Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-9x16c4-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_9x16c4__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 <= 9);
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
79 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
80 XNN_FORCE_REALIZATION(vsign_mask);
81 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);
82 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
83 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
84 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
85 // XNN_FORCE_REALIZATION(voutput_zero_point);
86 // XNN_FORCE_REALIZATION(voutput_min);
87 do {
88 __m512i vacc0x0123456789ABCDEF = _mm512_load_epi32(w);
89 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
90 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
91 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
92 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
93 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
94 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
95 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
96 __m512i vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
97 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
98 __m512i vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF;
99 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
100 __m512i vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF;
101 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
102 __m512i vacc7x0123456789ABCDEF = vacc0x0123456789ABCDEF;
103 __m512i vacc1x7x0123456789ABCDEF = _mm512_setzero_epi32();
104 __m512i vacc8x0123456789ABCDEF = vacc0x0123456789ABCDEF;
105 __m512i vacc1x8x0123456789ABCDEF = _mm512_setzero_epi32();
106 w = (const int32_t*) w + 16;
107
108 size_t p = ks;
109 do {
110 const int8_t* restrict a0 = a[0];
111 if XNN_UNPREDICTABLE(a0 != zero) {
112 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
113 }
114 const int8_t* restrict a1 = a[1];
115 if XNN_UNPREDICTABLE(a1 != zero) {
116 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
117 }
118 const int8_t* restrict a2 = a[2];
119 if XNN_UNPREDICTABLE(a2 != zero) {
120 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
121 }
122 const int8_t* restrict a3 = a[3];
123 if XNN_UNPREDICTABLE(a3 != zero) {
124 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
125 }
126 const int8_t* restrict a4 = a[4];
127 if XNN_UNPREDICTABLE(a4 != zero) {
128 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
129 }
130 const int8_t* restrict a5 = a[5];
131 if XNN_UNPREDICTABLE(a5 != zero) {
132 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
133 }
134 const int8_t* restrict a6 = a[6];
135 if XNN_UNPREDICTABLE(a6 != zero) {
136 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
137 }
138 const int8_t* restrict a7 = a[7];
139 if XNN_UNPREDICTABLE(a7 != zero) {
140 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
141 }
142 const int8_t* restrict a8 = a[8];
143 if XNN_UNPREDICTABLE(a8 != zero) {
144 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
145 }
146 a += 9;
147
148 size_t k = kc;
149 while (k >= 8 * sizeof(int8_t)) {
150 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
151 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
152 a0 += 8;
153 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
154 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
155 a1 += 8;
156 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
157 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
158 a2 += 8;
159 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
160 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
161 a3 += 8;
162 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
163 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
164 a4 += 8;
165 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
166 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
167 a5 += 8;
168 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
169 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
170 a6 += 8;
171 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
172 const __m512i va7x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7 + 4)), vsign_mask);
173 a7 += 8;
174 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
175 const __m512i va8x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8 + 4)), vsign_mask);
176 a8 += 8;
177
178 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
179 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
180
181 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
182 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
183 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
184 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
185 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
186 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
187 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
188 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEFx0123);
189 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEFx0123);
190 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
191 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
192 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
193 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
194 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
195 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
196 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
197 vacc1x7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x7x0123456789ABCDEF, va7x4567, vb0123456789ABCDEFx4567);
198 vacc1x8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x8x0123456789ABCDEF, va8x4567, vb0123456789ABCDEFx4567);
199
200 w = (const int8_t*) w + 128;
201 k -= 8 * sizeof(int8_t);
202 }
203
204 if (k != 0) {
205 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
206 a0 += 4;
207 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
208 a1 += 4;
209 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
210 a2 += 4;
211 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
212 a3 += 4;
213 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
214 a4 += 4;
215 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
216 a5 += 4;
217 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
218 a6 += 4;
219 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
220 a7 += 4;
221 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
222 a8 += 4;
223
224 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
225
226 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
227 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
228 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
229 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
230 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
231 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
232 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
233 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEF);
234 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEF);
235
236 w = (const int8_t*) w + 64;
237 k -= 4 * sizeof(int8_t);
238 }
239
240 p -= 9 * sizeof(void*);
241 } while (p != 0);
242
243 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
244 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
245 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
246 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
247 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
248 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
249 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
250 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, vacc1x7x0123456789ABCDEF);
251 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, vacc1x8x0123456789ABCDEF);
252
253 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
254 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
255 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
256 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
257 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
258 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
259 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
260 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
261 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
262
263 const __m512 vscale012345678ABCDEF = _mm512_load_ps(w);
264 w = (const float*) w + 16;
265 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale012345678ABCDEF);
266 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vscale012345678ABCDEF);
267 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vscale012345678ABCDEF);
268 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vscale012345678ABCDEF);
269 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vscale012345678ABCDEF);
270 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vscale012345678ABCDEF);
271 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vscale012345678ABCDEF);
272 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, vscale012345678ABCDEF);
273 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, vscale012345678ABCDEF);
274
275 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
276 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max_less_zero_point);
277 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max_less_zero_point);
278 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max_less_zero_point);
279 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max_less_zero_point);
280 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max_less_zero_point);
281 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max_less_zero_point);
282 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max_less_zero_point);
283 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max_less_zero_point);
284
285 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
286 vacc1x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled1x0123456789ABCDEF);
287 vacc2x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled2x0123456789ABCDEF);
288 vacc3x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled3x0123456789ABCDEF);
289 vacc4x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled4x0123456789ABCDEF);
290 vacc5x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled5x0123456789ABCDEF);
291 vacc6x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled6x0123456789ABCDEF);
292 vacc7x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled7x0123456789ABCDEF);
293 vacc8x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled8x0123456789ABCDEF);
294
295 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
296 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, voutput_zero_point);
297 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, voutput_zero_point);
298 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, voutput_zero_point);
299 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, voutput_zero_point);
300 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, voutput_zero_point);
301 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, voutput_zero_point);
302 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, voutput_zero_point);
303 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, voutput_zero_point);
304
305 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
306 __m128i vout1x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc1x0123456789ABCDEF);
307 __m128i vout2x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc2x0123456789ABCDEF);
308 __m128i vout3x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc3x0123456789ABCDEF);
309 __m128i vout4x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc4x0123456789ABCDEF);
310 __m128i vout5x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc5x0123456789ABCDEF);
311 __m128i vout6x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc6x0123456789ABCDEF);
312 __m128i vout7x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc7x0123456789ABCDEF);
313 __m128i vout8x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc8x0123456789ABCDEF);
314
315 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
316 vout1x0123456789ABCDEF = _mm_max_epi8(vout1x0123456789ABCDEF, voutput_min);
317 vout2x0123456789ABCDEF = _mm_max_epi8(vout2x0123456789ABCDEF, voutput_min);
318 vout3x0123456789ABCDEF = _mm_max_epi8(vout3x0123456789ABCDEF, voutput_min);
319 vout4x0123456789ABCDEF = _mm_max_epi8(vout4x0123456789ABCDEF, voutput_min);
320 vout5x0123456789ABCDEF = _mm_max_epi8(vout5x0123456789ABCDEF, voutput_min);
321 vout6x0123456789ABCDEF = _mm_max_epi8(vout6x0123456789ABCDEF, voutput_min);
322 vout7x0123456789ABCDEF = _mm_max_epi8(vout7x0123456789ABCDEF, voutput_min);
323 vout8x0123456789ABCDEF = _mm_max_epi8(vout8x0123456789ABCDEF, voutput_min);
324
325 if (nc >= 16) {
326 _mm_storeu_si128((__m128i*) c8, vout8x0123456789ABCDEF);
327 c8 = (int8_t*) ((uintptr_t) c8 + cn_stride);
328 _mm_storeu_si128((__m128i*) c7, vout7x0123456789ABCDEF);
329 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
330 _mm_storeu_si128((__m128i*) c6, vout6x0123456789ABCDEF);
331 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
332 _mm_storeu_si128((__m128i*) c5, vout5x0123456789ABCDEF);
333 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
334 _mm_storeu_si128((__m128i*) c4, vout4x0123456789ABCDEF);
335 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
336 _mm_storeu_si128((__m128i*) c3, vout3x0123456789ABCDEF);
337 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
338 _mm_storeu_si128((__m128i*) c2, vout2x0123456789ABCDEF);
339 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
340 _mm_storeu_si128((__m128i*) c1, vout1x0123456789ABCDEF);
341 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
342 _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);
343 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
344 a = (const int8_t**restrict) ((uintptr_t) a - ks);
345 nc -= 16;
346 } else {
347 // Prepare mask for valid 8-bit elements (depends on nc).
348 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
349 _mm_mask_storeu_epi8(c8, vmask, vout8x0123456789ABCDEF);
350 _mm_mask_storeu_epi8(c7, vmask, vout7x0123456789ABCDEF);
351 _mm_mask_storeu_epi8(c6, vmask, vout6x0123456789ABCDEF);
352 _mm_mask_storeu_epi8(c5, vmask, vout5x0123456789ABCDEF);
353 _mm_mask_storeu_epi8(c4, vmask, vout4x0123456789ABCDEF);
354 _mm_mask_storeu_epi8(c3, vmask, vout3x0123456789ABCDEF);
355 _mm_mask_storeu_epi8(c2, vmask, vout2x0123456789ABCDEF);
356 _mm_mask_storeu_epi8(c1, vmask, vout1x0123456789ABCDEF);
357 _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);
358 nc = 0;
359 }
360 } while (nc != 0);
361 }