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 <smmintrin.h>
9
10 #include <xnnpack/maxpool.h>
11
12
xnn_s8_maxpool_minmax_ukernel_9p8x__sse41_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)])13 void xnn_s8_maxpool_minmax_ukernel_9p8x__sse41_c16(
14 size_t output_pixels,
15 size_t kernel_elements,
16 size_t channels,
17 const int8_t** input,
18 size_t input_offset,
19 int8_t* output,
20 size_t input_increment,
21 size_t output_increment,
22 const union xnn_s8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(output_pixels != 0);
25 assert(kernel_elements != 0);
26 assert(channels != 0);
27
28 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4.max);
29 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4.min);
30
31 do {
32 int8_t* o = output;
33 {
34 const int8_t* i0 = *input++;
35 const int8_t* i1 = *input++;
36 const int8_t* i2 = *input++;
37 const int8_t* i3 = *input++;
38 const int8_t* i4 = *input++;
39 const int8_t* i5 = *input++;
40 const int8_t* i6 = *input++;
41 const int8_t* i7 = *input++;
42 const int8_t* i8 = *input++;
43 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
44 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
45 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
46 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
47 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
48 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
49 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
50 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
51 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
52 if (kernel_elements < 2) {
53 i1 = i0;
54 }
55 if (kernel_elements <= 2) {
56 i2 = i0;
57 }
58 if (kernel_elements < 4) {
59 i3 = i0;
60 }
61 if (kernel_elements <= 4) {
62 i4 = i0;
63 }
64 if (kernel_elements < 6) {
65 i5 = i0;
66 }
67 if (kernel_elements <= 6) {
68 i6 = i0;
69 }
70 if (kernel_elements < 8) {
71 i7 = i0;
72 }
73 if (kernel_elements <= 8) {
74 i8 = i0;
75 }
76
77 size_t c = channels;
78 for (; c >= 16; c -= 16) {
79 const __m128i vi0 = _mm_loadu_si128((const __m128i*) i0); i0 += 16;
80 const __m128i vi1 = _mm_loadu_si128((const __m128i*) i1); i1 += 16;
81 const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); i2 += 16;
82 const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); i3 += 16;
83 const __m128i vi4 = _mm_loadu_si128((const __m128i*) i4); i4 += 16;
84 const __m128i vi5 = _mm_loadu_si128((const __m128i*) i5); i5 += 16;
85 const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6); i6 += 16;
86 const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7); i7 += 16;
87 const __m128i vi8 = _mm_loadu_si128((const __m128i*) i8); i8 += 16;
88
89 const __m128i vmax018 = _mm_max_epi8(_mm_max_epi8(vi0, vi1), vi8);
90 const __m128i vmax23 = _mm_max_epi8(vi2, vi3);
91 const __m128i vmax45 = _mm_max_epi8(vi4, vi5);
92 const __m128i vmax67 = _mm_max_epi8(vi6, vi7);
93
94 const __m128i vmax2345 = _mm_max_epi8(vmax23, vmax45);
95 const __m128i vmax01678 = _mm_max_epi8(vmax018, vmax67);
96 __m128i vout = _mm_max_epi8(vmax2345, vmax01678);
97 vout = _mm_max_epi8(vout, voutput_min);
98 vout = _mm_min_epi8(vout, voutput_max);
99
100 _mm_storeu_si128((__m128i*) o, vout); o += 16;
101 }
102 if (c != 0) {
103 const __m128i vi0 = _mm_loadu_si128((const __m128i*) i0);
104 const __m128i vi1 = _mm_loadu_si128((const __m128i*) i1);
105 const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2);
106 const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3);
107 const __m128i vi4 = _mm_loadu_si128((const __m128i*) i4);
108 const __m128i vi5 = _mm_loadu_si128((const __m128i*) i5);
109 const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6);
110 const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7);
111 const __m128i vi8 = _mm_loadu_si128((const __m128i*) i8);
112
113 const __m128i vmax018 = _mm_max_epi8(_mm_max_epi8(vi0, vi1), vi8);
114 const __m128i vmax23 = _mm_max_epi8(vi2, vi3);
115 const __m128i vmax45 = _mm_max_epi8(vi4, vi5);
116 const __m128i vmax67 = _mm_max_epi8(vi6, vi7);
117
118 const __m128i vmax2345 = _mm_max_epi8(vmax23, vmax45);
119 const __m128i vmax01678 = _mm_max_epi8(vmax018, vmax67);
120 __m128i vout = _mm_max_epi8(vmax2345, vmax01678);
121 vout = _mm_max_epi8(vout, voutput_min);
122 vout = _mm_min_epi8(vout, voutput_max);
123
124 if (c & 8) {
125 _mm_storel_epi64((__m128i*) o, vout);
126 vout = _mm_unpackhi_epi64(vout, vout);
127 o += 8;
128 }
129 if (c & 4) {
130 *((uint32_t*) o) = (uint32_t) _mm_cvtsi128_si32(vout);
131 vout = _mm_srli_epi64(vout, 32);
132 o += 4;
133 }
134 if (c & 2) {
135 *((uint16_t*) o) = (uint16_t) _mm_extract_epi16(vout, 0);
136 vout = _mm_srli_epi32(vout, 16);
137 o += 2;
138 }
139 if (c & 1) {
140 *((int8_t*) o) = (int8_t) _mm_cvtsi128_si32(vout);
141 o += 1;
142 }
143 }
144 }
145
146 for (ptrdiff_t k = (ptrdiff_t) kernel_elements - 9; k > 0; k -= 8) {
147 const int8_t* i0 = *input++;
148 const int8_t* i1 = *input++;
149 const int8_t* i2 = *input++;
150 const int8_t* i3 = *input++;
151 const int8_t* i4 = *input++;
152 const int8_t* i5 = *input++;
153 const int8_t* i6 = *input++;
154 const int8_t* i7 = *input++;
155 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
156 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
157 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
158 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
159 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
160 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
161 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
162 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
163 if (k < 2) {
164 i1 = i0;
165 }
166 if (k <= 2) {
167 i2 = i0;
168 }
169 if (k < 4) {
170 i3 = i0;
171 }
172 if (k <= 4) {
173 i4 = i0;
174 }
175 if (k < 6) {
176 i5 = i0;
177 }
178 if (k <= 6) {
179 i6 = i0;
180 }
181 if (k < 8) {
182 i7 = i0;
183 }
184
185 o = output;
186 size_t c = channels;
187 for (; c >= 16; c -= 16) {
188 const __m128i vi0 = _mm_loadu_si128((const __m128i*) i0); i0 += 16;
189 const __m128i vi1 = _mm_loadu_si128((const __m128i*) i1); i1 += 16;
190 const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); i2 += 16;
191 const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); i3 += 16;
192 const __m128i vi4 = _mm_loadu_si128((const __m128i*) i4); i4 += 16;
193 const __m128i vi5 = _mm_loadu_si128((const __m128i*) i5); i5 += 16;
194 const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6); i6 += 16;
195 const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7); i7 += 16;
196 const __m128i vo = _mm_loadu_si128((const __m128i*) o);
197
198 const __m128i vmax01 = _mm_max_epi8(_mm_max_epi8(vi0, vi1), vo);
199 const __m128i vmax23 = _mm_max_epi8(vi2, vi3);
200 const __m128i vmax45 = _mm_max_epi8(vi4, vi5);
201 const __m128i vmax67 = _mm_max_epi8(vi6, vi7);
202
203 const __m128i vmax2345 = _mm_max_epi8(vmax23, vmax45);
204 const __m128i vmax0167 = _mm_max_epi8(vmax01, vmax67);
205 __m128i vout = _mm_max_epi8(vmax2345, vmax0167);
206 vout = _mm_max_epi8(vout, voutput_min);
207 vout = _mm_min_epi8(vout, voutput_max);
208
209 _mm_storeu_si128((__m128i*) o, vout);
210 o += 16;
211 }
212 if (c != 0) {
213 const __m128i vi0 = _mm_loadu_si128((const __m128i*) i0);
214 const __m128i vi1 = _mm_loadu_si128((const __m128i*) i1);
215 const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2);
216 const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3);
217 const __m128i vi4 = _mm_loadu_si128((const __m128i*) i4);
218 const __m128i vi5 = _mm_loadu_si128((const __m128i*) i5);
219 const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6);
220 const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7);
221 const __m128i vo = _mm_loadu_si128((const __m128i*) o);
222
223 const __m128i vmax01 = _mm_max_epi8(_mm_max_epi8(vi0, vi1), vo);
224 const __m128i vmax23 = _mm_max_epi8(vi2, vi3);
225 const __m128i vmax45 = _mm_max_epi8(vi4, vi5);
226 const __m128i vmax67 = _mm_max_epi8(vi6, vi7);
227
228 const __m128i vmax2345 = _mm_max_epi8(vmax23, vmax45);
229 const __m128i vmax0167 = _mm_max_epi8(vmax01, vmax67);
230 __m128i vout = _mm_max_epi8(vmax2345, vmax0167);
231 vout = _mm_max_epi8(vout, voutput_min);
232 vout = _mm_min_epi8(vout, voutput_max);
233
234 if (c & 8) {
235 _mm_storel_epi64((__m128i*) o, vout);
236 vout = _mm_unpackhi_epi64(vout, vout);
237 o += 8;
238 }
239 if (c & 4) {
240 *((uint32_t*) o) = (uint32_t) _mm_cvtsi128_si32(vout);
241 vout = _mm_srli_epi64(vout, 32);
242 o += 4;
243 }
244 if (c & 2) {
245 *((uint16_t*) o) = (uint16_t) _mm_extract_epi16(vout, 0);
246 vout = _mm_srli_epi32(vout, 16);
247 o += 2;
248 }
249 if (c & 1) {
250 *((int8_t*) o) = (int8_t) _mm_cvtsi128_si32(vout);
251 o += 1;
252 }
253 }
254 }
255 input = (const int8_t**) ((uintptr_t) input + input_increment);
256 output = (int8_t*) ((uintptr_t) o + output_increment);
257 } while (--output_pixels != 0);
258 }
259