1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <emmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/unaligned.h>
16
17
xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16_add16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16_add16(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qs8_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 int8_t* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
38 }
39 const int8_t* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
43 }
44 const int8_t* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
48 }
49 const int8_t* i3 = input[3];
50 assert(i3 != NULL);
51 if XNN_UNPREDICTABLE(i3 != zero) {
52 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
53 }
54 const int8_t* i4 = input[4];
55 assert(i4 != NULL);
56 if XNN_UNPREDICTABLE(i4 != zero) {
57 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
58 }
59 const int8_t* i5 = input[5];
60 assert(i5 != NULL);
61 if XNN_UNPREDICTABLE(i5 != zero) {
62 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
63 }
64 const int8_t* i6 = input[6];
65 assert(i6 != NULL);
66 if XNN_UNPREDICTABLE(i6 != zero) {
67 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
68 }
69 const int8_t* i7 = input[7];
70 assert(i7 != NULL);
71 if XNN_UNPREDICTABLE(i7 != zero) {
72 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
73 }
74 const int8_t* i8 = input[8];
75 assert(i8 != NULL);
76 if XNN_UNPREDICTABLE(i8 != zero) {
77 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
78 }
79 const int8_t* i9 = input[9];
80 assert(i9 != NULL);
81 if XNN_UNPREDICTABLE(i9 != zero) {
82 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
83 }
84 const int8_t* i10 = input[10];
85 assert(i10 != NULL);
86 if XNN_UNPREDICTABLE(i10 != zero) {
87 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
88 }
89 const int8_t* i11 = input[11];
90 assert(i11 != NULL);
91 if XNN_UNPREDICTABLE(i11 != zero) {
92 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
93 }
94 const int8_t* i12 = input[12];
95 assert(i12 != NULL);
96 if XNN_UNPREDICTABLE(i12 != zero) {
97 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
98 }
99 const int8_t* i13 = input[13];
100 assert(i13 != NULL);
101 if XNN_UNPREDICTABLE(i13 != zero) {
102 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
103 }
104 const int8_t* i14 = input[14];
105 assert(i14 != NULL);
106 if XNN_UNPREDICTABLE(i14 != zero) {
107 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
108 }
109 const int8_t* i15 = input[15];
110 assert(i15 != NULL);
111 if XNN_UNPREDICTABLE(i15 != zero) {
112 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
113 }
114 const int8_t* i16 = input[16];
115 assert(i16 != NULL);
116 if XNN_UNPREDICTABLE(i16 != zero) {
117 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
118 }
119 const int8_t* i17 = input[17];
120 assert(i17 != NULL);
121 if XNN_UNPREDICTABLE(i17 != zero) {
122 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
123 }
124 const int8_t* i18 = input[18];
125 assert(i18 != NULL);
126 if XNN_UNPREDICTABLE(i18 != zero) {
127 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
128 }
129 const int8_t* i19 = input[19];
130 assert(i19 != NULL);
131 if XNN_UNPREDICTABLE(i19 != zero) {
132 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
133 }
134 const int8_t* i20 = input[20];
135 assert(i20 != NULL);
136 if XNN_UNPREDICTABLE(i20 != zero) {
137 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
138 }
139 const int8_t* i21 = input[21];
140 assert(i21 != NULL);
141 if XNN_UNPREDICTABLE(i21 != zero) {
142 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
143 }
144 const int8_t* i22 = input[22];
145 assert(i22 != NULL);
146 if XNN_UNPREDICTABLE(i22 != zero) {
147 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
148 }
149 const int8_t* i23 = input[23];
150 assert(i23 != NULL);
151 if XNN_UNPREDICTABLE(i23 != zero) {
152 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
153 }
154 const int8_t* i24 = input[24];
155 assert(i24 != NULL);
156 if XNN_UNPREDICTABLE(i24 != zero) {
157 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
158 }
159 input = (const int8_t**) ((uintptr_t) input + input_stride);
160
161 size_t c = channels;
162 const void* w = weights;
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 vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
170 i0 += 8;
171
172 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
173 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
174
175 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
176
177
178 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
179 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
180 i1 += 8;
181
182 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
183 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
184
185 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
186
187 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
188 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
189 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
190
191 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
192 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
193 i2 += 8;
194
195 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
196 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
197
198 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
199
200
201 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
202 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
203 i3 += 8;
204
205 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
206 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
207
208 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
209
210 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
211 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
212 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
213
214 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
215 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
216 i4 += 8;
217
218 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
219 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
220
221 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
222
223
224 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
225 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
226 i5 += 8;
227
228 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
229 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
230
231 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
232
233 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
234 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
235 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
236
237 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
238 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
239 i6 += 8;
240
241 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
242 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
243
244 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
245
246
247 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
248 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
249 i7 += 8;
250
251 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
252 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
253
254 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
255
256 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
257 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
258 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
259
260 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
261 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
262 i8 += 8;
263
264 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
265 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
266
267 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
268
269
270 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
271 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
272 i9 += 8;
273
274 const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8);
275 const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8);
276
277 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
278
279 const __m128i vsignprod9x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
280 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod9x01234567));
281 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod9x01234567));
282
283 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
284 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
285 i10 += 8;
286
287 const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8);
288 const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8);
289
290 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
291
292
293 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
294 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
295 i11 += 8;
296
297 const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8);
298 const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8);
299
300 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
301
302 const __m128i vsignprod11x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
303 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod11x01234567));
304 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod11x01234567));
305
306 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
307 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
308 i12 += 8;
309
310 const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8);
311 const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8);
312
313 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
314
315
316 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
317 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
318 i13 += 8;
319
320 const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8);
321 const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8);
322
323 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
324
325 const __m128i vsignprod13x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
326 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod13x01234567));
327 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod13x01234567));
328
329 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
330 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
331 i14 += 8;
332
333 const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8);
334 const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8);
335
336 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
337
338
339 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
340 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
341 i15 += 8;
342
343 const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8);
344 const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8);
345
346 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
347
348 const __m128i vsignprod15x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
349 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod15x01234567));
350 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod15x01234567));
351
352 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
353 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
354 i16 += 8;
355
356 const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8);
357 const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8);
358
359 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
360
361
362 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
363 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
364 i17 += 8;
365
366 const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8);
367 const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8);
368
369 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
370
371 const __m128i vsignprod17x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
372 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod17x01234567));
373 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod17x01234567));
374
375 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
376 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
377 i18 += 8;
378
379 const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8);
380 const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8);
381
382 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
383
384
385 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
386 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
387 i19 += 8;
388
389 const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8);
390 const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8);
391
392 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
393
394 const __m128i vsignprod19x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
395 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod19x01234567));
396 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod19x01234567));
397
398 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
399 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
400 i20 += 8;
401
402 const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8);
403 const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8);
404
405 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
406
407
408 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
409 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
410 i21 += 8;
411
412 const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8);
413 const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8);
414
415 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
416
417 const __m128i vsignprod21x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
418 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod21x01234567));
419 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod21x01234567));
420
421 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
422 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
423 i22 += 8;
424
425 const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8);
426 const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8);
427
428 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
429
430
431 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
432 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
433 i23 += 8;
434
435 const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8);
436 const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8);
437
438 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
439
440 const __m128i vsignprod23x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
441 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod23x01234567));
442 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod23x01234567));
443
444 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
445 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
446 i24 += 8;
447
448 const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8);
449 const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8);
450
451 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
452
453 const __m128i vsignprod24x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
454 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod24x01234567));
455 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod24x01234567));
456
457 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t));
458
459 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
460 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
461
462 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
463 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
464 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
465
466 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
467 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
468 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
469
470 vacc0123 = _mm_cvtps_epi32(vscaled0123);
471 vacc4567 = _mm_cvtps_epi32(vscaled4567);
472
473 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
474 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
475
476 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
477 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
478
479 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
480
481
482 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
483 output += 8;
484 }
485 if XNN_UNLIKELY(c != 0) {
486 {
487 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
488 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
489
490
491 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
492 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
493
494 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
495 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
496
497 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
498
499
500 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
501 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
502
503 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
504 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
505
506 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
507
508 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
509 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
510 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
511
512 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
513 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
514
515 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
516 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
517
518 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
519
520
521 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
522 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
523
524 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
525 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
526
527 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
528
529 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
530 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
531 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
532
533 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
534 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
535
536 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
537 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
538
539 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
540
541
542 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
543 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
544
545 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
546 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
547
548 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
549
550 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
551 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
552 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
553
554 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
555 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
556
557 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
558 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
559
560 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
561
562
563 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
564 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
565
566 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
567 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
568
569 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
570
571 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
572 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
573 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
574
575 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
576 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
577
578 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
579 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
580
581 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
582
583
584 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
585 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
586
587 const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8);
588 const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8);
589
590 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
591
592 const __m128i vsignprod9x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
593 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod9x01234567));
594 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod9x01234567));
595
596 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
597 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
598
599 const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8);
600 const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8);
601
602 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
603
604
605 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
606 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
607
608 const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8);
609 const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8);
610
611 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
612
613 const __m128i vsignprod11x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
614 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod11x01234567));
615 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod11x01234567));
616
617 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
618 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
619
620 const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8);
621 const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8);
622
623 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
624
625
626 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
627 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
628
629 const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8);
630 const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8);
631
632 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
633
634 const __m128i vsignprod13x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
635 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod13x01234567));
636 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod13x01234567));
637
638 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
639 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
640
641 const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8);
642 const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8);
643
644 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
645
646
647 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
648 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
649
650 const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8);
651 const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8);
652
653 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
654
655 const __m128i vsignprod15x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
656 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod15x01234567));
657 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod15x01234567));
658
659 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
660 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
661
662 const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8);
663 const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8);
664
665 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
666
667
668 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
669 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
670
671 const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8);
672 const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8);
673
674 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
675
676 const __m128i vsignprod17x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
677 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod17x01234567));
678 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod17x01234567));
679
680 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
681 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
682
683 const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8);
684 const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8);
685
686 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
687
688
689 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
690 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
691
692 const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8);
693 const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8);
694
695 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
696
697 const __m128i vsignprod19x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
698 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod19x01234567));
699 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod19x01234567));
700
701 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
702 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
703
704 const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8);
705 const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8);
706
707 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
708
709
710 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
711 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
712
713 const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8);
714 const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8);
715
716 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
717
718 const __m128i vsignprod21x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
719 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod21x01234567));
720 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod21x01234567));
721
722 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
723 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
724
725 const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8);
726 const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8);
727
728 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
729
730
731 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
732 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
733
734 const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8);
735 const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8);
736
737 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
738
739 const __m128i vsignprod23x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
740 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod23x01234567));
741 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod23x01234567));
742
743 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
744 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
745
746 const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8);
747 const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8);
748
749 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
750
751 const __m128i vsignprod24x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
752 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod24x01234567));
753 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod24x01234567));
754
755
756 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
757 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
758
759 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
760 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
761 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
762
763 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
764 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
765 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
766
767 vacc0123 = _mm_cvtps_epi32(vscaled0123);
768 vacc4567 = _mm_cvtps_epi32(vscaled4567);
769
770
771 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
772 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
773
774 vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
775
776 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
777
778
779 if (c & 4) {
780 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
781 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
782 output += 4;
783 }
784 if (c & 2) {
785 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
786 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
787 output += 2;
788 }
789 if (c & 1) {
790 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
791 output += 1;
792 }
793 }
794 }
795
796 output = (int8_t*) ((uintptr_t) output + output_increment);
797 } while (--output_width != 0);
798 }
799