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 <emmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/unaligned.h>
16
17
xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse2_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)])18 void xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16(
19 size_t channels,
20 size_t output_width,
21 const uint8_t** input,
22 const void* weights,
23 uint8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const uint8_t* zero,
28 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 do {
34 const uint8_t* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
38 }
39 const uint8_t* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
43 }
44 const uint8_t* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
48 }
49 const uint8_t* i3 = input[3];
50 assert(i3 != NULL);
51 if XNN_UNPREDICTABLE(i3 != zero) {
52 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
53 }
54 const uint8_t* i4 = input[4];
55 assert(i4 != NULL);
56 if XNN_UNPREDICTABLE(i4 != zero) {
57 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
58 }
59 const uint8_t* i5 = input[5];
60 assert(i5 != NULL);
61 if XNN_UNPREDICTABLE(i5 != zero) {
62 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
63 }
64 const uint8_t* i6 = input[6];
65 assert(i6 != NULL);
66 if XNN_UNPREDICTABLE(i6 != zero) {
67 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
68 }
69 const uint8_t* i7 = input[7];
70 assert(i7 != NULL);
71 if XNN_UNPREDICTABLE(i7 != zero) {
72 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
73 }
74 const uint8_t* i8 = input[8];
75 assert(i8 != NULL);
76 if XNN_UNPREDICTABLE(i8 != zero) {
77 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
78 }
79 const uint8_t* i9 = input[9];
80 assert(i9 != NULL);
81 if XNN_UNPREDICTABLE(i9 != zero) {
82 i9 = (const uint8_t*) ((uintptr_t) i9 + input_offset);
83 }
84 const uint8_t* i10 = input[10];
85 assert(i10 != NULL);
86 if XNN_UNPREDICTABLE(i10 != zero) {
87 i10 = (const uint8_t*) ((uintptr_t) i10 + input_offset);
88 }
89 const uint8_t* i11 = input[11];
90 assert(i11 != NULL);
91 if XNN_UNPREDICTABLE(i11 != zero) {
92 i11 = (const uint8_t*) ((uintptr_t) i11 + input_offset);
93 }
94 const uint8_t* i12 = input[12];
95 assert(i12 != NULL);
96 if XNN_UNPREDICTABLE(i12 != zero) {
97 i12 = (const uint8_t*) ((uintptr_t) i12 + input_offset);
98 }
99 const uint8_t* i13 = input[13];
100 assert(i13 != NULL);
101 if XNN_UNPREDICTABLE(i13 != zero) {
102 i13 = (const uint8_t*) ((uintptr_t) i13 + input_offset);
103 }
104 const uint8_t* i14 = input[14];
105 assert(i14 != NULL);
106 if XNN_UNPREDICTABLE(i14 != zero) {
107 i14 = (const uint8_t*) ((uintptr_t) i14 + input_offset);
108 }
109 const uint8_t* i15 = input[15];
110 assert(i15 != NULL);
111 if XNN_UNPREDICTABLE(i15 != zero) {
112 i15 = (const uint8_t*) ((uintptr_t) i15 + input_offset);
113 }
114 const uint8_t* i16 = input[16];
115 assert(i16 != NULL);
116 if XNN_UNPREDICTABLE(i16 != zero) {
117 i16 = (const uint8_t*) ((uintptr_t) i16 + input_offset);
118 }
119 const uint8_t* i17 = input[17];
120 assert(i17 != NULL);
121 if XNN_UNPREDICTABLE(i17 != zero) {
122 i17 = (const uint8_t*) ((uintptr_t) i17 + input_offset);
123 }
124 const uint8_t* i18 = input[18];
125 assert(i18 != NULL);
126 if XNN_UNPREDICTABLE(i18 != zero) {
127 i18 = (const uint8_t*) ((uintptr_t) i18 + input_offset);
128 }
129 const uint8_t* i19 = input[19];
130 assert(i19 != NULL);
131 if XNN_UNPREDICTABLE(i19 != zero) {
132 i19 = (const uint8_t*) ((uintptr_t) i19 + input_offset);
133 }
134 const uint8_t* i20 = input[20];
135 assert(i20 != NULL);
136 if XNN_UNPREDICTABLE(i20 != zero) {
137 i20 = (const uint8_t*) ((uintptr_t) i20 + input_offset);
138 }
139 const uint8_t* i21 = input[21];
140 assert(i21 != NULL);
141 if XNN_UNPREDICTABLE(i21 != zero) {
142 i21 = (const uint8_t*) ((uintptr_t) i21 + input_offset);
143 }
144 const uint8_t* i22 = input[22];
145 assert(i22 != NULL);
146 if XNN_UNPREDICTABLE(i22 != zero) {
147 i22 = (const uint8_t*) ((uintptr_t) i22 + input_offset);
148 }
149 const uint8_t* i23 = input[23];
150 assert(i23 != NULL);
151 if XNN_UNPREDICTABLE(i23 != zero) {
152 i23 = (const uint8_t*) ((uintptr_t) i23 + input_offset);
153 }
154 const uint8_t* i24 = input[24];
155 assert(i24 != NULL);
156 if XNN_UNPREDICTABLE(i24 != zero) {
157 i24 = (const uint8_t*) ((uintptr_t) i24 + input_offset);
158 }
159 input = (const uint8_t**) ((uintptr_t) input + input_stride);
160
161 size_t c = channels;
162 const void* w = weights;
163 const __m128i vk_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
164 for (; c >= 8; c -= 8) {
165 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
166 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
167
168
169 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
170 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
171 i0 += 8;
172
173 const __m128i vzero = _mm_setzero_si128();
174 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
175 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x01234567, vzero), vk_zero_point);
176
177 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
178 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
179
180 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
181 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
182
183 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
184 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
185 i1 += 8;
186
187 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
188 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x01234567, vzero), vk_zero_point);
189
190 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
191 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
192
193 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
194 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
195
196 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
197 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
198 i2 += 8;
199
200 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
201 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x01234567, vzero), vk_zero_point);
202
203 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
204 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
205
206 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
207 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
208
209 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
210 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
211 i3 += 8;
212
213 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
214 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x01234567, vzero), vk_zero_point);
215
216 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
217 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
218
219 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
220 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
221
222 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
223 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
224 i4 += 8;
225
226 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
227 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x01234567, vzero), vk_zero_point);
228
229 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
230 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
231
232 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
233 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
234
235 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
236 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
237 i5 += 8;
238
239 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
240 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point);
241
242 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
243 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
244
245 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
246 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
247
248 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
249 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
250 i6 += 8;
251
252 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
253 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x01234567, vzero), vk_zero_point);
254
255 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
256 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
257
258 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
259 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
260
261 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
262 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
263 i7 += 8;
264
265 const __m128i vxi7x01234567 = _mm_unpacklo_epi8(vi7x01234567, vzero);
266 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x01234567, vzero), vk_zero_point);
267
268 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
269 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
270
271 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
272 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
273
274 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
275 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
276 i8 += 8;
277
278 const __m128i vxi8x01234567 = _mm_unpacklo_epi8(vi8x01234567, vzero);
279 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x01234567, vzero), vk_zero_point);
280
281 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
282 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
283
284 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
285 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
286
287 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
288 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
289 i9 += 8;
290
291 const __m128i vxi9x01234567 = _mm_unpacklo_epi8(vi9x01234567, vzero);
292 const __m128i vxk9x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk9x01234567, vzero), vk_zero_point);
293
294 const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
295 const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567);
296
297 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi));
298 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi));
299
300 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
301 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
302 i10 += 8;
303
304 const __m128i vxi10x01234567 = _mm_unpacklo_epi8(vi10x01234567, vzero);
305 const __m128i vxk10x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk10x01234567, vzero), vk_zero_point);
306
307 const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
308 const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567);
309
310 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi));
311 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi));
312
313 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
314 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
315 i11 += 8;
316
317 const __m128i vxi11x01234567 = _mm_unpacklo_epi8(vi11x01234567, vzero);
318 const __m128i vxk11x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk11x01234567, vzero), vk_zero_point);
319
320 const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
321 const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567);
322
323 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi));
324 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi));
325
326 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
327 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
328 i12 += 8;
329
330 const __m128i vxi12x01234567 = _mm_unpacklo_epi8(vi12x01234567, vzero);
331 const __m128i vxk12x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk12x01234567, vzero), vk_zero_point);
332
333 const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
334 const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567);
335
336 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi));
337 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi));
338
339 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
340 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
341 i13 += 8;
342
343 const __m128i vxi13x01234567 = _mm_unpacklo_epi8(vi13x01234567, vzero);
344 const __m128i vxk13x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk13x01234567, vzero), vk_zero_point);
345
346 const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
347 const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567);
348
349 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi));
350 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi));
351
352 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
353 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
354 i14 += 8;
355
356 const __m128i vxi14x01234567 = _mm_unpacklo_epi8(vi14x01234567, vzero);
357 const __m128i vxk14x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk14x01234567, vzero), vk_zero_point);
358
359 const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
360 const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567);
361
362 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi));
363 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi));
364
365 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
366 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
367 i15 += 8;
368
369 const __m128i vxi15x01234567 = _mm_unpacklo_epi8(vi15x01234567, vzero);
370 const __m128i vxk15x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk15x01234567, vzero), vk_zero_point);
371
372 const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
373 const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567);
374
375 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi));
376 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi));
377
378 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
379 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
380 i16 += 8;
381
382 const __m128i vxi16x01234567 = _mm_unpacklo_epi8(vi16x01234567, vzero);
383 const __m128i vxk16x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk16x01234567, vzero), vk_zero_point);
384
385 const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
386 const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567);
387
388 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi));
389 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi));
390
391 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
392 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
393 i17 += 8;
394
395 const __m128i vxi17x01234567 = _mm_unpacklo_epi8(vi17x01234567, vzero);
396 const __m128i vxk17x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk17x01234567, vzero), vk_zero_point);
397
398 const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
399 const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567);
400
401 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi));
402 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi));
403
404 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
405 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(uint8_t)));
406 i18 += 8;
407
408 const __m128i vxi18x01234567 = _mm_unpacklo_epi8(vi18x01234567, vzero);
409 const __m128i vxk18x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk18x01234567, vzero), vk_zero_point);
410
411 const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
412 const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567);
413
414 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi));
415 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi));
416
417 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
418 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(uint8_t)));
419 i19 += 8;
420
421 const __m128i vxi19x01234567 = _mm_unpacklo_epi8(vi19x01234567, vzero);
422 const __m128i vxk19x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk19x01234567, vzero), vk_zero_point);
423
424 const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
425 const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567);
426
427 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi));
428 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi));
429
430 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
431 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(uint8_t)));
432 i20 += 8;
433
434 const __m128i vxi20x01234567 = _mm_unpacklo_epi8(vi20x01234567, vzero);
435 const __m128i vxk20x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk20x01234567, vzero), vk_zero_point);
436
437 const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
438 const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567);
439
440 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi));
441 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi));
442
443 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
444 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(uint8_t)));
445 i21 += 8;
446
447 const __m128i vxi21x01234567 = _mm_unpacklo_epi8(vi21x01234567, vzero);
448 const __m128i vxk21x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk21x01234567, vzero), vk_zero_point);
449
450 const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
451 const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567);
452
453 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi));
454 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi));
455
456 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
457 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(uint8_t)));
458 i22 += 8;
459
460 const __m128i vxi22x01234567 = _mm_unpacklo_epi8(vi22x01234567, vzero);
461 const __m128i vxk22x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk22x01234567, vzero), vk_zero_point);
462
463 const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
464 const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567);
465
466 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi));
467 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi));
468
469 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
470 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(uint8_t)));
471 i23 += 8;
472
473 const __m128i vxi23x01234567 = _mm_unpacklo_epi8(vi23x01234567, vzero);
474 const __m128i vxk23x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk23x01234567, vzero), vk_zero_point);
475
476 const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
477 const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567);
478
479 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi));
480 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi));
481
482 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
483 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(uint8_t)));
484 i24 += 8;
485
486 const __m128i vxi24x01234567 = _mm_unpacklo_epi8(vi24x01234567, vzero);
487 const __m128i vxk24x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk24x01234567, vzero), vk_zero_point);
488
489 const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
490 const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567);
491
492 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi));
493 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi));
494
495 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(uint8_t));
496
497 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
498 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
499
500 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
501 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
502 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
503
504 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
505 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
506 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
507
508 vacc0123 = _mm_cvtps_epi32(vscaled0123);
509 vacc4567 = _mm_cvtps_epi32(vscaled4567);
510
511 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
512 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
513
514 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
515
516 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
517 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
518
519 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
520 output += 8;
521 }
522 if XNN_UNLIKELY(c != 0) {
523 {
524 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
525 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
526
527
528 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
529 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
530
531 const __m128i vzero = _mm_setzero_si128();
532 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
533 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x01234567, vzero), vk_zero_point);
534
535 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
536 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
537
538 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
539 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
540
541 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
542 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
543
544 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
545 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x01234567, vzero), vk_zero_point);
546
547 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
548 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
549
550 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
551 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
552
553 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
554 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
555
556 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
557 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x01234567, vzero), vk_zero_point);
558
559 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
560 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
561
562 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
563 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
564
565 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
566 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
567
568 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
569 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x01234567, vzero), vk_zero_point);
570
571 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
572 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
573
574 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
575 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
576
577 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
578 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
579
580 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
581 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x01234567, vzero), vk_zero_point);
582
583 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
584 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
585
586 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
587 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
588
589 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
590 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
591
592 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
593 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point);
594
595 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
596 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
597
598 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
599 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
600
601 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
602 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
603
604 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
605 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x01234567, vzero), vk_zero_point);
606
607 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
608 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
609
610 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
611 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
612
613 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
614 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
615
616 const __m128i vxi7x01234567 = _mm_unpacklo_epi8(vi7x01234567, vzero);
617 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x01234567, vzero), vk_zero_point);
618
619 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
620 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
621
622 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
623 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
624
625 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
626 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
627
628 const __m128i vxi8x01234567 = _mm_unpacklo_epi8(vi8x01234567, vzero);
629 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x01234567, vzero), vk_zero_point);
630
631 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
632 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
633
634 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
635 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
636
637 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
638 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
639
640 const __m128i vxi9x01234567 = _mm_unpacklo_epi8(vi9x01234567, vzero);
641 const __m128i vxk9x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk9x01234567, vzero), vk_zero_point);
642
643 const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
644 const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567);
645
646 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi));
647 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi));
648
649 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
650 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
651
652 const __m128i vxi10x01234567 = _mm_unpacklo_epi8(vi10x01234567, vzero);
653 const __m128i vxk10x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk10x01234567, vzero), vk_zero_point);
654
655 const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
656 const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567);
657
658 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi));
659 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi));
660
661 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
662 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
663
664 const __m128i vxi11x01234567 = _mm_unpacklo_epi8(vi11x01234567, vzero);
665 const __m128i vxk11x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk11x01234567, vzero), vk_zero_point);
666
667 const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
668 const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567);
669
670 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi));
671 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi));
672
673 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
674 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
675
676 const __m128i vxi12x01234567 = _mm_unpacklo_epi8(vi12x01234567, vzero);
677 const __m128i vxk12x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk12x01234567, vzero), vk_zero_point);
678
679 const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
680 const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567);
681
682 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi));
683 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi));
684
685 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
686 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
687
688 const __m128i vxi13x01234567 = _mm_unpacklo_epi8(vi13x01234567, vzero);
689 const __m128i vxk13x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk13x01234567, vzero), vk_zero_point);
690
691 const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
692 const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567);
693
694 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi));
695 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi));
696
697 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
698 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
699
700 const __m128i vxi14x01234567 = _mm_unpacklo_epi8(vi14x01234567, vzero);
701 const __m128i vxk14x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk14x01234567, vzero), vk_zero_point);
702
703 const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
704 const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567);
705
706 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi));
707 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi));
708
709 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
710 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
711
712 const __m128i vxi15x01234567 = _mm_unpacklo_epi8(vi15x01234567, vzero);
713 const __m128i vxk15x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk15x01234567, vzero), vk_zero_point);
714
715 const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
716 const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567);
717
718 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi));
719 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi));
720
721 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
722 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
723
724 const __m128i vxi16x01234567 = _mm_unpacklo_epi8(vi16x01234567, vzero);
725 const __m128i vxk16x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk16x01234567, vzero), vk_zero_point);
726
727 const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
728 const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567);
729
730 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi));
731 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi));
732
733 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
734 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
735
736 const __m128i vxi17x01234567 = _mm_unpacklo_epi8(vi17x01234567, vzero);
737 const __m128i vxk17x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk17x01234567, vzero), vk_zero_point);
738
739 const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
740 const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567);
741
742 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi));
743 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi));
744
745 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
746 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(uint8_t)));
747
748 const __m128i vxi18x01234567 = _mm_unpacklo_epi8(vi18x01234567, vzero);
749 const __m128i vxk18x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk18x01234567, vzero), vk_zero_point);
750
751 const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
752 const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567);
753
754 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi));
755 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi));
756
757 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
758 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(uint8_t)));
759
760 const __m128i vxi19x01234567 = _mm_unpacklo_epi8(vi19x01234567, vzero);
761 const __m128i vxk19x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk19x01234567, vzero), vk_zero_point);
762
763 const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
764 const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567);
765
766 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi));
767 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi));
768
769 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
770 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(uint8_t)));
771
772 const __m128i vxi20x01234567 = _mm_unpacklo_epi8(vi20x01234567, vzero);
773 const __m128i vxk20x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk20x01234567, vzero), vk_zero_point);
774
775 const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
776 const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567);
777
778 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi));
779 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi));
780
781 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
782 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(uint8_t)));
783
784 const __m128i vxi21x01234567 = _mm_unpacklo_epi8(vi21x01234567, vzero);
785 const __m128i vxk21x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk21x01234567, vzero), vk_zero_point);
786
787 const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
788 const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567);
789
790 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi));
791 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi));
792
793 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
794 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(uint8_t)));
795
796 const __m128i vxi22x01234567 = _mm_unpacklo_epi8(vi22x01234567, vzero);
797 const __m128i vxk22x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk22x01234567, vzero), vk_zero_point);
798
799 const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
800 const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567);
801
802 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi));
803 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi));
804
805 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
806 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(uint8_t)));
807
808 const __m128i vxi23x01234567 = _mm_unpacklo_epi8(vi23x01234567, vzero);
809 const __m128i vxk23x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk23x01234567, vzero), vk_zero_point);
810
811 const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
812 const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567);
813
814 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi));
815 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi));
816
817 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
818 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(uint8_t)));
819
820 const __m128i vxi24x01234567 = _mm_unpacklo_epi8(vi24x01234567, vzero);
821 const __m128i vxk24x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk24x01234567, vzero), vk_zero_point);
822
823 const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
824 const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567);
825
826 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi));
827 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi));
828
829
830 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
831 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
832
833 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
834 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
835 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
836
837 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
838 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
839 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
840
841 vacc0123 = _mm_cvtps_epi32(vscaled0123);
842 vacc4567 = _mm_cvtps_epi32(vscaled4567);
843
844
845 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
846 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
847
848 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
849
850 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
851
852 if (c & 4) {
853 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
854 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
855 output += 4;
856 }
857 if (c & 2) {
858 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
859 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
860 output += 2;
861 }
862 if (c & 1) {
863 *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567);
864 output += 1;
865 }
866 }
867 }
868
869 output = (uint8_t*) ((uintptr_t) output + output_increment);
870 } while (--output_width != 0);
871 }
872