1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv2d-chw/3x3p1-sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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 #include <xnnpack/math.h>
16
17
xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4(size_t input_height,size_t input_width,const float * input,const float * weights,const float * zero,float * output,uint32_t padding_top,const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4(
19 size_t input_height,
20 size_t input_width,
21 const float* input,
22 const float* weights,
23 const float* zero,
24 float* output,
25 uint32_t padding_top,
26 const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 assert(input_height != 0);
29 assert(input_width != 0);
30 assert(input_width % sizeof(float) == 0);
31 assert(padding_top == 1);
32
33 const __m128 vmask = _mm_load_ps((const float*) params->sse.mask);
34 const __m128 vmax = _mm_load_ps(params->sse.max);
35 const __m128 vmin = _mm_load_ps(params->sse.min);
36
37 const __m128 vbias = _mm_load1_ps(weights);
38 const __m128 vk00 = _mm_load1_ps(weights + 1);
39 const __m128 vk01 = _mm_load1_ps(weights + 2);
40 const __m128 vk02 = _mm_load1_ps(weights + 3);
41 const __m128 vk10 = _mm_load1_ps(weights + 4);
42 const __m128 vk11 = _mm_load1_ps(weights + 5);
43 const __m128 vk12 = _mm_load1_ps(weights + 6);
44 const __m128 vk20 = _mm_load1_ps(weights + 7);
45 const __m128 vk21 = _mm_load1_ps(weights + 8);
46 const __m128 vk22 = _mm_load1_ps(weights + 9);
47
48 const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
49
50 const float* i0 = zero;
51 const float* i1 = input;
52 const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
53 const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
54 const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
55
56 float* o0 = output;
57 float* o1 = (float*) ((uintptr_t) o0 + input_width);
58 float* o2 = (float*) ((uintptr_t) o1 + input_width);
59
60 size_t output_height = input_height;
61 do {
62 if XNN_UNPREDICTABLE(output_height < 2) {
63 i2 = zero;
64 o1 = o0;
65 }
66 if XNN_UNPREDICTABLE(output_height < 3) {
67 i3 = zero;
68 o2 = o1;
69 }
70 if XNN_UNPREDICTABLE(output_height < 4) {
71 i4 = zero;
72 }
73
74 // vi0x3012 = ( vi02, vi01, vi{M}0, vi{M}3 )
75 __m128 vi0x3012 = _mm_setzero_ps();
76 // vi1x3012 = ( vi12, vi11, vi{M}0, vi{M}3 )
77 __m128 vi1x3012 = _mm_setzero_ps();
78 // vi2x3012 = ( vi22, vi21, vi{M}0, vi{M}3 )
79 __m128 vi2x3012 = _mm_setzero_ps();
80 // vi3x3012 = ( vi32, vi31, vi{M}0, vi{M}3 )
81 __m128 vi3x3012 = _mm_setzero_ps();
82 // vi4x3012 = ( vi42, vi41, vi{M}0, vi{M}3 )
83 __m128 vi4x3012 = _mm_setzero_ps();
84
85 __m128 vi0x4567 = _mm_loadu_ps(i0);
86 i0 += 4;
87 __m128 vi1x4567 = _mm_loadu_ps(i1);
88 i1 += 4;
89 __m128 vi2x4567 = _mm_loadu_ps(i2);
90 i2 += 4;
91 __m128 vi3x4567 = _mm_loadu_ps(i3);
92 i3 += 4;
93 __m128 vi4x4567 = _mm_loadu_ps(i4);
94 i4 += 4;
95
96 size_t w = input_width;
97 for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
98 // vi0x89AB = ( vi0B, vi0A, vi09, vi08 )
99 const __m128 vi0x89AB = _mm_loadu_ps(i0);
100 i0 += 4;
101 // vi1x89AB = ( vi1B, vi1A, vi19, vi18 )
102 const __m128 vi1x89AB = _mm_loadu_ps(i1);
103 i1 += 4;
104 // vi2x89AB = ( vi2B, vi2A, vi29, vi28 )
105 const __m128 vi2x89AB = _mm_loadu_ps(i2);
106 i2 += 4;
107 // vi3x89AB = ( vi3B, vi3A, vi39, vi38 )
108 const __m128 vi3x89AB = _mm_loadu_ps(i3);
109 i3 += 4;
110 // vi4x89AB = ( vi4B, vi4A, vi49, vi48 )
111 const __m128 vi4x89AB = _mm_loadu_ps(i4);
112 i4 += 4;
113
114 // vi0x7456 = ( vi06, vi05, vi04, vi07 )
115 const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
116 // vi1x7456 = ( vi16, vi15, vi14, vi17 )
117 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
118 // vi2x7456 = ( vi26, vi25, vi24, vi27 )
119 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
120 // vi3x7456 = ( vi36, vi35, vi34, vi37 )
121 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3));
122 // vi4x7456 = ( vi46, vi45, vi44, vi47 )
123 const __m128 vi4x7456 = _mm_shuffle_ps(vi4x4567, vi4x4567, _MM_SHUFFLE(2, 1, 0, 3));
124
125 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
126 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
127 __m128 vo2p0 = _mm_add_ps(vbias, _mm_mul_ps(vi2x4567, vk01));
128 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk11));
129 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk11));
130 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk11));
131 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
132 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));
133 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x4567, vk21));
134
135 // vi0x3456 = ( vi06, vi05, vi04, vi03 )
136 const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
137 // vi1x3456 = ( vi16, vi15, vi14, vi13 )
138 const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
139 // vi2x3456 = ( vi26, vi25, vi24, vi23 )
140 const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
141 // vi3x3456 = ( vi36, vi35, vi34, vi33 )
142 const __m128 vi3x3456 = _mm_move_ss(vi3x7456, vi3x3012);
143 // vi4x3456 = ( vi46, vi45, vi44, vi43 )
144 const __m128 vi4x3456 = _mm_move_ss(vi4x7456, vi4x3012);
145
146 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk00));
147 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk00));
148 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk00));
149 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
150 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
151 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x3456, vk10));
152 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20));
153 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x3456, vk20));
154 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x3456, vk20));
155
156 vi0x3012 = vi0x7456;
157 vi1x3012 = vi1x7456;
158 vi2x3012 = vi2x7456;
159 vi3x3012 = vi3x7456;
160 vi4x3012 = vi4x7456;
161
162 // vi0x8567 = ( vi07, vi06, vi05, vi08 )
163 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB);
164 // vi1x8567 = ( vi17, vi16, vi15, vi18 )
165 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB);
166 // vi2x8567 = ( vi27, vi26, vi25, vi28 )
167 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB);
168 // vi3x8567 = ( vi37, vi36, vi35, vi38 )
169 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB);
170 // vi4x8567 = ( vi47, vi46, vi45, vi48 )
171 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB);
172
173 // vi0x5678 = ( vi08, vi07, vi06, vi05 )
174 const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
175 // vi1x5678 = ( vi18, vi17, vi16, vi15 )
176 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
177 // vi2x5678 = ( vi28, vi27, vi26, vi25 )
178 const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
179 // vi3x5678 = ( vi38, vi37, vi36, vi35 )
180 const __m128 vi3x5678 = _mm_shuffle_ps(vi3x8567, vi3x8567, _MM_SHUFFLE(0, 3, 2, 1));
181 // vi4x5678 = ( vi48, vi47, vi46, vi45 )
182 const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));
183
184 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
185 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
186 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x5678, vk02));
187 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk12));
188 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x5678, vk12));
189 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x5678, vk12));
190 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
191 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));
192 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk22));
193
194 vi0x4567 = vi0x89AB;
195 vi1x4567 = vi1x89AB;
196 vi2x4567 = vi2x89AB;
197 vi3x4567 = vi3x89AB;
198 vi4x4567 = vi4x89AB;
199
200
201 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
202 __m128 vo1 = _mm_max_ps(vo1p0, vmin);
203 __m128 vo2 = _mm_max_ps(vo2p0, vmin);
204
205 vo0 = _mm_min_ps(vo0, vmax);
206 vo1 = _mm_min_ps(vo1, vmax);
207 vo2 = _mm_min_ps(vo2, vmax);
208
209 _mm_storeu_ps(o2, vo2);
210 o2 += 4;
211 _mm_storeu_ps(o1, vo1);
212 o1 += 4;
213 _mm_storeu_ps(o0, vo0);
214 o0 += 4;
215 }
216 // Always process the last block of 1..4 pixels.
217 assert(w >= 1 * sizeof(float));
218 assert(w <= 4 * sizeof(float));
219 {
220 vi0x4567 = _mm_and_ps(vmask, vi0x4567);
221 vi1x4567 = _mm_and_ps(vmask, vi1x4567);
222 vi2x4567 = _mm_and_ps(vmask, vi2x4567);
223 vi3x4567 = _mm_and_ps(vmask, vi3x4567);
224 vi4x4567 = _mm_and_ps(vmask, vi4x4567);
225
226 // vi0x7456 = ( vi06, vi05, vi04, vi07 )
227 const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
228 // vi1x7456 = ( vi16, vi15, vi14, vi17 )
229 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
230 // vi2x7456 = ( vi26, vi25, vi24, vi27 )
231 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
232 // vi3x7456 = ( vi36, vi35, vi34, vi37 )
233 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3));
234 // vi4x7456 = ( vi46, vi45, vi44, vi47 )
235 const __m128 vi4x7456 = _mm_shuffle_ps(vi4x4567, vi4x4567, _MM_SHUFFLE(2, 1, 0, 3));
236
237 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
238 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
239 __m128 vo2p0 = _mm_add_ps(vbias, _mm_mul_ps(vi2x4567, vk01));
240 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk11));
241 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk11));
242 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk11));
243 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
244 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));
245 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x4567, vk21));
246
247 // vi0x3456 = ( vi06, vi05, vi04, vi03 )
248 const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
249 // vi1x3456 = ( vi16, vi15, vi14, vi13 )
250 const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
251 // vi2x3456 = ( vi26, vi25, vi24, vi23 )
252 const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
253 // vi3x3456 = ( vi36, vi35, vi34, vi33 )
254 const __m128 vi3x3456 = _mm_move_ss(vi3x7456, vi3x3012);
255 // vi4x3456 = ( vi46, vi45, vi44, vi43 )
256 const __m128 vi4x3456 = _mm_move_ss(vi4x7456, vi4x3012);
257
258 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk00));
259 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk00));
260 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk00));
261 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
262 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
263 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x3456, vk10));
264 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20));
265 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x3456, vk20));
266 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x3456, vk20));
267
268 const __m128 vzero = _mm_setzero_ps();
269 // vi0x8567 = ( vi07, vi06, vi05, 0.0 )
270 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero);
271 // vi1x8567 = ( vi17, vi16, vi15, 0.0 )
272 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero);
273 // vi2x8567 = ( vi27, vi26, vi25, 0.0 )
274 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero);
275 // vi3x8567 = ( vi37, vi36, vi35, 0.0 )
276 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero);
277 // vi4x8567 = ( vi47, vi46, vi45, 0.0 )
278 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero);
279
280 // vi0x5678 = ( vi08, vi07, vi06, vi05 )
281 const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
282 // vi1x5678 = ( vi18, vi17, vi16, vi15 )
283 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
284 // vi2x5678 = ( vi28, vi27, vi26, vi25 )
285 const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
286 // vi3x5678 = ( vi38, vi37, vi36, vi35 )
287 const __m128 vi3x5678 = _mm_shuffle_ps(vi3x8567, vi3x8567, _MM_SHUFFLE(0, 3, 2, 1));
288 // vi4x5678 = ( vi48, vi47, vi46, vi45 )
289 const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));
290
291 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
292 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
293 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x5678, vk02));
294 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk12));
295 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x5678, vk12));
296 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x5678, vk12));
297 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
298 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));
299 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk22));
300
301
302 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
303 __m128 vo1 = _mm_max_ps(vo1p0, vmin);
304 __m128 vo2 = _mm_max_ps(vo2p0, vmin);
305
306 vo0 = _mm_min_ps(vo0, vmax);
307 vo1 = _mm_min_ps(vo1, vmax);
308 vo2 = _mm_min_ps(vo2, vmax);
309
310 if XNN_LIKELY(w == 4 * sizeof(float)) {
311 _mm_storeu_ps(o2, vo2);
312 o2 += 4;
313 _mm_storeu_ps(o1, vo1);
314 o1 += 4;
315 _mm_storeu_ps(o0, vo0);
316 o0 += 4;
317 } else {
318 if (w & (2 * sizeof(float))) {
319 _mm_storel_pi((__m64*) o2, vo2);
320 o2 += 2;
321 _mm_storel_pi((__m64*) o1, vo1);
322 o1 += 2;
323 _mm_storel_pi((__m64*) o0, vo0);
324 o0 += 2;
325
326 vo0 = _mm_movehl_ps(vo0, vo0);
327 vo1 = _mm_movehl_ps(vo1, vo1);
328 vo2 = _mm_movehl_ps(vo2, vo2);
329 }
330 if (w & (1 * sizeof(float))) {
331 _mm_store_ss(o2, vo2);
332 o2 += 1;
333 _mm_store_ss(o1, vo1);
334 o1 += 1;
335 _mm_store_ss(o0, vo0);
336 o0 += 1;
337 }
338 }
339 }
340
341 i0 = (const float*) ((uintptr_t) i3 - input_decrement);
342 i1 = (const float*) ((uintptr_t) i4 - input_decrement);
343 i2 = (const float*) ((uintptr_t) i1 + input_width);
344 i3 = (const float*) ((uintptr_t) i2 + input_width);
345 i4 = (const float*) ((uintptr_t) i3 + input_width);
346
347 o0 = o2;
348 o1 = (float*) ((uintptr_t) o0 + input_width);
349 o2 = (float*) ((uintptr_t) o1 + input_width);
350
351 output_height = doz(output_height, 3);
352 } while (output_height != 0);
353 }
354