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 #include <xnnpack/unaligned.h>
21
22
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)])23 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__xop_mul16_add16(
24 size_t channels,
25 size_t output_width,
26 const int8_t** input,
27 const void* weights,
28 int8_t* output,
29 size_t input_stride,
30 size_t output_increment,
31 size_t input_offset,
32 const int8_t* zero,
33 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(channels != 0);
36 assert(output_width != 0);
37
38 do {
39 const int8_t* i0 = input[0];
40 assert(i0 != NULL);
41 if XNN_UNPREDICTABLE(i0 != zero) {
42 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
43 }
44 const int8_t* i1 = input[1];
45 assert(i1 != NULL);
46 if XNN_UNPREDICTABLE(i1 != zero) {
47 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
48 }
49 const int8_t* i2 = input[2];
50 assert(i2 != NULL);
51 if XNN_UNPREDICTABLE(i2 != zero) {
52 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
53 }
54 const int8_t* i3 = input[3];
55 assert(i3 != NULL);
56 if XNN_UNPREDICTABLE(i3 != zero) {
57 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
58 }
59 const int8_t* i4 = input[4];
60 assert(i4 != NULL);
61 if XNN_UNPREDICTABLE(i4 != zero) {
62 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
63 }
64 const int8_t* i5 = input[5];
65 assert(i5 != NULL);
66 if XNN_UNPREDICTABLE(i5 != zero) {
67 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
68 }
69 const int8_t* i6 = input[6];
70 assert(i6 != NULL);
71 if XNN_UNPREDICTABLE(i6 != zero) {
72 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
73 }
74 const int8_t* i7 = input[7];
75 assert(i7 != NULL);
76 if XNN_UNPREDICTABLE(i7 != zero) {
77 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
78 }
79 const int8_t* i8 = input[8];
80 assert(i8 != NULL);
81 if XNN_UNPREDICTABLE(i8 != zero) {
82 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
83 }
84 const int8_t* i9 = input[9];
85 assert(i9 != NULL);
86 if XNN_UNPREDICTABLE(i9 != zero) {
87 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
88 }
89 const int8_t* i10 = input[10];
90 assert(i10 != NULL);
91 if XNN_UNPREDICTABLE(i10 != zero) {
92 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
93 }
94 const int8_t* i11 = input[11];
95 assert(i11 != NULL);
96 if XNN_UNPREDICTABLE(i11 != zero) {
97 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
98 }
99 const int8_t* i12 = input[12];
100 assert(i12 != NULL);
101 if XNN_UNPREDICTABLE(i12 != zero) {
102 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
103 }
104 const int8_t* i13 = input[13];
105 assert(i13 != NULL);
106 if XNN_UNPREDICTABLE(i13 != zero) {
107 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
108 }
109 const int8_t* i14 = input[14];
110 assert(i14 != NULL);
111 if XNN_UNPREDICTABLE(i14 != zero) {
112 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
113 }
114 const int8_t* i15 = input[15];
115 assert(i15 != NULL);
116 if XNN_UNPREDICTABLE(i15 != zero) {
117 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
118 }
119 const int8_t* i16 = input[16];
120 assert(i16 != NULL);
121 if XNN_UNPREDICTABLE(i16 != zero) {
122 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
123 }
124 const int8_t* i17 = input[17];
125 assert(i17 != NULL);
126 if XNN_UNPREDICTABLE(i17 != zero) {
127 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
128 }
129 const int8_t* i18 = input[18];
130 assert(i18 != NULL);
131 if XNN_UNPREDICTABLE(i18 != zero) {
132 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
133 }
134 const int8_t* i19 = input[19];
135 assert(i19 != NULL);
136 if XNN_UNPREDICTABLE(i19 != zero) {
137 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
138 }
139 const int8_t* i20 = input[20];
140 assert(i20 != NULL);
141 if XNN_UNPREDICTABLE(i20 != zero) {
142 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
143 }
144 const int8_t* i21 = input[21];
145 assert(i21 != NULL);
146 if XNN_UNPREDICTABLE(i21 != zero) {
147 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
148 }
149 const int8_t* i22 = input[22];
150 assert(i22 != NULL);
151 if XNN_UNPREDICTABLE(i22 != zero) {
152 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
153 }
154 const int8_t* i23 = input[23];
155 assert(i23 != NULL);
156 if XNN_UNPREDICTABLE(i23 != zero) {
157 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
158 }
159 const int8_t* i24 = input[24];
160 assert(i24 != NULL);
161 if XNN_UNPREDICTABLE(i24 != zero) {
162 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
163 }
164 input = (const int8_t**) ((uintptr_t) input + input_stride);
165
166 size_t c = channels;
167 const void* w = weights;
168 for (; c >= 16; c -= 16) {
169 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
170 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
171 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
172 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
173
174
175 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
176 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
177 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
178 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
179 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
180 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
181 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
182 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
183 i0 += 16;
184
185
186 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
187 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
188
189
190 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
191 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
192 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
193 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
194 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
195 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
196 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
197 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
198 i1 += 16;
199
200
201 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
202 vprod89ABCDEF = _mm_macc_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF, vprod89ABCDEF);
203
204 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
205 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
206 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
207 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
208
209 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
210 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
211 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
212 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
213 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
214 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
215 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
216 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
217 i2 += 16;
218
219
220 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
221 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
222
223
224 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
225 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
226 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
227 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
228 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
229 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
230 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
231 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
232 i3 += 16;
233
234
235 vprod01234567 = _mm_macc_epi16(vxi3x01234567, vxk3x01234567, vprod01234567);
236 vprod89ABCDEF = _mm_macc_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF, vprod89ABCDEF);
237
238 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
239 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
240 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
241 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
242
243 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
244 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
245 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
246 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
247 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
248 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
249 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
250 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
251 i4 += 16;
252
253
254 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
255 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
256
257
258 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
259 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
260 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
261 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
262 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
263 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
264 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
265 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
266 i5 += 16;
267
268
269 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567);
270 vprod89ABCDEF = _mm_macc_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF, vprod89ABCDEF);
271
272 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
273 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
274 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
275 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
276
277 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
278 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
279 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
280 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
281 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
282 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
283 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
284 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
285 i6 += 16;
286
287
288 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
289 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
290
291
292 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
293 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
294 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
295 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
296 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
297 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
298 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
299 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
300 i7 += 16;
301
302
303 vprod01234567 = _mm_macc_epi16(vxi7x01234567, vxk7x01234567, vprod01234567);
304 vprod89ABCDEF = _mm_macc_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF, vprod89ABCDEF);
305
306 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
307 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
308 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
309 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
310
311 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
312 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
313 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
314 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
315 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
316 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
317 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
318 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
319 i8 += 16;
320
321
322 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
323 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
324
325
326 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
327 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
328 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
329 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
330 const __m128i vi9x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i9 + 8));
331 const __m128i vxi9x89ABCDEF = _mm_cvtepi8_epi16(vi9x89ABCDEF);
332 const __m128i vk9x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t)));
333 const __m128i vxk9x89ABCDEF = _mm_cvtepi8_epi16(vk9x89ABCDEF);
334 i9 += 16;
335
336
337 vprod01234567 = _mm_macc_epi16(vxi9x01234567, vxk9x01234567, vprod01234567);
338 vprod89ABCDEF = _mm_macc_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF, vprod89ABCDEF);
339
340 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
341 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
342 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
343 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
344
345 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
346 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
347 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)));
348 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
349 const __m128i vi10x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i10 + 8));
350 const __m128i vxi10x89ABCDEF = _mm_cvtepi8_epi16(vi10x89ABCDEF);
351 const __m128i vk10x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t)));
352 const __m128i vxk10x89ABCDEF = _mm_cvtepi8_epi16(vk10x89ABCDEF);
353 i10 += 16;
354
355
356 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
357 vprod89ABCDEF = _mm_mullo_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
358
359
360 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
361 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
362 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)));
363 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
364 const __m128i vi11x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i11 + 8));
365 const __m128i vxi11x89ABCDEF = _mm_cvtepi8_epi16(vi11x89ABCDEF);
366 const __m128i vk11x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t)));
367 const __m128i vxk11x89ABCDEF = _mm_cvtepi8_epi16(vk11x89ABCDEF);
368 i11 += 16;
369
370
371 vprod01234567 = _mm_macc_epi16(vxi11x01234567, vxk11x01234567, vprod01234567);
372 vprod89ABCDEF = _mm_macc_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF, vprod89ABCDEF);
373
374 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
375 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
376 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
377 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
378
379 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
380 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
381 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)));
382 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
383 const __m128i vi12x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i12 + 8));
384 const __m128i vxi12x89ABCDEF = _mm_cvtepi8_epi16(vi12x89ABCDEF);
385 const __m128i vk12x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t)));
386 const __m128i vxk12x89ABCDEF = _mm_cvtepi8_epi16(vk12x89ABCDEF);
387 i12 += 16;
388
389
390 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
391 vprod89ABCDEF = _mm_mullo_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
392
393
394 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
395 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
396 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)));
397 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
398 const __m128i vi13x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i13 + 8));
399 const __m128i vxi13x89ABCDEF = _mm_cvtepi8_epi16(vi13x89ABCDEF);
400 const __m128i vk13x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t)));
401 const __m128i vxk13x89ABCDEF = _mm_cvtepi8_epi16(vk13x89ABCDEF);
402 i13 += 16;
403
404
405 vprod01234567 = _mm_macc_epi16(vxi13x01234567, vxk13x01234567, vprod01234567);
406 vprod89ABCDEF = _mm_macc_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF, vprod89ABCDEF);
407
408 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
409 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
410 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
411 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
412
413 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
414 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
415 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)));
416 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
417 const __m128i vi14x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i14 + 8));
418 const __m128i vxi14x89ABCDEF = _mm_cvtepi8_epi16(vi14x89ABCDEF);
419 const __m128i vk14x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t)));
420 const __m128i vxk14x89ABCDEF = _mm_cvtepi8_epi16(vk14x89ABCDEF);
421 i14 += 16;
422
423
424 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
425 vprod89ABCDEF = _mm_mullo_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
426
427
428 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
429 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
430 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)));
431 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
432 const __m128i vi15x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i15 + 8));
433 const __m128i vxi15x89ABCDEF = _mm_cvtepi8_epi16(vi15x89ABCDEF);
434 const __m128i vk15x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t)));
435 const __m128i vxk15x89ABCDEF = _mm_cvtepi8_epi16(vk15x89ABCDEF);
436 i15 += 16;
437
438
439 vprod01234567 = _mm_macc_epi16(vxi15x01234567, vxk15x01234567, vprod01234567);
440 vprod89ABCDEF = _mm_macc_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF, vprod89ABCDEF);
441
442 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
443 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
444 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
445 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
446
447 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
448 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
449 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)));
450 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
451 const __m128i vi16x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i16 + 8));
452 const __m128i vxi16x89ABCDEF = _mm_cvtepi8_epi16(vi16x89ABCDEF);
453 const __m128i vk16x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t)));
454 const __m128i vxk16x89ABCDEF = _mm_cvtepi8_epi16(vk16x89ABCDEF);
455 i16 += 16;
456
457
458 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
459 vprod89ABCDEF = _mm_mullo_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
460
461
462 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
463 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
464 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)));
465 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
466 const __m128i vi17x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i17 + 8));
467 const __m128i vxi17x89ABCDEF = _mm_cvtepi8_epi16(vi17x89ABCDEF);
468 const __m128i vk17x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t)));
469 const __m128i vxk17x89ABCDEF = _mm_cvtepi8_epi16(vk17x89ABCDEF);
470 i17 += 16;
471
472
473 vprod01234567 = _mm_macc_epi16(vxi17x01234567, vxk17x01234567, vprod01234567);
474 vprod89ABCDEF = _mm_macc_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF, vprod89ABCDEF);
475
476 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
477 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
478 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
479 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
480
481 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
482 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
483 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)));
484 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
485 const __m128i vi18x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i18 + 8));
486 const __m128i vxi18x89ABCDEF = _mm_cvtepi8_epi16(vi18x89ABCDEF);
487 const __m128i vk18x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t)));
488 const __m128i vxk18x89ABCDEF = _mm_cvtepi8_epi16(vk18x89ABCDEF);
489 i18 += 16;
490
491
492 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
493 vprod89ABCDEF = _mm_mullo_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
494
495
496 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
497 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
498 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)));
499 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
500 const __m128i vi19x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i19 + 8));
501 const __m128i vxi19x89ABCDEF = _mm_cvtepi8_epi16(vi19x89ABCDEF);
502 const __m128i vk19x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t)));
503 const __m128i vxk19x89ABCDEF = _mm_cvtepi8_epi16(vk19x89ABCDEF);
504 i19 += 16;
505
506
507 vprod01234567 = _mm_macc_epi16(vxi19x01234567, vxk19x01234567, vprod01234567);
508 vprod89ABCDEF = _mm_macc_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF, vprod89ABCDEF);
509
510 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
511 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
512 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
513 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
514
515 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
516 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
517 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)));
518 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
519 const __m128i vi20x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i20 + 8));
520 const __m128i vxi20x89ABCDEF = _mm_cvtepi8_epi16(vi20x89ABCDEF);
521 const __m128i vk20x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t)));
522 const __m128i vxk20x89ABCDEF = _mm_cvtepi8_epi16(vk20x89ABCDEF);
523 i20 += 16;
524
525
526 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
527 vprod89ABCDEF = _mm_mullo_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
528
529
530 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
531 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
532 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)));
533 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
534 const __m128i vi21x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i21 + 8));
535 const __m128i vxi21x89ABCDEF = _mm_cvtepi8_epi16(vi21x89ABCDEF);
536 const __m128i vk21x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t)));
537 const __m128i vxk21x89ABCDEF = _mm_cvtepi8_epi16(vk21x89ABCDEF);
538 i21 += 16;
539
540
541 vprod01234567 = _mm_macc_epi16(vxi21x01234567, vxk21x01234567, vprod01234567);
542 vprod89ABCDEF = _mm_macc_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF, vprod89ABCDEF);
543
544 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
545 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
546 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
547 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
548
549 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
550 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
551 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)));
552 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
553 const __m128i vi22x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i22 + 8));
554 const __m128i vxi22x89ABCDEF = _mm_cvtepi8_epi16(vi22x89ABCDEF);
555 const __m128i vk22x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t)));
556 const __m128i vxk22x89ABCDEF = _mm_cvtepi8_epi16(vk22x89ABCDEF);
557 i22 += 16;
558
559
560 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
561 vprod89ABCDEF = _mm_mullo_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
562
563
564 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
565 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
566 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)));
567 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
568 const __m128i vi23x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i23 + 8));
569 const __m128i vxi23x89ABCDEF = _mm_cvtepi8_epi16(vi23x89ABCDEF);
570 const __m128i vk23x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t)));
571 const __m128i vxk23x89ABCDEF = _mm_cvtepi8_epi16(vk23x89ABCDEF);
572 i23 += 16;
573
574
575 vprod01234567 = _mm_macc_epi16(vxi23x01234567, vxk23x01234567, vprod01234567);
576 vprod89ABCDEF = _mm_macc_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF, vprod89ABCDEF);
577
578 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
579 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
580 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
581 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
582
583 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
584 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
585 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)));
586 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
587 const __m128i vi24x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i24 + 8));
588 const __m128i vxi24x89ABCDEF = _mm_cvtepi8_epi16(vi24x89ABCDEF);
589 const __m128i vk24x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t)));
590 const __m128i vxk24x89ABCDEF = _mm_cvtepi8_epi16(vk24x89ABCDEF);
591 i24 += 16;
592
593
594 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
595 vprod89ABCDEF = _mm_mullo_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
596
597 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
598 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
599 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
600 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
601
602 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
603
604 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
605 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
606 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
607 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
608
609 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
610 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
611 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
612 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
613 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
614
615 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
616 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
617 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
618 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
619 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
620
621 vacc0123 = _mm_cvtps_epi32(vscaled0123);
622 vacc4567 = _mm_cvtps_epi32(vscaled4567);
623 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
624 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
625
626 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
627 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
628 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
629
630
631 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
632
633 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
634 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
635
636 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
637 output += 16;
638 }
639 if XNN_UNLIKELY(c != 0) {
640 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
641 do {
642 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
643 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
644
645
646 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
647 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
648 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
649 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
650 i0 += 8;
651
652
653 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
654
655
656 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
657 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
658 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
659 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
660 i1 += 8;
661
662
663 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
664
665 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
666 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
667
668 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
669 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
670 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
671 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
672 i2 += 8;
673
674
675 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
676
677
678 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
679 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
680 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
681 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
682 i3 += 8;
683
684
685 vprod01234567 = _mm_macc_epi16(vxi3x01234567, vxk3x01234567, vprod01234567);
686
687 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
688 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
689
690 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
691 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
692 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
693 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
694 i4 += 8;
695
696
697 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
698
699
700 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
701 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
702 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
703 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
704 i5 += 8;
705
706
707 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567);
708
709 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
710 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
711
712 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
713 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
714 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
715 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
716 i6 += 8;
717
718
719 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
720
721
722 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
723 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
724 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
725 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
726 i7 += 8;
727
728
729 vprod01234567 = _mm_macc_epi16(vxi7x01234567, vxk7x01234567, vprod01234567);
730
731 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
732 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
733
734 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
735 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
736 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
737 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
738 i8 += 8;
739
740
741 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
742
743
744 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
745 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
746 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
747 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
748 i9 += 8;
749
750
751 vprod01234567 = _mm_macc_epi16(vxi9x01234567, vxk9x01234567, vprod01234567);
752
753 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
754 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
755
756 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
757 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
758 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) (k + 160));
759 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
760 i10 += 8;
761
762
763 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
764
765
766 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
767 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
768 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) (k + 176));
769 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
770 i11 += 8;
771
772
773 vprod01234567 = _mm_macc_epi16(vxi11x01234567, vxk11x01234567, vprod01234567);
774
775 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
776 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
777
778 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
779 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
780 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
781 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
782 i12 += 8;
783
784
785 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
786
787
788 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
789 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
790 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) (k + 208));
791 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
792 i13 += 8;
793
794
795 vprod01234567 = _mm_macc_epi16(vxi13x01234567, vxk13x01234567, vprod01234567);
796
797 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
798 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
799
800 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
801 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
802 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) (k + 224));
803 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
804 i14 += 8;
805
806
807 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
808
809
810 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
811 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
812 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) (k + 240));
813 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
814 i15 += 8;
815
816
817 vprod01234567 = _mm_macc_epi16(vxi15x01234567, vxk15x01234567, vprod01234567);
818
819 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
820 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
821
822 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
823 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
824 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) (k + 256));
825 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
826 i16 += 8;
827
828
829 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
830
831
832 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
833 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
834 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) (k + 272));
835 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
836 i17 += 8;
837
838
839 vprod01234567 = _mm_macc_epi16(vxi17x01234567, vxk17x01234567, vprod01234567);
840
841 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
842 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
843
844 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
845 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
846 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) (k + 288));
847 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
848 i18 += 8;
849
850
851 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
852
853
854 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
855 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
856 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) (k + 304));
857 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
858 i19 += 8;
859
860
861 vprod01234567 = _mm_macc_epi16(vxi19x01234567, vxk19x01234567, vprod01234567);
862
863 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
864 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
865
866 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
867 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
868 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) (k + 320));
869 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
870 i20 += 8;
871
872
873 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
874
875
876 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
877 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
878 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) (k + 336));
879 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
880 i21 += 8;
881
882
883 vprod01234567 = _mm_macc_epi16(vxi21x01234567, vxk21x01234567, vprod01234567);
884
885 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
886 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
887
888 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
889 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
890 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) (k + 352));
891 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
892 i22 += 8;
893
894
895 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
896
897
898 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
899 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
900 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) (k + 368));
901 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
902 i23 += 8;
903
904
905 vprod01234567 = _mm_macc_epi16(vxi23x01234567, vxk23x01234567, vprod01234567);
906
907 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
908 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
909
910 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
911 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
912 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) (k + 384));
913 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
914 i24 += 8;
915
916
917 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
918
919 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
920 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
921
922 k += 8;
923
924 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
925 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
926
927 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
928 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
929 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
930
931 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
932 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
933 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
934
935 vacc0123 = _mm_cvtps_epi32(vscaled0123);
936 vacc4567 = _mm_cvtps_epi32(vscaled4567);
937
938 w = (const void*) ((const int32_t*) w + 8);
939
940 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
941 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
942
943
944 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
945
946 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
947
948 if XNN_LIKELY(c >= 8) {
949 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
950 output += 8;
951 c -= 8;
952 } else {
953 if (c & 4) {
954 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
955 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
956 output += 4;
957 }
958 if (c & 2) {
959 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
960 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
961 output += 2;
962 }
963 if (c & 1) {
964 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
965 output += 1;
966 }
967 c = 0;
968 }
969 } while (c != 0);
970 }
971
972 output = (int8_t*) ((uintptr_t) output + output_increment);
973 } while (--output_width != 0);
974 }
975