker size_t input_stride,
24*4bdc9457SAndroid Build Coastguard Worker size_t output_increment,
25*4bdc9457SAndroid Build Coastguard Worker size_t input_offset,
26*4bdc9457SAndroid Build Coastguard Worker const int8_t* zero,
27*4bdc9457SAndroid Build Coastguard Worker const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28*4bdc9457SAndroid Build Coastguard Worker {
29*4bdc9457SAndroid Build Coastguard Worker assert(channels != 0);
30*4bdc9457SAndroid Build Coastguard Worker assert(output_width != 0);
31*4bdc9457SAndroid Build Coastguard Worker
32*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
33*4bdc9457SAndroid Build Coastguard Worker const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
34*4bdc9457SAndroid Build Coastguard Worker const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neon.output_min);
35*4bdc9457SAndroid Build Coastguard Worker const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neon.output_max);
36*4bdc9457SAndroid Build Coastguard Worker do {
37*4bdc9457SAndroid Build Coastguard Worker const int8_t* i0 = input[0];
38*4bdc9457SAndroid Build Coastguard Worker assert(i0 != NULL);
39*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i0 != zero) {
40*4bdc9457SAndroid Build Coastguard Worker i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
41*4bdc9457SAndroid Build Coastguard Worker }
42*4bdc9457SAndroid Build Coastguard Worker const int8_t* i1 = input[1];
43*4bdc9457SAndroid Build Coastguard Worker assert(i1 != NULL);
44*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i1 != zero) {
45*4bdc9457SAndroid Build Coastguard Worker i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
46*4bdc9457SAndroid Build Coastguard Worker }
47*4bdc9457SAndroid Build Coastguard Worker const int8_t* i2 = input[2];
48*4bdc9457SAndroid Build Coastguard Worker assert(i2 != NULL);
49*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i2 != zero) {
50*4bdc9457SAndroid Build Coastguard Worker i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
51*4bdc9457SAndroid Build Coastguard Worker }
52*4bdc9457SAndroid Build Coastguard Worker const int8_t* i3 = input[3];
53*4bdc9457SAndroid Build Coastguard Worker assert(i3 != NULL);
54*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i3 != zero) {
55*4bdc9457SAndroid Build Coastguard Worker i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
56*4bdc9457SAndroid Build Coastguard Worker }
57*4bdc9457SAndroid Build Coastguard Worker const int8_t* i4 = input[4];
58*4bdc9457SAndroid Build Coastguard Worker assert(i4 != NULL);
59*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i4 != zero) {
60*4bdc9457SAndroid Build Coastguard Worker i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
61*4bdc9457SAndroid Build Coastguard Worker }
62*4bdc9457SAndroid Build Coastguard Worker const int8_t* i5 = input[5];
63*4bdc9457SAndroid Build Coastguard Worker assert(i5 != NULL);
64*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i5 != zero) {
65*4bdc9457SAndroid Build Coastguard Worker i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
66*4bdc9457SAndroid Build Coastguard Worker }
67*4bdc9457SAndroid Build Coastguard Worker const int8_t* i6 = input[6];
68*4bdc9457SAndroid Build Coastguard Worker assert(i6 != NULL);
69*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i6 != zero) {
70*4bdc9457SAndroid Build Coastguard Worker i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
71*4bdc9457SAndroid Build Coastguard Worker }
72*4bdc9457SAndroid Build Coastguard Worker const int8_t* i7 = input[7];
73*4bdc9457SAndroid Build Coastguard Worker assert(i7 != NULL);
74*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i7 != zero) {
75*4bdc9457SAndroid Build Coastguard Worker i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
76*4bdc9457SAndroid Build Coastguard Worker }
77*4bdc9457SAndroid Build Coastguard Worker const int8_t* i8 = input[8];
78*4bdc9457SAndroid Build Coastguard Worker assert(i8 != NULL);
79*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i8 != zero) {
80*4bdc9457SAndroid Build Coastguard Worker i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
81*4bdc9457SAndroid Build Coastguard Worker }
82*4bdc9457SAndroid Build Coastguard Worker const int8_t* i9 = input[9];
83*4bdc9457SAndroid Build Coastguard Worker assert(i9 != NULL);
84*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i9 != zero) {
85*4bdc9457SAndroid Build Coastguard Worker i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
86*4bdc9457SAndroid Build Coastguard Worker }
87*4bdc9457SAndroid Build Coastguard Worker const int8_t* i10 = input[10];
88*4bdc9457SAndroid Build Coastguard Worker assert(i10 != NULL);
89*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i10 != zero) {
90*4bdc9457SAndroid Build Coastguard Worker i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
91*4bdc9457SAndroid Build Coastguard Worker }
92*4bdc9457SAndroid Build Coastguard Worker const int8_t* i11 = input[11];
93*4bdc9457SAndroid Build Coastguard Worker assert(i11 != NULL);
94*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i11 != zero) {
95*4bdc9457SAndroid Build Coastguard Worker i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
96*4bdc9457SAndroid Build Coastguard Worker }
97*4bdc9457SAndroid Build Coastguard Worker const int8_t* i12 = input[12];
98*4bdc9457SAndroid Build Coastguard Worker assert(i12 != NULL);
99*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i12 != zero) {
100*4bdc9457SAndroid Build Coastguard Worker i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
101*4bdc9457SAndroid Build Coastguard Worker }
102*4bdc9457SAndroid Build Coastguard Worker const int8_t* i13 = input[13];
103*4bdc9457SAndroid Build Coastguard Worker assert(i13 != NULL);
104*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i13 != zero) {
105*4bdc9457SAndroid Build Coastguard Worker i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
106*4bdc9457SAndroid Build Coastguard Worker }
107*4bdc9457SAndroid Build Coastguard Worker const int8_t* i14 = input[14];
108*4bdc9457SAndroid Build Coastguard Worker assert(i14 != NULL);
109*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i14 != zero) {
110*4bdc9457SAndroid Build Coastguard Worker i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
111*4bdc9457SAndroid Build Coastguard Worker }
112*4bdc9457SAndroid Build Coastguard Worker const int8_t* i15 = input[15];
113*4bdc9457SAndroid Build Coastguard Worker assert(i15 != NULL);
114*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i15 != zero) {
115*4bdc9457SAndroid Build Coastguard Worker i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
116*4bdc9457SAndroid Build Coastguard Worker }
117*4bdc9457SAndroid Build Coastguard Worker const int8_t* i16 = input[16];
118*4bdc9457SAndroid Build Coastguard Worker assert(i16 != NULL);
119*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i16 != zero) {
120*4bdc9457SAndroid Build Coastguard Worker i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
121*4bdc9457SAndroid Build Coastguard Worker }
122*4bdc9457SAndroid Build Coastguard Worker const int8_t* i17 = input[17];
123*4bdc9457SAndroid Build Coastguard Worker assert(i17 != NULL);
124*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i17 != zero) {
125*4bdc9457SAndroid Build Coastguard Worker i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
126*4bdc9457SAndroid Build Coastguard Worker }
127*4bdc9457SAndroid Build Coastguard Worker const int8_t* i18 = input[18];
128*4bdc9457SAndroid Build Coastguard Worker assert(i18 != NULL);
129*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i18 != zero) {
130*4bdc9457SAndroid Build Coastguard Worker i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
131*4bdc9457SAndroid Build Coastguard Worker }
132*4bdc9457SAndroid Build Coastguard Worker const int8_t* i19 = input[19];
133*4bdc9457SAndroid Build Coastguard Worker assert(i19 != NULL);
134*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i19 != zero) {
135*4bdc9457SAndroid Build Coastguard Worker i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
136*4bdc9457SAndroid Build Coastguard Worker }
137*4bdc9457SAndroid Build Coastguard Worker const int8_t* i20 = input[20];
138*4bdc9457SAndroid Build Coastguard Worker assert(i20 != NULL);
139*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i20 != zero) {
140*4bdc9457SAndroid Build Coastguard Worker i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
141*4bdc9457SAndroid Build Coastguard Worker }
142*4bdc9457SAndroid Build Coastguard Worker const int8_t* i21 = input[21];
143*4bdc9457SAndroid Build Coastguard Worker assert(i21 != NULL);
144*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i21 != zero) {
145*4bdc9457SAndroid Build Coastguard Worker i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
146*4bdc9457SAndroid Build Coastguard Worker }
147*4bdc9457SAndroid Build Coastguard Worker const int8_t* i22 = input[22];
148*4bdc9457SAndroid Build Coastguard Worker assert(i22 != NULL);
149*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i22 != zero) {
150*4bdc9457SAndroid Build Coastguard Worker i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
151*4bdc9457SAndroid Build Coastguard Worker }
152*4bdc9457SAndroid Build Coastguard Worker const int8_t* i23 = input[23];
153*4bdc9457SAndroid Build Coastguard Worker assert(i23 != NULL);
154*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i23 != zero) {
155*4bdc9457SAndroid Build Coastguard Worker i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
156*4bdc9457SAndroid Build Coastguard Worker }
157*4bdc9457SAndroid Build Coastguard Worker const int8_t* i24 = input[24];
158*4bdc9457SAndroid Build Coastguard Worker assert(i24 != NULL);
159*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i24 != zero) {
160*4bdc9457SAndroid Build Coastguard Worker i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
161*4bdc9457SAndroid Build Coastguard Worker }
162*4bdc9457SAndroid Build Coastguard Worker input = (const int8_t**) ((uintptr_t) input + input_stride);
163*4bdc9457SAndroid Build Coastguard Worker
164*4bdc9457SAndroid Build Coastguard Worker size_t c = channels;
165*4bdc9457SAndroid Build Coastguard Worker const void* w = weights;
166*4bdc9457SAndroid Build Coastguard Worker for (; c >= 16; c -= 16) {
167*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
168*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
169*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
170*4bdc9457SAndroid Build Coastguard Worker int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
171*4bdc9457SAndroid Build Coastguard Worker
172*4bdc9457SAndroid Build Coastguard Worker
173*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
174*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
175*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi0x89ABCDEF = vmovl_s8(vld1_s8(i0)); i0 += 8;
176*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk0x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
177*4bdc9457SAndroid Build Coastguard Worker
178*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
179*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
180*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF));
181*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi0x89ABCDEF), vget_high_s16(vk0x89ABCDEF));
182*4bdc9457SAndroid Build Coastguard Worker
183*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
184*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
185*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi1x89ABCDEF = vmovl_s8(vld1_s8(i1)); i1 += 8;
186*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk1x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
187*4bdc9457SAndroid Build Coastguard Worker
188*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
189*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
190*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF));
191*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF));
192*4bdc9457SAndroid Build Coastguard Worker
193*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
194*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
195*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi2x89ABCDEF = vmovl_s8(vld1_s8(i2)); i2 += 8;
196*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk2x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
197*4bdc9457SAndroid Build Coastguard Worker
198*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
199*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
200*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF));
201*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF));
202*4bdc9457SAndroid Build Coastguard Worker
203*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
204*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
205*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi3x89ABCDEF = vmovl_s8(vld1_s8(i3)); i3 += 8;
206*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk3x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
207*4bdc9457SAndroid Build Coastguard Worker
208*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
209*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
210*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi3x89ABCDEF), vget_low_s16(vk3x89ABCDEF));
211*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi3x89ABCDEF), vget_high_s16(vk3x89ABCDEF));
212*4bdc9457SAndroid Build Coastguard Worker
213*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
214*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
215*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi4x89ABCDEF = vmovl_s8(vld1_s8(i4)); i4 += 8;
216*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk4x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
217*4bdc9457SAndroid Build Coastguard Worker
218*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
219*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
220*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF));
221*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF));
222*4bdc9457SAndroid Build Coastguard Worker
223*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
224*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
225*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi5x89ABCDEF = vmovl_s8(vld1_s8(i5)); i5 += 8;
226*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk5x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
227*4bdc9457SAndroid Build Coastguard Worker
228*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
229*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
230*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi5x89ABCDEF), vget_low_s16(vk5x89ABCDEF));
231*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi5x89ABCDEF), vget_high_s16(vk5x89ABCDEF));
232*4bdc9457SAndroid Build Coastguard Worker
233*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6)); i6 += 8;
234*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
235*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi6x89ABCDEF = vmovl_s8(vld1_s8(i6)); i6 += 8;
236*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk6x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
237*4bdc9457SAndroid Build Coastguard Worker
238*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
239*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
240*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi6x89ABCDEF), vget_low_s16(vk6x89ABCDEF));
241*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi6x89ABCDEF), vget_high_s16(vk6x89ABCDEF));
242*4bdc9457SAndroid Build Coastguard Worker
243*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7)); i7 += 8;
244*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
245*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi7x89ABCDEF = vmovl_s8(vld1_s8(i7)); i7 += 8;
246*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk7x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
247*4bdc9457SAndroid Build Coastguard Worker
248*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
249*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
250*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi7x89ABCDEF), vget_low_s16(vk7x89ABCDEF));
251*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi7x89ABCDEF), vget_high_s16(vk7x89ABCDEF));
252*4bdc9457SAndroid Build Coastguard Worker
253*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8)); i8 += 8;
254*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
255*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi8x89ABCDEF = vmovl_s8(vld1_s8(i8)); i8 += 8;
256*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk8x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
257*4bdc9457SAndroid Build Coastguard Worker
258*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
259*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
260*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi8x89ABCDEF), vget_low_s16(vk8x89ABCDEF));
261*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi8x89ABCDEF), vget_high_s16(vk8x89ABCDEF));
262*4bdc9457SAndroid Build Coastguard Worker
263*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi9x01234567 = vmovl_s8(vld1_s8(i9)); i9 += 8;
264*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk9x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
265*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi9x89ABCDEF = vmovl_s8(vld1_s8(i9)); i9 += 8;
266*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk9x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
267*4bdc9457SAndroid Build Coastguard Worker
268*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi9x01234567), vget_low_s16(vk9x01234567));
269*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi9x01234567), vget_high_s16(vk9x01234567));
270*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi9x89ABCDEF), vget_low_s16(vk9x89ABCDEF));
271*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi9x89ABCDEF), vget_high_s16(vk9x89ABCDEF));
272*4bdc9457SAndroid Build Coastguard Worker
273*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi10x01234567 = vmovl_s8(vld1_s8(i10)); i10 += 8;
274*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk10x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
275*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi10x89ABCDEF = vmovl_s8(vld1_s8(i10)); i10 += 8;
276*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk10x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
277*4bdc9457SAndroid Build Coastguard Worker
278*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi10x01234567), vget_low_s16(vk10x01234567));
279*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi10x01234567), vget_high_s16(vk10x01234567));
280*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi10x89ABCDEF), vget_low_s16(vk10x89ABCDEF));
281*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi10x89ABCDEF), vget_high_s16(vk10x89ABCDEF));
282*4bdc9457SAndroid Build Coastguard Worker
283*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi11x01234567 = vmovl_s8(vld1_s8(i11)); i11 += 8;
284*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk11x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
285*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi11x89ABCDEF = vmovl_s8(vld1_s8(i11)); i11 += 8;
286*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk11x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
287*4bdc9457SAndroid Build Coastguard Worker
288*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi11x01234567), vget_low_s16(vk11x01234567));
289*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi11x01234567), vget_high_s16(vk11x01234567));
290*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi11x89ABCDEF), vget_low_s16(vk11x89ABCDEF));
291*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi11x89ABCDEF), vget_high_s16(vk11x89ABCDEF));
292*4bdc9457SAndroid Build Coastguard Worker
293*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi12x01234567 = vmovl_s8(vld1_s8(i12)); i12 += 8;
294*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk12x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
295*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi12x89ABCDEF = vmovl_s8(vld1_s8(i12)); i12 += 8;
296*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk12x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
297*4bdc9457SAndroid Build Coastguard Worker
298*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi12x01234567), vget_low_s16(vk12x01234567));
299*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi12x01234567), vget_high_s16(vk12x01234567));
300*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi12x89ABCDEF), vget_low_s16(vk12x89ABCDEF));
301*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi12x89ABCDEF), vget_high_s16(vk12x89ABCDEF));
302*4bdc9457SAndroid Build Coastguard Worker
303*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi13x01234567 = vmovl_s8(vld1_s8(i13)); i13 += 8;
304*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk13x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
305*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi13x89ABCDEF = vmovl_s8(vld1_s8(i13)); i13 += 8;
306*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk13x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
307*4bdc9457SAndroid Build Coastguard Worker
308*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi13x01234567), vget_low_s16(vk13x01234567));
309*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi13x01234567), vget_high_s16(vk13x01234567));
310*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi13x89ABCDEF), vget_low_s16(vk13x89ABCDEF));
311*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi13x89ABCDEF), vget_high_s16(vk13x89ABCDEF));
312*4bdc9457SAndroid Build Coastguard Worker
313*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi14x01234567 = vmovl_s8(vld1_s8(i14)); i14 += 8;
314*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk14x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
315*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi14x89ABCDEF = vmovl_s8(vld1_s8(i14)); i14 += 8;
316*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk14x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
317*4bdc9457SAndroid Build Coastguard Worker
318*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi14x01234567), vget_low_s16(vk14x01234567));
319*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi14x01234567), vget_high_s16(vk14x01234567));
320*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi14x89ABCDEF), vget_low_s16(vk14x89ABCDEF));
321*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi14x89ABCDEF), vget_high_s16(vk14x89ABCDEF));
322*4bdc9457SAndroid Build Coastguard Worker
323*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi15x01234567 = vmovl_s8(vld1_s8(i15)); i15 += 8;
324*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk15x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
325*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi15x89ABCDEF = vmovl_s8(vld1_s8(i15)); i15 += 8;
326*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk15x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
327*4bdc9457SAndroid Build Coastguard Worker
328*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi15x01234567), vget_low_s16(vk15x01234567));
329*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi15x01234567), vget_high_s16(vk15x01234567));
330*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi15x89ABCDEF), vget_low_s16(vk15x89ABCDEF));
331*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi15x89ABCDEF), vget_high_s16(vk15x89ABCDEF));
332*4bdc9457SAndroid Build Coastguard Worker
333*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi16x01234567 = vmovl_s8(vld1_s8(i16)); i16 += 8;
334*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk16x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
335*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi16x89ABCDEF = vmovl_s8(vld1_s8(i16)); i16 += 8;
336*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk16x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
337*4bdc9457SAndroid Build Coastguard Worker
338*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi16x01234567), vget_low_s16(vk16x01234567));
339*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi16x01234567), vget_high_s16(vk16x01234567));
340*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi16x89ABCDEF), vget_low_s16(vk16x89ABCDEF));
341*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi16x89ABCDEF), vget_high_s16(vk16x89ABCDEF));
342*4bdc9457SAndroid Build Coastguard Worker
343*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi17x01234567 = vmovl_s8(vld1_s8(i17)); i17 += 8;
344*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk17x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
345*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi17x89ABCDEF = vmovl_s8(vld1_s8(i17)); i17 += 8;
346*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk17x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
347*4bdc9457SAndroid Build Coastguard Worker
348*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi17x01234567), vget_low_s16(vk17x01234567));
349*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi17x01234567), vget_high_s16(vk17x01234567));
350*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi17x89ABCDEF), vget_low_s16(vk17x89ABCDEF));
351*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi17x89ABCDEF), vget_high_s16(vk17x89ABCDEF));
352*4bdc9457SAndroid Build Coastguard Worker
353*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi18x01234567 = vmovl_s8(vld1_s8(i18)); i18 += 8;
354*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk18x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
355*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi18x89ABCDEF = vmovl_s8(vld1_s8(i18)); i18 += 8;
356*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk18x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
357*4bdc9457SAndroid Build Coastguard Worker
358*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi18x01234567), vget_low_s16(vk18x01234567));
359*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi18x01234567), vget_high_s16(vk18x01234567));
360*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi18x89ABCDEF), vget_low_s16(vk18x89ABCDEF));
361*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi18x89ABCDEF), vget_high_s16(vk18x89ABCDEF));
362*4bdc9457SAndroid Build Coastguard Worker
363*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi19x01234567 = vmovl_s8(vld1_s8(i19)); i19 += 8;
364*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk19x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
365*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi19x89ABCDEF = vmovl_s8(vld1_s8(i19)); i19 += 8;
366*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk19x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
367*4bdc9457SAndroid Build Coastguard Worker
368*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi19x01234567), vget_low_s16(vk19x01234567));
369*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi19x01234567), vget_high_s16(vk19x01234567));
370*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi19x89ABCDEF), vget_low_s16(vk19x89ABCDEF));
371*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi19x89ABCDEF), vget_high_s16(vk19x89ABCDEF));
372*4bdc9457SAndroid Build Coastguard Worker
373*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi20x01234567 = vmovl_s8(vld1_s8(i20)); i20 += 8;
374*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk20x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
375*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi20x89ABCDEF = vmovl_s8(vld1_s8(i20)); i20 += 8;
376*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk20x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
377*4bdc9457SAndroid Build Coastguard Worker
378*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi20x01234567), vget_low_s16(vk20x01234567));
379*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi20x01234567), vget_high_s16(vk20x01234567));
380*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi20x89ABCDEF), vget_low_s16(vk20x89ABCDEF));
381*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi20x89ABCDEF), vget_high_s16(vk20x89ABCDEF));
382*4bdc9457SAndroid Build Coastguard Worker
383*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi21x01234567 = vmovl_s8(vld1_s8(i21)); i21 += 8;
384*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk21x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
385*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi21x89ABCDEF = vmovl_s8(vld1_s8(i21)); i21 += 8;
386*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk21x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
387*4bdc9457SAndroid Build Coastguard Worker
388*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi21x01234567), vget_low_s16(vk21x01234567));
389*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi21x01234567), vget_high_s16(vk21x01234567));
390*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi21x89ABCDEF), vget_low_s16(vk21x89ABCDEF));
391*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi21x89ABCDEF), vget_high_s16(vk21x89ABCDEF));
392*4bdc9457SAndroid Build Coastguard Worker
393*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi22x01234567 = vmovl_s8(vld1_s8(i22)); i22 += 8;
394*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
395*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi22x89ABCDEF = vmovl_s8(vld1_s8(i22)); i22 += 8;
396*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk22x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
397*4bdc9457SAndroid Build Coastguard Worker
398*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567));
399*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567));
400*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi22x89ABCDEF), vget_low_s16(vk22x89ABCDEF));
401*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi22x89ABCDEF), vget_high_s16(vk22x89ABCDEF));
402*4bdc9457SAndroid Build Coastguard Worker
403*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi23x01234567 = vmovl_s8(vld1_s8(i23)); i23 += 8;
404*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk23x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
405*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi23x89ABCDEF = vmovl_s8(vld1_s8(i23)); i23 += 8;
406*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk23x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
407*4bdc9457SAndroid Build Coastguard Worker
408*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi23x01234567), vget_low_s16(vk23x01234567));
409*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi23x01234567), vget_high_s16(vk23x01234567));
410*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi23x89ABCDEF), vget_low_s16(vk23x89ABCDEF));
411*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi23x89ABCDEF), vget_high_s16(vk23x89ABCDEF));
412*4bdc9457SAndroid Build Coastguard Worker
413*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi24x01234567 = vmovl_s8(vld1_s8(i24)); i24 += 8;
414*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk24x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
415*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi24x89ABCDEF = vmovl_s8(vld1_s8(i24)); i24 += 8;
416*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk24x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
417*4bdc9457SAndroid Build Coastguard Worker
418*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi24x01234567), vget_low_s16(vk24x01234567));
419*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi24x01234567), vget_high_s16(vk24x01234567));
420*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi24x89ABCDEF), vget_low_s16(vk24x89ABCDEF));
421*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi24x89ABCDEF), vget_high_s16(vk24x89ABCDEF));
422*4bdc9457SAndroid Build Coastguard Worker
423*4bdc9457SAndroid Build Coastguard Worker float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
424*4bdc9457SAndroid Build Coastguard Worker float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
425*4bdc9457SAndroid Build Coastguard Worker float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
426*4bdc9457SAndroid Build Coastguard Worker float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
427*4bdc9457SAndroid Build Coastguard Worker
428*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
429*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
430*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
431*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
432*4bdc9457SAndroid Build Coastguard Worker
433*4bdc9457SAndroid Build Coastguard Worker vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
434*4bdc9457SAndroid Build Coastguard Worker vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
435*4bdc9457SAndroid Build Coastguard Worker vfpacc89AB = vmulq_f32(vfpacc89AB, vscale89AB);
436*4bdc9457SAndroid Build Coastguard Worker vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscaleCDEF);
437*4bdc9457SAndroid Build Coastguard Worker
438*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
439*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
440*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc89AB, vmagic_bias));
441*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpaccCDEF, vmagic_bias));
442*4bdc9457SAndroid Build Coastguard Worker
443*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
444*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
445*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vqsubq_s32(vacc89AB, vmagic_bias_less_output_zero_point);
446*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vqsubq_s32(vaccCDEF, vmagic_bias_less_output_zero_point);
447*4bdc9457SAndroid Build Coastguard Worker
448*4bdc9457SAndroid Build Coastguard Worker #if XNN_ARCH_ARM64
449*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
450*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
451*4bdc9457SAndroid Build Coastguard Worker
452*4bdc9457SAndroid Build Coastguard Worker
453*4bdc9457SAndroid Build Coastguard Worker int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
454*4bdc9457SAndroid Build Coastguard Worker #else // !XNN_ARCH_ARM64
455*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
456*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
457*4bdc9457SAndroid Build Coastguard Worker
458*4bdc9457SAndroid Build Coastguard Worker
459*4bdc9457SAndroid Build Coastguard Worker int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
460*4bdc9457SAndroid Build Coastguard Worker #endif // !XNN_ARCH_ARM64
461*4bdc9457SAndroid Build Coastguard Worker
462*4bdc9457SAndroid Build Coastguard Worker vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
463*4bdc9457SAndroid Build Coastguard Worker
464*4bdc9457SAndroid Build Coastguard Worker vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
465*4bdc9457SAndroid Build Coastguard Worker
466*4bdc9457SAndroid Build Coastguard Worker vst1q_s8(output, vout0123456789ABCDEF); output += 16;
467*4bdc9457SAndroid Build Coastguard Worker }
468*4bdc9457SAndroid Build Coastguard Worker if XNN_UNLIKELY(c != 0) {
469*4bdc9457SAndroid Build Coastguard Worker const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
470*4bdc9457SAndroid Build Coastguard Worker do {
471*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
472*4bdc9457SAndroid Build Coastguard Worker int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
473*4bdc9457SAndroid Build Coastguard Worker
474*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
475*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(k)); k += 8;
476*4bdc9457SAndroid Build Coastguard Worker
477*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
478*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
479*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
480*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8((const void*) (k + 8)));
481*4bdc9457SAndroid Build Coastguard Worker
482*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
483*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
484*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
485*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8((const void*) (k + 24)));
486*4bdc9457SAndroid Build Coastguard Worker
487*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
488*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
489*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
490*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8((const void*) (k + 40)));
491*4bdc9457SAndroid Build Coastguard Worker
492*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
493*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
494*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
495*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8((const void*) (k + 56)));
496*4bdc9457SAndroid Build Coastguard Worker
497*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
498*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
499*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
500*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8((const void*) (k + 72)));
501*4bdc9457SAndroid Build Coastguard Worker
502*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
503*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
504*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6)); i6 += 8;
505*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8((const void*) (k + 88)));
506*4bdc9457SAndroid Build Coastguard Worker
507*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
508*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
509*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7)); i7 += 8;
510*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8((const void*) (k + 104)));
511*4bdc9457SAndroid Build Coastguard Worker
512*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
513*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
514*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8)); i8 += 8;
515*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8((const void*) (k + 120)));
516*4bdc9457SAndroid Build Coastguard Worker
517*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
518*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
519*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi9x01234567 = vmovl_s8(vld1_s8(i9)); i9 += 8;
520*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk9x01234567 = vmovl_s8(vld1_s8((const void*) (k + 136)));
521*4bdc9457SAndroid Build Coastguard Worker
522*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi9x01234567), vget_low_s16(vk9x01234567));
523*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi9x01234567), vget_high_s16(vk9x01234567));
524*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi10x01234567 = vmovl_s8(vld1_s8(i10)); i10 += 8;
525*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk10x01234567 = vmovl_s8(vld1_s8((const void*) (k + 152)));
526*4bdc9457SAndroid Build Coastguard Worker
527*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi10x01234567), vget_low_s16(vk10x01234567));
528*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi10x01234567), vget_high_s16(vk10x01234567));
529*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi11x01234567 = vmovl_s8(vld1_s8(i11)); i11 += 8;
530*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk11x01234567 = vmovl_s8(vld1_s8((const void*) (k + 168)));
531*4bdc9457SAndroid Build Coastguard Worker
532*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi11x01234567), vget_low_s16(vk11x01234567));
533*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi11x01234567), vget_high_s16(vk11x01234567));
534*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi12x01234567 = vmovl_s8(vld1_s8(i12)); i12 += 8;
535*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk12x01234567 = vmovl_s8(vld1_s8((const void*) (k + 184)));
536*4bdc9457SAndroid Build Coastguard Worker
537*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi12x01234567), vget_low_s16(vk12x01234567));
538*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi12x01234567), vget_high_s16(vk12x01234567));
539*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi13x01234567 = vmovl_s8(vld1_s8(i13)); i13 += 8;
540*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk13x01234567 = vmovl_s8(vld1_s8((const void*) (k + 200)));
541*4bdc9457SAndroid Build Coastguard Worker
542*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi13x01234567), vget_low_s16(vk13x01234567));
543*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi13x01234567), vget_high_s16(vk13x01234567));
544*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi14x01234567 = vmovl_s8(vld1_s8(i14)); i14 += 8;
545*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk14x01234567 = vmovl_s8(vld1_s8((const void*) (k + 216)));
546*4bdc9457SAndroid Build Coastguard Worker
547*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi14x01234567), vget_low_s16(vk14x01234567));
548*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi14x01234567), vget_high_s16(vk14x01234567));
549*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi15x01234567 = vmovl_s8(vld1_s8(i15)); i15 += 8;
550*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk15x01234567 = vmovl_s8(vld1_s8((const void*) (k + 232)));
551*4bdc9457SAndroid Build Coastguard Worker
552*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi15x01234567), vget_low_s16(vk15x01234567));
553*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi15x01234567), vget_high_s16(vk15x01234567));
554*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi16x01234567 = vmovl_s8(vld1_s8(i16)); i16 += 8;
555*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk16x01234567 = vmovl_s8(vld1_s8((const void*) (k + 248)));
556*4bdc9457SAndroid Build Coastguard Worker
557*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi16x01234567), vget_low_s16(vk16x01234567));
558*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi16x01234567), vget_high_s16(vk16x01234567));
559*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi17x01234567 = vmovl_s8(vld1_s8(i17)); i17 += 8;
560*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk17x01234567 = vmovl_s8(vld1_s8((const void*) (k + 264)));
561*4bdc9457SAndroid Build Coastguard Worker
562*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi17x01234567), vget_low_s16(vk17x01234567));
563*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi17x01234567), vget_high_s16(vk17x01234567));
564*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi18x01234567 = vmovl_s8(vld1_s8(i18)); i18 += 8;
565*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk18x01234567 = vmovl_s8(vld1_s8((const void*) (k + 280)));
566*4bdc9457SAndroid Build Coastguard Worker
567*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi18x01234567), vget_low_s16(vk18x01234567));
568*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi18x01234567), vget_high_s16(vk18x01234567));
569*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi19x01234567 = vmovl_s8(vld1_s8(i19)); i19 += 8;
570*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk19x01234567 = vmovl_s8(vld1_s8((const void*) (k + 296)));
571*4bdc9457SAndroid Build Coastguard Worker
572*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi19x01234567), vget_low_s16(vk19x01234567));
573*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi19x01234567), vget_high_s16(vk19x01234567));
574*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi20x01234567 = vmovl_s8(vld1_s8(i20)); i20 += 8;
575*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk20x01234567 = vmovl_s8(vld1_s8((const void*) (k + 312)));
576*4bdc9457SAndroid Build Coastguard Worker
577*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi20x01234567), vget_low_s16(vk20x01234567));
578*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi20x01234567), vget_high_s16(vk20x01234567));
579*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi21x01234567 = vmovl_s8(vld1_s8(i21)); i21 += 8;
580*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk21x01234567 = vmovl_s8(vld1_s8((const void*) (k + 328)));
581*4bdc9457SAndroid Build Coastguard Worker
582*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi21x01234567), vget_low_s16(vk21x01234567));
583*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi21x01234567), vget_high_s16(vk21x01234567));
584*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi22x01234567 = vmovl_s8(vld1_s8(i22)); i22 += 8;
585*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8((const void*) (k + 344)));
586*4bdc9457SAndroid Build Coastguard Worker
587*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567));
588*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567));
589*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi23x01234567 = vmovl_s8(vld1_s8(i23)); i23 += 8;
590*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk23x01234567 = vmovl_s8(vld1_s8((const void*) (k + 360)));
591*4bdc9457SAndroid Build Coastguard Worker
592*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi23x01234567), vget_low_s16(vk23x01234567));
593*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi23x01234567), vget_high_s16(vk23x01234567));
594*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vi24x01234567 = vmovl_s8(vld1_s8(i24)); i24 += 8;
595*4bdc9457SAndroid Build Coastguard Worker const int16x8_t vk24x01234567 = vmovl_s8(vld1_s8((const void*) (k + 376)));
596*4bdc9457SAndroid Build Coastguard Worker
597*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi24x01234567), vget_low_s16(vk24x01234567));
598*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi24x01234567), vget_high_s16(vk24x01234567));
599*4bdc9457SAndroid Build Coastguard Worker
600*4bdc9457SAndroid Build Coastguard Worker float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
601*4bdc9457SAndroid Build Coastguard Worker float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
602*4bdc9457SAndroid Build Coastguard Worker
603*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 400 * sizeof(int8_t)));
604*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 400 * sizeof(int8_t) + 4 * sizeof(float)));
605*4bdc9457SAndroid Build Coastguard Worker vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
606*4bdc9457SAndroid Build Coastguard Worker vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
607*4bdc9457SAndroid Build Coastguard Worker
608*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
609*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
610*4bdc9457SAndroid Build Coastguard Worker
611*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
612*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
613*4bdc9457SAndroid Build Coastguard Worker
614*4bdc9457SAndroid Build Coastguard Worker #if XNN_ARCH_ARM64
615*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
616*4bdc9457SAndroid Build Coastguard Worker #else
617*4bdc9457SAndroid Build Coastguard Worker int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
618*4bdc9457SAndroid Build Coastguard Worker #endif
619*4bdc9457SAndroid Build Coastguard Worker
620*4bdc9457SAndroid Build Coastguard Worker int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
621*4bdc9457SAndroid Build Coastguard Worker vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
622*4bdc9457SAndroid Build Coastguard Worker vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
623*4bdc9457SAndroid Build Coastguard Worker
624*4bdc9457SAndroid Build Coastguard Worker if XNN_LIKELY(c >= 8) {
625*4bdc9457SAndroid Build Coastguard Worker vst1_s8(output, vout01234567); output += 8;
626*4bdc9457SAndroid Build Coastguard Worker c -= 8;
627*4bdc9457SAndroid Build Coastguard Worker } else {
628*4bdc9457SAndroid Build Coastguard Worker if (c & 4) {
629*4bdc9457SAndroid Build Coastguard Worker vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
630*4bdc9457SAndroid Build Coastguard Worker vout01234567 = vext_s8(vout01234567, vout01234567, 4);
631*4bdc9457SAndroid Build Coastguard Worker }
632*4bdc9457SAndroid Build Coastguard Worker if (c & 2) {
633*4bdc9457SAndroid Build Coastguard Worker vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
634*4bdc9457SAndroid Build Coastguard Worker vout01234567 = vext_s8(vout01234567, vout01234567, 2);
635*4bdc9457SAndroid Build Coastguard Worker }
636*4bdc9457SAndroid Build Coastguard Worker if (c & 1) {
637*4bdc9457SAndroid Build Coastguard Worker vst1_lane_s8(output, vout01234567, 0); output += 1;
638*4bdc9457SAndroid Build Coastguard Worker }
639*4bdc9457SAndroid Build Coastguard Worker c = 0;
640*4bdc9457SAndroid Build Coastguard Worker }
641*4bdc9457SAndroid Build Coastguard Worker } while (c != 0);
642*4bdc9457SAndroid Build Coastguard Worker }
643*4bdc9457SAndroid Build Coastguard Worker
644*4bdc9457SAndroid Build Coastguard Worker output = (int8_t*) ((uintptr_t) output + output_increment);
645*4bdc9457SAndroid Build Coastguard Worker } while (--output_width != 0);
646*4bdc9457SAndroid Build Coastguard Worker }
647