Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-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_qs8_qc8w_igemm_minmax_fp32_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 int8_t* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 14);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 4 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr < 4) {
56 c3 = c2;
57 }
58 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
59 if XNN_UNPREDICTABLE(mr <= 4) {
60 c4 = c3;
61 }
62 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
63 if XNN_UNPREDICTABLE(mr < 6) {
64 c5 = c4;
65 }
66 int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
67 if XNN_UNPREDICTABLE(mr <= 6) {
68 c6 = c5;
69 }
70 int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride);
71 if XNN_UNPREDICTABLE(mr < 8) {
72 c7 = c6;
73 }
74 int8_t* c8 = (int8_t*) ((uintptr_t) c7 + cm_stride);
75 if XNN_UNPREDICTABLE(mr <= 8) {
76 c8 = c7;
77 }
78 int8_t* c9 = (int8_t*) ((uintptr_t) c8 + cm_stride);
79 if XNN_UNPREDICTABLE(mr < 10) {
80 c9 = c8;
81 }
82 int8_t* c10 = (int8_t*) ((uintptr_t) c9 + cm_stride);
83 if XNN_UNPREDICTABLE(mr <= 10) {
84 c10 = c9;
85 }
86 int8_t* c11 = (int8_t*) ((uintptr_t) c10 + cm_stride);
87 if XNN_UNPREDICTABLE(mr < 12) {
88 c11 = c10;
89 }
90 int8_t* c12 = (int8_t*) ((uintptr_t) c11 + cm_stride);
91 if XNN_UNPREDICTABLE(mr <= 12) {
92 c12 = c11;
93 }
94 int8_t* c13 = (int8_t*) ((uintptr_t) c12 + cm_stride);
95 if XNN_UNPREDICTABLE(mr != 14) {
96 c13 = c12;
97 }
98
99 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
100 XNN_FORCE_REALIZATION(vsign_mask);
101 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);
102 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
103 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
104 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
105 // XNN_FORCE_REALIZATION(voutput_zero_point);
106 // XNN_FORCE_REALIZATION(voutput_min);
107 do {
108 __m512i vacc0x0123456789ABCDEF = _mm512_load_epi32(w);
109 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
110 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
111 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
112 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
113 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
114 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
115 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
116 __m512i vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
117 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
118 __m512i vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF;
119 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
120 __m512i vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF;
121 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
122 __m512i vacc7x0123456789ABCDEF = vacc0x0123456789ABCDEF;
123 __m512i vacc1x7x0123456789ABCDEF = _mm512_setzero_epi32();
124 __m512i vacc8x0123456789ABCDEF = vacc0x0123456789ABCDEF;
125 __m512i vacc1x8x0123456789ABCDEF = _mm512_setzero_epi32();
126 __m512i vacc9x0123456789ABCDEF = vacc0x0123456789ABCDEF;
127 __m512i vacc1x9x0123456789ABCDEF = _mm512_setzero_epi32();
128 __m512i vacc10x0123456789ABCDEF = vacc0x0123456789ABCDEF;
129 __m512i vacc1x10x0123456789ABCDEF = _mm512_setzero_epi32();
130 __m512i vacc11x0123456789ABCDEF = vacc0x0123456789ABCDEF;
131 __m512i vacc1x11x0123456789ABCDEF = _mm512_setzero_epi32();
132 __m512i vacc12x0123456789ABCDEF = vacc0x0123456789ABCDEF;
133 __m512i vacc1x12x0123456789ABCDEF = _mm512_setzero_epi32();
134 __m512i vacc13x0123456789ABCDEF = vacc0x0123456789ABCDEF;
135 __m512i vacc1x13x0123456789ABCDEF = _mm512_setzero_epi32();
136 w = (const int32_t*) w + 16;
137
138 size_t p = ks;
139 do {
140 const int8_t* restrict a0 = a[0];
141 if XNN_UNPREDICTABLE(a0 != zero) {
142 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
143 }
144 const int8_t* restrict a1 = a[1];
145 if XNN_UNPREDICTABLE(a1 != zero) {
146 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
147 }
148 const int8_t* restrict a2 = a[2];
149 if XNN_UNPREDICTABLE(a2 != zero) {
150 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
151 }
152 const int8_t* restrict a3 = a[3];
153 if XNN_UNPREDICTABLE(a3 != zero) {
154 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
155 }
156 const int8_t* restrict a4 = a[4];
157 if XNN_UNPREDICTABLE(a4 != zero) {
158 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
159 }
160 const int8_t* restrict a5 = a[5];
161 if XNN_UNPREDICTABLE(a5 != zero) {
162 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
163 }
164 const int8_t* restrict a6 = a[6];
165 if XNN_UNPREDICTABLE(a6 != zero) {
166 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
167 }
168 const int8_t* restrict a7 = a[7];
169 if XNN_UNPREDICTABLE(a7 != zero) {
170 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
171 }
172 const int8_t* restrict a8 = a[8];
173 if XNN_UNPREDICTABLE(a8 != zero) {
174 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
175 }
176 const int8_t* restrict a9 = a[9];
177 if XNN_UNPREDICTABLE(a9 != zero) {
178 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
179 }
180 const int8_t* restrict a10 = a[10];
181 if XNN_UNPREDICTABLE(a10 != zero) {
182 a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
183 }
184 const int8_t* restrict a11 = a[11];
185 if XNN_UNPREDICTABLE(a11 != zero) {
186 a11 = (const int8_t*) ((uintptr_t) a11 + a_offset);
187 }
188 const int8_t* restrict a12 = a[12];
189 if XNN_UNPREDICTABLE(a12 != zero) {
190 a12 = (const int8_t*) ((uintptr_t) a12 + a_offset);
191 }
192 const int8_t* restrict a13 = a[13];
193 if XNN_UNPREDICTABLE(a13 != zero) {
194 a13 = (const int8_t*) ((uintptr_t) a13 + a_offset);
195 }
196 a += 14;
197
198 size_t k = kc;
199 while (k >= 8 * sizeof(int8_t)) {
200 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
201 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
202 a0 += 8;
203 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
204 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
205 a1 += 8;
206 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
207 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
208 a2 += 8;
209 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
210 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
211 a3 += 8;
212 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
213 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
214 a4 += 8;
215 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
216 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
217 a5 += 8;
218 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
219 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
220 a6 += 8;
221 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
222 const __m512i va7x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7 + 4)), vsign_mask);
223 a7 += 8;
224 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
225 const __m512i va8x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8 + 4)), vsign_mask);
226 a8 += 8;
227 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
228 const __m512i va9x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9 + 4)), vsign_mask);
229 a9 += 8;
230 const __m512i va10x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10)), vsign_mask);
231 const __m512i va10x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10 + 4)), vsign_mask);
232 a10 += 8;
233 const __m512i va11x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11)), vsign_mask);
234 const __m512i va11x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11 + 4)), vsign_mask);
235 a11 += 8;
236 const __m512i va12x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a12)), vsign_mask);
237 const __m512i va12x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a12 + 4)), vsign_mask);
238 a12 += 8;
239 const __m512i va13x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a13)), vsign_mask);
240 const __m512i va13x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a13 + 4)), vsign_mask);
241 a13 += 8;
242
243 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
244 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
245
246 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
247 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
248 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
249 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
250 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
251 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
252 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
253 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEFx0123);
254 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEFx0123);
255 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEFx0123);
256 vacc10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc10x0123456789ABCDEF, va10x0123, vb0123456789ABCDEFx0123);
257 vacc11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc11x0123456789ABCDEF, va11x0123, vb0123456789ABCDEFx0123);
258 vacc12x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc12x0123456789ABCDEF, va12x0123, vb0123456789ABCDEFx0123);
259 vacc13x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc13x0123456789ABCDEF, va13x0123, vb0123456789ABCDEFx0123);
260 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
261 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
262 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
263 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
264 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
265 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
266 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
267 vacc1x7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x7x0123456789ABCDEF, va7x4567, vb0123456789ABCDEFx4567);
268 vacc1x8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x8x0123456789ABCDEF, va8x4567, vb0123456789ABCDEFx4567);
269 vacc1x9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x9x0123456789ABCDEF, va9x4567, vb0123456789ABCDEFx4567);
270 vacc1x10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x10x0123456789ABCDEF, va10x4567, vb0123456789ABCDEFx4567);
271 vacc1x11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x11x0123456789ABCDEF, va11x4567, vb0123456789ABCDEFx4567);
272 vacc1x12x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x12x0123456789ABCDEF, va12x4567, vb0123456789ABCDEFx4567);
273 vacc1x13x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x13x0123456789ABCDEF, va13x4567, vb0123456789ABCDEFx4567);
274
275 w = (const int8_t*) w + 128;
276 k -= 8 * sizeof(int8_t);
277 }
278
279 if (k != 0) {
280 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
281 a0 += 4;
282 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
283 a1 += 4;
284 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
285 a2 += 4;
286 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
287 a3 += 4;
288 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
289 a4 += 4;
290 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
291 a5 += 4;
292 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
293 a6 += 4;
294 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
295 a7 += 4;
296 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
297 a8 += 4;
298 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
299 a9 += 4;
300 const __m512i va10x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10)), vsign_mask);
301 a10 += 4;
302 const __m512i va11x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11)), vsign_mask);
303 a11 += 4;
304 const __m512i va12x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a12)), vsign_mask);
305 a12 += 4;
306 const __m512i va13x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a13)), vsign_mask);
307 a13 += 4;
308
309 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
310
311 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
312 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
313 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
314 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
315 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
316 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
317 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
318 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEF);
319 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEF);
320 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEF);
321 vacc10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc10x0123456789ABCDEF, va10x0123, vb0123456789ABCDEF);
322 vacc11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc11x0123456789ABCDEF, va11x0123, vb0123456789ABCDEF);
323 vacc12x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc12x0123456789ABCDEF, va12x0123, vb0123456789ABCDEF);
324 vacc13x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc13x0123456789ABCDEF, va13x0123, vb0123456789ABCDEF);
325
326 w = (const int8_t*) w + 64;
327 k -= 4 * sizeof(int8_t);
328 }
329
330 p -= 14 * sizeof(void*);
331 } while (p != 0);
332
333 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
334 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
335 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
336 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
337 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
338 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
339 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
340 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, vacc1x7x0123456789ABCDEF);
341 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, vacc1x8x0123456789ABCDEF);
342 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, vacc1x9x0123456789ABCDEF);
343 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, vacc1x10x0123456789ABCDEF);
344 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, vacc1x11x0123456789ABCDEF);
345 vacc12x0123456789ABCDEF = _mm512_add_epi32(vacc12x0123456789ABCDEF, vacc1x12x0123456789ABCDEF);
346 vacc13x0123456789ABCDEF = _mm512_add_epi32(vacc13x0123456789ABCDEF, vacc1x13x0123456789ABCDEF);
347
348 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
349 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
350 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
351 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
352 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
353 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
354 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
355 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
356 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
357 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
358 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
359 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
360 __m512 vscaled12x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc12x0123456789ABCDEF);
361 __m512 vscaled13x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc13x0123456789ABCDEF);
362
363 const __m512 vscale012345678ABCDEF = _mm512_load_ps(w);
364 w = (const float*) w + 16;
365 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale012345678ABCDEF);
366 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vscale012345678ABCDEF);
367 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vscale012345678ABCDEF);
368 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vscale012345678ABCDEF);
369 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vscale012345678ABCDEF);
370 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vscale012345678ABCDEF);
371 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vscale012345678ABCDEF);
372 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, vscale012345678ABCDEF);
373 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, vscale012345678ABCDEF);
374 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, vscale012345678ABCDEF);
375 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, vscale012345678ABCDEF);
376 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, vscale012345678ABCDEF);
377 vscaled12x0123456789ABCDEF = _mm512_mul_ps(vscaled12x0123456789ABCDEF, vscale012345678ABCDEF);
378 vscaled13x0123456789ABCDEF = _mm512_mul_ps(vscaled13x0123456789ABCDEF, vscale012345678ABCDEF);
379
380 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
381 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max_less_zero_point);
382 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max_less_zero_point);
383 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max_less_zero_point);
384 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max_less_zero_point);
385 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max_less_zero_point);
386 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max_less_zero_point);
387 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max_less_zero_point);
388 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max_less_zero_point);
389 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max_less_zero_point);
390 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max_less_zero_point);
391 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max_less_zero_point);
392 vscaled12x0123456789ABCDEF = _mm512_min_ps(vscaled12x0123456789ABCDEF, voutput_max_less_zero_point);
393 vscaled13x0123456789ABCDEF = _mm512_min_ps(vscaled13x0123456789ABCDEF, voutput_max_less_zero_point);
394
395 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
396 vacc1x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled1x0123456789ABCDEF);
397 vacc2x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled2x0123456789ABCDEF);
398 vacc3x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled3x0123456789ABCDEF);
399 vacc4x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled4x0123456789ABCDEF);
400 vacc5x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled5x0123456789ABCDEF);
401 vacc6x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled6x0123456789ABCDEF);
402 vacc7x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled7x0123456789ABCDEF);
403 vacc8x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled8x0123456789ABCDEF);
404 vacc9x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled9x0123456789ABCDEF);
405 vacc10x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled10x0123456789ABCDEF);
406 vacc11x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled11x0123456789ABCDEF);
407 vacc12x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled12x0123456789ABCDEF);
408 vacc13x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled13x0123456789ABCDEF);
409
410 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
411 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, voutput_zero_point);
412 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, voutput_zero_point);
413 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, voutput_zero_point);
414 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, voutput_zero_point);
415 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, voutput_zero_point);
416 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, voutput_zero_point);
417 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, voutput_zero_point);
418 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, voutput_zero_point);
419 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, voutput_zero_point);
420 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, voutput_zero_point);
421 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, voutput_zero_point);
422 vacc12x0123456789ABCDEF = _mm512_add_epi32(vacc12x0123456789ABCDEF, voutput_zero_point);
423 vacc13x0123456789ABCDEF = _mm512_add_epi32(vacc13x0123456789ABCDEF, voutput_zero_point);
424
425 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
426 __m128i vout1x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc1x0123456789ABCDEF);
427 __m128i vout2x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc2x0123456789ABCDEF);
428 __m128i vout3x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc3x0123456789ABCDEF);
429 __m128i vout4x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc4x0123456789ABCDEF);
430 __m128i vout5x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc5x0123456789ABCDEF);
431 __m128i vout6x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc6x0123456789ABCDEF);
432 __m128i vout7x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc7x0123456789ABCDEF);
433 __m128i vout8x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc8x0123456789ABCDEF);
434 __m128i vout9x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc9x0123456789ABCDEF);
435 __m128i vout10x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc10x0123456789ABCDEF);
436 __m128i vout11x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc11x0123456789ABCDEF);
437 __m128i vout12x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc12x0123456789ABCDEF);
438 __m128i vout13x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc13x0123456789ABCDEF);
439
440 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
441 vout1x0123456789ABCDEF = _mm_max_epi8(vout1x0123456789ABCDEF, voutput_min);
442 vout2x0123456789ABCDEF = _mm_max_epi8(vout2x0123456789ABCDEF, voutput_min);
443 vout3x0123456789ABCDEF = _mm_max_epi8(vout3x0123456789ABCDEF, voutput_min);
444 vout4x0123456789ABCDEF = _mm_max_epi8(vout4x0123456789ABCDEF, voutput_min);
445 vout5x0123456789ABCDEF = _mm_max_epi8(vout5x0123456789ABCDEF, voutput_min);
446 vout6x0123456789ABCDEF = _mm_max_epi8(vout6x0123456789ABCDEF, voutput_min);
447 vout7x0123456789ABCDEF = _mm_max_epi8(vout7x0123456789ABCDEF, voutput_min);
448 vout8x0123456789ABCDEF = _mm_max_epi8(vout8x0123456789ABCDEF, voutput_min);
449 vout9x0123456789ABCDEF = _mm_max_epi8(vout9x0123456789ABCDEF, voutput_min);
450 vout10x0123456789ABCDEF = _mm_max_epi8(vout10x0123456789ABCDEF, voutput_min);
451 vout11x0123456789ABCDEF = _mm_max_epi8(vout11x0123456789ABCDEF, voutput_min);
452 vout12x0123456789ABCDEF = _mm_max_epi8(vout12x0123456789ABCDEF, voutput_min);
453 vout13x0123456789ABCDEF = _mm_max_epi8(vout13x0123456789ABCDEF, voutput_min);
454
455 if (nc >= 16) {
456 _mm_storeu_si128((__m128i*) c13, vout13x0123456789ABCDEF);
457 c13 = (int8_t*) ((uintptr_t) c13 + cn_stride);
458 _mm_storeu_si128((__m128i*) c12, vout12x0123456789ABCDEF);
459 c12 = (int8_t*) ((uintptr_t) c12 + cn_stride);
460 _mm_storeu_si128((__m128i*) c11, vout11x0123456789ABCDEF);
461 c11 = (int8_t*) ((uintptr_t) c11 + cn_stride);
462 _mm_storeu_si128((__m128i*) c10, vout10x0123456789ABCDEF);
463 c10 = (int8_t*) ((uintptr_t) c10 + cn_stride);
464 _mm_storeu_si128((__m128i*) c9, vout9x0123456789ABCDEF);
465 c9 = (int8_t*) ((uintptr_t) c9 + cn_stride);
466 _mm_storeu_si128((__m128i*) c8, vout8x0123456789ABCDEF);
467 c8 = (int8_t*) ((uintptr_t) c8 + cn_stride);
468 _mm_storeu_si128((__m128i*) c7, vout7x0123456789ABCDEF);
469 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
470 _mm_storeu_si128((__m128i*) c6, vout6x0123456789ABCDEF);
471 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
472 _mm_storeu_si128((__m128i*) c5, vout5x0123456789ABCDEF);
473 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
474 _mm_storeu_si128((__m128i*) c4, vout4x0123456789ABCDEF);
475 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
476 _mm_storeu_si128((__m128i*) c3, vout3x0123456789ABCDEF);
477 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
478 _mm_storeu_si128((__m128i*) c2, vout2x0123456789ABCDEF);
479 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
480 _mm_storeu_si128((__m128i*) c1, vout1x0123456789ABCDEF);
481 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
482 _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);
483 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
484 a = (const int8_t**restrict) ((uintptr_t) a - ks);
485 nc -= 16;
486 } else {
487 // Prepare mask for valid 8-bit elements (depends on nc).
488 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
489 _mm_mask_storeu_epi8(c13, vmask, vout13x0123456789ABCDEF);
490 _mm_mask_storeu_epi8(c12, vmask, vout12x0123456789ABCDEF);
491 _mm_mask_storeu_epi8(c11, vmask, vout11x0123456789ABCDEF);
492 _mm_mask_storeu_epi8(c10, vmask, vout10x0123456789ABCDEF);
493 _mm_mask_storeu_epi8(c9, vmask, vout9x0123456789ABCDEF);
494 _mm_mask_storeu_epi8(c8, vmask, vout8x0123456789ABCDEF);
495 _mm_mask_storeu_epi8(c7, vmask, vout7x0123456789ABCDEF);
496 _mm_mask_storeu_epi8(c6, vmask, vout6x0123456789ABCDEF);
497 _mm_mask_storeu_epi8(c5, vmask, vout5x0123456789ABCDEF);
498 _mm_mask_storeu_epi8(c4, vmask, vout4x0123456789ABCDEF);
499 _mm_mask_storeu_epi8(c3, vmask, vout3x0123456789ABCDEF);
500 _mm_mask_storeu_epi8(c2, vmask, vout2x0123456789ABCDEF);
501 _mm_mask_storeu_epi8(c1, vmask, vout1x0123456789ABCDEF);
502 _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);
503 nc = 0;
504 }
505 } while (nc != 0);
506 }