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 #include <xnnpack/unaligned.h>
16
17
xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse41_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__sse41_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 vxi0x01234567 = _mm_cvtepu8_epi16(vi0x01234567);
171 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
172 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk0x01234567), vk_zero_point);
173 i0 += 8;
174
175
176 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
177 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
178
179 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
180 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
181
182 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
183 const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(vi1x01234567);
184 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
185 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk1x01234567), vk_zero_point);
186 i1 += 8;
187
188
189 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
190 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
191
192 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
193 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
194
195 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
196 const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(vi2x01234567);
197 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
198 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk2x01234567), vk_zero_point);
199 i2 += 8;
200
201
202 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
203 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
204
205 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
206 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
207
208 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
209 const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(vi3x01234567);
210 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
211 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk3x01234567), vk_zero_point);
212 i3 += 8;
213
214
215 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
216 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
217
218 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
219 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
220
221 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
222 const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(vi4x01234567);
223 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
224 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk4x01234567), vk_zero_point);
225 i4 += 8;
226
227
228 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
229 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
230
231 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
232 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
233
234 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
235 const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(vi5x01234567);
236 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
237 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point);
238 i5 += 8;
239
240
241 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
242 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
243
244 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
245 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
246
247 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
248 const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(vi6x01234567);
249 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
250 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk6x01234567), vk_zero_point);
251 i6 += 8;
252
253
254 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
255 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
256
257 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
258 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
259
260 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
261 const __m128i vxi7x01234567 = _mm_cvtepu8_epi16(vi7x01234567);
262 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
263 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk7x01234567), vk_zero_point);
264 i7 += 8;
265
266
267 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
268 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
269
270 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
271 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
272
273 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
274 const __m128i vxi8x01234567 = _mm_cvtepu8_epi16(vi8x01234567);
275 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
276 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk8x01234567), vk_zero_point);
277 i8 += 8;
278
279
280 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
281 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
282
283 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
284 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
285
286 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
287 const __m128i vxi9x01234567 = _mm_cvtepu8_epi16(vi9x01234567);
288 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
289 const __m128i vxk9x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk9x01234567), vk_zero_point);
290 i9 += 8;
291
292
293 const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
294 const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567);
295
296 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi));
297 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi));
298
299 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
300 const __m128i vxi10x01234567 = _mm_cvtepu8_epi16(vi10x01234567);
301 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
302 const __m128i vxk10x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk10x01234567), vk_zero_point);
303 i10 += 8;
304
305
306 const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
307 const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567);
308
309 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi));
310 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi));
311
312 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
313 const __m128i vxi11x01234567 = _mm_cvtepu8_epi16(vi11x01234567);
314 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
315 const __m128i vxk11x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk11x01234567), vk_zero_point);
316 i11 += 8;
317
318
319 const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
320 const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567);
321
322 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi));
323 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi));
324
325 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
326 const __m128i vxi12x01234567 = _mm_cvtepu8_epi16(vi12x01234567);
327 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
328 const __m128i vxk12x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk12x01234567), vk_zero_point);
329 i12 += 8;
330
331
332 const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
333 const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567);
334
335 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi));
336 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi));
337
338 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
339 const __m128i vxi13x01234567 = _mm_cvtepu8_epi16(vi13x01234567);
340 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
341 const __m128i vxk13x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk13x01234567), vk_zero_point);
342 i13 += 8;
343
344
345 const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
346 const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567);
347
348 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi));
349 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi));
350
351 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
352 const __m128i vxi14x01234567 = _mm_cvtepu8_epi16(vi14x01234567);
353 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
354 const __m128i vxk14x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk14x01234567), vk_zero_point);
355 i14 += 8;
356
357
358 const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
359 const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567);
360
361 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi));
362 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi));
363
364 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
365 const __m128i vxi15x01234567 = _mm_cvtepu8_epi16(vi15x01234567);
366 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
367 const __m128i vxk15x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk15x01234567), vk_zero_point);
368 i15 += 8;
369
370
371 const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
372 const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567);
373
374 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi));
375 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi));
376
377 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
378 const __m128i vxi16x01234567 = _mm_cvtepu8_epi16(vi16x01234567);
379 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
380 const __m128i vxk16x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk16x01234567), vk_zero_point);
381 i16 += 8;
382
383
384 const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
385 const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567);
386
387 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi));
388 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi));
389
390 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
391 const __m128i vxi17x01234567 = _mm_cvtepu8_epi16(vi17x01234567);
392 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
393 const __m128i vxk17x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk17x01234567), vk_zero_point);
394 i17 += 8;
395
396
397 const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
398 const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567);
399
400 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi));
401 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi));
402
403 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
404 const __m128i vxi18x01234567 = _mm_cvtepu8_epi16(vi18x01234567);
405 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(uint8_t)));
406 const __m128i vxk18x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk18x01234567), vk_zero_point);
407 i18 += 8;
408
409
410 const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
411 const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567);
412
413 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi));
414 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi));
415
416 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
417 const __m128i vxi19x01234567 = _mm_cvtepu8_epi16(vi19x01234567);
418 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(uint8_t)));
419 const __m128i vxk19x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk19x01234567), vk_zero_point);
420 i19 += 8;
421
422
423 const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
424 const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567);
425
426 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi));
427 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi));
428
429 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
430 const __m128i vxi20x01234567 = _mm_cvtepu8_epi16(vi20x01234567);
431 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(uint8_t)));
432 const __m128i vxk20x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk20x01234567), vk_zero_point);
433 i20 += 8;
434
435
436 const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
437 const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567);
438
439 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi));
440 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi));
441
442 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
443 const __m128i vxi21x01234567 = _mm_cvtepu8_epi16(vi21x01234567);
444 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(uint8_t)));
445 const __m128i vxk21x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk21x01234567), vk_zero_point);
446 i21 += 8;
447
448
449 const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
450 const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567);
451
452 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi));
453 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi));
454
455 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
456 const __m128i vxi22x01234567 = _mm_cvtepu8_epi16(vi22x01234567);
457 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(uint8_t)));
458 const __m128i vxk22x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk22x01234567), vk_zero_point);
459 i22 += 8;
460
461
462 const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
463 const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567);
464
465 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi));
466 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi));
467
468 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
469 const __m128i vxi23x01234567 = _mm_cvtepu8_epi16(vi23x01234567);
470 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(uint8_t)));
471 const __m128i vxk23x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk23x01234567), vk_zero_point);
472 i23 += 8;
473
474
475 const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
476 const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567);
477
478 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi));
479 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi));
480
481 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
482 const __m128i vxi24x01234567 = _mm_cvtepu8_epi16(vi24x01234567);
483 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(uint8_t)));
484 const __m128i vxk24x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk24x01234567), vk_zero_point);
485 i24 += 8;
486
487
488 const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
489 const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567);
490
491 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi));
492 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi));
493
494 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(uint8_t));
495
496 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
497 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
498
499 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
500 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
501 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
502
503 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
504 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
505 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
506
507 vacc0123 = _mm_cvtps_epi32(vscaled0123);
508 vacc4567 = _mm_cvtps_epi32(vscaled4567);
509
510 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
511 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
512
513 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
514
515 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
516 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
517
518 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
519 output += 8;
520 }
521 if XNN_UNLIKELY(c != 0) {
522 {
523 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
524 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
525
526
527 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
528 const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(vi0x01234567);
529 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
530 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk0x01234567), vk_zero_point);
531
532
533 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
534 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
535
536 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
537 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
538
539 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
540 const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(vi1x01234567);
541 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
542 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk1x01234567), vk_zero_point);
543
544
545 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
546 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
547
548 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
549 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
550
551 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
552 const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(vi2x01234567);
553 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
554 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk2x01234567), vk_zero_point);
555
556
557 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
558 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
559
560 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
561 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
562
563 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
564 const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(vi3x01234567);
565 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
566 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk3x01234567), vk_zero_point);
567
568
569 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
570 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
571
572 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
573 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
574
575 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
576 const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(vi4x01234567);
577 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
578 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk4x01234567), vk_zero_point);
579
580
581 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
582 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
583
584 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
585 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
586
587 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
588 const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(vi5x01234567);
589 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
590 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point);
591
592
593 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
594 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
595
596 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
597 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
598
599 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
600 const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(vi6x01234567);
601 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
602 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk6x01234567), vk_zero_point);
603
604
605 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
606 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
607
608 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
609 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
610
611 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
612 const __m128i vxi7x01234567 = _mm_cvtepu8_epi16(vi7x01234567);
613 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
614 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk7x01234567), vk_zero_point);
615
616
617 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
618 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
619
620 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
621 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
622
623 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
624 const __m128i vxi8x01234567 = _mm_cvtepu8_epi16(vi8x01234567);
625 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
626 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk8x01234567), vk_zero_point);
627
628
629 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
630 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
631
632 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
633 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
634
635 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
636 const __m128i vxi9x01234567 = _mm_cvtepu8_epi16(vi9x01234567);
637 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
638 const __m128i vxk9x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk9x01234567), vk_zero_point);
639
640
641 const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
642 const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567);
643
644 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi));
645 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi));
646
647 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
648 const __m128i vxi10x01234567 = _mm_cvtepu8_epi16(vi10x01234567);
649 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
650 const __m128i vxk10x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk10x01234567), vk_zero_point);
651
652
653 const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
654 const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567);
655
656 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi));
657 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi));
658
659 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
660 const __m128i vxi11x01234567 = _mm_cvtepu8_epi16(vi11x01234567);
661 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
662 const __m128i vxk11x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk11x01234567), vk_zero_point);
663
664
665 const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
666 const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567);
667
668 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi));
669 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi));
670
671 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
672 const __m128i vxi12x01234567 = _mm_cvtepu8_epi16(vi12x01234567);
673 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
674 const __m128i vxk12x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk12x01234567), vk_zero_point);
675
676
677 const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
678 const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567);
679
680 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi));
681 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi));
682
683 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
684 const __m128i vxi13x01234567 = _mm_cvtepu8_epi16(vi13x01234567);
685 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
686 const __m128i vxk13x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk13x01234567), vk_zero_point);
687
688
689 const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
690 const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567);
691
692 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi));
693 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi));
694
695 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
696 const __m128i vxi14x01234567 = _mm_cvtepu8_epi16(vi14x01234567);
697 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
698 const __m128i vxk14x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk14x01234567), vk_zero_point);
699
700
701 const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
702 const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567);
703
704 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi));
705 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi));
706
707 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
708 const __m128i vxi15x01234567 = _mm_cvtepu8_epi16(vi15x01234567);
709 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
710 const __m128i vxk15x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk15x01234567), vk_zero_point);
711
712
713 const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
714 const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567);
715
716 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi));
717 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi));
718
719 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
720 const __m128i vxi16x01234567 = _mm_cvtepu8_epi16(vi16x01234567);
721 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
722 const __m128i vxk16x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk16x01234567), vk_zero_point);
723
724
725 const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
726 const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567);
727
728 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi));
729 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi));
730
731 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
732 const __m128i vxi17x01234567 = _mm_cvtepu8_epi16(vi17x01234567);
733 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
734 const __m128i vxk17x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk17x01234567), vk_zero_point);
735
736
737 const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
738 const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567);
739
740 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi));
741 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi));
742
743 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
744 const __m128i vxi18x01234567 = _mm_cvtepu8_epi16(vi18x01234567);
745 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(uint8_t)));
746 const __m128i vxk18x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk18x01234567), vk_zero_point);
747
748
749 const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
750 const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567);
751
752 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi));
753 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi));
754
755 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
756 const __m128i vxi19x01234567 = _mm_cvtepu8_epi16(vi19x01234567);
757 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(uint8_t)));
758 const __m128i vxk19x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk19x01234567), vk_zero_point);
759
760
761 const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
762 const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567);
763
764 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi));
765 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi));
766
767 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
768 const __m128i vxi20x01234567 = _mm_cvtepu8_epi16(vi20x01234567);
769 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(uint8_t)));
770 const __m128i vxk20x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk20x01234567), vk_zero_point);
771
772
773 const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
774 const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567);
775
776 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi));
777 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi));
778
779 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
780 const __m128i vxi21x01234567 = _mm_cvtepu8_epi16(vi21x01234567);
781 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(uint8_t)));
782 const __m128i vxk21x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk21x01234567), vk_zero_point);
783
784
785 const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
786 const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567);
787
788 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi));
789 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi));
790
791 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
792 const __m128i vxi22x01234567 = _mm_cvtepu8_epi16(vi22x01234567);
793 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(uint8_t)));
794 const __m128i vxk22x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk22x01234567), vk_zero_point);
795
796
797 const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
798 const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567);
799
800 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi));
801 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi));
802
803 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
804 const __m128i vxi23x01234567 = _mm_cvtepu8_epi16(vi23x01234567);
805 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(uint8_t)));
806 const __m128i vxk23x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk23x01234567), vk_zero_point);
807
808
809 const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
810 const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567);
811
812 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi));
813 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi));
814
815 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
816 const __m128i vxi24x01234567 = _mm_cvtepu8_epi16(vi24x01234567);
817 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(uint8_t)));
818 const __m128i vxk24x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk24x01234567), vk_zero_point);
819
820
821 const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
822 const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567);
823
824 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi));
825 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi));
826
827
828 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
829 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
830
831 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
832 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
833 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
834
835 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
836 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
837 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
838
839 vacc0123 = _mm_cvtps_epi32(vscaled0123);
840 vacc4567 = _mm_cvtps_epi32(vscaled4567);
841
842
843 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
844 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
845
846 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
847
848 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
849
850 if (c & 4) {
851 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
852 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
853 output += 4;
854 }
855 if (c & 2) {
856 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
857 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
858 output += 2;
859 }
860 if (c & 1) {
861 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
862 output += 1;
863 }
864 }
865 }
866
867 output = (uint8_t*) ((uintptr_t) output + output_increment);
868 } while (--output_width != 0);
869 }
870