1 // Copyright 2021 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <assert.h>
7
8 #include <emmintrin.h>
9
10 #include <xnnpack/maxpool.h>
11 #include <xnnpack/unaligned.h>
12
13
xnn_s8_maxpool_minmax_ukernel_9p8x__sse2_c16(size_t output_pixels,size_t kernel_elements,size_t channels,const int8_t ** input,size_t input_offset,int8_t * output,size_t input_increment,size_t output_increment,const union xnn_s8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])14 void xnn_s8_maxpool_minmax_ukernel_9p8x__sse2_c16(
15 size_t output_pixels,
16 size_t kernel_elements,
17 size_t channels,
18 const int8_t** input,
19 size_t input_offset,
20 int8_t* output,
21 size_t input_increment,
22 size_t output_increment,
23 const union xnn_s8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(output_pixels != 0);
26 assert(kernel_elements != 0);
27 assert(channels != 0);
28
29 const __m128i vbias = _mm_load_si128((const __m128i*) params->sse2.bias);
30 const __m128i voutput_max_with_bias = _mm_load_si128((const __m128i*) params->sse2.max_with_bias);
31 const __m128i voutput_min_with_bias = _mm_load_si128((const __m128i*) params->sse2.min_with_bias);
32
33 do {
34 int8_t* o = output;
35 {
36 const int8_t* i0 = *input++;
37 const int8_t* i1 = *input++;
38 const int8_t* i2 = *input++;
39 const int8_t* i3 = *input++;
40 const int8_t* i4 = *input++;
41 const int8_t* i5 = *input++;
42 const int8_t* i6 = *input++;
43 const int8_t* i7 = *input++;
44 const int8_t* i8 = *input++;
45 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
46 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
47 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
48 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
49 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
50 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
51 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
52 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
53 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
54 if (kernel_elements < 2) {
55 i1 = i0;
56 }
57 if (kernel_elements <= 2) {
58 i2 = i0;
59 }
60 if (kernel_elements < 4) {
61 i3 = i0;
62 }
63 if (kernel_elements <= 4) {
64 i4 = i0;
65 }
66 if (kernel_elements < 6) {
67 i5 = i0;
68 }
69 if (kernel_elements <= 6) {
70 i6 = i0;
71 }
72 if (kernel_elements < 8) {
73 i7 = i0;
74 }
75 if (kernel_elements <= 8) {
76 i8 = i0;
77 }
78
79 size_t c = channels;
80 for (; c >= 16; c -= 16) {
81 const __m128i vi0 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i0), vbias);
82 i0 += 16;
83 const __m128i vi1 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i1), vbias);
84 i1 += 16;
85 const __m128i vi2 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i2), vbias);
86 i2 += 16;
87 const __m128i vi3 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i3), vbias);
88 i3 += 16;
89 const __m128i vi4 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i4), vbias);
90 i4 += 16;
91 const __m128i vi5 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i5), vbias);
92 i5 += 16;
93 const __m128i vi6 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i6), vbias);
94 i6 += 16;
95 const __m128i vi7 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i7), vbias);
96 i7 += 16;
97 const __m128i vi8 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i8), vbias);
98 i8 += 16;
99
100 const __m128i vmax018 = _mm_max_epu8(_mm_max_epu8(vi0, vi1), vi8);
101 const __m128i vmax23 = _mm_max_epu8(vi2, vi3);
102 const __m128i vmax45 = _mm_max_epu8(vi4, vi5);
103 const __m128i vmax67 = _mm_max_epu8(vi6, vi7);
104
105 const __m128i vmax2345 = _mm_max_epu8(vmax23, vmax45);
106 const __m128i vmax01678 = _mm_max_epu8(vmax018, vmax67);
107 __m128i vout = _mm_max_epu8(vmax2345, vmax01678);
108 vout = _mm_max_epu8(vout, voutput_min_with_bias);
109 vout = _mm_min_epu8(vout, voutput_max_with_bias);
110 vout = _mm_xor_si128(vout, vbias);
111
112 _mm_storeu_si128((__m128i*) o, vout); o += 16;
113 }
114 if (c != 0) {
115 const __m128i vi0 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i0), vbias);
116 const __m128i vi1 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i1), vbias);
117 const __m128i vi2 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i2), vbias);
118 const __m128i vi3 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i3), vbias);
119 const __m128i vi4 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i4), vbias);
120 const __m128i vi5 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i5), vbias);
121 const __m128i vi6 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i6), vbias);
122 const __m128i vi7 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i7), vbias);
123 const __m128i vi8 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i8), vbias);
124
125 const __m128i vmax018 = _mm_max_epu8(_mm_max_epu8(vi0, vi1), vi8);
126 const __m128i vmax23 = _mm_max_epu8(vi2, vi3);
127 const __m128i vmax45 = _mm_max_epu8(vi4, vi5);
128 const __m128i vmax67 = _mm_max_epu8(vi6, vi7);
129
130 const __m128i vmax2345 = _mm_max_epu8(vmax23, vmax45);
131 const __m128i vmax01678 = _mm_max_epu8(vmax018, vmax67);
132 __m128i vout = _mm_max_epu8(vmax2345, vmax01678);
133 vout = _mm_max_epu8(vout, voutput_min_with_bias);
134 vout = _mm_min_epu8(vout, voutput_max_with_bias);
135 vout = _mm_xor_si128(vout, vbias);
136
137 if (c & 8) {
138 _mm_storel_epi64((__m128i*) o, vout);
139 vout = _mm_unpackhi_epi64(vout, vout);
140 o += 8;
141 }
142 if (c & 4) {
143 unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(vout));
144 vout = _mm_srli_epi64(vout, 32);
145 o += 4;
146 }
147 if (c & 2) {
148 unaligned_store_u16(o, (uint16_t) _mm_extract_epi16(vout, 0));
149 vout = _mm_srli_epi32(vout, 16);
150 o += 2;
151 }
152 if (c & 1) {
153 *((int8_t*) o) = (int8_t) _mm_cvtsi128_si32(vout);
154 o += 1;
155 }
156 }
157 }
158
159 for (ptrdiff_t k = (ptrdiff_t) kernel_elements - 9; k > 0; k -= 8) {
160 const int8_t* i0 = *input++;
161 const int8_t* i1 = *input++;
162 const int8_t* i2 = *input++;
163 const int8_t* i3 = *input++;
164 const int8_t* i4 = *input++;
165 const int8_t* i5 = *input++;
166 const int8_t* i6 = *input++;
167 const int8_t* i7 = *input++;
168 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
169 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
170 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
171 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
172 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
173 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
174 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
175 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
176 if (k < 2) {
177 i1 = i0;
178 }
179 if (k <= 2) {
180 i2 = i0;
181 }
182 if (k < 4) {
183 i3 = i0;
184 }
185 if (k <= 4) {
186 i4 = i0;
187 }
188 if (k < 6) {
189 i5 = i0;
190 }
191 if (k <= 6) {
192 i6 = i0;
193 }
194 if (k < 8) {
195 i7 = i0;
196 }
197
198 o = output;
199 size_t c = channels;
200 for (; c >= 16; c -= 16) {
201 const __m128i vi0 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i0), vbias);
202 i0 += 16;
203 const __m128i vi1 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i1), vbias);
204 i1 += 16;
205 const __m128i vi2 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i2), vbias);
206 i2 += 16;
207 const __m128i vi3 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i3), vbias);
208 i3 += 16;
209 const __m128i vi4 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i4), vbias);
210 i4 += 16;
211 const __m128i vi5 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i5), vbias);
212 i5 += 16;
213 const __m128i vi6 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i6), vbias);
214 i6 += 16;
215 const __m128i vi7 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i7), vbias);
216 i7 += 16;
217 const __m128i vo = _mm_xor_si128(_mm_loadu_si128((const __m128i*) o), vbias);
218
219 const __m128i vmax01 = _mm_max_epu8(_mm_max_epu8(vi0, vi1), vo);
220 const __m128i vmax23 = _mm_max_epu8(vi2, vi3);
221 const __m128i vmax45 = _mm_max_epu8(vi4, vi5);
222 const __m128i vmax67 = _mm_max_epu8(vi6, vi7);
223
224 const __m128i vmax2345 = _mm_max_epu8(vmax23, vmax45);
225 const __m128i vmax0167 = _mm_max_epu8(vmax01, vmax67);
226 __m128i vout = _mm_max_epu8(vmax2345, vmax0167);
227 vout = _mm_max_epu8(vout, voutput_min_with_bias);
228 vout = _mm_min_epu8(vout, voutput_max_with_bias);
229 vout = _mm_xor_si128(vout, vbias);
230
231 _mm_storeu_si128((__m128i*) o, vout);
232 o += 16;
233 }
234 if (c != 0) {
235 const __m128i vi0 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i0), vbias);
236 const __m128i vi1 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i1), vbias);
237 const __m128i vi2 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i2), vbias);
238 const __m128i vi3 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i3), vbias);
239 const __m128i vi4 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i4), vbias);
240 const __m128i vi5 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i5), vbias);
241 const __m128i vi6 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i6), vbias);
242 const __m128i vi7 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i7), vbias);
243 const __m128i vo = _mm_xor_si128(_mm_loadu_si128((const __m128i*) o), vbias);
244
245 const __m128i vmax01 = _mm_max_epu8(_mm_max_epu8(vi0, vi1), vo);
246 const __m128i vmax23 = _mm_max_epu8(vi2, vi3);
247 const __m128i vmax45 = _mm_max_epu8(vi4, vi5);
248 const __m128i vmax67 = _mm_max_epu8(vi6, vi7);
249
250 const __m128i vmax2345 = _mm_max_epu8(vmax23, vmax45);
251 const __m128i vmax0167 = _mm_max_epu8(vmax01, vmax67);
252 __m128i vout = _mm_max_epu8(vmax2345, vmax0167);
253 vout = _mm_max_epu8(vout, voutput_min_with_bias);
254 vout = _mm_min_epu8(vout, voutput_max_with_bias);
255 vout = _mm_xor_si128(vout, vbias);
256
257 if (c & 8) {
258 _mm_storel_epi64((__m128i*) o, vout);
259 vout = _mm_unpackhi_epi64(vout, vout);
260 o += 8;
261 }
262 if (c & 4) {
263 unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(vout));
264 vout = _mm_srli_epi64(vout, 32);
265 o += 4;
266 }
267 if (c & 2) {
268 unaligned_store_u16(o, (uint16_t) _mm_extract_epi16(vout, 0));
269 vout = _mm_srli_epi32(vout, 16);
270 o += 2;
271 }
272 if (c & 1) {
273 *o = (int8_t) _mm_cvtsi128_si32(vout);
274 o += 1;
275 }
276 }
277 }
278 input = (const int8_t**) ((uintptr_t) input + input_increment);
279 output = (int8_t*) ((uintptr_t) o + output_increment);
280 } while (--output_pixels != 0);
281 }
282