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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/dwconv.h>
20
21
xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__xop_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)])22 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__xop_mul16_add16(
23 size_t channels,
24 size_t output_width,
25 const int8_t** input,
26 const void* weights,
27 int8_t* output,
28 size_t input_stride,
29 size_t output_increment,
30 size_t input_offset,
31 const int8_t* zero,
32 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(channels != 0);
35 assert(output_width != 0);
36
37 do {
38 const int8_t* i0 = input[0];
39 assert(i0 != NULL);
40 if XNN_UNPREDICTABLE(i0 != zero) {
41 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
42 }
43 const int8_t* i1 = input[1];
44 assert(i1 != NULL);
45 if XNN_UNPREDICTABLE(i1 != zero) {
46 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
47 }
48 const int8_t* i2 = input[2];
49 assert(i2 != NULL);
50 if XNN_UNPREDICTABLE(i2 != zero) {
51 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
52 }
53 const int8_t* i3 = input[3];
54 assert(i3 != NULL);
55 if XNN_UNPREDICTABLE(i3 != zero) {
56 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
57 }
58 const int8_t* i4 = input[4];
59 assert(i4 != NULL);
60 if XNN_UNPREDICTABLE(i4 != zero) {
61 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
62 }
63 const int8_t* i5 = input[5];
64 assert(i5 != NULL);
65 if XNN_UNPREDICTABLE(i5 != zero) {
66 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
67 }
68 const int8_t* i6 = input[6];
69 assert(i6 != NULL);
70 if XNN_UNPREDICTABLE(i6 != zero) {
71 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
72 }
73 const int8_t* i7 = input[7];
74 assert(i7 != NULL);
75 if XNN_UNPREDICTABLE(i7 != zero) {
76 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
77 }
78 const int8_t* i8 = input[8];
79 assert(i8 != NULL);
80 if XNN_UNPREDICTABLE(i8 != zero) {
81 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
82 }
83 const int8_t* i9 = input[9];
84 assert(i9 != NULL);
85 if XNN_UNPREDICTABLE(i9 != zero) {
86 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
87 }
88 const int8_t* i10 = input[10];
89 assert(i10 != NULL);
90 if XNN_UNPREDICTABLE(i10 != zero) {
91 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
92 }
93 const int8_t* i11 = input[11];
94 assert(i11 != NULL);
95 if XNN_UNPREDICTABLE(i11 != zero) {
96 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
97 }
98 const int8_t* i12 = input[12];
99 assert(i12 != NULL);
100 if XNN_UNPREDICTABLE(i12 != zero) {
101 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
102 }
103 const int8_t* i13 = input[13];
104 assert(i13 != NULL);
105 if XNN_UNPREDICTABLE(i13 != zero) {
106 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
107 }
108 const int8_t* i14 = input[14];
109 assert(i14 != NULL);
110 if XNN_UNPREDICTABLE(i14 != zero) {
111 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
112 }
113 const int8_t* i15 = input[15];
114 assert(i15 != NULL);
115 if XNN_UNPREDICTABLE(i15 != zero) {
116 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
117 }
118 const int8_t* i16 = input[16];
119 assert(i16 != NULL);
120 if XNN_UNPREDICTABLE(i16 != zero) {
121 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
122 }
123 const int8_t* i17 = input[17];
124 assert(i17 != NULL);
125 if XNN_UNPREDICTABLE(i17 != zero) {
126 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
127 }
128 const int8_t* i18 = input[18];
129 assert(i18 != NULL);
130 if XNN_UNPREDICTABLE(i18 != zero) {
131 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
132 }
133 const int8_t* i19 = input[19];
134 assert(i19 != NULL);
135 if XNN_UNPREDICTABLE(i19 != zero) {
136 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
137 }
138 const int8_t* i20 = input[20];
139 assert(i20 != NULL);
140 if XNN_UNPREDICTABLE(i20 != zero) {
141 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
142 }
143 const int8_t* i21 = input[21];
144 assert(i21 != NULL);
145 if XNN_UNPREDICTABLE(i21 != zero) {
146 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
147 }
148 const int8_t* i22 = input[22];
149 assert(i22 != NULL);
150 if XNN_UNPREDICTABLE(i22 != zero) {
151 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
152 }
153 const int8_t* i23 = input[23];
154 assert(i23 != NULL);
155 if XNN_UNPREDICTABLE(i23 != zero) {
156 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
157 }
158 const int8_t* i24 = input[24];
159 assert(i24 != NULL);
160 if XNN_UNPREDICTABLE(i24 != zero) {
161 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
162 }
163 input = (const int8_t**) ((uintptr_t) input + input_stride);
164
165 size_t c = channels;
166 const void* w = weights;
167 for (; c >= 16; c -= 16) {
168 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
169 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
170 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
171 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
172
173
174 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
175 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
176 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
177 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
178 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
179 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
180 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
181 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
182 i0 += 16;
183
184
185 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
186 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
187
188
189 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
190 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
191 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
192 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
193 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
194 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
195 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
196 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
197 i1 += 16;
198
199
200 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
201 vprod89ABCDEF = _mm_macc_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF, vprod89ABCDEF);
202
203 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
204 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
205 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
206 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
207
208 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
209 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
210 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
211 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
212 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
213 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
214 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
215 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
216 i2 += 16;
217
218
219 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
220 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
221
222
223 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
224 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
225 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
226 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
227 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
228 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
229 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
230 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
231 i3 += 16;
232
233
234 vprod01234567 = _mm_macc_epi16(vxi3x01234567, vxk3x01234567, vprod01234567);
235 vprod89ABCDEF = _mm_macc_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF, vprod89ABCDEF);
236
237 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
238 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
239 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
240 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
241
242 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
243 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
244 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
245 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
246 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
247 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
248 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
249 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
250 i4 += 16;
251
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 vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
259 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
260 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
261 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
262 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
263 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
264 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
265 i5 += 16;
266
267
268 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567);
269 vprod89ABCDEF = _mm_macc_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF, vprod89ABCDEF);
270
271 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
272 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
273 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
274 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
275
276 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
277 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
278 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
279 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
280 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
281 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
282 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
283 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
284 i6 += 16;
285
286
287 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
288 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
289
290
291 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
292 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
293 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
294 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
295 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
296 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
297 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
298 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
299 i7 += 16;
300
301
302 vprod01234567 = _mm_macc_epi16(vxi7x01234567, vxk7x01234567, vprod01234567);
303 vprod89ABCDEF = _mm_macc_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF, vprod89ABCDEF);
304
305 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
306 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
307 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
308 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
309
310 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
311 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
312 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
313 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
314 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
315 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
316 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
317 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
318 i8 += 16;
319
320
321 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
322 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
323
324
325 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
326 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
327 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
328 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
329 const __m128i vi9x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i9 + 8));
330 const __m128i vxi9x89ABCDEF = _mm_cvtepi8_epi16(vi9x89ABCDEF);
331 const __m128i vk9x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t)));
332 const __m128i vxk9x89ABCDEF = _mm_cvtepi8_epi16(vk9x89ABCDEF);
333 i9 += 16;
334
335
336 vprod01234567 = _mm_macc_epi16(vxi9x01234567, vxk9x01234567, vprod01234567);
337 vprod89ABCDEF = _mm_macc_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF, vprod89ABCDEF);
338
339 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
340 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
341 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
342 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
343
344 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
345 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
346 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)));
347 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
348 const __m128i vi10x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i10 + 8));
349 const __m128i vxi10x89ABCDEF = _mm_cvtepi8_epi16(vi10x89ABCDEF);
350 const __m128i vk10x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t)));
351 const __m128i vxk10x89ABCDEF = _mm_cvtepi8_epi16(vk10x89ABCDEF);
352 i10 += 16;
353
354
355 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
356 vprod89ABCDEF = _mm_mullo_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
357
358
359 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
360 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
361 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)));
362 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
363 const __m128i vi11x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i11 + 8));
364 const __m128i vxi11x89ABCDEF = _mm_cvtepi8_epi16(vi11x89ABCDEF);
365 const __m128i vk11x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t)));
366 const __m128i vxk11x89ABCDEF = _mm_cvtepi8_epi16(vk11x89ABCDEF);
367 i11 += 16;
368
369
370 vprod01234567 = _mm_macc_epi16(vxi11x01234567, vxk11x01234567, vprod01234567);
371 vprod89ABCDEF = _mm_macc_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF, vprod89ABCDEF);
372
373 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
374 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
375 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
376 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
377
378 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
379 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
380 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)));
381 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
382 const __m128i vi12x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i12 + 8));
383 const __m128i vxi12x89ABCDEF = _mm_cvtepi8_epi16(vi12x89ABCDEF);
384 const __m128i vk12x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t)));
385 const __m128i vxk12x89ABCDEF = _mm_cvtepi8_epi16(vk12x89ABCDEF);
386 i12 += 16;
387
388
389 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
390 vprod89ABCDEF = _mm_mullo_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
391
392
393 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
394 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
395 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)));
396 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
397 const __m128i vi13x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i13 + 8));
398 const __m128i vxi13x89ABCDEF = _mm_cvtepi8_epi16(vi13x89ABCDEF);
399 const __m128i vk13x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t)));
400 const __m128i vxk13x89ABCDEF = _mm_cvtepi8_epi16(vk13x89ABCDEF);
401 i13 += 16;
402
403
404 vprod01234567 = _mm_macc_epi16(vxi13x01234567, vxk13x01234567, vprod01234567);
405 vprod89ABCDEF = _mm_macc_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF, vprod89ABCDEF);
406
407 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
408 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
409 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
410 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
411
412 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
413 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
414 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)));
415 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
416 const __m128i vi14x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i14 + 8));
417 const __m128i vxi14x89ABCDEF = _mm_cvtepi8_epi16(vi14x89ABCDEF);
418 const __m128i vk14x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t)));
419 const __m128i vxk14x89ABCDEF = _mm_cvtepi8_epi16(vk14x89ABCDEF);
420 i14 += 16;
421
422
423 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
424 vprod89ABCDEF = _mm_mullo_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
425
426
427 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
428 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
429 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)));
430 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
431 const __m128i vi15x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i15 + 8));
432 const __m128i vxi15x89ABCDEF = _mm_cvtepi8_epi16(vi15x89ABCDEF);
433 const __m128i vk15x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t)));
434 const __m128i vxk15x89ABCDEF = _mm_cvtepi8_epi16(vk15x89ABCDEF);
435 i15 += 16;
436
437
438 vprod01234567 = _mm_macc_epi16(vxi15x01234567, vxk15x01234567, vprod01234567);
439 vprod89ABCDEF = _mm_macc_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF, vprod89ABCDEF);
440
441 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
442 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
443 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
444 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
445
446 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
447 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
448 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)));
449 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
450 const __m128i vi16x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i16 + 8));
451 const __m128i vxi16x89ABCDEF = _mm_cvtepi8_epi16(vi16x89ABCDEF);
452 const __m128i vk16x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t)));
453 const __m128i vxk16x89ABCDEF = _mm_cvtepi8_epi16(vk16x89ABCDEF);
454 i16 += 16;
455
456
457 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
458 vprod89ABCDEF = _mm_mullo_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
459
460
461 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
462 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
463 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)));
464 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
465 const __m128i vi17x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i17 + 8));
466 const __m128i vxi17x89ABCDEF = _mm_cvtepi8_epi16(vi17x89ABCDEF);
467 const __m128i vk17x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t)));
468 const __m128i vxk17x89ABCDEF = _mm_cvtepi8_epi16(vk17x89ABCDEF);
469 i17 += 16;
470
471
472 vprod01234567 = _mm_macc_epi16(vxi17x01234567, vxk17x01234567, vprod01234567);
473 vprod89ABCDEF = _mm_macc_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF, vprod89ABCDEF);
474
475 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
476 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
477 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
478 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
479
480 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
481 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
482 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)));
483 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
484 const __m128i vi18x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i18 + 8));
485 const __m128i vxi18x89ABCDEF = _mm_cvtepi8_epi16(vi18x89ABCDEF);
486 const __m128i vk18x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t)));
487 const __m128i vxk18x89ABCDEF = _mm_cvtepi8_epi16(vk18x89ABCDEF);
488 i18 += 16;
489
490
491 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
492 vprod89ABCDEF = _mm_mullo_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
493
494
495 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
496 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
497 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)));
498 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
499 const __m128i vi19x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i19 + 8));
500 const __m128i vxi19x89ABCDEF = _mm_cvtepi8_epi16(vi19x89ABCDEF);
501 const __m128i vk19x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t)));
502 const __m128i vxk19x89ABCDEF = _mm_cvtepi8_epi16(vk19x89ABCDEF);
503 i19 += 16;
504
505
506 vprod01234567 = _mm_macc_epi16(vxi19x01234567, vxk19x01234567, vprod01234567);
507 vprod89ABCDEF = _mm_macc_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF, vprod89ABCDEF);
508
509 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
510 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
511 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
512 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
513
514 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
515 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
516 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)));
517 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
518 const __m128i vi20x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i20 + 8));
519 const __m128i vxi20x89ABCDEF = _mm_cvtepi8_epi16(vi20x89ABCDEF);
520 const __m128i vk20x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t)));
521 const __m128i vxk20x89ABCDEF = _mm_cvtepi8_epi16(vk20x89ABCDEF);
522 i20 += 16;
523
524
525 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
526 vprod89ABCDEF = _mm_mullo_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
527
528
529 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
530 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
531 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)));
532 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
533 const __m128i vi21x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i21 + 8));
534 const __m128i vxi21x89ABCDEF = _mm_cvtepi8_epi16(vi21x89ABCDEF);
535 const __m128i vk21x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t)));
536 const __m128i vxk21x89ABCDEF = _mm_cvtepi8_epi16(vk21x89ABCDEF);
537 i21 += 16;
538
539
540 vprod01234567 = _mm_macc_epi16(vxi21x01234567, vxk21x01234567, vprod01234567);
541 vprod89ABCDEF = _mm_macc_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF, vprod89ABCDEF);
542
543 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
544 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
545 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
546 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
547
548 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
549 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
550 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)));
551 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
552 const __m128i vi22x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i22 + 8));
553 const __m128i vxi22x89ABCDEF = _mm_cvtepi8_epi16(vi22x89ABCDEF);
554 const __m128i vk22x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t)));
555 const __m128i vxk22x89ABCDEF = _mm_cvtepi8_epi16(vk22x89ABCDEF);
556 i22 += 16;
557
558
559 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
560 vprod89ABCDEF = _mm_mullo_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
561
562
563 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
564 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
565 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)));
566 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
567 const __m128i vi23x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i23 + 8));
568 const __m128i vxi23x89ABCDEF = _mm_cvtepi8_epi16(vi23x89ABCDEF);
569 const __m128i vk23x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t)));
570 const __m128i vxk23x89ABCDEF = _mm_cvtepi8_epi16(vk23x89ABCDEF);
571 i23 += 16;
572
573
574 vprod01234567 = _mm_macc_epi16(vxi23x01234567, vxk23x01234567, vprod01234567);
575 vprod89ABCDEF = _mm_macc_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF, vprod89ABCDEF);
576
577 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
578 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
579 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
580 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
581
582 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
583 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
584 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)));
585 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
586 const __m128i vi24x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i24 + 8));
587 const __m128i vxi24x89ABCDEF = _mm_cvtepi8_epi16(vi24x89ABCDEF);
588 const __m128i vk24x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t)));
589 const __m128i vxk24x89ABCDEF = _mm_cvtepi8_epi16(vk24x89ABCDEF);
590 i24 += 16;
591
592
593 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
594 vprod89ABCDEF = _mm_mullo_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
595
596 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
597 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
598 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
599 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
600
601 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
602
603 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
604 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
605 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
606 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
607
608 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
609 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
610 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
611 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
612 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
613
614 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
615 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
616 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
617 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
618 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
619
620 vacc0123 = _mm_cvtps_epi32(vscaled0123);
621 vacc4567 = _mm_cvtps_epi32(vscaled4567);
622 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
623 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
624
625 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
626 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
627 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
628
629
630 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
631
632 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
633 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
634
635 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
636 output += 16;
637 }
638 if XNN_UNLIKELY(c != 0) {
639 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
640 do {
641 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
642 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
643
644
645 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
646 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
647 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
648 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
649 i0 += 8;
650
651
652 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
653
654
655 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
656 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
657 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
658 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
659 i1 += 8;
660
661
662 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
663
664 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
665 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
666
667 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
668 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
669 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
670 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
671 i2 += 8;
672
673
674 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
675
676
677 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
678 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
679 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
680 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
681 i3 += 8;
682
683
684 vprod01234567 = _mm_macc_epi16(vxi3x01234567, vxk3x01234567, vprod01234567);
685
686 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
687 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
688
689 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
690 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
691 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
692 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
693 i4 += 8;
694
695
696 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
697
698
699 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
700 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
701 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
702 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
703 i5 += 8;
704
705
706 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567);
707
708 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
709 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
710
711 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
712 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
713 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
714 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
715 i6 += 8;
716
717
718 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
719
720
721 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
722 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
723 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
724 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
725 i7 += 8;
726
727
728 vprod01234567 = _mm_macc_epi16(vxi7x01234567, vxk7x01234567, vprod01234567);
729
730 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
731 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
732
733 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
734 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
735 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
736 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
737 i8 += 8;
738
739
740 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
741
742
743 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
744 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
745 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
746 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
747 i9 += 8;
748
749
750 vprod01234567 = _mm_macc_epi16(vxi9x01234567, vxk9x01234567, vprod01234567);
751
752 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
753 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
754
755 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
756 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
757 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) (k + 160));
758 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
759 i10 += 8;
760
761
762 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
763
764
765 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
766 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
767 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) (k + 176));
768 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
769 i11 += 8;
770
771
772 vprod01234567 = _mm_macc_epi16(vxi11x01234567, vxk11x01234567, vprod01234567);
773
774 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
775 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
776
777 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
778 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
779 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
780 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
781 i12 += 8;
782
783
784 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
785
786
787 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
788 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
789 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) (k + 208));
790 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
791 i13 += 8;
792
793
794 vprod01234567 = _mm_macc_epi16(vxi13x01234567, vxk13x01234567, vprod01234567);
795
796 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
797 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
798
799 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
800 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
801 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) (k + 224));
802 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
803 i14 += 8;
804
805
806 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
807
808
809 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
810 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
811 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) (k + 240));
812 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
813 i15 += 8;
814
815
816 vprod01234567 = _mm_macc_epi16(vxi15x01234567, vxk15x01234567, vprod01234567);
817
818 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
819 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
820
821 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
822 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
823 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) (k + 256));
824 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
825 i16 += 8;
826
827
828 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
829
830
831 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
832 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
833 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) (k + 272));
834 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
835 i17 += 8;
836
837
838 vprod01234567 = _mm_macc_epi16(vxi17x01234567, vxk17x01234567, vprod01234567);
839
840 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
841 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
842
843 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
844 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
845 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) (k + 288));
846 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
847 i18 += 8;
848
849
850 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
851
852
853 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
854 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
855 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) (k + 304));
856 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
857 i19 += 8;
858
859
860 vprod01234567 = _mm_macc_epi16(vxi19x01234567, vxk19x01234567, vprod01234567);
861
862 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
863 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
864
865 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
866 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
867 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) (k + 320));
868 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
869 i20 += 8;
870
871
872 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
873
874
875 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
876 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
877 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) (k + 336));
878 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
879 i21 += 8;
880
881
882 vprod01234567 = _mm_macc_epi16(vxi21x01234567, vxk21x01234567, vprod01234567);
883
884 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
885 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
886
887 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
888 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
889 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) (k + 352));
890 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
891 i22 += 8;
892
893
894 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
895
896
897 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
898 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
899 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) (k + 368));
900 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
901 i23 += 8;
902
903
904 vprod01234567 = _mm_macc_epi16(vxi23x01234567, vxk23x01234567, vprod01234567);
905
906 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
907 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
908
909 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
910 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
911 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) (k + 384));
912 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
913 i24 += 8;
914
915
916 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
917
918 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
919 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
920
921 k += 8;
922
923 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
924 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
925
926 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
927 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
928 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
929
930 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
931 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
932 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
933
934 vacc0123 = _mm_cvtps_epi32(vscaled0123);
935 vacc4567 = _mm_cvtps_epi32(vscaled4567);
936
937 w = (const void*) ((const int32_t*) w + 8);
938
939 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
940 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
941
942
943 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
944
945 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
946
947 if XNN_LIKELY(c >= 8) {
948 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
949 output += 8;
950 c -= 8;
951 } else {
952 if (c & 4) {
953 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
954 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
955 output += 4;
956 }
957 if (c & 2) {
958 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
959 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
960 output += 2;
961 }
962 if (c & 1) {
963 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
964 output += 1;
965 }
966 c = 0;
967 }
968 } while (c != 0);
969 }
970
971 output = (int8_t*) ((uintptr_t) output + output_increment);
972 } while (--output_width != 0);
973 }
974