xref: /aosp_15_r20/external/XNNPACK/src/f16-dwconv/gen/up8x25-minmax-fma3-acc2.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-dwconv/up-fma3.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 <immintrin.h>
13 
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 
17 
xnn_f16_dwconv_minmax_ukernel_up8x25__fma3_acc2(size_t channels,size_t output_width,const void ** input,const void * weights,void * output,size_t input_stride,size_t output_increment,size_t input_offset,const void * zero,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_dwconv_minmax_ukernel_up8x25__fma3_acc2(
19     size_t channels,
20     size_t output_width,
21     const void** input,
22     const void* weights,
23     void* output,
24     size_t input_stride,
25     size_t output_increment,
26     size_t input_offset,
27     const void* zero,
28     const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(channels != 0);
31   assert(output_width != 0);
32 
33   const __m256 vmax = _mm256_load_ps(params->avx.max);
34   const __m256 vmin = _mm256_load_ps(params->avx.min);
35 
36   uint16_t* o = (uint16_t*) output;
37   do {
38     const uint16_t* i0 = input[0];
39     assert(i0 != NULL);
40     if XNN_UNPREDICTABLE(i0 != zero) {
41       i0 = (const uint16_t*) ((uintptr_t) i0 + input_offset);
42     }
43     const uint16_t* i1 = input[1];
44     assert(i1 != NULL);
45     if XNN_UNPREDICTABLE(i1 != zero) {
46       i1 = (const uint16_t*) ((uintptr_t) i1 + input_offset);
47     }
48     const uint16_t* i2 = input[2];
49     assert(i2 != NULL);
50     if XNN_UNPREDICTABLE(i2 != zero) {
51       i2 = (const uint16_t*) ((uintptr_t) i2 + input_offset);
52     }
53     const uint16_t* i3 = input[3];
54     assert(i3 != NULL);
55     if XNN_UNPREDICTABLE(i3 != zero) {
56       i3 = (const uint16_t*) ((uintptr_t) i3 + input_offset);
57     }
58     const uint16_t* i4 = input[4];
59     assert(i4 != NULL);
60     if XNN_UNPREDICTABLE(i4 != zero) {
61       i4 = (const uint16_t*) ((uintptr_t) i4 + input_offset);
62     }
63     const uint16_t* i5 = input[5];
64     assert(i5 != NULL);
65     if XNN_UNPREDICTABLE(i5 != zero) {
66       i5 = (const uint16_t*) ((uintptr_t) i5 + input_offset);
67     }
68     const uint16_t* i6 = input[6];
69     assert(i6 != NULL);
70     if XNN_UNPREDICTABLE(i6 != zero) {
71       i6 = (const uint16_t*) ((uintptr_t) i6 + input_offset);
72     }
73     const uint16_t* i7 = input[7];
74     assert(i7 != NULL);
75     if XNN_UNPREDICTABLE(i7 != zero) {
76       i7 = (const uint16_t*) ((uintptr_t) i7 + input_offset);
77     }
78     const uint16_t* i8 = input[8];
79     assert(i8 != NULL);
80     if XNN_UNPREDICTABLE(i8 != zero) {
81       i8 = (const uint16_t*) ((uintptr_t) i8 + input_offset);
82     }
83     const uint16_t* i9 = input[9];
84     assert(i9 != NULL);
85     if XNN_UNPREDICTABLE(i9 != zero) {
86       i9 = (const uint16_t*) ((uintptr_t) i9 + input_offset);
87     }
88     const uint16_t* i10 = input[10];
89     assert(i10 != NULL);
90     if XNN_UNPREDICTABLE(i10 != zero) {
91       i10 = (const uint16_t*) ((uintptr_t) i10 + input_offset);
92     }
93     const uint16_t* i11 = input[11];
94     assert(i11 != NULL);
95     if XNN_UNPREDICTABLE(i11 != zero) {
96       i11 = (const uint16_t*) ((uintptr_t) i11 + input_offset);
97     }
98     const uint16_t* i12 = input[12];
99     assert(i12 != NULL);
100     if XNN_UNPREDICTABLE(i12 != zero) {
101       i12 = (const uint16_t*) ((uintptr_t) i12 + input_offset);
102     }
103     const uint16_t* i13 = input[13];
104     assert(i13 != NULL);
105     if XNN_UNPREDICTABLE(i13 != zero) {
106       i13 = (const uint16_t*) ((uintptr_t) i13 + input_offset);
107     }
108     const uint16_t* i14 = input[14];
109     assert(i14 != NULL);
110     if XNN_UNPREDICTABLE(i14 != zero) {
111       i14 = (const uint16_t*) ((uintptr_t) i14 + input_offset);
112     }
113     const uint16_t* i15 = input[15];
114     assert(i15 != NULL);
115     if XNN_UNPREDICTABLE(i15 != zero) {
116       i15 = (const uint16_t*) ((uintptr_t) i15 + input_offset);
117     }
118     const uint16_t* i16 = input[16];
119     assert(i16 != NULL);
120     if XNN_UNPREDICTABLE(i16 != zero) {
121       i16 = (const uint16_t*) ((uintptr_t) i16 + input_offset);
122     }
123     const uint16_t* i17 = input[17];
124     assert(i17 != NULL);
125     if XNN_UNPREDICTABLE(i17 != zero) {
126       i17 = (const uint16_t*) ((uintptr_t) i17 + input_offset);
127     }
128     const uint16_t* i18 = input[18];
129     assert(i18 != NULL);
130     if XNN_UNPREDICTABLE(i18 != zero) {
131       i18 = (const uint16_t*) ((uintptr_t) i18 + input_offset);
132     }
133     const uint16_t* i19 = input[19];
134     assert(i19 != NULL);
135     if XNN_UNPREDICTABLE(i19 != zero) {
136       i19 = (const uint16_t*) ((uintptr_t) i19 + input_offset);
137     }
138     const uint16_t* i20 = input[20];
139     assert(i20 != NULL);
140     if XNN_UNPREDICTABLE(i20 != zero) {
141       i20 = (const uint16_t*) ((uintptr_t) i20 + input_offset);
142     }
143     const uint16_t* i21 = input[21];
144     assert(i21 != NULL);
145     if XNN_UNPREDICTABLE(i21 != zero) {
146       i21 = (const uint16_t*) ((uintptr_t) i21 + input_offset);
147     }
148     const uint16_t* i22 = input[22];
149     assert(i22 != NULL);
150     if XNN_UNPREDICTABLE(i22 != zero) {
151       i22 = (const uint16_t*) ((uintptr_t) i22 + input_offset);
152     }
153     const uint16_t* i23 = input[23];
154     assert(i23 != NULL);
155     if XNN_UNPREDICTABLE(i23 != zero) {
156       i23 = (const uint16_t*) ((uintptr_t) i23 + input_offset);
157     }
158     const uint16_t* i24 = input[24];
159     assert(i24 != NULL);
160     if XNN_UNPREDICTABLE(i24 != zero) {
161       i24 = (const uint16_t*) ((uintptr_t) i24 + input_offset);
162     }
163     input = (const void**) ((uintptr_t) input + input_stride);
164 
165     size_t c = channels;
166     const uint16_t* w = weights;
167     for (; c >= 8; c -= 8) {
168       __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
169 
170 
171       const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
172       i0 += 8;
173 
174       const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 8)));
175       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
176 
177       const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
178       i1 += 8;
179 
180       const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 16)));
181       __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
182 
183       const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
184       i2 += 8;
185 
186       const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 24)));
187       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
188 
189       const __m256 vi3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
190       i3 += 8;
191 
192       const __m256 vk3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 32)));
193       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
194 
195       const __m256 vi4x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
196       i4 += 8;
197 
198       const __m256 vk4x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 40)));
199       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
200 
201       const __m256 vi5x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
202       i5 += 8;
203 
204       const __m256 vk5x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 48)));
205       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
206 
207       const __m256 vi6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
208       i6 += 8;
209 
210       const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 56)));
211       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
212 
213       const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
214       i7 += 8;
215 
216       const __m256 vk7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 64)));
217       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
218 
219       const __m256 vi8x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));
220       i8 += 8;
221 
222       const __m256 vk8x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 72)));
223       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
224 
225       const __m256 vi9x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i9));
226       i9 += 8;
227 
228       const __m256 vk9x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 80)));
229       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
230 
231       const __m256 vi10x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i10));
232       i10 += 8;
233 
234       const __m256 vk10x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 88)));
235       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
236 
237       const __m256 vi11x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i11));
238       i11 += 8;
239 
240       const __m256 vk11x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 96)));
241       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
242 
243       const __m256 vi12x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i12));
244       i12 += 8;
245 
246       const __m256 vk12x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 104)));
247       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
248 
249       const __m256 vi13x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i13));
250       i13 += 8;
251 
252       const __m256 vk13x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 112)));
253       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
254 
255       const __m256 vi14x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i14));
256       i14 += 8;
257 
258       const __m256 vk14x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 120)));
259       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
260 
261       const __m256 vi15x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i15));
262       i15 += 8;
263 
264       const __m256 vk15x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 128)));
265       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
266 
267       const __m256 vi16x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i16));
268       i16 += 8;
269 
270       const __m256 vk16x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 136)));
271       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
272 
273       const __m256 vi17x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i17));
274       i17 += 8;
275 
276       const __m256 vk17x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 144)));
277       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
278 
279       const __m256 vi18x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i18));
280       i18 += 8;
281 
282       const __m256 vk18x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 152)));
283       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
284 
285       const __m256 vi19x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i19));
286       i19 += 8;
287 
288       const __m256 vk19x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 160)));
289       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
290 
291       const __m256 vi20x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i20));
292       i20 += 8;
293 
294       const __m256 vk20x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 168)));
295       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
296 
297       const __m256 vi21x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i21));
298       i21 += 8;
299 
300       const __m256 vk21x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 176)));
301       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
302 
303       const __m256 vi22x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i22));
304       i22 += 8;
305 
306       const __m256 vk22x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 184)));
307       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
308 
309       const __m256 vi23x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i23));
310       i23 += 8;
311 
312       const __m256 vk23x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 192)));
313       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
314 
315       const __m256 vi24x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i24));
316       i24 += 8;
317 
318       const __m256 vk24x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 200)));
319       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
320 
321       w += 208;
322 
323       // Add up all accumulators to vacc01234567p0
324       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
325 
326       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
327       vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
328 
329       _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC));
330       o += 8;
331     }
332     if XNN_UNLIKELY(c != 0) {
333       assert(c >= 1);
334       assert(c <= 7);
335 
336       __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
337 
338       const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
339 
340       const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 8)));
341       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
342 
343       const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
344 
345       const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 16)));
346       __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
347 
348       const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
349 
350       const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 24)));
351       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
352 
353       const __m256 vi3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
354 
355       const __m256 vk3x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 32)));
356       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
357 
358       const __m256 vi4x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
359 
360       const __m256 vk4x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 40)));
361       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
362 
363       const __m256 vi5x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
364 
365       const __m256 vk5x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 48)));
366       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
367 
368       const __m256 vi6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
369 
370       const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 56)));
371       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
372 
373       const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
374 
375       const __m256 vk7x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 64)));
376       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
377 
378       const __m256 vi8x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));
379 
380       const __m256 vk8x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 72)));
381       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
382 
383       const __m256 vi9x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i9));
384 
385       const __m256 vk9x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 80)));
386       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
387 
388       const __m256 vi10x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i10));
389 
390       const __m256 vk10x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 88)));
391       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
392 
393       const __m256 vi11x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i11));
394 
395       const __m256 vk11x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 96)));
396       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
397 
398       const __m256 vi12x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i12));
399 
400       const __m256 vk12x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 104)));
401       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
402 
403       const __m256 vi13x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i13));
404 
405       const __m256 vk13x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 112)));
406       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
407 
408       const __m256 vi14x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i14));
409 
410       const __m256 vk14x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 120)));
411       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
412 
413       const __m256 vi15x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i15));
414 
415       const __m256 vk15x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 128)));
416       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
417 
418       const __m256 vi16x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i16));
419 
420       const __m256 vk16x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 136)));
421       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
422 
423       const __m256 vi17x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i17));
424 
425       const __m256 vk17x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 144)));
426       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
427 
428       const __m256 vi18x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i18));
429 
430       const __m256 vk18x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 152)));
431       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
432 
433       const __m256 vi19x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i19));
434 
435       const __m256 vk19x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 160)));
436       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
437 
438       const __m256 vi20x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i20));
439 
440       const __m256 vk20x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 168)));
441       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
442 
443       const __m256 vi21x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i21));
444 
445       const __m256 vk21x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 176)));
446       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
447 
448       const __m256 vi22x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i22));
449 
450       const __m256 vk22x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 184)));
451       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
452 
453       const __m256 vi23x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i23));
454 
455       const __m256 vk23x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 192)));
456       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
457 
458       const __m256 vi24x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i24));
459 
460       const __m256 vk24x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 200)));
461       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
462 
463       // Add up all accumulators to vacc01234567p0
464       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
465 
466       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
467       vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
468 
469       __m128i vh01234567 = _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC);
470       if (c & 4) {
471         _mm_storel_epi64((__m128i*) o, vh01234567);
472         vh01234567 = _mm_unpackhi_epi64(vh01234567, vh01234567);
473         o += 4;
474       }
475       if (c & 2) {
476         _mm_storeu_si32(o, vh01234567);
477         vh01234567 = _mm_srli_epi64(vh01234567, 32);
478         o += 2;
479       }
480       if (c & 1) {
481         *o = (uint16_t) _mm_extract_epi16(vh01234567, 0);
482         o += 1;
483       }
484     }
485 
486     o = (uint16_t*) ((uintptr_t) o + output_increment);
487   } while (--output_width != 0);
488 }
489