1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv/up-sse.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2019 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 
16 
xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,size_t input_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2(
18     size_t channels,
19     size_t output_width,
20     const float** input,
21     const float* weights,
22     float* output,
23     size_t input_stride,
24     size_t output_increment,
25     size_t input_offset,
26     const float* zero,
27     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
28 {
29   assert(channels != 0);
30   assert(output_width != 0);
31 
32   const __m128 vmax = _mm_load_ps(params->sse.max);
33   const __m128 vmin = _mm_load_ps(params->sse.min);
34   do {
35     const float* i0 = input[0];
36     assert(i0 != NULL);
37     if XNN_UNPREDICTABLE(i0 != zero) {
38       i0 = (const float*) ((uintptr_t) i0 + input_offset);
39     }
40     const float* i1 = input[1];
41     assert(i1 != NULL);
42     if XNN_UNPREDICTABLE(i1 != zero) {
43       i1 = (const float*) ((uintptr_t) i1 + input_offset);
44     }
45     const float* i2 = input[2];
46     assert(i2 != NULL);
47     if XNN_UNPREDICTABLE(i2 != zero) {
48       i2 = (const float*) ((uintptr_t) i2 + input_offset);
49     }
50     const float* i3 = input[3];
51     assert(i3 != NULL);
52     if XNN_UNPREDICTABLE(i3 != zero) {
53       i3 = (const float*) ((uintptr_t) i3 + input_offset);
54     }
55     const float* i4 = input[4];
56     assert(i4 != NULL);
57     if XNN_UNPREDICTABLE(i4 != zero) {
58       i4 = (const float*) ((uintptr_t) i4 + input_offset);
59     }
60     const float* i5 = input[5];
61     assert(i5 != NULL);
62     if XNN_UNPREDICTABLE(i5 != zero) {
63       i5 = (const float*) ((uintptr_t) i5 + input_offset);
64     }
65     const float* i6 = input[6];
66     assert(i6 != NULL);
67     if XNN_UNPREDICTABLE(i6 != zero) {
68       i6 = (const float*) ((uintptr_t) i6 + input_offset);
69     }
70     const float* i7 = input[7];
71     assert(i7 != NULL);
72     if XNN_UNPREDICTABLE(i7 != zero) {
73       i7 = (const float*) ((uintptr_t) i7 + input_offset);
74     }
75     const float* i8 = input[8];
76     assert(i8 != NULL);
77     if XNN_UNPREDICTABLE(i8 != zero) {
78       i8 = (const float*) ((uintptr_t) i8 + input_offset);
79     }
80     const float* i9 = input[9];
81     assert(i9 != NULL);
82     if XNN_UNPREDICTABLE(i9 != zero) {
83       i9 = (const float*) ((uintptr_t) i9 + input_offset);
84     }
85     const float* i10 = input[10];
86     assert(i10 != NULL);
87     if XNN_UNPREDICTABLE(i10 != zero) {
88       i10 = (const float*) ((uintptr_t) i10 + input_offset);
89     }
90     const float* i11 = input[11];
91     assert(i11 != NULL);
92     if XNN_UNPREDICTABLE(i11 != zero) {
93       i11 = (const float*) ((uintptr_t) i11 + input_offset);
94     }
95     const float* i12 = input[12];
96     assert(i12 != NULL);
97     if XNN_UNPREDICTABLE(i12 != zero) {
98       i12 = (const float*) ((uintptr_t) i12 + input_offset);
99     }
100     const float* i13 = input[13];
101     assert(i13 != NULL);
102     if XNN_UNPREDICTABLE(i13 != zero) {
103       i13 = (const float*) ((uintptr_t) i13 + input_offset);
104     }
105     const float* i14 = input[14];
106     assert(i14 != NULL);
107     if XNN_UNPREDICTABLE(i14 != zero) {
108       i14 = (const float*) ((uintptr_t) i14 + input_offset);
109     }
110     const float* i15 = input[15];
111     assert(i15 != NULL);
112     if XNN_UNPREDICTABLE(i15 != zero) {
113       i15 = (const float*) ((uintptr_t) i15 + input_offset);
114     }
115     const float* i16 = input[16];
116     assert(i16 != NULL);
117     if XNN_UNPREDICTABLE(i16 != zero) {
118       i16 = (const float*) ((uintptr_t) i16 + input_offset);
119     }
120     const float* i17 = input[17];
121     assert(i17 != NULL);
122     if XNN_UNPREDICTABLE(i17 != zero) {
123       i17 = (const float*) ((uintptr_t) i17 + input_offset);
124     }
125     const float* i18 = input[18];
126     assert(i18 != NULL);
127     if XNN_UNPREDICTABLE(i18 != zero) {
128       i18 = (const float*) ((uintptr_t) i18 + input_offset);
129     }
130     const float* i19 = input[19];
131     assert(i19 != NULL);
132     if XNN_UNPREDICTABLE(i19 != zero) {
133       i19 = (const float*) ((uintptr_t) i19 + input_offset);
134     }
135     const float* i20 = input[20];
136     assert(i20 != NULL);
137     if XNN_UNPREDICTABLE(i20 != zero) {
138       i20 = (const float*) ((uintptr_t) i20 + input_offset);
139     }
140     const float* i21 = input[21];
141     assert(i21 != NULL);
142     if XNN_UNPREDICTABLE(i21 != zero) {
143       i21 = (const float*) ((uintptr_t) i21 + input_offset);
144     }
145     const float* i22 = input[22];
146     assert(i22 != NULL);
147     if XNN_UNPREDICTABLE(i22 != zero) {
148       i22 = (const float*) ((uintptr_t) i22 + input_offset);
149     }
150     const float* i23 = input[23];
151     assert(i23 != NULL);
152     if XNN_UNPREDICTABLE(i23 != zero) {
153       i23 = (const float*) ((uintptr_t) i23 + input_offset);
154     }
155     const float* i24 = input[24];
156     assert(i24 != NULL);
157     if XNN_UNPREDICTABLE(i24 != zero) {
158       i24 = (const float*) ((uintptr_t) i24 + input_offset);
159     }
160     input = (const float**) ((uintptr_t) input + input_stride);
161 
162     size_t c = channels;
163     const float* w = weights;
164     for (; c >= 8; c -= 8) {
165       __m128 vacc0123p0 = _mm_load_ps(w);
166       __m128 vacc4567p0 = _mm_load_ps(w + 4);
167 
168 
169       const __m128 vi0x0123 = _mm_loadu_ps(i0);
170       const __m128 vi0x4567 = _mm_loadu_ps(i0 + 4);
171       i0 += 8;
172 
173       const __m128 vk0x0123 = _mm_load_ps(w + 8);
174       const __m128 vk0x4567 = _mm_load_ps(w + 12);
175       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
176       vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi0x4567, vk0x4567));
177 
178       const __m128 vi1x0123 = _mm_loadu_ps(i1);
179       const __m128 vi1x4567 = _mm_loadu_ps(i1 + 4);
180       i1 += 8;
181 
182       const __m128 vk1x0123 = _mm_load_ps(w + 16);
183       const __m128 vk1x4567 = _mm_load_ps(w + 20);
184       __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
185       __m128 vacc4567p1 = _mm_mul_ps(vi1x4567, vk1x4567);
186 
187       const __m128 vi2x0123 = _mm_loadu_ps(i2);
188       const __m128 vi2x4567 = _mm_loadu_ps(i2 + 4);
189       i2 += 8;
190 
191       const __m128 vk2x0123 = _mm_load_ps(w + 24);
192       const __m128 vk2x4567 = _mm_load_ps(w + 28);
193       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
194       vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567));
195 
196       const __m128 vi3x0123 = _mm_loadu_ps(i3);
197       const __m128 vi3x4567 = _mm_loadu_ps(i3 + 4);
198       i3 += 8;
199 
200       const __m128 vk3x0123 = _mm_load_ps(w + 32);
201       const __m128 vk3x4567 = _mm_load_ps(w + 36);
202       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
203       vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi3x4567, vk3x4567));
204 
205       const __m128 vi4x0123 = _mm_loadu_ps(i4);
206       const __m128 vi4x4567 = _mm_loadu_ps(i4 + 4);
207       i4 += 8;
208 
209       const __m128 vk4x0123 = _mm_load_ps(w + 40);
210       const __m128 vk4x4567 = _mm_load_ps(w + 44);
211       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123));
212       vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi4x4567, vk4x4567));
213 
214       const __m128 vi5x0123 = _mm_loadu_ps(i5);
215       const __m128 vi5x4567 = _mm_loadu_ps(i5 + 4);
216       i5 += 8;
217 
218       const __m128 vk5x0123 = _mm_load_ps(w + 48);
219       const __m128 vk5x4567 = _mm_load_ps(w + 52);
220       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123));
221       vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi5x4567, vk5x4567));
222 
223       const __m128 vi6x0123 = _mm_loadu_ps(i6);
224       const __m128 vi6x4567 = _mm_loadu_ps(i6 + 4);
225       i6 += 8;
226 
227       const __m128 vk6x0123 = _mm_load_ps(w + 56);
228       const __m128 vk6x4567 = _mm_load_ps(w + 60);
229       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123));
230       vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi6x4567, vk6x4567));
231 
232       const __m128 vi7x0123 = _mm_loadu_ps(i7);
233       const __m128 vi7x4567 = _mm_loadu_ps(i7 + 4);
234       i7 += 8;
235 
236       const __m128 vk7x0123 = _mm_load_ps(w + 64);
237       const __m128 vk7x4567 = _mm_load_ps(w + 68);
238       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123));
239       vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi7x4567, vk7x4567));
240 
241       const __m128 vi8x0123 = _mm_loadu_ps(i8);
242       const __m128 vi8x4567 = _mm_loadu_ps(i8 + 4);
243       i8 += 8;
244 
245       const __m128 vk8x0123 = _mm_load_ps(w + 72);
246       const __m128 vk8x4567 = _mm_load_ps(w + 76);
247       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123));
248       vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi8x4567, vk8x4567));
249 
250       const __m128 vi9x0123 = _mm_loadu_ps(i9);
251       const __m128 vi9x4567 = _mm_loadu_ps(i9 + 4);
252       i9 += 8;
253 
254       const __m128 vk9x0123 = _mm_load_ps(w + 80);
255       const __m128 vk9x4567 = _mm_load_ps(w + 84);
256       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi9x0123, vk9x0123));
257       vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi9x4567, vk9x4567));
258 
259       const __m128 vi10x0123 = _mm_loadu_ps(i10);
260       const __m128 vi10x4567 = _mm_loadu_ps(i10 + 4);
261       i10 += 8;
262 
263       const __m128 vk10x0123 = _mm_load_ps(w + 88);
264       const __m128 vk10x4567 = _mm_load_ps(w + 92);
265       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi10x0123, vk10x0123));
266       vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi10x4567, vk10x4567));
267 
268       const __m128 vi11x0123 = _mm_loadu_ps(i11);
269       const __m128 vi11x4567 = _mm_loadu_ps(i11 + 4);
270       i11 += 8;
271 
272       const __m128 vk11x0123 = _mm_load_ps(w + 96);
273       const __m128 vk11x4567 = _mm_load_ps(w + 100);
274       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi11x0123, vk11x0123));
275       vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi11x4567, vk11x4567));
276 
277       const __m128 vi12x0123 = _mm_loadu_ps(i12);
278       const __m128 vi12x4567 = _mm_loadu_ps(i12 + 4);
279       i12 += 8;
280 
281       const __m128 vk12x0123 = _mm_load_ps(w + 104);
282       const __m128 vk12x4567 = _mm_load_ps(w + 108);
283       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi12x0123, vk12x0123));
284       vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi12x4567, vk12x4567));
285 
286       const __m128 vi13x0123 = _mm_loadu_ps(i13);
287       const __m128 vi13x4567 = _mm_loadu_ps(i13 + 4);
288       i13 += 8;
289 
290       const __m128 vk13x0123 = _mm_load_ps(w + 112);
291       const __m128 vk13x4567 = _mm_load_ps(w + 116);
292       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi13x0123, vk13x0123));
293       vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi13x4567, vk13x4567));
294 
295       const __m128 vi14x0123 = _mm_loadu_ps(i14);
296       const __m128 vi14x4567 = _mm_loadu_ps(i14 + 4);
297       i14 += 8;
298 
299       const __m128 vk14x0123 = _mm_load_ps(w + 120);
300       const __m128 vk14x4567 = _mm_load_ps(w + 124);
301       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi14x0123, vk14x0123));
302       vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi14x4567, vk14x4567));
303 
304       const __m128 vi15x0123 = _mm_loadu_ps(i15);
305       const __m128 vi15x4567 = _mm_loadu_ps(i15 + 4);
306       i15 += 8;
307 
308       const __m128 vk15x0123 = _mm_load_ps(w + 128);
309       const __m128 vk15x4567 = _mm_load_ps(w + 132);
310       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi15x0123, vk15x0123));
311       vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi15x4567, vk15x4567));
312 
313       const __m128 vi16x0123 = _mm_loadu_ps(i16);
314       const __m128 vi16x4567 = _mm_loadu_ps(i16 + 4);
315       i16 += 8;
316 
317       const __m128 vk16x0123 = _mm_load_ps(w + 136);
318       const __m128 vk16x4567 = _mm_load_ps(w + 140);
319       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi16x0123, vk16x0123));
320       vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi16x4567, vk16x4567));
321 
322       const __m128 vi17x0123 = _mm_loadu_ps(i17);
323       const __m128 vi17x4567 = _mm_loadu_ps(i17 + 4);
324       i17 += 8;
325 
326       const __m128 vk17x0123 = _mm_load_ps(w + 144);
327       const __m128 vk17x4567 = _mm_load_ps(w + 148);
328       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi17x0123, vk17x0123));
329       vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi17x4567, vk17x4567));
330 
331       const __m128 vi18x0123 = _mm_loadu_ps(i18);
332       const __m128 vi18x4567 = _mm_loadu_ps(i18 + 4);
333       i18 += 8;
334 
335       const __m128 vk18x0123 = _mm_load_ps(w + 152);
336       const __m128 vk18x4567 = _mm_load_ps(w + 156);
337       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi18x0123, vk18x0123));
338       vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi18x4567, vk18x4567));
339 
340       const __m128 vi19x0123 = _mm_loadu_ps(i19);
341       const __m128 vi19x4567 = _mm_loadu_ps(i19 + 4);
342       i19 += 8;
343 
344       const __m128 vk19x0123 = _mm_load_ps(w + 160);
345       const __m128 vk19x4567 = _mm_load_ps(w + 164);
346       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi19x0123, vk19x0123));
347       vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi19x4567, vk19x4567));
348 
349       const __m128 vi20x0123 = _mm_loadu_ps(i20);
350       const __m128 vi20x4567 = _mm_loadu_ps(i20 + 4);
351       i20 += 8;
352 
353       const __m128 vk20x0123 = _mm_load_ps(w + 168);
354       const __m128 vk20x4567 = _mm_load_ps(w + 172);
355       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi20x0123, vk20x0123));
356       vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi20x4567, vk20x4567));
357 
358       const __m128 vi21x0123 = _mm_loadu_ps(i21);
359       const __m128 vi21x4567 = _mm_loadu_ps(i21 + 4);
360       i21 += 8;
361 
362       const __m128 vk21x0123 = _mm_load_ps(w + 176);
363       const __m128 vk21x4567 = _mm_load_ps(w + 180);
364       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi21x0123, vk21x0123));
365       vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi21x4567, vk21x4567));
366 
367       const __m128 vi22x0123 = _mm_loadu_ps(i22);
368       const __m128 vi22x4567 = _mm_loadu_ps(i22 + 4);
369       i22 += 8;
370 
371       const __m128 vk22x0123 = _mm_load_ps(w + 184);
372       const __m128 vk22x4567 = _mm_load_ps(w + 188);
373       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi22x0123, vk22x0123));
374       vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi22x4567, vk22x4567));
375 
376       const __m128 vi23x0123 = _mm_loadu_ps(i23);
377       const __m128 vi23x4567 = _mm_loadu_ps(i23 + 4);
378       i23 += 8;
379 
380       const __m128 vk23x0123 = _mm_load_ps(w + 192);
381       const __m128 vk23x4567 = _mm_load_ps(w + 196);
382       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi23x0123, vk23x0123));
383       vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi23x4567, vk23x4567));
384 
385       const __m128 vi24x0123 = _mm_loadu_ps(i24);
386       const __m128 vi24x4567 = _mm_loadu_ps(i24 + 4);
387       i24 += 8;
388 
389       const __m128 vk24x0123 = _mm_load_ps(w + 200);
390       const __m128 vk24x4567 = _mm_load_ps(w + 204);
391       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123));
392       vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi24x4567, vk24x4567));
393 
394       w += 208;
395 
396       // Add up all accumulators to vacc01234567p0
397       vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
398       vacc4567p0 = _mm_add_ps(vacc4567p0, vacc4567p1);
399 
400       __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
401       __m128 vacc4567 = _mm_max_ps(vacc4567p0, vmin);
402       vacc0123 = _mm_min_ps(vacc0123, vmax);
403       vacc4567 = _mm_min_ps(vacc4567, vmax);
404 
405       _mm_storeu_ps(output, vacc0123);
406       _mm_storeu_ps(output + 4, vacc4567);
407       output += 8;
408     }
409     for (; c >= 4; c -= 4) {
410       __m128 vacc0123p0 = _mm_load_ps(w);
411 
412       const __m128 vi0x0123 = _mm_loadu_ps(i0);
413       i0 += 4;
414 
415       const __m128 vk0x0123 = _mm_load_ps(w + 8);
416       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
417 
418       const __m128 vi1x0123 = _mm_loadu_ps(i1);
419       i1 += 4;
420 
421       const __m128 vk1x0123 = _mm_load_ps(w + 16);
422       __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
423 
424       const __m128 vi2x0123 = _mm_loadu_ps(i2);
425       i2 += 4;
426 
427       const __m128 vk2x0123 = _mm_load_ps(w + 24);
428       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
429 
430       const __m128 vi3x0123 = _mm_loadu_ps(i3);
431       i3 += 4;
432 
433       const __m128 vk3x0123 = _mm_load_ps(w + 32);
434       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
435 
436       const __m128 vi4x0123 = _mm_loadu_ps(i4);
437       i4 += 4;
438 
439       const __m128 vk4x0123 = _mm_load_ps(w + 40);
440       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123));
441 
442       const __m128 vi5x0123 = _mm_loadu_ps(i5);
443       i5 += 4;
444 
445       const __m128 vk5x0123 = _mm_load_ps(w + 48);
446       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123));
447 
448       const __m128 vi6x0123 = _mm_loadu_ps(i6);
449       i6 += 4;
450 
451       const __m128 vk6x0123 = _mm_load_ps(w + 56);
452       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123));
453 
454       const __m128 vi7x0123 = _mm_loadu_ps(i7);
455       i7 += 4;
456 
457       const __m128 vk7x0123 = _mm_load_ps(w + 64);
458       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123));
459 
460       const __m128 vi8x0123 = _mm_loadu_ps(i8);
461       i8 += 4;
462 
463       const __m128 vk8x0123 = _mm_load_ps(w + 72);
464       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123));
465 
466       const __m128 vi9x0123 = _mm_loadu_ps(i9);
467       i9 += 4;
468 
469       const __m128 vk9x0123 = _mm_load_ps(w + 80);
470       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi9x0123, vk9x0123));
471 
472       const __m128 vi10x0123 = _mm_loadu_ps(i10);
473       i10 += 4;
474 
475       const __m128 vk10x0123 = _mm_load_ps(w + 88);
476       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi10x0123, vk10x0123));
477 
478       const __m128 vi11x0123 = _mm_loadu_ps(i11);
479       i11 += 4;
480 
481       const __m128 vk11x0123 = _mm_load_ps(w + 96);
482       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi11x0123, vk11x0123));
483 
484       const __m128 vi12x0123 = _mm_loadu_ps(i12);
485       i12 += 4;
486 
487       const __m128 vk12x0123 = _mm_load_ps(w + 104);
488       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi12x0123, vk12x0123));
489 
490       const __m128 vi13x0123 = _mm_loadu_ps(i13);
491       i13 += 4;
492 
493       const __m128 vk13x0123 = _mm_load_ps(w + 112);
494       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi13x0123, vk13x0123));
495 
496       const __m128 vi14x0123 = _mm_loadu_ps(i14);
497       i14 += 4;
498 
499       const __m128 vk14x0123 = _mm_load_ps(w + 120);
500       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi14x0123, vk14x0123));
501 
502       const __m128 vi15x0123 = _mm_loadu_ps(i15);
503       i15 += 4;
504 
505       const __m128 vk15x0123 = _mm_load_ps(w + 128);
506       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi15x0123, vk15x0123));
507 
508       const __m128 vi16x0123 = _mm_loadu_ps(i16);
509       i16 += 4;
510 
511       const __m128 vk16x0123 = _mm_load_ps(w + 136);
512       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi16x0123, vk16x0123));
513 
514       const __m128 vi17x0123 = _mm_loadu_ps(i17);
515       i17 += 4;
516 
517       const __m128 vk17x0123 = _mm_load_ps(w + 144);
518       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi17x0123, vk17x0123));
519 
520       const __m128 vi18x0123 = _mm_loadu_ps(i18);
521       i18 += 4;
522 
523       const __m128 vk18x0123 = _mm_load_ps(w + 152);
524       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi18x0123, vk18x0123));
525 
526       const __m128 vi19x0123 = _mm_loadu_ps(i19);
527       i19 += 4;
528 
529       const __m128 vk19x0123 = _mm_load_ps(w + 160);
530       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi19x0123, vk19x0123));
531 
532       const __m128 vi20x0123 = _mm_loadu_ps(i20);
533       i20 += 4;
534 
535       const __m128 vk20x0123 = _mm_load_ps(w + 168);
536       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi20x0123, vk20x0123));
537 
538       const __m128 vi21x0123 = _mm_loadu_ps(i21);
539       i21 += 4;
540 
541       const __m128 vk21x0123 = _mm_load_ps(w + 176);
542       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi21x0123, vk21x0123));
543 
544       const __m128 vi22x0123 = _mm_loadu_ps(i22);
545       i22 += 4;
546 
547       const __m128 vk22x0123 = _mm_load_ps(w + 184);
548       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi22x0123, vk22x0123));
549 
550       const __m128 vi23x0123 = _mm_loadu_ps(i23);
551       i23 += 4;
552 
553       const __m128 vk23x0123 = _mm_load_ps(w + 192);
554       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi23x0123, vk23x0123));
555 
556       const __m128 vi24x0123 = _mm_loadu_ps(i24);
557       i24 += 4;
558 
559       const __m128 vk24x0123 = _mm_load_ps(w + 200);
560       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123));
561 
562       w += 4;
563 
564       // Add up all accumulators to vacc01234567p0
565       vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
566 
567       __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
568       vacc0123 = _mm_min_ps(vacc0123, vmax);
569 
570       _mm_storeu_ps(output, vacc0123);
571       output += 4;
572     }
573     if XNN_UNLIKELY(c != 0) {
574       __m128 vacc0123p0 = _mm_load_ps(w);
575 
576       const __m128 vi0x0123 = _mm_loadu_ps(i0);
577       const __m128 vk0x0123 = _mm_load_ps(w + 8);
578       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
579 
580       const __m128 vi1x0123 = _mm_loadu_ps(i1);
581       const __m128 vk1x0123 = _mm_load_ps(w + 16);
582       __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
583 
584       const __m128 vi2x0123 = _mm_loadu_ps(i2);
585       const __m128 vk2x0123 = _mm_load_ps(w + 24);
586       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
587 
588       const __m128 vi3x0123 = _mm_loadu_ps(i3);
589       const __m128 vk3x0123 = _mm_load_ps(w + 32);
590       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
591 
592       const __m128 vi4x0123 = _mm_loadu_ps(i4);
593       const __m128 vk4x0123 = _mm_load_ps(w + 40);
594       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123));
595 
596       const __m128 vi5x0123 = _mm_loadu_ps(i5);
597       const __m128 vk5x0123 = _mm_load_ps(w + 48);
598       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123));
599 
600       const __m128 vi6x0123 = _mm_loadu_ps(i6);
601       const __m128 vk6x0123 = _mm_load_ps(w + 56);
602       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123));
603 
604       const __m128 vi7x0123 = _mm_loadu_ps(i7);
605       const __m128 vk7x0123 = _mm_load_ps(w + 64);
606       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123));
607 
608       const __m128 vi8x0123 = _mm_loadu_ps(i8);
609       const __m128 vk8x0123 = _mm_load_ps(w + 72);
610       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123));
611 
612       const __m128 vi9x0123 = _mm_loadu_ps(i9);
613       const __m128 vk9x0123 = _mm_load_ps(w + 80);
614       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi9x0123, vk9x0123));
615 
616       const __m128 vi10x0123 = _mm_loadu_ps(i10);
617       const __m128 vk10x0123 = _mm_load_ps(w + 88);
618       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi10x0123, vk10x0123));
619 
620       const __m128 vi11x0123 = _mm_loadu_ps(i11);
621       const __m128 vk11x0123 = _mm_load_ps(w + 96);
622       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi11x0123, vk11x0123));
623 
624       const __m128 vi12x0123 = _mm_loadu_ps(i12);
625       const __m128 vk12x0123 = _mm_load_ps(w + 104);
626       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi12x0123, vk12x0123));
627 
628       const __m128 vi13x0123 = _mm_loadu_ps(i13);
629       const __m128 vk13x0123 = _mm_load_ps(w + 112);
630       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi13x0123, vk13x0123));
631 
632       const __m128 vi14x0123 = _mm_loadu_ps(i14);
633       const __m128 vk14x0123 = _mm_load_ps(w + 120);
634       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi14x0123, vk14x0123));
635 
636       const __m128 vi15x0123 = _mm_loadu_ps(i15);
637       const __m128 vk15x0123 = _mm_load_ps(w + 128);
638       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi15x0123, vk15x0123));
639 
640       const __m128 vi16x0123 = _mm_loadu_ps(i16);
641       const __m128 vk16x0123 = _mm_load_ps(w + 136);
642       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi16x0123, vk16x0123));
643 
644       const __m128 vi17x0123 = _mm_loadu_ps(i17);
645       const __m128 vk17x0123 = _mm_load_ps(w + 144);
646       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi17x0123, vk17x0123));
647 
648       const __m128 vi18x0123 = _mm_loadu_ps(i18);
649       const __m128 vk18x0123 = _mm_load_ps(w + 152);
650       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi18x0123, vk18x0123));
651 
652       const __m128 vi19x0123 = _mm_loadu_ps(i19);
653       const __m128 vk19x0123 = _mm_load_ps(w + 160);
654       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi19x0123, vk19x0123));
655 
656       const __m128 vi20x0123 = _mm_loadu_ps(i20);
657       const __m128 vk20x0123 = _mm_load_ps(w + 168);
658       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi20x0123, vk20x0123));
659 
660       const __m128 vi21x0123 = _mm_loadu_ps(i21);
661       const __m128 vk21x0123 = _mm_load_ps(w + 176);
662       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi21x0123, vk21x0123));
663 
664       const __m128 vi22x0123 = _mm_loadu_ps(i22);
665       const __m128 vk22x0123 = _mm_load_ps(w + 184);
666       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi22x0123, vk22x0123));
667 
668       const __m128 vi23x0123 = _mm_loadu_ps(i23);
669       const __m128 vk23x0123 = _mm_load_ps(w + 192);
670       vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi23x0123, vk23x0123));
671 
672       const __m128 vi24x0123 = _mm_loadu_ps(i24);
673       const __m128 vk24x0123 = _mm_load_ps(w + 200);
674       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123));
675 
676       // Add up all accumulators to vacc01234567p0
677       vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
678 
679       __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
680       vacc0123 = _mm_min_ps(vacc0123, vmax);
681 
682       if (c & 2) {
683         _mm_storel_pi((__m64*) output, vacc0123);
684         vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
685         output += 2;
686       }
687       if (c & 1) {
688         _mm_store_ss(output, vacc0123);
689         output += 1;
690       }
691     }
692 
693     output = (float*) ((uintptr_t) output + output_increment);
694   } while (--output_width != 0);
695 }
696