xref: /aosp_15_r20/external/XNNPACK/src/f16-dwconv/gen/up16x25-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_up16x25__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_up16x25__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 >= 16; c -= 16) {
168       __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
169       __m256 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 8)));
170 
171 
172       const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
173       const __m256 vi0x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i0 + 8)));
174       i0 += 16;
175 
176       const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 16)));
177       const __m256 vk0x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 24)));
178       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
179       vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x89ABCDEF, vk0x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
180 
181       const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
182       const __m256 vi1x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i1 + 8)));
183       i1 += 16;
184 
185       const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 32)));
186       const __m256 vk1x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 40)));
187       __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
188       __m256 vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x89ABCDEF, vk1x89ABCDEF), _MM_FROUND_NO_EXC));
189 
190       const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
191       const __m256 vi2x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i2 + 8)));
192       i2 += 16;
193 
194       const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 48)));
195       const __m256 vk2x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 56)));
196       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
197       vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
198 
199       const __m256 vi3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
200       const __m256 vi3x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i3 + 8)));
201       i3 += 16;
202 
203       const __m256 vk3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 64)));
204       const __m256 vk3x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 72)));
205       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
206       vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3x89ABCDEF, vk3x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
207 
208       const __m256 vi4x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
209       const __m256 vi4x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i4 + 8)));
210       i4 += 16;
211 
212       const __m256 vk4x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 80)));
213       const __m256 vk4x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 88)));
214       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
215       vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi4x89ABCDEF, vk4x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
216 
217       const __m256 vi5x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
218       const __m256 vi5x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i5 + 8)));
219       i5 += 16;
220 
221       const __m256 vk5x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 96)));
222       const __m256 vk5x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 104)));
223       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
224       vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi5x89ABCDEF, vk5x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
225 
226       const __m256 vi6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
227       const __m256 vi6x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i6 + 8)));
228       i6 += 16;
229 
230       const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 112)));
231       const __m256 vk6x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 120)));
232       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
233       vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi6x89ABCDEF, vk6x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
234 
235       const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
236       const __m256 vi7x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i7 + 8)));
237       i7 += 16;
238 
239       const __m256 vk7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 128)));
240       const __m256 vk7x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 136)));
241       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
242       vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi7x89ABCDEF, vk7x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
243 
244       const __m256 vi8x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));
245       const __m256 vi8x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i8 + 8)));
246       i8 += 16;
247 
248       const __m256 vk8x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 144)));
249       const __m256 vk8x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 152)));
250       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
251       vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi8x89ABCDEF, vk8x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
252 
253       const __m256 vi9x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i9));
254       const __m256 vi9x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i9 + 8)));
255       i9 += 16;
256 
257       const __m256 vk9x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 160)));
258       const __m256 vk9x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 168)));
259       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
260       vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi9x89ABCDEF, vk9x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
261 
262       const __m256 vi10x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i10));
263       const __m256 vi10x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i10 + 8)));
264       i10 += 16;
265 
266       const __m256 vk10x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 176)));
267       const __m256 vk10x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 184)));
268       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
269       vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi10x89ABCDEF, vk10x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
270 
271       const __m256 vi11x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i11));
272       const __m256 vi11x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i11 + 8)));
273       i11 += 16;
274 
275       const __m256 vk11x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 192)));
276       const __m256 vk11x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 200)));
277       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
278       vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi11x89ABCDEF, vk11x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
279 
280       const __m256 vi12x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i12));
281       const __m256 vi12x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i12 + 8)));
282       i12 += 16;
283 
284       const __m256 vk12x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 208)));
285       const __m256 vk12x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 216)));
286       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
287       vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi12x89ABCDEF, vk12x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
288 
289       const __m256 vi13x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i13));
290       const __m256 vi13x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i13 + 8)));
291       i13 += 16;
292 
293       const __m256 vk13x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 224)));
294       const __m256 vk13x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 232)));
295       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
296       vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi13x89ABCDEF, vk13x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
297 
298       const __m256 vi14x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i14));
299       const __m256 vi14x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i14 + 8)));
300       i14 += 16;
301 
302       const __m256 vk14x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 240)));
303       const __m256 vk14x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 248)));
304       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
305       vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi14x89ABCDEF, vk14x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
306 
307       const __m256 vi15x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i15));
308       const __m256 vi15x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i15 + 8)));
309       i15 += 16;
310 
311       const __m256 vk15x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 256)));
312       const __m256 vk15x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 264)));
313       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
314       vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi15x89ABCDEF, vk15x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
315 
316       const __m256 vi16x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i16));
317       const __m256 vi16x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i16 + 8)));
318       i16 += 16;
319 
320       const __m256 vk16x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 272)));
321       const __m256 vk16x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 280)));
322       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
323       vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi16x89ABCDEF, vk16x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
324 
325       const __m256 vi17x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i17));
326       const __m256 vi17x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i17 + 8)));
327       i17 += 16;
328 
329       const __m256 vk17x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 288)));
330       const __m256 vk17x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 296)));
331       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
332       vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi17x89ABCDEF, vk17x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
333 
334       const __m256 vi18x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i18));
335       const __m256 vi18x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i18 + 8)));
336       i18 += 16;
337 
338       const __m256 vk18x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 304)));
339       const __m256 vk18x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 312)));
340       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
341       vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi18x89ABCDEF, vk18x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
342 
343       const __m256 vi19x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i19));
344       const __m256 vi19x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i19 + 8)));
345       i19 += 16;
346 
347       const __m256 vk19x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 320)));
348       const __m256 vk19x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 328)));
349       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
350       vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi19x89ABCDEF, vk19x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
351 
352       const __m256 vi20x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i20));
353       const __m256 vi20x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i20 + 8)));
354       i20 += 16;
355 
356       const __m256 vk20x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 336)));
357       const __m256 vk20x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 344)));
358       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
359       vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi20x89ABCDEF, vk20x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
360 
361       const __m256 vi21x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i21));
362       const __m256 vi21x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i21 + 8)));
363       i21 += 16;
364 
365       const __m256 vk21x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 352)));
366       const __m256 vk21x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 360)));
367       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
368       vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi21x89ABCDEF, vk21x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
369 
370       const __m256 vi22x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i22));
371       const __m256 vi22x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i22 + 8)));
372       i22 += 16;
373 
374       const __m256 vk22x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 368)));
375       const __m256 vk22x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 376)));
376       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
377       vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x89ABCDEF, vk22x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
378 
379       const __m256 vi23x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i23));
380       const __m256 vi23x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i23 + 8)));
381       i23 += 16;
382 
383       const __m256 vk23x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 384)));
384       const __m256 vk23x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 392)));
385       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
386       vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi23x89ABCDEF, vk23x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
387 
388       const __m256 vi24x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i24));
389       const __m256 vi24x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i24 + 8)));
390       i24 += 16;
391 
392       const __m256 vk24x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 400)));
393       const __m256 vk24x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 408)));
394       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
395       vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi24x89ABCDEF, vk24x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
396 
397       w += 416;
398 
399       // Add up all accumulators to vacc0123456789ABCDEFp0
400       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
401       vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc89ABCDEFp0, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
402 
403       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
404       __m256 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEFp0, vmin);
405       vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
406       vacc89ABCDEF = _mm256_min_ps(vacc89ABCDEF, vmax);
407 
408       _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC));
409       _mm_storeu_si128((__m128i*) (o + 8), _mm256_cvtps_ph(vacc89ABCDEF, _MM_FROUND_NO_EXC));
410       o += 16;
411     }
412     for (; c >= 8; c -= 8) {
413       __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
414 
415       const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
416       i0 += 8;
417 
418       const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 16)));
419       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
420 
421       const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
422       i1 += 8;
423 
424       const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 32)));
425       __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
426 
427       const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
428       i2 += 8;
429 
430       const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 48)));
431       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
432 
433       const __m256 vi3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
434       i3 += 8;
435 
436       const __m256 vk3x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 64)));
437       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
438 
439       const __m256 vi4x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
440       i4 += 8;
441 
442       const __m256 vk4x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 80)));
443       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
444 
445       const __m256 vi5x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
446       i5 += 8;
447 
448       const __m256 vk5x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 96)));
449       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
450 
451       const __m256 vi6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
452       i6 += 8;
453 
454       const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 112)));
455       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
456 
457       const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
458       i7 += 8;
459 
460       const __m256 vk7x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 128)));
461       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
462 
463       const __m256 vi8x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));
464       i8 += 8;
465 
466       const __m256 vk8x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 144)));
467       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
468 
469       const __m256 vi9x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i9));
470       i9 += 8;
471 
472       const __m256 vk9x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 160)));
473       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
474 
475       const __m256 vi10x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i10));
476       i10 += 8;
477 
478       const __m256 vk10x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 176)));
479       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
480 
481       const __m256 vi11x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i11));
482       i11 += 8;
483 
484       const __m256 vk11x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 192)));
485       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
486 
487       const __m256 vi12x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i12));
488       i12 += 8;
489 
490       const __m256 vk12x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 208)));
491       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
492 
493       const __m256 vi13x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i13));
494       i13 += 8;
495 
496       const __m256 vk13x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 224)));
497       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
498 
499       const __m256 vi14x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i14));
500       i14 += 8;
501 
502       const __m256 vk14x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 240)));
503       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
504 
505       const __m256 vi15x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i15));
506       i15 += 8;
507 
508       const __m256 vk15x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 256)));
509       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
510 
511       const __m256 vi16x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i16));
512       i16 += 8;
513 
514       const __m256 vk16x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 272)));
515       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
516 
517       const __m256 vi17x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i17));
518       i17 += 8;
519 
520       const __m256 vk17x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 288)));
521       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
522 
523       const __m256 vi18x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i18));
524       i18 += 8;
525 
526       const __m256 vk18x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 304)));
527       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
528 
529       const __m256 vi19x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i19));
530       i19 += 8;
531 
532       const __m256 vk19x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 320)));
533       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
534 
535       const __m256 vi20x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i20));
536       i20 += 8;
537 
538       const __m256 vk20x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 336)));
539       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
540 
541       const __m256 vi21x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i21));
542       i21 += 8;
543 
544       const __m256 vk21x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 352)));
545       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
546 
547       const __m256 vi22x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i22));
548       i22 += 8;
549 
550       const __m256 vk22x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 368)));
551       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
552 
553       const __m256 vi23x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i23));
554       i23 += 8;
555 
556       const __m256 vk23x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 384)));
557       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
558 
559       const __m256 vi24x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i24));
560       i24 += 8;
561 
562       const __m256 vk24x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 400)));
563       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
564 
565       w += 8;
566 
567       // Add up all accumulators to vacc01234567p0
568       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
569 
570       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
571       vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
572 
573       _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC));
574       o += 8;
575     }
576     if XNN_UNLIKELY(c != 0) {
577       assert(c >= 1);
578       assert(c <= 7);
579 
580       __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
581 
582       const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
583 
584       const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 16)));
585       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
586 
587       const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
588 
589       const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 32)));
590       __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
591 
592       const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
593 
594       const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 48)));
595       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
596 
597       const __m256 vi3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
598 
599       const __m256 vk3x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 64)));
600       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
601 
602       const __m256 vi4x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
603 
604       const __m256 vk4x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 80)));
605       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
606 
607       const __m256 vi5x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
608 
609       const __m256 vk5x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 96)));
610       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
611 
612       const __m256 vi6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
613 
614       const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 112)));
615       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
616 
617       const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
618 
619       const __m256 vk7x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 128)));
620       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
621 
622       const __m256 vi8x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));
623 
624       const __m256 vk8x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 144)));
625       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
626 
627       const __m256 vi9x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i9));
628 
629       const __m256 vk9x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 160)));
630       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
631 
632       const __m256 vi10x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i10));
633 
634       const __m256 vk10x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 176)));
635       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
636 
637       const __m256 vi11x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i11));
638 
639       const __m256 vk11x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 192)));
640       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
641 
642       const __m256 vi12x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i12));
643 
644       const __m256 vk12x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 208)));
645       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
646 
647       const __m256 vi13x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i13));
648 
649       const __m256 vk13x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 224)));
650       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
651 
652       const __m256 vi14x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i14));
653 
654       const __m256 vk14x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 240)));
655       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
656 
657       const __m256 vi15x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i15));
658 
659       const __m256 vk15x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 256)));
660       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
661 
662       const __m256 vi16x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i16));
663 
664       const __m256 vk16x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 272)));
665       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
666 
667       const __m256 vi17x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i17));
668 
669       const __m256 vk17x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 288)));
670       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
671 
672       const __m256 vi18x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i18));
673 
674       const __m256 vk18x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 304)));
675       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
676 
677       const __m256 vi19x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i19));
678 
679       const __m256 vk19x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 320)));
680       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
681 
682       const __m256 vi20x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i20));
683 
684       const __m256 vk20x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 336)));
685       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
686 
687       const __m256 vi21x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i21));
688 
689       const __m256 vk21x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 352)));
690       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
691 
692       const __m256 vi22x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i22));
693 
694       const __m256 vk22x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 368)));
695       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
696 
697       const __m256 vi23x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i23));
698 
699       const __m256 vk23x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 384)));
700       vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
701 
702       const __m256 vi24x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i24));
703 
704       const __m256 vk24x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 400)));
705       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
706 
707       // Add up all accumulators to vacc01234567p0
708       vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
709 
710       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
711       vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
712 
713       __m128i vh01234567 = _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC);
714       if (c & 4) {
715         _mm_storel_epi64((__m128i*) o, vh01234567);
716         vh01234567 = _mm_unpackhi_epi64(vh01234567, vh01234567);
717         o += 4;
718       }
719       if (c & 2) {
720         _mm_storeu_si32(o, vh01234567);
721         vh01234567 = _mm_srli_epi64(vh01234567, 32);
722         o += 2;
723       }
724       if (c & 1) {
725         *o = (uint16_t) _mm_extract_epi16(vh01234567, 0);
726         o += 1;
727       }
728     }
729 
730     o = (uint16_t*) ((uintptr_t) o + output_increment);
731   } while (--output_width != 0);
732 }
733