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 <emmintrin.h>
9
10 #include <xnnpack/argmaxpool.h>
11
12
xnn_f32_argmaxpool_ukernel_9x__sse2_c4(size_t output_pixels,size_t pooling_elements,size_t channels,const float ** input,size_t input_offset,float * output,uint32_t * index,size_t input_increment,size_t output_increment,const union xnn_f32_output_params params[restrict static1])13 void xnn_f32_argmaxpool_ukernel_9x__sse2_c4(
14 size_t output_pixels,
15 size_t pooling_elements,
16 size_t channels,
17 const float** input,
18 size_t input_offset,
19 float* output,
20 uint32_t* index,
21 size_t input_increment,
22 size_t output_increment,
23 const union xnn_f32_output_params params[restrict static 1])
24 {
25 assert(output_pixels != 0);
26 assert(pooling_elements != 0);
27 assert(pooling_elements <= 9);
28 assert(channels != 0);
29
30 const __m128 voutput_max = _mm_load_ps(params->sse.max);
31 const __m128 voutput_min = _mm_load_ps(params->sse.min);
32 do {
33 float* o = output;
34 uint32_t* i = index;
35
36 const float* i0 = input[0];
37 const float* i1 = input[1];
38 const float* i2 = input[2];
39 const float* i3 = input[3];
40 const float* i4 = input[4];
41 const float* i5 = input[5];
42 const float* i6 = input[6];
43 const float* i7 = input[7];
44 const float* i8 = input[8];
45 i0 = (const float*) ((uintptr_t) i0 + input_offset);
46 i1 = (const float*) ((uintptr_t) i1 + input_offset);
47 i2 = (const float*) ((uintptr_t) i2 + input_offset);
48 i3 = (const float*) ((uintptr_t) i3 + input_offset);
49 i4 = (const float*) ((uintptr_t) i4 + input_offset);
50 i5 = (const float*) ((uintptr_t) i5 + input_offset);
51 i6 = (const float*) ((uintptr_t) i6 + input_offset);
52 i7 = (const float*) ((uintptr_t) i7 + input_offset);
53 i8 = (const float*) ((uintptr_t) i8 + input_offset);
54 if (pooling_elements < 2) {
55 i1 = i0;
56 }
57 if (pooling_elements <= 2) {
58 i2 = i0;
59 }
60 if (pooling_elements < 4) {
61 i3 = i0;
62 }
63 if (pooling_elements <= 4) {
64 i4 = i0;
65 }
66 if (pooling_elements < 6) {
67 i5 = i0;
68 }
69 if (pooling_elements <= 6) {
70 i6 = i0;
71 }
72 if (pooling_elements < 8) {
73 i7 = i0;
74 }
75 if (pooling_elements <= 8) {
76 i8 = i0;
77 }
78
79 size_t c = channels;
80 for (; c >= 4; c -= 4) {
81 const __m128 vi0 = _mm_loadu_ps(i0);
82 i0 += 4;
83 const __m128 vi1 = _mm_loadu_ps(i1);
84 i1 += 4;
85 const __m128 vi2 = _mm_loadu_ps(i2);
86 i2 += 4;
87 const __m128 vi3 = _mm_loadu_ps(i3);
88 i3 += 4;
89 const __m128 vi4 = _mm_loadu_ps(i4);
90 i4 += 4;
91 const __m128 vi5 = _mm_loadu_ps(i5);
92 i5 += 4;
93 const __m128 vi6 = _mm_loadu_ps(i6);
94 i6 += 4;
95 const __m128 vi7 = _mm_loadu_ps(i7);
96 i7 += 4;
97 const __m128 vi8 = _mm_loadu_ps(i8);
98 i8 += 4;
99
100 __m128 vmax = vi0;
101 __m128i vidx = _mm_setzero_si128();
102
103 const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax));
104 vmax = _mm_max_ps(vi1, vmax);
105 vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, _mm_set1_epi32(1)));
106
107 const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax));
108 vmax = _mm_max_ps(vi2, vmax);
109 vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, _mm_set1_epi32(2)));
110
111 const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax));
112 vmax = _mm_max_ps(vi3, vmax);
113 vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, _mm_set1_epi32(3)));
114
115 const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax));
116 vmax = _mm_max_ps(vi4, vmax);
117 vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, _mm_set1_epi32(4)));
118
119 const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax));
120 vmax = _mm_max_ps(vi5, vmax);
121 vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, _mm_set1_epi32(5)));
122
123 const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax));
124 vmax = _mm_max_ps(vi6, vmax);
125 vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, _mm_set1_epi32(6)));
126
127 const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax));
128 vmax = _mm_max_ps(vi7, vmax);
129 vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, _mm_set1_epi32(7)));
130
131 const __m128i vm8 = _mm_castps_si128(_mm_cmpgt_ps(vi8, vmax));
132 vmax = _mm_max_ps(vi8, vmax);
133 vidx = _mm_or_si128(_mm_andnot_si128(vm8, vidx), _mm_and_si128(vm8, _mm_set1_epi32(8)));
134
135 const __m128 vout = _mm_max_ps(_mm_min_ps(vmax, voutput_max), voutput_min);
136
137 _mm_storeu_ps(o, vout);
138 o += 4;
139 _mm_storeu_si128((__m128i*) i, vidx);
140 i += 4;
141 }
142 if (c != 0) {
143 const __m128 vi0 = _mm_loadu_ps(i0);
144 const __m128 vi1 = _mm_loadu_ps(i1);
145 const __m128 vi2 = _mm_loadu_ps(i2);
146 const __m128 vi3 = _mm_loadu_ps(i3);
147 const __m128 vi4 = _mm_loadu_ps(i4);
148 const __m128 vi5 = _mm_loadu_ps(i5);
149 const __m128 vi6 = _mm_loadu_ps(i6);
150 const __m128 vi7 = _mm_loadu_ps(i7);
151 const __m128 vi8 = _mm_loadu_ps(i8);
152
153 __m128 vmax = vi0;
154 __m128i vidx = _mm_setzero_si128();
155
156 const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax));
157 vmax = _mm_max_ps(vi1, vmax);
158 vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, _mm_set1_epi32(1)));
159
160 const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax));
161 vmax = _mm_max_ps(vi2, vmax);
162 vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, _mm_set1_epi32(2)));
163
164 const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax));
165 vmax = _mm_max_ps(vi3, vmax);
166 vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, _mm_set1_epi32(3)));
167
168 const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax));
169 vmax = _mm_max_ps(vi4, vmax);
170 vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, _mm_set1_epi32(4)));
171
172 const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax));
173 vmax = _mm_max_ps(vi5, vmax);
174 vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, _mm_set1_epi32(5)));
175
176 const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax));
177 vmax = _mm_max_ps(vi6, vmax);
178 vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, _mm_set1_epi32(6)));
179
180 const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax));
181 vmax = _mm_max_ps(vi7, vmax);
182 vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, _mm_set1_epi32(7)));
183
184 const __m128i vm8 = _mm_castps_si128(_mm_cmpgt_ps(vi8, vmax));
185 vmax = _mm_max_ps(vi8, vmax);
186 vidx = _mm_or_si128(_mm_andnot_si128(vm8, vidx), _mm_and_si128(vm8, _mm_set1_epi32(8)));
187
188 __m128 vout = _mm_max_ps(_mm_min_ps(vmax, voutput_max), voutput_min);
189
190 if (c & 2) {
191 _mm_store_sd((double*) o, _mm_castps_pd(vout));
192 _mm_storel_epi64((__m128i*) i, vidx);
193 vout = _mm_movehl_ps(vout, vout);
194 vidx = _mm_unpackhi_epi64(vidx, vidx);
195 o += 2;
196 i += 2;
197 }
198 if (c & 1) {
199 _mm_store_ss(o, vout);
200 *i = (uint32_t) _mm_cvtsi128_si32(vidx);
201 o += 1;
202 i += 1;
203 }
204 }
205 input = (const float**) ((uintptr_t) input + input_increment);
206 output = (float*) ((uintptr_t) o + output_increment);
207 index = (uint32_t*) i;
208 } while (--output_pixels != 0);
209 }
210