1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-neon-mul8.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_qc8_dwconv_minmax_fp32_ukernel_up8x9__neonv8_mul8_ld64(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)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__neonv8_mul8_ld64(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qc8_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 int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
34 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neonv8.output_min);
35 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neonv8.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 >= 8; c -= 8) {
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
90 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
91 const int8x8_t vk0x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
92
93 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
94
95 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
96 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
97 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
98 const int8x8_t vk1x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
99
100 vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
101
102 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
103 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
104 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
105 const int8x8_t vk2x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
106
107 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
108
109 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
110 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
111 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
112 const int8x8_t vk3x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
113
114 vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
115
116 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
117 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
118 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
119 const int8x8_t vk4x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
120
121 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
122
123 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
124 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
125 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
126 const int8x8_t vk5x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
127
128 vprod01234567 = vmull_s8(vi5x01234567, vk5x01234567);
129
130 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
131 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
132 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
133 const int8x8_t vk6x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
134
135 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
136
137 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
138 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
139 const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
140 const int8x8_t vk7x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
141
142 vprod01234567 = vmull_s8(vi7x01234567, vk7x01234567);
143
144 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
145 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
146 const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
147 const int8x8_t vk8x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
148
149 vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
150
151 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
152 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
153
154 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
155 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
156
157 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
158 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
159
160 vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
161 vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
162
163 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
164 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
165
166 #if XNN_ARCH_ARM64
167 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
168
169 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
170
171 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
172 #else // !XNN_ARCH_ARM64
173 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
174
175 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
176
177 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
178 #endif // !XNN_ARCH_ARM64
179
180 vout01234567 = vmax_s8(vout01234567, voutput_min);
181
182 vout01234567 = vmin_s8(vout01234567, voutput_max);
183
184 vst1_s8(output, vout01234567); output += 8;
185 }
186 if XNN_UNLIKELY(c != 0) {
187 {
188 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
189 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
190
191 const int8x8_t vi0x01234567 = vld1_s8(i0);
192 const int8x8_t vk0x01234567 = vld1_s8(w);
193
194 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
195
196 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
197 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
198 const int8x8_t vi1x01234567 = vld1_s8(i1);
199 const int8x8_t vk1x01234567 = vld1_s8((const void*) ((const int8_t*) w + 8));
200
201 vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
202
203 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
204 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
205 const int8x8_t vi2x01234567 = vld1_s8(i2);
206 const int8x8_t vk2x01234567 = vld1_s8((const void*) ((const int8_t*) w + 16));
207
208 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
209
210 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
211 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
212 const int8x8_t vi3x01234567 = vld1_s8(i3);
213 const int8x8_t vk3x01234567 = vld1_s8((const void*) ((const int8_t*) w + 24));
214
215 vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
216
217 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
218 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
219 const int8x8_t vi4x01234567 = vld1_s8(i4);
220 const int8x8_t vk4x01234567 = vld1_s8((const void*) ((const int8_t*) w + 32));
221
222 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
223
224 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
225 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
226 const int8x8_t vi5x01234567 = vld1_s8(i5);
227 const int8x8_t vk5x01234567 = vld1_s8((const void*) ((const int8_t*) w + 40));
228
229 vprod01234567 = vmull_s8(vi5x01234567, vk5x01234567);
230
231 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
232 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
233 const int8x8_t vi6x01234567 = vld1_s8(i6);
234 const int8x8_t vk6x01234567 = vld1_s8((const void*) ((const int8_t*) w + 48));
235
236 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
237
238 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
239 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
240 const int8x8_t vi7x01234567 = vld1_s8(i7);
241 const int8x8_t vk7x01234567 = vld1_s8((const void*) ((const int8_t*) w + 56));
242
243 vprod01234567 = vmull_s8(vi7x01234567, vk7x01234567);
244
245 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
246 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
247 const int8x8_t vi8x01234567 = vld1_s8(i8);
248 const int8x8_t vk8x01234567 = vld1_s8((const void*) ((const int8_t*) w + 64));
249
250 vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
251
252 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
253 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
254
255 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
256 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
257
258 const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 0 * sizeof(int32_t) + 72 * sizeof(int8_t)));
259 const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 0 * sizeof(int32_t) + 72 * sizeof(int8_t) + 4 * sizeof(float)));
260 vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
261 vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
262
263 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
264 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
265
266 #if XNN_ARCH_ARM64
267 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
268 #else
269 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
270 #endif
271 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
272
273 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
274 vout01234567 = vmax_s8(vout01234567, voutput_min);
275 vout01234567 = vmin_s8(vout01234567, voutput_max);
276
277 if (c & 4) {
278 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
279 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
280 }
281 if (c & 2) {
282 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
283 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
284 }
285 if (c & 1) {
286 vst1_lane_s8(output, vout01234567, 0); output += 1;
287 }
288 }
289 }
290
291 output = (int8_t*) ((uintptr_t) output + output_increment);
292 } while (--output_width != 0);
293 }
294