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/gavgpool.h>
11 #include <xnnpack/math.h>
12
13
xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4(size_t rows,size_t channels,const float * input,size_t input_stride,const float * zero,float * buffer,float * output,const union xnn_f32_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])14 void xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4(
15 size_t rows,
16 size_t channels,
17 const float* input,
18 size_t input_stride,
19 const float* zero,
20 float* buffer,
21 float* output,
22 const union xnn_f32_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(rows > 7);
25 assert(channels != 0);
26
27 const float* i0 = input;
28 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
29 const float* i2 = (const float*) ((uintptr_t) i1 + input_stride);
30 const float* i3 = (const float*) ((uintptr_t) i2 + input_stride);
31 const float* i4 = (const float*) ((uintptr_t) i3 + input_stride);
32 const float* i5 = (const float*) ((uintptr_t) i4 + input_stride);
33 const float* i6 = (const float*) ((uintptr_t) i5 + input_stride);
34 const size_t packed_channels = round_up_po2(channels, 4);
35 const size_t input_increment = 7 * input_stride - packed_channels * sizeof(float);
36
37 float* b = buffer;
38 for (size_t c = 0; c < channels; c += 4) {
39 const __m128 vi0 = _mm_loadu_ps(i0);
40 i0 += 4;
41 const __m128 vi1 = _mm_loadu_ps(i1);
42 i1 += 4;
43 const __m128 vi2 = _mm_loadu_ps(i2);
44 i2 += 4;
45 const __m128 vi3 = _mm_loadu_ps(i3);
46 i3 += 4;
47 const __m128 vi4 = _mm_loadu_ps(i4);
48 i4 += 4;
49 const __m128 vi5 = _mm_loadu_ps(i5);
50 i5 += 4;
51 const __m128 vi6 = _mm_loadu_ps(i6);
52 i6 += 4;
53
54 const __m128 vsum01 = _mm_add_ps(vi0, vi1);
55 const __m128 vsum23 = _mm_add_ps(vi2, vi3);
56 const __m128 vsum45 = _mm_add_ps(vi4, vi5);
57
58 const __m128 vsum016 = _mm_add_ps(vsum01, vi6);
59 const __m128 vsum2345 = _mm_add_ps(vsum23, vsum45);
60
61 const __m128 vsum = _mm_add_ps(vsum016, vsum2345);
62
63 _mm_store_ps(b, vsum); b += 4;
64 }
65 for (rows -= 7; rows > 7; rows -= 7) {
66 b = buffer;
67
68 i0 = (const float*) ((uintptr_t) i0 + input_increment);
69 i1 = (const float*) ((uintptr_t) i1 + input_increment);
70 i2 = (const float*) ((uintptr_t) i2 + input_increment);
71 i3 = (const float*) ((uintptr_t) i3 + input_increment);
72 i4 = (const float*) ((uintptr_t) i4 + input_increment);
73 i5 = (const float*) ((uintptr_t) i5 + input_increment);
74 i6 = (const float*) ((uintptr_t) i6 + input_increment);
75
76 for (size_t c = 0; c < channels; c += 4) {
77 const __m128 vi0 = _mm_loadu_ps(i0);
78 i0 += 4;
79 const __m128 vi1 = _mm_loadu_ps(i1);
80 i1 += 4;
81 const __m128 vi2 = _mm_loadu_ps(i2);
82 i2 += 4;
83 const __m128 vi3 = _mm_loadu_ps(i3);
84 i3 += 4;
85 const __m128 vi4 = _mm_loadu_ps(i4);
86 i4 += 4;
87 const __m128 vi5 = _mm_loadu_ps(i5);
88 i5 += 4;
89 const __m128 vi6 = _mm_loadu_ps(i6);
90 i6 += 4;
91 const __m128 vacc = _mm_load_ps(b);
92
93 const __m128 vsum01 = _mm_add_ps(vi0, vi1);
94 const __m128 vsum23 = _mm_add_ps(vi2, vi3);
95 const __m128 vsum45 = _mm_add_ps(vi4, vi5);
96 const __m128 vsum6a = _mm_add_ps(vi6, vacc);
97
98 const __m128 vsum0123 = _mm_add_ps(vsum01, vsum23);
99 const __m128 vsum456a = _mm_add_ps(vsum45, vsum6a);
100
101 const __m128 vsum = _mm_add_ps(vsum0123, vsum456a);
102
103 _mm_store_ps(b, vsum); b += 4;
104 }
105 }
106
107 i0 = (const float*) ((uintptr_t) i0 + input_increment);
108 i1 = (const float*) ((uintptr_t) i1 + input_increment);
109 if (rows < 2) {
110 i1 = zero;
111 }
112 i2 = (const float*) ((uintptr_t) i2 + input_increment);
113 if (rows <= 2) {
114 i2 = zero;
115 }
116 i3 = (const float*) ((uintptr_t) i3 + input_increment);
117 if (rows < 4) {
118 i3 = zero;
119 }
120 i4 = (const float*) ((uintptr_t) i4 + input_increment);
121 if (rows <= 4) {
122 i4 = zero;
123 }
124 i5 = (const float*) ((uintptr_t) i5 + input_increment);
125 if (rows < 6) {
126 i5 = zero;
127 }
128 i6 = (const float*) ((uintptr_t) i6 + input_increment);
129 if (rows <= 6) {
130 i6 = zero;
131 }
132 const __m128 vscale = _mm_load_ps(params->sse.scale);
133 const __m128 vmin = _mm_load_ps(params->sse.min);
134 const __m128 vmax = _mm_load_ps(params->sse.max);
135
136 b = buffer;
137 while (channels >= 4) {
138 const __m128 vi0 = _mm_loadu_ps(i0);
139 i0 += 4;
140 const __m128 vi1 = _mm_loadu_ps(i1);
141 i1 += 4;
142 const __m128 vi2 = _mm_loadu_ps(i2);
143 i2 += 4;
144 const __m128 vi3 = _mm_loadu_ps(i3);
145 i3 += 4;
146 const __m128 vi4 = _mm_loadu_ps(i4);
147 i4 += 4;
148 const __m128 vi5 = _mm_loadu_ps(i5);
149 i5 += 4;
150 const __m128 vi6 = _mm_loadu_ps(i6);
151 i6 += 4;
152 const __m128 vacc = _mm_load_ps(b);
153 b += 4;
154
155 const __m128 vsum01 = _mm_add_ps(vi0, vi1);
156 const __m128 vsum23 = _mm_add_ps(vi2, vi3);
157 const __m128 vsum45 = _mm_add_ps(vi4, vi5);
158 const __m128 vsum6a = _mm_add_ps(vi6, vacc);
159
160 const __m128 vsum0123 = _mm_add_ps(vsum01, vsum23);
161 const __m128 vsum456a = _mm_add_ps(vsum45, vsum6a);
162
163 const __m128 vsum = _mm_add_ps(vsum0123, vsum456a);
164
165 __m128 vout = _mm_mul_ps(vsum, vscale);
166 vout = _mm_max_ps(vout, vmin);
167 vout = _mm_min_ps(vout, vmax);
168
169 _mm_storeu_ps(output, vout);
170 output += 4;
171
172 channels -= 4;
173 }
174 if (channels != 0) {
175 const __m128 vi0 = _mm_loadu_ps(i0);
176 const __m128 vi1 = _mm_loadu_ps(i1);
177 const __m128 vi2 = _mm_loadu_ps(i2);
178 const __m128 vi3 = _mm_loadu_ps(i3);
179 const __m128 vi4 = _mm_loadu_ps(i4);
180 const __m128 vi5 = _mm_loadu_ps(i5);
181 const __m128 vi6 = _mm_loadu_ps(i6);
182 const __m128 vacc = _mm_loadu_ps(b);
183
184 const __m128 vsum01 = _mm_add_ps(vi0, vi1);
185 const __m128 vsum23 = _mm_add_ps(vi2, vi3);
186 const __m128 vsum45 = _mm_add_ps(vi4, vi5);
187 const __m128 vsum6a = _mm_add_ps(vi6, vacc);
188
189 const __m128 vsum0123 = _mm_add_ps(vsum01, vsum23);
190 const __m128 vsum456a = _mm_add_ps(vsum45, vsum6a);
191
192 const __m128 vsum = _mm_add_ps(vsum0123, vsum456a);
193
194 __m128 vout = _mm_mul_ps(vsum, vscale);
195 vout = _mm_max_ps(vout, vmin);
196 vout = _mm_min_ps(vout, vmax);
197
198 if (channels & 2) {
199 _mm_storel_pi((__m64*) output, vout);
200 vout = _mm_movehl_ps(vout, vout);
201 output += 2;
202 }
203 if (channels & 1) {
204 _mm_store_ss(output, vout);
205 }
206 }
207 }
208