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 #include <xnnpack/unaligned.h>
12
13
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)])14 void xnn_s8_maxpool_minmax_ukernel_9p8x__sse41_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 voutput_max = _mm_load_si128((const __m128i*) params->sse4.max);
30 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4.min);
31
32 do {
33 int8_t* o = output;
34 {
35 const int8_t* i0 = *input++;
36 const int8_t* i1 = *input++;
37 const int8_t* i2 = *input++;
38 const int8_t* i3 = *input++;
39 const int8_t* i4 = *input++;
40 const int8_t* i5 = *input++;
41 const int8_t* i6 = *input++;
42 const int8_t* i7 = *input++;
43 const int8_t* i8 = *input++;
44 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
45 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
46 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
47 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
48 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
49 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
50 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
51 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
52 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
53 if (kernel_elements < 2) {
54 i1 = i0;
55 }
56 if (kernel_elements <= 2) {
57 i2 = i0;
58 }
59 if (kernel_elements < 4) {
60 i3 = i0;
61 }
62 if (kernel_elements <= 4) {
63 i4 = i0;
64 }
65 if (kernel_elements < 6) {
66 i5 = i0;
67 }
68 if (kernel_elements <= 6) {
69 i6 = i0;
70 }
71 if (kernel_elements < 8) {
72 i7 = i0;
73 }
74 if (kernel_elements <= 8) {
75 i8 = i0;
76 }
77
78 size_t c = channels;
79 for (; c >= 16; c -= 16) {
80 const __m128i vi0 = _mm_loadu_si128((const __m128i*) i0); i0 += 16;
81 const __m128i vi1 = _mm_loadu_si128((const __m128i*) i1); i1 += 16;
82 const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); i2 += 16;
83 const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); i3 += 16;
84 const __m128i vi4 = _mm_loadu_si128((const __m128i*) i4); i4 += 16;
85 const __m128i vi5 = _mm_loadu_si128((const __m128i*) i5); i5 += 16;
86 const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6); i6 += 16;
87 const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7); i7 += 16;
88 const __m128i vi8 = _mm_loadu_si128((const __m128i*) i8); i8 += 16;
89
90 const __m128i vmax018 = _mm_max_epi8(_mm_max_epi8(vi0, vi1), vi8);
91 const __m128i vmax23 = _mm_max_epi8(vi2, vi3);
92 const __m128i vmax45 = _mm_max_epi8(vi4, vi5);
93 const __m128i vmax67 = _mm_max_epi8(vi6, vi7);
94
95 const __m128i vmax2345 = _mm_max_epi8(vmax23, vmax45);
96 const __m128i vmax01678 = _mm_max_epi8(vmax018, vmax67);
97 __m128i vout = _mm_max_epi8(vmax2345, vmax01678);
98 vout = _mm_max_epi8(vout, voutput_min);
99 vout = _mm_min_epi8(vout, voutput_max);
100
101 _mm_storeu_si128((__m128i*) o, vout); o += 16;
102 }
103 if (c != 0) {
104 const __m128i vi0 = _mm_loadu_si128((const __m128i*) i0);
105 const __m128i vi1 = _mm_loadu_si128((const __m128i*) i1);
106 const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2);
107 const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3);
108 const __m128i vi4 = _mm_loadu_si128((const __m128i*) i4);
109 const __m128i vi5 = _mm_loadu_si128((const __m128i*) i5);
110 const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6);
111 const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7);
112 const __m128i vi8 = _mm_loadu_si128((const __m128i*) i8);
113
114 const __m128i vmax018 = _mm_max_epi8(_mm_max_epi8(vi0, vi1), vi8);
115 const __m128i vmax23 = _mm_max_epi8(vi2, vi3);
116 const __m128i vmax45 = _mm_max_epi8(vi4, vi5);
117 const __m128i vmax67 = _mm_max_epi8(vi6, vi7);
118
119 const __m128i vmax2345 = _mm_max_epi8(vmax23, vmax45);
120 const __m128i vmax01678 = _mm_max_epi8(vmax018, vmax67);
121 __m128i vout = _mm_max_epi8(vmax2345, vmax01678);
122 vout = _mm_max_epi8(vout, voutput_min);
123 vout = _mm_min_epi8(vout, voutput_max);
124
125 if (c & 8) {
126 _mm_storel_epi64((__m128i*) o, vout);
127 vout = _mm_unpackhi_epi64(vout, vout);
128 o += 8;
129 }
130 if (c & 4) {
131 unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(vout));
132 vout = _mm_srli_epi64(vout, 32);
133 o += 4;
134 }
135 if (c & 2) {
136 unaligned_store_u16(o, (uint16_t) _mm_extract_epi16(vout, 0));
137 vout = _mm_srli_epi32(vout, 16);
138 o += 2;
139 }
140 if (c & 1) {
141 *o = (int8_t) _mm_cvtsi128_si32(vout);
142 o += 1;
143 }
144 }
145 }
146
147 for (ptrdiff_t k = (ptrdiff_t) kernel_elements - 9; k > 0; k -= 8) {
148 const int8_t* i0 = *input++;
149 const int8_t* i1 = *input++;
150 const int8_t* i2 = *input++;
151 const int8_t* i3 = *input++;
152 const int8_t* i4 = *input++;
153 const int8_t* i5 = *input++;
154 const int8_t* i6 = *input++;
155 const int8_t* i7 = *input++;
156 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
157 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
158 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
159 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
160 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
161 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
162 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
163 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
164 if (k < 2) {
165 i1 = i0;
166 }
167 if (k <= 2) {
168 i2 = i0;
169 }
170 if (k < 4) {
171 i3 = i0;
172 }
173 if (k <= 4) {
174 i4 = i0;
175 }
176 if (k < 6) {
177 i5 = i0;
178 }
179 if (k <= 6) {
180 i6 = i0;
181 }
182 if (k < 8) {
183 i7 = i0;
184 }
185
186 o = output;
187 size_t c = channels;
188 for (; c >= 16; c -= 16) {
189 const __m128i vi0 = _mm_loadu_si128((const __m128i*) i0); i0 += 16;
190 const __m128i vi1 = _mm_loadu_si128((const __m128i*) i1); i1 += 16;
191 const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); i2 += 16;
192 const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); i3 += 16;
193 const __m128i vi4 = _mm_loadu_si128((const __m128i*) i4); i4 += 16;
194 const __m128i vi5 = _mm_loadu_si128((const __m128i*) i5); i5 += 16;
195 const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6); i6 += 16;
196 const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7); i7 += 16;
197 const __m128i vo = _mm_loadu_si128((const __m128i*) o);
198
199 const __m128i vmax01 = _mm_max_epi8(_mm_max_epi8(vi0, vi1), vo);
200 const __m128i vmax23 = _mm_max_epi8(vi2, vi3);
201 const __m128i vmax45 = _mm_max_epi8(vi4, vi5);
202 const __m128i vmax67 = _mm_max_epi8(vi6, vi7);
203
204 const __m128i vmax2345 = _mm_max_epi8(vmax23, vmax45);
205 const __m128i vmax0167 = _mm_max_epi8(vmax01, vmax67);
206 __m128i vout = _mm_max_epi8(vmax2345, vmax0167);
207 vout = _mm_max_epi8(vout, voutput_min);
208 vout = _mm_min_epi8(vout, voutput_max);
209
210 _mm_storeu_si128((__m128i*) o, vout);
211 o += 16;
212 }
213 if (c != 0) {
214 const __m128i vi0 = _mm_loadu_si128((const __m128i*) i0);
215 const __m128i vi1 = _mm_loadu_si128((const __m128i*) i1);
216 const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2);
217 const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3);
218 const __m128i vi4 = _mm_loadu_si128((const __m128i*) i4);
219 const __m128i vi5 = _mm_loadu_si128((const __m128i*) i5);
220 const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6);
221 const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7);
222 const __m128i vo = _mm_loadu_si128((const __m128i*) o);
223
224 const __m128i vmax01 = _mm_max_epi8(_mm_max_epi8(vi0, vi1), vo);
225 const __m128i vmax23 = _mm_max_epi8(vi2, vi3);
226 const __m128i vmax45 = _mm_max_epi8(vi4, vi5);
227 const __m128i vmax67 = _mm_max_epi8(vi6, vi7);
228
229 const __m128i vmax2345 = _mm_max_epi8(vmax23, vmax45);
230 const __m128i vmax0167 = _mm_max_epi8(vmax01, vmax67);
231 __m128i vout = _mm_max_epi8(vmax2345, vmax0167);
232 vout = _mm_max_epi8(vout, voutput_min);
233 vout = _mm_min_epi8(vout, voutput_max);
234
235 if (c & 8) {
236 _mm_storel_epi64((__m128i*) o, vout);
237 vout = _mm_unpackhi_epi64(vout, vout);
238 o += 8;
239 }
240 if (c & 4) {
241 unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(vout));
242 vout = _mm_srli_epi64(vout, 32);
243 o += 4;
244 }
245 if (c & 2) {
246 unaligned_store_u16(o, (uint16_t) _mm_extract_epi16(vout, 0));
247 vout = _mm_srli_epi32(vout, 16);
248 o += 2;
249 }
250 if (c & 1) {
251 *o = (int8_t) _mm_cvtsi128_si32(vout);
252 o += 1;
253 }
254 }
255 }
256 input = (const int8_t**) ((uintptr_t) input + input_increment);
257 output = (int8_t*) ((uintptr_t) o + output_increment);
258 } while (--output_pixels != 0);
259 }
260