1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul32.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 <immintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16
17
xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32(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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32(
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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 do {
34 const int8_t* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
38 }
39 const int8_t* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
43 }
44 const int8_t* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
48 }
49 const int8_t* i3 = input[3];
50 assert(i3 != NULL);
51 if XNN_UNPREDICTABLE(i3 != zero) {
52 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
53 }
54 const int8_t* i4 = input[4];
55 assert(i4 != NULL);
56 if XNN_UNPREDICTABLE(i4 != zero) {
57 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
58 }
59 const int8_t* i5 = input[5];
60 assert(i5 != NULL);
61 if XNN_UNPREDICTABLE(i5 != zero) {
62 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
63 }
64 const int8_t* i6 = input[6];
65 assert(i6 != NULL);
66 if XNN_UNPREDICTABLE(i6 != zero) {
67 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
68 }
69 const int8_t* i7 = input[7];
70 assert(i7 != NULL);
71 if XNN_UNPREDICTABLE(i7 != zero) {
72 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
73 }
74 const int8_t* i8 = input[8];
75 assert(i8 != NULL);
76 if XNN_UNPREDICTABLE(i8 != zero) {
77 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
78 }
79 const int8_t* i9 = input[9];
80 assert(i9 != NULL);
81 if XNN_UNPREDICTABLE(i9 != zero) {
82 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
83 }
84 const int8_t* i10 = input[10];
85 assert(i10 != NULL);
86 if XNN_UNPREDICTABLE(i10 != zero) {
87 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
88 }
89 const int8_t* i11 = input[11];
90 assert(i11 != NULL);
91 if XNN_UNPREDICTABLE(i11 != zero) {
92 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
93 }
94 const int8_t* i12 = input[12];
95 assert(i12 != NULL);
96 if XNN_UNPREDICTABLE(i12 != zero) {
97 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
98 }
99 const int8_t* i13 = input[13];
100 assert(i13 != NULL);
101 if XNN_UNPREDICTABLE(i13 != zero) {
102 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
103 }
104 const int8_t* i14 = input[14];
105 assert(i14 != NULL);
106 if XNN_UNPREDICTABLE(i14 != zero) {
107 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
108 }
109 const int8_t* i15 = input[15];
110 assert(i15 != NULL);
111 if XNN_UNPREDICTABLE(i15 != zero) {
112 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
113 }
114 const int8_t* i16 = input[16];
115 assert(i16 != NULL);
116 if XNN_UNPREDICTABLE(i16 != zero) {
117 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
118 }
119 const int8_t* i17 = input[17];
120 assert(i17 != NULL);
121 if XNN_UNPREDICTABLE(i17 != zero) {
122 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
123 }
124 const int8_t* i18 = input[18];
125 assert(i18 != NULL);
126 if XNN_UNPREDICTABLE(i18 != zero) {
127 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
128 }
129 const int8_t* i19 = input[19];
130 assert(i19 != NULL);
131 if XNN_UNPREDICTABLE(i19 != zero) {
132 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
133 }
134 const int8_t* i20 = input[20];
135 assert(i20 != NULL);
136 if XNN_UNPREDICTABLE(i20 != zero) {
137 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
138 }
139 const int8_t* i21 = input[21];
140 assert(i21 != NULL);
141 if XNN_UNPREDICTABLE(i21 != zero) {
142 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
143 }
144 const int8_t* i22 = input[22];
145 assert(i22 != NULL);
146 if XNN_UNPREDICTABLE(i22 != zero) {
147 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
148 }
149 const int8_t* i23 = input[23];
150 assert(i23 != NULL);
151 if XNN_UNPREDICTABLE(i23 != zero) {
152 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
153 }
154 const int8_t* i24 = input[24];
155 assert(i24 != NULL);
156 if XNN_UNPREDICTABLE(i24 != zero) {
157 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
158 }
159 input = (const int8_t**) ((uintptr_t) input + input_stride);
160
161 size_t c = channels;
162 const void* w = weights;
163 for (; c >= 8; c -= 8) {
164 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
165 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
166
167
168 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
169 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))));
170 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 4));
171 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 4 * sizeof(int8_t))));
172 i0 += 8;
173
174 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
175 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
176
177 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
178 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))));
179 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 4));
180 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 12 * sizeof(int8_t))));
181 i1 += 8;
182
183 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
184 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
185
186 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
187 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))));
188 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 4));
189 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 20 * sizeof(int8_t))));
190 i2 += 8;
191
192 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
193 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
194
195 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
196 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))));
197 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 4));
198 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 28 * sizeof(int8_t))));
199 i3 += 8;
200
201 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
202 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
203
204 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
205 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))));
206 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 4));
207 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 36 * sizeof(int8_t))));
208 i4 += 8;
209
210 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
211 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
212
213 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
214 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))));
215 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 4));
216 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 44 * sizeof(int8_t))));
217 i5 += 8;
218
219 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
220 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
221
222 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
223 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t))));
224 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 4));
225 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 52 * sizeof(int8_t))));
226 i6 += 8;
227
228 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
229 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567));
230
231 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
232 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t))));
233 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 4));
234 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 60 * sizeof(int8_t))));
235 i7 += 8;
236
237 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
238 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi7x4567, vk7x4567));
239
240 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
241 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t))));
242 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 4));
243 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 68 * sizeof(int8_t))));
244 i8 += 8;
245
246 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
247 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi8x4567, vk8x4567));
248
249 const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i9));
250 const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t))));
251 const __m128i vi9x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i9 + 4));
252 const __m128i vk9x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 76 * sizeof(int8_t))));
253 i9 += 8;
254
255 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi9x0123, vk9x0123));
256 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi9x4567, vk9x4567));
257
258 const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i10));
259 const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t))));
260 const __m128i vi10x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i10 + 4));
261 const __m128i vk10x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 84 * sizeof(int8_t))));
262 i10 += 8;
263
264 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi10x0123, vk10x0123));
265 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi10x4567, vk10x4567));
266
267 const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i11));
268 const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t))));
269 const __m128i vi11x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i11 + 4));
270 const __m128i vk11x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 92 * sizeof(int8_t))));
271 i11 += 8;
272
273 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi11x0123, vk11x0123));
274 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi11x4567, vk11x4567));
275
276 const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i12));
277 const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t))));
278 const __m128i vi12x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i12 + 4));
279 const __m128i vk12x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 100 * sizeof(int8_t))));
280 i12 += 8;
281
282 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi12x0123, vk12x0123));
283 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi12x4567, vk12x4567));
284
285 const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i13));
286 const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t))));
287 const __m128i vi13x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i13 + 4));
288 const __m128i vk13x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 108 * sizeof(int8_t))));
289 i13 += 8;
290
291 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi13x0123, vk13x0123));
292 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi13x4567, vk13x4567));
293
294 const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i14));
295 const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t))));
296 const __m128i vi14x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i14 + 4));
297 const __m128i vk14x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 116 * sizeof(int8_t))));
298 i14 += 8;
299
300 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi14x0123, vk14x0123));
301 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi14x4567, vk14x4567));
302
303 const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i15));
304 const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t))));
305 const __m128i vi15x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i15 + 4));
306 const __m128i vk15x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 124 * sizeof(int8_t))));
307 i15 += 8;
308
309 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi15x0123, vk15x0123));
310 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi15x4567, vk15x4567));
311
312 const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i16));
313 const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t))));
314 const __m128i vi16x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i16 + 4));
315 const __m128i vk16x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 132 * sizeof(int8_t))));
316 i16 += 8;
317
318 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi16x0123, vk16x0123));
319 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi16x4567, vk16x4567));
320
321 const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i17));
322 const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t))));
323 const __m128i vi17x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i17 + 4));
324 const __m128i vk17x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 140 * sizeof(int8_t))));
325 i17 += 8;
326
327 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi17x0123, vk17x0123));
328 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi17x4567, vk17x4567));
329
330 const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i18));
331 const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t))));
332 const __m128i vi18x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i18 + 4));
333 const __m128i vk18x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 148 * sizeof(int8_t))));
334 i18 += 8;
335
336 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi18x0123, vk18x0123));
337 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi18x4567, vk18x4567));
338
339 const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i19));
340 const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t))));
341 const __m128i vi19x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i19 + 4));
342 const __m128i vk19x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 156 * sizeof(int8_t))));
343 i19 += 8;
344
345 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi19x0123, vk19x0123));
346 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi19x4567, vk19x4567));
347
348 const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i20));
349 const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t))));
350 const __m128i vi20x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i20 + 4));
351 const __m128i vk20x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 164 * sizeof(int8_t))));
352 i20 += 8;
353
354 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi20x0123, vk20x0123));
355 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi20x4567, vk20x4567));
356
357 const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i21));
358 const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t))));
359 const __m128i vi21x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i21 + 4));
360 const __m128i vk21x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 172 * sizeof(int8_t))));
361 i21 += 8;
362
363 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi21x0123, vk21x0123));
364 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi21x4567, vk21x4567));
365
366 const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i22));
367 const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t))));
368 const __m128i vi22x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i22 + 4));
369 const __m128i vk22x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 180 * sizeof(int8_t))));
370 i22 += 8;
371
372 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi22x0123, vk22x0123));
373 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi22x4567, vk22x4567));
374
375 const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i23));
376 const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t))));
377 const __m128i vi23x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i23 + 4));
378 const __m128i vk23x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 188 * sizeof(int8_t))));
379 i23 += 8;
380
381 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi23x0123, vk23x0123));
382 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi23x4567, vk23x4567));
383
384 const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i24));
385 const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t))));
386 const __m128i vi24x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i24 + 4));
387 const __m128i vk24x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 196 * sizeof(int8_t))));
388 i24 += 8;
389
390 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi24x0123, vk24x0123));
391 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi24x4567, vk24x4567));
392
393 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t));
394
395 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
396 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
397
398 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
399 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
400 w = (const void*) ((const float*) w + 8);
401 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
402 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
403
404 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
405 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
406 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
407
408 vacc0123 = _mm_cvtps_epi32(vscaled0123);
409 vacc4567 = _mm_cvtps_epi32(vscaled4567);
410
411 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
412 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
413
414 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4.output_min);
415 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
416 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
417
418 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
419 output += 8;
420 }
421 if XNN_UNLIKELY(c != 0) {
422 const int8_t* k = (const int8_t*) ((const int32_t*) w + 8);
423 do {
424 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
425
426 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
427 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(k));
428 i0 += 4;
429
430 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
431 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
432 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 8)));
433 i1 += 4;
434
435 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
436 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
437 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 16)));
438 i2 += 4;
439
440 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
441 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
442 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 24)));
443 i3 += 4;
444
445 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
446 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
447 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 32)));
448 i4 += 4;
449
450 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
451 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
452 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 40)));
453 i5 += 4;
454
455 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
456 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
457 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 48)));
458 i6 += 4;
459
460 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
461 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
462 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 56)));
463 i7 += 4;
464
465 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
466 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
467 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 64)));
468 i8 += 4;
469
470 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
471 const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i9));
472 const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 72)));
473 i9 += 4;
474
475 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi9x0123, vk9x0123));
476 const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i10));
477 const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 80)));
478 i10 += 4;
479
480 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi10x0123, vk10x0123));
481 const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i11));
482 const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 88)));
483 i11 += 4;
484
485 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi11x0123, vk11x0123));
486 const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i12));
487 const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 96)));
488 i12 += 4;
489
490 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi12x0123, vk12x0123));
491 const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i13));
492 const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 104)));
493 i13 += 4;
494
495 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi13x0123, vk13x0123));
496 const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i14));
497 const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 112)));
498 i14 += 4;
499
500 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi14x0123, vk14x0123));
501 const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i15));
502 const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 120)));
503 i15 += 4;
504
505 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi15x0123, vk15x0123));
506 const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i16));
507 const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 128)));
508 i16 += 4;
509
510 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi16x0123, vk16x0123));
511 const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i17));
512 const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 136)));
513 i17 += 4;
514
515 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi17x0123, vk17x0123));
516 const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i18));
517 const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 144)));
518 i18 += 4;
519
520 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi18x0123, vk18x0123));
521 const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i19));
522 const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 152)));
523 i19 += 4;
524
525 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi19x0123, vk19x0123));
526 const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i20));
527 const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 160)));
528 i20 += 4;
529
530 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi20x0123, vk20x0123));
531 const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i21));
532 const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 168)));
533 i21 += 4;
534
535 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi21x0123, vk21x0123));
536 const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i22));
537 const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 176)));
538 i22 += 4;
539
540 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi22x0123, vk22x0123));
541 const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i23));
542 const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 184)));
543 i23 += 4;
544
545 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi23x0123, vk23x0123));
546 const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i24));
547 const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 192)));
548 i24 += 4;
549
550 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi24x0123, vk24x0123));
551
552 k += 4;
553
554 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
555 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t)));
556 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
557 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->sse4.output_max_less_zero_point));
558 vacc0123 = _mm_cvtps_epi32(vscaled0123);
559
560 w = (const void*) ((const int32_t*) w + 4);
561
562 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
563 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
564
565 vout0123 = _mm_packs_epi16(vout0123, vout0123);
566 vout0123 = _mm_max_epi8(vout0123, _mm_load_si128((const __m128i*) params->sse4.output_min));
567
568 if XNN_LIKELY(c >= 4) {
569 _mm_storeu_si32(output, vout0123);
570 output += 4;
571 c -= 4;
572 } else {
573 if (c & 2) {
574 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123, 0);
575 vout0123 = _mm_srli_epi32(vout0123, 16);
576 output += 2;
577 }
578 if (c & 1) {
579 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
580 output += 1;
581 }
582 c = 0;
583 }
584 } while (c != 0);
585 }
586
587 output = (int8_t*) ((uintptr_t) output + output_increment);
588 } while (--output_width != 0);
589 }
590