1 // Copyright (c) Facebook, Inc. and its affiliates.
2 // All rights reserved.
3 //
4 // Copyright 2019 Google LLC
5 //
6 // This source code is licensed under the BSD-style license found in the
7 // LICENSE file in the root directory of this source tree.
8
9 #include <assert.h>
10
11 #include <arm_neon.h>
12
13 #include <xnnpack/avgpool.h>
14 #include <xnnpack/common.h>
15
16
xnn_q8_avgpool_ukernel_mp9p8q__neon(size_t n,size_t ks,size_t kc,const uint8_t ** input,const uint8_t * zero,int32_t * buffer,uint8_t * output,size_t input_increment,size_t output_increment,const union xnn_q8_avgpool_params params[restrict static1])17 void xnn_q8_avgpool_ukernel_mp9p8q__neon(
18 size_t n,
19 size_t ks,
20 size_t kc,
21 const uint8_t** input,
22 const uint8_t* zero,
23 int32_t* buffer,
24 uint8_t* output,
25 size_t input_increment,
26 size_t output_increment,
27 const union xnn_q8_avgpool_params params[restrict static 1])
28 {
29 assert(n != 0);
30 assert(ks > 9);
31 assert(kc != 0);
32
33 const int32x4_t vbias = vld1q_dup_s32(¶ms->neon.bias);
34 #if XNN_ARCH_ARM64
35 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->neon.multiplier);
36 #else
37 const int32x2_t vmultiplier = vld1_dup_s32(¶ms->neon.multiplier);
38 #endif
39 const int64x2_t vleft_shift = vld1q_dup_s64(¶ms->neon.left_shift);
40 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neon.output_zero_point);
41 const uint8x8_t voutput_min = vld1_dup_u8(¶ms->neon.output_min);
42 const uint8x8_t voutput_max = vld1_dup_u8(¶ms->neon.output_max);
43
44 do {
45 {
46 const uint8_t* i0 = *input++;
47 const uint8_t* i1 = *input++;
48 const uint8_t* i2 = *input++;
49 const uint8_t* i3 = *input++;
50 const uint8_t* i4 = *input++;
51 const uint8_t* i5 = *input++;
52 const uint8_t* i6 = *input++;
53 const uint8_t* i7 = *input++;
54 const uint8_t* i8 = *input++;
55
56 int32_t* acc = buffer;
57 for (size_t k = 0; k < kc; k += 8) {
58 const uint8x8_t vi0 = vld1_u8(i0); i0 += 8;
59 const uint8x8_t vi1 = vld1_u8(i1); i1 += 8;
60 const uint8x8_t vi2 = vld1_u8(i2); i2 += 8;
61 const uint8x8_t vi3 = vld1_u8(i3); i3 += 8;
62 const uint8x8_t vi4 = vld1_u8(i4); i4 += 8;
63 const uint8x8_t vi5 = vld1_u8(i5); i5 += 8;
64 const uint8x8_t vi6 = vld1_u8(i6); i6 += 8;
65 const uint8x8_t vi7 = vld1_u8(i7); i7 += 8;
66 const uint8x8_t vi8 = vld1_u8(i8); i8 += 8;
67
68 const uint16x8_t vsum018 = vaddw_u8(vaddl_u8(vi0, vi1), vi8);
69 const uint16x8_t vsum23 = vaddl_u8(vi2, vi3);
70 const uint16x8_t vsum45 = vaddl_u8(vi4, vi5);
71 const uint16x8_t vsum67 = vaddl_u8(vi6, vi7);
72
73 const uint16x8_t vsum2345 = vaddq_u16(vsum23, vsum45);
74 const uint16x8_t vsum01678 = vaddq_u16(vsum018, vsum67);
75 const uint16x8_t vsum = vaddq_u16(vsum2345, vsum01678);
76
77 const int32x4_t vacc_lo = vaddw_s16(vbias, vreinterpret_s16_u16(vget_low_u16(vsum)));
78 const int32x4_t vacc_hi = vaddw_s16(vbias, vreinterpret_s16_u16(vget_high_u16(vsum)));
79
80 vst1q_s32(acc, vacc_lo); acc += 4;
81 vst1q_s32(acc, vacc_hi); acc += 4;
82 }
83 }
84
85 size_t m = ks;
86 for (m -= 9; m > 8; m -= 8) {
87 const uint8_t* i0 = *input++;
88 const uint8_t* i1 = *input++;
89 const uint8_t* i2 = *input++;
90 const uint8_t* i3 = *input++;
91 const uint8_t* i4 = *input++;
92 const uint8_t* i5 = *input++;
93 const uint8_t* i6 = *input++;
94 const uint8_t* i7 = *input++;
95
96 int32_t* acc = buffer;
97 for (size_t k = 0; k < kc; k += 8) {
98 const uint8x8_t vi0 = vld1_u8(i0); i0 += 8;
99 const uint8x8_t vi1 = vld1_u8(i1); i1 += 8;
100 const uint8x8_t vi2 = vld1_u8(i2); i2 += 8;
101 const uint8x8_t vi3 = vld1_u8(i3); i3 += 8;
102 const uint8x8_t vi4 = vld1_u8(i4); i4 += 8;
103 const uint8x8_t vi5 = vld1_u8(i5); i5 += 8;
104 const uint8x8_t vi6 = vld1_u8(i6); i6 += 8;
105 const uint8x8_t vi7 = vld1_u8(i7); i7 += 8;
106 int32x4_t vacc_lo = vld1q_s32(acc);
107 int32x4_t vacc_hi = vld1q_s32(acc + 4);
108
109 const uint16x8_t vsum01 = vaddl_u8(vi0, vi1);
110 const uint16x8_t vsum23 = vaddl_u8(vi2, vi3);
111 const uint16x8_t vsum45 = vaddl_u8(vi4, vi5);
112 const uint16x8_t vsum67 = vaddl_u8(vi6, vi7);
113
114 const uint16x8_t vsum0123 = vaddq_u16(vsum01, vsum23);
115 const uint16x8_t vsum4567 = vaddq_u16(vsum45, vsum67);
116 const uint16x8_t vsum = vaddq_u16(vsum0123, vsum4567);
117
118 vacc_lo = vaddw_s16(vacc_lo, vreinterpret_s16_u16(vget_low_u16(vsum)));
119 vacc_hi = vaddw_s16(vacc_hi, vreinterpret_s16_u16(vget_high_u16(vsum)));
120
121 vst1q_s32(acc, vacc_lo); acc += 4;
122 vst1q_s32(acc, vacc_hi); acc += 4;
123 }
124 }
125
126 {
127 const uint8_t* i0 = input[0];
128 const uint8_t* i1 = input[1];
129 const uint8_t* i2 = input[2];
130 const uint8_t* i3 = input[3];
131 const uint8_t* i4 = input[4];
132 const uint8_t* i5 = input[5];
133 const uint8_t* i6 = input[6];
134 const uint8_t* i7 = input[7];
135 input = (const uint8_t**) ((uintptr_t) input + input_increment);
136 if (m < 2) {
137 i1 = zero;
138 }
139 if (m <= 2) {
140 i2 = zero;
141 }
142 if (m < 4) {
143 i3 = zero;
144 }
145 if (m <= 4) {
146 i4 = zero;
147 }
148 if (m < 6) {
149 i5 = zero;
150 }
151 if (m <= 6) {
152 i6 = zero;
153 }
154 if (m != 8) {
155 i7 = zero;
156 }
157
158 size_t k = kc;
159 int32_t* acc = buffer;
160 while (k >= 8) {
161 const uint8x8_t vi0 = vld1_u8(i0); i0 += 8;
162 const uint8x8_t vi1 = vld1_u8(i1); i1 += 8;
163 const uint8x8_t vi2 = vld1_u8(i2); i2 += 8;
164 const uint8x8_t vi3 = vld1_u8(i3); i3 += 8;
165 const uint8x8_t vi4 = vld1_u8(i4); i4 += 8;
166 const uint8x8_t vi5 = vld1_u8(i5); i5 += 8;
167 const uint8x8_t vi6 = vld1_u8(i6); i6 += 8;
168 const uint8x8_t vi7 = vld1_u8(i7); i7 += 8;
169 int32x4_t vacc_lo = vld1q_s32(acc); acc += 4;
170 int32x4_t vacc_hi = vld1q_s32(acc); acc += 4;
171
172 const int16x8_t vsum01 = vreinterpretq_s16_u16(vaddl_u8(vi0, vi1));
173 const int16x8_t vsum23 = vreinterpretq_s16_u16(vaddl_u8(vi2, vi3));
174 const int16x8_t vsum45 = vreinterpretq_s16_u16(vaddl_u8(vi4, vi5));
175 const int16x8_t vsum67 = vreinterpretq_s16_u16(vaddl_u8(vi6, vi7));
176
177 const int16x8_t vsum0123 = vaddq_s16(vsum01, vsum23);
178 const int16x8_t vsum4567 = vaddq_s16(vsum45, vsum67);
179 const int16x8_t vsum = vaddq_s16(vsum0123, vsum4567);
180
181 vacc_lo = vaddw_s16(vacc_lo, vget_low_s16(vsum));
182 vacc_hi = vaddw_s16(vacc_hi, vget_high_s16(vsum));
183
184 const int32x4_t vneg_mask_lo = vreinterpretq_s32_u32(vcltq_s32(vacc_lo, vmovq_n_s32(0)));
185 const int32x4_t vneg_mask_hi = vreinterpretq_s32_u32(vcltq_s32(vacc_hi, vmovq_n_s32(0)));
186
187 #if XNN_ARCH_ARM64
188 const int64x2_t vproduct01 = vmull_s32(vget_low_s32(vacc_lo), vget_low_s32(vmultiplier));
189 const int64x2_t vproduct23 = vmull_high_s32(vacc_lo, vmultiplier);
190 const int64x2_t vproduct45 = vmull_s32(vget_low_s32(vacc_hi), vget_low_s32(vmultiplier));
191 const int64x2_t vproduct67 = vmull_high_s32(vacc_hi, vmultiplier);
192
193 const int64x2_t vadjusted_product01 = vaddw_s32(vproduct01, vget_low_s32(vneg_mask_lo));
194 const int64x2_t vadjusted_product23 = vaddw_high_s32(vproduct23, vneg_mask_lo);
195 const int64x2_t vadjusted_product45 = vaddw_s32(vproduct45, vget_low_s32(vneg_mask_hi));
196 const int64x2_t vadjusted_product67 = vaddw_high_s32(vproduct67, vneg_mask_hi);
197 #else
198 const int64x2_t vproduct01 = vmull_s32(vget_low_s32(vacc_lo), vmultiplier);
199 const int64x2_t vproduct23 = vmull_s32(vget_high_s32(vacc_lo), vmultiplier);
200 const int64x2_t vproduct45 = vmull_s32(vget_low_s32(vacc_hi), vmultiplier);
201 const int64x2_t vproduct67 = vmull_s32(vget_high_s32(vacc_hi), vmultiplier);
202
203 const int64x2_t vadjusted_product01 = vaddw_s32(vproduct01, vget_low_s32(vneg_mask_lo));
204 const int64x2_t vadjusted_product23 = vaddw_s32(vproduct23, vget_high_s32(vneg_mask_lo));
205 const int64x2_t vadjusted_product45 = vaddw_s32(vproduct45, vget_low_s32(vneg_mask_hi));
206 const int64x2_t vadjusted_product67 = vaddw_s32(vproduct67, vget_high_s32(vneg_mask_hi));
207 #endif
208
209 const int64x2_t vscaled_acc01 = vrshlq_s64(vadjusted_product01, vleft_shift);
210 const int64x2_t vscaled_acc23 = vrshlq_s64(vadjusted_product23, vleft_shift);
211 const int64x2_t vscaled_acc45 = vrshlq_s64(vadjusted_product45, vleft_shift);
212 const int64x2_t vscaled_acc67 = vrshlq_s64(vadjusted_product67, vleft_shift);
213
214 #if XNN_ARCH_ARM64
215 vacc_lo = vuzp1q_s32(vreinterpretq_s32_s64(vscaled_acc01), vreinterpretq_s32_s64(vscaled_acc23));
216 vacc_hi = vuzp1q_s32(vreinterpretq_s32_s64(vscaled_acc45), vreinterpretq_s32_s64(vscaled_acc67));
217
218 const int16x8_t vacc = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc_lo), vacc_hi), voutput_zero_point);
219 #else
220 vacc_lo = vcombine_s32(vmovn_s64(vscaled_acc01), vmovn_s64(vscaled_acc23));
221 vacc_hi = vcombine_s32(vmovn_s64(vscaled_acc45), vmovn_s64(vscaled_acc67));
222
223 const int16x8_t vacc = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi)), voutput_zero_point);
224 #endif
225
226 uint8x8_t vout = vqmovun_s16(vacc);
227 vout = vmax_u8(vout, voutput_min);
228 vout = vmin_u8(vout, voutput_max);
229
230 vst1_u8(output, vout); output += 8;
231
232 k -= 8;
233 }
234 if (k != 0) {
235 const uint8x8_t vi0 = vld1_u8(i0);
236 const uint8x8_t vi1 = vld1_u8(i1);
237 const uint8x8_t vi2 = vld1_u8(i2);
238 const uint8x8_t vi3 = vld1_u8(i3);
239 const uint8x8_t vi4 = vld1_u8(i4);
240 const uint8x8_t vi5 = vld1_u8(i5);
241 const uint8x8_t vi6 = vld1_u8(i6);
242 const uint8x8_t vi7 = vld1_u8(i7);
243 int32x4_t vacc_lo = vld1q_s32(acc); acc += 4;
244 int32x4_t vacc_hi = vld1q_s32(acc);
245
246 const int16x8_t vsum01 = vreinterpretq_s16_u16(vaddl_u8(vi0, vi1));
247 const int16x8_t vsum23 = vreinterpretq_s16_u16(vaddl_u8(vi2, vi3));
248 const int16x8_t vsum45 = vreinterpretq_s16_u16(vaddl_u8(vi4, vi5));
249 const int16x8_t vsum67 = vreinterpretq_s16_u16(vaddl_u8(vi6, vi7));
250
251 const int16x8_t vsum0123 = vaddq_s16(vsum01, vsum23);
252 const int16x8_t vsum4567 = vaddq_s16(vsum45, vsum67);
253 const int16x8_t vsum = vaddq_s16(vsum0123, vsum4567);
254
255 vacc_lo = vaddw_s16(vacc_lo, vget_low_s16(vsum));
256 vacc_hi = vaddw_s16(vacc_hi, vget_high_s16(vsum));
257
258 const int32x4_t vneg_mask_lo = vreinterpretq_s32_u32(vcltq_s32(vacc_lo, vmovq_n_s32(0)));
259 const int32x4_t vneg_mask_hi = vreinterpretq_s32_u32(vcltq_s32(vacc_hi, vmovq_n_s32(0)));
260
261 #if XNN_ARCH_ARM64
262 const int64x2_t vproduct01 = vmull_s32(vget_low_s32(vacc_lo), vget_low_s32(vmultiplier));
263 const int64x2_t vproduct23 = vmull_high_s32(vacc_lo, vmultiplier);
264 const int64x2_t vproduct45 = vmull_s32(vget_low_s32(vacc_hi), vget_low_s32(vmultiplier));
265 const int64x2_t vproduct67 = vmull_high_s32(vacc_hi, vmultiplier);
266
267 const int64x2_t vadjusted_product01 = vaddw_s32(vproduct01, vget_low_s32(vneg_mask_lo));
268 const int64x2_t vadjusted_product23 = vaddw_high_s32(vproduct23, vneg_mask_lo);
269 const int64x2_t vadjusted_product45 = vaddw_s32(vproduct45, vget_low_s32(vneg_mask_hi));
270 const int64x2_t vadjusted_product67 = vaddw_high_s32(vproduct67, vneg_mask_hi);
271 #else
272 const int64x2_t vproduct01 = vmull_s32(vget_low_s32(vacc_lo), vmultiplier);
273 const int64x2_t vproduct23 = vmull_s32(vget_high_s32(vacc_lo), vmultiplier);
274 const int64x2_t vproduct45 = vmull_s32(vget_low_s32(vacc_hi), vmultiplier);
275 const int64x2_t vproduct67 = vmull_s32(vget_high_s32(vacc_hi), vmultiplier);
276
277 const int64x2_t vadjusted_product01 = vaddw_s32(vproduct01, vget_low_s32(vneg_mask_lo));
278 const int64x2_t vadjusted_product23 = vaddw_s32(vproduct23, vget_high_s32(vneg_mask_lo));
279 const int64x2_t vadjusted_product45 = vaddw_s32(vproduct45, vget_low_s32(vneg_mask_hi));
280 const int64x2_t vadjusted_product67 = vaddw_s32(vproduct67, vget_high_s32(vneg_mask_hi));
281 #endif
282
283 const int64x2_t vscaled_acc01 = vrshlq_s64(vadjusted_product01, vleft_shift);
284 const int64x2_t vscaled_acc23 = vrshlq_s64(vadjusted_product23, vleft_shift);
285 const int64x2_t vscaled_acc45 = vrshlq_s64(vadjusted_product45, vleft_shift);
286 const int64x2_t vscaled_acc67 = vrshlq_s64(vadjusted_product67, vleft_shift);
287
288 #if XNN_ARCH_ARM64
289 vacc_lo = vuzp1q_s32(vreinterpretq_s32_s64(vscaled_acc01), vreinterpretq_s32_s64(vscaled_acc23));
290 vacc_hi = vuzp1q_s32(vreinterpretq_s32_s64(vscaled_acc45), vreinterpretq_s32_s64(vscaled_acc67));
291
292 const int16x8_t vacc = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc_lo), vacc_hi), voutput_zero_point);
293 #else
294 vacc_lo = vcombine_s32(vmovn_s64(vscaled_acc01), vmovn_s64(vscaled_acc23));
295 vacc_hi = vcombine_s32(vmovn_s64(vscaled_acc45), vmovn_s64(vscaled_acc67));
296
297 const int16x8_t vacc = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi)), voutput_zero_point);
298 #endif
299
300 uint8x8_t vout = vqmovun_s16(vacc);
301 vout = vmax_u8(vout, voutput_min);
302 vout = vmin_u8(vout, voutput_max);
303
304 if (k & 4) {
305 vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_u8(vout), 0); output += 4;
306 vout = vext_u8(vout, vout, 4);
307 }
308 if (k & 2) {
309 vst1_lane_u16(__builtin_assume_aligned(output, 1), vreinterpret_u16_u8(vout), 0); output += 2;
310 vout = vext_u8(vout, vout, 2);
311 }
312 if (k & 1) {
313 vst1_lane_u8(output, vout, 0); output += 1;
314 }
315 }
316 }
317 output = (uint8_t*) ((uintptr_t) output + output_increment);
318 } while (--n != 0);
319 }
320