1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv/up-sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <xmmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15
16
xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,size_t input_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2(
18 size_t channels,
19 size_t output_width,
20 const float** input,
21 const float* weights,
22 float* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const float* zero,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const __m128 vmax = _mm_load_ps(params->sse.max);
33 const __m128 vmin = _mm_load_ps(params->sse.min);
34 do {
35 const float* i0 = input[0];
36 assert(i0 != NULL);
37 if XNN_UNPREDICTABLE(i0 != zero) {
38 i0 = (const float*) ((uintptr_t) i0 + input_offset);
39 }
40 const float* i1 = input[1];
41 assert(i1 != NULL);
42 if XNN_UNPREDICTABLE(i1 != zero) {
43 i1 = (const float*) ((uintptr_t) i1 + input_offset);
44 }
45 const float* i2 = input[2];
46 assert(i2 != NULL);
47 if XNN_UNPREDICTABLE(i2 != zero) {
48 i2 = (const float*) ((uintptr_t) i2 + input_offset);
49 }
50 const float* i3 = input[3];
51 assert(i3 != NULL);
52 if XNN_UNPREDICTABLE(i3 != zero) {
53 i3 = (const float*) ((uintptr_t) i3 + input_offset);
54 }
55 const float* i4 = input[4];
56 assert(i4 != NULL);
57 if XNN_UNPREDICTABLE(i4 != zero) {
58 i4 = (const float*) ((uintptr_t) i4 + input_offset);
59 }
60 const float* i5 = input[5];
61 assert(i5 != NULL);
62 if XNN_UNPREDICTABLE(i5 != zero) {
63 i5 = (const float*) ((uintptr_t) i5 + input_offset);
64 }
65 const float* i6 = input[6];
66 assert(i6 != NULL);
67 if XNN_UNPREDICTABLE(i6 != zero) {
68 i6 = (const float*) ((uintptr_t) i6 + input_offset);
69 }
70 const float* i7 = input[7];
71 assert(i7 != NULL);
72 if XNN_UNPREDICTABLE(i7 != zero) {
73 i7 = (const float*) ((uintptr_t) i7 + input_offset);
74 }
75 const float* i8 = input[8];
76 assert(i8 != NULL);
77 if XNN_UNPREDICTABLE(i8 != zero) {
78 i8 = (const float*) ((uintptr_t) i8 + input_offset);
79 }
80 input = (const float**) ((uintptr_t) input + input_stride);
81
82 size_t c = channels;
83 const float* w = weights;
84 for (; c >= 8; c -= 8) {
85 __m128 vacc0123p0 = _mm_load_ps(w);
86 __m128 vacc4567p0 = _mm_load_ps(w + 4);
87
88
89 const __m128 vi0x0123 = _mm_loadu_ps(i0);
90 const __m128 vi0x4567 = _mm_loadu_ps(i0 + 4);
91 i0 += 8;
92
93 const __m128 vk0x0123 = _mm_load_ps(w + 8);
94 const __m128 vk0x4567 = _mm_load_ps(w + 12);
95 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
96 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi0x4567, vk0x4567));
97
98 const __m128 vi1x0123 = _mm_loadu_ps(i1);
99 const __m128 vi1x4567 = _mm_loadu_ps(i1 + 4);
100 i1 += 8;
101
102 const __m128 vk1x0123 = _mm_load_ps(w + 16);
103 const __m128 vk1x4567 = _mm_load_ps(w + 20);
104 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
105 __m128 vacc4567p1 = _mm_mul_ps(vi1x4567, vk1x4567);
106
107 const __m128 vi2x0123 = _mm_loadu_ps(i2);
108 const __m128 vi2x4567 = _mm_loadu_ps(i2 + 4);
109 i2 += 8;
110
111 const __m128 vk2x0123 = _mm_load_ps(w + 24);
112 const __m128 vk2x4567 = _mm_load_ps(w + 28);
113 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
114 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567));
115
116 const __m128 vi3x0123 = _mm_loadu_ps(i3);
117 const __m128 vi3x4567 = _mm_loadu_ps(i3 + 4);
118 i3 += 8;
119
120 const __m128 vk3x0123 = _mm_load_ps(w + 32);
121 const __m128 vk3x4567 = _mm_load_ps(w + 36);
122 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
123 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi3x4567, vk3x4567));
124
125 const __m128 vi4x0123 = _mm_loadu_ps(i4);
126 const __m128 vi4x4567 = _mm_loadu_ps(i4 + 4);
127 i4 += 8;
128
129 const __m128 vk4x0123 = _mm_load_ps(w + 40);
130 const __m128 vk4x4567 = _mm_load_ps(w + 44);
131 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123));
132 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi4x4567, vk4x4567));
133
134 const __m128 vi5x0123 = _mm_loadu_ps(i5);
135 const __m128 vi5x4567 = _mm_loadu_ps(i5 + 4);
136 i5 += 8;
137
138 const __m128 vk5x0123 = _mm_load_ps(w + 48);
139 const __m128 vk5x4567 = _mm_load_ps(w + 52);
140 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123));
141 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi5x4567, vk5x4567));
142
143 const __m128 vi6x0123 = _mm_loadu_ps(i6);
144 const __m128 vi6x4567 = _mm_loadu_ps(i6 + 4);
145 i6 += 8;
146
147 const __m128 vk6x0123 = _mm_load_ps(w + 56);
148 const __m128 vk6x4567 = _mm_load_ps(w + 60);
149 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123));
150 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi6x4567, vk6x4567));
151
152 const __m128 vi7x0123 = _mm_loadu_ps(i7);
153 const __m128 vi7x4567 = _mm_loadu_ps(i7 + 4);
154 i7 += 8;
155
156 const __m128 vk7x0123 = _mm_load_ps(w + 64);
157 const __m128 vk7x4567 = _mm_load_ps(w + 68);
158 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123));
159 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi7x4567, vk7x4567));
160
161 const __m128 vi8x0123 = _mm_loadu_ps(i8);
162 const __m128 vi8x4567 = _mm_loadu_ps(i8 + 4);
163 i8 += 8;
164
165 const __m128 vk8x0123 = _mm_load_ps(w + 72);
166 const __m128 vk8x4567 = _mm_load_ps(w + 76);
167 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123));
168 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi8x4567, vk8x4567));
169
170 w += 80;
171
172 // Add up all accumulators to vacc01234567p0
173 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
174 vacc4567p0 = _mm_add_ps(vacc4567p0, vacc4567p1);
175
176 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
177 __m128 vacc4567 = _mm_max_ps(vacc4567p0, vmin);
178 vacc0123 = _mm_min_ps(vacc0123, vmax);
179 vacc4567 = _mm_min_ps(vacc4567, vmax);
180
181 _mm_storeu_ps(output, vacc0123);
182 _mm_storeu_ps(output + 4, vacc4567);
183 output += 8;
184 }
185 for (; c >= 4; c -= 4) {
186 __m128 vacc0123p0 = _mm_load_ps(w);
187
188 const __m128 vi0x0123 = _mm_loadu_ps(i0);
189 i0 += 4;
190
191 const __m128 vk0x0123 = _mm_load_ps(w + 8);
192 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
193
194 const __m128 vi1x0123 = _mm_loadu_ps(i1);
195 i1 += 4;
196
197 const __m128 vk1x0123 = _mm_load_ps(w + 16);
198 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
199
200 const __m128 vi2x0123 = _mm_loadu_ps(i2);
201 i2 += 4;
202
203 const __m128 vk2x0123 = _mm_load_ps(w + 24);
204 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
205
206 const __m128 vi3x0123 = _mm_loadu_ps(i3);
207 i3 += 4;
208
209 const __m128 vk3x0123 = _mm_load_ps(w + 32);
210 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
211
212 const __m128 vi4x0123 = _mm_loadu_ps(i4);
213 i4 += 4;
214
215 const __m128 vk4x0123 = _mm_load_ps(w + 40);
216 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123));
217
218 const __m128 vi5x0123 = _mm_loadu_ps(i5);
219 i5 += 4;
220
221 const __m128 vk5x0123 = _mm_load_ps(w + 48);
222 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123));
223
224 const __m128 vi6x0123 = _mm_loadu_ps(i6);
225 i6 += 4;
226
227 const __m128 vk6x0123 = _mm_load_ps(w + 56);
228 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123));
229
230 const __m128 vi7x0123 = _mm_loadu_ps(i7);
231 i7 += 4;
232
233 const __m128 vk7x0123 = _mm_load_ps(w + 64);
234 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123));
235
236 const __m128 vi8x0123 = _mm_loadu_ps(i8);
237 i8 += 4;
238
239 const __m128 vk8x0123 = _mm_load_ps(w + 72);
240 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123));
241
242 w += 4;
243
244 // Add up all accumulators to vacc01234567p0
245 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
246
247 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
248 vacc0123 = _mm_min_ps(vacc0123, vmax);
249
250 _mm_storeu_ps(output, vacc0123);
251 output += 4;
252 }
253 if XNN_UNLIKELY(c != 0) {
254 __m128 vacc0123p0 = _mm_load_ps(w);
255
256 const __m128 vi0x0123 = _mm_loadu_ps(i0);
257 const __m128 vk0x0123 = _mm_load_ps(w + 8);
258 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
259
260 const __m128 vi1x0123 = _mm_loadu_ps(i1);
261 const __m128 vk1x0123 = _mm_load_ps(w + 16);
262 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
263
264 const __m128 vi2x0123 = _mm_loadu_ps(i2);
265 const __m128 vk2x0123 = _mm_load_ps(w + 24);
266 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
267
268 const __m128 vi3x0123 = _mm_loadu_ps(i3);
269 const __m128 vk3x0123 = _mm_load_ps(w + 32);
270 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
271
272 const __m128 vi4x0123 = _mm_loadu_ps(i4);
273 const __m128 vk4x0123 = _mm_load_ps(w + 40);
274 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123));
275
276 const __m128 vi5x0123 = _mm_loadu_ps(i5);
277 const __m128 vk5x0123 = _mm_load_ps(w + 48);
278 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123));
279
280 const __m128 vi6x0123 = _mm_loadu_ps(i6);
281 const __m128 vk6x0123 = _mm_load_ps(w + 56);
282 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123));
283
284 const __m128 vi7x0123 = _mm_loadu_ps(i7);
285 const __m128 vk7x0123 = _mm_load_ps(w + 64);
286 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123));
287
288 const __m128 vi8x0123 = _mm_loadu_ps(i8);
289 const __m128 vk8x0123 = _mm_load_ps(w + 72);
290 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123));
291
292 // Add up all accumulators to vacc01234567p0
293 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
294
295 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
296 vacc0123 = _mm_min_ps(vacc0123, vmax);
297
298 if (c & 2) {
299 _mm_storel_pi((__m64*) output, vacc0123);
300 vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
301 output += 2;
302 }
303 if (c & 1) {
304 _mm_store_ss(output, vacc0123);
305 output += 1;
306 }
307 }
308
309 output = (float*) ((uintptr_t) output + output_increment);
310 } while (--output_width != 0);
311 }
312