1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gavgpool/multipass-neon.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/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17
18
xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16(size_t rows,size_t channels,const int8_t * input,size_t input_stride,const int8_t * zero,int32_t * buffer,int8_t * output,const union xnn_qs8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16(
20 size_t rows,
21 size_t channels,
22 const int8_t* input,
23 size_t input_stride,
24 const int8_t* zero,
25 int32_t* buffer,
26 int8_t* output,
27 const union xnn_qs8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(rows > 7);
30 assert(channels != 0);
31
32 const int8_t* i0 = input;
33 const int8_t* i1 = (const int8_t*) ((uintptr_t) i0 + input_stride);
34 const int8_t* i2 = (const int8_t*) ((uintptr_t) i1 + input_stride);
35 const int8_t* i3 = (const int8_t*) ((uintptr_t) i2 + input_stride);
36 const int8_t* i4 = (const int8_t*) ((uintptr_t) i3 + input_stride);
37 const int8_t* i5 = (const int8_t*) ((uintptr_t) i4 + input_stride);
38 const int8_t* i6 = (const int8_t*) ((uintptr_t) i5 + input_stride);
39 const size_t input_increment = 7 * input_stride - round_up_po2(channels, 16) * sizeof(int8_t);
40
41 const int32x4_t vinit_bias = vld1q_dup_s32(¶ms->fp32_neonv8.init_bias);
42 int32_t* b = buffer;
43 size_t c = channels;
44 for (; c != 0; c = doz(c, 16)) {
45 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
46 const int8x8_t vi0x89ABCDEF = vld1_s8(i0); i0 += 8;
47 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
48 const int8x8_t vi1x89ABCDEF = vld1_s8(i1); i1 += 8;
49
50 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
51 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567);
52 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8;
53 int16x8_t vsum89ABCDEF = vaddl_s8(vi0x89ABCDEF, vi1x89ABCDEF);
54
55 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
56 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567);
57 const int8x8_t vi3x89ABCDEF = vld1_s8(i3); i3 += 8;
58 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi2x89ABCDEF);
59 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
60 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567);
61 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8;
62 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi3x89ABCDEF);
63 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
64 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567);
65 const int8x8_t vi5x89ABCDEF = vld1_s8(i5); i5 += 8;
66 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi4x89ABCDEF);
67 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
68 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567);
69 const int8x8_t vi6x89ABCDEF = vld1_s8(i6); i6 += 8;
70 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi5x89ABCDEF);
71 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567);
72 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi6x89ABCDEF);
73
74 const int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567));
75 const int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567));
76 const int32x4_t vacc89AB = vaddw_s16(vinit_bias, vget_low_s16(vsum89ABCDEF));
77 const int32x4_t vaccCDEF = vaddw_s16(vinit_bias, vget_high_s16(vsum89ABCDEF));
78
79 vst1q_s32(b, vacc0123); b += 4;
80 vst1q_s32(b, vacc4567); b += 4;
81 vst1q_s32(b, vacc89AB); b += 4;
82 vst1q_s32(b, vaccCDEF); b += 4;
83 }
84
85 for (rows -= 7; rows > 7; rows -= 7) {
86 i0 = (const int8_t*) ((uintptr_t) i0 + input_increment);
87 i1 = (const int8_t*) ((uintptr_t) i1 + input_increment);
88 i2 = (const int8_t*) ((uintptr_t) i2 + input_increment);
89 i3 = (const int8_t*) ((uintptr_t) i3 + input_increment);
90 i4 = (const int8_t*) ((uintptr_t) i4 + input_increment);
91 i5 = (const int8_t*) ((uintptr_t) i5 + input_increment);
92 i6 = (const int8_t*) ((uintptr_t) i6 + input_increment);
93
94 int32_t* b = buffer;
95 size_t c = channels;
96 for (; c != 0; c = doz(c, 16)) {
97 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
98 const int8x8_t vi0x89ABCDEF = vld1_s8(i0); i0 += 8;
99 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
100 const int8x8_t vi1x89ABCDEF = vld1_s8(i1); i1 += 8;
101
102 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
103 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567);
104 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8;
105 int16x8_t vsum89ABCDEF = vaddl_s8(vi0x89ABCDEF, vi1x89ABCDEF);
106
107 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
108 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567);
109 const int8x8_t vi3x89ABCDEF = vld1_s8(i3); i3 += 8;
110 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi2x89ABCDEF);
111 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
112 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567);
113 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8;
114 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi3x89ABCDEF);
115 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
116 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567);
117 const int8x8_t vi5x89ABCDEF = vld1_s8(i5); i5 += 8;
118 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi4x89ABCDEF);
119 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
120 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567);
121 const int8x8_t vi6x89ABCDEF = vld1_s8(i6); i6 += 8;
122 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi5x89ABCDEF);
123 int32x4_t vacc0123 = vld1q_s32(b);
124 int32x4_t vacc4567 = vld1q_s32(b + 4);
125 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567);
126 int32x4_t vacc89AB = vld1q_s32(b + 8);
127 int32x4_t vaccCDEF = vld1q_s32(b + 12);
128 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi6x89ABCDEF);
129
130 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vsum01234567));
131 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vsum01234567));
132 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vsum89ABCDEF));
133 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vsum89ABCDEF));
134
135 vst1q_s32(b, vacc0123); b += 4;
136 vst1q_s32(b, vacc4567); b += 4;
137 vst1q_s32(b, vacc89AB); b += 4;
138 vst1q_s32(b, vaccCDEF); b += 4;
139 }
140 }
141
142 i0 = (const int8_t*) ((uintptr_t) i0 + input_increment);
143 i1 = (const int8_t*) ((uintptr_t) i1 + input_increment);
144 if XNN_UNPREDICTABLE(rows < 2) {
145 i1 = zero;
146 }
147 i2 = (const int8_t*) ((uintptr_t) i2 + input_increment);
148 if XNN_UNPREDICTABLE(rows <= 2) {
149 i2 = zero;
150 }
151 i3 = (const int8_t*) ((uintptr_t) i3 + input_increment);
152 if XNN_UNPREDICTABLE(rows < 4) {
153 i3 = zero;
154 }
155 i4 = (const int8_t*) ((uintptr_t) i4 + input_increment);
156 if XNN_UNPREDICTABLE(rows <= 4) {
157 i4 = zero;
158 }
159 i5 = (const int8_t*) ((uintptr_t) i5 + input_increment);
160 if XNN_UNPREDICTABLE(rows < 6) {
161 i5 = zero;
162 }
163 i6 = (const int8_t*) ((uintptr_t) i6 + input_increment);
164 if XNN_UNPREDICTABLE(rows <= 6) {
165 i6 = zero;
166 }
167
168 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neonv8.scale);
169 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
170 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neonv8.output_min);
171 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neonv8.output_max);
172 for (; channels >= 16; channels -= 16) {
173 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
174 const int8x8_t vi0x89ABCDEF = vld1_s8(i0); i0 += 8;
175 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
176 const int8x8_t vi1x89ABCDEF = vld1_s8(i1); i1 += 8;
177
178 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
179 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567);
180 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8;
181 int16x8_t vsum89ABCDEF = vaddl_s8(vi0x89ABCDEF, vi1x89ABCDEF);
182
183 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
184 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567);
185 const int8x8_t vi3x89ABCDEF = vld1_s8(i3); i3 += 8;
186 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi2x89ABCDEF);
187 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
188 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567);
189 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8;
190 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi3x89ABCDEF);
191 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
192 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567);
193 const int8x8_t vi5x89ABCDEF = vld1_s8(i5); i5 += 8;
194 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi4x89ABCDEF);
195 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
196 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567);
197 const int8x8_t vi6x89ABCDEF = vld1_s8(i6); i6 += 8;
198 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi5x89ABCDEF);
199 int32x4_t vacc0123 = vld1q_s32(buffer); buffer += 4;
200 int32x4_t vacc4567 = vld1q_s32(buffer); buffer += 4;
201 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567);
202 int32x4_t vacc89AB = vld1q_s32(buffer); buffer += 4;
203 int32x4_t vaccCDEF = vld1q_s32(buffer); buffer += 4;
204 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi6x89ABCDEF);
205
206 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vsum01234567));
207 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vsum01234567));
208 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vsum89ABCDEF));
209 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vsum89ABCDEF));
210
211 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
212 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
213 float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
214 float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
215
216 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
217 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
218 vfpacc89AB = vmulq_f32(vfpacc89AB, vscale);
219 vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscale);
220
221 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
222 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
223 vacc89AB = vcvtnq_s32_f32(vfpacc89AB);
224 vaccCDEF = vcvtnq_s32_f32(vfpaccCDEF);
225
226 #if XNN_ARCH_ARM64
227 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
228 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
229 #else // !XNN_ARCH_ARM64
230 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
231 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
232 #endif // !XNN_ARCH_ARM64
233
234 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
235 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
236
237 #if XNN_ARCH_ARM64
238 int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
239 #else // !XNN_ARCH_ARM64
240 int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
241 #endif // !XNN_ARCH_ARM64
242
243 vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
244
245 vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
246
247 vst1q_s8(output, vout0123456789ABCDEF); output += 16;
248 }
249 if XNN_UNLIKELY(channels != 0) {
250 do {
251 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
252 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
253 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
254 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567);
255
256 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
257 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567);
258 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
259 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567);
260 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
261 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567);
262 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
263 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567);
264 int32x4_t vacc0123 = vld1q_s32(buffer); buffer += 4;
265 int32x4_t vacc4567 = vld1q_s32(buffer); buffer += 4;
266 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567);
267
268 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vsum01234567));
269 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vsum01234567));
270
271 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
272 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
273
274 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
275 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
276
277 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
278 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
279
280 #if XNN_ARCH_ARM64
281 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
282 #else
283 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
284 #endif
285 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
286
287 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
288 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
289 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
290
291 if XNN_LIKELY(channels >= 8) {
292 vst1_s8(output, vout01234567); output += 8;
293 channels -= 8;
294 } else {
295 if (channels & 4) {
296 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
297 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
298 }
299 if (channels & 2) {
300 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
301 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
302 }
303 if (channels & 1) {
304 vst1_lane_s8(output, vout01234567, 0); output += 1;
305 }
306 channels = 0;
307 }
308 } while (channels != 0);
309 }
310 }
311