1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <smmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15
16
xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul16_add16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul16_add16(
18 size_t channels,
19 size_t output_width,
20 const int8_t** input,
21 const void* weights,
22 int8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const int8_t* zero,
27 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 do {
33 const int8_t* i0 = input[0];
34 assert(i0 != NULL);
35 if XNN_UNPREDICTABLE(i0 != zero) {
36 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
37 }
38 const int8_t* i1 = input[1];
39 assert(i1 != NULL);
40 if XNN_UNPREDICTABLE(i1 != zero) {
41 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
42 }
43 const int8_t* i2 = input[2];
44 assert(i2 != NULL);
45 if XNN_UNPREDICTABLE(i2 != zero) {
46 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
47 }
48 const int8_t* i3 = input[3];
49 assert(i3 != NULL);
50 if XNN_UNPREDICTABLE(i3 != zero) {
51 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
52 }
53 const int8_t* i4 = input[4];
54 assert(i4 != NULL);
55 if XNN_UNPREDICTABLE(i4 != zero) {
56 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
57 }
58 const int8_t* i5 = input[5];
59 assert(i5 != NULL);
60 if XNN_UNPREDICTABLE(i5 != zero) {
61 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
62 }
63 const int8_t* i6 = input[6];
64 assert(i6 != NULL);
65 if XNN_UNPREDICTABLE(i6 != zero) {
66 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
67 }
68 const int8_t* i7 = input[7];
69 assert(i7 != NULL);
70 if XNN_UNPREDICTABLE(i7 != zero) {
71 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
72 }
73 const int8_t* i8 = input[8];
74 assert(i8 != NULL);
75 if XNN_UNPREDICTABLE(i8 != zero) {
76 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
77 }
78 const int8_t* i9 = input[9];
79 assert(i9 != NULL);
80 if XNN_UNPREDICTABLE(i9 != zero) {
81 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
82 }
83 const int8_t* i10 = input[10];
84 assert(i10 != NULL);
85 if XNN_UNPREDICTABLE(i10 != zero) {
86 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
87 }
88 const int8_t* i11 = input[11];
89 assert(i11 != NULL);
90 if XNN_UNPREDICTABLE(i11 != zero) {
91 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
92 }
93 const int8_t* i12 = input[12];
94 assert(i12 != NULL);
95 if XNN_UNPREDICTABLE(i12 != zero) {
96 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
97 }
98 const int8_t* i13 = input[13];
99 assert(i13 != NULL);
100 if XNN_UNPREDICTABLE(i13 != zero) {
101 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
102 }
103 const int8_t* i14 = input[14];
104 assert(i14 != NULL);
105 if XNN_UNPREDICTABLE(i14 != zero) {
106 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
107 }
108 const int8_t* i15 = input[15];
109 assert(i15 != NULL);
110 if XNN_UNPREDICTABLE(i15 != zero) {
111 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
112 }
113 const int8_t* i16 = input[16];
114 assert(i16 != NULL);
115 if XNN_UNPREDICTABLE(i16 != zero) {
116 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
117 }
118 const int8_t* i17 = input[17];
119 assert(i17 != NULL);
120 if XNN_UNPREDICTABLE(i17 != zero) {
121 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
122 }
123 const int8_t* i18 = input[18];
124 assert(i18 != NULL);
125 if XNN_UNPREDICTABLE(i18 != zero) {
126 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
127 }
128 const int8_t* i19 = input[19];
129 assert(i19 != NULL);
130 if XNN_UNPREDICTABLE(i19 != zero) {
131 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
132 }
133 const int8_t* i20 = input[20];
134 assert(i20 != NULL);
135 if XNN_UNPREDICTABLE(i20 != zero) {
136 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
137 }
138 const int8_t* i21 = input[21];
139 assert(i21 != NULL);
140 if XNN_UNPREDICTABLE(i21 != zero) {
141 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
142 }
143 const int8_t* i22 = input[22];
144 assert(i22 != NULL);
145 if XNN_UNPREDICTABLE(i22 != zero) {
146 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
147 }
148 const int8_t* i23 = input[23];
149 assert(i23 != NULL);
150 if XNN_UNPREDICTABLE(i23 != zero) {
151 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
152 }
153 const int8_t* i24 = input[24];
154 assert(i24 != NULL);
155 if XNN_UNPREDICTABLE(i24 != zero) {
156 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
157 }
158 input = (const int8_t**) ((uintptr_t) input + input_stride);
159
160 size_t c = channels;
161 const void* w = weights;
162 for (; c >= 16; c -= 16) {
163 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
164 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
165 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
166 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
167
168
169 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
170 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
171 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
172 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
173 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
174 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
175 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
176 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
177 i0 += 16;
178
179
180 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
181 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
182
183
184 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
185 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
186 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
187 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
188 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
189 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
190 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
191 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
192 i1 += 16;
193
194
195 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
196 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF));
197
198 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
199 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
200 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
201 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
202
203 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
204 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
205 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
206 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
207 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
208 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
209 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
210 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
211 i2 += 16;
212
213
214 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
215 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
216
217
218 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
219 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
220 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
221 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
222 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
223 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
224 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
225 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
226 i3 += 16;
227
228
229 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
230 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF));
231
232 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
233 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
234 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
235 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
236
237 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
238 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
239 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
240 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
241 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
242 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
243 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
244 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
245 i4 += 16;
246
247
248 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
249 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
250
251
252 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
253 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
254 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
255 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
256 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
257 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
258 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
259 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
260 i5 += 16;
261
262
263 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
264 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF));
265
266 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
267 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
268 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
269 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
270
271 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
272 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
273 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
274 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
275 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
276 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
277 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
278 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
279 i6 += 16;
280
281
282 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
283 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
284
285
286 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
287 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
288 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
289 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
290 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
291 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
292 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
293 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
294 i7 += 16;
295
296
297 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
298 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF));
299
300 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
301 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
302 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
303 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
304
305 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
306 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
307 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
308 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
309 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
310 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
311 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
312 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
313 i8 += 16;
314
315
316 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
317 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
318
319
320 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
321 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
322 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
323 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
324 const __m128i vi9x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i9 + 8));
325 const __m128i vxi9x89ABCDEF = _mm_cvtepi8_epi16(vi9x89ABCDEF);
326 const __m128i vk9x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t)));
327 const __m128i vxk9x89ABCDEF = _mm_cvtepi8_epi16(vk9x89ABCDEF);
328 i9 += 16;
329
330
331 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
332 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF));
333
334 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
335 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
336 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
337 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
338
339 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
340 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
341 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)));
342 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
343 const __m128i vi10x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i10 + 8));
344 const __m128i vxi10x89ABCDEF = _mm_cvtepi8_epi16(vi10x89ABCDEF);
345 const __m128i vk10x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t)));
346 const __m128i vxk10x89ABCDEF = _mm_cvtepi8_epi16(vk10x89ABCDEF);
347 i10 += 16;
348
349
350 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
351 vprod89ABCDEF = _mm_mullo_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
352
353
354 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
355 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
356 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)));
357 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
358 const __m128i vi11x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i11 + 8));
359 const __m128i vxi11x89ABCDEF = _mm_cvtepi8_epi16(vi11x89ABCDEF);
360 const __m128i vk11x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t)));
361 const __m128i vxk11x89ABCDEF = _mm_cvtepi8_epi16(vk11x89ABCDEF);
362 i11 += 16;
363
364
365 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
366 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF));
367
368 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
369 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
370 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
371 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
372
373 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
374 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
375 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)));
376 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
377 const __m128i vi12x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i12 + 8));
378 const __m128i vxi12x89ABCDEF = _mm_cvtepi8_epi16(vi12x89ABCDEF);
379 const __m128i vk12x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t)));
380 const __m128i vxk12x89ABCDEF = _mm_cvtepi8_epi16(vk12x89ABCDEF);
381 i12 += 16;
382
383
384 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
385 vprod89ABCDEF = _mm_mullo_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
386
387
388 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
389 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
390 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)));
391 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
392 const __m128i vi13x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i13 + 8));
393 const __m128i vxi13x89ABCDEF = _mm_cvtepi8_epi16(vi13x89ABCDEF);
394 const __m128i vk13x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t)));
395 const __m128i vxk13x89ABCDEF = _mm_cvtepi8_epi16(vk13x89ABCDEF);
396 i13 += 16;
397
398
399 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
400 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF));
401
402 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
403 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
404 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
405 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
406
407 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
408 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
409 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)));
410 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
411 const __m128i vi14x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i14 + 8));
412 const __m128i vxi14x89ABCDEF = _mm_cvtepi8_epi16(vi14x89ABCDEF);
413 const __m128i vk14x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t)));
414 const __m128i vxk14x89ABCDEF = _mm_cvtepi8_epi16(vk14x89ABCDEF);
415 i14 += 16;
416
417
418 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
419 vprod89ABCDEF = _mm_mullo_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
420
421
422 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
423 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
424 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)));
425 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
426 const __m128i vi15x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i15 + 8));
427 const __m128i vxi15x89ABCDEF = _mm_cvtepi8_epi16(vi15x89ABCDEF);
428 const __m128i vk15x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t)));
429 const __m128i vxk15x89ABCDEF = _mm_cvtepi8_epi16(vk15x89ABCDEF);
430 i15 += 16;
431
432
433 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
434 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF));
435
436 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
437 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
438 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
439 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
440
441 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
442 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
443 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)));
444 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
445 const __m128i vi16x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i16 + 8));
446 const __m128i vxi16x89ABCDEF = _mm_cvtepi8_epi16(vi16x89ABCDEF);
447 const __m128i vk16x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t)));
448 const __m128i vxk16x89ABCDEF = _mm_cvtepi8_epi16(vk16x89ABCDEF);
449 i16 += 16;
450
451
452 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
453 vprod89ABCDEF = _mm_mullo_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
454
455
456 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
457 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
458 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)));
459 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
460 const __m128i vi17x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i17 + 8));
461 const __m128i vxi17x89ABCDEF = _mm_cvtepi8_epi16(vi17x89ABCDEF);
462 const __m128i vk17x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t)));
463 const __m128i vxk17x89ABCDEF = _mm_cvtepi8_epi16(vk17x89ABCDEF);
464 i17 += 16;
465
466
467 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
468 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF));
469
470 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
471 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
472 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
473 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
474
475 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
476 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
477 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)));
478 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
479 const __m128i vi18x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i18 + 8));
480 const __m128i vxi18x89ABCDEF = _mm_cvtepi8_epi16(vi18x89ABCDEF);
481 const __m128i vk18x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t)));
482 const __m128i vxk18x89ABCDEF = _mm_cvtepi8_epi16(vk18x89ABCDEF);
483 i18 += 16;
484
485
486 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
487 vprod89ABCDEF = _mm_mullo_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
488
489
490 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
491 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
492 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)));
493 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
494 const __m128i vi19x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i19 + 8));
495 const __m128i vxi19x89ABCDEF = _mm_cvtepi8_epi16(vi19x89ABCDEF);
496 const __m128i vk19x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t)));
497 const __m128i vxk19x89ABCDEF = _mm_cvtepi8_epi16(vk19x89ABCDEF);
498 i19 += 16;
499
500
501 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
502 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF));
503
504 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
505 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
506 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
507 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
508
509 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
510 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
511 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)));
512 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
513 const __m128i vi20x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i20 + 8));
514 const __m128i vxi20x89ABCDEF = _mm_cvtepi8_epi16(vi20x89ABCDEF);
515 const __m128i vk20x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t)));
516 const __m128i vxk20x89ABCDEF = _mm_cvtepi8_epi16(vk20x89ABCDEF);
517 i20 += 16;
518
519
520 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
521 vprod89ABCDEF = _mm_mullo_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
522
523
524 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
525 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
526 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)));
527 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
528 const __m128i vi21x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i21 + 8));
529 const __m128i vxi21x89ABCDEF = _mm_cvtepi8_epi16(vi21x89ABCDEF);
530 const __m128i vk21x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t)));
531 const __m128i vxk21x89ABCDEF = _mm_cvtepi8_epi16(vk21x89ABCDEF);
532 i21 += 16;
533
534
535 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
536 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF));
537
538 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
539 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
540 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
541 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
542
543 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
544 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
545 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)));
546 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
547 const __m128i vi22x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i22 + 8));
548 const __m128i vxi22x89ABCDEF = _mm_cvtepi8_epi16(vi22x89ABCDEF);
549 const __m128i vk22x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t)));
550 const __m128i vxk22x89ABCDEF = _mm_cvtepi8_epi16(vk22x89ABCDEF);
551 i22 += 16;
552
553
554 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
555 vprod89ABCDEF = _mm_mullo_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
556
557
558 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
559 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
560 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)));
561 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
562 const __m128i vi23x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i23 + 8));
563 const __m128i vxi23x89ABCDEF = _mm_cvtepi8_epi16(vi23x89ABCDEF);
564 const __m128i vk23x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t)));
565 const __m128i vxk23x89ABCDEF = _mm_cvtepi8_epi16(vk23x89ABCDEF);
566 i23 += 16;
567
568
569 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
570 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF));
571
572 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
573 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
574 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
575 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
576
577 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
578 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
579 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)));
580 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
581 const __m128i vi24x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i24 + 8));
582 const __m128i vxi24x89ABCDEF = _mm_cvtepi8_epi16(vi24x89ABCDEF);
583 const __m128i vk24x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t)));
584 const __m128i vxk24x89ABCDEF = _mm_cvtepi8_epi16(vk24x89ABCDEF);
585 i24 += 16;
586
587
588 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
589 vprod89ABCDEF = _mm_mullo_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
590
591 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
592 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
593 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
594 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
595
596 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
597
598 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
599 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
600 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
601 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
602
603 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
604 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
605 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
606 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
607 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
608
609 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
610 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
611 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
612 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
613 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
614
615 vacc0123 = _mm_cvtps_epi32(vscaled0123);
616 vacc4567 = _mm_cvtps_epi32(vscaled4567);
617 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
618 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
619
620 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
621 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
622 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
623
624
625 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
626
627 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
628 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
629
630 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
631 output += 16;
632 }
633 if XNN_UNLIKELY(c != 0) {
634 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
635 do {
636 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
637 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
638
639
640 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
641 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
642 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
643 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
644 i0 += 8;
645
646
647 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
648
649
650 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
651 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
652 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
653 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
654 i1 += 8;
655
656
657 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
658
659 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
660 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
661
662 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
663 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
664 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
665 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
666 i2 += 8;
667
668
669 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
670
671
672 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
673 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
674 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
675 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
676 i3 += 8;
677
678
679 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
680
681 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
682 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
683
684 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
685 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
686 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
687 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
688 i4 += 8;
689
690
691 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
692
693
694 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
695 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
696 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
697 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
698 i5 += 8;
699
700
701 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
702
703 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
704 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
705
706 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
707 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
708 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
709 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
710 i6 += 8;
711
712
713 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
714
715
716 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
717 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
718 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
719 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
720 i7 += 8;
721
722
723 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
724
725 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
726 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
727
728 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
729 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
730 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
731 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
732 i8 += 8;
733
734
735 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
736
737
738 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
739 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
740 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
741 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
742 i9 += 8;
743
744
745 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
746
747 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
748 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
749
750 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
751 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
752 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) (k + 160));
753 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
754 i10 += 8;
755
756
757 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
758
759
760 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
761 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
762 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) (k + 176));
763 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
764 i11 += 8;
765
766
767 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
768
769 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
770 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
771
772 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
773 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
774 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
775 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
776 i12 += 8;
777
778
779 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
780
781
782 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
783 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
784 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) (k + 208));
785 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
786 i13 += 8;
787
788
789 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
790
791 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
792 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
793
794 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
795 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
796 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) (k + 224));
797 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
798 i14 += 8;
799
800
801 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
802
803
804 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
805 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
806 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) (k + 240));
807 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
808 i15 += 8;
809
810
811 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
812
813 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
814 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
815
816 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
817 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
818 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) (k + 256));
819 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
820 i16 += 8;
821
822
823 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
824
825
826 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
827 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
828 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) (k + 272));
829 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
830 i17 += 8;
831
832
833 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
834
835 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
836 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
837
838 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
839 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
840 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) (k + 288));
841 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
842 i18 += 8;
843
844
845 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
846
847
848 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
849 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
850 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) (k + 304));
851 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
852 i19 += 8;
853
854
855 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
856
857 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
858 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
859
860 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
861 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
862 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) (k + 320));
863 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
864 i20 += 8;
865
866
867 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
868
869
870 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
871 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
872 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) (k + 336));
873 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
874 i21 += 8;
875
876
877 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
878
879 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
880 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
881
882 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
883 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
884 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) (k + 352));
885 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
886 i22 += 8;
887
888
889 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
890
891
892 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
893 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
894 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) (k + 368));
895 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
896 i23 += 8;
897
898
899 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
900
901 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
902 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
903
904 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
905 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
906 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) (k + 384));
907 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
908 i24 += 8;
909
910
911 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
912
913 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
914 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
915
916 k += 8;
917
918 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
919 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
920
921 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
922 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
923 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
924
925 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
926 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
927 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
928
929 vacc0123 = _mm_cvtps_epi32(vscaled0123);
930 vacc4567 = _mm_cvtps_epi32(vscaled4567);
931
932 w = (const void*) ((const int32_t*) w + 8);
933
934 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
935 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
936
937
938 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
939
940 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
941
942 if XNN_LIKELY(c >= 8) {
943 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
944 output += 8;
945 c -= 8;
946 } else {
947 if (c & 4) {
948 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
949 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
950 output += 4;
951 }
952 if (c & 2) {
953 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
954 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
955 output += 2;
956 }
957 if (c & 1) {
958 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
959 output += 1;
960 }
961 c = 0;
962 }
963 } while (c != 0);
964 }
965
966 output = (int8_t*) ((uintptr_t) output + output_increment);
967 } while (--output_width != 0);
968 }
969