Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-14x16c4-minmax-avx512vnni-prfm.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 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_14x16c4__avx512vnni_prfm(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 size_t ks,
27 const int8_t** restrict a,
28 const void* restrict w,
29 int8_t* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 14);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 4 * sizeof(int8_t));
46 int8_t* c0 = c;
47 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 c3 = c2;
58 }
59 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 c4 = c3;
62 }
63 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
64 if XNN_UNPREDICTABLE(mr < 6) {
65 c5 = c4;
66 }
67 int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
68 if XNN_UNPREDICTABLE(mr <= 6) {
69 c6 = c5;
70 }
71 int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride);
72 if XNN_UNPREDICTABLE(mr < 8) {
73 c7 = c6;
74 }
75 int8_t* c8 = (int8_t*) ((uintptr_t) c7 + cm_stride);
76 if XNN_UNPREDICTABLE(mr <= 8) {
77 c8 = c7;
78 }
79 int8_t* c9 = (int8_t*) ((uintptr_t) c8 + cm_stride);
80 if XNN_UNPREDICTABLE(mr < 10) {
81 c9 = c8;
82 }
83 int8_t* c10 = (int8_t*) ((uintptr_t) c9 + cm_stride);
84 if XNN_UNPREDICTABLE(mr <= 10) {
85 c10 = c9;
86 }
87 int8_t* c11 = (int8_t*) ((uintptr_t) c10 + cm_stride);
88 if XNN_UNPREDICTABLE(mr < 12) {
89 c11 = c10;
90 }
91 int8_t* c12 = (int8_t*) ((uintptr_t) c11 + cm_stride);
92 if XNN_UNPREDICTABLE(mr <= 12) {
93 c12 = c11;
94 }
95 int8_t* c13 = (int8_t*) ((uintptr_t) c12 + cm_stride);
96 if XNN_UNPREDICTABLE(mr != 14) {
97 c13 = c12;
98 }
99
100 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
101 XNN_FORCE_REALIZATION(vsign_mask);
102 const __m512 voutput_max_less_zero_point = _mm512_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
103 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
104 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
105 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
106 // XNN_FORCE_REALIZATION(voutput_zero_point);
107 // XNN_FORCE_REALIZATION(voutput_min);
108 do {
109 __m512i vacc0x0123456789ABCDEF = _mm512_load_epi32(w);
110 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
111 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
112 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
113 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
114 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
115 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
116 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
117 __m512i vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
118 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
119 __m512i vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF;
120 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
121 __m512i vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF;
122 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
123 __m512i vacc7x0123456789ABCDEF = vacc0x0123456789ABCDEF;
124 __m512i vacc1x7x0123456789ABCDEF = _mm512_setzero_epi32();
125 __m512i vacc8x0123456789ABCDEF = vacc0x0123456789ABCDEF;
126 __m512i vacc1x8x0123456789ABCDEF = _mm512_setzero_epi32();
127 __m512i vacc9x0123456789ABCDEF = vacc0x0123456789ABCDEF;
128 __m512i vacc1x9x0123456789ABCDEF = _mm512_setzero_epi32();
129 __m512i vacc10x0123456789ABCDEF = vacc0x0123456789ABCDEF;
130 __m512i vacc1x10x0123456789ABCDEF = _mm512_setzero_epi32();
131 __m512i vacc11x0123456789ABCDEF = vacc0x0123456789ABCDEF;
132 __m512i vacc1x11x0123456789ABCDEF = _mm512_setzero_epi32();
133 __m512i vacc12x0123456789ABCDEF = vacc0x0123456789ABCDEF;
134 __m512i vacc1x12x0123456789ABCDEF = _mm512_setzero_epi32();
135 __m512i vacc13x0123456789ABCDEF = vacc0x0123456789ABCDEF;
136 __m512i vacc1x13x0123456789ABCDEF = _mm512_setzero_epi32();
137 w = (const int32_t*) w + 16;
138
139 size_t p = ks;
140 do {
141 const int8_t* restrict a0 = a[0];
142 if XNN_UNPREDICTABLE(a0 != zero) {
143 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
144 }
145 const int8_t* restrict a1 = a[1];
146 if XNN_UNPREDICTABLE(a1 != zero) {
147 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
148 }
149 const int8_t* restrict a2 = a[2];
150 if XNN_UNPREDICTABLE(a2 != zero) {
151 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
152 }
153 const int8_t* restrict a3 = a[3];
154 if XNN_UNPREDICTABLE(a3 != zero) {
155 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
156 }
157 const int8_t* restrict a4 = a[4];
158 if XNN_UNPREDICTABLE(a4 != zero) {
159 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
160 }
161 const int8_t* restrict a5 = a[5];
162 if XNN_UNPREDICTABLE(a5 != zero) {
163 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
164 }
165 const int8_t* restrict a6 = a[6];
166 if XNN_UNPREDICTABLE(a6 != zero) {
167 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
168 }
169 const int8_t* restrict a7 = a[7];
170 if XNN_UNPREDICTABLE(a7 != zero) {
171 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
172 }
173 const int8_t* restrict a8 = a[8];
174 if XNN_UNPREDICTABLE(a8 != zero) {
175 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
176 }
177 const int8_t* restrict a9 = a[9];
178 if XNN_UNPREDICTABLE(a9 != zero) {
179 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
180 }
181 const int8_t* restrict a10 = a[10];
182 if XNN_UNPREDICTABLE(a10 != zero) {
183 a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
184 }
185 const int8_t* restrict a11 = a[11];
186 if XNN_UNPREDICTABLE(a11 != zero) {
187 a11 = (const int8_t*) ((uintptr_t) a11 + a_offset);
188 }
189 const int8_t* restrict a12 = a[12];
190 if XNN_UNPREDICTABLE(a12 != zero) {
191 a12 = (const int8_t*) ((uintptr_t) a12 + a_offset);
192 }
193 const int8_t* restrict a13 = a[13];
194 if XNN_UNPREDICTABLE(a13 != zero) {
195 a13 = (const int8_t*) ((uintptr_t) a13 + a_offset);
196 }
197 a += 14;
198
199 size_t k = kc;
200 while (k >= 8 * sizeof(int8_t)) {
201 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
202 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
203 a0 += 8;
204 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
205 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
206 a1 += 8;
207 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
208 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
209 a2 += 8;
210 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
211 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
212 a3 += 8;
213 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
214 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
215 a4 += 8;
216 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
217 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
218 a5 += 8;
219 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
220 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
221 a6 += 8;
222 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
223 const __m512i va7x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7 + 4)), vsign_mask);
224 a7 += 8;
225 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
226 const __m512i va8x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8 + 4)), vsign_mask);
227 a8 += 8;
228 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
229 const __m512i va9x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9 + 4)), vsign_mask);
230 a9 += 8;
231 const __m512i va10x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10)), vsign_mask);
232 const __m512i va10x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10 + 4)), vsign_mask);
233 a10 += 8;
234 const __m512i va11x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11)), vsign_mask);
235 const __m512i va11x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11 + 4)), vsign_mask);
236 a11 += 8;
237 const __m512i va12x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a12)), vsign_mask);
238 const __m512i va12x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a12 + 4)), vsign_mask);
239 a12 += 8;
240 const __m512i va13x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a13)), vsign_mask);
241 const __m512i va13x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a13 + 4)), vsign_mask);
242 a13 += 8;
243
244 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
245 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
246 xnn_prefetch_to_l1((const int8_t*) w + 896);
247
248 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
249 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
250 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
251 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
252 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
253 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
254 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
255 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEFx0123);
256 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEFx0123);
257 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEFx0123);
258 vacc10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc10x0123456789ABCDEF, va10x0123, vb0123456789ABCDEFx0123);
259 vacc11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc11x0123456789ABCDEF, va11x0123, vb0123456789ABCDEFx0123);
260 vacc12x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc12x0123456789ABCDEF, va12x0123, vb0123456789ABCDEFx0123);
261 vacc13x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc13x0123456789ABCDEF, va13x0123, vb0123456789ABCDEFx0123);
262 xnn_prefetch_to_l1((const int8_t*) w + 960);
263 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
264 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
265 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
266 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
267 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
268 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
269 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
270 vacc1x7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x7x0123456789ABCDEF, va7x4567, vb0123456789ABCDEFx4567);
271 vacc1x8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x8x0123456789ABCDEF, va8x4567, vb0123456789ABCDEFx4567);
272 vacc1x9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x9x0123456789ABCDEF, va9x4567, vb0123456789ABCDEFx4567);
273 vacc1x10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x10x0123456789ABCDEF, va10x4567, vb0123456789ABCDEFx4567);
274 vacc1x11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x11x0123456789ABCDEF, va11x4567, vb0123456789ABCDEFx4567);
275 vacc1x12x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x12x0123456789ABCDEF, va12x4567, vb0123456789ABCDEFx4567);
276 vacc1x13x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x13x0123456789ABCDEF, va13x4567, vb0123456789ABCDEFx4567);
277
278 w = (const int8_t*) w + 128;
279 k -= 8 * sizeof(int8_t);
280 }
281
282 if (k != 0) {
283 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
284 a0 += 4;
285 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
286 a1 += 4;
287 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
288 a2 += 4;
289 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
290 a3 += 4;
291 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
292 a4 += 4;
293 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
294 a5 += 4;
295 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
296 a6 += 4;
297 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
298 a7 += 4;
299 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
300 a8 += 4;
301 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
302 a9 += 4;
303 const __m512i va10x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10)), vsign_mask);
304 a10 += 4;
305 const __m512i va11x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11)), vsign_mask);
306 a11 += 4;
307 const __m512i va12x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a12)), vsign_mask);
308 a12 += 4;
309 const __m512i va13x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a13)), vsign_mask);
310 a13 += 4;
311
312 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
313 xnn_prefetch_to_l1((const int8_t*) w + 960);
314
315 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
316 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
317 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
318 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
319 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
320 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
321 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
322 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEF);
323 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEF);
324 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEF);
325 vacc10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc10x0123456789ABCDEF, va10x0123, vb0123456789ABCDEF);
326 vacc11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc11x0123456789ABCDEF, va11x0123, vb0123456789ABCDEF);
327 vacc12x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc12x0123456789ABCDEF, va12x0123, vb0123456789ABCDEF);
328 vacc13x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc13x0123456789ABCDEF, va13x0123, vb0123456789ABCDEF);
329
330 w = (const int8_t*) w + 64;
331 k -= 4 * sizeof(int8_t);
332 }
333
334 p -= 14 * sizeof(void*);
335 } while (p != 0);
336
337 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
338 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
339 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
340 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
341 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
342 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
343 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
344 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, vacc1x7x0123456789ABCDEF);
345 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, vacc1x8x0123456789ABCDEF);
346 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, vacc1x9x0123456789ABCDEF);
347 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, vacc1x10x0123456789ABCDEF);
348 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, vacc1x11x0123456789ABCDEF);
349 vacc12x0123456789ABCDEF = _mm512_add_epi32(vacc12x0123456789ABCDEF, vacc1x12x0123456789ABCDEF);
350 vacc13x0123456789ABCDEF = _mm512_add_epi32(vacc13x0123456789ABCDEF, vacc1x13x0123456789ABCDEF);
351
352 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
353 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
354 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
355 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
356 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
357 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
358 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
359 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
360 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
361 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
362 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
363 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
364 __m512 vscaled12x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc12x0123456789ABCDEF);
365 __m512 vscaled13x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc13x0123456789ABCDEF);
366
367 const __m512 vscale012345678ABCDEF = _mm512_load_ps(w);
368 w = (const float*) w + 16;
369 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale012345678ABCDEF);
370 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vscale012345678ABCDEF);
371 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vscale012345678ABCDEF);
372 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vscale012345678ABCDEF);
373 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vscale012345678ABCDEF);
374 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vscale012345678ABCDEF);
375 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vscale012345678ABCDEF);
376 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, vscale012345678ABCDEF);
377 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, vscale012345678ABCDEF);
378 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, vscale012345678ABCDEF);
379 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, vscale012345678ABCDEF);
380 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, vscale012345678ABCDEF);
381 vscaled12x0123456789ABCDEF = _mm512_mul_ps(vscaled12x0123456789ABCDEF, vscale012345678ABCDEF);
382 vscaled13x0123456789ABCDEF = _mm512_mul_ps(vscaled13x0123456789ABCDEF, vscale012345678ABCDEF);
383
384 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
385 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max_less_zero_point);
386 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max_less_zero_point);
387 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max_less_zero_point);
388 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max_less_zero_point);
389 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max_less_zero_point);
390 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max_less_zero_point);
391 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max_less_zero_point);
392 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max_less_zero_point);
393 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max_less_zero_point);
394 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max_less_zero_point);
395 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max_less_zero_point);
396 vscaled12x0123456789ABCDEF = _mm512_min_ps(vscaled12x0123456789ABCDEF, voutput_max_less_zero_point);
397 vscaled13x0123456789ABCDEF = _mm512_min_ps(vscaled13x0123456789ABCDEF, voutput_max_less_zero_point);
398
399 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
400 vacc1x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled1x0123456789ABCDEF);
401 vacc2x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled2x0123456789ABCDEF);
402 vacc3x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled3x0123456789ABCDEF);
403 vacc4x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled4x0123456789ABCDEF);
404 vacc5x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled5x0123456789ABCDEF);
405 vacc6x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled6x0123456789ABCDEF);
406 vacc7x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled7x0123456789ABCDEF);
407 vacc8x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled8x0123456789ABCDEF);
408 vacc9x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled9x0123456789ABCDEF);
409 vacc10x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled10x0123456789ABCDEF);
410 vacc11x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled11x0123456789ABCDEF);
411 vacc12x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled12x0123456789ABCDEF);
412 vacc13x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled13x0123456789ABCDEF);
413
414 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
415 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, voutput_zero_point);
416 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, voutput_zero_point);
417 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, voutput_zero_point);
418 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, voutput_zero_point);
419 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, voutput_zero_point);
420 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, voutput_zero_point);
421 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, voutput_zero_point);
422 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, voutput_zero_point);
423 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, voutput_zero_point);
424 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, voutput_zero_point);
425 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, voutput_zero_point);
426 vacc12x0123456789ABCDEF = _mm512_add_epi32(vacc12x0123456789ABCDEF, voutput_zero_point);
427 vacc13x0123456789ABCDEF = _mm512_add_epi32(vacc13x0123456789ABCDEF, voutput_zero_point);
428
429 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
430 __m128i vout1x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc1x0123456789ABCDEF);
431 __m128i vout2x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc2x0123456789ABCDEF);
432 __m128i vout3x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc3x0123456789ABCDEF);
433 __m128i vout4x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc4x0123456789ABCDEF);
434 __m128i vout5x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc5x0123456789ABCDEF);
435 __m128i vout6x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc6x0123456789ABCDEF);
436 __m128i vout7x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc7x0123456789ABCDEF);
437 __m128i vout8x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc8x0123456789ABCDEF);
438 __m128i vout9x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc9x0123456789ABCDEF);
439 __m128i vout10x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc10x0123456789ABCDEF);
440 __m128i vout11x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc11x0123456789ABCDEF);
441 __m128i vout12x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc12x0123456789ABCDEF);
442 __m128i vout13x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc13x0123456789ABCDEF);
443
444 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
445 vout1x0123456789ABCDEF = _mm_max_epi8(vout1x0123456789ABCDEF, voutput_min);
446 vout2x0123456789ABCDEF = _mm_max_epi8(vout2x0123456789ABCDEF, voutput_min);
447 vout3x0123456789ABCDEF = _mm_max_epi8(vout3x0123456789ABCDEF, voutput_min);
448 vout4x0123456789ABCDEF = _mm_max_epi8(vout4x0123456789ABCDEF, voutput_min);
449 vout5x0123456789ABCDEF = _mm_max_epi8(vout5x0123456789ABCDEF, voutput_min);
450 vout6x0123456789ABCDEF = _mm_max_epi8(vout6x0123456789ABCDEF, voutput_min);
451 vout7x0123456789ABCDEF = _mm_max_epi8(vout7x0123456789ABCDEF, voutput_min);
452 vout8x0123456789ABCDEF = _mm_max_epi8(vout8x0123456789ABCDEF, voutput_min);
453 vout9x0123456789ABCDEF = _mm_max_epi8(vout9x0123456789ABCDEF, voutput_min);
454 vout10x0123456789ABCDEF = _mm_max_epi8(vout10x0123456789ABCDEF, voutput_min);
455 vout11x0123456789ABCDEF = _mm_max_epi8(vout11x0123456789ABCDEF, voutput_min);
456 vout12x0123456789ABCDEF = _mm_max_epi8(vout12x0123456789ABCDEF, voutput_min);
457 vout13x0123456789ABCDEF = _mm_max_epi8(vout13x0123456789ABCDEF, voutput_min);
458
459 if (nc >= 16) {
460 _mm_storeu_si128((__m128i*) c13, vout13x0123456789ABCDEF);
461 c13 = (int8_t*) ((uintptr_t) c13 + cn_stride);
462 _mm_storeu_si128((__m128i*) c12, vout12x0123456789ABCDEF);
463 c12 = (int8_t*) ((uintptr_t) c12 + cn_stride);
464 _mm_storeu_si128((__m128i*) c11, vout11x0123456789ABCDEF);
465 c11 = (int8_t*) ((uintptr_t) c11 + cn_stride);
466 _mm_storeu_si128((__m128i*) c10, vout10x0123456789ABCDEF);
467 c10 = (int8_t*) ((uintptr_t) c10 + cn_stride);
468 _mm_storeu_si128((__m128i*) c9, vout9x0123456789ABCDEF);
469 c9 = (int8_t*) ((uintptr_t) c9 + cn_stride);
470 _mm_storeu_si128((__m128i*) c8, vout8x0123456789ABCDEF);
471 c8 = (int8_t*) ((uintptr_t) c8 + cn_stride);
472 _mm_storeu_si128((__m128i*) c7, vout7x0123456789ABCDEF);
473 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
474 _mm_storeu_si128((__m128i*) c6, vout6x0123456789ABCDEF);
475 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
476 _mm_storeu_si128((__m128i*) c5, vout5x0123456789ABCDEF);
477 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
478 _mm_storeu_si128((__m128i*) c4, vout4x0123456789ABCDEF);
479 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
480 _mm_storeu_si128((__m128i*) c3, vout3x0123456789ABCDEF);
481 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
482 _mm_storeu_si128((__m128i*) c2, vout2x0123456789ABCDEF);
483 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
484 _mm_storeu_si128((__m128i*) c1, vout1x0123456789ABCDEF);
485 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
486 _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);
487 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
488 a = (const int8_t**restrict) ((uintptr_t) a - ks);
489 nc -= 16;
490 } else {
491 // Prepare mask for valid 8-bit elements (depends on nc).
492 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
493 _mm_mask_storeu_epi8(c13, vmask, vout13x0123456789ABCDEF);
494 _mm_mask_storeu_epi8(c12, vmask, vout12x0123456789ABCDEF);
495 _mm_mask_storeu_epi8(c11, vmask, vout11x0123456789ABCDEF);
496 _mm_mask_storeu_epi8(c10, vmask, vout10x0123456789ABCDEF);
497 _mm_mask_storeu_epi8(c9, vmask, vout9x0123456789ABCDEF);
498 _mm_mask_storeu_epi8(c8, vmask, vout8x0123456789ABCDEF);
499 _mm_mask_storeu_epi8(c7, vmask, vout7x0123456789ABCDEF);
500 _mm_mask_storeu_epi8(c6, vmask, vout6x0123456789ABCDEF);
501 _mm_mask_storeu_epi8(c5, vmask, vout5x0123456789ABCDEF);
502 _mm_mask_storeu_epi8(c4, vmask, vout4x0123456789ABCDEF);
503 _mm_mask_storeu_epi8(c3, vmask, vout3x0123456789ABCDEF);
504 _mm_mask_storeu_epi8(c2, vmask, vout2x0123456789ABCDEF);
505 _mm_mask_storeu_epi8(c1, vmask, vout1x0123456789ABCDEF);
506 _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);
507 nc = 0;
508 }
509 } while (nc != 0);
510 }