Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-8x16c4-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_8x16c4__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 <= 8);
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
76 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
77 XNN_FORCE_REALIZATION(vsign_mask);
78 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);
79 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
80 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
81 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
82 // XNN_FORCE_REALIZATION(voutput_zero_point);
83 // XNN_FORCE_REALIZATION(voutput_min);
84 do {
85 __m512i vacc0x0123456789ABCDEF = _mm512_load_epi32(w);
86 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
87 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
88 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
89 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
90 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
91 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
92 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
93 __m512i vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
94 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
95 __m512i vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF;
96 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
97 __m512i vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF;
98 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
99 __m512i vacc7x0123456789ABCDEF = vacc0x0123456789ABCDEF;
100 __m512i vacc1x7x0123456789ABCDEF = _mm512_setzero_epi32();
101 w = (const int32_t*) w + 16;
102
103 size_t p = ks;
104 do {
105 const int8_t* restrict a0 = a[0];
106 if XNN_UNPREDICTABLE(a0 != zero) {
107 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
108 }
109 const int8_t* restrict a1 = a[1];
110 if XNN_UNPREDICTABLE(a1 != zero) {
111 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
112 }
113 const int8_t* restrict a2 = a[2];
114 if XNN_UNPREDICTABLE(a2 != zero) {
115 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
116 }
117 const int8_t* restrict a3 = a[3];
118 if XNN_UNPREDICTABLE(a3 != zero) {
119 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
120 }
121 const int8_t* restrict a4 = a[4];
122 if XNN_UNPREDICTABLE(a4 != zero) {
123 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
124 }
125 const int8_t* restrict a5 = a[5];
126 if XNN_UNPREDICTABLE(a5 != zero) {
127 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
128 }
129 const int8_t* restrict a6 = a[6];
130 if XNN_UNPREDICTABLE(a6 != zero) {
131 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
132 }
133 const int8_t* restrict a7 = a[7];
134 if XNN_UNPREDICTABLE(a7 != zero) {
135 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
136 }
137 a += 8;
138
139 size_t k = kc;
140 while (k >= 8 * sizeof(int8_t)) {
141 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
142 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
143 a0 += 8;
144 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
145 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
146 a1 += 8;
147 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
148 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
149 a2 += 8;
150 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
151 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
152 a3 += 8;
153 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
154 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
155 a4 += 8;
156 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
157 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
158 a5 += 8;
159 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
160 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
161 a6 += 8;
162 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
163 const __m512i va7x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7 + 4)), vsign_mask);
164 a7 += 8;
165
166 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
167 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
168 xnn_prefetch_to_l1((const int8_t*) w + 896);
169
170 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
171 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
172 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
173 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
174 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
175 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
176 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
177 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEFx0123);
178 xnn_prefetch_to_l1((const int8_t*) w + 960);
179 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
180 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
181 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
182 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
183 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
184 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
185 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
186 vacc1x7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x7x0123456789ABCDEF, va7x4567, vb0123456789ABCDEFx4567);
187
188 w = (const int8_t*) w + 128;
189 k -= 8 * sizeof(int8_t);
190 }
191
192 if (k != 0) {
193 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
194 a0 += 4;
195 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
196 a1 += 4;
197 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
198 a2 += 4;
199 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
200 a3 += 4;
201 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
202 a4 += 4;
203 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
204 a5 += 4;
205 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
206 a6 += 4;
207 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
208 a7 += 4;
209
210 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
211 xnn_prefetch_to_l1((const int8_t*) w + 960);
212
213 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
214 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
215 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
216 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
217 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
218 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
219 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
220 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEF);
221
222 w = (const int8_t*) w + 64;
223 k -= 4 * sizeof(int8_t);
224 }
225
226 p -= 8 * sizeof(void*);
227 } while (p != 0);
228
229 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
230 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
231 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
232 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
233 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
234 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
235 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
236 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, vacc1x7x0123456789ABCDEF);
237
238 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
239 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
240 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
241 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
242 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
243 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
244 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
245 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
246
247 const __m512 vscale012345678ABCDEF = _mm512_load_ps(w);
248 w = (const float*) w + 16;
249 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale012345678ABCDEF);
250 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vscale012345678ABCDEF);
251 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vscale012345678ABCDEF);
252 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vscale012345678ABCDEF);
253 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vscale012345678ABCDEF);
254 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vscale012345678ABCDEF);
255 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vscale012345678ABCDEF);
256 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, vscale012345678ABCDEF);
257
258 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
259 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max_less_zero_point);
260 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max_less_zero_point);
261 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max_less_zero_point);
262 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max_less_zero_point);
263 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max_less_zero_point);
264 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max_less_zero_point);
265 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max_less_zero_point);
266
267 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
268 vacc1x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled1x0123456789ABCDEF);
269 vacc2x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled2x0123456789ABCDEF);
270 vacc3x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled3x0123456789ABCDEF);
271 vacc4x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled4x0123456789ABCDEF);
272 vacc5x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled5x0123456789ABCDEF);
273 vacc6x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled6x0123456789ABCDEF);
274 vacc7x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled7x0123456789ABCDEF);
275
276 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
277 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, voutput_zero_point);
278 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, voutput_zero_point);
279 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, voutput_zero_point);
280 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, voutput_zero_point);
281 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, voutput_zero_point);
282 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, voutput_zero_point);
283 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, voutput_zero_point);
284
285 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
286 __m128i vout1x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc1x0123456789ABCDEF);
287 __m128i vout2x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc2x0123456789ABCDEF);
288 __m128i vout3x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc3x0123456789ABCDEF);
289 __m128i vout4x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc4x0123456789ABCDEF);
290 __m128i vout5x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc5x0123456789ABCDEF);
291 __m128i vout6x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc6x0123456789ABCDEF);
292 __m128i vout7x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc7x0123456789ABCDEF);
293
294 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
295 vout1x0123456789ABCDEF = _mm_max_epi8(vout1x0123456789ABCDEF, voutput_min);
296 vout2x0123456789ABCDEF = _mm_max_epi8(vout2x0123456789ABCDEF, voutput_min);
297 vout3x0123456789ABCDEF = _mm_max_epi8(vout3x0123456789ABCDEF, voutput_min);
298 vout4x0123456789ABCDEF = _mm_max_epi8(vout4x0123456789ABCDEF, voutput_min);
299 vout5x0123456789ABCDEF = _mm_max_epi8(vout5x0123456789ABCDEF, voutput_min);
300 vout6x0123456789ABCDEF = _mm_max_epi8(vout6x0123456789ABCDEF, voutput_min);
301 vout7x0123456789ABCDEF = _mm_max_epi8(vout7x0123456789ABCDEF, voutput_min);
302
303 if (nc >= 16) {
304 _mm_storeu_si128((__m128i*) c7, vout7x0123456789ABCDEF);
305 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
306 _mm_storeu_si128((__m128i*) c6, vout6x0123456789ABCDEF);
307 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
308 _mm_storeu_si128((__m128i*) c5, vout5x0123456789ABCDEF);
309 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
310 _mm_storeu_si128((__m128i*) c4, vout4x0123456789ABCDEF);
311 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
312 _mm_storeu_si128((__m128i*) c3, vout3x0123456789ABCDEF);
313 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
314 _mm_storeu_si128((__m128i*) c2, vout2x0123456789ABCDEF);
315 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
316 _mm_storeu_si128((__m128i*) c1, vout1x0123456789ABCDEF);
317 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
318 _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);
319 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
320 a = (const int8_t**restrict) ((uintptr_t) a - ks);
321 nc -= 16;
322 } else {
323 // Prepare mask for valid 8-bit elements (depends on nc).
324 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
325 _mm_mask_storeu_epi8(c7, vmask, vout7x0123456789ABCDEF);
326 _mm_mask_storeu_epi8(c6, vmask, vout6x0123456789ABCDEF);
327 _mm_mask_storeu_epi8(c5, vmask, vout5x0123456789ABCDEF);
328 _mm_mask_storeu_epi8(c4, vmask, vout4x0123456789ABCDEF);
329 _mm_mask_storeu_epi8(c3, vmask, vout3x0123456789ABCDEF);
330 _mm_mask_storeu_epi8(c2, vmask, vout2x0123456789ABCDEF);
331 _mm_mask_storeu_epi8(c1, vmask, vout1x0123456789ABCDEF);
332 _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);
333 nc = 0;
334 }
335 } while (nc != 0);
336 }