1 // Auto-generated file. Do not edit!
2 // Template: src/f16-gavgpool/multipass-neonfp16arith.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/gavgpool.h>
15 #include <xnnpack/math.h>
16
17
xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16(size_t rows,size_t channels,const void * input,size_t input_stride,const void * zero,void * buffer,void * output,const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16(
19 size_t rows,
20 size_t channels,
21 const void* input,
22 size_t input_stride,
23 const void* zero,
24 void* buffer,
25 void* output,
26 const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
27 {
28 assert(rows > 7);
29 assert(channels != 0);
30
31 const __fp16* i0 = input;
32 const __fp16* i1 = (const __fp16*) ((uintptr_t) i0 + input_stride);
33 const __fp16* i2 = (const __fp16*) ((uintptr_t) i1 + input_stride);
34 const __fp16* i3 = (const __fp16*) ((uintptr_t) i2 + input_stride);
35 const __fp16* i4 = (const __fp16*) ((uintptr_t) i3 + input_stride);
36 const __fp16* i5 = (const __fp16*) ((uintptr_t) i4 + input_stride);
37 const __fp16* i6 = (const __fp16*) ((uintptr_t) i5 + input_stride);
38 const size_t input_increment = 7 * input_stride - round_up_po2(channels, 8) * sizeof(__fp16);
39
40 __fp16* b = buffer;
41 size_t c = channels;
42 for (; c >= 16; c -= 16) {
43 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
44 const float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8;
45 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
46 const float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8;
47
48 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
49 float16x8_t vacc01234567 = vaddq_f16(vi0x01234567, vi1x01234567);
50 const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8;
51 float16x8_t vacc89ABCDEF = vaddq_f16(vi0x89ABCDEF, vi1x89ABCDEF);
52
53 const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
54 vacc01234567 = vaddq_f16(vacc01234567, vi2x01234567);
55 const float16x8_t vi3x89ABCDEF = vld1q_f16(i3); i3 += 8;
56 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi2x89ABCDEF);
57 const float16x8_t vi4x01234567 = vld1q_f16(i4); i4 += 8;
58 vacc01234567 = vaddq_f16(vacc01234567, vi3x01234567);
59 const float16x8_t vi4x89ABCDEF = vld1q_f16(i4); i4 += 8;
60 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi3x89ABCDEF);
61 const float16x8_t vi5x01234567 = vld1q_f16(i5); i5 += 8;
62 vacc01234567 = vaddq_f16(vacc01234567, vi4x01234567);
63 const float16x8_t vi5x89ABCDEF = vld1q_f16(i5); i5 += 8;
64 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi4x89ABCDEF);
65 const float16x8_t vi6x01234567 = vld1q_f16(i6); i6 += 8;
66 vacc01234567 = vaddq_f16(vacc01234567, vi5x01234567);
67 const float16x8_t vi6x89ABCDEF = vld1q_f16(i6); i6 += 8;
68 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi5x89ABCDEF);
69 vacc01234567 = vaddq_f16(vacc01234567, vi6x01234567);
70 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi6x89ABCDEF);
71
72 vst1q_f16(b, vacc01234567); b += 8;
73 vst1q_f16(b, vacc89ABCDEF); b += 8;
74 }
75 if XNN_UNLIKELY(c != 0) {
76 do {
77 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
78 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
79 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
80 float16x8_t vacc01234567 = vaddq_f16(vi0x01234567, vi1x01234567);
81
82 const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
83 vacc01234567 = vaddq_f16(vacc01234567, vi2x01234567);
84 const float16x8_t vi4x01234567 = vld1q_f16(i4); i4 += 8;
85 vacc01234567 = vaddq_f16(vacc01234567, vi3x01234567);
86 const float16x8_t vi5x01234567 = vld1q_f16(i5); i5 += 8;
87 vacc01234567 = vaddq_f16(vacc01234567, vi4x01234567);
88 const float16x8_t vi6x01234567 = vld1q_f16(i6); i6 += 8;
89 vacc01234567 = vaddq_f16(vacc01234567, vi5x01234567);
90 vacc01234567 = vaddq_f16(vacc01234567, vi6x01234567);
91
92 vst1q_f16(b, vacc01234567); b += 8;
93
94 c = doz(c, 8);
95 } while (c != 0);
96 }
97
98 for (rows -= 7; rows > 7; rows -= 7) {
99 i0 = (const __fp16*) ((uintptr_t) i0 + input_increment);
100 i1 = (const __fp16*) ((uintptr_t) i1 + input_increment);
101 i2 = (const __fp16*) ((uintptr_t) i2 + input_increment);
102 i3 = (const __fp16*) ((uintptr_t) i3 + input_increment);
103 i4 = (const __fp16*) ((uintptr_t) i4 + input_increment);
104 i5 = (const __fp16*) ((uintptr_t) i5 + input_increment);
105 i6 = (const __fp16*) ((uintptr_t) i6 + input_increment);
106
107 __fp16* b = buffer;
108 size_t c = channels;
109 for (; c >= 16; c -= 16) {
110 float16x8_t vacc01234567 = vld1q_f16(b);
111 float16x8_t vacc89ABCDEF = vld1q_f16(b + 8);
112
113 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
114 const float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8;
115
116 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
117 vacc01234567 = vaddq_f16(vacc01234567, vi0x01234567);
118 const float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8;
119 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi0x89ABCDEF);
120 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
121 vacc01234567 = vaddq_f16(vacc01234567, vi1x01234567);
122 const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8;
123 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi1x89ABCDEF);
124 const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
125 vacc01234567 = vaddq_f16(vacc01234567, vi2x01234567);
126 const float16x8_t vi3x89ABCDEF = vld1q_f16(i3); i3 += 8;
127 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi2x89ABCDEF);
128 const float16x8_t vi4x01234567 = vld1q_f16(i4); i4 += 8;
129 vacc01234567 = vaddq_f16(vacc01234567, vi3x01234567);
130 const float16x8_t vi4x89ABCDEF = vld1q_f16(i4); i4 += 8;
131 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi3x89ABCDEF);
132 const float16x8_t vi5x01234567 = vld1q_f16(i5); i5 += 8;
133 vacc01234567 = vaddq_f16(vacc01234567, vi4x01234567);
134 const float16x8_t vi5x89ABCDEF = vld1q_f16(i5); i5 += 8;
135 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi4x89ABCDEF);
136 const float16x8_t vi6x01234567 = vld1q_f16(i6); i6 += 8;
137 vacc01234567 = vaddq_f16(vacc01234567, vi5x01234567);
138 const float16x8_t vi6x89ABCDEF = vld1q_f16(i6); i6 += 8;
139 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi5x89ABCDEF);
140 vacc01234567 = vaddq_f16(vacc01234567, vi6x01234567);
141 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi6x89ABCDEF);
142
143 vst1q_f16(b, vacc01234567); b += 8;
144 vst1q_f16(b, vacc89ABCDEF); b += 8;
145 }
146 if XNN_UNLIKELY(c != 0) {
147 do {
148 float16x8_t vacc01234567 = vld1q_f16(b);
149 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
150
151 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
152 vacc01234567 = vaddq_f16(vacc01234567, vi0x01234567);
153 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
154 vacc01234567 = vaddq_f16(vacc01234567, vi1x01234567);
155 const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
156 vacc01234567 = vaddq_f16(vacc01234567, vi2x01234567);
157 const float16x8_t vi4x01234567 = vld1q_f16(i4); i4 += 8;
158 vacc01234567 = vaddq_f16(vacc01234567, vi3x01234567);
159 const float16x8_t vi5x01234567 = vld1q_f16(i5); i5 += 8;
160 vacc01234567 = vaddq_f16(vacc01234567, vi4x01234567);
161 const float16x8_t vi6x01234567 = vld1q_f16(i6); i6 += 8;
162 vacc01234567 = vaddq_f16(vacc01234567, vi5x01234567);
163 vacc01234567 = vaddq_f16(vacc01234567, vi6x01234567);
164
165 vst1q_f16(b, vacc01234567); b += 8;
166
167 c = doz(c, 8);
168 } while (c != 0);
169 }
170 }
171
172 i0 = (const __fp16*) ((uintptr_t) i0 + input_increment);
173 i1 = (const __fp16*) ((uintptr_t) i1 + input_increment);
174 if XNN_UNPREDICTABLE(rows < 2) {
175 i1 = (const __fp16*) zero;
176 }
177 i2 = (const __fp16*) ((uintptr_t) i2 + input_increment);
178 if XNN_UNPREDICTABLE(rows <= 2) {
179 i2 = (const __fp16*) zero;
180 }
181 i3 = (const __fp16*) ((uintptr_t) i3 + input_increment);
182 if XNN_UNPREDICTABLE(rows < 4) {
183 i3 = (const __fp16*) zero;
184 }
185 i4 = (const __fp16*) ((uintptr_t) i4 + input_increment);
186 if XNN_UNPREDICTABLE(rows <= 4) {
187 i4 = (const __fp16*) zero;
188 }
189 i5 = (const __fp16*) ((uintptr_t) i5 + input_increment);
190 if XNN_UNPREDICTABLE(rows < 6) {
191 i5 = (const __fp16*) zero;
192 }
193 i6 = (const __fp16*) ((uintptr_t) i6 + input_increment);
194 if XNN_UNPREDICTABLE(rows <= 6) {
195 i6 = (const __fp16*) zero;
196 }
197
198 const float16x8_t vscale = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.scale));
199 const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.min));
200 const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.max));
201 for (; channels >= 16; channels -= 16) {
202 float16x8_t vacc01234567 = vld1q_f16(buffer); buffer = (__fp16*) buffer + 8;
203 float16x8_t vacc89ABCDEF = vld1q_f16(buffer); buffer = (__fp16*) buffer + 8;
204
205 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
206 const float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8;
207
208 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
209 vacc01234567 = vaddq_f16(vacc01234567, vi0x01234567);
210 const float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8;
211 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi0x89ABCDEF);
212 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
213 vacc01234567 = vaddq_f16(vacc01234567, vi1x01234567);
214 const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8;
215 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi1x89ABCDEF);
216 const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
217 vacc01234567 = vaddq_f16(vacc01234567, vi2x01234567);
218 const float16x8_t vi3x89ABCDEF = vld1q_f16(i3); i3 += 8;
219 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi2x89ABCDEF);
220 const float16x8_t vi4x01234567 = vld1q_f16(i4); i4 += 8;
221 vacc01234567 = vaddq_f16(vacc01234567, vi3x01234567);
222 const float16x8_t vi4x89ABCDEF = vld1q_f16(i4); i4 += 8;
223 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi3x89ABCDEF);
224 const float16x8_t vi5x01234567 = vld1q_f16(i5); i5 += 8;
225 vacc01234567 = vaddq_f16(vacc01234567, vi4x01234567);
226 const float16x8_t vi5x89ABCDEF = vld1q_f16(i5); i5 += 8;
227 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi4x89ABCDEF);
228 const float16x8_t vi6x01234567 = vld1q_f16(i6); i6 += 8;
229 vacc01234567 = vaddq_f16(vacc01234567, vi5x01234567);
230 const float16x8_t vi6x89ABCDEF = vld1q_f16(i6); i6 += 8;
231 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi5x89ABCDEF);
232 vacc01234567 = vaddq_f16(vacc01234567, vi6x01234567);
233 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi6x89ABCDEF);
234
235 vacc01234567 = vmulq_f16(vacc01234567, vscale);
236 vacc89ABCDEF = vmulq_f16(vacc89ABCDEF, vscale);
237
238 vacc01234567 = vmaxq_f16(vacc01234567, vmin);
239 vacc89ABCDEF = vmaxq_f16(vacc89ABCDEF, vmin);
240
241 vacc01234567 = vminq_f16(vacc01234567, vmax);
242 vacc89ABCDEF = vminq_f16(vacc89ABCDEF, vmax);
243
244 vst1q_f16(output, vacc01234567); output = (__fp16*) output + 8;
245 vst1q_f16(output, vacc89ABCDEF); output = (__fp16*) output + 8;
246 }
247 if XNN_UNLIKELY(channels != 0) {
248 do {
249 float16x8_t vacc01234567 = vld1q_f16(buffer); buffer = (__fp16*) buffer + 8;
250
251 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
252 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
253 vacc01234567 = vaddq_f16(vacc01234567, vi0x01234567);
254 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
255 vacc01234567 = vaddq_f16(vacc01234567, vi1x01234567);
256 const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
257 vacc01234567 = vaddq_f16(vacc01234567, vi2x01234567);
258 const float16x8_t vi4x01234567 = vld1q_f16(i4); i4 += 8;
259 vacc01234567 = vaddq_f16(vacc01234567, vi3x01234567);
260 const float16x8_t vi5x01234567 = vld1q_f16(i5); i5 += 8;
261 vacc01234567 = vaddq_f16(vacc01234567, vi4x01234567);
262 const float16x8_t vi6x01234567 = vld1q_f16(i6); i6 += 8;
263 vacc01234567 = vaddq_f16(vacc01234567, vi5x01234567);
264 vacc01234567 = vaddq_f16(vacc01234567, vi6x01234567);
265
266 vacc01234567 = vmulq_f16(vacc01234567, vscale);
267 vacc01234567 = vmaxq_f16(vacc01234567, vmin);
268 vacc01234567 = vminq_f16(vacc01234567, vmax);
269
270 if XNN_LIKELY(channels >= 8) {
271 vst1q_f16(output, vacc01234567); output = (__fp16*) output + 8;
272 channels -= 8;
273 } else {
274 float16x4_t vacc0123 = vget_low_f16(vacc01234567);
275 if (channels & 4) {
276 vst1_f16(output, vacc0123); output = (__fp16*) output + 4;
277 vacc0123 = vget_high_f16(vacc01234567);
278 }
279 if (channels & 2) {
280 vst1_lane_u32(output, vreinterpret_u32_f16(vacc0123), 0); output = (__fp16*) output + 2;
281 vacc0123 = vext_f16(vacc0123, vacc0123, 2);
282 }
283 if (channels & 1) {
284 vst1_lane_f16(output, vacc0123, 0); output = (__fp16*) output + 1;
285 }
286 channels = 0;
287 }
288 } while (channels != 0);
289 }
290 }
291