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_4x4(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_4x4(
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   const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
56 
57   float* o0 = output;
58   float* o1 = (float*) ((uintptr_t) o0 + input_width);
59   float* o2 = (float*) ((uintptr_t) o1 + input_width);
60   float* o3 = (float*) ((uintptr_t) o2 + input_width);
61 
62   size_t output_height = input_height;
63   do {
64     if XNN_UNPREDICTABLE(output_height < 2) {
65       i2 = zero;
66       o1 = o0;
67     }
68     if XNN_UNPREDICTABLE(output_height < 3) {
69       i3 = zero;
70       o2 = o1;
71     }
72     if XNN_UNPREDICTABLE(output_height < 4) {
73       i4 = zero;
74       o3 = o2;
75     }
76     if XNN_UNPREDICTABLE(output_height < 5) {
77       i5 = zero;
78     }
79 
80     // vi0x3012 = ( vi02, vi01, vi{M}0, vi{M}3 )
81     __m128 vi0x3012 = _mm_setzero_ps();
82     // vi1x3012 = ( vi12, vi11, vi{M}0, vi{M}3 )
83     __m128 vi1x3012 = _mm_setzero_ps();
84     // vi2x3012 = ( vi22, vi21, vi{M}0, vi{M}3 )
85     __m128 vi2x3012 = _mm_setzero_ps();
86     // vi3x3012 = ( vi32, vi31, vi{M}0, vi{M}3 )
87     __m128 vi3x3012 = _mm_setzero_ps();
88     // vi4x3012 = ( vi42, vi41, vi{M}0, vi{M}3 )
89     __m128 vi4x3012 = _mm_setzero_ps();
90     // vi5x3012 = ( vi52, vi51, vi{M}0, vi{M}3 )
91     __m128 vi5x3012 = _mm_setzero_ps();
92 
93     __m128 vi0x4567 = _mm_loadu_ps(i0);
94     i0 += 4;
95     __m128 vi1x4567 = _mm_loadu_ps(i1);
96     i1 += 4;
97     __m128 vi2x4567 = _mm_loadu_ps(i2);
98     i2 += 4;
99     __m128 vi3x4567 = _mm_loadu_ps(i3);
100     i3 += 4;
101     __m128 vi4x4567 = _mm_loadu_ps(i4);
102     i4 += 4;
103     __m128 vi5x4567 = _mm_loadu_ps(i5);
104     i5 += 4;
105 
106     size_t w = input_width;
107     for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
108       // vi0x89AB = ( vi0B, vi0A, vi09, vi08 )
109       const __m128 vi0x89AB = _mm_loadu_ps(i0);
110       i0 += 4;
111       // vi1x89AB = ( vi1B, vi1A, vi19, vi18 )
112       const __m128 vi1x89AB = _mm_loadu_ps(i1);
113       i1 += 4;
114       // vi2x89AB = ( vi2B, vi2A, vi29, vi28 )
115       const __m128 vi2x89AB = _mm_loadu_ps(i2);
116       i2 += 4;
117       // vi3x89AB = ( vi3B, vi3A, vi39, vi38 )
118       const __m128 vi3x89AB = _mm_loadu_ps(i3);
119       i3 += 4;
120       // vi4x89AB = ( vi4B, vi4A, vi49, vi48 )
121       const __m128 vi4x89AB = _mm_loadu_ps(i4);
122       i4 += 4;
123       // vi5x89AB = ( vi5B, vi5A, vi59, vi58 )
124       const __m128 vi5x89AB = _mm_loadu_ps(i5);
125       i5 += 4;
126 
127       // vi0x7456 = ( vi06, vi05, vi04, vi07 )
128       const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
129       // vi1x7456 = ( vi16, vi15, vi14, vi17 )
130       const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
131       // vi2x7456 = ( vi26, vi25, vi24, vi27 )
132       const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
133       // vi3x7456 = ( vi36, vi35, vi34, vi37 )
134       const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3));
135       // vi4x7456 = ( vi46, vi45, vi44, vi47 )
136       const __m128 vi4x7456 = _mm_shuffle_ps(vi4x4567, vi4x4567, _MM_SHUFFLE(2, 1, 0, 3));
137       // vi5x7456 = ( vi56, vi55, vi54, vi57 )
138       const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3));
139 
140       __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
141       __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
142       __m128 vo2p0 = _mm_add_ps(vbias, _mm_mul_ps(vi2x4567, vk01));
143       __m128 vo3p0 = _mm_add_ps(vbias, _mm_mul_ps(vi3x4567, vk01));
144       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk11));
145       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk11));
146       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk11));
147       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x4567, vk11));
148       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
149       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));
150       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x4567, vk21));
151       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x4567, vk21));
152 
153       // vi0x3456 = ( vi06, vi05, vi04, vi03 )
154       const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
155       // vi1x3456 = ( vi16, vi15, vi14, vi13 )
156       const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
157       // vi2x3456 = ( vi26, vi25, vi24, vi23 )
158       const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
159       // vi3x3456 = ( vi36, vi35, vi34, vi33 )
160       const __m128 vi3x3456 = _mm_move_ss(vi3x7456, vi3x3012);
161       // vi4x3456 = ( vi46, vi45, vi44, vi43 )
162       const __m128 vi4x3456 = _mm_move_ss(vi4x7456, vi4x3012);
163       // vi5x3456 = ( vi56, vi55, vi54, vi53 )
164       const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012);
165 
166       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk00));
167       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk00));
168       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk00));
169       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x3456, vk00));
170       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
171       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
172       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x3456, vk10));
173       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x3456, vk10));
174       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20));
175       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x3456, vk20));
176       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x3456, vk20));
177       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk20));
178 
179       vi0x3012 = vi0x7456;
180       vi1x3012 = vi1x7456;
181       vi2x3012 = vi2x7456;
182       vi3x3012 = vi3x7456;
183       vi4x3012 = vi4x7456;
184       vi5x3012 = vi5x7456;
185 
186       // vi0x8567 = ( vi07, vi06, vi05, vi08 )
187       const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB);
188       // vi1x8567 = ( vi17, vi16, vi15, vi18 )
189       const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB);
190       // vi2x8567 = ( vi27, vi26, vi25, vi28 )
191       const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB);
192       // vi3x8567 = ( vi37, vi36, vi35, vi38 )
193       const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB);
194       // vi4x8567 = ( vi47, vi46, vi45, vi48 )
195       const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB);
196       // vi5x8567 = ( vi57, vi56, vi55, vi58 )
197       const __m128 vi5x8567 = _mm_move_ss(vi5x4567, vi5x89AB);
198 
199       // vi0x5678 = ( vi08, vi07, vi06, vi05 )
200       const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
201       // vi1x5678 = ( vi18, vi17, vi16, vi15 )
202       const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
203       // vi2x5678 = ( vi28, vi27, vi26, vi25 )
204       const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
205       // vi3x5678 = ( vi38, vi37, vi36, vi35 )
206       const __m128 vi3x5678 = _mm_shuffle_ps(vi3x8567, vi3x8567, _MM_SHUFFLE(0, 3, 2, 1));
207       // vi4x5678 = ( vi48, vi47, vi46, vi45 )
208       const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));
209       // vi5x5678 = ( vi58, vi57, vi56, vi55 )
210       const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));
211 
212       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
213       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
214       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x5678, vk02));
215       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x5678, vk02));
216       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk12));
217       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x5678, vk12));
218       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x5678, vk12));
219       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x5678, vk12));
220       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
221       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));
222       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk22));
223       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk22));
224 
225       vi0x4567 = vi0x89AB;
226       vi1x4567 = vi1x89AB;
227       vi2x4567 = vi2x89AB;
228       vi3x4567 = vi3x89AB;
229       vi4x4567 = vi4x89AB;
230       vi5x4567 = vi5x89AB;
231 
232 
233       __m128 vo0 = _mm_max_ps(vo0p0, vmin);
234       __m128 vo1 = _mm_max_ps(vo1p0, vmin);
235       __m128 vo2 = _mm_max_ps(vo2p0, vmin);
236       __m128 vo3 = _mm_max_ps(vo3p0, vmin);
237 
238       vo0 = _mm_min_ps(vo0, vmax);
239       vo1 = _mm_min_ps(vo1, vmax);
240       vo2 = _mm_min_ps(vo2, vmax);
241       vo3 = _mm_min_ps(vo3, vmax);
242 
243       _mm_storeu_ps(o3, vo3);
244       o3 += 4;
245       _mm_storeu_ps(o2, vo2);
246       o2 += 4;
247       _mm_storeu_ps(o1, vo1);
248       o1 += 4;
249       _mm_storeu_ps(o0, vo0);
250       o0 += 4;
251     }
252     // Always process the last block of 1..4 pixels.
253     assert(w >= 1 * sizeof(float));
254     assert(w <= 4 * sizeof(float));
255     {
256       vi0x4567 = _mm_and_ps(vmask, vi0x4567);
257       vi1x4567 = _mm_and_ps(vmask, vi1x4567);
258       vi2x4567 = _mm_and_ps(vmask, vi2x4567);
259       vi3x4567 = _mm_and_ps(vmask, vi3x4567);
260       vi4x4567 = _mm_and_ps(vmask, vi4x4567);
261       vi5x4567 = _mm_and_ps(vmask, vi5x4567);
262 
263       // vi0x7456 = ( vi06, vi05, vi04, vi07 )
264       const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
265       // vi1x7456 = ( vi16, vi15, vi14, vi17 )
266       const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
267       // vi2x7456 = ( vi26, vi25, vi24, vi27 )
268       const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
269       // vi3x7456 = ( vi36, vi35, vi34, vi37 )
270       const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3));
271       // vi4x7456 = ( vi46, vi45, vi44, vi47 )
272       const __m128 vi4x7456 = _mm_shuffle_ps(vi4x4567, vi4x4567, _MM_SHUFFLE(2, 1, 0, 3));
273       // vi5x7456 = ( vi56, vi55, vi54, vi57 )
274       const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3));
275 
276       __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
277       __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
278       __m128 vo2p0 = _mm_add_ps(vbias, _mm_mul_ps(vi2x4567, vk01));
279       __m128 vo3p0 = _mm_add_ps(vbias, _mm_mul_ps(vi3x4567, vk01));
280       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk11));
281       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk11));
282       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk11));
283       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x4567, vk11));
284       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
285       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));
286       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x4567, vk21));
287       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x4567, vk21));
288 
289       // vi0x3456 = ( vi06, vi05, vi04, vi03 )
290       const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
291       // vi1x3456 = ( vi16, vi15, vi14, vi13 )
292       const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
293       // vi2x3456 = ( vi26, vi25, vi24, vi23 )
294       const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
295       // vi3x3456 = ( vi36, vi35, vi34, vi33 )
296       const __m128 vi3x3456 = _mm_move_ss(vi3x7456, vi3x3012);
297       // vi4x3456 = ( vi46, vi45, vi44, vi43 )
298       const __m128 vi4x3456 = _mm_move_ss(vi4x7456, vi4x3012);
299       // vi5x3456 = ( vi56, vi55, vi54, vi53 )
300       const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012);
301 
302       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk00));
303       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk00));
304       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk00));
305       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x3456, vk00));
306       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
307       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
308       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x3456, vk10));
309       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x3456, vk10));
310       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20));
311       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x3456, vk20));
312       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x3456, vk20));
313       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk20));
314 
315       const __m128 vzero = _mm_setzero_ps();
316       // vi0x8567 = ( vi07, vi06, vi05, 0.0 )
317       const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero);
318       // vi1x8567 = ( vi17, vi16, vi15, 0.0 )
319       const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero);
320       // vi2x8567 = ( vi27, vi26, vi25, 0.0 )
321       const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero);
322       // vi3x8567 = ( vi37, vi36, vi35, 0.0 )
323       const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero);
324       // vi4x8567 = ( vi47, vi46, vi45, 0.0 )
325       const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero);
326       // vi5x8567 = ( vi57, vi56, vi55, 0.0 )
327       const __m128 vi5x8567 = _mm_move_ss(vi5x4567, vzero);
328 
329       // vi0x5678 = ( vi08, vi07, vi06, vi05 )
330       const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
331       // vi1x5678 = ( vi18, vi17, vi16, vi15 )
332       const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
333       // vi2x5678 = ( vi28, vi27, vi26, vi25 )
334       const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
335       // vi3x5678 = ( vi38, vi37, vi36, vi35 )
336       const __m128 vi3x5678 = _mm_shuffle_ps(vi3x8567, vi3x8567, _MM_SHUFFLE(0, 3, 2, 1));
337       // vi4x5678 = ( vi48, vi47, vi46, vi45 )
338       const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));
339       // vi5x5678 = ( vi58, vi57, vi56, vi55 )
340       const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));
341 
342       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
343       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
344       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x5678, vk02));
345       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x5678, vk02));
346       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk12));
347       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x5678, vk12));
348       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x5678, vk12));
349       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x5678, vk12));
350       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
351       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));
352       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk22));
353       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk22));
354 
355 
356       __m128 vo0 = _mm_max_ps(vo0p0, vmin);
357       __m128 vo1 = _mm_max_ps(vo1p0, vmin);
358       __m128 vo2 = _mm_max_ps(vo2p0, vmin);
359       __m128 vo3 = _mm_max_ps(vo3p0, vmin);
360 
361       vo0 = _mm_min_ps(vo0, vmax);
362       vo1 = _mm_min_ps(vo1, vmax);
363       vo2 = _mm_min_ps(vo2, vmax);
364       vo3 = _mm_min_ps(vo3, vmax);
365 
366       if XNN_LIKELY(w == 4 * sizeof(float)) {
367         _mm_storeu_ps(o3, vo3);
368         o3 += 4;
369         _mm_storeu_ps(o2, vo2);
370         o2 += 4;
371         _mm_storeu_ps(o1, vo1);
372         o1 += 4;
373         _mm_storeu_ps(o0, vo0);
374         o0 += 4;
375       } else {
376         if (w & (2 * sizeof(float))) {
377           _mm_storel_pi((__m64*) o3, vo3);
378           o3 += 2;
379           _mm_storel_pi((__m64*) o2, vo2);
380           o2 += 2;
381           _mm_storel_pi((__m64*) o1, vo1);
382           o1 += 2;
383           _mm_storel_pi((__m64*) o0, vo0);
384           o0 += 2;
385 
386           vo0 = _mm_movehl_ps(vo0, vo0);
387           vo1 = _mm_movehl_ps(vo1, vo1);
388           vo2 = _mm_movehl_ps(vo2, vo2);
389           vo3 = _mm_movehl_ps(vo3, vo3);
390         }
391         if (w & (1 * sizeof(float))) {
392           _mm_store_ss(o3, vo3);
393           o3 += 1;
394           _mm_store_ss(o2, vo2);
395           o2 += 1;
396           _mm_store_ss(o1, vo1);
397           o1 += 1;
398           _mm_store_ss(o0, vo0);
399           o0 += 1;
400         }
401       }
402     }
403 
404     i0 = (const float*) ((uintptr_t) i4 - input_decrement);
405     i1 = (const float*) ((uintptr_t) i5 - input_decrement);
406     i2 = (const float*) ((uintptr_t) i1 + input_width);
407     i3 = (const float*) ((uintptr_t) i2 + input_width);
408     i4 = (const float*) ((uintptr_t) i3 + input_width);
409     i5 = (const float*) ((uintptr_t) i4 + input_width);
410 
411     o0 = o3;
412     o1 = (float*) ((uintptr_t) o0 + input_width);
413     o2 = (float*) ((uintptr_t) o1 + input_width);
414     o3 = (float*) ((uintptr_t) o2 + input_width);
415 
416     output_height = doz(output_height, 4);
417   } while (output_height != 0);
418 }
419