1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv2d-chw/5x5s2p2-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_5x5s2p2__sse_1x4_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_5x5s2p2__sse_1x4_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 >= 1);
32   assert(padding_top <= 2);
33 
34   const __m128 vmask_even = _mm_load_ps((const float*) params->sse.mask_even);
35   const __m128 vmask_odd  = _mm_load_ps((const float*) params->sse.mask_odd);
36   const __m128 vmax = _mm_load_ps(params->sse.max);
37   const __m128 vmin = _mm_load_ps(params->sse.min);
38 
39   const __m128 vbias = _mm_load1_ps(weights);
40   const __m128 vk00 = _mm_load1_ps(weights + 1);
41   const __m128 vk01 = _mm_load1_ps(weights + 2);
42   const __m128 vk02 = _mm_load1_ps(weights + 3);
43   const __m128 vk03 = _mm_load1_ps(weights + 4);
44   const __m128 vk04 = _mm_load1_ps(weights + 5);
45   const __m128 vk10 = _mm_load1_ps(weights + 6);
46   const __m128 vk11 = _mm_load1_ps(weights + 7);
47   const __m128 vk12 = _mm_load1_ps(weights + 8);
48   const __m128 vk13 = _mm_load1_ps(weights + 9);
49   const __m128 vk14 = _mm_load1_ps(weights + 10);
50   const __m128 vk20 = _mm_load1_ps(weights + 11);
51   const __m128 vk21 = _mm_load1_ps(weights + 12);
52   const __m128 vk22 = _mm_load1_ps(weights + 13);
53   const __m128 vk23 = _mm_load1_ps(weights + 14);
54   const __m128 vk24 = _mm_load1_ps(weights + 15);
55   const __m128 vk30 = _mm_load1_ps(weights + 16);
56   const __m128 vk31 = _mm_load1_ps(weights + 17);
57   const __m128 vk32 = _mm_load1_ps(weights + 18);
58   const __m128 vk33 = _mm_load1_ps(weights + 19);
59   const __m128 vk34 = _mm_load1_ps(weights + 20);
60   const __m128 vk40 = _mm_load1_ps(weights + 21);
61   const __m128 vk41 = _mm_load1_ps(weights + 22);
62   const __m128 vk42 = _mm_load1_ps(weights + 23);
63   const __m128 vk43 = _mm_load1_ps(weights + 24);
64   const __m128 vk44 = _mm_load1_ps(weights + 25);
65 
66   const uint32_t padding_top_less_1 = padding_top - 1;
67   const size_t input_decrement = round_up_po2(input_width, 8 * sizeof(float));
68 
69   const float* i0 = zero;
70   const float* i1 = (const float*) ((uintptr_t) input - ((-padding_top_less_1) & input_width));
71   const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
72   if XNN_UNPREDICTABLE(padding_top_less_1 != 0) {
73     i1 = zero;
74   }
75   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
76   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
77 
78 
79   float* o0 = output;
80 
81   size_t padded_input_height = input_height + (padding_top_less_1 + 1) + 2 /* padding bottom */;
82   size_t output_height = (padded_input_height - 5 /* kernel size */ + 2 /* subsampling */) / 2;
83   do {
84     if XNN_UNPREDICTABLE(padded_input_height < 6) {
85       i3 = zero;
86     }
87     if XNN_UNPREDICTABLE(padded_input_height < 7) {
88       i4 = zero;
89     }
90 
91     __m128 vi0x6024 = _mm_setzero_ps();
92     __m128 vi1x6024 = _mm_setzero_ps();
93     __m128 vi2x6024 = _mm_setzero_ps();
94     __m128 vi3x6024 = _mm_setzero_ps();
95     __m128 vi4x6024 = _mm_setzero_ps();
96 
97     __m128 vi0x7135 = _mm_setzero_ps();
98     __m128 vi1x7135 = _mm_setzero_ps();
99     __m128 vi2x7135 = _mm_setzero_ps();
100     __m128 vi3x7135 = _mm_setzero_ps();
101     __m128 vi4x7135 = _mm_setzero_ps();
102 
103     const __m128 vi0x89AB = _mm_loadu_ps(i0);
104     const __m128 vi0xCDEF = _mm_loadu_ps(i0 + 4);
105     i0 += 8;
106     const __m128 vi1x89AB = _mm_loadu_ps(i1);
107     const __m128 vi1xCDEF = _mm_loadu_ps(i1 + 4);
108     i1 += 8;
109     const __m128 vi2x89AB = _mm_loadu_ps(i2);
110     const __m128 vi2xCDEF = _mm_loadu_ps(i2 + 4);
111     i2 += 8;
112     const __m128 vi3x89AB = _mm_loadu_ps(i3);
113     const __m128 vi3xCDEF = _mm_loadu_ps(i3 + 4);
114     i3 += 8;
115     const __m128 vi4x89AB = _mm_loadu_ps(i4);
116     const __m128 vi4xCDEF = _mm_loadu_ps(i4 + 4);
117     i4 += 8;
118 
119     __m128 vi0x8ACE = _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(2, 0, 2, 0));
120     __m128 vi0x9BDF = _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(3, 1, 3, 1));
121     __m128 vi1x8ACE = _mm_shuffle_ps(vi1x89AB, vi1xCDEF, _MM_SHUFFLE(2, 0, 2, 0));
122     __m128 vi1x9BDF = _mm_shuffle_ps(vi1x89AB, vi1xCDEF, _MM_SHUFFLE(3, 1, 3, 1));
123     __m128 vi2x8ACE = _mm_shuffle_ps(vi2x89AB, vi2xCDEF, _MM_SHUFFLE(2, 0, 2, 0));
124     __m128 vi2x9BDF = _mm_shuffle_ps(vi2x89AB, vi2xCDEF, _MM_SHUFFLE(3, 1, 3, 1));
125     __m128 vi3x8ACE = _mm_shuffle_ps(vi3x89AB, vi3xCDEF, _MM_SHUFFLE(2, 0, 2, 0));
126     __m128 vi3x9BDF = _mm_shuffle_ps(vi3x89AB, vi3xCDEF, _MM_SHUFFLE(3, 1, 3, 1));
127     __m128 vi4x8ACE = _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(2, 0, 2, 0));
128     __m128 vi4x9BDF = _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(3, 1, 3, 1));
129 
130     size_t w = input_width;
131     for (; w > 8 * sizeof(float); w -= 8 * sizeof(float)) {
132       __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x8ACE, vk02));
133       __m128 vo0p1 = _mm_mul_ps(vi1x8ACE, vk12);
134       __m128 vo0p2 = _mm_mul_ps(vi2x8ACE, vk22);
135       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x8ACE, vk32));
136       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x8ACE, vk42));
137 
138       const __m128 vi0xE8AC = _mm_shuffle_ps(vi0x8ACE, vi0x8ACE, _MM_SHUFFLE(2, 1, 0, 3));
139       const __m128 vi1xE8AC = _mm_shuffle_ps(vi1x8ACE, vi1x8ACE, _MM_SHUFFLE(2, 1, 0, 3));
140       const __m128 vi2xE8AC = _mm_shuffle_ps(vi2x8ACE, vi2x8ACE, _MM_SHUFFLE(2, 1, 0, 3));
141       const __m128 vi3xE8AC = _mm_shuffle_ps(vi3x8ACE, vi3x8ACE, _MM_SHUFFLE(2, 1, 0, 3));
142       const __m128 vi4xE8AC = _mm_shuffle_ps(vi4x8ACE, vi4x8ACE, _MM_SHUFFLE(2, 1, 0, 3));
143 
144       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0x9BDF, vk03));
145       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x9BDF, vk13));
146       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x9BDF, vk23));
147       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi3x9BDF, vk33));
148       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x9BDF, vk43));
149 
150       const __m128 vi0x68AC = _mm_move_ss(vi0xE8AC, vi0x6024);
151       vi0x6024 = vi0xE8AC;
152       const __m128 vi1x68AC = _mm_move_ss(vi1xE8AC, vi1x6024);
153       vi1x6024 = vi1xE8AC;
154       const __m128 vi2x68AC = _mm_move_ss(vi2xE8AC, vi2x6024);
155       vi2x6024 = vi2xE8AC;
156       const __m128 vi3x68AC = _mm_move_ss(vi3xE8AC, vi3x6024);
157       vi3x6024 = vi3xE8AC;
158       const __m128 vi4x68AC = _mm_move_ss(vi4xE8AC, vi4x6024);
159       vi4x6024 = vi4xE8AC;
160 
161       const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
162       const __m128 vi1xF9BD = _mm_shuffle_ps(vi1x9BDF, vi1x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
163       const __m128 vi2xF9BD = _mm_shuffle_ps(vi2x9BDF, vi2x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
164       const __m128 vi3xF9BD = _mm_shuffle_ps(vi3x9BDF, vi3x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
165       const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
166 
167       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x68AC, vk00));
168       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi1x68AC, vk10));
169       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x68AC, vk20));
170       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi3x68AC, vk30));
171       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi4x68AC, vk40));
172 
173       const __m128 vi0xGHIJ = _mm_loadu_ps(i0);
174       const __m128 vi0xKLMN = _mm_loadu_ps(i0 + 4);
175       i0 += 8;
176       const __m128 vi1xGHIJ = _mm_loadu_ps(i1);
177       const __m128 vi1xKLMN = _mm_loadu_ps(i1 + 4);
178       i1 += 8;
179       const __m128 vi2xGHIJ = _mm_loadu_ps(i2);
180       const __m128 vi2xKLMN = _mm_loadu_ps(i2 + 4);
181       i2 += 8;
182       const __m128 vi3xGHIJ = _mm_loadu_ps(i3);
183       const __m128 vi3xKLMN = _mm_loadu_ps(i3 + 4);
184       i3 += 8;
185       const __m128 vi4xGHIJ = _mm_loadu_ps(i4);
186       const __m128 vi4xKLMN = _mm_loadu_ps(i4 + 4);
187       i4 += 8;
188 
189       const __m128 vi0x79BD = _mm_move_ss(vi0xF9BD, vi0x7135);
190       vi0x7135 = vi0xF9BD;
191       const __m128 vi1x79BD = _mm_move_ss(vi1xF9BD, vi1x7135);
192       vi1x7135 = vi1xF9BD;
193       const __m128 vi2x79BD = _mm_move_ss(vi2xF9BD, vi2x7135);
194       vi2x7135 = vi2xF9BD;
195       const __m128 vi3x79BD = _mm_move_ss(vi3xF9BD, vi3x7135);
196       vi3x7135 = vi3xF9BD;
197       const __m128 vi4x79BD = _mm_move_ss(vi4xF9BD, vi4x7135);
198       vi4x7135 = vi4xF9BD;
199 
200       const __m128 vi0xGIKM = _mm_shuffle_ps(vi0xGHIJ, vi0xKLMN, _MM_SHUFFLE(2, 0, 2, 0));
201       const __m128 vi0xHJLN = _mm_shuffle_ps(vi0xGHIJ, vi0xKLMN, _MM_SHUFFLE(3, 1, 3, 1));
202       vi0x9BDF = vi0xHJLN;
203       const __m128 vi1xGIKM = _mm_shuffle_ps(vi1xGHIJ, vi1xKLMN, _MM_SHUFFLE(2, 0, 2, 0));
204       const __m128 vi1xHJLN = _mm_shuffle_ps(vi1xGHIJ, vi1xKLMN, _MM_SHUFFLE(3, 1, 3, 1));
205       vi1x9BDF = vi1xHJLN;
206       const __m128 vi2xGIKM = _mm_shuffle_ps(vi2xGHIJ, vi2xKLMN, _MM_SHUFFLE(2, 0, 2, 0));
207       const __m128 vi2xHJLN = _mm_shuffle_ps(vi2xGHIJ, vi2xKLMN, _MM_SHUFFLE(3, 1, 3, 1));
208       vi2x9BDF = vi2xHJLN;
209       const __m128 vi3xGIKM = _mm_shuffle_ps(vi3xGHIJ, vi3xKLMN, _MM_SHUFFLE(2, 0, 2, 0));
210       const __m128 vi3xHJLN = _mm_shuffle_ps(vi3xGHIJ, vi3xKLMN, _MM_SHUFFLE(3, 1, 3, 1));
211       vi3x9BDF = vi3xHJLN;
212       const __m128 vi4xGIKM = _mm_shuffle_ps(vi4xGHIJ, vi4xKLMN, _MM_SHUFFLE(2, 0, 2, 0));
213       const __m128 vi4xHJLN = _mm_shuffle_ps(vi4xGHIJ, vi4xKLMN, _MM_SHUFFLE(3, 1, 3, 1));
214       vi4x9BDF = vi4xHJLN;
215 
216       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x79BD, vk01));
217       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x79BD, vk11));
218       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi2x79BD, vk21));
219       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x79BD, vk31));
220       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x79BD, vk41));
221 
222       const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vi0xGIKM);
223       vi0x8ACE = vi0xGIKM;
224       const __m128 vi1xGACE = _mm_move_ss(vi1x8ACE, vi1xGIKM);
225       vi1x8ACE = vi1xGIKM;
226       const __m128 vi2xGACE = _mm_move_ss(vi2x8ACE, vi2xGIKM);
227       vi2x8ACE = vi2xGIKM;
228       const __m128 vi3xGACE = _mm_move_ss(vi3x8ACE, vi3xGIKM);
229       vi3x8ACE = vi3xGIKM;
230       const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vi4xGIKM);
231       vi4x8ACE = vi4xGIKM;
232 
233       const __m128 vi0xACEG = _mm_shuffle_ps(vi0xGACE, vi0xGACE, _MM_SHUFFLE(0, 3, 2, 1));
234       const __m128 vi1xACEG = _mm_shuffle_ps(vi1xGACE, vi1xGACE, _MM_SHUFFLE(0, 3, 2, 1));
235       const __m128 vi2xACEG = _mm_shuffle_ps(vi2xGACE, vi2xGACE, _MM_SHUFFLE(0, 3, 2, 1));
236       const __m128 vi3xACEG = _mm_shuffle_ps(vi3xGACE, vi3xGACE, _MM_SHUFFLE(0, 3, 2, 1));
237       const __m128 vi4xACEG = _mm_shuffle_ps(vi4xGACE, vi4xGACE, _MM_SHUFFLE(0, 3, 2, 1));
238 
239       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0xACEG, vk04));
240       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1xACEG, vk14));
241       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2xACEG, vk24));
242       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi3xACEG, vk34));
243       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4xACEG, vk44));
244 
245       vo0p0 = _mm_add_ps(vo0p0, vo0p1);
246       vo0p0 = _mm_add_ps(vo0p0, vo0p2);
247 
248       __m128 vo0 = _mm_max_ps(vo0p0, vmin);
249 
250       vo0 = _mm_min_ps(vo0, vmax);
251 
252       _mm_storeu_ps(o0, vo0);
253       o0 += 4;
254     }
255     // Last block has 1-8 pixels to process.
256     assert(w <= 8 * sizeof(float));
257     assert(w >= 1 * sizeof(float));
258     {
259       vi0x8ACE = _mm_and_ps(vi0x8ACE, vmask_even);
260       vi0x9BDF = _mm_and_ps(vi0x9BDF, vmask_odd);
261       vi1x8ACE = _mm_and_ps(vi1x8ACE, vmask_even);
262       vi1x9BDF = _mm_and_ps(vi1x9BDF, vmask_odd);
263       vi2x8ACE = _mm_and_ps(vi2x8ACE, vmask_even);
264       vi2x9BDF = _mm_and_ps(vi2x9BDF, vmask_odd);
265       vi3x8ACE = _mm_and_ps(vi3x8ACE, vmask_even);
266       vi3x9BDF = _mm_and_ps(vi3x9BDF, vmask_odd);
267       vi4x8ACE = _mm_and_ps(vi4x8ACE, vmask_even);
268       vi4x9BDF = _mm_and_ps(vi4x9BDF, vmask_odd);
269 
270       __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x8ACE, vk02));
271       __m128 vo0p1 = _mm_mul_ps(vi1x8ACE, vk12);
272       __m128 vo0p2 = _mm_mul_ps(vi2x8ACE, vk22);
273       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x8ACE, vk32));
274       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x8ACE, vk42));
275 
276       const __m128 vi0xE8AC = _mm_shuffle_ps(vi0x8ACE, vi0x8ACE, _MM_SHUFFLE(2, 1, 0, 3));
277       const __m128 vi1xE8AC = _mm_shuffle_ps(vi1x8ACE, vi1x8ACE, _MM_SHUFFLE(2, 1, 0, 3));
278       const __m128 vi2xE8AC = _mm_shuffle_ps(vi2x8ACE, vi2x8ACE, _MM_SHUFFLE(2, 1, 0, 3));
279       const __m128 vi3xE8AC = _mm_shuffle_ps(vi3x8ACE, vi3x8ACE, _MM_SHUFFLE(2, 1, 0, 3));
280       const __m128 vi4xE8AC = _mm_shuffle_ps(vi4x8ACE, vi4x8ACE, _MM_SHUFFLE(2, 1, 0, 3));
281 
282       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0x9BDF, vk03));
283       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x9BDF, vk13));
284       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x9BDF, vk23));
285       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi3x9BDF, vk33));
286       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x9BDF, vk43));
287 
288       const __m128 vi0x68AC = _mm_move_ss(vi0xE8AC, vi0x6024);
289       const __m128 vi1x68AC = _mm_move_ss(vi1xE8AC, vi1x6024);
290       const __m128 vi2x68AC = _mm_move_ss(vi2xE8AC, vi2x6024);
291       const __m128 vi3x68AC = _mm_move_ss(vi3xE8AC, vi3x6024);
292       const __m128 vi4x68AC = _mm_move_ss(vi4xE8AC, vi4x6024);
293 
294       const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
295       const __m128 vi1xF9BD = _mm_shuffle_ps(vi1x9BDF, vi1x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
296       const __m128 vi2xF9BD = _mm_shuffle_ps(vi2x9BDF, vi2x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
297       const __m128 vi3xF9BD = _mm_shuffle_ps(vi3x9BDF, vi3x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
298       const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
299 
300       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x68AC, vk00));
301       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi1x68AC, vk10));
302       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x68AC, vk20));
303       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi3x68AC, vk30));
304       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi4x68AC, vk40));
305 
306       const __m128 vi0x79BD = _mm_move_ss(vi0xF9BD, vi0x7135);
307       const __m128 vi1x79BD = _mm_move_ss(vi1xF9BD, vi1x7135);
308       const __m128 vi2x79BD = _mm_move_ss(vi2xF9BD, vi2x7135);
309       const __m128 vi3x79BD = _mm_move_ss(vi3xF9BD, vi3x7135);
310       const __m128 vi4x79BD = _mm_move_ss(vi4xF9BD, vi4x7135);
311 
312       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x79BD, vk01));
313       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x79BD, vk11));
314       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi2x79BD, vk21));
315       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x79BD, vk31));
316       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x79BD, vk41));
317 
318       const __m128 vzero = _mm_setzero_ps();
319       const __m128 vi0xGACE = _mm_move_ss(vi0x8ACE, vzero);
320       const __m128 vi1xGACE = _mm_move_ss(vi1x8ACE, vzero);
321       const __m128 vi2xGACE = _mm_move_ss(vi2x8ACE, vzero);
322       const __m128 vi3xGACE = _mm_move_ss(vi3x8ACE, vzero);
323       const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vzero);
324 
325       const __m128 vi0xACEG = _mm_shuffle_ps(vi0xGACE, vi0xGACE, _MM_SHUFFLE(0, 3, 2, 1));
326       const __m128 vi1xACEG = _mm_shuffle_ps(vi1xGACE, vi1xGACE, _MM_SHUFFLE(0, 3, 2, 1));
327       const __m128 vi2xACEG = _mm_shuffle_ps(vi2xGACE, vi2xGACE, _MM_SHUFFLE(0, 3, 2, 1));
328       const __m128 vi3xACEG = _mm_shuffle_ps(vi3xGACE, vi3xGACE, _MM_SHUFFLE(0, 3, 2, 1));
329       const __m128 vi4xACEG = _mm_shuffle_ps(vi4xGACE, vi4xGACE, _MM_SHUFFLE(0, 3, 2, 1));
330 
331       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0xACEG, vk04));
332       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1xACEG, vk14));
333       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2xACEG, vk24));
334       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi3xACEG, vk34));
335       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4xACEG, vk44));
336 
337       vo0p0 = _mm_add_ps(vo0p0, vo0p1);
338       vo0p0 = _mm_add_ps(vo0p0, vo0p2);
339 
340       __m128 vo0 = _mm_max_ps(vo0p0, vmin);
341 
342       vo0 = _mm_min_ps(vo0, vmax);
343 
344       size_t w_tmp = (w + 1 * sizeof(float)) / (2 * sizeof(float));
345       if XNN_LIKELY(w_tmp >= 4) {
346         _mm_storeu_ps(o0, vo0);
347         o0 += 4;
348       } else {
349         if (w_tmp & 2) {
350           _mm_storel_pi((__m64*) o0, vo0);
351           o0 += 2;
352 
353           vo0 = _mm_movehl_ps(vo0, vo0);
354         }
355         if (w_tmp & 1) {
356           _mm_store_ss(o0, vo0);
357           o0 += 1;
358         }
359       }
360     }
361 
362     i0 = (const float*) ((uintptr_t) i2 - input_decrement);
363     i1 = (const float*) ((uintptr_t) i3 - input_decrement);
364     i2 = (const float*) ((uintptr_t) i4 - input_decrement);
365     i3 = (const float*) ((uintptr_t) i2 + input_width);
366     i4 = (const float*) ((uintptr_t) i3 + input_width);
367 
368 
369     output_height -= 1;
370     padded_input_height -= 2;
371   } while (output_height != 0);
372 }
373