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_5x4(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_5x4(
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   const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
73   const float* i7 = (const float*) ((uintptr_t) i6 + input_width);
74   const float* i8 = (const float*) ((uintptr_t) i7 + input_width);
75 
76   float* o0 = output;
77   float* o1 = (float*) ((uintptr_t) o0 + input_width);
78   float* o2 = (float*) ((uintptr_t) o1 + input_width);
79   float* o3 = (float*) ((uintptr_t) o2 + input_width);
80   float* o4 = (float*) ((uintptr_t) o3 + input_width);
81 
82   size_t output_height = input_height;
83   do {
84     if XNN_UNPREDICTABLE(output_height < 2) {
85       i3 = zero;
86       o1 = o0;
87     }
88     if XNN_UNPREDICTABLE(output_height < 3) {
89       i4 = zero;
90       o2 = o1;
91     }
92     if XNN_UNPREDICTABLE(output_height < 4) {
93       i5 = zero;
94       o3 = o2;
95     }
96     if XNN_UNPREDICTABLE(output_height < 5) {
97       i6 = zero;
98       o4 = o3;
99     }
100     if XNN_UNPREDICTABLE(output_height < 6) {
101       i7 = zero;
102     }
103     if XNN_UNPREDICTABLE(output_height < 7) {
104       i8 = zero;
105     }
106 
107     __m128 vi0x3012 = _mm_setzero_ps();
108     __m128 vi1x3012 = _mm_setzero_ps();
109     __m128 vi2x3012 = _mm_setzero_ps();
110     __m128 vi3x3012 = _mm_setzero_ps();
111     __m128 vi4x3012 = _mm_setzero_ps();
112     __m128 vi5x3012 = _mm_setzero_ps();
113     __m128 vi6x3012 = _mm_setzero_ps();
114     __m128 vi7x3012 = _mm_setzero_ps();
115     __m128 vi8x3012 = _mm_setzero_ps();
116 
117     __m128 vi0x4567 = _mm_loadu_ps(i0);
118     i0 += 4;
119     __m128 vi1x4567 = _mm_loadu_ps(i1);
120     i1 += 4;
121     __m128 vi2x4567 = _mm_loadu_ps(i2);
122     i2 += 4;
123     __m128 vi3x4567 = _mm_loadu_ps(i3);
124     i3 += 4;
125     __m128 vi4x4567 = _mm_loadu_ps(i4);
126     i4 += 4;
127     __m128 vi5x4567 = _mm_loadu_ps(i5);
128     i5 += 4;
129     __m128 vi6x4567 = _mm_loadu_ps(i6);
130     i6 += 4;
131     __m128 vi7x4567 = _mm_loadu_ps(i7);
132     i7 += 4;
133     __m128 vi8x4567 = _mm_loadu_ps(i8);
134     i8 += 4;
135 
136     size_t w = input_width;
137     for (; w > 8 * sizeof(float); w -= 4 * sizeof(float)) {
138       __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk02));
139       __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk02));
140       __m128 vo2p0 = _mm_add_ps(vbias, _mm_mul_ps(vi2x4567, vk02));
141       __m128 vo3p0 = _mm_add_ps(vbias, _mm_mul_ps(vi3x4567, vk02));
142       __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk02));
143       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk12));
144       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk12));
145       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk12));
146       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x4567, vk12));
147       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x4567, vk12));
148       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk22));
149       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk22));
150       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x4567, vk22));
151       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x4567, vk22));
152       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x4567, vk22));
153       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x4567, vk32));
154       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x4567, vk32));
155       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x4567, vk32));
156       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x4567, vk32));
157       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x4567, vk32));
158       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x4567, vk42));
159       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x4567, vk42));
160       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x4567, vk42));
161       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x4567, vk42));
162       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x4567, vk42));
163 
164       const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
165       const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
166       const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
167       const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3));
168       const __m128 vi4x7456 = _mm_shuffle_ps(vi4x4567, vi4x4567, _MM_SHUFFLE(2, 1, 0, 3));
169       const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3));
170       const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3));
171       const __m128 vi7x7456 = _mm_shuffle_ps(vi7x4567, vi7x4567, _MM_SHUFFLE(2, 1, 0, 3));
172       const __m128 vi8x7456 = _mm_shuffle_ps(vi8x4567, vi8x4567, _MM_SHUFFLE(2, 1, 0, 3));
173 
174       const __m128 vi0x89AB = _mm_loadu_ps(i0);
175       i0 += 4;
176       const __m128 vi1x89AB = _mm_loadu_ps(i1);
177       i1 += 4;
178       const __m128 vi2x89AB = _mm_loadu_ps(i2);
179       i2 += 4;
180       const __m128 vi3x89AB = _mm_loadu_ps(i3);
181       i3 += 4;
182       const __m128 vi4x89AB = _mm_loadu_ps(i4);
183       i4 += 4;
184       const __m128 vi5x89AB = _mm_loadu_ps(i5);
185       i5 += 4;
186       const __m128 vi6x89AB = _mm_loadu_ps(i6);
187       i6 += 4;
188       const __m128 vi7x89AB = _mm_loadu_ps(i7);
189       i7 += 4;
190       const __m128 vi8x89AB = _mm_loadu_ps(i8);
191       i8 += 4;
192 
193       const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
194       const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
195       const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
196       const __m128 vi3x3456 = _mm_move_ss(vi3x7456, vi3x3012);
197       const __m128 vi4x3456 = _mm_move_ss(vi4x7456, vi4x3012);
198       const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012);
199       const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012);
200       const __m128 vi7x3456 = _mm_move_ss(vi7x7456, vi7x3012);
201       const __m128 vi8x3456 = _mm_move_ss(vi8x7456, vi8x3012);
202 
203       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk01));
204       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk01));
205       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk01));
206       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x3456, vk01));
207       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x3456, vk01));
208       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk11));
209       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk11));
210       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x3456, vk11));
211       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x3456, vk11));
212       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk11));
213       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk21));
214       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x3456, vk21));
215       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x3456, vk21));
216       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk21));
217       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x3456, vk21));
218       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x3456, vk31));
219       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x3456, vk31));
220       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x3456, vk31));
221       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x3456, vk31));
222       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x3456, vk31));
223       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x3456, vk41));
224       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x3456, vk41));
225       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x3456, vk41));
226       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x3456, vk41));
227       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x3456, vk41));
228 
229       const __m128 vi0x2345 = _mm_shuffle_ps(vi0x3012, vi0x7456, _MM_SHUFFLE(2, 1, 0, 3));
230       vi0x3012 = vi0x7456;
231       const __m128 vi1x2345 = _mm_shuffle_ps(vi1x3012, vi1x7456, _MM_SHUFFLE(2, 1, 0, 3));
232       vi1x3012 = vi1x7456;
233       const __m128 vi2x2345 = _mm_shuffle_ps(vi2x3012, vi2x7456, _MM_SHUFFLE(2, 1, 0, 3));
234       vi2x3012 = vi2x7456;
235       const __m128 vi3x2345 = _mm_shuffle_ps(vi3x3012, vi3x7456, _MM_SHUFFLE(2, 1, 0, 3));
236       vi3x3012 = vi3x7456;
237       const __m128 vi4x2345 = _mm_shuffle_ps(vi4x3012, vi4x7456, _MM_SHUFFLE(2, 1, 0, 3));
238       vi4x3012 = vi4x7456;
239       const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3));
240       vi5x3012 = vi5x7456;
241       const __m128 vi6x2345 = _mm_shuffle_ps(vi6x3012, vi6x7456, _MM_SHUFFLE(2, 1, 0, 3));
242       vi6x3012 = vi6x7456;
243       const __m128 vi7x2345 = _mm_shuffle_ps(vi7x3012, vi7x7456, _MM_SHUFFLE(2, 1, 0, 3));
244       vi7x3012 = vi7x7456;
245       const __m128 vi8x2345 = _mm_shuffle_ps(vi8x3012, vi8x7456, _MM_SHUFFLE(2, 1, 0, 3));
246       vi8x3012 = vi8x7456;
247 
248       const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB);
249       vi0x4567 = vi0x89AB;
250       const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB);
251       vi1x4567 = vi1x89AB;
252       const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB);
253       vi2x4567 = vi2x89AB;
254       const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB);
255       vi3x4567 = vi3x89AB;
256       const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB);
257       vi4x4567 = vi4x89AB;
258       const __m128 vi5x8567 = _mm_move_ss(vi5x4567, vi5x89AB);
259       vi5x4567 = vi5x89AB;
260       const __m128 vi6x8567 = _mm_move_ss(vi6x4567, vi6x89AB);
261       vi6x4567 = vi6x89AB;
262       const __m128 vi7x8567 = _mm_move_ss(vi7x4567, vi7x89AB);
263       vi7x4567 = vi7x89AB;
264       const __m128 vi8x8567 = _mm_move_ss(vi8x4567, vi8x89AB);
265       vi8x4567 = vi8x89AB;
266 
267       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x2345, vk00));
268       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x2345, vk00));
269       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x2345, vk00));
270       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x2345, vk00));
271       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x2345, vk00));
272       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x2345, vk10));
273       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x2345, vk10));
274       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x2345, vk10));
275       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x2345, vk10));
276       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x2345, vk10));
277       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x2345, vk20));
278       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x2345, vk20));
279       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x2345, vk20));
280       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x2345, vk20));
281       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x2345, vk20));
282       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x2345, vk30));
283       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x2345, vk30));
284       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x2345, vk30));
285       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x2345, vk30));
286       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x2345, vk30));
287       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x2345, vk40));
288       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x2345, vk40));
289       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x2345, vk40));
290       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x2345, vk40));
291       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x2345, vk40));
292 
293       const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
294       const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
295       const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
296       const __m128 vi3x5678 = _mm_shuffle_ps(vi3x8567, vi3x8567, _MM_SHUFFLE(0, 3, 2, 1));
297       const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));
298       const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));
299       const __m128 vi6x5678 = _mm_shuffle_ps(vi6x8567, vi6x8567, _MM_SHUFFLE(0, 3, 2, 1));
300       const __m128 vi7x5678 = _mm_shuffle_ps(vi7x8567, vi7x8567, _MM_SHUFFLE(0, 3, 2, 1));
301       const __m128 vi8x5678 = _mm_shuffle_ps(vi8x8567, vi8x8567, _MM_SHUFFLE(0, 3, 2, 1));
302 
303       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk03));
304       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk03));
305       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x5678, vk03));
306       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x5678, vk03));
307       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x5678, vk03));
308       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk13));
309       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x5678, vk13));
310       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x5678, vk13));
311       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x5678, vk13));
312       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x5678, vk13));
313       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk23));
314       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk23));
315       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk23));
316       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk23));
317       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x5678, vk23));
318       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x5678, vk33));
319       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x5678, vk33));
320       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33));
321       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x5678, vk33));
322       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x5678, vk33));
323       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43));
324       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43));
325       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x5678, vk43));
326       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x5678, vk43));
327       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x5678, vk43));
328 
329       const __m128 vi0x6789 = _mm_shuffle_ps(vi0x5678, vi0x89AB, _MM_SHUFFLE(1, 0, 2, 1));
330       const __m128 vi1x6789 = _mm_shuffle_ps(vi1x5678, vi1x89AB, _MM_SHUFFLE(1, 0, 2, 1));
331       const __m128 vi2x6789 = _mm_shuffle_ps(vi2x5678, vi2x89AB, _MM_SHUFFLE(1, 0, 2, 1));
332       const __m128 vi3x6789 = _mm_shuffle_ps(vi3x5678, vi3x89AB, _MM_SHUFFLE(1, 0, 2, 1));
333       const __m128 vi4x6789 = _mm_shuffle_ps(vi4x5678, vi4x89AB, _MM_SHUFFLE(1, 0, 2, 1));
334       const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1));
335       const __m128 vi6x6789 = _mm_shuffle_ps(vi6x5678, vi6x89AB, _MM_SHUFFLE(1, 0, 2, 1));
336       const __m128 vi7x6789 = _mm_shuffle_ps(vi7x5678, vi7x89AB, _MM_SHUFFLE(1, 0, 2, 1));
337       const __m128 vi8x6789 = _mm_shuffle_ps(vi8x5678, vi8x89AB, _MM_SHUFFLE(1, 0, 2, 1));
338 
339       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x6789, vk04));
340       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x6789, vk04));
341       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x6789, vk04));
342       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x6789, vk04));
343       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x6789, vk04));
344       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x6789, vk14));
345       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x6789, vk14));
346       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x6789, vk14));
347       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x6789, vk14));
348       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x6789, vk14));
349       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x6789, vk24));
350       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x6789, vk24));
351       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x6789, vk24));
352       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24));
353       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x6789, vk24));
354       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x6789, vk34));
355       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x6789, vk34));
356       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x6789, vk34));
357       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x6789, vk34));
358       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x6789, vk34));
359       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x6789, vk44));
360       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44));
361       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x6789, vk44));
362       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x6789, vk44));
363       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x6789, vk44));
364 
365 
366       __m128 vo0 = _mm_max_ps(vo0p0, vmin);
367       __m128 vo1 = _mm_max_ps(vo1p0, vmin);
368       __m128 vo2 = _mm_max_ps(vo2p0, vmin);
369       __m128 vo3 = _mm_max_ps(vo3p0, vmin);
370       __m128 vo4 = _mm_max_ps(vo4p0, vmin);
371 
372       vo0 = _mm_min_ps(vo0, vmax);
373       vo1 = _mm_min_ps(vo1, vmax);
374       vo2 = _mm_min_ps(vo2, vmax);
375       vo3 = _mm_min_ps(vo3, vmax);
376       vo4 = _mm_min_ps(vo4, vmax);
377 
378       _mm_storeu_ps(o4, vo4);
379       o4 += 4;
380       _mm_storeu_ps(o3, vo3);
381       o3 += 4;
382       _mm_storeu_ps(o2, vo2);
383       o2 += 4;
384       _mm_storeu_ps(o1, vo1);
385       o1 += 4;
386       _mm_storeu_ps(o0, vo0);
387       o0 += 4;
388     }
389     // Always process the last block of 5..8 pixels.
390     if XNN_LIKELY(w > 4 * sizeof(float)) {
391       __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk02));
392       __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk02));
393       __m128 vo2p0 = _mm_add_ps(vbias, _mm_mul_ps(vi2x4567, vk02));
394       __m128 vo3p0 = _mm_add_ps(vbias, _mm_mul_ps(vi3x4567, vk02));
395       __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk02));
396       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk12));
397       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk12));
398       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk12));
399       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x4567, vk12));
400       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x4567, vk12));
401       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk22));
402       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk22));
403       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x4567, vk22));
404       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x4567, vk22));
405       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x4567, vk22));
406       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x4567, vk32));
407       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x4567, vk32));
408       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x4567, vk32));
409       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x4567, vk32));
410       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x4567, vk32));
411       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x4567, vk42));
412       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x4567, vk42));
413       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x4567, vk42));
414       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x4567, vk42));
415       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x4567, vk42));
416 
417       const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
418       const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
419       const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
420       const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3));
421       const __m128 vi4x7456 = _mm_shuffle_ps(vi4x4567, vi4x4567, _MM_SHUFFLE(2, 1, 0, 3));
422       const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3));
423       const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3));
424       const __m128 vi7x7456 = _mm_shuffle_ps(vi7x4567, vi7x4567, _MM_SHUFFLE(2, 1, 0, 3));
425       const __m128 vi8x7456 = _mm_shuffle_ps(vi8x4567, vi8x4567, _MM_SHUFFLE(2, 1, 0, 3));
426 
427       const __m128 vi0x89AB = _mm_and_ps(_mm_loadu_ps(i0), vmask);
428       i0 += 4;
429       const __m128 vi1x89AB = _mm_and_ps(_mm_loadu_ps(i1), vmask);
430       i1 += 4;
431       const __m128 vi2x89AB = _mm_and_ps(_mm_loadu_ps(i2), vmask);
432       i2 += 4;
433       const __m128 vi3x89AB = _mm_and_ps(_mm_loadu_ps(i3), vmask);
434       i3 += 4;
435       const __m128 vi4x89AB = _mm_and_ps(_mm_loadu_ps(i4), vmask);
436       i4 += 4;
437       const __m128 vi5x89AB = _mm_and_ps(_mm_loadu_ps(i5), vmask);
438       i5 += 4;
439       const __m128 vi6x89AB = _mm_and_ps(_mm_loadu_ps(i6), vmask);
440       i6 += 4;
441       const __m128 vi7x89AB = _mm_and_ps(_mm_loadu_ps(i7), vmask);
442       i7 += 4;
443       const __m128 vi8x89AB = _mm_and_ps(_mm_loadu_ps(i8), vmask);
444       i8 += 4;
445 
446       const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
447       const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
448       const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
449       const __m128 vi3x3456 = _mm_move_ss(vi3x7456, vi3x3012);
450       const __m128 vi4x3456 = _mm_move_ss(vi4x7456, vi4x3012);
451       const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012);
452       const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012);
453       const __m128 vi7x3456 = _mm_move_ss(vi7x7456, vi7x3012);
454       const __m128 vi8x3456 = _mm_move_ss(vi8x7456, vi8x3012);
455 
456       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk01));
457       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk01));
458       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk01));
459       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x3456, vk01));
460       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x3456, vk01));
461       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk11));
462       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk11));
463       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x3456, vk11));
464       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x3456, vk11));
465       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk11));
466       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk21));
467       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x3456, vk21));
468       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x3456, vk21));
469       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk21));
470       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x3456, vk21));
471       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x3456, vk31));
472       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x3456, vk31));
473       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x3456, vk31));
474       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x3456, vk31));
475       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x3456, vk31));
476       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x3456, vk41));
477       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x3456, vk41));
478       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x3456, vk41));
479       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x3456, vk41));
480       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x3456, vk41));
481 
482       const __m128 vi0x2345 = _mm_shuffle_ps(vi0x3012, vi0x7456, _MM_SHUFFLE(2, 1, 0, 3));
483       vi0x3012 = vi0x7456;
484       const __m128 vi1x2345 = _mm_shuffle_ps(vi1x3012, vi1x7456, _MM_SHUFFLE(2, 1, 0, 3));
485       vi1x3012 = vi1x7456;
486       const __m128 vi2x2345 = _mm_shuffle_ps(vi2x3012, vi2x7456, _MM_SHUFFLE(2, 1, 0, 3));
487       vi2x3012 = vi2x7456;
488       const __m128 vi3x2345 = _mm_shuffle_ps(vi3x3012, vi3x7456, _MM_SHUFFLE(2, 1, 0, 3));
489       vi3x3012 = vi3x7456;
490       const __m128 vi4x2345 = _mm_shuffle_ps(vi4x3012, vi4x7456, _MM_SHUFFLE(2, 1, 0, 3));
491       vi4x3012 = vi4x7456;
492       const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3));
493       vi5x3012 = vi5x7456;
494       const __m128 vi6x2345 = _mm_shuffle_ps(vi6x3012, vi6x7456, _MM_SHUFFLE(2, 1, 0, 3));
495       vi6x3012 = vi6x7456;
496       const __m128 vi7x2345 = _mm_shuffle_ps(vi7x3012, vi7x7456, _MM_SHUFFLE(2, 1, 0, 3));
497       vi7x3012 = vi7x7456;
498       const __m128 vi8x2345 = _mm_shuffle_ps(vi8x3012, vi8x7456, _MM_SHUFFLE(2, 1, 0, 3));
499       vi8x3012 = vi8x7456;
500 
501       const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB);
502       vi0x4567 = vi0x89AB;
503       const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB);
504       vi1x4567 = vi1x89AB;
505       const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB);
506       vi2x4567 = vi2x89AB;
507       const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB);
508       vi3x4567 = vi3x89AB;
509       const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB);
510       vi4x4567 = vi4x89AB;
511       const __m128 vi5x8567 = _mm_move_ss(vi5x4567, vi5x89AB);
512       vi5x4567 = vi5x89AB;
513       const __m128 vi6x8567 = _mm_move_ss(vi6x4567, vi6x89AB);
514       vi6x4567 = vi6x89AB;
515       const __m128 vi7x8567 = _mm_move_ss(vi7x4567, vi7x89AB);
516       vi7x4567 = vi7x89AB;
517       const __m128 vi8x8567 = _mm_move_ss(vi8x4567, vi8x89AB);
518       vi8x4567 = vi8x89AB;
519 
520       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x2345, vk00));
521       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x2345, vk00));
522       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x2345, vk00));
523       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x2345, vk00));
524       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x2345, vk00));
525       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x2345, vk10));
526       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x2345, vk10));
527       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x2345, vk10));
528       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x2345, vk10));
529       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x2345, vk10));
530       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x2345, vk20));
531       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x2345, vk20));
532       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x2345, vk20));
533       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x2345, vk20));
534       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x2345, vk20));
535       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x2345, vk30));
536       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x2345, vk30));
537       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x2345, vk30));
538       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x2345, vk30));
539       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x2345, vk30));
540       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x2345, vk40));
541       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x2345, vk40));
542       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x2345, vk40));
543       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x2345, vk40));
544       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x2345, vk40));
545 
546       const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
547       const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
548       const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
549       const __m128 vi3x5678 = _mm_shuffle_ps(vi3x8567, vi3x8567, _MM_SHUFFLE(0, 3, 2, 1));
550       const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));
551       const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));
552       const __m128 vi6x5678 = _mm_shuffle_ps(vi6x8567, vi6x8567, _MM_SHUFFLE(0, 3, 2, 1));
553       const __m128 vi7x5678 = _mm_shuffle_ps(vi7x8567, vi7x8567, _MM_SHUFFLE(0, 3, 2, 1));
554       const __m128 vi8x5678 = _mm_shuffle_ps(vi8x8567, vi8x8567, _MM_SHUFFLE(0, 3, 2, 1));
555 
556       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk03));
557       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk03));
558       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x5678, vk03));
559       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x5678, vk03));
560       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x5678, vk03));
561       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk13));
562       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x5678, vk13));
563       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x5678, vk13));
564       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x5678, vk13));
565       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x5678, vk13));
566       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk23));
567       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk23));
568       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk23));
569       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk23));
570       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x5678, vk23));
571       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x5678, vk33));
572       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x5678, vk33));
573       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33));
574       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x5678, vk33));
575       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x5678, vk33));
576       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43));
577       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43));
578       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x5678, vk43));
579       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x5678, vk43));
580       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x5678, vk43));
581 
582       const __m128 vi0x6789 = _mm_shuffle_ps(vi0x5678, vi0x89AB, _MM_SHUFFLE(1, 0, 2, 1));
583       const __m128 vi1x6789 = _mm_shuffle_ps(vi1x5678, vi1x89AB, _MM_SHUFFLE(1, 0, 2, 1));
584       const __m128 vi2x6789 = _mm_shuffle_ps(vi2x5678, vi2x89AB, _MM_SHUFFLE(1, 0, 2, 1));
585       const __m128 vi3x6789 = _mm_shuffle_ps(vi3x5678, vi3x89AB, _MM_SHUFFLE(1, 0, 2, 1));
586       const __m128 vi4x6789 = _mm_shuffle_ps(vi4x5678, vi4x89AB, _MM_SHUFFLE(1, 0, 2, 1));
587       const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1));
588       const __m128 vi6x6789 = _mm_shuffle_ps(vi6x5678, vi6x89AB, _MM_SHUFFLE(1, 0, 2, 1));
589       const __m128 vi7x6789 = _mm_shuffle_ps(vi7x5678, vi7x89AB, _MM_SHUFFLE(1, 0, 2, 1));
590       const __m128 vi8x6789 = _mm_shuffle_ps(vi8x5678, vi8x89AB, _MM_SHUFFLE(1, 0, 2, 1));
591 
592       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x6789, vk04));
593       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x6789, vk04));
594       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x6789, vk04));
595       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x6789, vk04));
596       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x6789, vk04));
597       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x6789, vk14));
598       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x6789, vk14));
599       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x6789, vk14));
600       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x6789, vk14));
601       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x6789, vk14));
602       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x6789, vk24));
603       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x6789, vk24));
604       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x6789, vk24));
605       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24));
606       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x6789, vk24));
607       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x6789, vk34));
608       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x6789, vk34));
609       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x6789, vk34));
610       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x6789, vk34));
611       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x6789, vk34));
612       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x6789, vk44));
613       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44));
614       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x6789, vk44));
615       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x6789, vk44));
616       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x6789, vk44));
617 
618 
619       __m128 vo0 = _mm_max_ps(vo0p0, vmin);
620       __m128 vo1 = _mm_max_ps(vo1p0, vmin);
621       __m128 vo2 = _mm_max_ps(vo2p0, vmin);
622       __m128 vo3 = _mm_max_ps(vo3p0, vmin);
623       __m128 vo4 = _mm_max_ps(vo4p0, vmin);
624 
625       vo0 = _mm_min_ps(vo0, vmax);
626       vo1 = _mm_min_ps(vo1, vmax);
627       vo2 = _mm_min_ps(vo2, vmax);
628       vo3 = _mm_min_ps(vo3, vmax);
629       vo4 = _mm_min_ps(vo4, vmax);
630 
631       _mm_storeu_ps(o4, vo4);
632       o4 += 4;
633       _mm_storeu_ps(o3, vo3);
634       o3 += 4;
635       _mm_storeu_ps(o2, vo2);
636       o2 += 4;
637       _mm_storeu_ps(o1, vo1);
638       o1 += 4;
639       _mm_storeu_ps(o0, vo0);
640       o0 += 4;
641 
642       w -= 4 * sizeof(float);
643     }
644     assert(w >= 1 * sizeof(float));
645     assert(w <= 4 * sizeof(float));
646     {
647       vi0x4567 = _mm_and_ps(vi0x4567, vmask);
648       vi1x4567 = _mm_and_ps(vi1x4567, vmask);
649       vi2x4567 = _mm_and_ps(vi2x4567, vmask);
650       vi3x4567 = _mm_and_ps(vi3x4567, vmask);
651       vi4x4567 = _mm_and_ps(vi4x4567, vmask);
652       vi5x4567 = _mm_and_ps(vi5x4567, vmask);
653       vi6x4567 = _mm_and_ps(vi6x4567, vmask);
654       vi7x4567 = _mm_and_ps(vi7x4567, vmask);
655       vi8x4567 = _mm_and_ps(vi8x4567, vmask);
656 
657       __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk02));
658       __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk02));
659       __m128 vo2p0 = _mm_add_ps(vbias, _mm_mul_ps(vi2x4567, vk02));
660       __m128 vo3p0 = _mm_add_ps(vbias, _mm_mul_ps(vi3x4567, vk02));
661       __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk02));
662       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk12));
663       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk12));
664       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk12));
665       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x4567, vk12));
666       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x4567, vk12));
667       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk22));
668       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk22));
669       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x4567, vk22));
670       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x4567, vk22));
671       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x4567, vk22));
672       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x4567, vk32));
673       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x4567, vk32));
674       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x4567, vk32));
675       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x4567, vk32));
676       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x4567, vk32));
677       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x4567, vk42));
678       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x4567, vk42));
679       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x4567, vk42));
680       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x4567, vk42));
681       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x4567, vk42));
682 
683       const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
684       const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
685       const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
686       const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3));
687       const __m128 vi4x7456 = _mm_shuffle_ps(vi4x4567, vi4x4567, _MM_SHUFFLE(2, 1, 0, 3));
688       const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3));
689       const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3));
690       const __m128 vi7x7456 = _mm_shuffle_ps(vi7x4567, vi7x4567, _MM_SHUFFLE(2, 1, 0, 3));
691       const __m128 vi8x7456 = _mm_shuffle_ps(vi8x4567, vi8x4567, _MM_SHUFFLE(2, 1, 0, 3));
692 
693       const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
694       const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
695       const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
696       const __m128 vi3x3456 = _mm_move_ss(vi3x7456, vi3x3012);
697       const __m128 vi4x3456 = _mm_move_ss(vi4x7456, vi4x3012);
698       const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012);
699       const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012);
700       const __m128 vi7x3456 = _mm_move_ss(vi7x7456, vi7x3012);
701       const __m128 vi8x3456 = _mm_move_ss(vi8x7456, vi8x3012);
702 
703       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk01));
704       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk01));
705       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk01));
706       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x3456, vk01));
707       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x3456, vk01));
708       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk11));
709       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk11));
710       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x3456, vk11));
711       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x3456, vk11));
712       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk11));
713       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk21));
714       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x3456, vk21));
715       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x3456, vk21));
716       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk21));
717       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x3456, vk21));
718       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x3456, vk31));
719       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x3456, vk31));
720       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x3456, vk31));
721       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x3456, vk31));
722       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x3456, vk31));
723       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x3456, vk41));
724       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x3456, vk41));
725       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x3456, vk41));
726       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x3456, vk41));
727       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x3456, vk41));
728 
729       const __m128 vi0x2345 = _mm_shuffle_ps(vi0x3012, vi0x7456, _MM_SHUFFLE(2, 1, 0, 3));
730       const __m128 vi1x2345 = _mm_shuffle_ps(vi1x3012, vi1x7456, _MM_SHUFFLE(2, 1, 0, 3));
731       const __m128 vi2x2345 = _mm_shuffle_ps(vi2x3012, vi2x7456, _MM_SHUFFLE(2, 1, 0, 3));
732       const __m128 vi3x2345 = _mm_shuffle_ps(vi3x3012, vi3x7456, _MM_SHUFFLE(2, 1, 0, 3));
733       const __m128 vi4x2345 = _mm_shuffle_ps(vi4x3012, vi4x7456, _MM_SHUFFLE(2, 1, 0, 3));
734       const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3));
735       const __m128 vi6x2345 = _mm_shuffle_ps(vi6x3012, vi6x7456, _MM_SHUFFLE(2, 1, 0, 3));
736       const __m128 vi7x2345 = _mm_shuffle_ps(vi7x3012, vi7x7456, _MM_SHUFFLE(2, 1, 0, 3));
737       const __m128 vi8x2345 = _mm_shuffle_ps(vi8x3012, vi8x7456, _MM_SHUFFLE(2, 1, 0, 3));
738 
739       const __m128 vzero = _mm_setzero_ps();
740       const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero);
741       const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero);
742       const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero);
743       const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero);
744       const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero);
745       const __m128 vi5x8567 = _mm_move_ss(vi5x4567, vzero);
746       const __m128 vi6x8567 = _mm_move_ss(vi6x4567, vzero);
747       const __m128 vi7x8567 = _mm_move_ss(vi7x4567, vzero);
748       const __m128 vi8x8567 = _mm_move_ss(vi8x4567, vzero);
749 
750       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x2345, vk00));
751       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x2345, vk00));
752       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x2345, vk00));
753       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x2345, vk00));
754       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x2345, vk00));
755       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x2345, vk10));
756       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x2345, vk10));
757       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x2345, vk10));
758       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x2345, vk10));
759       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x2345, vk10));
760       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x2345, vk20));
761       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x2345, vk20));
762       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x2345, vk20));
763       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x2345, vk20));
764       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x2345, vk20));
765       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x2345, vk30));
766       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x2345, vk30));
767       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x2345, vk30));
768       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x2345, vk30));
769       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x2345, vk30));
770       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x2345, vk40));
771       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x2345, vk40));
772       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x2345, vk40));
773       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x2345, vk40));
774       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x2345, vk40));
775 
776       const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
777       const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
778       const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
779       const __m128 vi3x5678 = _mm_shuffle_ps(vi3x8567, vi3x8567, _MM_SHUFFLE(0, 3, 2, 1));
780       const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));
781       const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));
782       const __m128 vi6x5678 = _mm_shuffle_ps(vi6x8567, vi6x8567, _MM_SHUFFLE(0, 3, 2, 1));
783       const __m128 vi7x5678 = _mm_shuffle_ps(vi7x8567, vi7x8567, _MM_SHUFFLE(0, 3, 2, 1));
784       const __m128 vi8x5678 = _mm_shuffle_ps(vi8x8567, vi8x8567, _MM_SHUFFLE(0, 3, 2, 1));
785 
786       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk03));
787       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk03));
788       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x5678, vk03));
789       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x5678, vk03));
790       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x5678, vk03));
791       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk13));
792       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x5678, vk13));
793       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x5678, vk13));
794       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x5678, vk13));
795       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x5678, vk13));
796       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk23));
797       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk23));
798       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk23));
799       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk23));
800       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x5678, vk23));
801       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x5678, vk33));
802       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x5678, vk33));
803       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33));
804       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x5678, vk33));
805       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x5678, vk33));
806       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43));
807       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43));
808       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x5678, vk43));
809       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x5678, vk43));
810       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x5678, vk43));
811 
812       const __m128 vi0x6789 = _mm_shuffle_ps(vi0x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
813       const __m128 vi1x6789 = _mm_shuffle_ps(vi1x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
814       const __m128 vi2x6789 = _mm_shuffle_ps(vi2x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
815       const __m128 vi3x6789 = _mm_shuffle_ps(vi3x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
816       const __m128 vi4x6789 = _mm_shuffle_ps(vi4x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
817       const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
818       const __m128 vi6x6789 = _mm_shuffle_ps(vi6x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
819       const __m128 vi7x6789 = _mm_shuffle_ps(vi7x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
820       const __m128 vi8x6789 = _mm_shuffle_ps(vi8x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
821 
822       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x6789, vk04));
823       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x6789, vk04));
824       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x6789, vk04));
825       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x6789, vk04));
826       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x6789, vk04));
827       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x6789, vk14));
828       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x6789, vk14));
829       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x6789, vk14));
830       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x6789, vk14));
831       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x6789, vk14));
832       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x6789, vk24));
833       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x6789, vk24));
834       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x6789, vk24));
835       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24));
836       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x6789, vk24));
837       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x6789, vk34));
838       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x6789, vk34));
839       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x6789, vk34));
840       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x6789, vk34));
841       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi7x6789, vk34));
842       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x6789, vk44));
843       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44));
844       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x6789, vk44));
845       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x6789, vk44));
846       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x6789, vk44));
847 
848 
849       __m128 vo0 = _mm_max_ps(vo0p0, vmin);
850       __m128 vo1 = _mm_max_ps(vo1p0, vmin);
851       __m128 vo2 = _mm_max_ps(vo2p0, vmin);
852       __m128 vo3 = _mm_max_ps(vo3p0, vmin);
853       __m128 vo4 = _mm_max_ps(vo4p0, vmin);
854 
855       vo0 = _mm_min_ps(vo0, vmax);
856       vo1 = _mm_min_ps(vo1, vmax);
857       vo2 = _mm_min_ps(vo2, vmax);
858       vo3 = _mm_min_ps(vo3, vmax);
859       vo4 = _mm_min_ps(vo4, vmax);
860 
861       if XNN_LIKELY(w & (4 * sizeof(float))) {
862         _mm_storeu_ps(o4, vo4);
863         o4 += 4;
864         _mm_storeu_ps(o3, vo3);
865         o3 += 4;
866         _mm_storeu_ps(o2, vo2);
867         o2 += 4;
868         _mm_storeu_ps(o1, vo1);
869         o1 += 4;
870         _mm_storeu_ps(o0, vo0);
871         o0 += 4;
872       } else {
873         if (w & (2 * sizeof(float))) {
874           _mm_storel_pi((__m64*) o4, vo4);
875           o4 += 2;
876           _mm_storel_pi((__m64*) o3, vo3);
877           o3 += 2;
878           _mm_storel_pi((__m64*) o2, vo2);
879           o2 += 2;
880           _mm_storel_pi((__m64*) o1, vo1);
881           o1 += 2;
882           _mm_storel_pi((__m64*) o0, vo0);
883           o0 += 2;
884 
885           vo0 = _mm_movehl_ps(vo0, vo0);
886           vo1 = _mm_movehl_ps(vo1, vo1);
887           vo2 = _mm_movehl_ps(vo2, vo2);
888           vo3 = _mm_movehl_ps(vo3, vo3);
889           vo4 = _mm_movehl_ps(vo4, vo4);
890         }
891         if (w & (1 * sizeof(float))) {
892           _mm_store_ss(o4, vo4);
893           o4 += 1;
894           _mm_store_ss(o3, vo3);
895           o3 += 1;
896           _mm_store_ss(o2, vo2);
897           o2 += 1;
898           _mm_store_ss(o1, vo1);
899           o1 += 1;
900           _mm_store_ss(o0, vo0);
901           o0 += 1;
902         }
903       }
904     }
905 
906     i0 = (const float*) ((uintptr_t) i5 - input_decrement);
907     i1 = (const float*) ((uintptr_t) i6 - input_decrement);
908     i2 = (const float*) ((uintptr_t) i1 + input_width);
909     i3 = (const float*) ((uintptr_t) i2 + input_width);
910     i4 = (const float*) ((uintptr_t) i3 + input_width);
911     i5 = (const float*) ((uintptr_t) i4 + input_width);
912     i6 = (const float*) ((uintptr_t) i5 + input_width);
913     i7 = (const float*) ((uintptr_t) i6 + input_width);
914     i8 = (const float*) ((uintptr_t) i7 + input_width);
915 
916     o0 = o4;
917     o1 = (float*) ((uintptr_t) o0 + input_width);
918     o2 = (float*) ((uintptr_t) o1 + input_width);
919     o3 = (float*) ((uintptr_t) o2 + input_width);
920     o4 = (float*) ((uintptr_t) o3 + input_width);
921 
922     output_height = doz(output_height, 5);
923   } while (output_height != 0);
924 }
925