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 #include <xnnpack/intrinsics-polyfill.h>
16
17
xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__neonv8_mul16(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)])18 void xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__neonv8_mul16(
19 size_t channels,
20 size_t output_width,
21 const uint8_t** input,
22 const void* weights,
23 uint8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const uint8_t* zero,
28 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->fp32_neonv8.kernel_zero_point);
34 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neonv8.scale);
35 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
36 const uint8x8_t voutput_min = vld1_dup_u8(¶ms->fp32_neonv8.output_min);
37 const uint8x8_t voutput_max = vld1_dup_u8(¶ms->fp32_neonv8.output_max);
38 do {
39 const uint8_t* i0 = input[0];
40 assert(i0 != NULL);
41 if XNN_UNPREDICTABLE(i0 != zero) {
42 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
43 }
44 const uint8_t* i1 = input[1];
45 assert(i1 != NULL);
46 if XNN_UNPREDICTABLE(i1 != zero) {
47 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
48 }
49 const uint8_t* i2 = input[2];
50 assert(i2 != NULL);
51 if XNN_UNPREDICTABLE(i2 != zero) {
52 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
53 }
54 const uint8_t* i3 = input[3];
55 assert(i3 != NULL);
56 if XNN_UNPREDICTABLE(i3 != zero) {
57 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
58 }
59 const uint8_t* i4 = input[4];
60 assert(i4 != NULL);
61 if XNN_UNPREDICTABLE(i4 != zero) {
62 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
63 }
64 const uint8_t* i5 = input[5];
65 assert(i5 != NULL);
66 if XNN_UNPREDICTABLE(i5 != zero) {
67 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
68 }
69 const uint8_t* i6 = input[6];
70 assert(i6 != NULL);
71 if XNN_UNPREDICTABLE(i6 != zero) {
72 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
73 }
74 const uint8_t* i7 = input[7];
75 assert(i7 != NULL);
76 if XNN_UNPREDICTABLE(i7 != zero) {
77 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
78 }
79 const uint8_t* i8 = input[8];
80 assert(i8 != NULL);
81 if XNN_UNPREDICTABLE(i8 != zero) {
82 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
83 }
84 input = (const uint8_t**) ((uintptr_t) input + input_stride);
85
86 size_t c = channels;
87 const void* w = weights;
88 for (; c >= 8; c -= 8) {
89 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
90 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
91
92
93 const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
94 const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
95
96 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
97 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
98
99 const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
100 const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
101
102 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
103 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
104
105 const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
106 const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
107
108 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
109 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
110
111 const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
112 const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
113
114 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
115 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
116
117 const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
118 const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
119
120 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
121 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
122
123 const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
124 const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
125
126 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
127 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
128
129 const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
130 const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
131
132 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
133 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
134
135 const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
136 const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
137
138 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
139 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
140
141 const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
142 const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
143
144 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
145 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
146
147 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
148 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
149
150 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
151 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
152
153 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
154 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
155
156 #if XNN_ARCH_ARM64
157 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
158
159 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
160
161 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
162 #else // !XNN_ARCH_ARM64
163 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
164
165 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
166
167 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
168 #endif // !XNN_ARCH_ARM64
169
170 vout01234567 = vmax_u8(vout01234567, voutput_min);
171
172 vout01234567 = vmin_u8(vout01234567, voutput_max);
173
174 vst1_u8(output, vout01234567); output += 8;
175 }
176 if XNN_UNLIKELY(c != 0) {
177 {
178 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
179 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
180
181 const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0)));
182 const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point));
183
184 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
185 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
186 const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1)));
187 const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 8)), vkernel_zero_point));
188
189 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
190 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
191 const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2)));
192 const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 16)), vkernel_zero_point));
193
194 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
195 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
196 const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3)));
197 const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 24)), vkernel_zero_point));
198
199 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
200 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
201 const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4)));
202 const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 32)), vkernel_zero_point));
203
204 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
205 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
206 const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5)));
207 const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 40)), vkernel_zero_point));
208
209 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
210 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
211 const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6)));
212 const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 48)), vkernel_zero_point));
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 const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7)));
217 const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 56)), vkernel_zero_point));
218
219 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
220 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
221 const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8)));
222 const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 64)), vkernel_zero_point));
223
224 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
225 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
226
227 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
228 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
229
230 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
231 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
232
233 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
234 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
235
236 #if XNN_ARCH_ARM64
237 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
238 #else
239 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
240 #endif
241 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
242
243 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
244 vout01234567 = vmax_u8(vout01234567, voutput_min);
245 vout01234567 = vmin_u8(vout01234567, voutput_max);
246
247 if (c & 4) {
248 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
249 vout01234567 = vext_u8(vout01234567, vout01234567, 4);
250 }
251 if (c & 2) {
252 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
253 vout01234567 = vext_u8(vout01234567, vout01234567, 2);
254 }
255 if (c & 1) {
256 vst1_lane_u8(output, vout01234567, 0); output += 1;
257 }
258 }
259 }
260
261 output = (uint8_t*) ((uintptr_t) output + output_increment);
262 } while (--output_width != 0);
263 }
264