Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-14x16c4-minmax-avx512vnni.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx16c4-avx512vnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/unaligned.h"
19
20
21 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_14x16c4__avx512vnni(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const int8_t* zero_data,
34 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
35 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 14);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 4 * sizeof(int8_t));
47 float* c0 = c;
48 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 c1 = c0;
51 }
52 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 c2 = c1;
55 }
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 c3 = c2;
59 }
60 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
61 if XNN_UNPREDICTABLE(mr <= 4) {
62 c4 = c3;
63 }
64 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
65 if XNN_UNPREDICTABLE(mr < 6) {
66 c5 = c4;
67 }
68 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
69 if XNN_UNPREDICTABLE(mr <= 6) {
70 c6 = c5;
71 }
72 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
73 if XNN_UNPREDICTABLE(mr < 8) {
74 c7 = c6;
75 }
76 float* c8 = (float*) ((uintptr_t) c7 + cm_stride);
77 if XNN_UNPREDICTABLE(mr <= 8) {
78 c8 = c7;
79 }
80 float* c9 = (float*) ((uintptr_t) c8 + cm_stride);
81 if XNN_UNPREDICTABLE(mr < 10) {
82 c9 = c8;
83 }
84 float* c10 = (float*) ((uintptr_t) c9 + cm_stride);
85 if XNN_UNPREDICTABLE(mr <= 10) {
86 c10 = c9;
87 }
88 float* c11 = (float*) ((uintptr_t) c10 + cm_stride);
89 if XNN_UNPREDICTABLE(mr < 12) {
90 c11 = c10;
91 }
92 float* c12 = (float*) ((uintptr_t) c11 + cm_stride);
93 if XNN_UNPREDICTABLE(mr <= 12) {
94 c12 = c11;
95 }
96 float* c13 = (float*) ((uintptr_t) c12 + cm_stride);
97 if XNN_UNPREDICTABLE(mr != 14) {
98 c13 = c12;
99 }
100
101 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
102 XNN_FORCE_REALIZATION(vsign_mask);
103 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point + 128);
104 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
105 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
106 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
107 XNN_FORCE_REALIZATION(vinput_zero_point);
108 // XNN_FORCE_REALIZATION(vinput_inv_scale);
109 // XNN_FORCE_REALIZATION(voutput_min);
110 // XNN_FORCE_REALIZATION(voutput_max);
111 do {
112 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
113 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
114 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
115 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
116 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
117 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
118 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
119 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
120 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
121 __m512i vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
122 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
123 __m512i vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF;
124 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
125 __m512i vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF;
126 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
127 __m512i vacc7x0123456789ABCDEF = vacc0x0123456789ABCDEF;
128 __m512i vacc1x7x0123456789ABCDEF = _mm512_setzero_epi32();
129 __m512i vacc8x0123456789ABCDEF = vacc0x0123456789ABCDEF;
130 __m512i vacc1x8x0123456789ABCDEF = _mm512_setzero_epi32();
131 __m512i vacc9x0123456789ABCDEF = vacc0x0123456789ABCDEF;
132 __m512i vacc1x9x0123456789ABCDEF = _mm512_setzero_epi32();
133 __m512i vacc10x0123456789ABCDEF = vacc0x0123456789ABCDEF;
134 __m512i vacc1x10x0123456789ABCDEF = _mm512_setzero_epi32();
135 __m512i vacc11x0123456789ABCDEF = vacc0x0123456789ABCDEF;
136 __m512i vacc1x11x0123456789ABCDEF = _mm512_setzero_epi32();
137 __m512i vacc12x0123456789ABCDEF = vacc0x0123456789ABCDEF;
138 __m512i vacc1x12x0123456789ABCDEF = _mm512_setzero_epi32();
139 __m512i vacc13x0123456789ABCDEF = vacc0x0123456789ABCDEF;
140 __m512i vacc1x13x0123456789ABCDEF = _mm512_setzero_epi32();
141 w = (const int32_t*) w + 16;
142
143 size_t p = ks;
144 do {
145 const int8_t* restrict a0 = a[0];
146 if XNN_UNPREDICTABLE(a0 != zero) {
147 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
148 } else {
149 a0 = zero_data;
150 }
151 const int8_t* restrict a1 = a[1];
152 if XNN_UNPREDICTABLE(a1 != zero) {
153 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
154 } else {
155 a1 = zero_data;
156 }
157 const int8_t* restrict a2 = a[2];
158 if XNN_UNPREDICTABLE(a2 != zero) {
159 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
160 } else {
161 a2 = zero_data;
162 }
163 const int8_t* restrict a3 = a[3];
164 if XNN_UNPREDICTABLE(a3 != zero) {
165 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
166 } else {
167 a3 = zero_data;
168 }
169 const int8_t* restrict a4 = a[4];
170 if XNN_UNPREDICTABLE(a4 != zero) {
171 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
172 } else {
173 a4 = zero_data;
174 }
175 const int8_t* restrict a5 = a[5];
176 if XNN_UNPREDICTABLE(a5 != zero) {
177 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
178 } else {
179 a5 = zero_data;
180 }
181 const int8_t* restrict a6 = a[6];
182 if XNN_UNPREDICTABLE(a6 != zero) {
183 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
184 } else {
185 a6 = zero_data;
186 }
187 const int8_t* restrict a7 = a[7];
188 if XNN_UNPREDICTABLE(a7 != zero) {
189 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
190 } else {
191 a7 = zero_data;
192 }
193 const int8_t* restrict a8 = a[8];
194 if XNN_UNPREDICTABLE(a8 != zero) {
195 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
196 } else {
197 a8 = zero_data;
198 }
199 const int8_t* restrict a9 = a[9];
200 if XNN_UNPREDICTABLE(a9 != zero) {
201 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
202 } else {
203 a9 = zero_data;
204 }
205 const int8_t* restrict a10 = a[10];
206 if XNN_UNPREDICTABLE(a10 != zero) {
207 a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
208 } else {
209 a10 = zero_data;
210 }
211 const int8_t* restrict a11 = a[11];
212 if XNN_UNPREDICTABLE(a11 != zero) {
213 a11 = (const int8_t*) ((uintptr_t) a11 + a_offset);
214 } else {
215 a11 = zero_data;
216 }
217 const int8_t* restrict a12 = a[12];
218 if XNN_UNPREDICTABLE(a12 != zero) {
219 a12 = (const int8_t*) ((uintptr_t) a12 + a_offset);
220 } else {
221 a12 = zero_data;
222 }
223 const int8_t* restrict a13 = a[13];
224 if XNN_UNPREDICTABLE(a13 != zero) {
225 a13 = (const int8_t*) ((uintptr_t) a13 + a_offset);
226 } else {
227 a13 = zero_data;
228 }
229 a += 14;
230
231 size_t k = kc;
232 while (k >= 8 * sizeof(int8_t)) {
233 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
234 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
235 a0 += 8;
236 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
237 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
238 a1 += 8;
239 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
240 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
241 a2 += 8;
242 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
243 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
244 a3 += 8;
245 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
246 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
247 a4 += 8;
248 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
249 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
250 a5 += 8;
251 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
252 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
253 a6 += 8;
254 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
255 const __m512i va7x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7 + 4)), vsign_mask);
256 a7 += 8;
257 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
258 const __m512i va8x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8 + 4)), vsign_mask);
259 a8 += 8;
260 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
261 const __m512i va9x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9 + 4)), vsign_mask);
262 a9 += 8;
263 const __m512i va10x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10)), vsign_mask);
264 const __m512i va10x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10 + 4)), vsign_mask);
265 a10 += 8;
266 const __m512i va11x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11)), vsign_mask);
267 const __m512i va11x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11 + 4)), vsign_mask);
268 a11 += 8;
269 const __m512i va12x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a12)), vsign_mask);
270 const __m512i va12x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a12 + 4)), vsign_mask);
271 a12 += 8;
272 const __m512i va13x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a13)), vsign_mask);
273 const __m512i va13x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a13 + 4)), vsign_mask);
274 a13 += 8;
275
276 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
277 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
278
279 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
280 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
281 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
282 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
283 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
284 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
285 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
286 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEFx0123);
287 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEFx0123);
288 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEFx0123);
289 vacc10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc10x0123456789ABCDEF, va10x0123, vb0123456789ABCDEFx0123);
290 vacc11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc11x0123456789ABCDEF, va11x0123, vb0123456789ABCDEFx0123);
291 vacc12x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc12x0123456789ABCDEF, va12x0123, vb0123456789ABCDEFx0123);
292 vacc13x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc13x0123456789ABCDEF, va13x0123, vb0123456789ABCDEFx0123);
293 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
294 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
295 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
296 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
297 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
298 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
299 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
300 vacc1x7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x7x0123456789ABCDEF, va7x4567, vb0123456789ABCDEFx4567);
301 vacc1x8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x8x0123456789ABCDEF, va8x4567, vb0123456789ABCDEFx4567);
302 vacc1x9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x9x0123456789ABCDEF, va9x4567, vb0123456789ABCDEFx4567);
303 vacc1x10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x10x0123456789ABCDEF, va10x4567, vb0123456789ABCDEFx4567);
304 vacc1x11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x11x0123456789ABCDEF, va11x4567, vb0123456789ABCDEFx4567);
305 vacc1x12x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x12x0123456789ABCDEF, va12x4567, vb0123456789ABCDEFx4567);
306 vacc1x13x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x13x0123456789ABCDEF, va13x4567, vb0123456789ABCDEFx4567);
307
308 w = (const int8_t*) w + 128;
309 k -= 8 * sizeof(int8_t);
310 }
311
312 if (k != 0) {
313 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
314 a0 += 4;
315 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
316 a1 += 4;
317 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
318 a2 += 4;
319 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
320 a3 += 4;
321 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
322 a4 += 4;
323 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
324 a5 += 4;
325 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
326 a6 += 4;
327 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
328 a7 += 4;
329 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
330 a8 += 4;
331 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
332 a9 += 4;
333 const __m512i va10x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10)), vsign_mask);
334 a10 += 4;
335 const __m512i va11x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11)), vsign_mask);
336 a11 += 4;
337 const __m512i va12x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a12)), vsign_mask);
338 a12 += 4;
339 const __m512i va13x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a13)), vsign_mask);
340 a13 += 4;
341
342 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
343
344 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
345 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
346 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
347 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
348 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
349 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
350 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
351 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEF);
352 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEF);
353 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEF);
354 vacc10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc10x0123456789ABCDEF, va10x0123, vb0123456789ABCDEF);
355 vacc11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc11x0123456789ABCDEF, va11x0123, vb0123456789ABCDEF);
356 vacc12x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc12x0123456789ABCDEF, va12x0123, vb0123456789ABCDEF);
357 vacc13x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc13x0123456789ABCDEF, va13x0123, vb0123456789ABCDEF);
358
359 w = (const int8_t*) w + 64;
360 k -= 4 * sizeof(int8_t);
361 }
362
363 p -= 14 * sizeof(void*);
364 } while (p != 0);
365
366 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
367 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
368 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
369 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
370 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
371 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
372 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
373 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, vacc1x7x0123456789ABCDEF);
374 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, vacc1x8x0123456789ABCDEF);
375 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, vacc1x9x0123456789ABCDEF);
376 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, vacc1x10x0123456789ABCDEF);
377 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, vacc1x11x0123456789ABCDEF);
378 vacc12x0123456789ABCDEF = _mm512_add_epi32(vacc12x0123456789ABCDEF, vacc1x12x0123456789ABCDEF);
379 vacc13x0123456789ABCDEF = _mm512_add_epi32(vacc13x0123456789ABCDEF, vacc1x13x0123456789ABCDEF);
380
381 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
382 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
383 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
384 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
385 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
386 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
387 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
388 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
389 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
390 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
391 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
392 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
393 __m512 vscaled12x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc12x0123456789ABCDEF);
394 __m512 vscaled13x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc13x0123456789ABCDEF);
395
396 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
397 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
398 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
399 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
400 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vinput_inv_scale);
401 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vinput_inv_scale);
402 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vinput_inv_scale);
403 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, vinput_inv_scale);
404 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, vinput_inv_scale);
405 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, vinput_inv_scale);
406 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, vinput_inv_scale);
407 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, vinput_inv_scale);
408 vscaled12x0123456789ABCDEF = _mm512_mul_ps(vscaled12x0123456789ABCDEF, vinput_inv_scale);
409 vscaled13x0123456789ABCDEF = _mm512_mul_ps(vscaled13x0123456789ABCDEF, vinput_inv_scale);
410
411 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
412 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
413 w = (const float*) w + 32;
414
415 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
416 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
417 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
418 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
419 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
420 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
421 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
422 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
423 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
424 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vscaled9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
425 vscaled10x0123456789ABCDEF = _mm512_fmadd_ps(vscaled10x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
426 vscaled11x0123456789ABCDEF = _mm512_fmadd_ps(vscaled11x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
427 vscaled12x0123456789ABCDEF = _mm512_fmadd_ps(vscaled12x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
428 vscaled13x0123456789ABCDEF = _mm512_fmadd_ps(vscaled13x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
429
430 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
431 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
432 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
433 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
434 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
435 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
436 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
437 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
438 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
439 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
440 vscaled10x0123456789ABCDEF = _mm512_max_ps(vscaled10x0123456789ABCDEF, voutput_min);
441 vscaled11x0123456789ABCDEF = _mm512_max_ps(vscaled11x0123456789ABCDEF, voutput_min);
442 vscaled12x0123456789ABCDEF = _mm512_max_ps(vscaled12x0123456789ABCDEF, voutput_min);
443 vscaled13x0123456789ABCDEF = _mm512_max_ps(vscaled13x0123456789ABCDEF, voutput_min);
444
445 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
446 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
447 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
448 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
449 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
450 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
451 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
452 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
453 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
454 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
455 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max);
456 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max);
457 vscaled12x0123456789ABCDEF = _mm512_min_ps(vscaled12x0123456789ABCDEF, voutput_max);
458 vscaled13x0123456789ABCDEF = _mm512_min_ps(vscaled13x0123456789ABCDEF, voutput_max);
459
460 if XNN_LIKELY(nc >= 16) {
461 _mm512_storeu_ps(c13, vscaled13x0123456789ABCDEF);
462 c13 = (float*) ((uintptr_t) c13 + cn_stride);
463 _mm512_storeu_ps(c12, vscaled12x0123456789ABCDEF);
464 c12 = (float*) ((uintptr_t) c12 + cn_stride);
465 _mm512_storeu_ps(c11, vscaled11x0123456789ABCDEF);
466 c11 = (float*) ((uintptr_t) c11 + cn_stride);
467 _mm512_storeu_ps(c10, vscaled10x0123456789ABCDEF);
468 c10 = (float*) ((uintptr_t) c10 + cn_stride);
469 _mm512_storeu_ps(c9, vscaled9x0123456789ABCDEF);
470 c9 = (float*) ((uintptr_t) c9 + cn_stride);
471 _mm512_storeu_ps(c8, vscaled8x0123456789ABCDEF);
472 c8 = (float*) ((uintptr_t) c8 + cn_stride);
473 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
474 c7 = (float*) ((uintptr_t) c7 + cn_stride);
475 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
476 c6 = (float*) ((uintptr_t) c6 + cn_stride);
477 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
478 c5 = (float*) ((uintptr_t) c5 + cn_stride);
479 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
480 c4 = (float*) ((uintptr_t) c4 + cn_stride);
481 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
482 c3 = (float*) ((uintptr_t) c3 + cn_stride);
483 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
484 c2 = (float*) ((uintptr_t) c2 + cn_stride);
485 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
486 c1 = (float*) ((uintptr_t) c1 + cn_stride);
487 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
488 c0 = (float*) ((uintptr_t) c0 + cn_stride);
489 a = (const int8_t**restrict) ((uintptr_t) a - ks);
490 nc -= 16;
491 } else {
492 // Prepare mask for valid 32-bit elements (depends on nc).
493 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
494 _mm512_mask_storeu_ps(c13, vmask, vscaled13x0123456789ABCDEF);
495 _mm512_mask_storeu_ps(c12, vmask, vscaled12x0123456789ABCDEF);
496 _mm512_mask_storeu_ps(c11, vmask, vscaled11x0123456789ABCDEF);
497 _mm512_mask_storeu_ps(c10, vmask, vscaled10x0123456789ABCDEF);
498 _mm512_mask_storeu_ps(c9, vmask, vscaled9x0123456789ABCDEF);
499 _mm512_mask_storeu_ps(c8, vmask, vscaled8x0123456789ABCDEF);
500 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
501 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
502 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
503 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
504 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
505 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
506 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
507 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
508 nc = 0;
509 }
510 } while (nc != 0);
511 }