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_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul16(size_t channels,size_t output_width,const uint8_t ** input,const void * weights,uint8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul16(
18 size_t channels,
19 size_t output_width,
20 const uint8_t** input,
21 const void* weights,
22 uint8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const uint8_t* zero,
27 const union xnn_qu8_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 uint8_t* i0 = input[0];
34 assert(i0 != NULL);
35 if XNN_UNPREDICTABLE(i0 != zero) {
36 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
37 }
38 const uint8_t* i1 = input[1];
39 assert(i1 != NULL);
40 if XNN_UNPREDICTABLE(i1 != zero) {
41 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
42 }
43 const uint8_t* i2 = input[2];
44 assert(i2 != NULL);
45 if XNN_UNPREDICTABLE(i2 != zero) {
46 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
47 }
48 const uint8_t* i3 = input[3];
49 assert(i3 != NULL);
50 if XNN_UNPREDICTABLE(i3 != zero) {
51 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
52 }
53 const uint8_t* i4 = input[4];
54 assert(i4 != NULL);
55 if XNN_UNPREDICTABLE(i4 != zero) {
56 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
57 }
58 const uint8_t* i5 = input[5];
59 assert(i5 != NULL);
60 if XNN_UNPREDICTABLE(i5 != zero) {
61 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
62 }
63 const uint8_t* i6 = input[6];
64 assert(i6 != NULL);
65 if XNN_UNPREDICTABLE(i6 != zero) {
66 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
67 }
68 const uint8_t* i7 = input[7];
69 assert(i7 != NULL);
70 if XNN_UNPREDICTABLE(i7 != zero) {
71 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
72 }
73 const uint8_t* i8 = input[8];
74 assert(i8 != NULL);
75 if XNN_UNPREDICTABLE(i8 != zero) {
76 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
77 }
78 const uint8_t* i9 = input[9];
79 assert(i9 != NULL);
80 if XNN_UNPREDICTABLE(i9 != zero) {
81 i9 = (const uint8_t*) ((uintptr_t) i9 + input_offset);
82 }
83 const uint8_t* i10 = input[10];
84 assert(i10 != NULL);
85 if XNN_UNPREDICTABLE(i10 != zero) {
86 i10 = (const uint8_t*) ((uintptr_t) i10 + input_offset);
87 }
88 const uint8_t* i11 = input[11];
89 assert(i11 != NULL);
90 if XNN_UNPREDICTABLE(i11 != zero) {
91 i11 = (const uint8_t*) ((uintptr_t) i11 + input_offset);
92 }
93 const uint8_t* i12 = input[12];
94 assert(i12 != NULL);
95 if XNN_UNPREDICTABLE(i12 != zero) {
96 i12 = (const uint8_t*) ((uintptr_t) i12 + input_offset);
97 }
98 const uint8_t* i13 = input[13];
99 assert(i13 != NULL);
100 if XNN_UNPREDICTABLE(i13 != zero) {
101 i13 = (const uint8_t*) ((uintptr_t) i13 + input_offset);
102 }
103 const uint8_t* i14 = input[14];
104 assert(i14 != NULL);
105 if XNN_UNPREDICTABLE(i14 != zero) {
106 i14 = (const uint8_t*) ((uintptr_t) i14 + input_offset);
107 }
108 const uint8_t* i15 = input[15];
109 assert(i15 != NULL);
110 if XNN_UNPREDICTABLE(i15 != zero) {
111 i15 = (const uint8_t*) ((uintptr_t) i15 + input_offset);
112 }
113 const uint8_t* i16 = input[16];
114 assert(i16 != NULL);
115 if XNN_UNPREDICTABLE(i16 != zero) {
116 i16 = (const uint8_t*) ((uintptr_t) i16 + input_offset);
117 }
118 const uint8_t* i17 = input[17];
119 assert(i17 != NULL);
120 if XNN_UNPREDICTABLE(i17 != zero) {
121 i17 = (const uint8_t*) ((uintptr_t) i17 + input_offset);
122 }
123 const uint8_t* i18 = input[18];
124 assert(i18 != NULL);
125 if XNN_UNPREDICTABLE(i18 != zero) {
126 i18 = (const uint8_t*) ((uintptr_t) i18 + input_offset);
127 }
128 const uint8_t* i19 = input[19];
129 assert(i19 != NULL);
130 if XNN_UNPREDICTABLE(i19 != zero) {
131 i19 = (const uint8_t*) ((uintptr_t) i19 + input_offset);
132 }
133 const uint8_t* i20 = input[20];
134 assert(i20 != NULL);
135 if XNN_UNPREDICTABLE(i20 != zero) {
136 i20 = (const uint8_t*) ((uintptr_t) i20 + input_offset);
137 }
138 const uint8_t* i21 = input[21];
139 assert(i21 != NULL);
140 if XNN_UNPREDICTABLE(i21 != zero) {
141 i21 = (const uint8_t*) ((uintptr_t) i21 + input_offset);
142 }
143 const uint8_t* i22 = input[22];
144 assert(i22 != NULL);
145 if XNN_UNPREDICTABLE(i22 != zero) {
146 i22 = (const uint8_t*) ((uintptr_t) i22 + input_offset);
147 }
148 const uint8_t* i23 = input[23];
149 assert(i23 != NULL);
150 if XNN_UNPREDICTABLE(i23 != zero) {
151 i23 = (const uint8_t*) ((uintptr_t) i23 + input_offset);
152 }
153 const uint8_t* i24 = input[24];
154 assert(i24 != NULL);
155 if XNN_UNPREDICTABLE(i24 != zero) {
156 i24 = (const uint8_t*) ((uintptr_t) i24 + input_offset);
157 }
158 input = (const uint8_t**) ((uintptr_t) input + input_stride);
159
160 size_t c = channels;
161 const void* w = weights;
162 const __m128i vk_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
163 for (; c >= 16; c -= 16) {
164 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
165 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
166 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
167 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
168
169
170 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
171 const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(vi0x01234567);
172 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
173 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk0x01234567), vk_zero_point);
174 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
175 const __m128i vxi0x89ABCDEF = _mm_cvtepu8_epi16(vi0x89ABCDEF);
176 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
177 const __m128i vxk0x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk0x89ABCDEF), vk_zero_point);
178 i0 += 16;
179
180
181 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
182 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
183 const __m128i vprod0x89ABCDEFlo = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
184 const __m128i vprod0x89ABCDEFhi = _mm_mulhi_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
185
186 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
187 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
188 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
189 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
190
191 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
192 const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(vi1x01234567);
193 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
194 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk1x01234567), vk_zero_point);
195 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
196 const __m128i vxi1x89ABCDEF = _mm_cvtepu8_epi16(vi1x89ABCDEF);
197 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
198 const __m128i vxk1x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk1x89ABCDEF), vk_zero_point);
199 i1 += 16;
200
201
202 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
203 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
204 const __m128i vprod1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
205 const __m128i vprod1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
206
207 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
208 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
209 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
210 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
211
212 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
213 const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(vi2x01234567);
214 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
215 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk2x01234567), vk_zero_point);
216 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
217 const __m128i vxi2x89ABCDEF = _mm_cvtepu8_epi16(vi2x89ABCDEF);
218 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
219 const __m128i vxk2x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk2x89ABCDEF), vk_zero_point);
220 i2 += 16;
221
222
223 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
224 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
225 const __m128i vprod2x89ABCDEFlo = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
226 const __m128i vprod2x89ABCDEFhi = _mm_mulhi_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
227
228 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
229 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
230 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
231 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
232
233 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
234 const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(vi3x01234567);
235 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
236 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk3x01234567), vk_zero_point);
237 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
238 const __m128i vxi3x89ABCDEF = _mm_cvtepu8_epi16(vi3x89ABCDEF);
239 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
240 const __m128i vxk3x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk3x89ABCDEF), vk_zero_point);
241 i3 += 16;
242
243
244 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
245 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
246 const __m128i vprod3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
247 const __m128i vprod3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
248
249 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
250 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
251 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
252 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
253
254 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
255 const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(vi4x01234567);
256 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
257 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk4x01234567), vk_zero_point);
258 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
259 const __m128i vxi4x89ABCDEF = _mm_cvtepu8_epi16(vi4x89ABCDEF);
260 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
261 const __m128i vxk4x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk4x89ABCDEF), vk_zero_point);
262 i4 += 16;
263
264
265 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
266 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
267 const __m128i vprod4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
268 const __m128i vprod4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
269
270 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
271 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
272 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
273 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
274
275 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
276 const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(vi5x01234567);
277 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
278 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point);
279 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
280 const __m128i vxi5x89ABCDEF = _mm_cvtepu8_epi16(vi5x89ABCDEF);
281 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
282 const __m128i vxk5x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x89ABCDEF), vk_zero_point);
283 i5 += 16;
284
285
286 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
287 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
288 const __m128i vprod5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
289 const __m128i vprod5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
290
291 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
292 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
293 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
294 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
295
296 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
297 const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(vi6x01234567);
298 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
299 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk6x01234567), vk_zero_point);
300 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
301 const __m128i vxi6x89ABCDEF = _mm_cvtepu8_epi16(vi6x89ABCDEF);
302 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
303 const __m128i vxk6x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk6x89ABCDEF), vk_zero_point);
304 i6 += 16;
305
306
307 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
308 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
309 const __m128i vprod6x89ABCDEFlo = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
310 const __m128i vprod6x89ABCDEFhi = _mm_mulhi_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
311
312 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
313 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
314 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
315 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
316
317 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
318 const __m128i vxi7x01234567 = _mm_cvtepu8_epi16(vi7x01234567);
319 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
320 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk7x01234567), vk_zero_point);
321 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
322 const __m128i vxi7x89ABCDEF = _mm_cvtepu8_epi16(vi7x89ABCDEF);
323 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
324 const __m128i vxk7x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk7x89ABCDEF), vk_zero_point);
325 i7 += 16;
326
327
328 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
329 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
330 const __m128i vprod7x89ABCDEFlo = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
331 const __m128i vprod7x89ABCDEFhi = _mm_mulhi_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
332
333 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
334 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
335 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
336 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
337
338 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
339 const __m128i vxi8x01234567 = _mm_cvtepu8_epi16(vi8x01234567);
340 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
341 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk8x01234567), vk_zero_point);
342 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
343 const __m128i vxi8x89ABCDEF = _mm_cvtepu8_epi16(vi8x89ABCDEF);
344 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
345 const __m128i vxk8x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk8x89ABCDEF), vk_zero_point);
346 i8 += 16;
347
348
349 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
350 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
351 const __m128i vprod8x89ABCDEFlo = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
352 const __m128i vprod8x89ABCDEFhi = _mm_mulhi_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
353
354 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
355 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
356 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
357 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
358
359 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
360 const __m128i vxi9x01234567 = _mm_cvtepu8_epi16(vi9x01234567);
361 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(uint8_t)));
362 const __m128i vxk9x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk9x01234567), vk_zero_point);
363 const __m128i vi9x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i9 + 8));
364 const __m128i vxi9x89ABCDEF = _mm_cvtepu8_epi16(vi9x89ABCDEF);
365 const __m128i vk9x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(uint8_t)));
366 const __m128i vxk9x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk9x89ABCDEF), vk_zero_point);
367 i9 += 16;
368
369
370 const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
371 const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567);
372 const __m128i vprod9x89ABCDEFlo = _mm_mullo_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF);
373 const __m128i vprod9x89ABCDEFhi = _mm_mulhi_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF);
374
375 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi));
376 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi));
377 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod9x89ABCDEFlo, vprod9x89ABCDEFhi));
378 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod9x89ABCDEFlo, vprod9x89ABCDEFhi));
379
380 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
381 const __m128i vxi10x01234567 = _mm_cvtepu8_epi16(vi10x01234567);
382 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(uint8_t)));
383 const __m128i vxk10x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk10x01234567), vk_zero_point);
384 const __m128i vi10x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i10 + 8));
385 const __m128i vxi10x89ABCDEF = _mm_cvtepu8_epi16(vi10x89ABCDEF);
386 const __m128i vk10x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(uint8_t)));
387 const __m128i vxk10x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk10x89ABCDEF), vk_zero_point);
388 i10 += 16;
389
390
391 const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
392 const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567);
393 const __m128i vprod10x89ABCDEFlo = _mm_mullo_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
394 const __m128i vprod10x89ABCDEFhi = _mm_mulhi_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
395
396 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi));
397 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi));
398 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod10x89ABCDEFlo, vprod10x89ABCDEFhi));
399 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod10x89ABCDEFlo, vprod10x89ABCDEFhi));
400
401 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
402 const __m128i vxi11x01234567 = _mm_cvtepu8_epi16(vi11x01234567);
403 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(uint8_t)));
404 const __m128i vxk11x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk11x01234567), vk_zero_point);
405 const __m128i vi11x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i11 + 8));
406 const __m128i vxi11x89ABCDEF = _mm_cvtepu8_epi16(vi11x89ABCDEF);
407 const __m128i vk11x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(uint8_t)));
408 const __m128i vxk11x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk11x89ABCDEF), vk_zero_point);
409 i11 += 16;
410
411
412 const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
413 const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567);
414 const __m128i vprod11x89ABCDEFlo = _mm_mullo_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF);
415 const __m128i vprod11x89ABCDEFhi = _mm_mulhi_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF);
416
417 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi));
418 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi));
419 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod11x89ABCDEFlo, vprod11x89ABCDEFhi));
420 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod11x89ABCDEFlo, vprod11x89ABCDEFhi));
421
422 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
423 const __m128i vxi12x01234567 = _mm_cvtepu8_epi16(vi12x01234567);
424 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(uint8_t)));
425 const __m128i vxk12x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk12x01234567), vk_zero_point);
426 const __m128i vi12x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i12 + 8));
427 const __m128i vxi12x89ABCDEF = _mm_cvtepu8_epi16(vi12x89ABCDEF);
428 const __m128i vk12x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(uint8_t)));
429 const __m128i vxk12x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk12x89ABCDEF), vk_zero_point);
430 i12 += 16;
431
432
433 const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
434 const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567);
435 const __m128i vprod12x89ABCDEFlo = _mm_mullo_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
436 const __m128i vprod12x89ABCDEFhi = _mm_mulhi_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
437
438 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi));
439 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi));
440 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod12x89ABCDEFlo, vprod12x89ABCDEFhi));
441 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod12x89ABCDEFlo, vprod12x89ABCDEFhi));
442
443 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
444 const __m128i vxi13x01234567 = _mm_cvtepu8_epi16(vi13x01234567);
445 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(uint8_t)));
446 const __m128i vxk13x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk13x01234567), vk_zero_point);
447 const __m128i vi13x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i13 + 8));
448 const __m128i vxi13x89ABCDEF = _mm_cvtepu8_epi16(vi13x89ABCDEF);
449 const __m128i vk13x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(uint8_t)));
450 const __m128i vxk13x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk13x89ABCDEF), vk_zero_point);
451 i13 += 16;
452
453
454 const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
455 const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567);
456 const __m128i vprod13x89ABCDEFlo = _mm_mullo_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF);
457 const __m128i vprod13x89ABCDEFhi = _mm_mulhi_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF);
458
459 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi));
460 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi));
461 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod13x89ABCDEFlo, vprod13x89ABCDEFhi));
462 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod13x89ABCDEFlo, vprod13x89ABCDEFhi));
463
464 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
465 const __m128i vxi14x01234567 = _mm_cvtepu8_epi16(vi14x01234567);
466 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(uint8_t)));
467 const __m128i vxk14x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk14x01234567), vk_zero_point);
468 const __m128i vi14x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i14 + 8));
469 const __m128i vxi14x89ABCDEF = _mm_cvtepu8_epi16(vi14x89ABCDEF);
470 const __m128i vk14x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(uint8_t)));
471 const __m128i vxk14x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk14x89ABCDEF), vk_zero_point);
472 i14 += 16;
473
474
475 const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
476 const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567);
477 const __m128i vprod14x89ABCDEFlo = _mm_mullo_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
478 const __m128i vprod14x89ABCDEFhi = _mm_mulhi_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
479
480 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi));
481 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi));
482 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod14x89ABCDEFlo, vprod14x89ABCDEFhi));
483 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod14x89ABCDEFlo, vprod14x89ABCDEFhi));
484
485 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
486 const __m128i vxi15x01234567 = _mm_cvtepu8_epi16(vi15x01234567);
487 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(uint8_t)));
488 const __m128i vxk15x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk15x01234567), vk_zero_point);
489 const __m128i vi15x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i15 + 8));
490 const __m128i vxi15x89ABCDEF = _mm_cvtepu8_epi16(vi15x89ABCDEF);
491 const __m128i vk15x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(uint8_t)));
492 const __m128i vxk15x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk15x89ABCDEF), vk_zero_point);
493 i15 += 16;
494
495
496 const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
497 const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567);
498 const __m128i vprod15x89ABCDEFlo = _mm_mullo_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF);
499 const __m128i vprod15x89ABCDEFhi = _mm_mulhi_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF);
500
501 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi));
502 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi));
503 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod15x89ABCDEFlo, vprod15x89ABCDEFhi));
504 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod15x89ABCDEFlo, vprod15x89ABCDEFhi));
505
506 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
507 const __m128i vxi16x01234567 = _mm_cvtepu8_epi16(vi16x01234567);
508 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(uint8_t)));
509 const __m128i vxk16x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk16x01234567), vk_zero_point);
510 const __m128i vi16x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i16 + 8));
511 const __m128i vxi16x89ABCDEF = _mm_cvtepu8_epi16(vi16x89ABCDEF);
512 const __m128i vk16x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(uint8_t)));
513 const __m128i vxk16x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk16x89ABCDEF), vk_zero_point);
514 i16 += 16;
515
516
517 const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
518 const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567);
519 const __m128i vprod16x89ABCDEFlo = _mm_mullo_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
520 const __m128i vprod16x89ABCDEFhi = _mm_mulhi_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
521
522 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi));
523 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi));
524 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod16x89ABCDEFlo, vprod16x89ABCDEFhi));
525 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod16x89ABCDEFlo, vprod16x89ABCDEFhi));
526
527 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
528 const __m128i vxi17x01234567 = _mm_cvtepu8_epi16(vi17x01234567);
529 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(uint8_t)));
530 const __m128i vxk17x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk17x01234567), vk_zero_point);
531 const __m128i vi17x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i17 + 8));
532 const __m128i vxi17x89ABCDEF = _mm_cvtepu8_epi16(vi17x89ABCDEF);
533 const __m128i vk17x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(uint8_t)));
534 const __m128i vxk17x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk17x89ABCDEF), vk_zero_point);
535 i17 += 16;
536
537
538 const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
539 const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567);
540 const __m128i vprod17x89ABCDEFlo = _mm_mullo_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF);
541 const __m128i vprod17x89ABCDEFhi = _mm_mulhi_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF);
542
543 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi));
544 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi));
545 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod17x89ABCDEFlo, vprod17x89ABCDEFhi));
546 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod17x89ABCDEFlo, vprod17x89ABCDEFhi));
547
548 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
549 const __m128i vxi18x01234567 = _mm_cvtepu8_epi16(vi18x01234567);
550 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(uint8_t)));
551 const __m128i vxk18x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk18x01234567), vk_zero_point);
552 const __m128i vi18x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i18 + 8));
553 const __m128i vxi18x89ABCDEF = _mm_cvtepu8_epi16(vi18x89ABCDEF);
554 const __m128i vk18x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(uint8_t)));
555 const __m128i vxk18x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk18x89ABCDEF), vk_zero_point);
556 i18 += 16;
557
558
559 const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
560 const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567);
561 const __m128i vprod18x89ABCDEFlo = _mm_mullo_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
562 const __m128i vprod18x89ABCDEFhi = _mm_mulhi_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
563
564 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi));
565 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi));
566 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod18x89ABCDEFlo, vprod18x89ABCDEFhi));
567 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod18x89ABCDEFlo, vprod18x89ABCDEFhi));
568
569 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
570 const __m128i vxi19x01234567 = _mm_cvtepu8_epi16(vi19x01234567);
571 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(uint8_t)));
572 const __m128i vxk19x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk19x01234567), vk_zero_point);
573 const __m128i vi19x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i19 + 8));
574 const __m128i vxi19x89ABCDEF = _mm_cvtepu8_epi16(vi19x89ABCDEF);
575 const __m128i vk19x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(uint8_t)));
576 const __m128i vxk19x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk19x89ABCDEF), vk_zero_point);
577 i19 += 16;
578
579
580 const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
581 const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567);
582 const __m128i vprod19x89ABCDEFlo = _mm_mullo_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF);
583 const __m128i vprod19x89ABCDEFhi = _mm_mulhi_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF);
584
585 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi));
586 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi));
587 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod19x89ABCDEFlo, vprod19x89ABCDEFhi));
588 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod19x89ABCDEFlo, vprod19x89ABCDEFhi));
589
590 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
591 const __m128i vxi20x01234567 = _mm_cvtepu8_epi16(vi20x01234567);
592 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(uint8_t)));
593 const __m128i vxk20x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk20x01234567), vk_zero_point);
594 const __m128i vi20x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i20 + 8));
595 const __m128i vxi20x89ABCDEF = _mm_cvtepu8_epi16(vi20x89ABCDEF);
596 const __m128i vk20x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(uint8_t)));
597 const __m128i vxk20x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk20x89ABCDEF), vk_zero_point);
598 i20 += 16;
599
600
601 const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
602 const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567);
603 const __m128i vprod20x89ABCDEFlo = _mm_mullo_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
604 const __m128i vprod20x89ABCDEFhi = _mm_mulhi_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
605
606 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi));
607 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi));
608 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod20x89ABCDEFlo, vprod20x89ABCDEFhi));
609 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod20x89ABCDEFlo, vprod20x89ABCDEFhi));
610
611 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
612 const __m128i vxi21x01234567 = _mm_cvtepu8_epi16(vi21x01234567);
613 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(uint8_t)));
614 const __m128i vxk21x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk21x01234567), vk_zero_point);
615 const __m128i vi21x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i21 + 8));
616 const __m128i vxi21x89ABCDEF = _mm_cvtepu8_epi16(vi21x89ABCDEF);
617 const __m128i vk21x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(uint8_t)));
618 const __m128i vxk21x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk21x89ABCDEF), vk_zero_point);
619 i21 += 16;
620
621
622 const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
623 const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567);
624 const __m128i vprod21x89ABCDEFlo = _mm_mullo_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF);
625 const __m128i vprod21x89ABCDEFhi = _mm_mulhi_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF);
626
627 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi));
628 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi));
629 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod21x89ABCDEFlo, vprod21x89ABCDEFhi));
630 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod21x89ABCDEFlo, vprod21x89ABCDEFhi));
631
632 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
633 const __m128i vxi22x01234567 = _mm_cvtepu8_epi16(vi22x01234567);
634 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(uint8_t)));
635 const __m128i vxk22x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk22x01234567), vk_zero_point);
636 const __m128i vi22x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i22 + 8));
637 const __m128i vxi22x89ABCDEF = _mm_cvtepu8_epi16(vi22x89ABCDEF);
638 const __m128i vk22x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(uint8_t)));
639 const __m128i vxk22x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk22x89ABCDEF), vk_zero_point);
640 i22 += 16;
641
642
643 const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
644 const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567);
645 const __m128i vprod22x89ABCDEFlo = _mm_mullo_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
646 const __m128i vprod22x89ABCDEFhi = _mm_mulhi_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
647
648 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi));
649 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi));
650 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod22x89ABCDEFlo, vprod22x89ABCDEFhi));
651 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod22x89ABCDEFlo, vprod22x89ABCDEFhi));
652
653 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
654 const __m128i vxi23x01234567 = _mm_cvtepu8_epi16(vi23x01234567);
655 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(uint8_t)));
656 const __m128i vxk23x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk23x01234567), vk_zero_point);
657 const __m128i vi23x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i23 + 8));
658 const __m128i vxi23x89ABCDEF = _mm_cvtepu8_epi16(vi23x89ABCDEF);
659 const __m128i vk23x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(uint8_t)));
660 const __m128i vxk23x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk23x89ABCDEF), vk_zero_point);
661 i23 += 16;
662
663
664 const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
665 const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567);
666 const __m128i vprod23x89ABCDEFlo = _mm_mullo_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF);
667 const __m128i vprod23x89ABCDEFhi = _mm_mulhi_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF);
668
669 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi));
670 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi));
671 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod23x89ABCDEFlo, vprod23x89ABCDEFhi));
672 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod23x89ABCDEFlo, vprod23x89ABCDEFhi));
673
674 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
675 const __m128i vxi24x01234567 = _mm_cvtepu8_epi16(vi24x01234567);
676 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(uint8_t)));
677 const __m128i vxk24x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk24x01234567), vk_zero_point);
678 const __m128i vi24x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i24 + 8));
679 const __m128i vxi24x89ABCDEF = _mm_cvtepu8_epi16(vi24x89ABCDEF);
680 const __m128i vk24x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(uint8_t)));
681 const __m128i vxk24x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk24x89ABCDEF), vk_zero_point);
682 i24 += 16;
683
684
685 const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
686 const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567);
687 const __m128i vprod24x89ABCDEFlo = _mm_mullo_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
688 const __m128i vprod24x89ABCDEFhi = _mm_mulhi_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
689
690 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi));
691 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi));
692 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod24x89ABCDEFlo, vprod24x89ABCDEFhi));
693 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod24x89ABCDEFlo, vprod24x89ABCDEFhi));
694
695 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(uint8_t));
696
697 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
698 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
699 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
700 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
701
702 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
703 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
704 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
705 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
706 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
707
708 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
709 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
710 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
711 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
712 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
713
714 vacc0123 = _mm_cvtps_epi32(vscaled0123);
715 vacc4567 = _mm_cvtps_epi32(vscaled4567);
716 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
717 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
718
719 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
720 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
721 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
722
723 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
724
725 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
726 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
727
728 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
729 output += 16;
730 }
731 if XNN_UNLIKELY(c != 0) {
732 const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 16);
733 do {
734 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
735 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
736
737
738 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
739 const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(vi0x01234567);
740 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
741 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk0x01234567), vk_zero_point);
742 i0 += 8;
743
744
745 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
746 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
747
748 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
749 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
750
751 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
752 const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(vi1x01234567);
753 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
754 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk1x01234567), vk_zero_point);
755 i1 += 8;
756
757
758 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
759 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
760
761 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
762 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
763
764 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
765 const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(vi2x01234567);
766 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
767 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk2x01234567), vk_zero_point);
768 i2 += 8;
769
770
771 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
772 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
773
774 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
775 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
776
777 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
778 const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(vi3x01234567);
779 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
780 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk3x01234567), vk_zero_point);
781 i3 += 8;
782
783
784 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
785 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
786
787 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
788 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
789
790 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
791 const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(vi4x01234567);
792 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
793 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk4x01234567), vk_zero_point);
794 i4 += 8;
795
796
797 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
798 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
799
800 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
801 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
802
803 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
804 const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(vi5x01234567);
805 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
806 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point);
807 i5 += 8;
808
809
810 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
811 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
812
813 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
814 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
815
816 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
817 const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(vi6x01234567);
818 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
819 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk6x01234567), vk_zero_point);
820 i6 += 8;
821
822
823 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
824 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
825
826 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
827 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
828
829 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
830 const __m128i vxi7x01234567 = _mm_cvtepu8_epi16(vi7x01234567);
831 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
832 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk7x01234567), vk_zero_point);
833 i7 += 8;
834
835
836 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
837 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
838
839 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
840 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
841
842 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
843 const __m128i vxi8x01234567 = _mm_cvtepu8_epi16(vi8x01234567);
844 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
845 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk8x01234567), vk_zero_point);
846 i8 += 8;
847
848
849 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
850 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
851
852 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
853 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
854
855 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
856 const __m128i vxi9x01234567 = _mm_cvtepu8_epi16(vi9x01234567);
857 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
858 const __m128i vxk9x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk9x01234567), vk_zero_point);
859 i9 += 8;
860
861
862 const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
863 const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567);
864
865 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi));
866 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi));
867
868 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
869 const __m128i vxi10x01234567 = _mm_cvtepu8_epi16(vi10x01234567);
870 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) (k + 160));
871 const __m128i vxk10x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk10x01234567), vk_zero_point);
872 i10 += 8;
873
874
875 const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
876 const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567);
877
878 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi));
879 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi));
880
881 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
882 const __m128i vxi11x01234567 = _mm_cvtepu8_epi16(vi11x01234567);
883 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) (k + 176));
884 const __m128i vxk11x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk11x01234567), vk_zero_point);
885 i11 += 8;
886
887
888 const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
889 const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567);
890
891 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi));
892 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi));
893
894 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
895 const __m128i vxi12x01234567 = _mm_cvtepu8_epi16(vi12x01234567);
896 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
897 const __m128i vxk12x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk12x01234567), vk_zero_point);
898 i12 += 8;
899
900
901 const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
902 const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567);
903
904 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi));
905 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi));
906
907 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
908 const __m128i vxi13x01234567 = _mm_cvtepu8_epi16(vi13x01234567);
909 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) (k + 208));
910 const __m128i vxk13x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk13x01234567), vk_zero_point);
911 i13 += 8;
912
913
914 const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
915 const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567);
916
917 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi));
918 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi));
919
920 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
921 const __m128i vxi14x01234567 = _mm_cvtepu8_epi16(vi14x01234567);
922 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) (k + 224));
923 const __m128i vxk14x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk14x01234567), vk_zero_point);
924 i14 += 8;
925
926
927 const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
928 const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567);
929
930 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi));
931 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi));
932
933 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
934 const __m128i vxi15x01234567 = _mm_cvtepu8_epi16(vi15x01234567);
935 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) (k + 240));
936 const __m128i vxk15x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk15x01234567), vk_zero_point);
937 i15 += 8;
938
939
940 const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
941 const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567);
942
943 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi));
944 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi));
945
946 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
947 const __m128i vxi16x01234567 = _mm_cvtepu8_epi16(vi16x01234567);
948 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) (k + 256));
949 const __m128i vxk16x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk16x01234567), vk_zero_point);
950 i16 += 8;
951
952
953 const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
954 const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567);
955
956 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi));
957 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi));
958
959 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
960 const __m128i vxi17x01234567 = _mm_cvtepu8_epi16(vi17x01234567);
961 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) (k + 272));
962 const __m128i vxk17x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk17x01234567), vk_zero_point);
963 i17 += 8;
964
965
966 const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
967 const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567);
968
969 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi));
970 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi));
971
972 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
973 const __m128i vxi18x01234567 = _mm_cvtepu8_epi16(vi18x01234567);
974 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) (k + 288));
975 const __m128i vxk18x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk18x01234567), vk_zero_point);
976 i18 += 8;
977
978
979 const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
980 const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567);
981
982 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi));
983 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi));
984
985 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
986 const __m128i vxi19x01234567 = _mm_cvtepu8_epi16(vi19x01234567);
987 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) (k + 304));
988 const __m128i vxk19x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk19x01234567), vk_zero_point);
989 i19 += 8;
990
991
992 const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
993 const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567);
994
995 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi));
996 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi));
997
998 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
999 const __m128i vxi20x01234567 = _mm_cvtepu8_epi16(vi20x01234567);
1000 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) (k + 320));
1001 const __m128i vxk20x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk20x01234567), vk_zero_point);
1002 i20 += 8;
1003
1004
1005 const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
1006 const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567);
1007
1008 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi));
1009 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi));
1010
1011 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
1012 const __m128i vxi21x01234567 = _mm_cvtepu8_epi16(vi21x01234567);
1013 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) (k + 336));
1014 const __m128i vxk21x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk21x01234567), vk_zero_point);
1015 i21 += 8;
1016
1017
1018 const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
1019 const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567);
1020
1021 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi));
1022 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi));
1023
1024 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
1025 const __m128i vxi22x01234567 = _mm_cvtepu8_epi16(vi22x01234567);
1026 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) (k + 352));
1027 const __m128i vxk22x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk22x01234567), vk_zero_point);
1028 i22 += 8;
1029
1030
1031 const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
1032 const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567);
1033
1034 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi));
1035 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi));
1036
1037 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
1038 const __m128i vxi23x01234567 = _mm_cvtepu8_epi16(vi23x01234567);
1039 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) (k + 368));
1040 const __m128i vxk23x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk23x01234567), vk_zero_point);
1041 i23 += 8;
1042
1043
1044 const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
1045 const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567);
1046
1047 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi));
1048 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi));
1049
1050 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
1051 const __m128i vxi24x01234567 = _mm_cvtepu8_epi16(vi24x01234567);
1052 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) (k + 384));
1053 const __m128i vxk24x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk24x01234567), vk_zero_point);
1054 i24 += 8;
1055
1056
1057 const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
1058 const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567);
1059
1060 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi));
1061 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi));
1062
1063 k += 8;
1064
1065 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
1066 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
1067
1068 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
1069 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
1070 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
1071
1072 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
1073 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
1074 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
1075
1076 vacc0123 = _mm_cvtps_epi32(vscaled0123);
1077 vacc4567 = _mm_cvtps_epi32(vscaled4567);
1078
1079 w = (const void*) ((const int32_t*) w + 8);
1080
1081 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
1082 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
1083
1084 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
1085
1086 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
1087
1088 if XNN_LIKELY(c >= 8) {
1089 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
1090 output += 8;
1091 c -= 8;
1092 } else {
1093 if (c & 4) {
1094 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
1095 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
1096 output += 4;
1097 }
1098 if (c & 2) {
1099 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
1100 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
1101 output += 2;
1102 }
1103 if (c & 1) {
1104 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
1105 output += 1;
1106 }
1107 c = 0;
1108 }
1109 } while (c != 0);
1110 }
1111
1112 output = (uint8_t*) ((uintptr_t) output + output_increment);
1113 } while (--output_width != 0);
1114 }
1115