1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-avx512skx-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_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32(size_t channels,size_t output_width,const uint8_t ** input,const void * weights,uint8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32(
19 size_t channels,
20 size_t output_width,
21 const uint8_t** input,
22 const void* weights,
23 uint8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const uint8_t* zero,
28 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_MSAN
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const __m512 vscale = _mm512_load_ps(params->fp32_avx512.scale);
34 const __m512 voutput_max_less_zero_point = _mm512_load_ps(params->fp32_avx512.output_max_less_zero_point);
35 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx512.output_zero_point);
36 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx512.output_min);
37
38 const __m512i vk_zero_point = _mm512_cvtepu16_epi32(_mm256_load_si256((const __m256i*) params->fp32_avx512.kernel_zero_point));
39 do {
40 const uint8_t* i0 = input[0];
41 assert(i0 != NULL);
42 if XNN_UNPREDICTABLE(i0 != zero) {
43 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
44 }
45 const uint8_t* i1 = input[1];
46 assert(i1 != NULL);
47 if XNN_UNPREDICTABLE(i1 != zero) {
48 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
49 }
50 const uint8_t* i2 = input[2];
51 assert(i2 != NULL);
52 if XNN_UNPREDICTABLE(i2 != zero) {
53 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
54 }
55 const uint8_t* i3 = input[3];
56 assert(i3 != NULL);
57 if XNN_UNPREDICTABLE(i3 != zero) {
58 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
59 }
60 const uint8_t* i4 = input[4];
61 assert(i4 != NULL);
62 if XNN_UNPREDICTABLE(i4 != zero) {
63 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
64 }
65 const uint8_t* i5 = input[5];
66 assert(i5 != NULL);
67 if XNN_UNPREDICTABLE(i5 != zero) {
68 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
69 }
70 const uint8_t* i6 = input[6];
71 assert(i6 != NULL);
72 if XNN_UNPREDICTABLE(i6 != zero) {
73 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
74 }
75 const uint8_t* i7 = input[7];
76 assert(i7 != NULL);
77 if XNN_UNPREDICTABLE(i7 != zero) {
78 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
79 }
80 const uint8_t* i8 = input[8];
81 assert(i8 != NULL);
82 if XNN_UNPREDICTABLE(i8 != zero) {
83 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
84 }
85 const uint8_t* i9 = input[9];
86 assert(i9 != NULL);
87 if XNN_UNPREDICTABLE(i9 != zero) {
88 i9 = (const uint8_t*) ((uintptr_t) i9 + input_offset);
89 }
90 const uint8_t* i10 = input[10];
91 assert(i10 != NULL);
92 if XNN_UNPREDICTABLE(i10 != zero) {
93 i10 = (const uint8_t*) ((uintptr_t) i10 + input_offset);
94 }
95 const uint8_t* i11 = input[11];
96 assert(i11 != NULL);
97 if XNN_UNPREDICTABLE(i11 != zero) {
98 i11 = (const uint8_t*) ((uintptr_t) i11 + input_offset);
99 }
100 const uint8_t* i12 = input[12];
101 assert(i12 != NULL);
102 if XNN_UNPREDICTABLE(i12 != zero) {
103 i12 = (const uint8_t*) ((uintptr_t) i12 + input_offset);
104 }
105 const uint8_t* i13 = input[13];
106 assert(i13 != NULL);
107 if XNN_UNPREDICTABLE(i13 != zero) {
108 i13 = (const uint8_t*) ((uintptr_t) i13 + input_offset);
109 }
110 const uint8_t* i14 = input[14];
111 assert(i14 != NULL);
112 if XNN_UNPREDICTABLE(i14 != zero) {
113 i14 = (const uint8_t*) ((uintptr_t) i14 + input_offset);
114 }
115 const uint8_t* i15 = input[15];
116 assert(i15 != NULL);
117 if XNN_UNPREDICTABLE(i15 != zero) {
118 i15 = (const uint8_t*) ((uintptr_t) i15 + input_offset);
119 }
120 const uint8_t* i16 = input[16];
121 assert(i16 != NULL);
122 if XNN_UNPREDICTABLE(i16 != zero) {
123 i16 = (const uint8_t*) ((uintptr_t) i16 + input_offset);
124 }
125 const uint8_t* i17 = input[17];
126 assert(i17 != NULL);
127 if XNN_UNPREDICTABLE(i17 != zero) {
128 i17 = (const uint8_t*) ((uintptr_t) i17 + input_offset);
129 }
130 const uint8_t* i18 = input[18];
131 assert(i18 != NULL);
132 if XNN_UNPREDICTABLE(i18 != zero) {
133 i18 = (const uint8_t*) ((uintptr_t) i18 + input_offset);
134 }
135 const uint8_t* i19 = input[19];
136 assert(i19 != NULL);
137 if XNN_UNPREDICTABLE(i19 != zero) {
138 i19 = (const uint8_t*) ((uintptr_t) i19 + input_offset);
139 }
140 const uint8_t* i20 = input[20];
141 assert(i20 != NULL);
142 if XNN_UNPREDICTABLE(i20 != zero) {
143 i20 = (const uint8_t*) ((uintptr_t) i20 + input_offset);
144 }
145 const uint8_t* i21 = input[21];
146 assert(i21 != NULL);
147 if XNN_UNPREDICTABLE(i21 != zero) {
148 i21 = (const uint8_t*) ((uintptr_t) i21 + input_offset);
149 }
150 const uint8_t* i22 = input[22];
151 assert(i22 != NULL);
152 if XNN_UNPREDICTABLE(i22 != zero) {
153 i22 = (const uint8_t*) ((uintptr_t) i22 + input_offset);
154 }
155 const uint8_t* i23 = input[23];
156 assert(i23 != NULL);
157 if XNN_UNPREDICTABLE(i23 != zero) {
158 i23 = (const uint8_t*) ((uintptr_t) i23 + input_offset);
159 }
160 const uint8_t* i24 = input[24];
161 assert(i24 != NULL);
162 if XNN_UNPREDICTABLE(i24 != zero) {
163 i24 = (const uint8_t*) ((uintptr_t) i24 + input_offset);
164 }
165 input = (const uint8_t**) ((uintptr_t) input + input_stride);
166
167 size_t c = channels;
168 const void* w = weights;
169 for (; c >= 16; c -= 16) {
170 __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w);
171
172
173 const __m512i vi0x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i0));
174 const __m512i vk0x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(uint8_t)))), vk_zero_point);
175 i0 += 16;
176
177 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
178
179 const __m512i vi1x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i1));
180 const __m512i vk1x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(uint8_t)))), vk_zero_point);
181 i1 += 16;
182
183 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
184
185 const __m512i vi2x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i2));
186 const __m512i vk2x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(uint8_t)))), vk_zero_point);
187 i2 += 16;
188
189 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
190
191 const __m512i vi3x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i3));
192 const __m512i vk3x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(uint8_t)))), vk_zero_point);
193 i3 += 16;
194
195 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
196
197 const __m512i vi4x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i4));
198 const __m512i vk4x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(uint8_t)))), vk_zero_point);
199 i4 += 16;
200
201 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
202
203 const __m512i vi5x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i5));
204 const __m512i vk5x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(uint8_t)))), vk_zero_point);
205 i5 += 16;
206
207 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
208
209 const __m512i vi6x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i6));
210 const __m512i vk6x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(uint8_t)))), vk_zero_point);
211 i6 += 16;
212
213 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
214
215 const __m512i vi7x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i7));
216 const __m512i vk7x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(uint8_t)))), vk_zero_point);
217 i7 += 16;
218
219 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
220
221 const __m512i vi8x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i8));
222 const __m512i vk8x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(uint8_t)))), vk_zero_point);
223 i8 += 16;
224
225 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF));
226
227 const __m512i vi9x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i9));
228 const __m512i vk9x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(uint8_t)))), vk_zero_point);
229 i9 += 16;
230
231 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF));
232
233 const __m512i vi10x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i10));
234 const __m512i vk10x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(uint8_t)))), vk_zero_point);
235 i10 += 16;
236
237 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF));
238
239 const __m512i vi11x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i11));
240 const __m512i vk11x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(uint8_t)))), vk_zero_point);
241 i11 += 16;
242
243 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF));
244
245 const __m512i vi12x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i12));
246 const __m512i vk12x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(uint8_t)))), vk_zero_point);
247 i12 += 16;
248
249 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF));
250
251 const __m512i vi13x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i13));
252 const __m512i vk13x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(uint8_t)))), vk_zero_point);
253 i13 += 16;
254
255 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF));
256
257 const __m512i vi14x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i14));
258 const __m512i vk14x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(uint8_t)))), vk_zero_point);
259 i14 += 16;
260
261 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF));
262
263 const __m512i vi15x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i15));
264 const __m512i vk15x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(uint8_t)))), vk_zero_point);
265 i15 += 16;
266
267 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF));
268
269 const __m512i vi16x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i16));
270 const __m512i vk16x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(uint8_t)))), vk_zero_point);
271 i16 += 16;
272
273 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF));
274
275 const __m512i vi17x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i17));
276 const __m512i vk17x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(uint8_t)))), vk_zero_point);
277 i17 += 16;
278
279 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF));
280
281 const __m512i vi18x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i18));
282 const __m512i vk18x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(uint8_t)))), vk_zero_point);
283 i18 += 16;
284
285 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF));
286
287 const __m512i vi19x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i19));
288 const __m512i vk19x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(uint8_t)))), vk_zero_point);
289 i19 += 16;
290
291 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF));
292
293 const __m512i vi20x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i20));
294 const __m512i vk20x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(uint8_t)))), vk_zero_point);
295 i20 += 16;
296
297 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF));
298
299 const __m512i vi21x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i21));
300 const __m512i vk21x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(uint8_t)))), vk_zero_point);
301 i21 += 16;
302
303 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF));
304
305 const __m512i vi22x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i22));
306 const __m512i vk22x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(uint8_t)))), vk_zero_point);
307 i22 += 16;
308
309 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF));
310
311 const __m512i vi23x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i23));
312 const __m512i vk23x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(uint8_t)))), vk_zero_point);
313 i23 += 16;
314
315 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF));
316
317 const __m512i vi24x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i24));
318 const __m512i vk24x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(uint8_t)))), vk_zero_point);
319 i24 += 16;
320
321 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF));
322
323 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(uint8_t));
324
325 __m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF);
326
327 vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale);
328
329 vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point);
330
331 vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF);
332
333 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), voutput_zero_point);
334
335 const __m128i vout012389AB = _mm256_castsi256_si128(vout012389AB4567CDEF);
336 const __m128i vout4567CDEF = _mm256_extracti128_si256(vout012389AB4567CDEF, 1);
337 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packus_epi16(vout012389AB, vout4567CDEF), _MM_SHUFFLE(3, 1, 2, 0));
338
339 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
340
341 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
342 output += 16;
343 }
344 if XNN_UNLIKELY(c != 0) {
345 // Prepare mask for valid 8-bit elements (depends on nc).
346 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << (c & 15)) - UINT32_C(1)));
347 {
348 __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w);
349
350
351 const __m512i vi0x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i0));
352 const __m512i vk0x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(uint8_t)))), vk_zero_point);
353
354 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
355
356 const __m512i vi1x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i1));
357 const __m512i vk1x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(uint8_t)))), vk_zero_point);
358
359 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
360
361 const __m512i vi2x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i2));
362 const __m512i vk2x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(uint8_t)))), vk_zero_point);
363
364 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
365
366 const __m512i vi3x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i3));
367 const __m512i vk3x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(uint8_t)))), vk_zero_point);
368
369 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
370
371 const __m512i vi4x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i4));
372 const __m512i vk4x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(uint8_t)))), vk_zero_point);
373
374 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
375
376 const __m512i vi5x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i5));
377 const __m512i vk5x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(uint8_t)))), vk_zero_point);
378
379 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
380
381 const __m512i vi6x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i6));
382 const __m512i vk6x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(uint8_t)))), vk_zero_point);
383
384 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
385
386 const __m512i vi7x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i7));
387 const __m512i vk7x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(uint8_t)))), vk_zero_point);
388
389 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
390
391 const __m512i vi8x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i8));
392 const __m512i vk8x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(uint8_t)))), vk_zero_point);
393
394 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF));
395
396 const __m512i vi9x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i9));
397 const __m512i vk9x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(uint8_t)))), vk_zero_point);
398
399 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF));
400
401 const __m512i vi10x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i10));
402 const __m512i vk10x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(uint8_t)))), vk_zero_point);
403
404 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF));
405
406 const __m512i vi11x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i11));
407 const __m512i vk11x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(uint8_t)))), vk_zero_point);
408
409 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF));
410
411 const __m512i vi12x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i12));
412 const __m512i vk12x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(uint8_t)))), vk_zero_point);
413
414 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF));
415
416 const __m512i vi13x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i13));
417 const __m512i vk13x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(uint8_t)))), vk_zero_point);
418
419 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF));
420
421 const __m512i vi14x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i14));
422 const __m512i vk14x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(uint8_t)))), vk_zero_point);
423
424 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF));
425
426 const __m512i vi15x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i15));
427 const __m512i vk15x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(uint8_t)))), vk_zero_point);
428
429 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF));
430
431 const __m512i vi16x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i16));
432 const __m512i vk16x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(uint8_t)))), vk_zero_point);
433
434 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF));
435
436 const __m512i vi17x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i17));
437 const __m512i vk17x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(uint8_t)))), vk_zero_point);
438
439 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF));
440
441 const __m512i vi18x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i18));
442 const __m512i vk18x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(uint8_t)))), vk_zero_point);
443
444 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF));
445
446 const __m512i vi19x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i19));
447 const __m512i vk19x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(uint8_t)))), vk_zero_point);
448
449 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF));
450
451 const __m512i vi20x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i20));
452 const __m512i vk20x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(uint8_t)))), vk_zero_point);
453
454 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF));
455
456 const __m512i vi21x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i21));
457 const __m512i vk21x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(uint8_t)))), vk_zero_point);
458
459 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF));
460
461 const __m512i vi22x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i22));
462 const __m512i vk22x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(uint8_t)))), vk_zero_point);
463
464 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF));
465
466 const __m512i vi23x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i23));
467 const __m512i vk23x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(uint8_t)))), vk_zero_point);
468
469 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF));
470
471 const __m512i vi24x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i24));
472 const __m512i vk24x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(uint8_t)))), vk_zero_point);
473
474 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF));
475
476
477 __m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF);
478 vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale);
479 vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point);
480 vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF);
481
482
483 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), voutput_zero_point);
484
485 const __m128i vout012389AB = _mm256_castsi256_si128(vout012389AB4567CDEF);
486 const __m128i vout4567CDEF = _mm256_extracti128_si256(vout012389AB4567CDEF, 1);
487 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packus_epi16(vout012389AB, vout4567CDEF), _MM_SHUFFLE(3, 1, 2, 0));
488 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
489
490 _mm_mask_storeu_epi8(output, vmask, vout0123456789ABCDEF);
491 output = (uint8_t*) ((uintptr_t) output + c);
492 }
493 }
494
495 output = (uint8_t*) ((uintptr_t) output + output_increment);
496 } while (--output_width != 0);
497 }
498