1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gavgpool/unipass-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/unaligned.h>
16
17
xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c24(size_t rows,size_t channels,const uint8_t * input,size_t input_stride,const uint8_t * zero,uint8_t * output,const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c24(
19 size_t rows,
20 size_t channels,
21 const uint8_t* input,
22 size_t input_stride,
23 const uint8_t* zero,
24 uint8_t* output,
25 const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
26 {
27 assert(rows != 0);
28 assert(rows <= 7);
29 assert(channels != 0);
30
31 const uint8_t* i0 = input;
32 const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
33 if XNN_UNPREDICTABLE(rows < 2) {
34 i1 = zero;
35 }
36 const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
37 if XNN_UNPREDICTABLE(rows <= 2) {
38 i2 = zero;
39 }
40 const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
41 if XNN_UNPREDICTABLE(rows < 4) {
42 i3 = zero;
43 }
44 const uint8_t* i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
45 if XNN_UNPREDICTABLE(rows <= 4) {
46 i4 = zero;
47 }
48 const uint8_t* i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
49 if XNN_UNPREDICTABLE(rows < 6) {
50 i5 = zero;
51 }
52 const uint8_t* i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride);
53 if XNN_UNPREDICTABLE(rows <= 6) {
54 i6 = zero;
55 }
56
57 const __m128i vinit_bias = _mm_load_si128((const __m128i*) params->fp32_sse2.init_bias);
58 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
59 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
60 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
61 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
62 const __m128i vzero = _mm_setzero_si128();
63 for (; channels >= 24; channels -= 24) {
64
65 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
66 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
67 const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16));
68 i0 += 24;
69
70 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
71 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
72 const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, vzero);
73 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
74 const __m128i vxi0xGHIJKLMN = _mm_unpacklo_epi8(vi0xGHIJKLMN, vzero);
75 const __m128i vi1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i1 + 16));
76 i1 += 24;
77
78 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
79 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
80 const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, vzero);
81 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
82 const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, vzero);
83 const __m128i vi2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i2 + 16));
84 i2 += 24;
85
86 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
87 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
88 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
89 __m128i vacc89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
90 const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, vzero);
91 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
92 __m128i vaccGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN);
93 const __m128i vxi2xGHIJKLMN = _mm_unpacklo_epi8(vi2xGHIJKLMN, vzero);
94 const __m128i vi3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i3 + 16));
95 i3 += 24;
96
97 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
98 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
99 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
100 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi2x89ABCDEF);
101 const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, vzero);
102 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
103 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi2xGHIJKLMN);
104 const __m128i vxi3xGHIJKLMN = _mm_unpacklo_epi8(vi3xGHIJKLMN, vzero);
105 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16));
106 i4 += 24;
107
108 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
109 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
110 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
111 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi3x89ABCDEF);
112 const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, vzero);
113 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
114 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi3xGHIJKLMN);
115 const __m128i vxi4xGHIJKLMN = _mm_unpacklo_epi8(vi4xGHIJKLMN, vzero);
116 const __m128i vi5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i5 + 16));
117 i5 += 24;
118
119 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
120 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
121 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
122 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi4x89ABCDEF);
123 const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, vzero);
124 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
125 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi4xGHIJKLMN);
126 const __m128i vxi5xGHIJKLMN = _mm_unpacklo_epi8(vi5xGHIJKLMN, vzero);
127 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16));
128 i6 += 24;
129
130 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
131 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
132 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi5x89ABCDEF);
133 const __m128i vxi6x89ABCDEF = _mm_unpacklo_epi8(vi6x89ABCDEF, vzero);
134 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi5xGHIJKLMN);
135 const __m128i vxi6xGHIJKLMN = _mm_unpacklo_epi8(vi6xGHIJKLMN, vzero);
136
137 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
138 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi6x89ABCDEF);
139 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi6xGHIJKLMN);
140
141 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
142 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
143 __m128i vacc89AB = _mm_unpacklo_epi16(vacc89ABCDEF, vzero);
144 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero);
145 __m128i vaccGHIJ = _mm_unpacklo_epi16(vaccGHIJKLMN, vzero);
146 __m128i vaccKLMN = _mm_unpackhi_epi16(vaccGHIJKLMN, vzero);
147
148 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
149 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
150 vacc89AB = _mm_add_epi32(vacc89AB, vinit_bias);
151 vaccCDEF = _mm_add_epi32(vaccCDEF, vinit_bias);
152 vaccGHIJ = _mm_add_epi32(vaccGHIJ, vinit_bias);
153 vaccKLMN = _mm_add_epi32(vaccKLMN, vinit_bias);
154
155 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
156 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
157 __m128 vfpacc89AB = _mm_cvtepi32_ps(vacc89AB);
158 __m128 vfpaccCDEF = _mm_cvtepi32_ps(vaccCDEF);
159 __m128 vfpaccGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
160 __m128 vfpaccKLMN = _mm_cvtepi32_ps(vaccKLMN);
161
162 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
163 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
164 vfpacc89AB = _mm_mul_ps(vfpacc89AB, vscale);
165 vfpaccCDEF = _mm_mul_ps(vfpaccCDEF, vscale);
166 vfpaccGHIJ = _mm_mul_ps(vfpaccGHIJ, vscale);
167 vfpaccKLMN = _mm_mul_ps(vfpaccKLMN, vscale);
168
169 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
170 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
171 vfpacc89AB = _mm_min_ps(vfpacc89AB, voutput_max_less_zero_point);
172 vfpaccCDEF = _mm_min_ps(vfpaccCDEF, voutput_max_less_zero_point);
173 vfpaccGHIJ = _mm_min_ps(vfpaccGHIJ, voutput_max_less_zero_point);
174 vfpaccKLMN = _mm_min_ps(vfpaccKLMN, voutput_max_less_zero_point);
175
176 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
177 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
178 vacc89AB = _mm_cvtps_epi32(vfpacc89AB);
179 vaccCDEF = _mm_cvtps_epi32(vfpaccCDEF);
180 vaccGHIJ = _mm_cvtps_epi32(vfpaccGHIJ);
181 vaccKLMN = _mm_cvtps_epi32(vfpaccKLMN);
182
183 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
184 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
185 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
186
187
188 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
189 __m128i voutGHIJKLMNGHIJKLMN = _mm_packus_epi16(voutGHIJKLMN, voutGHIJKLMN);
190
191 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
192 voutGHIJKLMNGHIJKLMN = _mm_max_epu8(voutGHIJKLMNGHIJKLMN, voutput_min);
193
194 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
195 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
196 output += 24;
197 }
198 if XNN_UNLIKELY(channels != 0) {
199 do {
200
201 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
202 i0 += 8;
203
204 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
205 i1 += 8;
206
207 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
208 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
209 i2 += 8;
210
211 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
212 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
213 i3 += 8;
214
215 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
216 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
217 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
218 i4 += 8;
219
220 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
221 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
222 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
223 i5 += 8;
224
225 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
226 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
227 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
228 i6 += 8;
229
230 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
231 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
232
233 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
234 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
235
236 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
237
238 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
239 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
240
241 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
242 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
243
244 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
245 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
246
247 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
248 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
249
250 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
251 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
252
253 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
254 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
255
256 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
257
258 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
259 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
260
261 if XNN_LIKELY(channels >= 8) {
262 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
263 output += 8;
264 channels -= 8;
265 } else {
266 if (channels & 4) {
267 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
268 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
269 output += 4;
270 }
271 uint32_t vout0123 = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
272 if (channels & 2) {
273 unaligned_store_u16(output, (uint16_t) vout0123);
274 vout0123 >>= 16;
275 output += 2;
276 }
277 if (channels & 1) {
278 *output = (uint8_t) vout0123;
279 output += 1;
280 }
281 channels = 0;
282 }
283 } while (channels != 0);
284 }
285 }
286