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_qc8_dwconv_minmax_fp32_ukernel_up16x25__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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__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_qc8_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 >= 16; c -= 16) {
164 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
165 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
166 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
167 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
168
169
170 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
171 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
172 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
173 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
174 i0 += 16;
175
176 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
177 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
178 const __m128i vxi0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x89ABCDEF, vi0x89ABCDEF), 8);
179 const __m128i vxk0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x89ABCDEF, vk0x89ABCDEF), 8);
180
181 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
182 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
183
184
185 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
186 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
187 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
188 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
189 i1 += 16;
190
191 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
192 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
193 const __m128i vxi1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x89ABCDEF, vi1x89ABCDEF), 8);
194 const __m128i vxk1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x89ABCDEF, vk1x89ABCDEF), 8);
195
196 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
197 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF));
198
199 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
200 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
201 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
202 const __m128i vsignprod1x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
203 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod1x89ABCDEF));
204 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod1x89ABCDEF));
205
206 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
207 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
208 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
209 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
210 i2 += 16;
211
212 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
213 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
214 const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8);
215 const __m128i vxk2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x89ABCDEF, vk2x89ABCDEF), 8);
216
217 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
218 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
219
220
221 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
222 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
223 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
224 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
225 i3 += 16;
226
227 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
228 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
229 const __m128i vxi3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x89ABCDEF, vi3x89ABCDEF), 8);
230 const __m128i vxk3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x89ABCDEF, vk3x89ABCDEF), 8);
231
232 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
233 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF));
234
235 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
236 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
237 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
238 const __m128i vsignprod3x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
239 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod3x89ABCDEF));
240 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod3x89ABCDEF));
241
242 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
243 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
244 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
245 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
246 i4 += 16;
247
248 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
249 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
250 const __m128i vxi4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x89ABCDEF, vi4x89ABCDEF), 8);
251 const __m128i vxk4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x89ABCDEF, vk4x89ABCDEF), 8);
252
253 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
254 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
255
256
257 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
258 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
259 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
260 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
261 i5 += 16;
262
263 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
264 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
265 const __m128i vxi5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x89ABCDEF, vi5x89ABCDEF), 8);
266 const __m128i vxk5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x89ABCDEF, vk5x89ABCDEF), 8);
267
268 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
269 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF));
270
271 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
272 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
273 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
274 const __m128i vsignprod5x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
275 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod5x89ABCDEF));
276 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod5x89ABCDEF));
277
278 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
279 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
280 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
281 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
282 i6 += 16;
283
284 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
285 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
286 const __m128i vxi6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x89ABCDEF, vi6x89ABCDEF), 8);
287 const __m128i vxk6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x89ABCDEF, vk6x89ABCDEF), 8);
288
289 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
290 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
291
292
293 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
294 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
295 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
296 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
297 i7 += 16;
298
299 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
300 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
301 const __m128i vxi7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x89ABCDEF, vi7x89ABCDEF), 8);
302 const __m128i vxk7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x89ABCDEF, vk7x89ABCDEF), 8);
303
304 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
305 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF));
306
307 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
308 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
309 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
310 const __m128i vsignprod7x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
311 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod7x89ABCDEF));
312 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod7x89ABCDEF));
313
314 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
315 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
316 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
317 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
318 i8 += 16;
319
320 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
321 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
322 const __m128i vxi8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x89ABCDEF, vi8x89ABCDEF), 8);
323 const __m128i vxk8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x89ABCDEF, vk8x89ABCDEF), 8);
324
325 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
326 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
327
328
329 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
330 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
331 const __m128i vi9x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i9 + 8));
332 const __m128i vk9x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t)));
333 i9 += 16;
334
335 const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8);
336 const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8);
337 const __m128i vxi9x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x89ABCDEF, vi9x89ABCDEF), 8);
338 const __m128i vxk9x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x89ABCDEF, vk9x89ABCDEF), 8);
339
340 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
341 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF));
342
343 const __m128i vsignprod9x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
344 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod9x01234567));
345 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod9x01234567));
346 const __m128i vsignprod9x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
347 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod9x89ABCDEF));
348 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod9x89ABCDEF));
349
350 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
351 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)));
352 const __m128i vi10x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i10 + 8));
353 const __m128i vk10x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t)));
354 i10 += 16;
355
356 const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8);
357 const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8);
358 const __m128i vxi10x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x89ABCDEF, vi10x89ABCDEF), 8);
359 const __m128i vxk10x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x89ABCDEF, vk10x89ABCDEF), 8);
360
361 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
362 vprod89ABCDEF = _mm_mullo_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
363
364
365 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
366 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)));
367 const __m128i vi11x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i11 + 8));
368 const __m128i vk11x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t)));
369 i11 += 16;
370
371 const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8);
372 const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8);
373 const __m128i vxi11x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x89ABCDEF, vi11x89ABCDEF), 8);
374 const __m128i vxk11x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x89ABCDEF, vk11x89ABCDEF), 8);
375
376 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
377 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF));
378
379 const __m128i vsignprod11x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
380 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod11x01234567));
381 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod11x01234567));
382 const __m128i vsignprod11x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
383 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod11x89ABCDEF));
384 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod11x89ABCDEF));
385
386 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
387 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)));
388 const __m128i vi12x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i12 + 8));
389 const __m128i vk12x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t)));
390 i12 += 16;
391
392 const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8);
393 const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8);
394 const __m128i vxi12x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x89ABCDEF, vi12x89ABCDEF), 8);
395 const __m128i vxk12x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x89ABCDEF, vk12x89ABCDEF), 8);
396
397 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
398 vprod89ABCDEF = _mm_mullo_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
399
400
401 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
402 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)));
403 const __m128i vi13x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i13 + 8));
404 const __m128i vk13x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t)));
405 i13 += 16;
406
407 const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8);
408 const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8);
409 const __m128i vxi13x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x89ABCDEF, vi13x89ABCDEF), 8);
410 const __m128i vxk13x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x89ABCDEF, vk13x89ABCDEF), 8);
411
412 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
413 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF));
414
415 const __m128i vsignprod13x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
416 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod13x01234567));
417 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod13x01234567));
418 const __m128i vsignprod13x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
419 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod13x89ABCDEF));
420 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod13x89ABCDEF));
421
422 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
423 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)));
424 const __m128i vi14x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i14 + 8));
425 const __m128i vk14x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t)));
426 i14 += 16;
427
428 const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8);
429 const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8);
430 const __m128i vxi14x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x89ABCDEF, vi14x89ABCDEF), 8);
431 const __m128i vxk14x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x89ABCDEF, vk14x89ABCDEF), 8);
432
433 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
434 vprod89ABCDEF = _mm_mullo_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
435
436
437 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
438 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)));
439 const __m128i vi15x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i15 + 8));
440 const __m128i vk15x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t)));
441 i15 += 16;
442
443 const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8);
444 const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8);
445 const __m128i vxi15x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x89ABCDEF, vi15x89ABCDEF), 8);
446 const __m128i vxk15x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x89ABCDEF, vk15x89ABCDEF), 8);
447
448 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
449 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF));
450
451 const __m128i vsignprod15x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
452 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod15x01234567));
453 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod15x01234567));
454 const __m128i vsignprod15x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
455 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod15x89ABCDEF));
456 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod15x89ABCDEF));
457
458 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
459 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)));
460 const __m128i vi16x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i16 + 8));
461 const __m128i vk16x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t)));
462 i16 += 16;
463
464 const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8);
465 const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8);
466 const __m128i vxi16x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x89ABCDEF, vi16x89ABCDEF), 8);
467 const __m128i vxk16x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x89ABCDEF, vk16x89ABCDEF), 8);
468
469 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
470 vprod89ABCDEF = _mm_mullo_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
471
472
473 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
474 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)));
475 const __m128i vi17x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i17 + 8));
476 const __m128i vk17x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t)));
477 i17 += 16;
478
479 const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8);
480 const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8);
481 const __m128i vxi17x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x89ABCDEF, vi17x89ABCDEF), 8);
482 const __m128i vxk17x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x89ABCDEF, vk17x89ABCDEF), 8);
483
484 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
485 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF));
486
487 const __m128i vsignprod17x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
488 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod17x01234567));
489 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod17x01234567));
490 const __m128i vsignprod17x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
491 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod17x89ABCDEF));
492 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod17x89ABCDEF));
493
494 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
495 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)));
496 const __m128i vi18x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i18 + 8));
497 const __m128i vk18x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t)));
498 i18 += 16;
499
500 const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8);
501 const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8);
502 const __m128i vxi18x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x89ABCDEF, vi18x89ABCDEF), 8);
503 const __m128i vxk18x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x89ABCDEF, vk18x89ABCDEF), 8);
504
505 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
506 vprod89ABCDEF = _mm_mullo_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
507
508
509 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
510 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)));
511 const __m128i vi19x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i19 + 8));
512 const __m128i vk19x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t)));
513 i19 += 16;
514
515 const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8);
516 const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8);
517 const __m128i vxi19x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x89ABCDEF, vi19x89ABCDEF), 8);
518 const __m128i vxk19x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x89ABCDEF, vk19x89ABCDEF), 8);
519
520 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
521 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF));
522
523 const __m128i vsignprod19x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
524 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod19x01234567));
525 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod19x01234567));
526 const __m128i vsignprod19x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
527 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod19x89ABCDEF));
528 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod19x89ABCDEF));
529
530 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
531 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)));
532 const __m128i vi20x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i20 + 8));
533 const __m128i vk20x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t)));
534 i20 += 16;
535
536 const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8);
537 const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8);
538 const __m128i vxi20x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x89ABCDEF, vi20x89ABCDEF), 8);
539 const __m128i vxk20x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x89ABCDEF, vk20x89ABCDEF), 8);
540
541 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
542 vprod89ABCDEF = _mm_mullo_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
543
544
545 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
546 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)));
547 const __m128i vi21x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i21 + 8));
548 const __m128i vk21x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t)));
549 i21 += 16;
550
551 const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8);
552 const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8);
553 const __m128i vxi21x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x89ABCDEF, vi21x89ABCDEF), 8);
554 const __m128i vxk21x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x89ABCDEF, vk21x89ABCDEF), 8);
555
556 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
557 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF));
558
559 const __m128i vsignprod21x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
560 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod21x01234567));
561 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod21x01234567));
562 const __m128i vsignprod21x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
563 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod21x89ABCDEF));
564 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod21x89ABCDEF));
565
566 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
567 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)));
568 const __m128i vi22x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i22 + 8));
569 const __m128i vk22x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t)));
570 i22 += 16;
571
572 const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8);
573 const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8);
574 const __m128i vxi22x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x89ABCDEF, vi22x89ABCDEF), 8);
575 const __m128i vxk22x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x89ABCDEF, vk22x89ABCDEF), 8);
576
577 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
578 vprod89ABCDEF = _mm_mullo_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
579
580
581 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
582 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)));
583 const __m128i vi23x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i23 + 8));
584 const __m128i vk23x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t)));
585 i23 += 16;
586
587 const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8);
588 const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8);
589 const __m128i vxi23x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x89ABCDEF, vi23x89ABCDEF), 8);
590 const __m128i vxk23x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x89ABCDEF, vk23x89ABCDEF), 8);
591
592 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
593 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF));
594
595 const __m128i vsignprod23x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
596 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod23x01234567));
597 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod23x01234567));
598 const __m128i vsignprod23x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
599 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod23x89ABCDEF));
600 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod23x89ABCDEF));
601
602 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
603 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)));
604 const __m128i vi24x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i24 + 8));
605 const __m128i vk24x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t)));
606 i24 += 16;
607
608 const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8);
609 const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8);
610 const __m128i vxi24x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x89ABCDEF, vi24x89ABCDEF), 8);
611 const __m128i vxk24x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x89ABCDEF, vk24x89ABCDEF), 8);
612
613 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
614 vprod89ABCDEF = _mm_mullo_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
615
616 const __m128i vsignprod24x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
617 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod24x01234567));
618 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod24x01234567));
619 const __m128i vsignprod24x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
620 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod24x89ABCDEF));
621 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod24x89ABCDEF));
622
623 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
624
625 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
626 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
627 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
628 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
629
630 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
631 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
632 const __m128 vscale89AB = _mm_loadu_ps((const float*) w + 8);
633 const __m128 vscaleCDEF = _mm_loadu_ps((const float*) w + 12);
634 w = (const void*) ((const float*) w + 16);
635 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
636 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
637 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale89AB);
638 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscaleCDEF);
639
640 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
641 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
642 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
643 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
644 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
645
646 vacc0123 = _mm_cvtps_epi32(vscaled0123);
647 vacc4567 = _mm_cvtps_epi32(vscaled4567);
648 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
649 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
650
651 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
652 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
653 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
654
655 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
656 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
657 vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
658
659 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
660
661
662 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
663 output += 16;
664 }
665 if XNN_UNLIKELY(c != 0) {
666 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
667 do {
668 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
669 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
670
671
672 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
673 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
674 i0 += 8;
675
676 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
677 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
678
679 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
680
681
682 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
683 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
684 i1 += 8;
685
686 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
687 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
688
689 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
690
691 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
692 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
693 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
694
695 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
696 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
697 i2 += 8;
698
699 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
700 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
701
702 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
703
704
705 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
706 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
707 i3 += 8;
708
709 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
710 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
711
712 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
713
714 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
715 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
716 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
717
718 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
719 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
720 i4 += 8;
721
722 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
723 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
724
725 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
726
727
728 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
729 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
730 i5 += 8;
731
732 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
733 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
734
735 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
736
737 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
738 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
739 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
740
741 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
742 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
743 i6 += 8;
744
745 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
746 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
747
748 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
749
750
751 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
752 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
753 i7 += 8;
754
755 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
756 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
757
758 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
759
760 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
761 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
762 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
763
764 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
765 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
766 i8 += 8;
767
768 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
769 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
770
771 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
772
773
774 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
775 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
776 i9 += 8;
777
778 const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8);
779 const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8);
780
781 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
782
783 const __m128i vsignprod9x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
784 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod9x01234567));
785 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod9x01234567));
786
787 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
788 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) (k + 160));
789 i10 += 8;
790
791 const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8);
792 const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8);
793
794 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
795
796
797 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
798 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) (k + 176));
799 i11 += 8;
800
801 const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8);
802 const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8);
803
804 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
805
806 const __m128i vsignprod11x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
807 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod11x01234567));
808 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod11x01234567));
809
810 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
811 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
812 i12 += 8;
813
814 const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8);
815 const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8);
816
817 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
818
819
820 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
821 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) (k + 208));
822 i13 += 8;
823
824 const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8);
825 const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8);
826
827 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
828
829 const __m128i vsignprod13x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
830 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod13x01234567));
831 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod13x01234567));
832
833 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
834 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) (k + 224));
835 i14 += 8;
836
837 const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8);
838 const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8);
839
840 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
841
842
843 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
844 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) (k + 240));
845 i15 += 8;
846
847 const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8);
848 const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8);
849
850 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
851
852 const __m128i vsignprod15x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
853 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod15x01234567));
854 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod15x01234567));
855
856 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
857 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) (k + 256));
858 i16 += 8;
859
860 const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8);
861 const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8);
862
863 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
864
865
866 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
867 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) (k + 272));
868 i17 += 8;
869
870 const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8);
871 const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8);
872
873 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
874
875 const __m128i vsignprod17x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
876 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod17x01234567));
877 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod17x01234567));
878
879 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
880 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) (k + 288));
881 i18 += 8;
882
883 const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8);
884 const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8);
885
886 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
887
888
889 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
890 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) (k + 304));
891 i19 += 8;
892
893 const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8);
894 const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8);
895
896 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
897
898 const __m128i vsignprod19x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
899 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod19x01234567));
900 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod19x01234567));
901
902 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
903 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) (k + 320));
904 i20 += 8;
905
906 const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8);
907 const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8);
908
909 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
910
911
912 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
913 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) (k + 336));
914 i21 += 8;
915
916 const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8);
917 const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8);
918
919 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
920
921 const __m128i vsignprod21x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
922 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod21x01234567));
923 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod21x01234567));
924
925 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
926 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) (k + 352));
927 i22 += 8;
928
929 const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8);
930 const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8);
931
932 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
933
934
935 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
936 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) (k + 368));
937 i23 += 8;
938
939 const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8);
940 const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8);
941
942 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
943
944 const __m128i vsignprod23x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
945 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod23x01234567));
946 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod23x01234567));
947
948 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
949 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) (k + 384));
950 i24 += 8;
951
952 const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8);
953 const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8);
954
955 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
956
957 const __m128i vsignprod24x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
958 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod24x01234567));
959 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod24x01234567));
960
961 k += 8;
962
963 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
964 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
965
966 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t)));
967 const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t) + 4 * sizeof(float)));
968 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
969 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
970
971 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
972 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
973 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
974
975 vacc0123 = _mm_cvtps_epi32(vscaled0123);
976 vacc4567 = _mm_cvtps_epi32(vscaled4567);
977
978 w = (const void*) ((const int32_t*) w + 8);
979
980 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
981 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
982
983 vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
984
985 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
986
987
988 if XNN_LIKELY(c >= 8) {
989 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
990 output += 8;
991 c -= 8;
992 } else {
993 if (c & 4) {
994 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
995 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
996 output += 4;
997 }
998 if (c & 2) {
999 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
1000 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
1001 output += 2;
1002 }
1003 if (c & 1) {
1004 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
1005 output += 1;
1006 }
1007 c = 0;
1008 }
1009 } while (c != 0);
1010 }
1011
1012 output = (int8_t*) ((uintptr_t) output + output_increment);
1013 } while (--output_width != 0);
1014 }
1015