1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker // Template: src/qs8-dwconv/unipass-scalar.c.in
3*4bdc9457SAndroid Build Coastguard Worker // Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2021 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 #include <math.h>
12*4bdc9457SAndroid Build Coastguard Worker
13*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/dwconv.h>
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/math.h>
15*4bdc9457SAndroid Build Coastguard Worker
16*4bdc9457SAndroid Build Coastguard Worker
xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf(size_t channels,size_t output_width,const uint8_t ** input,const void * weights,uint8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17*4bdc9457SAndroid Build Coastguard Worker void xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf(
18*4bdc9457SAndroid Build Coastguard Worker size_t channels,
19*4bdc9457SAndroid Build Coastguard Worker size_t output_width,
20*4bdc9457SAndroid Build Coastguard Worker const uint8_t** input,
21*4bdc9457SAndroid Build Coastguard Worker const void* weights,
22*4bdc9457SAndroid Build Coastguard Worker uint8_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 uint8_t* zero,
27*4bdc9457SAndroid Build Coastguard Worker const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
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 float vscale = params->fp32_scalar_lrintf.scale;
33*4bdc9457SAndroid Build Coastguard Worker const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
34*4bdc9457SAndroid Build Coastguard Worker const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
35*4bdc9457SAndroid Build Coastguard Worker const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
36*4bdc9457SAndroid Build Coastguard Worker const int32_t vkernel_zero_point = params->fp32_scalar_lrintf.kernel_zero_point;
37*4bdc9457SAndroid Build Coastguard Worker do {
38*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i0 = input[0];
39*4bdc9457SAndroid Build Coastguard Worker assert(i0 != NULL);
40*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i0 != zero) {
41*4bdc9457SAndroid Build Coastguard Worker i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
42*4bdc9457SAndroid Build Coastguard Worker }
43*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i1 = input[1];
44*4bdc9457SAndroid Build Coastguard Worker assert(i1 != NULL);
45*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i1 != zero) {
46*4bdc9457SAndroid Build Coastguard Worker i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
47*4bdc9457SAndroid Build Coastguard Worker }
48*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i2 = input[2];
49*4bdc9457SAndroid Build Coastguard Worker assert(i2 != NULL);
50*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i2 != zero) {
51*4bdc9457SAndroid Build Coastguard Worker i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
52*4bdc9457SAndroid Build Coastguard Worker }
53*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i3 = input[3];
54*4bdc9457SAndroid Build Coastguard Worker assert(i3 != NULL);
55*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i3 != zero) {
56*4bdc9457SAndroid Build Coastguard Worker i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
57*4bdc9457SAndroid Build Coastguard Worker }
58*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i4 = input[4];
59*4bdc9457SAndroid Build Coastguard Worker assert(i4 != NULL);
60*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i4 != zero) {
61*4bdc9457SAndroid Build Coastguard Worker i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
62*4bdc9457SAndroid Build Coastguard Worker }
63*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i5 = input[5];
64*4bdc9457SAndroid Build Coastguard Worker assert(i5 != NULL);
65*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i5 != zero) {
66*4bdc9457SAndroid Build Coastguard Worker i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
67*4bdc9457SAndroid Build Coastguard Worker }
68*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i6 = input[6];
69*4bdc9457SAndroid Build Coastguard Worker assert(i6 != NULL);
70*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i6 != zero) {
71*4bdc9457SAndroid Build Coastguard Worker i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
72*4bdc9457SAndroid Build Coastguard Worker }
73*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i7 = input[7];
74*4bdc9457SAndroid Build Coastguard Worker assert(i7 != NULL);
75*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i7 != zero) {
76*4bdc9457SAndroid Build Coastguard Worker i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
77*4bdc9457SAndroid Build Coastguard Worker }
78*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i8 = input[8];
79*4bdc9457SAndroid Build Coastguard Worker assert(i8 != NULL);
80*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i8 != zero) {
81*4bdc9457SAndroid Build Coastguard Worker i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
82*4bdc9457SAndroid Build Coastguard Worker }
83*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i9 = input[9];
84*4bdc9457SAndroid Build Coastguard Worker assert(i9 != NULL);
85*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i9 != zero) {
86*4bdc9457SAndroid Build Coastguard Worker i9 = (const uint8_t*) ((uintptr_t) i9 + input_offset);
87*4bdc9457SAndroid Build Coastguard Worker }
88*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i10 = input[10];
89*4bdc9457SAndroid Build Coastguard Worker assert(i10 != NULL);
90*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i10 != zero) {
91*4bdc9457SAndroid Build Coastguard Worker i10 = (const uint8_t*) ((uintptr_t) i10 + input_offset);
92*4bdc9457SAndroid Build Coastguard Worker }
93*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i11 = input[11];
94*4bdc9457SAndroid Build Coastguard Worker assert(i11 != NULL);
95*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i11 != zero) {
96*4bdc9457SAndroid Build Coastguard Worker i11 = (const uint8_t*) ((uintptr_t) i11 + input_offset);
97*4bdc9457SAndroid Build Coastguard Worker }
98*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i12 = input[12];
99*4bdc9457SAndroid Build Coastguard Worker assert(i12 != NULL);
100*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i12 != zero) {
101*4bdc9457SAndroid Build Coastguard Worker i12 = (const uint8_t*) ((uintptr_t) i12 + input_offset);
102*4bdc9457SAndroid Build Coastguard Worker }
103*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i13 = input[13];
104*4bdc9457SAndroid Build Coastguard Worker assert(i13 != NULL);
105*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i13 != zero) {
106*4bdc9457SAndroid Build Coastguard Worker i13 = (const uint8_t*) ((uintptr_t) i13 + input_offset);
107*4bdc9457SAndroid Build Coastguard Worker }
108*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i14 = input[14];
109*4bdc9457SAndroid Build Coastguard Worker assert(i14 != NULL);
110*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i14 != zero) {
111*4bdc9457SAndroid Build Coastguard Worker i14 = (const uint8_t*) ((uintptr_t) i14 + input_offset);
112*4bdc9457SAndroid Build Coastguard Worker }
113*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i15 = input[15];
114*4bdc9457SAndroid Build Coastguard Worker assert(i15 != NULL);
115*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i15 != zero) {
116*4bdc9457SAndroid Build Coastguard Worker i15 = (const uint8_t*) ((uintptr_t) i15 + input_offset);
117*4bdc9457SAndroid Build Coastguard Worker }
118*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i16 = input[16];
119*4bdc9457SAndroid Build Coastguard Worker assert(i16 != NULL);
120*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i16 != zero) {
121*4bdc9457SAndroid Build Coastguard Worker i16 = (const uint8_t*) ((uintptr_t) i16 + input_offset);
122*4bdc9457SAndroid Build Coastguard Worker }
123*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i17 = input[17];
124*4bdc9457SAndroid Build Coastguard Worker assert(i17 != NULL);
125*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i17 != zero) {
126*4bdc9457SAndroid Build Coastguard Worker i17 = (const uint8_t*) ((uintptr_t) i17 + input_offset);
127*4bdc9457SAndroid Build Coastguard Worker }
128*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i18 = input[18];
129*4bdc9457SAndroid Build Coastguard Worker assert(i18 != NULL);
130*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i18 != zero) {
131*4bdc9457SAndroid Build Coastguard Worker i18 = (const uint8_t*) ((uintptr_t) i18 + input_offset);
132*4bdc9457SAndroid Build Coastguard Worker }
133*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i19 = input[19];
134*4bdc9457SAndroid Build Coastguard Worker assert(i19 != NULL);
135*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i19 != zero) {
136*4bdc9457SAndroid Build Coastguard Worker i19 = (const uint8_t*) ((uintptr_t) i19 + input_offset);
137*4bdc9457SAndroid Build Coastguard Worker }
138*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i20 = input[20];
139*4bdc9457SAndroid Build Coastguard Worker assert(i20 != NULL);
140*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i20 != zero) {
141*4bdc9457SAndroid Build Coastguard Worker i20 = (const uint8_t*) ((uintptr_t) i20 + input_offset);
142*4bdc9457SAndroid Build Coastguard Worker }
143*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i21 = input[21];
144*4bdc9457SAndroid Build Coastguard Worker assert(i21 != NULL);
145*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i21 != zero) {
146*4bdc9457SAndroid Build Coastguard Worker i21 = (const uint8_t*) ((uintptr_t) i21 + input_offset);
147*4bdc9457SAndroid Build Coastguard Worker }
148*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i22 = input[22];
149*4bdc9457SAndroid Build Coastguard Worker assert(i22 != NULL);
150*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i22 != zero) {
151*4bdc9457SAndroid Build Coastguard Worker i22 = (const uint8_t*) ((uintptr_t) i22 + input_offset);
152*4bdc9457SAndroid Build Coastguard Worker }
153*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i23 = input[23];
154*4bdc9457SAndroid Build Coastguard Worker assert(i23 != NULL);
155*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i23 != zero) {
156*4bdc9457SAndroid Build Coastguard Worker i23 = (const uint8_t*) ((uintptr_t) i23 + input_offset);
157*4bdc9457SAndroid Build Coastguard Worker }
158*4bdc9457SAndroid Build Coastguard Worker const uint8_t* i24 = input[24];
159*4bdc9457SAndroid Build Coastguard Worker assert(i24 != NULL);
160*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i24 != zero) {
161*4bdc9457SAndroid Build Coastguard Worker i24 = (const uint8_t*) ((uintptr_t) i24 + input_offset);
162*4bdc9457SAndroid Build Coastguard Worker }
163*4bdc9457SAndroid Build Coastguard Worker input = (const uint8_t**) ((uintptr_t) input + input_stride);
164*4bdc9457SAndroid Build Coastguard Worker
165*4bdc9457SAndroid Build Coastguard Worker size_t c = channels;
166*4bdc9457SAndroid Build Coastguard Worker const void* w = weights;
167*4bdc9457SAndroid Build Coastguard Worker for (; c >= 4; c -= 4) {
168*4bdc9457SAndroid Build Coastguard Worker int32_t vacc0 = ((const int32_t*) w)[0];
169*4bdc9457SAndroid Build Coastguard Worker int32_t vacc1 = ((const int32_t*) w)[1];
170*4bdc9457SAndroid Build Coastguard Worker int32_t vacc2 = ((const int32_t*) w)[2];
171*4bdc9457SAndroid Build Coastguard Worker int32_t vacc3 = ((const int32_t*) w)[3];
172*4bdc9457SAndroid Build Coastguard Worker
173*4bdc9457SAndroid Build Coastguard Worker
174*4bdc9457SAndroid Build Coastguard Worker const int32_t vi0x0 = (int32_t) (uint32_t) i0[0];
175*4bdc9457SAndroid Build Coastguard Worker const int32_t vi0x1 = (int32_t) (uint32_t) i0[1];
176*4bdc9457SAndroid Build Coastguard Worker const int32_t vi0x2 = (int32_t) (uint32_t) i0[2];
177*4bdc9457SAndroid Build Coastguard Worker const int32_t vi0x3 = (int32_t) (uint32_t) i0[3];
178*4bdc9457SAndroid Build Coastguard Worker i0 += 4;
179*4bdc9457SAndroid Build Coastguard Worker
180*4bdc9457SAndroid Build Coastguard Worker const int32_t vk0x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[0] - vkernel_zero_point;
181*4bdc9457SAndroid Build Coastguard Worker const int32_t vk0x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[1] - vkernel_zero_point;
182*4bdc9457SAndroid Build Coastguard Worker const int32_t vk0x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[2] - vkernel_zero_point;
183*4bdc9457SAndroid Build Coastguard Worker const int32_t vk0x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[3] - vkernel_zero_point;
184*4bdc9457SAndroid Build Coastguard Worker
185*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi0x0 * vk0x0;
186*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi0x1 * vk0x1;
187*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi0x2 * vk0x2;
188*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi0x3 * vk0x3;
189*4bdc9457SAndroid Build Coastguard Worker
190*4bdc9457SAndroid Build Coastguard Worker const int32_t vi1x0 = (int32_t) (uint32_t) i1[0];
191*4bdc9457SAndroid Build Coastguard Worker const int32_t vi1x1 = (int32_t) (uint32_t) i1[1];
192*4bdc9457SAndroid Build Coastguard Worker const int32_t vi1x2 = (int32_t) (uint32_t) i1[2];
193*4bdc9457SAndroid Build Coastguard Worker const int32_t vi1x3 = (int32_t) (uint32_t) i1[3];
194*4bdc9457SAndroid Build Coastguard Worker i1 += 4;
195*4bdc9457SAndroid Build Coastguard Worker
196*4bdc9457SAndroid Build Coastguard Worker const int32_t vk1x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[4] - vkernel_zero_point;
197*4bdc9457SAndroid Build Coastguard Worker const int32_t vk1x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[5] - vkernel_zero_point;
198*4bdc9457SAndroid Build Coastguard Worker const int32_t vk1x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[6] - vkernel_zero_point;
199*4bdc9457SAndroid Build Coastguard Worker const int32_t vk1x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[7] - vkernel_zero_point;
200*4bdc9457SAndroid Build Coastguard Worker
201*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi1x0 * vk1x0;
202*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi1x1 * vk1x1;
203*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi1x2 * vk1x2;
204*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi1x3 * vk1x3;
205*4bdc9457SAndroid Build Coastguard Worker
206*4bdc9457SAndroid Build Coastguard Worker const int32_t vi2x0 = (int32_t) (uint32_t) i2[0];
207*4bdc9457SAndroid Build Coastguard Worker const int32_t vi2x1 = (int32_t) (uint32_t) i2[1];
208*4bdc9457SAndroid Build Coastguard Worker const int32_t vi2x2 = (int32_t) (uint32_t) i2[2];
209*4bdc9457SAndroid Build Coastguard Worker const int32_t vi2x3 = (int32_t) (uint32_t) i2[3];
210*4bdc9457SAndroid Build Coastguard Worker i2 += 4;
211*4bdc9457SAndroid Build Coastguard Worker
212*4bdc9457SAndroid Build Coastguard Worker const int32_t vk2x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[8] - vkernel_zero_point;
213*4bdc9457SAndroid Build Coastguard Worker const int32_t vk2x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[9] - vkernel_zero_point;
214*4bdc9457SAndroid Build Coastguard Worker const int32_t vk2x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[10] - vkernel_zero_point;
215*4bdc9457SAndroid Build Coastguard Worker const int32_t vk2x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[11] - vkernel_zero_point;
216*4bdc9457SAndroid Build Coastguard Worker
217*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi2x0 * vk2x0;
218*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi2x1 * vk2x1;
219*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi2x2 * vk2x2;
220*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi2x3 * vk2x3;
221*4bdc9457SAndroid Build Coastguard Worker
222*4bdc9457SAndroid Build Coastguard Worker const int32_t vi3x0 = (int32_t) (uint32_t) i3[0];
223*4bdc9457SAndroid Build Coastguard Worker const int32_t vi3x1 = (int32_t) (uint32_t) i3[1];
224*4bdc9457SAndroid Build Coastguard Worker const int32_t vi3x2 = (int32_t) (uint32_t) i3[2];
225*4bdc9457SAndroid Build Coastguard Worker const int32_t vi3x3 = (int32_t) (uint32_t) i3[3];
226*4bdc9457SAndroid Build Coastguard Worker i3 += 4;
227*4bdc9457SAndroid Build Coastguard Worker
228*4bdc9457SAndroid Build Coastguard Worker const int32_t vk3x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[12] - vkernel_zero_point;
229*4bdc9457SAndroid Build Coastguard Worker const int32_t vk3x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[13] - vkernel_zero_point;
230*4bdc9457SAndroid Build Coastguard Worker const int32_t vk3x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[14] - vkernel_zero_point;
231*4bdc9457SAndroid Build Coastguard Worker const int32_t vk3x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[15] - vkernel_zero_point;
232*4bdc9457SAndroid Build Coastguard Worker
233*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi3x0 * vk3x0;
234*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi3x1 * vk3x1;
235*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi3x2 * vk3x2;
236*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi3x3 * vk3x3;
237*4bdc9457SAndroid Build Coastguard Worker
238*4bdc9457SAndroid Build Coastguard Worker const int32_t vi4x0 = (int32_t) (uint32_t) i4[0];
239*4bdc9457SAndroid Build Coastguard Worker const int32_t vi4x1 = (int32_t) (uint32_t) i4[1];
240*4bdc9457SAndroid Build Coastguard Worker const int32_t vi4x2 = (int32_t) (uint32_t) i4[2];
241*4bdc9457SAndroid Build Coastguard Worker const int32_t vi4x3 = (int32_t) (uint32_t) i4[3];
242*4bdc9457SAndroid Build Coastguard Worker i4 += 4;
243*4bdc9457SAndroid Build Coastguard Worker
244*4bdc9457SAndroid Build Coastguard Worker const int32_t vk4x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[16] - vkernel_zero_point;
245*4bdc9457SAndroid Build Coastguard Worker const int32_t vk4x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[17] - vkernel_zero_point;
246*4bdc9457SAndroid Build Coastguard Worker const int32_t vk4x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[18] - vkernel_zero_point;
247*4bdc9457SAndroid Build Coastguard Worker const int32_t vk4x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[19] - vkernel_zero_point;
248*4bdc9457SAndroid Build Coastguard Worker
249*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi4x0 * vk4x0;
250*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi4x1 * vk4x1;
251*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi4x2 * vk4x2;
252*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi4x3 * vk4x3;
253*4bdc9457SAndroid Build Coastguard Worker
254*4bdc9457SAndroid Build Coastguard Worker const int32_t vi5x0 = (int32_t) (uint32_t) i5[0];
255*4bdc9457SAndroid Build Coastguard Worker const int32_t vi5x1 = (int32_t) (uint32_t) i5[1];
256*4bdc9457SAndroid Build Coastguard Worker const int32_t vi5x2 = (int32_t) (uint32_t) i5[2];
257*4bdc9457SAndroid Build Coastguard Worker const int32_t vi5x3 = (int32_t) (uint32_t) i5[3];
258*4bdc9457SAndroid Build Coastguard Worker i5 += 4;
259*4bdc9457SAndroid Build Coastguard Worker
260*4bdc9457SAndroid Build Coastguard Worker const int32_t vk5x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[20] - vkernel_zero_point;
261*4bdc9457SAndroid Build Coastguard Worker const int32_t vk5x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[21] - vkernel_zero_point;
262*4bdc9457SAndroid Build Coastguard Worker const int32_t vk5x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[22] - vkernel_zero_point;
263*4bdc9457SAndroid Build Coastguard Worker const int32_t vk5x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[23] - vkernel_zero_point;
264*4bdc9457SAndroid Build Coastguard Worker
265*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi5x0 * vk5x0;
266*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi5x1 * vk5x1;
267*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi5x2 * vk5x2;
268*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi5x3 * vk5x3;
269*4bdc9457SAndroid Build Coastguard Worker
270*4bdc9457SAndroid Build Coastguard Worker const int32_t vi6x0 = (int32_t) (uint32_t) i6[0];
271*4bdc9457SAndroid Build Coastguard Worker const int32_t vi6x1 = (int32_t) (uint32_t) i6[1];
272*4bdc9457SAndroid Build Coastguard Worker const int32_t vi6x2 = (int32_t) (uint32_t) i6[2];
273*4bdc9457SAndroid Build Coastguard Worker const int32_t vi6x3 = (int32_t) (uint32_t) i6[3];
274*4bdc9457SAndroid Build Coastguard Worker i6 += 4;
275*4bdc9457SAndroid Build Coastguard Worker
276*4bdc9457SAndroid Build Coastguard Worker const int32_t vk6x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[24] - vkernel_zero_point;
277*4bdc9457SAndroid Build Coastguard Worker const int32_t vk6x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[25] - vkernel_zero_point;
278*4bdc9457SAndroid Build Coastguard Worker const int32_t vk6x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[26] - vkernel_zero_point;
279*4bdc9457SAndroid Build Coastguard Worker const int32_t vk6x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[27] - vkernel_zero_point;
280*4bdc9457SAndroid Build Coastguard Worker
281*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi6x0 * vk6x0;
282*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi6x1 * vk6x1;
283*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi6x2 * vk6x2;
284*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi6x3 * vk6x3;
285*4bdc9457SAndroid Build Coastguard Worker
286*4bdc9457SAndroid Build Coastguard Worker const int32_t vi7x0 = (int32_t) (uint32_t) i7[0];
287*4bdc9457SAndroid Build Coastguard Worker const int32_t vi7x1 = (int32_t) (uint32_t) i7[1];
288*4bdc9457SAndroid Build Coastguard Worker const int32_t vi7x2 = (int32_t) (uint32_t) i7[2];
289*4bdc9457SAndroid Build Coastguard Worker const int32_t vi7x3 = (int32_t) (uint32_t) i7[3];
290*4bdc9457SAndroid Build Coastguard Worker i7 += 4;
291*4bdc9457SAndroid Build Coastguard Worker
292*4bdc9457SAndroid Build Coastguard Worker const int32_t vk7x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[28] - vkernel_zero_point;
293*4bdc9457SAndroid Build Coastguard Worker const int32_t vk7x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[29] - vkernel_zero_point;
294*4bdc9457SAndroid Build Coastguard Worker const int32_t vk7x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[30] - vkernel_zero_point;
295*4bdc9457SAndroid Build Coastguard Worker const int32_t vk7x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[31] - vkernel_zero_point;
296*4bdc9457SAndroid Build Coastguard Worker
297*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi7x0 * vk7x0;
298*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi7x1 * vk7x1;
299*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi7x2 * vk7x2;
300*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi7x3 * vk7x3;
301*4bdc9457SAndroid Build Coastguard Worker
302*4bdc9457SAndroid Build Coastguard Worker const int32_t vi8x0 = (int32_t) (uint32_t) i8[0];
303*4bdc9457SAndroid Build Coastguard Worker const int32_t vi8x1 = (int32_t) (uint32_t) i8[1];
304*4bdc9457SAndroid Build Coastguard Worker const int32_t vi8x2 = (int32_t) (uint32_t) i8[2];
305*4bdc9457SAndroid Build Coastguard Worker const int32_t vi8x3 = (int32_t) (uint32_t) i8[3];
306*4bdc9457SAndroid Build Coastguard Worker i8 += 4;
307*4bdc9457SAndroid Build Coastguard Worker
308*4bdc9457SAndroid Build Coastguard Worker const int32_t vk8x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[32] - vkernel_zero_point;
309*4bdc9457SAndroid Build Coastguard Worker const int32_t vk8x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[33] - vkernel_zero_point;
310*4bdc9457SAndroid Build Coastguard Worker const int32_t vk8x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[34] - vkernel_zero_point;
311*4bdc9457SAndroid Build Coastguard Worker const int32_t vk8x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[35] - vkernel_zero_point;
312*4bdc9457SAndroid Build Coastguard Worker
313*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi8x0 * vk8x0;
314*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi8x1 * vk8x1;
315*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi8x2 * vk8x2;
316*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi8x3 * vk8x3;
317*4bdc9457SAndroid Build Coastguard Worker
318*4bdc9457SAndroid Build Coastguard Worker const int32_t vi9x0 = (int32_t) (uint32_t) i9[0];
319*4bdc9457SAndroid Build Coastguard Worker const int32_t vi9x1 = (int32_t) (uint32_t) i9[1];
320*4bdc9457SAndroid Build Coastguard Worker const int32_t vi9x2 = (int32_t) (uint32_t) i9[2];
321*4bdc9457SAndroid Build Coastguard Worker const int32_t vi9x3 = (int32_t) (uint32_t) i9[3];
322*4bdc9457SAndroid Build Coastguard Worker i9 += 4;
323*4bdc9457SAndroid Build Coastguard Worker
324*4bdc9457SAndroid Build Coastguard Worker const int32_t vk9x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[36] - vkernel_zero_point;
325*4bdc9457SAndroid Build Coastguard Worker const int32_t vk9x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[37] - vkernel_zero_point;
326*4bdc9457SAndroid Build Coastguard Worker const int32_t vk9x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[38] - vkernel_zero_point;
327*4bdc9457SAndroid Build Coastguard Worker const int32_t vk9x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[39] - vkernel_zero_point;
328*4bdc9457SAndroid Build Coastguard Worker
329*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi9x0 * vk9x0;
330*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi9x1 * vk9x1;
331*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi9x2 * vk9x2;
332*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi9x3 * vk9x3;
333*4bdc9457SAndroid Build Coastguard Worker
334*4bdc9457SAndroid Build Coastguard Worker const int32_t vi10x0 = (int32_t) (uint32_t) i10[0];
335*4bdc9457SAndroid Build Coastguard Worker const int32_t vi10x1 = (int32_t) (uint32_t) i10[1];
336*4bdc9457SAndroid Build Coastguard Worker const int32_t vi10x2 = (int32_t) (uint32_t) i10[2];
337*4bdc9457SAndroid Build Coastguard Worker const int32_t vi10x3 = (int32_t) (uint32_t) i10[3];
338*4bdc9457SAndroid Build Coastguard Worker i10 += 4;
339*4bdc9457SAndroid Build Coastguard Worker
340*4bdc9457SAndroid Build Coastguard Worker const int32_t vk10x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[40] - vkernel_zero_point;
341*4bdc9457SAndroid Build Coastguard Worker const int32_t vk10x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[41] - vkernel_zero_point;
342*4bdc9457SAndroid Build Coastguard Worker const int32_t vk10x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[42] - vkernel_zero_point;
343*4bdc9457SAndroid Build Coastguard Worker const int32_t vk10x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[43] - vkernel_zero_point;
344*4bdc9457SAndroid Build Coastguard Worker
345*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi10x0 * vk10x0;
346*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi10x1 * vk10x1;
347*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi10x2 * vk10x2;
348*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi10x3 * vk10x3;
349*4bdc9457SAndroid Build Coastguard Worker
350*4bdc9457SAndroid Build Coastguard Worker const int32_t vi11x0 = (int32_t) (uint32_t) i11[0];
351*4bdc9457SAndroid Build Coastguard Worker const int32_t vi11x1 = (int32_t) (uint32_t) i11[1];
352*4bdc9457SAndroid Build Coastguard Worker const int32_t vi11x2 = (int32_t) (uint32_t) i11[2];
353*4bdc9457SAndroid Build Coastguard Worker const int32_t vi11x3 = (int32_t) (uint32_t) i11[3];
354*4bdc9457SAndroid Build Coastguard Worker i11 += 4;
355*4bdc9457SAndroid Build Coastguard Worker
356*4bdc9457SAndroid Build Coastguard Worker const int32_t vk11x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[44] - vkernel_zero_point;
357*4bdc9457SAndroid Build Coastguard Worker const int32_t vk11x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[45] - vkernel_zero_point;
358*4bdc9457SAndroid Build Coastguard Worker const int32_t vk11x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[46] - vkernel_zero_point;
359*4bdc9457SAndroid Build Coastguard Worker const int32_t vk11x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[47] - vkernel_zero_point;
360*4bdc9457SAndroid Build Coastguard Worker
361*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi11x0 * vk11x0;
362*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi11x1 * vk11x1;
363*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi11x2 * vk11x2;
364*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi11x3 * vk11x3;
365*4bdc9457SAndroid Build Coastguard Worker
366*4bdc9457SAndroid Build Coastguard Worker const int32_t vi12x0 = (int32_t) (uint32_t) i12[0];
367*4bdc9457SAndroid Build Coastguard Worker const int32_t vi12x1 = (int32_t) (uint32_t) i12[1];
368*4bdc9457SAndroid Build Coastguard Worker const int32_t vi12x2 = (int32_t) (uint32_t) i12[2];
369*4bdc9457SAndroid Build Coastguard Worker const int32_t vi12x3 = (int32_t) (uint32_t) i12[3];
370*4bdc9457SAndroid Build Coastguard Worker i12 += 4;
371*4bdc9457SAndroid Build Coastguard Worker
372*4bdc9457SAndroid Build Coastguard Worker const int32_t vk12x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[48] - vkernel_zero_point;
373*4bdc9457SAndroid Build Coastguard Worker const int32_t vk12x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[49] - vkernel_zero_point;
374*4bdc9457SAndroid Build Coastguard Worker const int32_t vk12x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[50] - vkernel_zero_point;
375*4bdc9457SAndroid Build Coastguard Worker const int32_t vk12x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[51] - vkernel_zero_point;
376*4bdc9457SAndroid Build Coastguard Worker
377*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi12x0 * vk12x0;
378*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi12x1 * vk12x1;
379*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi12x2 * vk12x2;
380*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi12x3 * vk12x3;
381*4bdc9457SAndroid Build Coastguard Worker
382*4bdc9457SAndroid Build Coastguard Worker const int32_t vi13x0 = (int32_t) (uint32_t) i13[0];
383*4bdc9457SAndroid Build Coastguard Worker const int32_t vi13x1 = (int32_t) (uint32_t) i13[1];
384*4bdc9457SAndroid Build Coastguard Worker const int32_t vi13x2 = (int32_t) (uint32_t) i13[2];
385*4bdc9457SAndroid Build Coastguard Worker const int32_t vi13x3 = (int32_t) (uint32_t) i13[3];
386*4bdc9457SAndroid Build Coastguard Worker i13 += 4;
387*4bdc9457SAndroid Build Coastguard Worker
388*4bdc9457SAndroid Build Coastguard Worker const int32_t vk13x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[52] - vkernel_zero_point;
389*4bdc9457SAndroid Build Coastguard Worker const int32_t vk13x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[53] - vkernel_zero_point;
390*4bdc9457SAndroid Build Coastguard Worker const int32_t vk13x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[54] - vkernel_zero_point;
391*4bdc9457SAndroid Build Coastguard Worker const int32_t vk13x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[55] - vkernel_zero_point;
392*4bdc9457SAndroid Build Coastguard Worker
393*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi13x0 * vk13x0;
394*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi13x1 * vk13x1;
395*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi13x2 * vk13x2;
396*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi13x3 * vk13x3;
397*4bdc9457SAndroid Build Coastguard Worker
398*4bdc9457SAndroid Build Coastguard Worker const int32_t vi14x0 = (int32_t) (uint32_t) i14[0];
399*4bdc9457SAndroid Build Coastguard Worker const int32_t vi14x1 = (int32_t) (uint32_t) i14[1];
400*4bdc9457SAndroid Build Coastguard Worker const int32_t vi14x2 = (int32_t) (uint32_t) i14[2];
401*4bdc9457SAndroid Build Coastguard Worker const int32_t vi14x3 = (int32_t) (uint32_t) i14[3];
402*4bdc9457SAndroid Build Coastguard Worker i14 += 4;
403*4bdc9457SAndroid Build Coastguard Worker
404*4bdc9457SAndroid Build Coastguard Worker const int32_t vk14x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[56] - vkernel_zero_point;
405*4bdc9457SAndroid Build Coastguard Worker const int32_t vk14x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[57] - vkernel_zero_point;
406*4bdc9457SAndroid Build Coastguard Worker const int32_t vk14x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[58] - vkernel_zero_point;
407*4bdc9457SAndroid Build Coastguard Worker const int32_t vk14x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[59] - vkernel_zero_point;
408*4bdc9457SAndroid Build Coastguard Worker
409*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi14x0 * vk14x0;
410*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi14x1 * vk14x1;
411*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi14x2 * vk14x2;
412*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi14x3 * vk14x3;
413*4bdc9457SAndroid Build Coastguard Worker
414*4bdc9457SAndroid Build Coastguard Worker const int32_t vi15x0 = (int32_t) (uint32_t) i15[0];
415*4bdc9457SAndroid Build Coastguard Worker const int32_t vi15x1 = (int32_t) (uint32_t) i15[1];
416*4bdc9457SAndroid Build Coastguard Worker const int32_t vi15x2 = (int32_t) (uint32_t) i15[2];
417*4bdc9457SAndroid Build Coastguard Worker const int32_t vi15x3 = (int32_t) (uint32_t) i15[3];
418*4bdc9457SAndroid Build Coastguard Worker i15 += 4;
419*4bdc9457SAndroid Build Coastguard Worker
420*4bdc9457SAndroid Build Coastguard Worker const int32_t vk15x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[60] - vkernel_zero_point;
421*4bdc9457SAndroid Build Coastguard Worker const int32_t vk15x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[61] - vkernel_zero_point;
422*4bdc9457SAndroid Build Coastguard Worker const int32_t vk15x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[62] - vkernel_zero_point;
423*4bdc9457SAndroid Build Coastguard Worker const int32_t vk15x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[63] - vkernel_zero_point;
424*4bdc9457SAndroid Build Coastguard Worker
425*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi15x0 * vk15x0;
426*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi15x1 * vk15x1;
427*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi15x2 * vk15x2;
428*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi15x3 * vk15x3;
429*4bdc9457SAndroid Build Coastguard Worker
430*4bdc9457SAndroid Build Coastguard Worker const int32_t vi16x0 = (int32_t) (uint32_t) i16[0];
431*4bdc9457SAndroid Build Coastguard Worker const int32_t vi16x1 = (int32_t) (uint32_t) i16[1];
432*4bdc9457SAndroid Build Coastguard Worker const int32_t vi16x2 = (int32_t) (uint32_t) i16[2];
433*4bdc9457SAndroid Build Coastguard Worker const int32_t vi16x3 = (int32_t) (uint32_t) i16[3];
434*4bdc9457SAndroid Build Coastguard Worker i16 += 4;
435*4bdc9457SAndroid Build Coastguard Worker
436*4bdc9457SAndroid Build Coastguard Worker const int32_t vk16x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[64] - vkernel_zero_point;
437*4bdc9457SAndroid Build Coastguard Worker const int32_t vk16x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[65] - vkernel_zero_point;
438*4bdc9457SAndroid Build Coastguard Worker const int32_t vk16x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[66] - vkernel_zero_point;
439*4bdc9457SAndroid Build Coastguard Worker const int32_t vk16x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[67] - vkernel_zero_point;
440*4bdc9457SAndroid Build Coastguard Worker
441*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi16x0 * vk16x0;
442*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi16x1 * vk16x1;
443*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi16x2 * vk16x2;
444*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi16x3 * vk16x3;
445*4bdc9457SAndroid Build Coastguard Worker
446*4bdc9457SAndroid Build Coastguard Worker const int32_t vi17x0 = (int32_t) (uint32_t) i17[0];
447*4bdc9457SAndroid Build Coastguard Worker const int32_t vi17x1 = (int32_t) (uint32_t) i17[1];
448*4bdc9457SAndroid Build Coastguard Worker const int32_t vi17x2 = (int32_t) (uint32_t) i17[2];
449*4bdc9457SAndroid Build Coastguard Worker const int32_t vi17x3 = (int32_t) (uint32_t) i17[3];
450*4bdc9457SAndroid Build Coastguard Worker i17 += 4;
451*4bdc9457SAndroid Build Coastguard Worker
452*4bdc9457SAndroid Build Coastguard Worker const int32_t vk17x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[68] - vkernel_zero_point;
453*4bdc9457SAndroid Build Coastguard Worker const int32_t vk17x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[69] - vkernel_zero_point;
454*4bdc9457SAndroid Build Coastguard Worker const int32_t vk17x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[70] - vkernel_zero_point;
455*4bdc9457SAndroid Build Coastguard Worker const int32_t vk17x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[71] - vkernel_zero_point;
456*4bdc9457SAndroid Build Coastguard Worker
457*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi17x0 * vk17x0;
458*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi17x1 * vk17x1;
459*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi17x2 * vk17x2;
460*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi17x3 * vk17x3;
461*4bdc9457SAndroid Build Coastguard Worker
462*4bdc9457SAndroid Build Coastguard Worker const int32_t vi18x0 = (int32_t) (uint32_t) i18[0];
463*4bdc9457SAndroid Build Coastguard Worker const int32_t vi18x1 = (int32_t) (uint32_t) i18[1];
464*4bdc9457SAndroid Build Coastguard Worker const int32_t vi18x2 = (int32_t) (uint32_t) i18[2];
465*4bdc9457SAndroid Build Coastguard Worker const int32_t vi18x3 = (int32_t) (uint32_t) i18[3];
466*4bdc9457SAndroid Build Coastguard Worker i18 += 4;
467*4bdc9457SAndroid Build Coastguard Worker
468*4bdc9457SAndroid Build Coastguard Worker const int32_t vk18x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[72] - vkernel_zero_point;
469*4bdc9457SAndroid Build Coastguard Worker const int32_t vk18x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[73] - vkernel_zero_point;
470*4bdc9457SAndroid Build Coastguard Worker const int32_t vk18x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[74] - vkernel_zero_point;
471*4bdc9457SAndroid Build Coastguard Worker const int32_t vk18x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[75] - vkernel_zero_point;
472*4bdc9457SAndroid Build Coastguard Worker
473*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi18x0 * vk18x0;
474*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi18x1 * vk18x1;
475*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi18x2 * vk18x2;
476*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi18x3 * vk18x3;
477*4bdc9457SAndroid Build Coastguard Worker
478*4bdc9457SAndroid Build Coastguard Worker const int32_t vi19x0 = (int32_t) (uint32_t) i19[0];
479*4bdc9457SAndroid Build Coastguard Worker const int32_t vi19x1 = (int32_t) (uint32_t) i19[1];
480*4bdc9457SAndroid Build Coastguard Worker const int32_t vi19x2 = (int32_t) (uint32_t) i19[2];
481*4bdc9457SAndroid Build Coastguard Worker const int32_t vi19x3 = (int32_t) (uint32_t) i19[3];
482*4bdc9457SAndroid Build Coastguard Worker i19 += 4;
483*4bdc9457SAndroid Build Coastguard Worker
484*4bdc9457SAndroid Build Coastguard Worker const int32_t vk19x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[76] - vkernel_zero_point;
485*4bdc9457SAndroid Build Coastguard Worker const int32_t vk19x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[77] - vkernel_zero_point;
486*4bdc9457SAndroid Build Coastguard Worker const int32_t vk19x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[78] - vkernel_zero_point;
487*4bdc9457SAndroid Build Coastguard Worker const int32_t vk19x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[79] - vkernel_zero_point;
488*4bdc9457SAndroid Build Coastguard Worker
489*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi19x0 * vk19x0;
490*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi19x1 * vk19x1;
491*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi19x2 * vk19x2;
492*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi19x3 * vk19x3;
493*4bdc9457SAndroid Build Coastguard Worker
494*4bdc9457SAndroid Build Coastguard Worker const int32_t vi20x0 = (int32_t) (uint32_t) i20[0];
495*4bdc9457SAndroid Build Coastguard Worker const int32_t vi20x1 = (int32_t) (uint32_t) i20[1];
496*4bdc9457SAndroid Build Coastguard Worker const int32_t vi20x2 = (int32_t) (uint32_t) i20[2];
497*4bdc9457SAndroid Build Coastguard Worker const int32_t vi20x3 = (int32_t) (uint32_t) i20[3];
498*4bdc9457SAndroid Build Coastguard Worker i20 += 4;
499*4bdc9457SAndroid Build Coastguard Worker
500*4bdc9457SAndroid Build Coastguard Worker const int32_t vk20x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[80] - vkernel_zero_point;
501*4bdc9457SAndroid Build Coastguard Worker const int32_t vk20x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[81] - vkernel_zero_point;
502*4bdc9457SAndroid Build Coastguard Worker const int32_t vk20x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[82] - vkernel_zero_point;
503*4bdc9457SAndroid Build Coastguard Worker const int32_t vk20x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[83] - vkernel_zero_point;
504*4bdc9457SAndroid Build Coastguard Worker
505*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi20x0 * vk20x0;
506*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi20x1 * vk20x1;
507*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi20x2 * vk20x2;
508*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi20x3 * vk20x3;
509*4bdc9457SAndroid Build Coastguard Worker
510*4bdc9457SAndroid Build Coastguard Worker const int32_t vi21x0 = (int32_t) (uint32_t) i21[0];
511*4bdc9457SAndroid Build Coastguard Worker const int32_t vi21x1 = (int32_t) (uint32_t) i21[1];
512*4bdc9457SAndroid Build Coastguard Worker const int32_t vi21x2 = (int32_t) (uint32_t) i21[2];
513*4bdc9457SAndroid Build Coastguard Worker const int32_t vi21x3 = (int32_t) (uint32_t) i21[3];
514*4bdc9457SAndroid Build Coastguard Worker i21 += 4;
515*4bdc9457SAndroid Build Coastguard Worker
516*4bdc9457SAndroid Build Coastguard Worker const int32_t vk21x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[84] - vkernel_zero_point;
517*4bdc9457SAndroid Build Coastguard Worker const int32_t vk21x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[85] - vkernel_zero_point;
518*4bdc9457SAndroid Build Coastguard Worker const int32_t vk21x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[86] - vkernel_zero_point;
519*4bdc9457SAndroid Build Coastguard Worker const int32_t vk21x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[87] - vkernel_zero_point;
520*4bdc9457SAndroid Build Coastguard Worker
521*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi21x0 * vk21x0;
522*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi21x1 * vk21x1;
523*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi21x2 * vk21x2;
524*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi21x3 * vk21x3;
525*4bdc9457SAndroid Build Coastguard Worker
526*4bdc9457SAndroid Build Coastguard Worker const int32_t vi22x0 = (int32_t) (uint32_t) i22[0];
527*4bdc9457SAndroid Build Coastguard Worker const int32_t vi22x1 = (int32_t) (uint32_t) i22[1];
528*4bdc9457SAndroid Build Coastguard Worker const int32_t vi22x2 = (int32_t) (uint32_t) i22[2];
529*4bdc9457SAndroid Build Coastguard Worker const int32_t vi22x3 = (int32_t) (uint32_t) i22[3];
530*4bdc9457SAndroid Build Coastguard Worker i22 += 4;
531*4bdc9457SAndroid Build Coastguard Worker
532*4bdc9457SAndroid Build Coastguard Worker const int32_t vk22x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[88] - vkernel_zero_point;
533*4bdc9457SAndroid Build Coastguard Worker const int32_t vk22x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[89] - vkernel_zero_point;
534*4bdc9457SAndroid Build Coastguard Worker const int32_t vk22x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[90] - vkernel_zero_point;
535*4bdc9457SAndroid Build Coastguard Worker const int32_t vk22x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[91] - vkernel_zero_point;
536*4bdc9457SAndroid Build Coastguard Worker
537*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi22x0 * vk22x0;
538*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi22x1 * vk22x1;
539*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi22x2 * vk22x2;
540*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi22x3 * vk22x3;
541*4bdc9457SAndroid Build Coastguard Worker
542*4bdc9457SAndroid Build Coastguard Worker const int32_t vi23x0 = (int32_t) (uint32_t) i23[0];
543*4bdc9457SAndroid Build Coastguard Worker const int32_t vi23x1 = (int32_t) (uint32_t) i23[1];
544*4bdc9457SAndroid Build Coastguard Worker const int32_t vi23x2 = (int32_t) (uint32_t) i23[2];
545*4bdc9457SAndroid Build Coastguard Worker const int32_t vi23x3 = (int32_t) (uint32_t) i23[3];
546*4bdc9457SAndroid Build Coastguard Worker i23 += 4;
547*4bdc9457SAndroid Build Coastguard Worker
548*4bdc9457SAndroid Build Coastguard Worker const int32_t vk23x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[92] - vkernel_zero_point;
549*4bdc9457SAndroid Build Coastguard Worker const int32_t vk23x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[93] - vkernel_zero_point;
550*4bdc9457SAndroid Build Coastguard Worker const int32_t vk23x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[94] - vkernel_zero_point;
551*4bdc9457SAndroid Build Coastguard Worker const int32_t vk23x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[95] - vkernel_zero_point;
552*4bdc9457SAndroid Build Coastguard Worker
553*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi23x0 * vk23x0;
554*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi23x1 * vk23x1;
555*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi23x2 * vk23x2;
556*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi23x3 * vk23x3;
557*4bdc9457SAndroid Build Coastguard Worker
558*4bdc9457SAndroid Build Coastguard Worker const int32_t vi24x0 = (int32_t) (uint32_t) i24[0];
559*4bdc9457SAndroid Build Coastguard Worker const int32_t vi24x1 = (int32_t) (uint32_t) i24[1];
560*4bdc9457SAndroid Build Coastguard Worker const int32_t vi24x2 = (int32_t) (uint32_t) i24[2];
561*4bdc9457SAndroid Build Coastguard Worker const int32_t vi24x3 = (int32_t) (uint32_t) i24[3];
562*4bdc9457SAndroid Build Coastguard Worker i24 += 4;
563*4bdc9457SAndroid Build Coastguard Worker
564*4bdc9457SAndroid Build Coastguard Worker const int32_t vk24x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[96] - vkernel_zero_point;
565*4bdc9457SAndroid Build Coastguard Worker const int32_t vk24x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[97] - vkernel_zero_point;
566*4bdc9457SAndroid Build Coastguard Worker const int32_t vk24x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[98] - vkernel_zero_point;
567*4bdc9457SAndroid Build Coastguard Worker const int32_t vk24x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[99] - vkernel_zero_point;
568*4bdc9457SAndroid Build Coastguard Worker
569*4bdc9457SAndroid Build Coastguard Worker vacc0 += vi24x0 * vk24x0;
570*4bdc9457SAndroid Build Coastguard Worker vacc1 += vi24x1 * vk24x1;
571*4bdc9457SAndroid Build Coastguard Worker vacc2 += vi24x2 * vk24x2;
572*4bdc9457SAndroid Build Coastguard Worker vacc3 += vi24x3 * vk24x3;
573*4bdc9457SAndroid Build Coastguard Worker
574*4bdc9457SAndroid Build Coastguard Worker w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t) + 100 * sizeof(uint8_t));
575*4bdc9457SAndroid Build Coastguard Worker
576*4bdc9457SAndroid Build Coastguard Worker float vfpacc0 = (float) vacc0;
577*4bdc9457SAndroid Build Coastguard Worker float vfpacc1 = (float) vacc1;
578*4bdc9457SAndroid Build Coastguard Worker float vfpacc2 = (float) vacc2;
579*4bdc9457SAndroid Build Coastguard Worker float vfpacc3 = (float) vacc3;
580*4bdc9457SAndroid Build Coastguard Worker
581*4bdc9457SAndroid Build Coastguard Worker vfpacc0 *= vscale;
582*4bdc9457SAndroid Build Coastguard Worker vfpacc1 *= vscale;
583*4bdc9457SAndroid Build Coastguard Worker vfpacc2 *= vscale;
584*4bdc9457SAndroid Build Coastguard Worker vfpacc3 *= vscale;
585*4bdc9457SAndroid Build Coastguard Worker
586*4bdc9457SAndroid Build Coastguard Worker vfpacc0 = math_max_f32(vfpacc0, voutput_min_less_zero_point);
587*4bdc9457SAndroid Build Coastguard Worker vfpacc1 = math_max_f32(vfpacc1, voutput_min_less_zero_point);
588*4bdc9457SAndroid Build Coastguard Worker vfpacc2 = math_max_f32(vfpacc2, voutput_min_less_zero_point);
589*4bdc9457SAndroid Build Coastguard Worker vfpacc3 = math_max_f32(vfpacc3, voutput_min_less_zero_point);
590*4bdc9457SAndroid Build Coastguard Worker
591*4bdc9457SAndroid Build Coastguard Worker vfpacc0 = math_min_f32(vfpacc0, voutput_max_less_zero_point);
592*4bdc9457SAndroid Build Coastguard Worker vfpacc1 = math_min_f32(vfpacc1, voutput_max_less_zero_point);
593*4bdc9457SAndroid Build Coastguard Worker vfpacc2 = math_min_f32(vfpacc2, voutput_max_less_zero_point);
594*4bdc9457SAndroid Build Coastguard Worker vfpacc3 = math_min_f32(vfpacc3, voutput_max_less_zero_point);
595*4bdc9457SAndroid Build Coastguard Worker
596*4bdc9457SAndroid Build Coastguard Worker const int32_t vrndacc0 = (int32_t) lrintf(vfpacc0);
597*4bdc9457SAndroid Build Coastguard Worker const int32_t vrndacc1 = (int32_t) lrintf(vfpacc1);
598*4bdc9457SAndroid Build Coastguard Worker const int32_t vrndacc2 = (int32_t) lrintf(vfpacc2);
599*4bdc9457SAndroid Build Coastguard Worker const int32_t vrndacc3 = (int32_t) lrintf(vfpacc3);
600*4bdc9457SAndroid Build Coastguard Worker
601*4bdc9457SAndroid Build Coastguard Worker int32_t vout0 = (int32_t) vrndacc0 + voutput_zero_point;
602*4bdc9457SAndroid Build Coastguard Worker int32_t vout1 = (int32_t) vrndacc1 + voutput_zero_point;
603*4bdc9457SAndroid Build Coastguard Worker int32_t vout2 = (int32_t) vrndacc2 + voutput_zero_point;
604*4bdc9457SAndroid Build Coastguard Worker int32_t vout3 = (int32_t) vrndacc3 + voutput_zero_point;
605*4bdc9457SAndroid Build Coastguard Worker
606*4bdc9457SAndroid Build Coastguard Worker output[0] = (uint8_t) vout0;
607*4bdc9457SAndroid Build Coastguard Worker output[1] = (uint8_t) vout1;
608*4bdc9457SAndroid Build Coastguard Worker output[2] = (uint8_t) vout2;
609*4bdc9457SAndroid Build Coastguard Worker output[3] = (uint8_t) vout3;
610*4bdc9457SAndroid Build Coastguard Worker output += 4;
611*4bdc9457SAndroid Build Coastguard Worker }
612*4bdc9457SAndroid Build Coastguard Worker if XNN_UNLIKELY(c != 0) {
613*4bdc9457SAndroid Build Coastguard Worker const uint8_t* k = (const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t));
614*4bdc9457SAndroid Build Coastguard Worker do {
615*4bdc9457SAndroid Build Coastguard Worker int32_t vacc = *((const int32_t*) w);
616*4bdc9457SAndroid Build Coastguard Worker w = (const void*) ((uintptr_t) w + sizeof(int32_t));
617*4bdc9457SAndroid Build Coastguard Worker
618*4bdc9457SAndroid Build Coastguard Worker const int32_t vi0 = (int32_t) (uint32_t) *i0++;
619*4bdc9457SAndroid Build Coastguard Worker const int32_t vk0 = (int32_t) (uint32_t) k[0] - vkernel_zero_point;
620*4bdc9457SAndroid Build Coastguard Worker vacc += vi0 * vk0;
621*4bdc9457SAndroid Build Coastguard Worker const int32_t vi1 = (int32_t) (uint32_t) *i1++;
622*4bdc9457SAndroid Build Coastguard Worker const int32_t vk1 = (int32_t) (uint32_t) k[4] - vkernel_zero_point;
623*4bdc9457SAndroid Build Coastguard Worker vacc += vi1 * vk1;
624*4bdc9457SAndroid Build Coastguard Worker const int32_t vi2 = (int32_t) (uint32_t) *i2++;
625*4bdc9457SAndroid Build Coastguard Worker const int32_t vk2 = (int32_t) (uint32_t) k[8] - vkernel_zero_point;
626*4bdc9457SAndroid Build Coastguard Worker vacc += vi2 * vk2;
627*4bdc9457SAndroid Build Coastguard Worker const int32_t vi3 = (int32_t) (uint32_t) *i3++;
628*4bdc9457SAndroid Build Coastguard Worker const int32_t vk3 = (int32_t) (uint32_t) k[12] - vkernel_zero_point;
629*4bdc9457SAndroid Build Coastguard Worker vacc += vi3 * vk3;
630*4bdc9457SAndroid Build Coastguard Worker const int32_t vi4 = (int32_t) (uint32_t) *i4++;
631*4bdc9457SAndroid Build Coastguard Worker const int32_t vk4 = (int32_t) (uint32_t) k[16] - vkernel_zero_point;
632*4bdc9457SAndroid Build Coastguard Worker vacc += vi4 * vk4;
633*4bdc9457SAndroid Build Coastguard Worker const int32_t vi5 = (int32_t) (uint32_t) *i5++;
634*4bdc9457SAndroid Build Coastguard Worker const int32_t vk5 = (int32_t) (uint32_t) k[20] - vkernel_zero_point;
635*4bdc9457SAndroid Build Coastguard Worker vacc += vi5 * vk5;
636*4bdc9457SAndroid Build Coastguard Worker const int32_t vi6 = (int32_t) (uint32_t) *i6++;
637*4bdc9457SAndroid Build Coastguard Worker const int32_t vk6 = (int32_t) (uint32_t) k[24] - vkernel_zero_point;
638*4bdc9457SAndroid Build Coastguard Worker vacc += vi6 * vk6;
639*4bdc9457SAndroid Build Coastguard Worker const int32_t vi7 = (int32_t) (uint32_t) *i7++;
640*4bdc9457SAndroid Build Coastguard Worker const int32_t vk7 = (int32_t) (uint32_t) k[28] - vkernel_zero_point;
641*4bdc9457SAndroid Build Coastguard Worker vacc += vi7 * vk7;
642*4bdc9457SAndroid Build Coastguard Worker const int32_t vi8 = (int32_t) (uint32_t) *i8++;
643*4bdc9457SAndroid Build Coastguard Worker const int32_t vk8 = (int32_t) (uint32_t) k[32] - vkernel_zero_point;
644*4bdc9457SAndroid Build Coastguard Worker vacc += vi8 * vk8;
645*4bdc9457SAndroid Build Coastguard Worker const int32_t vi9 = (int32_t) (uint32_t) *i9++;
646*4bdc9457SAndroid Build Coastguard Worker const int32_t vk9 = (int32_t) (uint32_t) k[36] - vkernel_zero_point;
647*4bdc9457SAndroid Build Coastguard Worker vacc += vi9 * vk9;
648*4bdc9457SAndroid Build Coastguard Worker const int32_t vi10 = (int32_t) (uint32_t) *i10++;
649*4bdc9457SAndroid Build Coastguard Worker const int32_t vk10 = (int32_t) (uint32_t) k[40] - vkernel_zero_point;
650*4bdc9457SAndroid Build Coastguard Worker vacc += vi10 * vk10;
651*4bdc9457SAndroid Build Coastguard Worker const int32_t vi11 = (int32_t) (uint32_t) *i11++;
652*4bdc9457SAndroid Build Coastguard Worker const int32_t vk11 = (int32_t) (uint32_t) k[44] - vkernel_zero_point;
653*4bdc9457SAndroid Build Coastguard Worker vacc += vi11 * vk11;
654*4bdc9457SAndroid Build Coastguard Worker const int32_t vi12 = (int32_t) (uint32_t) *i12++;
655*4bdc9457SAndroid Build Coastguard Worker const int32_t vk12 = (int32_t) (uint32_t) k[48] - vkernel_zero_point;
656*4bdc9457SAndroid Build Coastguard Worker vacc += vi12 * vk12;
657*4bdc9457SAndroid Build Coastguard Worker const int32_t vi13 = (int32_t) (uint32_t) *i13++;
658*4bdc9457SAndroid Build Coastguard Worker const int32_t vk13 = (int32_t) (uint32_t) k[52] - vkernel_zero_point;
659*4bdc9457SAndroid Build Coastguard Worker vacc += vi13 * vk13;
660*4bdc9457SAndroid Build Coastguard Worker const int32_t vi14 = (int32_t) (uint32_t) *i14++;
661*4bdc9457SAndroid Build Coastguard Worker const int32_t vk14 = (int32_t) (uint32_t) k[56] - vkernel_zero_point;
662*4bdc9457SAndroid Build Coastguard Worker vacc += vi14 * vk14;
663*4bdc9457SAndroid Build Coastguard Worker const int32_t vi15 = (int32_t) (uint32_t) *i15++;
664*4bdc9457SAndroid Build Coastguard Worker const int32_t vk15 = (int32_t) (uint32_t) k[60] - vkernel_zero_point;
665*4bdc9457SAndroid Build Coastguard Worker vacc += vi15 * vk15;
666*4bdc9457SAndroid Build Coastguard Worker const int32_t vi16 = (int32_t) (uint32_t) *i16++;
667*4bdc9457SAndroid Build Coastguard Worker const int32_t vk16 = (int32_t) (uint32_t) k[64] - vkernel_zero_point;
668*4bdc9457SAndroid Build Coastguard Worker vacc += vi16 * vk16;
669*4bdc9457SAndroid Build Coastguard Worker const int32_t vi17 = (int32_t) (uint32_t) *i17++;
670*4bdc9457SAndroid Build Coastguard Worker const int32_t vk17 = (int32_t) (uint32_t) k[68] - vkernel_zero_point;
671*4bdc9457SAndroid Build Coastguard Worker vacc += vi17 * vk17;
672*4bdc9457SAndroid Build Coastguard Worker const int32_t vi18 = (int32_t) (uint32_t) *i18++;
673*4bdc9457SAndroid Build Coastguard Worker const int32_t vk18 = (int32_t) (uint32_t) k[72] - vkernel_zero_point;
674*4bdc9457SAndroid Build Coastguard Worker vacc += vi18 * vk18;
675*4bdc9457SAndroid Build Coastguard Worker const int32_t vi19 = (int32_t) (uint32_t) *i19++;
676*4bdc9457SAndroid Build Coastguard Worker const int32_t vk19 = (int32_t) (uint32_t) k[76] - vkernel_zero_point;
677*4bdc9457SAndroid Build Coastguard Worker vacc += vi19 * vk19;
678*4bdc9457SAndroid Build Coastguard Worker const int32_t vi20 = (int32_t) (uint32_t) *i20++;
679*4bdc9457SAndroid Build Coastguard Worker const int32_t vk20 = (int32_t) (uint32_t) k[80] - vkernel_zero_point;
680*4bdc9457SAndroid Build Coastguard Worker vacc += vi20 * vk20;
681*4bdc9457SAndroid Build Coastguard Worker const int32_t vi21 = (int32_t) (uint32_t) *i21++;
682*4bdc9457SAndroid Build Coastguard Worker const int32_t vk21 = (int32_t) (uint32_t) k[84] - vkernel_zero_point;
683*4bdc9457SAndroid Build Coastguard Worker vacc += vi21 * vk21;
684*4bdc9457SAndroid Build Coastguard Worker const int32_t vi22 = (int32_t) (uint32_t) *i22++;
685*4bdc9457SAndroid Build Coastguard Worker const int32_t vk22 = (int32_t) (uint32_t) k[88] - vkernel_zero_point;
686*4bdc9457SAndroid Build Coastguard Worker vacc += vi22 * vk22;
687*4bdc9457SAndroid Build Coastguard Worker const int32_t vi23 = (int32_t) (uint32_t) *i23++;
688*4bdc9457SAndroid Build Coastguard Worker const int32_t vk23 = (int32_t) (uint32_t) k[92] - vkernel_zero_point;
689*4bdc9457SAndroid Build Coastguard Worker vacc += vi23 * vk23;
690*4bdc9457SAndroid Build Coastguard Worker const int32_t vi24 = (int32_t) (uint32_t) *i24++;
691*4bdc9457SAndroid Build Coastguard Worker const int32_t vk24 = (int32_t) (uint32_t) k[96] - vkernel_zero_point;
692*4bdc9457SAndroid Build Coastguard Worker vacc += vi24 * vk24;
693*4bdc9457SAndroid Build Coastguard Worker k += 1;
694*4bdc9457SAndroid Build Coastguard Worker
695*4bdc9457SAndroid Build Coastguard Worker float vfpacc = (float) vacc * vscale;
696*4bdc9457SAndroid Build Coastguard Worker
697*4bdc9457SAndroid Build Coastguard Worker vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
698*4bdc9457SAndroid Build Coastguard Worker vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
699*4bdc9457SAndroid Build Coastguard Worker const int32_t vrndacc = (int32_t) lrintf(vfpacc);
700*4bdc9457SAndroid Build Coastguard Worker int32_t vout = vrndacc + voutput_zero_point;
701*4bdc9457SAndroid Build Coastguard Worker
702*4bdc9457SAndroid Build Coastguard Worker *output++ = (uint8_t) vout;
703*4bdc9457SAndroid Build Coastguard Worker } while (--c != 0);
704*4bdc9457SAndroid Build Coastguard Worker }
705*4bdc9457SAndroid Build Coastguard Worker
706*4bdc9457SAndroid Build Coastguard Worker output = (uint8_t*) ((uintptr_t) output + output_increment);
707*4bdc9457SAndroid Build Coastguard Worker } while (--output_width != 0);
708*4bdc9457SAndroid Build Coastguard Worker }
709