• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv2d-chw/5x5p2-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_5x5p2__sse_2x4_acc3(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_5x5p2__sse_2x4_acc3(
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 == 2);
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 vk03 = _mm_load1_ps(weights + 4);
42   const __m128 vk04 = _mm_load1_ps(weights + 5);
43   const __m128 vk10 = _mm_load1_ps(weights + 6);
44   const __m128 vk11 = _mm_load1_ps(weights + 7);
45   const __m128 vk12 = _mm_load1_ps(weights + 8);
46   const __m128 vk13 = _mm_load1_ps(weights + 9);
47   const __m128 vk14 = _mm_load1_ps(weights + 10);
48   const __m128 vk20 = _mm_load1_ps(weights + 11);
49   const __m128 vk21 = _mm_load1_ps(weights + 12);
50   const __m128 vk22 = _mm_load1_ps(weights + 13);
51   const __m128 vk23 = _mm_load1_ps(weights + 14);
52   const __m128 vk24 = _mm_load1_ps(weights + 15);
53   const __m128 vk30 = _mm_load1_ps(weights + 16);
54   const __m128 vk31 = _mm_load1_ps(weights + 17);
55   const __m128 vk32 = _mm_load1_ps(weights + 18);
56   const __m128 vk33 = _mm_load1_ps(weights + 19);
57   const __m128 vk34 = _mm_load1_ps(weights + 20);
58   const __m128 vk40 = _mm_load1_ps(weights + 21);
59   const __m128 vk41 = _mm_load1_ps(weights + 22);
60   const __m128 vk42 = _mm_load1_ps(weights + 23);
61   const __m128 vk43 = _mm_load1_ps(weights + 24);
62   const __m128 vk44 = _mm_load1_ps(weights + 25);
63 
64   const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
65 
66   const float* i0 = zero;
67   const float* i1 = zero;
68   const float* i2 = input;
69   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
70   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
71   const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
72 
73   float* o0 = output;
74   float* o1 = (float*) ((uintptr_t) o0 + input_width);
75 
76   size_t output_height = input_height;
77   do {
78     if XNN_UNPREDICTABLE(output_height < 2) {
79       i3 = zero;
80       o1 = o0;
81     }
82     if XNN_UNPREDICTABLE(output_height < 3) {
83       i4 = zero;
84     }
85     if XNN_UNPREDICTABLE(output_height < 4) {
86       i5 = zero;
87     }
88 
89     __m128 vi0x3012 = _mm_setzero_ps();
90     __m128 vi1x3012 = _mm_setzero_ps();
91     __m128 vi2x3012 = _mm_setzero_ps();
92     __m128 vi3x3012 = _mm_setzero_ps();
93     __m128 vi4x3012 = _mm_setzero_ps();
94     __m128 vi5x3012 = _mm_setzero_ps();
95 
96     __m128 vi0x4567 = _mm_loadu_ps(i0);
97     i0 += 4;
98     __m128 vi1x4567 = _mm_loadu_ps(i1);
99     i1 += 4;
100     __m128 vi2x4567 = _mm_loadu_ps(i2);
101     i2 += 4;
102     __m128 vi3x4567 = _mm_loadu_ps(i3);
103     i3 += 4;
104     __m128 vi4x4567 = _mm_loadu_ps(i4);
105     i4 += 4;
106     __m128 vi5x4567 = _mm_loadu_ps(i5);
107     i5 += 4;
108 
109     size_t w = input_width;
110     for (; w > 8 * sizeof(float); w -= 4 * sizeof(float)) {
111       __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk02));
112       __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk02));
113       __m128 vo0p1 = _mm_mul_ps(vi1x4567, vk12);
114       __m128 vo1p1 = _mm_mul_ps(vi2x4567, vk12);
115       __m128 vo0p2 = _mm_mul_ps(vi2x4567, vk22);
116       __m128 vo1p2 = _mm_mul_ps(vi3x4567, vk22);
117       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x4567, vk32));
118       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x4567, vk32));
119       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x4567, vk42));
120       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5x4567, vk42));
121 
122       const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
123       const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
124       const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
125       const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3));
126       const __m128 vi4x7456 = _mm_shuffle_ps(vi4x4567, vi4x4567, _MM_SHUFFLE(2, 1, 0, 3));
127       const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3));
128 
129       const __m128 vi0x89AB = _mm_loadu_ps(i0);
130       i0 += 4;
131       const __m128 vi1x89AB = _mm_loadu_ps(i1);
132       i1 += 4;
133       const __m128 vi2x89AB = _mm_loadu_ps(i2);
134       i2 += 4;
135       const __m128 vi3x89AB = _mm_loadu_ps(i3);
136       i3 += 4;
137       const __m128 vi4x89AB = _mm_loadu_ps(i4);
138       i4 += 4;
139       const __m128 vi5x89AB = _mm_loadu_ps(i5);
140       i5 += 4;
141 
142       const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
143       const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
144       const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
145       const __m128 vi3x3456 = _mm_move_ss(vi3x7456, vi3x3012);
146       const __m128 vi4x3456 = _mm_move_ss(vi4x7456, vi4x3012);
147       const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012);
148 
149       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0x3456, vk01));
150       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi1x3456, vk01));
151       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk11));
152       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk11));
153       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk21));
154       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi3x3456, vk21));
155       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi3x3456, vk31));
156       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi4x3456, vk31));
157       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x3456, vk41));
158       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x3456, vk41));
159 
160       const __m128 vi0x2345 = _mm_shuffle_ps(vi0x3012, vi0x7456, _MM_SHUFFLE(2, 1, 0, 3));
161       vi0x3012 = vi0x7456;
162       const __m128 vi1x2345 = _mm_shuffle_ps(vi1x3012, vi1x7456, _MM_SHUFFLE(2, 1, 0, 3));
163       vi1x3012 = vi1x7456;
164       const __m128 vi2x2345 = _mm_shuffle_ps(vi2x3012, vi2x7456, _MM_SHUFFLE(2, 1, 0, 3));
165       vi2x3012 = vi2x7456;
166       const __m128 vi3x2345 = _mm_shuffle_ps(vi3x3012, vi3x7456, _MM_SHUFFLE(2, 1, 0, 3));
167       vi3x3012 = vi3x7456;
168       const __m128 vi4x2345 = _mm_shuffle_ps(vi4x3012, vi4x7456, _MM_SHUFFLE(2, 1, 0, 3));
169       vi4x3012 = vi4x7456;
170       const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3));
171       vi5x3012 = vi5x7456;
172 
173       const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB);
174       vi0x4567 = vi0x89AB;
175       const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB);
176       vi1x4567 = vi1x89AB;
177       const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB);
178       vi2x4567 = vi2x89AB;
179       const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB);
180       vi3x4567 = vi3x89AB;
181       const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB);
182       vi4x4567 = vi4x89AB;
183       const __m128 vi5x8567 = _mm_move_ss(vi5x4567, vi5x89AB);
184       vi5x4567 = vi5x89AB;
185 
186       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x2345, vk00));
187       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi1x2345, vk00));
188       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi1x2345, vk10));
189       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi2x2345, vk10));
190       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x2345, vk20));
191       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x2345, vk20));
192       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi3x2345, vk30));
193       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi4x2345, vk30));
194       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi4x2345, vk40));
195       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi5x2345, vk40));
196 
197       const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
198       const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
199       const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
200       const __m128 vi3x5678 = _mm_shuffle_ps(vi3x8567, vi3x8567, _MM_SHUFFLE(0, 3, 2, 1));
201       const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));
202       const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));
203 
204       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk03));
205       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk03));
206       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x5678, vk13));
207       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi2x5678, vk13));
208       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi2x5678, vk23));
209       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi3x5678, vk23));
210       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x5678, vk33));
211       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x5678, vk33));
212       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x5678, vk43));
213       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5x5678, vk43));
214 
215       const __m128 vi0x6789 = _mm_shuffle_ps(vi0x5678, vi0x89AB, _MM_SHUFFLE(1, 0, 2, 1));
216       const __m128 vi1x6789 = _mm_shuffle_ps(vi1x5678, vi1x89AB, _MM_SHUFFLE(1, 0, 2, 1));
217       const __m128 vi2x6789 = _mm_shuffle_ps(vi2x5678, vi2x89AB, _MM_SHUFFLE(1, 0, 2, 1));
218       const __m128 vi3x6789 = _mm_shuffle_ps(vi3x5678, vi3x89AB, _MM_SHUFFLE(1, 0, 2, 1));
219       const __m128 vi4x6789 = _mm_shuffle_ps(vi4x5678, vi4x89AB, _MM_SHUFFLE(1, 0, 2, 1));
220       const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1));
221 
222       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0x6789, vk04));
223       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi1x6789, vk04));
224       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x6789, vk14));
225       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x6789, vk14));
226       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x6789, vk24));
227       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi3x6789, vk24));
228       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi3x6789, vk34));
229       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi4x6789, vk34));
230       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x6789, vk44));
231       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44));
232 
233       vo0p0 = _mm_add_ps(vo0p0, vo0p1);
234       vo1p0 = _mm_add_ps(vo1p0, vo1p1);
235       vo0p0 = _mm_add_ps(vo0p0, vo0p2);
236       vo1p0 = _mm_add_ps(vo1p0, vo1p2);
237 
238       __m128 vo0 = _mm_max_ps(vo0p0, vmin);
239       __m128 vo1 = _mm_max_ps(vo1p0, vmin);
240 
241       vo0 = _mm_min_ps(vo0, vmax);
242       vo1 = _mm_min_ps(vo1, vmax);
243 
244       _mm_storeu_ps(o1, vo1);
245       o1 += 4;
246       _mm_storeu_ps(o0, vo0);
247       o0 += 4;
248     }
249     // Always process the last block of 5..8 pixels.
250     if XNN_LIKELY(w > 4 * sizeof(float)) {
251       __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk02));
252       __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk02));
253       __m128 vo0p1 = _mm_mul_ps(vi1x4567, vk12);
254       __m128 vo1p1 = _mm_mul_ps(vi2x4567, vk12);
255       __m128 vo0p2 = _mm_mul_ps(vi2x4567, vk22);
256       __m128 vo1p2 = _mm_mul_ps(vi3x4567, vk22);
257       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x4567, vk32));
258       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x4567, vk32));
259       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x4567, vk42));
260       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5x4567, vk42));
261 
262       const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
263       const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
264       const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
265       const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3));
266       const __m128 vi4x7456 = _mm_shuffle_ps(vi4x4567, vi4x4567, _MM_SHUFFLE(2, 1, 0, 3));
267       const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3));
268 
269       const __m128 vi0x89AB = _mm_and_ps(_mm_loadu_ps(i0), vmask);
270       i0 += 4;
271       const __m128 vi1x89AB = _mm_and_ps(_mm_loadu_ps(i1), vmask);
272       i1 += 4;
273       const __m128 vi2x89AB = _mm_and_ps(_mm_loadu_ps(i2), vmask);
274       i2 += 4;
275       const __m128 vi3x89AB = _mm_and_ps(_mm_loadu_ps(i3), vmask);
276       i3 += 4;
277       const __m128 vi4x89AB = _mm_and_ps(_mm_loadu_ps(i4), vmask);
278       i4 += 4;
279       const __m128 vi5x89AB = _mm_and_ps(_mm_loadu_ps(i5), vmask);
280       i5 += 4;
281 
282       const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
283       const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
284       const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
285       const __m128 vi3x3456 = _mm_move_ss(vi3x7456, vi3x3012);
286       const __m128 vi4x3456 = _mm_move_ss(vi4x7456, vi4x3012);
287       const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012);
288 
289       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0x3456, vk01));
290       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi1x3456, vk01));
291       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk11));
292       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk11));
293       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk21));
294       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi3x3456, vk21));
295       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi3x3456, vk31));
296       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi4x3456, vk31));
297       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x3456, vk41));
298       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x3456, vk41));
299 
300       const __m128 vi0x2345 = _mm_shuffle_ps(vi0x3012, vi0x7456, _MM_SHUFFLE(2, 1, 0, 3));
301       vi0x3012 = vi0x7456;
302       const __m128 vi1x2345 = _mm_shuffle_ps(vi1x3012, vi1x7456, _MM_SHUFFLE(2, 1, 0, 3));
303       vi1x3012 = vi1x7456;
304       const __m128 vi2x2345 = _mm_shuffle_ps(vi2x3012, vi2x7456, _MM_SHUFFLE(2, 1, 0, 3));
305       vi2x3012 = vi2x7456;
306       const __m128 vi3x2345 = _mm_shuffle_ps(vi3x3012, vi3x7456, _MM_SHUFFLE(2, 1, 0, 3));
307       vi3x3012 = vi3x7456;
308       const __m128 vi4x2345 = _mm_shuffle_ps(vi4x3012, vi4x7456, _MM_SHUFFLE(2, 1, 0, 3));
309       vi4x3012 = vi4x7456;
310       const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3));
311       vi5x3012 = vi5x7456;
312 
313       const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB);
314       vi0x4567 = vi0x89AB;
315       const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB);
316       vi1x4567 = vi1x89AB;
317       const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB);
318       vi2x4567 = vi2x89AB;
319       const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB);
320       vi3x4567 = vi3x89AB;
321       const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB);
322       vi4x4567 = vi4x89AB;
323       const __m128 vi5x8567 = _mm_move_ss(vi5x4567, vi5x89AB);
324       vi5x4567 = vi5x89AB;
325 
326       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x2345, vk00));
327       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi1x2345, vk00));
328       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi1x2345, vk10));
329       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi2x2345, vk10));
330       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x2345, vk20));
331       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x2345, vk20));
332       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi3x2345, vk30));
333       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi4x2345, vk30));
334       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi4x2345, vk40));
335       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi5x2345, vk40));
336 
337       const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
338       const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
339       const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
340       const __m128 vi3x5678 = _mm_shuffle_ps(vi3x8567, vi3x8567, _MM_SHUFFLE(0, 3, 2, 1));
341       const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));
342       const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));
343 
344       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk03));
345       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk03));
346       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x5678, vk13));
347       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi2x5678, vk13));
348       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi2x5678, vk23));
349       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi3x5678, vk23));
350       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x5678, vk33));
351       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x5678, vk33));
352       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x5678, vk43));
353       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5x5678, vk43));
354 
355       const __m128 vi0x6789 = _mm_shuffle_ps(vi0x5678, vi0x89AB, _MM_SHUFFLE(1, 0, 2, 1));
356       const __m128 vi1x6789 = _mm_shuffle_ps(vi1x5678, vi1x89AB, _MM_SHUFFLE(1, 0, 2, 1));
357       const __m128 vi2x6789 = _mm_shuffle_ps(vi2x5678, vi2x89AB, _MM_SHUFFLE(1, 0, 2, 1));
358       const __m128 vi3x6789 = _mm_shuffle_ps(vi3x5678, vi3x89AB, _MM_SHUFFLE(1, 0, 2, 1));
359       const __m128 vi4x6789 = _mm_shuffle_ps(vi4x5678, vi4x89AB, _MM_SHUFFLE(1, 0, 2, 1));
360       const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1));
361 
362       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0x6789, vk04));
363       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi1x6789, vk04));
364       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x6789, vk14));
365       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x6789, vk14));
366       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x6789, vk24));
367       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi3x6789, vk24));
368       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi3x6789, vk34));
369       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi4x6789, vk34));
370       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x6789, vk44));
371       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44));
372 
373       vo0p0 = _mm_add_ps(vo0p0, vo0p1);
374       vo1p0 = _mm_add_ps(vo1p0, vo1p1);
375       vo0p0 = _mm_add_ps(vo0p0, vo0p2);
376       vo1p0 = _mm_add_ps(vo1p0, vo1p2);
377 
378       __m128 vo0 = _mm_max_ps(vo0p0, vmin);
379       __m128 vo1 = _mm_max_ps(vo1p0, vmin);
380 
381       vo0 = _mm_min_ps(vo0, vmax);
382       vo1 = _mm_min_ps(vo1, vmax);
383 
384       _mm_storeu_ps(o1, vo1);
385       o1 += 4;
386       _mm_storeu_ps(o0, vo0);
387       o0 += 4;
388 
389       w -= 4 * sizeof(float);
390     }
391     assert(w >= 1 * sizeof(float));
392     assert(w <= 4 * sizeof(float));
393     {
394       vi0x4567 = _mm_and_ps(vi0x4567, vmask);
395       vi1x4567 = _mm_and_ps(vi1x4567, vmask);
396       vi2x4567 = _mm_and_ps(vi2x4567, vmask);
397       vi3x4567 = _mm_and_ps(vi3x4567, vmask);
398       vi4x4567 = _mm_and_ps(vi4x4567, vmask);
399       vi5x4567 = _mm_and_ps(vi5x4567, vmask);
400 
401       __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk02));
402       __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk02));
403       __m128 vo0p1 = _mm_mul_ps(vi1x4567, vk12);
404       __m128 vo1p1 = _mm_mul_ps(vi2x4567, vk12);
405       __m128 vo0p2 = _mm_mul_ps(vi2x4567, vk22);
406       __m128 vo1p2 = _mm_mul_ps(vi3x4567, vk22);
407       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x4567, vk32));
408       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x4567, vk32));
409       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x4567, vk42));
410       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5x4567, vk42));
411 
412       const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
413       const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
414       const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
415       const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3));
416       const __m128 vi4x7456 = _mm_shuffle_ps(vi4x4567, vi4x4567, _MM_SHUFFLE(2, 1, 0, 3));
417       const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3));
418 
419       const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
420       const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
421       const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
422       const __m128 vi3x3456 = _mm_move_ss(vi3x7456, vi3x3012);
423       const __m128 vi4x3456 = _mm_move_ss(vi4x7456, vi4x3012);
424       const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012);
425 
426       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0x3456, vk01));
427       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi1x3456, vk01));
428       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk11));
429       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk11));
430       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk21));
431       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi3x3456, vk21));
432       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi3x3456, vk31));
433       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi4x3456, vk31));
434       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x3456, vk41));
435       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x3456, vk41));
436 
437       const __m128 vi0x2345 = _mm_shuffle_ps(vi0x3012, vi0x7456, _MM_SHUFFLE(2, 1, 0, 3));
438       const __m128 vi1x2345 = _mm_shuffle_ps(vi1x3012, vi1x7456, _MM_SHUFFLE(2, 1, 0, 3));
439       const __m128 vi2x2345 = _mm_shuffle_ps(vi2x3012, vi2x7456, _MM_SHUFFLE(2, 1, 0, 3));
440       const __m128 vi3x2345 = _mm_shuffle_ps(vi3x3012, vi3x7456, _MM_SHUFFLE(2, 1, 0, 3));
441       const __m128 vi4x2345 = _mm_shuffle_ps(vi4x3012, vi4x7456, _MM_SHUFFLE(2, 1, 0, 3));
442       const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3));
443 
444       const __m128 vzero = _mm_setzero_ps();
445       const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero);
446       const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero);
447       const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero);
448       const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero);
449       const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero);
450       const __m128 vi5x8567 = _mm_move_ss(vi5x4567, vzero);
451 
452       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x2345, vk00));
453       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi1x2345, vk00));
454       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi1x2345, vk10));
455       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi2x2345, vk10));
456       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x2345, vk20));
457       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x2345, vk20));
458       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi3x2345, vk30));
459       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi4x2345, vk30));
460       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi4x2345, vk40));
461       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi5x2345, vk40));
462 
463       const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
464       const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
465       const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
466       const __m128 vi3x5678 = _mm_shuffle_ps(vi3x8567, vi3x8567, _MM_SHUFFLE(0, 3, 2, 1));
467       const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));
468       const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));
469 
470       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk03));
471       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk03));
472       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x5678, vk13));
473       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi2x5678, vk13));
474       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi2x5678, vk23));
475       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi3x5678, vk23));
476       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x5678, vk33));
477       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x5678, vk33));
478       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x5678, vk43));
479       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5x5678, vk43));
480 
481       const __m128 vi0x6789 = _mm_shuffle_ps(vi0x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
482       const __m128 vi1x6789 = _mm_shuffle_ps(vi1x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
483       const __m128 vi2x6789 = _mm_shuffle_ps(vi2x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
484       const __m128 vi3x6789 = _mm_shuffle_ps(vi3x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
485       const __m128 vi4x6789 = _mm_shuffle_ps(vi4x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
486       const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
487 
488       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0x6789, vk04));
489       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi1x6789, vk04));
490       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x6789, vk14));
491       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x6789, vk14));
492       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x6789, vk24));
493       vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi3x6789, vk24));
494       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi3x6789, vk34));
495       vo1p2 = _mm_add_ps(vo1p2, _mm_mul_ps(vi4x6789, vk34));
496       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x6789, vk44));
497       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44));
498 
499       vo0p0 = _mm_add_ps(vo0p0, vo0p1);
500       vo1p0 = _mm_add_ps(vo1p0, vo1p1);
501       vo0p0 = _mm_add_ps(vo0p0, vo0p2);
502       vo1p0 = _mm_add_ps(vo1p0, vo1p2);
503 
504       __m128 vo0 = _mm_max_ps(vo0p0, vmin);
505       __m128 vo1 = _mm_max_ps(vo1p0, vmin);
506 
507       vo0 = _mm_min_ps(vo0, vmax);
508       vo1 = _mm_min_ps(vo1, vmax);
509 
510       if XNN_LIKELY(w & (4 * sizeof(float))) {
511         _mm_storeu_ps(o1, vo1);
512         o1 += 4;
513         _mm_storeu_ps(o0, vo0);
514         o0 += 4;
515       } else {
516         if (w & (2 * sizeof(float))) {
517           _mm_storel_pi((__m64*) o1, vo1);
518           o1 += 2;
519           _mm_storel_pi((__m64*) o0, vo0);
520           o0 += 2;
521 
522           vo0 = _mm_movehl_ps(vo0, vo0);
523           vo1 = _mm_movehl_ps(vo1, vo1);
524         }
525         if (w & (1 * sizeof(float))) {
526           _mm_store_ss(o1, vo1);
527           o1 += 1;
528           _mm_store_ss(o0, vo0);
529           o0 += 1;
530         }
531       }
532     }
533 
534     i0 = (const float*) ((uintptr_t) i2 - input_decrement);
535     i1 = (const float*) ((uintptr_t) i3 - input_decrement);
536     i2 = (const float*) ((uintptr_t) i1 + input_width);
537     i3 = (const float*) ((uintptr_t) i2 + input_width);
538     i4 = (const float*) ((uintptr_t) i3 + input_width);
539     i5 = (const float*) ((uintptr_t) i4 + input_width);
540 
541     o0 = o1;
542     o1 = (float*) ((uintptr_t) o0 + input_width);
543 
544     output_height = doz(output_height, 2);
545   } while (output_height != 0);
546 }
547