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_up8x25__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_up8x25__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 >= 8; c -= 8) {
164 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
165 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
166
167
168 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
169 const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(vi0x01234567);
170 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
171 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk0x01234567), vk_zero_point);
172 i0 += 8;
173
174
175 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
176 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
177
178 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
179 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
180
181 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
182 const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(vi1x01234567);
183 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
184 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk1x01234567), vk_zero_point);
185 i1 += 8;
186
187
188 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
189 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
190
191 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
192 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
193
194 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
195 const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(vi2x01234567);
196 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
197 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk2x01234567), vk_zero_point);
198 i2 += 8;
199
200
201 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
202 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
203
204 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
205 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
206
207 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
208 const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(vi3x01234567);
209 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
210 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk3x01234567), vk_zero_point);
211 i3 += 8;
212
213
214 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
215 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
216
217 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
218 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
219
220 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
221 const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(vi4x01234567);
222 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
223 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk4x01234567), vk_zero_point);
224 i4 += 8;
225
226
227 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
228 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
229
230 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
231 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
232
233 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
234 const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(vi5x01234567);
235 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
236 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point);
237 i5 += 8;
238
239
240 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
241 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
242
243 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
244 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
245
246 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
247 const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(vi6x01234567);
248 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
249 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk6x01234567), vk_zero_point);
250 i6 += 8;
251
252
253 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
254 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
255
256 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
257 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
258
259 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
260 const __m128i vxi7x01234567 = _mm_cvtepu8_epi16(vi7x01234567);
261 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
262 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk7x01234567), vk_zero_point);
263 i7 += 8;
264
265
266 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
267 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
268
269 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
270 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
271
272 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
273 const __m128i vxi8x01234567 = _mm_cvtepu8_epi16(vi8x01234567);
274 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
275 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk8x01234567), vk_zero_point);
276 i8 += 8;
277
278
279 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
280 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
281
282 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
283 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
284
285 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
286 const __m128i vxi9x01234567 = _mm_cvtepu8_epi16(vi9x01234567);
287 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
288 const __m128i vxk9x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk9x01234567), vk_zero_point);
289 i9 += 8;
290
291
292 const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
293 const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567);
294
295 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi));
296 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi));
297
298 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
299 const __m128i vxi10x01234567 = _mm_cvtepu8_epi16(vi10x01234567);
300 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
301 const __m128i vxk10x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk10x01234567), vk_zero_point);
302 i10 += 8;
303
304
305 const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
306 const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567);
307
308 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi));
309 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi));
310
311 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
312 const __m128i vxi11x01234567 = _mm_cvtepu8_epi16(vi11x01234567);
313 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
314 const __m128i vxk11x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk11x01234567), vk_zero_point);
315 i11 += 8;
316
317
318 const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
319 const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567);
320
321 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi));
322 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi));
323
324 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
325 const __m128i vxi12x01234567 = _mm_cvtepu8_epi16(vi12x01234567);
326 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
327 const __m128i vxk12x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk12x01234567), vk_zero_point);
328 i12 += 8;
329
330
331 const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
332 const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567);
333
334 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi));
335 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi));
336
337 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
338 const __m128i vxi13x01234567 = _mm_cvtepu8_epi16(vi13x01234567);
339 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
340 const __m128i vxk13x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk13x01234567), vk_zero_point);
341 i13 += 8;
342
343
344 const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
345 const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567);
346
347 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi));
348 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi));
349
350 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
351 const __m128i vxi14x01234567 = _mm_cvtepu8_epi16(vi14x01234567);
352 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
353 const __m128i vxk14x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk14x01234567), vk_zero_point);
354 i14 += 8;
355
356
357 const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
358 const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567);
359
360 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi));
361 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi));
362
363 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
364 const __m128i vxi15x01234567 = _mm_cvtepu8_epi16(vi15x01234567);
365 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
366 const __m128i vxk15x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk15x01234567), vk_zero_point);
367 i15 += 8;
368
369
370 const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
371 const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567);
372
373 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi));
374 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi));
375
376 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
377 const __m128i vxi16x01234567 = _mm_cvtepu8_epi16(vi16x01234567);
378 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
379 const __m128i vxk16x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk16x01234567), vk_zero_point);
380 i16 += 8;
381
382
383 const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
384 const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567);
385
386 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi));
387 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi));
388
389 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
390 const __m128i vxi17x01234567 = _mm_cvtepu8_epi16(vi17x01234567);
391 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
392 const __m128i vxk17x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk17x01234567), vk_zero_point);
393 i17 += 8;
394
395
396 const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
397 const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567);
398
399 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi));
400 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi));
401
402 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
403 const __m128i vxi18x01234567 = _mm_cvtepu8_epi16(vi18x01234567);
404 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(uint8_t)));
405 const __m128i vxk18x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk18x01234567), vk_zero_point);
406 i18 += 8;
407
408
409 const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
410 const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567);
411
412 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi));
413 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi));
414
415 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
416 const __m128i vxi19x01234567 = _mm_cvtepu8_epi16(vi19x01234567);
417 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(uint8_t)));
418 const __m128i vxk19x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk19x01234567), vk_zero_point);
419 i19 += 8;
420
421
422 const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
423 const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567);
424
425 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi));
426 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi));
427
428 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
429 const __m128i vxi20x01234567 = _mm_cvtepu8_epi16(vi20x01234567);
430 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(uint8_t)));
431 const __m128i vxk20x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk20x01234567), vk_zero_point);
432 i20 += 8;
433
434
435 const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
436 const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567);
437
438 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi));
439 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi));
440
441 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
442 const __m128i vxi21x01234567 = _mm_cvtepu8_epi16(vi21x01234567);
443 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(uint8_t)));
444 const __m128i vxk21x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk21x01234567), vk_zero_point);
445 i21 += 8;
446
447
448 const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
449 const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567);
450
451 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi));
452 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi));
453
454 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
455 const __m128i vxi22x01234567 = _mm_cvtepu8_epi16(vi22x01234567);
456 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(uint8_t)));
457 const __m128i vxk22x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk22x01234567), vk_zero_point);
458 i22 += 8;
459
460
461 const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
462 const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567);
463
464 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi));
465 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi));
466
467 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
468 const __m128i vxi23x01234567 = _mm_cvtepu8_epi16(vi23x01234567);
469 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(uint8_t)));
470 const __m128i vxk23x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk23x01234567), vk_zero_point);
471 i23 += 8;
472
473
474 const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
475 const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567);
476
477 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi));
478 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi));
479
480 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
481 const __m128i vxi24x01234567 = _mm_cvtepu8_epi16(vi24x01234567);
482 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(uint8_t)));
483 const __m128i vxk24x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk24x01234567), vk_zero_point);
484 i24 += 8;
485
486
487 const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
488 const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567);
489
490 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi));
491 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi));
492
493 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(uint8_t));
494
495 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
496 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
497
498 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
499 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
500 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
501
502 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
503 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
504 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
505
506 vacc0123 = _mm_cvtps_epi32(vscaled0123);
507 vacc4567 = _mm_cvtps_epi32(vscaled4567);
508
509 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
510 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
511
512 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
513
514 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
515 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
516
517 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
518 output += 8;
519 }
520 if XNN_UNLIKELY(c != 0) {
521 {
522 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
523 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
524
525
526 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
527 const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(vi0x01234567);
528 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
529 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk0x01234567), vk_zero_point);
530
531
532 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
533 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
534
535 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
536 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
537
538 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
539 const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(vi1x01234567);
540 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
541 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk1x01234567), vk_zero_point);
542
543
544 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
545 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
546
547 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
548 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
549
550 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
551 const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(vi2x01234567);
552 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
553 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk2x01234567), vk_zero_point);
554
555
556 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
557 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
558
559 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
560 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
561
562 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
563 const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(vi3x01234567);
564 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
565 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk3x01234567), vk_zero_point);
566
567
568 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
569 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
570
571 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
572 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
573
574 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
575 const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(vi4x01234567);
576 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
577 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk4x01234567), vk_zero_point);
578
579
580 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
581 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
582
583 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
584 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
585
586 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
587 const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(vi5x01234567);
588 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
589 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point);
590
591
592 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
593 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
594
595 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
596 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
597
598 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
599 const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(vi6x01234567);
600 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
601 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk6x01234567), vk_zero_point);
602
603
604 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
605 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
606
607 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
608 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
609
610 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
611 const __m128i vxi7x01234567 = _mm_cvtepu8_epi16(vi7x01234567);
612 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
613 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk7x01234567), vk_zero_point);
614
615
616 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
617 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
618
619 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
620 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
621
622 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
623 const __m128i vxi8x01234567 = _mm_cvtepu8_epi16(vi8x01234567);
624 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
625 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk8x01234567), vk_zero_point);
626
627
628 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
629 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
630
631 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
632 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
633
634 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
635 const __m128i vxi9x01234567 = _mm_cvtepu8_epi16(vi9x01234567);
636 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
637 const __m128i vxk9x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk9x01234567), vk_zero_point);
638
639
640 const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
641 const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567);
642
643 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi));
644 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi));
645
646 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
647 const __m128i vxi10x01234567 = _mm_cvtepu8_epi16(vi10x01234567);
648 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
649 const __m128i vxk10x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk10x01234567), vk_zero_point);
650
651
652 const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
653 const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567);
654
655 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi));
656 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi));
657
658 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
659 const __m128i vxi11x01234567 = _mm_cvtepu8_epi16(vi11x01234567);
660 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
661 const __m128i vxk11x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk11x01234567), vk_zero_point);
662
663
664 const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
665 const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567);
666
667 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi));
668 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi));
669
670 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
671 const __m128i vxi12x01234567 = _mm_cvtepu8_epi16(vi12x01234567);
672 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
673 const __m128i vxk12x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk12x01234567), vk_zero_point);
674
675
676 const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
677 const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567);
678
679 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi));
680 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi));
681
682 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
683 const __m128i vxi13x01234567 = _mm_cvtepu8_epi16(vi13x01234567);
684 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
685 const __m128i vxk13x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk13x01234567), vk_zero_point);
686
687
688 const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
689 const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567);
690
691 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi));
692 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi));
693
694 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
695 const __m128i vxi14x01234567 = _mm_cvtepu8_epi16(vi14x01234567);
696 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
697 const __m128i vxk14x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk14x01234567), vk_zero_point);
698
699
700 const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
701 const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567);
702
703 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi));
704 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi));
705
706 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
707 const __m128i vxi15x01234567 = _mm_cvtepu8_epi16(vi15x01234567);
708 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
709 const __m128i vxk15x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk15x01234567), vk_zero_point);
710
711
712 const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
713 const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567);
714
715 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi));
716 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi));
717
718 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
719 const __m128i vxi16x01234567 = _mm_cvtepu8_epi16(vi16x01234567);
720 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
721 const __m128i vxk16x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk16x01234567), vk_zero_point);
722
723
724 const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
725 const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567);
726
727 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi));
728 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi));
729
730 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
731 const __m128i vxi17x01234567 = _mm_cvtepu8_epi16(vi17x01234567);
732 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
733 const __m128i vxk17x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk17x01234567), vk_zero_point);
734
735
736 const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
737 const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567);
738
739 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi));
740 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi));
741
742 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
743 const __m128i vxi18x01234567 = _mm_cvtepu8_epi16(vi18x01234567);
744 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(uint8_t)));
745 const __m128i vxk18x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk18x01234567), vk_zero_point);
746
747
748 const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
749 const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567);
750
751 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi));
752 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi));
753
754 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
755 const __m128i vxi19x01234567 = _mm_cvtepu8_epi16(vi19x01234567);
756 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(uint8_t)));
757 const __m128i vxk19x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk19x01234567), vk_zero_point);
758
759
760 const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
761 const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567);
762
763 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi));
764 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi));
765
766 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
767 const __m128i vxi20x01234567 = _mm_cvtepu8_epi16(vi20x01234567);
768 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(uint8_t)));
769 const __m128i vxk20x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk20x01234567), vk_zero_point);
770
771
772 const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
773 const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567);
774
775 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi));
776 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi));
777
778 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
779 const __m128i vxi21x01234567 = _mm_cvtepu8_epi16(vi21x01234567);
780 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(uint8_t)));
781 const __m128i vxk21x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk21x01234567), vk_zero_point);
782
783
784 const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
785 const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567);
786
787 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi));
788 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi));
789
790 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
791 const __m128i vxi22x01234567 = _mm_cvtepu8_epi16(vi22x01234567);
792 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(uint8_t)));
793 const __m128i vxk22x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk22x01234567), vk_zero_point);
794
795
796 const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
797 const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567);
798
799 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi));
800 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi));
801
802 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
803 const __m128i vxi23x01234567 = _mm_cvtepu8_epi16(vi23x01234567);
804 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(uint8_t)));
805 const __m128i vxk23x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk23x01234567), vk_zero_point);
806
807
808 const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
809 const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567);
810
811 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi));
812 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi));
813
814 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
815 const __m128i vxi24x01234567 = _mm_cvtepu8_epi16(vi24x01234567);
816 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(uint8_t)));
817 const __m128i vxk24x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk24x01234567), vk_zero_point);
818
819
820 const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
821 const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567);
822
823 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi));
824 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi));
825
826
827 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
828 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
829
830 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
831 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
832 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
833
834 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
835 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
836 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
837
838 vacc0123 = _mm_cvtps_epi32(vscaled0123);
839 vacc4567 = _mm_cvtps_epi32(vscaled4567);
840
841
842 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
843 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
844
845 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
846
847 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
848
849 if (c & 4) {
850 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
851 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
852 output += 4;
853 }
854 if (c & 2) {
855 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
856 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
857 output += 2;
858 }
859 if (c & 1) {
860 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
861 output += 1;
862 }
863 }
864 }
865
866 output = (uint8_t*) ((uintptr_t) output + output_increment);
867 } while (--output_width != 0);
868 }
869