1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gavgpool/multipass-sse2.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 <emmintrin.h>
13
14 #include <xnnpack/gavgpool.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17
18
xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c8(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__sse2_c8(
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, 8) * sizeof(int8_t);
40
41 const __m128i vinit_bias = _mm_load_si128((const __m128i*) params->fp32_sse2.init_bias);
42 int32_t* b = buffer;
43 size_t c = channels;
44 for (; c != 0; c = doz(c, 8)) {
45
46 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
47 i0 += 8;
48
49 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
50 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
51 i1 += 8;
52
53 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
54 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
55 i2 += 8;
56
57 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
58 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
59 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
60 i3 += 8;
61
62 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
63 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
64 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
65 i4 += 8;
66
67 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
68 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
69 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
70 i5 += 8;
71
72 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
73 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
74 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
75 i6 += 8;
76
77 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
78 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
79
80 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
81
82 const __m128i vsgnacc01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc01234567);
83 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vsgnacc01234567);
84 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vsgnacc01234567);
85
86 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
87 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
88
89 _mm_store_si128((__m128i*) b, vacc0123);
90 _mm_store_si128((__m128i*) (b + 4), vacc4567);
91 b += 8;
92 }
93
94 for (rows -= 7; rows > 7; rows -= 7) {
95 i0 = (const int8_t*) ((uintptr_t) i0 + input_increment);
96 i1 = (const int8_t*) ((uintptr_t) i1 + input_increment);
97 i2 = (const int8_t*) ((uintptr_t) i2 + input_increment);
98 i3 = (const int8_t*) ((uintptr_t) i3 + input_increment);
99 i4 = (const int8_t*) ((uintptr_t) i4 + input_increment);
100 i5 = (const int8_t*) ((uintptr_t) i5 + input_increment);
101 i6 = (const int8_t*) ((uintptr_t) i6 + input_increment);
102
103 int32_t* b = buffer;
104 size_t c = channels;
105 for (; c != 0; c = doz(c, 8)) {
106
107 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
108 i0 += 8;
109
110 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
111 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
112 i1 += 8;
113
114 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
115 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
116 i2 += 8;
117
118 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
119 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
120 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
121 i3 += 8;
122
123 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
124 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
125 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
126 i4 += 8;
127
128 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
129 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
130 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
131 i5 += 8;
132
133 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
134 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
135 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
136 i6 += 8;
137
138 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
139 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
140
141 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
142
143 const __m128i vsgnacc01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc01234567);
144 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vsgnacc01234567);
145 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vsgnacc01234567);
146
147 vacc0123 = _mm_add_epi32(vacc0123, _mm_load_si128((const __m128i*) b));
148 vacc4567 = _mm_add_epi32(vacc4567, _mm_load_si128((const __m128i*) (b + 4)));
149
150 _mm_store_si128((__m128i*) b, vacc0123);
151 _mm_store_si128((__m128i*) (b + 4), vacc4567);
152 b += 8;
153 }
154 }
155
156 i0 = (const int8_t*) ((uintptr_t) i0 + input_increment);
157 i1 = (const int8_t*) ((uintptr_t) i1 + input_increment);
158 if XNN_UNPREDICTABLE(rows < 2) {
159 i1 = zero;
160 }
161 i2 = (const int8_t*) ((uintptr_t) i2 + input_increment);
162 if XNN_UNPREDICTABLE(rows <= 2) {
163 i2 = zero;
164 }
165 i3 = (const int8_t*) ((uintptr_t) i3 + input_increment);
166 if XNN_UNPREDICTABLE(rows < 4) {
167 i3 = zero;
168 }
169 i4 = (const int8_t*) ((uintptr_t) i4 + input_increment);
170 if XNN_UNPREDICTABLE(rows <= 4) {
171 i4 = zero;
172 }
173 i5 = (const int8_t*) ((uintptr_t) i5 + input_increment);
174 if XNN_UNPREDICTABLE(rows < 6) {
175 i5 = zero;
176 }
177 i6 = (const int8_t*) ((uintptr_t) i6 + input_increment);
178 if XNN_UNPREDICTABLE(rows <= 6) {
179 i6 = zero;
180 }
181
182 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
183 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
184 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
185 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
186 for (; channels >= 8; channels -= 8) {
187
188 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
189 i0 += 8;
190
191 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
192 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
193 i1 += 8;
194
195 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
196 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
197 i2 += 8;
198
199 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
200 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
201 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
202 i3 += 8;
203
204 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
205 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
206 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
207 i4 += 8;
208
209 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
210 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
211 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
212 i5 += 8;
213
214 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
215 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
216 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
217 i6 += 8;
218
219 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
220 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
221
222 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
223
224 const __m128i vsgnacc01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc01234567);
225 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vsgnacc01234567);
226 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vsgnacc01234567);
227
228 vacc0123 = _mm_add_epi32(vacc0123, _mm_load_si128((const __m128i*) buffer));
229 vacc4567 = _mm_add_epi32(vacc4567, _mm_load_si128((const __m128i*) (buffer + 4)));
230 buffer += 8;
231
232 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
233 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
234
235 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
236 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
237
238 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
239 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
240
241 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
242 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
243
244 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
245
246 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
247
248 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
249
250
251 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
252 output += 8;
253 }
254 if XNN_UNLIKELY(channels != 0) {
255 {
256
257 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
258 i0 += 8;
259
260 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
261 i1 += 8;
262
263 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
264 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
265 i2 += 8;
266
267 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
268 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
269 i3 += 8;
270
271 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
272 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
273 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
274 i4 += 8;
275
276 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
277 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
278 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
279 i5 += 8;
280
281 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
282 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
283 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
284 i6 += 8;
285
286 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
287 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
288
289 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
290 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
291
292 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
293
294 const __m128i vsgnacc01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc01234567);
295 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vsgnacc01234567);
296 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vsgnacc01234567);
297
298 vacc0123 = _mm_add_epi32(vacc0123, _mm_load_si128((const __m128i*) buffer));
299 vacc4567 = _mm_add_epi32(vacc4567, _mm_load_si128((const __m128i*) (buffer + 4)));
300 buffer += 8;
301
302 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
303 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
304
305 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
306 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
307
308 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
309 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
310
311 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
312 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
313
314 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
315 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
316
317 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
318
319 if (channels & 4) {
320 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
321 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
322 output += 4;
323 }
324 uint32_t vout0123 = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
325 if (channels & 2) {
326 unaligned_store_u16(output, (uint16_t) vout0123);
327 vout0123 >>= 16;
328 output += 2;
329 }
330 if (channels & 1) {
331 *output = (int8_t) vout0123;
332 }
333 }
334 }
335 }
336