1 // Copyright 2019 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 <xmmintrin.h>
9
10 #include <xnnpack/maxpool.h>
11
12
xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4(size_t output_pixels,size_t kernel_elements,size_t channels,const float ** input,size_t input_offset,float * output,size_t input_increment,size_t output_increment,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])13 void xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4(
14 size_t output_pixels,
15 size_t kernel_elements,
16 size_t channels,
17 const float** input,
18 size_t input_offset,
19 float* output,
20 size_t input_increment,
21 size_t output_increment,
22 const union xnn_f32_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 __m128 voutput_max = _mm_load_ps(params->sse.max);
29 const __m128 voutput_min = _mm_load_ps(params->sse.min);
30 do {
31 float* o = output;
32 {
33 const float* i0 = *input++;
34 const float* i1 = *input++;
35 const float* i2 = *input++;
36 const float* i3 = *input++;
37 const float* i4 = *input++;
38 const float* i5 = *input++;
39 const float* i6 = *input++;
40 const float* i7 = *input++;
41 const float* i8 = *input++;
42 i0 = (const float*) ((uintptr_t) i0 + input_offset);
43 i1 = (const float*) ((uintptr_t) i1 + input_offset);
44 i2 = (const float*) ((uintptr_t) i2 + input_offset);
45 i3 = (const float*) ((uintptr_t) i3 + input_offset);
46 i4 = (const float*) ((uintptr_t) i4 + input_offset);
47 i5 = (const float*) ((uintptr_t) i5 + input_offset);
48 i6 = (const float*) ((uintptr_t) i6 + input_offset);
49 i7 = (const float*) ((uintptr_t) i7 + input_offset);
50 i8 = (const float*) ((uintptr_t) i8 + input_offset);
51 if (kernel_elements < 2) {
52 i1 = i0;
53 }
54 if (kernel_elements <= 2) {
55 i2 = i0;
56 }
57 if (kernel_elements < 4) {
58 i3 = i0;
59 }
60 if (kernel_elements <= 4) {
61 i4 = i0;
62 }
63 if (kernel_elements < 6) {
64 i5 = i0;
65 }
66 if (kernel_elements <= 6) {
67 i6 = i0;
68 }
69 if (kernel_elements < 8) {
70 i7 = i0;
71 }
72 if (kernel_elements <= 8) {
73 i8 = i0;
74 }
75
76 size_t c = channels;
77 for (; c >= 4; c -= 4) {
78 const __m128 vi0 = _mm_loadu_ps(i0);
79 i0 += 4;
80 const __m128 vi1 = _mm_loadu_ps(i1);
81 i1 += 4;
82 const __m128 vi2 = _mm_loadu_ps(i2);
83 i2 += 4;
84 const __m128 vi3 = _mm_loadu_ps(i3);
85 i3 += 4;
86 const __m128 vi4 = _mm_loadu_ps(i4);
87 i4 += 4;
88 const __m128 vi5 = _mm_loadu_ps(i5);
89 i5 += 4;
90 const __m128 vi6 = _mm_loadu_ps(i6);
91 i6 += 4;
92 const __m128 vi7 = _mm_loadu_ps(i7);
93 i7 += 4;
94 const __m128 vi8 = _mm_loadu_ps(i8);
95 i8 += 4;
96
97 const __m128 vmax018 = _mm_max_ps(_mm_max_ps(vi0, vi1), vi8);
98 const __m128 vmax23 = _mm_max_ps(vi2, vi3);
99 const __m128 vmax45 = _mm_max_ps(vi4, vi5);
100 const __m128 vmax67 = _mm_max_ps(vi6, vi7);
101
102 const __m128 vmax2345 = _mm_max_ps(vmax23, vmax45);
103 const __m128 vmax01678 = _mm_max_ps(vmax018, vmax67);
104 const __m128 vmax = _mm_max_ps(vmax2345, vmax01678);
105 const __m128 vout = _mm_max_ps(_mm_min_ps(vmax, voutput_max), voutput_min);
106
107 _mm_storeu_ps(o, vout);
108 o += 4;
109 }
110 if (c != 0) {
111 const __m128 vi0 = _mm_loadu_ps(i0);
112 i0 += 4;
113 const __m128 vi1 = _mm_loadu_ps(i1);
114 i1 += 4;
115 const __m128 vi2 = _mm_loadu_ps(i2);
116 i2 += 4;
117 const __m128 vi3 = _mm_loadu_ps(i3);
118 i3 += 4;
119 const __m128 vi4 = _mm_loadu_ps(i4);
120 i4 += 4;
121 const __m128 vi5 = _mm_loadu_ps(i5);
122 i5 += 4;
123 const __m128 vi6 = _mm_loadu_ps(i6);
124 i6 += 4;
125 const __m128 vi7 = _mm_loadu_ps(i7);
126 i7 += 4;
127 const __m128 vi8 = _mm_loadu_ps(i8);
128 i8 += 4;
129
130 const __m128 vmax018 = _mm_max_ps(_mm_max_ps(vi0, vi1), vi8);
131 const __m128 vmax23 = _mm_max_ps(vi2, vi3);
132 const __m128 vmax45 = _mm_max_ps(vi4, vi5);
133 const __m128 vmax67 = _mm_max_ps(vi6, vi7);
134
135 const __m128 vmax2345 = _mm_max_ps(vmax23, vmax45);
136 const __m128 vmax01678 = _mm_max_ps(vmax018, vmax67);
137 const __m128 vmax = _mm_max_ps(vmax2345, vmax01678);
138 __m128 vout = _mm_max_ps(_mm_min_ps(vmax, voutput_max), voutput_min);
139
140 if (c & 2) {
141 _mm_storel_pi((__m64*) o, vout);
142 o += 2;
143 vout = _mm_movehl_ps(vout, vout);
144 }
145 if (c & 1) {
146 _mm_store_ss(o, vout);
147 o += 1;
148 }
149 }
150 }
151
152 for (ptrdiff_t k = (ptrdiff_t) kernel_elements - 9; k > 0; k -= 8) {
153 const float* i0 = *input++;
154 const float* i1 = *input++;
155 const float* i2 = *input++;
156 const float* i3 = *input++;
157 const float* i4 = *input++;
158 const float* i5 = *input++;
159 const float* i6 = *input++;
160 const float* i7 = *input++;
161 i0 = (const float*) ((uintptr_t) i0 + input_offset);
162 i1 = (const float*) ((uintptr_t) i1 + input_offset);
163 i2 = (const float*) ((uintptr_t) i2 + input_offset);
164 i3 = (const float*) ((uintptr_t) i3 + input_offset);
165 i4 = (const float*) ((uintptr_t) i4 + input_offset);
166 i5 = (const float*) ((uintptr_t) i5 + input_offset);
167 i6 = (const float*) ((uintptr_t) i6 + input_offset);
168 i7 = (const float*) ((uintptr_t) i7 + input_offset);
169 if (k < 2) {
170 i1 = i0;
171 }
172 if (k <= 2) {
173 i2 = i0;
174 }
175 if (k < 4) {
176 i3 = i0;
177 }
178 if (k <= 4) {
179 i4 = i0;
180 }
181 if (k < 6) {
182 i5 = i0;
183 }
184 if (k <= 6) {
185 i6 = i0;
186 }
187 if (k < 8) {
188 i7 = i0;
189 }
190
191 o = output;
192 size_t c = channels;
193 for (; c >= 4; c -= 4) {
194 const __m128 vi0 = _mm_loadu_ps(i0);
195 i0 += 4;
196 const __m128 vi1 = _mm_loadu_ps(i1);
197 i1 += 4;
198 const __m128 vi2 = _mm_loadu_ps(i2);
199 i2 += 4;
200 const __m128 vi3 = _mm_loadu_ps(i3);
201 i3 += 4;
202 const __m128 vi4 = _mm_loadu_ps(i4);
203 i4 += 4;
204 const __m128 vi5 = _mm_loadu_ps(i5);
205 i5 += 4;
206 const __m128 vi6 = _mm_loadu_ps(i6);
207 i6 += 4;
208 const __m128 vi7 = _mm_loadu_ps(i7);
209 i7 += 4;
210 const __m128 vo = _mm_loadu_ps(o);
211
212 const __m128 vmax01 = _mm_max_ps(_mm_max_ps(vi0, vi1), vo);
213 const __m128 vmax23 = _mm_max_ps(vi2, vi3);
214 const __m128 vmax45 = _mm_max_ps(vi4, vi5);
215 const __m128 vmax67 = _mm_max_ps(vi6, vi7);
216
217 const __m128 vmax2345 = _mm_max_ps(vmax23, vmax45);
218 const __m128 vmax0167 = _mm_max_ps(vmax01, vmax67);
219 const __m128 vmax = _mm_max_ps(vmax2345, vmax0167);
220 const __m128 vout = _mm_max_ps(_mm_min_ps(vmax, voutput_max), voutput_min);
221
222 _mm_storeu_ps(o, vout);
223 o += 4;
224 }
225 if (c != 0) {
226 const __m128 vi0 = _mm_loadu_ps(i0);
227 const __m128 vi1 = _mm_loadu_ps(i1);
228 const __m128 vi2 = _mm_loadu_ps(i2);
229 const __m128 vi3 = _mm_loadu_ps(i3);
230 const __m128 vi4 = _mm_loadu_ps(i4);
231 const __m128 vi5 = _mm_loadu_ps(i5);
232 const __m128 vi6 = _mm_loadu_ps(i6);
233 const __m128 vi7 = _mm_loadu_ps(i7);
234 const __m128 vo = _mm_loadu_ps(o);
235
236 const __m128 vmax01 = _mm_max_ps(_mm_max_ps(vi0, vi1), vo);
237 const __m128 vmax23 = _mm_max_ps(vi2, vi3);
238 const __m128 vmax45 = _mm_max_ps(vi4, vi5);
239 const __m128 vmax67 = _mm_max_ps(vi6, vi7);
240
241 const __m128 vmax2345 = _mm_max_ps(vmax23, vmax45);
242 const __m128 vmax0167 = _mm_max_ps(vmax01, vmax67);
243 const __m128 vmax = _mm_max_ps(vmax2345, vmax0167);
244 __m128 vout = _mm_max_ps(_mm_min_ps(vmax, voutput_max), voutput_min);
245
246 if (c & 2) {
247 _mm_storel_pi((__m64*) o, vout);
248 o += 2;
249 vout = _mm_movehl_ps(vout, vout);
250 }
251 if (c & 1) {
252 _mm_store_ss(o, vout);
253 o += 1;
254 }
255 }
256 }
257 input = (const float**) ((uintptr_t) input + input_increment);
258 output = (float*) ((uintptr_t) o + output_increment);
259 } while (--output_pixels != 0);
260 }
261