1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-neon-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include <xnnpack/dwconv.h>
15
16
xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__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 void xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16(
18 size_t channels,
19 size_t output_width,
20 const int8_t** input,
21 const void* weights,
22 int8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const int8_t* zero,
27 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
33 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
34 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neon.output_min);
35 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neon.output_max);
36 do {
37 const int8_t* i0 = input[0];
38 assert(i0 != NULL);
39 if XNN_UNPREDICTABLE(i0 != zero) {
40 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
41 }
42 const int8_t* i1 = input[1];
43 assert(i1 != NULL);
44 if XNN_UNPREDICTABLE(i1 != zero) {
45 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
46 }
47 const int8_t* i2 = input[2];
48 assert(i2 != NULL);
49 if XNN_UNPREDICTABLE(i2 != zero) {
50 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
51 }
52 const int8_t* i3 = input[3];
53 assert(i3 != NULL);
54 if XNN_UNPREDICTABLE(i3 != zero) {
55 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
56 }
57 const int8_t* i4 = input[4];
58 assert(i4 != NULL);
59 if XNN_UNPREDICTABLE(i4 != zero) {
60 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
61 }
62 const int8_t* i5 = input[5];
63 assert(i5 != NULL);
64 if XNN_UNPREDICTABLE(i5 != zero) {
65 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
66 }
67 const int8_t* i6 = input[6];
68 assert(i6 != NULL);
69 if XNN_UNPREDICTABLE(i6 != zero) {
70 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
71 }
72 const int8_t* i7 = input[7];
73 assert(i7 != NULL);
74 if XNN_UNPREDICTABLE(i7 != zero) {
75 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
76 }
77 const int8_t* i8 = input[8];
78 assert(i8 != NULL);
79 if XNN_UNPREDICTABLE(i8 != zero) {
80 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
81 }
82 input = (const int8_t**) ((uintptr_t) input + input_stride);
83
84 size_t c = channels;
85 const void* w = weights;
86 for (; c >= 32; c -= 32) {
87 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
88 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
89 int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
90 int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
91 int32x4_t vaccGHIJ = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
92 int32x4_t vaccKLMN = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
93 int32x4_t vaccOPQR = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
94 int32x4_t vaccSTUV = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
95
96
97 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
98 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
99 const int16x8_t vi0x89ABCDEF = vmovl_s8(vld1_s8(i0)); i0 += 8;
100 const int16x8_t vk0x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
101 const int16x8_t vi0xGHIJKLMN = vmovl_s8(vld1_s8(i0)); i0 += 8;
102 const int16x8_t vk0xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
103 const int16x8_t vi0xOPQRSTUV = vmovl_s8(vld1_s8(i0)); i0 += 8;
104 const int16x8_t vk0xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
105
106 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
107 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
108 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF));
109 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi0x89ABCDEF), vget_high_s16(vk0x89ABCDEF));
110 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi0xGHIJKLMN), vget_low_s16(vk0xGHIJKLMN));
111 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi0xGHIJKLMN), vget_high_s16(vk0xGHIJKLMN));
112 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi0xOPQRSTUV), vget_low_s16(vk0xOPQRSTUV));
113 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi0xOPQRSTUV), vget_high_s16(vk0xOPQRSTUV));
114
115 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
116 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
117 const int16x8_t vi1x89ABCDEF = vmovl_s8(vld1_s8(i1)); i1 += 8;
118 const int16x8_t vk1x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
119 const int16x8_t vi1xGHIJKLMN = vmovl_s8(vld1_s8(i1)); i1 += 8;
120 const int16x8_t vk1xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
121 const int16x8_t vi1xOPQRSTUV = vmovl_s8(vld1_s8(i1)); i1 += 8;
122 const int16x8_t vk1xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
123
124 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
125 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
126 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF));
127 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF));
128 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi1xGHIJKLMN), vget_low_s16(vk1xGHIJKLMN));
129 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi1xGHIJKLMN), vget_high_s16(vk1xGHIJKLMN));
130 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi1xOPQRSTUV), vget_low_s16(vk1xOPQRSTUV));
131 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi1xOPQRSTUV), vget_high_s16(vk1xOPQRSTUV));
132
133 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
134 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
135 const int16x8_t vi2x89ABCDEF = vmovl_s8(vld1_s8(i2)); i2 += 8;
136 const int16x8_t vk2x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
137 const int16x8_t vi2xGHIJKLMN = vmovl_s8(vld1_s8(i2)); i2 += 8;
138 const int16x8_t vk2xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
139 const int16x8_t vi2xOPQRSTUV = vmovl_s8(vld1_s8(i2)); i2 += 8;
140 const int16x8_t vk2xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
141
142 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
143 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
144 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF));
145 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF));
146 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi2xGHIJKLMN), vget_low_s16(vk2xGHIJKLMN));
147 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi2xGHIJKLMN), vget_high_s16(vk2xGHIJKLMN));
148 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi2xOPQRSTUV), vget_low_s16(vk2xOPQRSTUV));
149 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi2xOPQRSTUV), vget_high_s16(vk2xOPQRSTUV));
150
151 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
152 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
153 const int16x8_t vi3x89ABCDEF = vmovl_s8(vld1_s8(i3)); i3 += 8;
154 const int16x8_t vk3x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
155 const int16x8_t vi3xGHIJKLMN = vmovl_s8(vld1_s8(i3)); i3 += 8;
156 const int16x8_t vk3xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
157 const int16x8_t vi3xOPQRSTUV = vmovl_s8(vld1_s8(i3)); i3 += 8;
158 const int16x8_t vk3xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
159
160 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
161 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
162 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi3x89ABCDEF), vget_low_s16(vk3x89ABCDEF));
163 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi3x89ABCDEF), vget_high_s16(vk3x89ABCDEF));
164 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN));
165 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN));
166 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi3xOPQRSTUV), vget_low_s16(vk3xOPQRSTUV));
167 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi3xOPQRSTUV), vget_high_s16(vk3xOPQRSTUV));
168
169 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
170 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
171 const int16x8_t vi4x89ABCDEF = vmovl_s8(vld1_s8(i4)); i4 += 8;
172 const int16x8_t vk4x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
173 const int16x8_t vi4xGHIJKLMN = vmovl_s8(vld1_s8(i4)); i4 += 8;
174 const int16x8_t vk4xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
175 const int16x8_t vi4xOPQRSTUV = vmovl_s8(vld1_s8(i4)); i4 += 8;
176 const int16x8_t vk4xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
177
178 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
179 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
180 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF));
181 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF));
182 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi4xGHIJKLMN), vget_low_s16(vk4xGHIJKLMN));
183 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi4xGHIJKLMN), vget_high_s16(vk4xGHIJKLMN));
184 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV));
185 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV));
186
187 const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
188 const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
189 const int16x8_t vi5x89ABCDEF = vmovl_s8(vld1_s8(i5)); i5 += 8;
190 const int16x8_t vk5x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
191 const int16x8_t vi5xGHIJKLMN = vmovl_s8(vld1_s8(i5)); i5 += 8;
192 const int16x8_t vk5xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
193 const int16x8_t vi5xOPQRSTUV = vmovl_s8(vld1_s8(i5)); i5 += 8;
194 const int16x8_t vk5xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
195
196 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
197 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
198 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi5x89ABCDEF), vget_low_s16(vk5x89ABCDEF));
199 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi5x89ABCDEF), vget_high_s16(vk5x89ABCDEF));
200 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi5xGHIJKLMN), vget_low_s16(vk5xGHIJKLMN));
201 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi5xGHIJKLMN), vget_high_s16(vk5xGHIJKLMN));
202 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi5xOPQRSTUV), vget_low_s16(vk5xOPQRSTUV));
203 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi5xOPQRSTUV), vget_high_s16(vk5xOPQRSTUV));
204
205 const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6)); i6 += 8;
206 const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
207 const int16x8_t vi6x89ABCDEF = vmovl_s8(vld1_s8(i6)); i6 += 8;
208 const int16x8_t vk6x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
209 const int16x8_t vi6xGHIJKLMN = vmovl_s8(vld1_s8(i6)); i6 += 8;
210 const int16x8_t vk6xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
211 const int16x8_t vi6xOPQRSTUV = vmovl_s8(vld1_s8(i6)); i6 += 8;
212 const int16x8_t vk6xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
213
214 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
215 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
216 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi6x89ABCDEF), vget_low_s16(vk6x89ABCDEF));
217 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi6x89ABCDEF), vget_high_s16(vk6x89ABCDEF));
218 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi6xGHIJKLMN), vget_low_s16(vk6xGHIJKLMN));
219 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi6xGHIJKLMN), vget_high_s16(vk6xGHIJKLMN));
220 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi6xOPQRSTUV), vget_low_s16(vk6xOPQRSTUV));
221 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi6xOPQRSTUV), vget_high_s16(vk6xOPQRSTUV));
222
223 const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7)); i7 += 8;
224 const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
225 const int16x8_t vi7x89ABCDEF = vmovl_s8(vld1_s8(i7)); i7 += 8;
226 const int16x8_t vk7x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
227 const int16x8_t vi7xGHIJKLMN = vmovl_s8(vld1_s8(i7)); i7 += 8;
228 const int16x8_t vk7xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
229 const int16x8_t vi7xOPQRSTUV = vmovl_s8(vld1_s8(i7)); i7 += 8;
230 const int16x8_t vk7xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
231
232 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
233 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
234 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi7x89ABCDEF), vget_low_s16(vk7x89ABCDEF));
235 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi7x89ABCDEF), vget_high_s16(vk7x89ABCDEF));
236 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN));
237 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN));
238 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi7xOPQRSTUV), vget_low_s16(vk7xOPQRSTUV));
239 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi7xOPQRSTUV), vget_high_s16(vk7xOPQRSTUV));
240
241 const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8)); i8 += 8;
242 const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
243 const int16x8_t vi8x89ABCDEF = vmovl_s8(vld1_s8(i8)); i8 += 8;
244 const int16x8_t vk8x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
245 const int16x8_t vi8xGHIJKLMN = vmovl_s8(vld1_s8(i8)); i8 += 8;
246 const int16x8_t vk8xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
247 const int16x8_t vi8xOPQRSTUV = vmovl_s8(vld1_s8(i8)); i8 += 8;
248 const int16x8_t vk8xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
249
250 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
251 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
252 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi8x89ABCDEF), vget_low_s16(vk8x89ABCDEF));
253 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi8x89ABCDEF), vget_high_s16(vk8x89ABCDEF));
254 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi8xGHIJKLMN), vget_low_s16(vk8xGHIJKLMN));
255 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi8xGHIJKLMN), vget_high_s16(vk8xGHIJKLMN));
256 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi8xOPQRSTUV), vget_low_s16(vk8xOPQRSTUV));
257 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi8xOPQRSTUV), vget_high_s16(vk8xOPQRSTUV));
258
259 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
260 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
261 float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
262 float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
263 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ);
264 float32x4_t vfpaccKLMN = vcvtq_f32_s32(vaccKLMN);
265 float32x4_t vfpaccOPQR = vcvtq_f32_s32(vaccOPQR);
266 float32x4_t vfpaccSTUV = vcvtq_f32_s32(vaccSTUV);
267
268 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
269 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
270 const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
271 const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
272 const float32x4_t vscaleGHIJ = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
273 const float32x4_t vscaleKLMN = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
274 const float32x4_t vscaleOPQR = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
275 const float32x4_t vscaleSTUV = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
276
277 vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
278 vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
279 vfpacc89AB = vmulq_f32(vfpacc89AB, vscale89AB);
280 vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscaleCDEF);
281 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscaleGHIJ);
282 vfpaccKLMN = vmulq_f32(vfpaccKLMN, vscaleKLMN);
283 vfpaccOPQR = vmulq_f32(vfpaccOPQR, vscaleOPQR);
284 vfpaccSTUV = vmulq_f32(vfpaccSTUV, vscaleSTUV);
285
286 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
287 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
288 vacc89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc89AB, vmagic_bias));
289 vaccCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpaccCDEF, vmagic_bias));
290 vaccGHIJ = vreinterpretq_s32_f32(vaddq_f32(vfpaccGHIJ, vmagic_bias));
291 vaccKLMN = vreinterpretq_s32_f32(vaddq_f32(vfpaccKLMN, vmagic_bias));
292 vaccOPQR = vreinterpretq_s32_f32(vaddq_f32(vfpaccOPQR, vmagic_bias));
293 vaccSTUV = vreinterpretq_s32_f32(vaddq_f32(vfpaccSTUV, vmagic_bias));
294
295 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
296 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
297 vacc89AB = vqsubq_s32(vacc89AB, vmagic_bias_less_output_zero_point);
298 vaccCDEF = vqsubq_s32(vaccCDEF, vmagic_bias_less_output_zero_point);
299 vaccGHIJ = vqsubq_s32(vaccGHIJ, vmagic_bias_less_output_zero_point);
300 vaccKLMN = vqsubq_s32(vaccKLMN, vmagic_bias_less_output_zero_point);
301 vaccOPQR = vqsubq_s32(vaccOPQR, vmagic_bias_less_output_zero_point);
302 vaccSTUV = vqsubq_s32(vaccSTUV, vmagic_bias_less_output_zero_point);
303
304 #if XNN_ARCH_ARM64
305 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
306 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
307 int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN);
308 int16x8_t vaccOPQRSTUV = vqmovn_high_s32(vqmovn_s32(vaccOPQR), vaccSTUV);
309
310
311 int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
312 int8x16_t voutGHIJKLMNOPQRSTUV = vqmovn_high_s16(vqmovn_s16(vaccGHIJKLMN), vaccOPQRSTUV);
313 #else // !XNN_ARCH_ARM64
314 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
315 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
316 int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN));
317 int16x8_t vaccOPQRSTUV = vcombine_s16(vqmovn_s32(vaccOPQR), vqmovn_s32(vaccSTUV));
318
319
320 int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
321 int8x16_t voutGHIJKLMNOPQRSTUV = vcombine_s8(vqmovn_s16(vaccGHIJKLMN), vqmovn_s16(vaccOPQRSTUV));
322 #endif // !XNN_ARCH_ARM64
323
324 vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
325 voutGHIJKLMNOPQRSTUV = vmaxq_s8(voutGHIJKLMNOPQRSTUV, voutput_min);
326
327 vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
328 voutGHIJKLMNOPQRSTUV = vminq_s8(voutGHIJKLMNOPQRSTUV, voutput_max);
329
330 vst1q_s8(output, vout0123456789ABCDEF); output += 16;
331 vst1q_s8(output, voutGHIJKLMNOPQRSTUV); output += 16;
332 }
333 if XNN_UNLIKELY(c != 0) {
334 const int8_t* k = (const int8_t*) ((const int32_t*) w + 32);
335 do {
336 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
337 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
338
339 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
340 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(k)); k += 8;
341
342 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
343 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
344 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
345 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8((const void*) (k + 24)));
346
347 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
348 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
349 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
350 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8((const void*) (k + 56)));
351
352 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
353 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
354 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
355 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8((const void*) (k + 88)));
356
357 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
358 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
359 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
360 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8((const void*) (k + 120)));
361
362 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
363 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
364 const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
365 const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8((const void*) (k + 152)));
366
367 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
368 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
369 const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6)); i6 += 8;
370 const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8((const void*) (k + 184)));
371
372 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
373 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
374 const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7)); i7 += 8;
375 const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8((const void*) (k + 216)));
376
377 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
378 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
379 const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8)); i8 += 8;
380 const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8((const void*) (k + 248)));
381
382 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
383 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
384
385 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
386 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
387
388 const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 24 * sizeof(int32_t) + 288 * sizeof(int8_t)));
389 const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 24 * sizeof(int32_t) + 288 * sizeof(int8_t) + 4 * sizeof(float)));
390 vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
391 vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
392
393 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
394 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
395
396 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
397 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
398
399 #if XNN_ARCH_ARM64
400 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
401 #else
402 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
403 #endif
404
405 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
406 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
407 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
408
409 if XNN_LIKELY(c >= 8) {
410 vst1_s8(output, vout01234567); output += 8;
411 c -= 8;
412 } else {
413 if (c & 4) {
414 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
415 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
416 }
417 if (c & 2) {
418 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
419 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
420 }
421 if (c & 1) {
422 vst1_lane_s8(output, vout01234567, 0); output += 1;
423 }
424 c = 0;
425 }
426 } while (c != 0);
427 }
428
429 output = (int8_t*) ((uintptr_t) output + output_increment);
430 } while (--output_width != 0);
431 }
432