1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <smmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/unaligned.h>
16
17
xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul16_add16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul16_add16(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 do {
34 const int8_t* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
38 }
39 const int8_t* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
43 }
44 const int8_t* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
48 }
49 const int8_t* i3 = input[3];
50 assert(i3 != NULL);
51 if XNN_UNPREDICTABLE(i3 != zero) {
52 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
53 }
54 const int8_t* i4 = input[4];
55 assert(i4 != NULL);
56 if XNN_UNPREDICTABLE(i4 != zero) {
57 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
58 }
59 const int8_t* i5 = input[5];
60 assert(i5 != NULL);
61 if XNN_UNPREDICTABLE(i5 != zero) {
62 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
63 }
64 const int8_t* i6 = input[6];
65 assert(i6 != NULL);
66 if XNN_UNPREDICTABLE(i6 != zero) {
67 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
68 }
69 const int8_t* i7 = input[7];
70 assert(i7 != NULL);
71 if XNN_UNPREDICTABLE(i7 != zero) {
72 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
73 }
74 const int8_t* i8 = input[8];
75 assert(i8 != NULL);
76 if XNN_UNPREDICTABLE(i8 != zero) {
77 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
78 }
79 const int8_t* i9 = input[9];
80 assert(i9 != NULL);
81 if XNN_UNPREDICTABLE(i9 != zero) {
82 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
83 }
84 const int8_t* i10 = input[10];
85 assert(i10 != NULL);
86 if XNN_UNPREDICTABLE(i10 != zero) {
87 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
88 }
89 const int8_t* i11 = input[11];
90 assert(i11 != NULL);
91 if XNN_UNPREDICTABLE(i11 != zero) {
92 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
93 }
94 const int8_t* i12 = input[12];
95 assert(i12 != NULL);
96 if XNN_UNPREDICTABLE(i12 != zero) {
97 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
98 }
99 const int8_t* i13 = input[13];
100 assert(i13 != NULL);
101 if XNN_UNPREDICTABLE(i13 != zero) {
102 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
103 }
104 const int8_t* i14 = input[14];
105 assert(i14 != NULL);
106 if XNN_UNPREDICTABLE(i14 != zero) {
107 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
108 }
109 const int8_t* i15 = input[15];
110 assert(i15 != NULL);
111 if XNN_UNPREDICTABLE(i15 != zero) {
112 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
113 }
114 const int8_t* i16 = input[16];
115 assert(i16 != NULL);
116 if XNN_UNPREDICTABLE(i16 != zero) {
117 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
118 }
119 const int8_t* i17 = input[17];
120 assert(i17 != NULL);
121 if XNN_UNPREDICTABLE(i17 != zero) {
122 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
123 }
124 const int8_t* i18 = input[18];
125 assert(i18 != NULL);
126 if XNN_UNPREDICTABLE(i18 != zero) {
127 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
128 }
129 const int8_t* i19 = input[19];
130 assert(i19 != NULL);
131 if XNN_UNPREDICTABLE(i19 != zero) {
132 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
133 }
134 const int8_t* i20 = input[20];
135 assert(i20 != NULL);
136 if XNN_UNPREDICTABLE(i20 != zero) {
137 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
138 }
139 const int8_t* i21 = input[21];
140 assert(i21 != NULL);
141 if XNN_UNPREDICTABLE(i21 != zero) {
142 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
143 }
144 const int8_t* i22 = input[22];
145 assert(i22 != NULL);
146 if XNN_UNPREDICTABLE(i22 != zero) {
147 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
148 }
149 const int8_t* i23 = input[23];
150 assert(i23 != NULL);
151 if XNN_UNPREDICTABLE(i23 != zero) {
152 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
153 }
154 const int8_t* i24 = input[24];
155 assert(i24 != NULL);
156 if XNN_UNPREDICTABLE(i24 != zero) {
157 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
158 }
159 input = (const int8_t**) ((uintptr_t) input + input_stride);
160
161 size_t c = channels;
162 const void* w = weights;
163 for (; c >= 16; c -= 16) {
164 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
165 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
166 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
167 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
168
169
170 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
171 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
172 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
173 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
174 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
175 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
176 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
177 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
178 i0 += 16;
179
180
181 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
182 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
183
184
185 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
186 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
187 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
188 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
189 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
190 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
191 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
192 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
193 i1 += 16;
194
195
196 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
197 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF));
198
199 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
200 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
201 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
202 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
203
204 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
205 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
206 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
207 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
208 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
209 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
210 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
211 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
212 i2 += 16;
213
214
215 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
216 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
217
218
219 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
220 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
221 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
222 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
223 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
224 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
225 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
226 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
227 i3 += 16;
228
229
230 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
231 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF));
232
233 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
234 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
235 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
236 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
237
238 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
239 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
240 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
241 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
242 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
243 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
244 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
245 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
246 i4 += 16;
247
248
249 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
250 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
251
252
253 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
254 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
255 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
256 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
257 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
258 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
259 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
260 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
261 i5 += 16;
262
263
264 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
265 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF));
266
267 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
268 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
269 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
270 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
271
272 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
273 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
274 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
275 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
276 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
277 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
278 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
279 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
280 i6 += 16;
281
282
283 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
284 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
285
286
287 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
288 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
289 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
290 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
291 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
292 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
293 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
294 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
295 i7 += 16;
296
297
298 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
299 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF));
300
301 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
302 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
303 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
304 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
305
306 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
307 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
308 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
309 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
310 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
311 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
312 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
313 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
314 i8 += 16;
315
316
317 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
318 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
319
320
321 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
322 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
323 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
324 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
325 const __m128i vi9x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i9 + 8));
326 const __m128i vxi9x89ABCDEF = _mm_cvtepi8_epi16(vi9x89ABCDEF);
327 const __m128i vk9x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t)));
328 const __m128i vxk9x89ABCDEF = _mm_cvtepi8_epi16(vk9x89ABCDEF);
329 i9 += 16;
330
331
332 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
333 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF));
334
335 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
336 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
337 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
338 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
339
340 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
341 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
342 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)));
343 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
344 const __m128i vi10x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i10 + 8));
345 const __m128i vxi10x89ABCDEF = _mm_cvtepi8_epi16(vi10x89ABCDEF);
346 const __m128i vk10x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t)));
347 const __m128i vxk10x89ABCDEF = _mm_cvtepi8_epi16(vk10x89ABCDEF);
348 i10 += 16;
349
350
351 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
352 vprod89ABCDEF = _mm_mullo_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
353
354
355 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
356 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
357 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)));
358 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
359 const __m128i vi11x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i11 + 8));
360 const __m128i vxi11x89ABCDEF = _mm_cvtepi8_epi16(vi11x89ABCDEF);
361 const __m128i vk11x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t)));
362 const __m128i vxk11x89ABCDEF = _mm_cvtepi8_epi16(vk11x89ABCDEF);
363 i11 += 16;
364
365
366 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
367 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF));
368
369 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
370 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
371 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
372 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
373
374 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
375 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
376 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)));
377 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
378 const __m128i vi12x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i12 + 8));
379 const __m128i vxi12x89ABCDEF = _mm_cvtepi8_epi16(vi12x89ABCDEF);
380 const __m128i vk12x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t)));
381 const __m128i vxk12x89ABCDEF = _mm_cvtepi8_epi16(vk12x89ABCDEF);
382 i12 += 16;
383
384
385 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
386 vprod89ABCDEF = _mm_mullo_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
387
388
389 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
390 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
391 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)));
392 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
393 const __m128i vi13x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i13 + 8));
394 const __m128i vxi13x89ABCDEF = _mm_cvtepi8_epi16(vi13x89ABCDEF);
395 const __m128i vk13x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t)));
396 const __m128i vxk13x89ABCDEF = _mm_cvtepi8_epi16(vk13x89ABCDEF);
397 i13 += 16;
398
399
400 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
401 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF));
402
403 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
404 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
405 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
406 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
407
408 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
409 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
410 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)));
411 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
412 const __m128i vi14x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i14 + 8));
413 const __m128i vxi14x89ABCDEF = _mm_cvtepi8_epi16(vi14x89ABCDEF);
414 const __m128i vk14x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t)));
415 const __m128i vxk14x89ABCDEF = _mm_cvtepi8_epi16(vk14x89ABCDEF);
416 i14 += 16;
417
418
419 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
420 vprod89ABCDEF = _mm_mullo_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
421
422
423 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
424 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
425 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)));
426 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
427 const __m128i vi15x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i15 + 8));
428 const __m128i vxi15x89ABCDEF = _mm_cvtepi8_epi16(vi15x89ABCDEF);
429 const __m128i vk15x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t)));
430 const __m128i vxk15x89ABCDEF = _mm_cvtepi8_epi16(vk15x89ABCDEF);
431 i15 += 16;
432
433
434 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
435 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF));
436
437 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
438 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
439 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
440 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
441
442 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
443 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
444 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)));
445 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
446 const __m128i vi16x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i16 + 8));
447 const __m128i vxi16x89ABCDEF = _mm_cvtepi8_epi16(vi16x89ABCDEF);
448 const __m128i vk16x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t)));
449 const __m128i vxk16x89ABCDEF = _mm_cvtepi8_epi16(vk16x89ABCDEF);
450 i16 += 16;
451
452
453 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
454 vprod89ABCDEF = _mm_mullo_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
455
456
457 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
458 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
459 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)));
460 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
461 const __m128i vi17x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i17 + 8));
462 const __m128i vxi17x89ABCDEF = _mm_cvtepi8_epi16(vi17x89ABCDEF);
463 const __m128i vk17x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t)));
464 const __m128i vxk17x89ABCDEF = _mm_cvtepi8_epi16(vk17x89ABCDEF);
465 i17 += 16;
466
467
468 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
469 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF));
470
471 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
472 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
473 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
474 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
475
476 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
477 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
478 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)));
479 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
480 const __m128i vi18x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i18 + 8));
481 const __m128i vxi18x89ABCDEF = _mm_cvtepi8_epi16(vi18x89ABCDEF);
482 const __m128i vk18x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t)));
483 const __m128i vxk18x89ABCDEF = _mm_cvtepi8_epi16(vk18x89ABCDEF);
484 i18 += 16;
485
486
487 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
488 vprod89ABCDEF = _mm_mullo_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
489
490
491 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
492 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
493 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)));
494 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
495 const __m128i vi19x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i19 + 8));
496 const __m128i vxi19x89ABCDEF = _mm_cvtepi8_epi16(vi19x89ABCDEF);
497 const __m128i vk19x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t)));
498 const __m128i vxk19x89ABCDEF = _mm_cvtepi8_epi16(vk19x89ABCDEF);
499 i19 += 16;
500
501
502 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
503 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF));
504
505 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
506 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
507 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
508 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
509
510 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
511 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
512 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)));
513 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
514 const __m128i vi20x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i20 + 8));
515 const __m128i vxi20x89ABCDEF = _mm_cvtepi8_epi16(vi20x89ABCDEF);
516 const __m128i vk20x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t)));
517 const __m128i vxk20x89ABCDEF = _mm_cvtepi8_epi16(vk20x89ABCDEF);
518 i20 += 16;
519
520
521 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
522 vprod89ABCDEF = _mm_mullo_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
523
524
525 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
526 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
527 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)));
528 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
529 const __m128i vi21x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i21 + 8));
530 const __m128i vxi21x89ABCDEF = _mm_cvtepi8_epi16(vi21x89ABCDEF);
531 const __m128i vk21x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t)));
532 const __m128i vxk21x89ABCDEF = _mm_cvtepi8_epi16(vk21x89ABCDEF);
533 i21 += 16;
534
535
536 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
537 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF));
538
539 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
540 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
541 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
542 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
543
544 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
545 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
546 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)));
547 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
548 const __m128i vi22x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i22 + 8));
549 const __m128i vxi22x89ABCDEF = _mm_cvtepi8_epi16(vi22x89ABCDEF);
550 const __m128i vk22x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t)));
551 const __m128i vxk22x89ABCDEF = _mm_cvtepi8_epi16(vk22x89ABCDEF);
552 i22 += 16;
553
554
555 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
556 vprod89ABCDEF = _mm_mullo_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
557
558
559 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
560 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
561 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)));
562 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
563 const __m128i vi23x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i23 + 8));
564 const __m128i vxi23x89ABCDEF = _mm_cvtepi8_epi16(vi23x89ABCDEF);
565 const __m128i vk23x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t)));
566 const __m128i vxk23x89ABCDEF = _mm_cvtepi8_epi16(vk23x89ABCDEF);
567 i23 += 16;
568
569
570 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
571 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF));
572
573 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
574 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
575 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
576 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
577
578 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
579 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
580 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)));
581 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
582 const __m128i vi24x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i24 + 8));
583 const __m128i vxi24x89ABCDEF = _mm_cvtepi8_epi16(vi24x89ABCDEF);
584 const __m128i vk24x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t)));
585 const __m128i vxk24x89ABCDEF = _mm_cvtepi8_epi16(vk24x89ABCDEF);
586 i24 += 16;
587
588
589 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
590 vprod89ABCDEF = _mm_mullo_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
591
592 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
593 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
594 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
595 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
596
597 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
598
599 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
600 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
601 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
602 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
603
604 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
605 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
606 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
607 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
608 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
609
610 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
611 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
612 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
613 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
614 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
615
616 vacc0123 = _mm_cvtps_epi32(vscaled0123);
617 vacc4567 = _mm_cvtps_epi32(vscaled4567);
618 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
619 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
620
621 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
622 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
623 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
624
625
626 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
627
628 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
629 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
630
631 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
632 output += 16;
633 }
634 if XNN_UNLIKELY(c != 0) {
635 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
636 do {
637 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
638 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
639
640
641 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
642 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
643 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
644 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
645 i0 += 8;
646
647
648 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
649
650
651 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
652 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
653 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
654 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
655 i1 += 8;
656
657
658 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
659
660 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
661 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
662
663 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
664 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
665 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
666 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
667 i2 += 8;
668
669
670 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
671
672
673 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
674 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
675 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
676 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
677 i3 += 8;
678
679
680 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
681
682 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
683 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
684
685 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
686 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
687 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
688 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
689 i4 += 8;
690
691
692 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
693
694
695 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
696 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
697 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
698 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
699 i5 += 8;
700
701
702 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
703
704 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
705 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
706
707 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
708 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
709 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
710 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
711 i6 += 8;
712
713
714 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
715
716
717 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
718 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
719 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
720 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
721 i7 += 8;
722
723
724 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
725
726 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
727 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
728
729 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
730 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
731 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
732 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
733 i8 += 8;
734
735
736 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
737
738
739 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
740 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
741 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
742 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
743 i9 += 8;
744
745
746 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
747
748 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
749 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
750
751 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
752 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
753 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) (k + 160));
754 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
755 i10 += 8;
756
757
758 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
759
760
761 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
762 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
763 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) (k + 176));
764 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
765 i11 += 8;
766
767
768 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
769
770 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
771 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
772
773 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
774 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
775 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
776 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
777 i12 += 8;
778
779
780 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
781
782
783 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
784 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
785 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) (k + 208));
786 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
787 i13 += 8;
788
789
790 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
791
792 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
793 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
794
795 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
796 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
797 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) (k + 224));
798 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
799 i14 += 8;
800
801
802 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
803
804
805 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
806 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
807 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) (k + 240));
808 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
809 i15 += 8;
810
811
812 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
813
814 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
815 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
816
817 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
818 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
819 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) (k + 256));
820 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
821 i16 += 8;
822
823
824 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
825
826
827 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
828 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
829 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) (k + 272));
830 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
831 i17 += 8;
832
833
834 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
835
836 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
837 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
838
839 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
840 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
841 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) (k + 288));
842 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
843 i18 += 8;
844
845
846 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
847
848
849 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
850 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
851 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) (k + 304));
852 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
853 i19 += 8;
854
855
856 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
857
858 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
859 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
860
861 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
862 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
863 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) (k + 320));
864 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
865 i20 += 8;
866
867
868 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
869
870
871 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
872 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
873 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) (k + 336));
874 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
875 i21 += 8;
876
877
878 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
879
880 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
881 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
882
883 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
884 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
885 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) (k + 352));
886 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
887 i22 += 8;
888
889
890 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
891
892
893 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
894 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
895 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) (k + 368));
896 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
897 i23 += 8;
898
899
900 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
901
902 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
903 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
904
905 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
906 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
907 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) (k + 384));
908 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
909 i24 += 8;
910
911
912 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
913
914 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
915 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
916
917 k += 8;
918
919 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
920 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
921
922 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
923 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
924 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
925
926 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
927 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
928 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
929
930 vacc0123 = _mm_cvtps_epi32(vscaled0123);
931 vacc4567 = _mm_cvtps_epi32(vscaled4567);
932
933 w = (const void*) ((const int32_t*) w + 8);
934
935 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
936 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
937
938
939 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
940
941 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
942
943 if XNN_LIKELY(c >= 8) {
944 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
945 output += 8;
946 c -= 8;
947 } else {
948 if (c & 4) {
949 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
950 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
951 output += 4;
952 }
953 if (c & 2) {
954 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
955 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
956 output += 2;
957 }
958 if (c & 1) {
959 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
960 output += 1;
961 }
962 c = 0;
963 }
964 } while (c != 0);
965 }
966
967 output = (int8_t*) ((uintptr_t) output + output_increment);
968 } while (--output_width != 0);
969 }
970