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_up32x25__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_up32x25__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 >= 32; c -= 32) {
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 __m256 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 16)));
171 __m256 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 24)));
172
173
174 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
175 const __m256 vi0x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i0 + 8)));
176 const __m256 vi0xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i0 + 16)));
177 const __m256 vi0xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i0 + 24)));
178 i0 += 32;
179
180 const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 32)));
181 const __m256 vk0x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 40)));
182 const __m256 vk0xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 48)));
183 const __m256 vk0xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 56)));
184 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
185 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x89ABCDEF, vk0x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
186 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0xGHIJKLMN, vk0xGHIJKLMN, vaccGHIJKLMNp0), _MM_FROUND_NO_EXC));
187 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0xOPQRSTUV, vk0xOPQRSTUV, vaccOPQRSTUVp0), _MM_FROUND_NO_EXC));
188
189 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
190 const __m256 vi1x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i1 + 8)));
191 const __m256 vi1xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i1 + 16)));
192 const __m256 vi1xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i1 + 24)));
193 i1 += 32;
194
195 const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 64)));
196 const __m256 vk1x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 72)));
197 const __m256 vk1xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 80)));
198 const __m256 vk1xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 88)));
199 __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
200 __m256 vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x89ABCDEF, vk1x89ABCDEF), _MM_FROUND_NO_EXC));
201 __m256 vaccGHIJKLMNp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1xGHIJKLMN, vk1xGHIJKLMN), _MM_FROUND_NO_EXC));
202 __m256 vaccOPQRSTUVp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1xOPQRSTUV, vk1xOPQRSTUV), _MM_FROUND_NO_EXC));
203
204 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
205 const __m256 vi2x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i2 + 8)));
206 const __m256 vi2xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i2 + 16)));
207 const __m256 vi2xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i2 + 24)));
208 i2 += 32;
209
210 const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 96)));
211 const __m256 vk2x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 104)));
212 const __m256 vk2xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 112)));
213 const __m256 vk2xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 120)));
214 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
215 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
216 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2xGHIJKLMN, vk2xGHIJKLMN, vaccGHIJKLMNp0), _MM_FROUND_NO_EXC));
217 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2xOPQRSTUV, vk2xOPQRSTUV, vaccOPQRSTUVp0), _MM_FROUND_NO_EXC));
218
219 const __m256 vi3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
220 const __m256 vi3x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i3 + 8)));
221 const __m256 vi3xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i3 + 16)));
222 const __m256 vi3xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i3 + 24)));
223 i3 += 32;
224
225 const __m256 vk3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 128)));
226 const __m256 vk3x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 136)));
227 const __m256 vk3xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 144)));
228 const __m256 vk3xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 152)));
229 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
230 vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3x89ABCDEF, vk3x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
231 vaccGHIJKLMNp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3xGHIJKLMN, vk3xGHIJKLMN, vaccGHIJKLMNp1), _MM_FROUND_NO_EXC));
232 vaccOPQRSTUVp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3xOPQRSTUV, vk3xOPQRSTUV, vaccOPQRSTUVp1), _MM_FROUND_NO_EXC));
233
234 const __m256 vi4x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
235 const __m256 vi4x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i4 + 8)));
236 const __m256 vi4xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i4 + 16)));
237 const __m256 vi4xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i4 + 24)));
238 i4 += 32;
239
240 const __m256 vk4x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 160)));
241 const __m256 vk4x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 168)));
242 const __m256 vk4xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 176)));
243 const __m256 vk4xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 184)));
244 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
245 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi4x89ABCDEF, vk4x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
246 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi4xGHIJKLMN, vk4xGHIJKLMN, vaccGHIJKLMNp0), _MM_FROUND_NO_EXC));
247 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi4xOPQRSTUV, vk4xOPQRSTUV, vaccOPQRSTUVp0), _MM_FROUND_NO_EXC));
248
249 const __m256 vi5x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
250 const __m256 vi5x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i5 + 8)));
251 const __m256 vi5xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i5 + 16)));
252 const __m256 vi5xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i5 + 24)));
253 i5 += 32;
254
255 const __m256 vk5x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 192)));
256 const __m256 vk5x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 200)));
257 const __m256 vk5xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 208)));
258 const __m256 vk5xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 216)));
259 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
260 vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi5x89ABCDEF, vk5x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
261 vaccGHIJKLMNp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi5xGHIJKLMN, vk5xGHIJKLMN, vaccGHIJKLMNp1), _MM_FROUND_NO_EXC));
262 vaccOPQRSTUVp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi5xOPQRSTUV, vk5xOPQRSTUV, vaccOPQRSTUVp1), _MM_FROUND_NO_EXC));
263
264 const __m256 vi6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
265 const __m256 vi6x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i6 + 8)));
266 const __m256 vi6xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i6 + 16)));
267 const __m256 vi6xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i6 + 24)));
268 i6 += 32;
269
270 const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 224)));
271 const __m256 vk6x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 232)));
272 const __m256 vk6xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 240)));
273 const __m256 vk6xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 248)));
274 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
275 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi6x89ABCDEF, vk6x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
276 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi6xGHIJKLMN, vk6xGHIJKLMN, vaccGHIJKLMNp0), _MM_FROUND_NO_EXC));
277 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi6xOPQRSTUV, vk6xOPQRSTUV, vaccOPQRSTUVp0), _MM_FROUND_NO_EXC));
278
279 const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
280 const __m256 vi7x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i7 + 8)));
281 const __m256 vi7xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i7 + 16)));
282 const __m256 vi7xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i7 + 24)));
283 i7 += 32;
284
285 const __m256 vk7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 256)));
286 const __m256 vk7x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 264)));
287 const __m256 vk7xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 272)));
288 const __m256 vk7xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 280)));
289 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
290 vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi7x89ABCDEF, vk7x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
291 vaccGHIJKLMNp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi7xGHIJKLMN, vk7xGHIJKLMN, vaccGHIJKLMNp1), _MM_FROUND_NO_EXC));
292 vaccOPQRSTUVp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi7xOPQRSTUV, vk7xOPQRSTUV, vaccOPQRSTUVp1), _MM_FROUND_NO_EXC));
293
294 const __m256 vi8x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));
295 const __m256 vi8x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i8 + 8)));
296 const __m256 vi8xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i8 + 16)));
297 const __m256 vi8xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i8 + 24)));
298 i8 += 32;
299
300 const __m256 vk8x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 288)));
301 const __m256 vk8x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 296)));
302 const __m256 vk8xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 304)));
303 const __m256 vk8xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 312)));
304 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
305 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi8x89ABCDEF, vk8x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
306 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi8xGHIJKLMN, vk8xGHIJKLMN, vaccGHIJKLMNp0), _MM_FROUND_NO_EXC));
307 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi8xOPQRSTUV, vk8xOPQRSTUV, vaccOPQRSTUVp0), _MM_FROUND_NO_EXC));
308
309 const __m256 vi9x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i9));
310 const __m256 vi9x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i9 + 8)));
311 const __m256 vi9xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i9 + 16)));
312 const __m256 vi9xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i9 + 24)));
313 i9 += 32;
314
315 const __m256 vk9x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 320)));
316 const __m256 vk9x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 328)));
317 const __m256 vk9xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 336)));
318 const __m256 vk9xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 344)));
319 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
320 vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi9x89ABCDEF, vk9x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
321 vaccGHIJKLMNp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi9xGHIJKLMN, vk9xGHIJKLMN, vaccGHIJKLMNp1), _MM_FROUND_NO_EXC));
322 vaccOPQRSTUVp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi9xOPQRSTUV, vk9xOPQRSTUV, vaccOPQRSTUVp1), _MM_FROUND_NO_EXC));
323
324 const __m256 vi10x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i10));
325 const __m256 vi10x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i10 + 8)));
326 const __m256 vi10xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i10 + 16)));
327 const __m256 vi10xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i10 + 24)));
328 i10 += 32;
329
330 const __m256 vk10x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 352)));
331 const __m256 vk10x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 360)));
332 const __m256 vk10xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 368)));
333 const __m256 vk10xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 376)));
334 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
335 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi10x89ABCDEF, vk10x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
336 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi10xGHIJKLMN, vk10xGHIJKLMN, vaccGHIJKLMNp0), _MM_FROUND_NO_EXC));
337 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi10xOPQRSTUV, vk10xOPQRSTUV, vaccOPQRSTUVp0), _MM_FROUND_NO_EXC));
338
339 const __m256 vi11x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i11));
340 const __m256 vi11x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i11 + 8)));
341 const __m256 vi11xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i11 + 16)));
342 const __m256 vi11xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i11 + 24)));
343 i11 += 32;
344
345 const __m256 vk11x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 384)));
346 const __m256 vk11x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 392)));
347 const __m256 vk11xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 400)));
348 const __m256 vk11xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 408)));
349 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
350 vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi11x89ABCDEF, vk11x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
351 vaccGHIJKLMNp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi11xGHIJKLMN, vk11xGHIJKLMN, vaccGHIJKLMNp1), _MM_FROUND_NO_EXC));
352 vaccOPQRSTUVp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi11xOPQRSTUV, vk11xOPQRSTUV, vaccOPQRSTUVp1), _MM_FROUND_NO_EXC));
353
354 const __m256 vi12x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i12));
355 const __m256 vi12x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i12 + 8)));
356 const __m256 vi12xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i12 + 16)));
357 const __m256 vi12xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i12 + 24)));
358 i12 += 32;
359
360 const __m256 vk12x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 416)));
361 const __m256 vk12x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 424)));
362 const __m256 vk12xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 432)));
363 const __m256 vk12xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 440)));
364 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
365 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi12x89ABCDEF, vk12x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
366 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi12xGHIJKLMN, vk12xGHIJKLMN, vaccGHIJKLMNp0), _MM_FROUND_NO_EXC));
367 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi12xOPQRSTUV, vk12xOPQRSTUV, vaccOPQRSTUVp0), _MM_FROUND_NO_EXC));
368
369 const __m256 vi13x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i13));
370 const __m256 vi13x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i13 + 8)));
371 const __m256 vi13xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i13 + 16)));
372 const __m256 vi13xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i13 + 24)));
373 i13 += 32;
374
375 const __m256 vk13x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 448)));
376 const __m256 vk13x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 456)));
377 const __m256 vk13xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 464)));
378 const __m256 vk13xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 472)));
379 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
380 vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi13x89ABCDEF, vk13x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
381 vaccGHIJKLMNp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi13xGHIJKLMN, vk13xGHIJKLMN, vaccGHIJKLMNp1), _MM_FROUND_NO_EXC));
382 vaccOPQRSTUVp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi13xOPQRSTUV, vk13xOPQRSTUV, vaccOPQRSTUVp1), _MM_FROUND_NO_EXC));
383
384 const __m256 vi14x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i14));
385 const __m256 vi14x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i14 + 8)));
386 const __m256 vi14xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i14 + 16)));
387 const __m256 vi14xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i14 + 24)));
388 i14 += 32;
389
390 const __m256 vk14x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 480)));
391 const __m256 vk14x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 488)));
392 const __m256 vk14xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 496)));
393 const __m256 vk14xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 504)));
394 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
395 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi14x89ABCDEF, vk14x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
396 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi14xGHIJKLMN, vk14xGHIJKLMN, vaccGHIJKLMNp0), _MM_FROUND_NO_EXC));
397 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi14xOPQRSTUV, vk14xOPQRSTUV, vaccOPQRSTUVp0), _MM_FROUND_NO_EXC));
398
399 const __m256 vi15x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i15));
400 const __m256 vi15x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i15 + 8)));
401 const __m256 vi15xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i15 + 16)));
402 const __m256 vi15xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i15 + 24)));
403 i15 += 32;
404
405 const __m256 vk15x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 512)));
406 const __m256 vk15x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 520)));
407 const __m256 vk15xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 528)));
408 const __m256 vk15xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 536)));
409 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
410 vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi15x89ABCDEF, vk15x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
411 vaccGHIJKLMNp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi15xGHIJKLMN, vk15xGHIJKLMN, vaccGHIJKLMNp1), _MM_FROUND_NO_EXC));
412 vaccOPQRSTUVp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi15xOPQRSTUV, vk15xOPQRSTUV, vaccOPQRSTUVp1), _MM_FROUND_NO_EXC));
413
414 const __m256 vi16x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i16));
415 const __m256 vi16x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i16 + 8)));
416 const __m256 vi16xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i16 + 16)));
417 const __m256 vi16xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i16 + 24)));
418 i16 += 32;
419
420 const __m256 vk16x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 544)));
421 const __m256 vk16x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 552)));
422 const __m256 vk16xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 560)));
423 const __m256 vk16xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 568)));
424 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
425 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi16x89ABCDEF, vk16x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
426 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi16xGHIJKLMN, vk16xGHIJKLMN, vaccGHIJKLMNp0), _MM_FROUND_NO_EXC));
427 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi16xOPQRSTUV, vk16xOPQRSTUV, vaccOPQRSTUVp0), _MM_FROUND_NO_EXC));
428
429 const __m256 vi17x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i17));
430 const __m256 vi17x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i17 + 8)));
431 const __m256 vi17xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i17 + 16)));
432 const __m256 vi17xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i17 + 24)));
433 i17 += 32;
434
435 const __m256 vk17x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 576)));
436 const __m256 vk17x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 584)));
437 const __m256 vk17xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 592)));
438 const __m256 vk17xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 600)));
439 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
440 vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi17x89ABCDEF, vk17x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
441 vaccGHIJKLMNp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi17xGHIJKLMN, vk17xGHIJKLMN, vaccGHIJKLMNp1), _MM_FROUND_NO_EXC));
442 vaccOPQRSTUVp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi17xOPQRSTUV, vk17xOPQRSTUV, vaccOPQRSTUVp1), _MM_FROUND_NO_EXC));
443
444 const __m256 vi18x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i18));
445 const __m256 vi18x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i18 + 8)));
446 const __m256 vi18xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i18 + 16)));
447 const __m256 vi18xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i18 + 24)));
448 i18 += 32;
449
450 const __m256 vk18x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 608)));
451 const __m256 vk18x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 616)));
452 const __m256 vk18xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 624)));
453 const __m256 vk18xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 632)));
454 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
455 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi18x89ABCDEF, vk18x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
456 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi18xGHIJKLMN, vk18xGHIJKLMN, vaccGHIJKLMNp0), _MM_FROUND_NO_EXC));
457 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi18xOPQRSTUV, vk18xOPQRSTUV, vaccOPQRSTUVp0), _MM_FROUND_NO_EXC));
458
459 const __m256 vi19x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i19));
460 const __m256 vi19x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i19 + 8)));
461 const __m256 vi19xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i19 + 16)));
462 const __m256 vi19xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i19 + 24)));
463 i19 += 32;
464
465 const __m256 vk19x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 640)));
466 const __m256 vk19x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 648)));
467 const __m256 vk19xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 656)));
468 const __m256 vk19xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 664)));
469 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
470 vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi19x89ABCDEF, vk19x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
471 vaccGHIJKLMNp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi19xGHIJKLMN, vk19xGHIJKLMN, vaccGHIJKLMNp1), _MM_FROUND_NO_EXC));
472 vaccOPQRSTUVp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi19xOPQRSTUV, vk19xOPQRSTUV, vaccOPQRSTUVp1), _MM_FROUND_NO_EXC));
473
474 const __m256 vi20x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i20));
475 const __m256 vi20x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i20 + 8)));
476 const __m256 vi20xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i20 + 16)));
477 const __m256 vi20xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i20 + 24)));
478 i20 += 32;
479
480 const __m256 vk20x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 672)));
481 const __m256 vk20x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 680)));
482 const __m256 vk20xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 688)));
483 const __m256 vk20xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 696)));
484 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
485 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi20x89ABCDEF, vk20x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
486 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi20xGHIJKLMN, vk20xGHIJKLMN, vaccGHIJKLMNp0), _MM_FROUND_NO_EXC));
487 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi20xOPQRSTUV, vk20xOPQRSTUV, vaccOPQRSTUVp0), _MM_FROUND_NO_EXC));
488
489 const __m256 vi21x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i21));
490 const __m256 vi21x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i21 + 8)));
491 const __m256 vi21xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i21 + 16)));
492 const __m256 vi21xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i21 + 24)));
493 i21 += 32;
494
495 const __m256 vk21x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 704)));
496 const __m256 vk21x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 712)));
497 const __m256 vk21xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 720)));
498 const __m256 vk21xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 728)));
499 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
500 vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi21x89ABCDEF, vk21x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
501 vaccGHIJKLMNp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi21xGHIJKLMN, vk21xGHIJKLMN, vaccGHIJKLMNp1), _MM_FROUND_NO_EXC));
502 vaccOPQRSTUVp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi21xOPQRSTUV, vk21xOPQRSTUV, vaccOPQRSTUVp1), _MM_FROUND_NO_EXC));
503
504 const __m256 vi22x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i22));
505 const __m256 vi22x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i22 + 8)));
506 const __m256 vi22xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i22 + 16)));
507 const __m256 vi22xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i22 + 24)));
508 i22 += 32;
509
510 const __m256 vk22x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 736)));
511 const __m256 vk22x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 744)));
512 const __m256 vk22xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 752)));
513 const __m256 vk22xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 760)));
514 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
515 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x89ABCDEF, vk22x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
516 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22xGHIJKLMN, vk22xGHIJKLMN, vaccGHIJKLMNp0), _MM_FROUND_NO_EXC));
517 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22xOPQRSTUV, vk22xOPQRSTUV, vaccOPQRSTUVp0), _MM_FROUND_NO_EXC));
518
519 const __m256 vi23x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i23));
520 const __m256 vi23x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i23 + 8)));
521 const __m256 vi23xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i23 + 16)));
522 const __m256 vi23xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i23 + 24)));
523 i23 += 32;
524
525 const __m256 vk23x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 768)));
526 const __m256 vk23x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 776)));
527 const __m256 vk23xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 784)));
528 const __m256 vk23xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 792)));
529 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
530 vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi23x89ABCDEF, vk23x89ABCDEF, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
531 vaccGHIJKLMNp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi23xGHIJKLMN, vk23xGHIJKLMN, vaccGHIJKLMNp1), _MM_FROUND_NO_EXC));
532 vaccOPQRSTUVp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi23xOPQRSTUV, vk23xOPQRSTUV, vaccOPQRSTUVp1), _MM_FROUND_NO_EXC));
533
534 const __m256 vi24x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i24));
535 const __m256 vi24x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i24 + 8)));
536 const __m256 vi24xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i24 + 16)));
537 const __m256 vi24xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i24 + 24)));
538 i24 += 32;
539
540 const __m256 vk24x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 800)));
541 const __m256 vk24x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 808)));
542 const __m256 vk24xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 816)));
543 const __m256 vk24xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 824)));
544 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
545 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi24x89ABCDEF, vk24x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
546 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi24xGHIJKLMN, vk24xGHIJKLMN, vaccGHIJKLMNp0), _MM_FROUND_NO_EXC));
547 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi24xOPQRSTUV, vk24xOPQRSTUV, vaccOPQRSTUVp0), _MM_FROUND_NO_EXC));
548
549 w += 832;
550
551 // Add up all accumulators to vacc0123456789ABCDEFGHIJKLMNOPQRSTUVp0
552 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
553 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc89ABCDEFp0, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
554 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vaccGHIJKLMNp0, vaccGHIJKLMNp1), _MM_FROUND_NO_EXC));
555 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vaccOPQRSTUVp0, vaccOPQRSTUVp1), _MM_FROUND_NO_EXC));
556
557 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
558 __m256 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEFp0, vmin);
559 __m256 vaccGHIJKLMN = _mm256_max_ps(vaccGHIJKLMNp0, vmin);
560 __m256 vaccOPQRSTUV = _mm256_max_ps(vaccOPQRSTUVp0, vmin);
561 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
562 vacc89ABCDEF = _mm256_min_ps(vacc89ABCDEF, vmax);
563 vaccGHIJKLMN = _mm256_min_ps(vaccGHIJKLMN, vmax);
564 vaccOPQRSTUV = _mm256_min_ps(vaccOPQRSTUV, vmax);
565
566 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC));
567 _mm_storeu_si128((__m128i*) (o + 8), _mm256_cvtps_ph(vacc89ABCDEF, _MM_FROUND_NO_EXC));
568 _mm_storeu_si128((__m128i*) (o + 16), _mm256_cvtps_ph(vaccGHIJKLMN, _MM_FROUND_NO_EXC));
569 _mm_storeu_si128((__m128i*) (o + 24), _mm256_cvtps_ph(vaccOPQRSTUV, _MM_FROUND_NO_EXC));
570 o += 32;
571 }
572 for (; c >= 8; c -= 8) {
573 __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
574
575 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
576 i0 += 8;
577
578 const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 32)));
579 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
580
581 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
582 i1 += 8;
583
584 const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 64)));
585 __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
586
587 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
588 i2 += 8;
589
590 const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 96)));
591 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
592
593 const __m256 vi3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
594 i3 += 8;
595
596 const __m256 vk3x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 128)));
597 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
598
599 const __m256 vi4x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
600 i4 += 8;
601
602 const __m256 vk4x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 160)));
603 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
604
605 const __m256 vi5x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
606 i5 += 8;
607
608 const __m256 vk5x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 192)));
609 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
610
611 const __m256 vi6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
612 i6 += 8;
613
614 const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 224)));
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 i7 += 8;
619
620 const __m256 vk7x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 256)));
621 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
622
623 const __m256 vi8x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));
624 i8 += 8;
625
626 const __m256 vk8x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 288)));
627 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
628
629 const __m256 vi9x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i9));
630 i9 += 8;
631
632 const __m256 vk9x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 320)));
633 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
634
635 const __m256 vi10x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i10));
636 i10 += 8;
637
638 const __m256 vk10x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 352)));
639 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
640
641 const __m256 vi11x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i11));
642 i11 += 8;
643
644 const __m256 vk11x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 384)));
645 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
646
647 const __m256 vi12x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i12));
648 i12 += 8;
649
650 const __m256 vk12x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 416)));
651 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
652
653 const __m256 vi13x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i13));
654 i13 += 8;
655
656 const __m256 vk13x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 448)));
657 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
658
659 const __m256 vi14x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i14));
660 i14 += 8;
661
662 const __m256 vk14x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 480)));
663 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
664
665 const __m256 vi15x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i15));
666 i15 += 8;
667
668 const __m256 vk15x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 512)));
669 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
670
671 const __m256 vi16x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i16));
672 i16 += 8;
673
674 const __m256 vk16x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 544)));
675 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
676
677 const __m256 vi17x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i17));
678 i17 += 8;
679
680 const __m256 vk17x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 576)));
681 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
682
683 const __m256 vi18x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i18));
684 i18 += 8;
685
686 const __m256 vk18x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 608)));
687 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
688
689 const __m256 vi19x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i19));
690 i19 += 8;
691
692 const __m256 vk19x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 640)));
693 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
694
695 const __m256 vi20x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i20));
696 i20 += 8;
697
698 const __m256 vk20x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 672)));
699 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
700
701 const __m256 vi21x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i21));
702 i21 += 8;
703
704 const __m256 vk21x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 704)));
705 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
706
707 const __m256 vi22x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i22));
708 i22 += 8;
709
710 const __m256 vk22x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 736)));
711 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
712
713 const __m256 vi23x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i23));
714 i23 += 8;
715
716 const __m256 vk23x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 768)));
717 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
718
719 const __m256 vi24x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i24));
720 i24 += 8;
721
722 const __m256 vk24x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 800)));
723 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
724
725 w += 8;
726
727 // Add up all accumulators to vacc01234567p0
728 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
729
730 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
731 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
732
733 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC));
734 o += 8;
735 }
736 if XNN_UNLIKELY(c != 0) {
737 assert(c >= 1);
738 assert(c <= 7);
739
740 __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
741
742 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
743
744 const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 32)));
745 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
746
747 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
748
749 const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 64)));
750 __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
751
752 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
753
754 const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 96)));
755 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
756
757 const __m256 vi3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
758
759 const __m256 vk3x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 128)));
760 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
761
762 const __m256 vi4x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
763
764 const __m256 vk4x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 160)));
765 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
766
767 const __m256 vi5x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
768
769 const __m256 vk5x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 192)));
770 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
771
772 const __m256 vi6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
773
774 const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 224)));
775 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
776
777 const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
778
779 const __m256 vk7x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 256)));
780 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
781
782 const __m256 vi8x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));
783
784 const __m256 vk8x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 288)));
785 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
786
787 const __m256 vi9x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i9));
788
789 const __m256 vk9x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 320)));
790 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
791
792 const __m256 vi10x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i10));
793
794 const __m256 vk10x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 352)));
795 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
796
797 const __m256 vi11x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i11));
798
799 const __m256 vk11x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 384)));
800 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
801
802 const __m256 vi12x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i12));
803
804 const __m256 vk12x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 416)));
805 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
806
807 const __m256 vi13x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i13));
808
809 const __m256 vk13x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 448)));
810 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
811
812 const __m256 vi14x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i14));
813
814 const __m256 vk14x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 480)));
815 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
816
817 const __m256 vi15x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i15));
818
819 const __m256 vk15x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 512)));
820 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
821
822 const __m256 vi16x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i16));
823
824 const __m256 vk16x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 544)));
825 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
826
827 const __m256 vi17x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i17));
828
829 const __m256 vk17x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 576)));
830 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
831
832 const __m256 vi18x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i18));
833
834 const __m256 vk18x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 608)));
835 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
836
837 const __m256 vi19x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i19));
838
839 const __m256 vk19x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 640)));
840 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
841
842 const __m256 vi20x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i20));
843
844 const __m256 vk20x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 672)));
845 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
846
847 const __m256 vi21x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i21));
848
849 const __m256 vk21x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 704)));
850 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
851
852 const __m256 vi22x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i22));
853
854 const __m256 vk22x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 736)));
855 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
856
857 const __m256 vi23x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i23));
858
859 const __m256 vk23x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 768)));
860 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
861
862 const __m256 vi24x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i24));
863
864 const __m256 vk24x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 800)));
865 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
866
867 // Add up all accumulators to vacc01234567p0
868 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
869
870 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
871 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
872
873 __m128i vh01234567 = _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC);
874 if (c & 4) {
875 _mm_storel_epi64((__m128i*) o, vh01234567);
876 vh01234567 = _mm_unpackhi_epi64(vh01234567, vh01234567);
877 o += 4;
878 }
879 if (c & 2) {
880 _mm_storeu_si32(o, vh01234567);
881 vh01234567 = _mm_srli_epi64(vh01234567, 32);
882 o += 2;
883 }
884 if (c & 1) {
885 *o = (uint16_t) _mm_extract_epi16(vh01234567, 0);
886 o += 1;
887 }
888 }
889
890 o = (uint16_t*) ((uintptr_t) o + output_increment);
891 } while (--output_width != 0);
892 }
893