Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-8x16c8-minmax-fp32-avx512skx-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx16c8-avx512skx.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/igemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/prefetch.h"
19
20
21 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_8x16c8__avx512skx_prfm(
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 <= 8);
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, 8 * 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
75 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
76 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);
77 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
78 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
79 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
80 // XNN_FORCE_REALIZATION(voutput_zero_point);
81 // XNN_FORCE_REALIZATION(voutput_min);
82 do {
83 __m512i vacc0x0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
84 __m512i vacc0x4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
85 __m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
86 __m512i vacc0xCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
87 __m512i vacc1x0123 = vacc0x0123;
88 __m512i vacc1x4567 = vacc0x4567;
89 __m512i vacc1x89AB = vacc0x89AB;
90 __m512i vacc1xCDEF = vacc0xCDEF;
91 __m512i vacc2x0123 = vacc0x0123;
92 __m512i vacc2x4567 = vacc0x4567;
93 __m512i vacc2x89AB = vacc0x89AB;
94 __m512i vacc2xCDEF = vacc0xCDEF;
95 __m512i vacc3x0123 = vacc0x0123;
96 __m512i vacc3x4567 = vacc0x4567;
97 __m512i vacc3x89AB = vacc0x89AB;
98 __m512i vacc3xCDEF = vacc0xCDEF;
99 __m512i vacc4x0123 = vacc0x0123;
100 __m512i vacc4x4567 = vacc0x4567;
101 __m512i vacc4x89AB = vacc0x89AB;
102 __m512i vacc4xCDEF = vacc0xCDEF;
103 __m512i vacc5x0123 = vacc0x0123;
104 __m512i vacc5x4567 = vacc0x4567;
105 __m512i vacc5x89AB = vacc0x89AB;
106 __m512i vacc5xCDEF = vacc0xCDEF;
107 __m512i vacc6x0123 = vacc0x0123;
108 __m512i vacc6x4567 = vacc0x4567;
109 __m512i vacc6x89AB = vacc0x89AB;
110 __m512i vacc6xCDEF = vacc0xCDEF;
111 __m512i vacc7x0123 = vacc0x0123;
112 __m512i vacc7x4567 = vacc0x4567;
113 __m512i vacc7x89AB = vacc0x89AB;
114 __m512i vacc7xCDEF = vacc0xCDEF;
115 w = (const int32_t*) w + 16;
116
117 size_t p = ks;
118 do {
119 const int8_t* restrict a0 = a[0];
120 if XNN_UNPREDICTABLE(a0 != zero) {
121 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
122 }
123 const int8_t* restrict a1 = a[1];
124 if XNN_UNPREDICTABLE(a1 != zero) {
125 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
126 }
127 const int8_t* restrict a2 = a[2];
128 if XNN_UNPREDICTABLE(a2 != zero) {
129 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
130 }
131 const int8_t* restrict a3 = a[3];
132 if XNN_UNPREDICTABLE(a3 != zero) {
133 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
134 }
135 const int8_t* restrict a4 = a[4];
136 if XNN_UNPREDICTABLE(a4 != zero) {
137 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
138 }
139 const int8_t* restrict a5 = a[5];
140 if XNN_UNPREDICTABLE(a5 != zero) {
141 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
142 }
143 const int8_t* restrict a6 = a[6];
144 if XNN_UNPREDICTABLE(a6 != zero) {
145 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
146 }
147 const int8_t* restrict a7 = a[7];
148 if XNN_UNPREDICTABLE(a7 != zero) {
149 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
150 }
151 a += 8;
152
153 size_t k = kc;
154
155 while (k >= 8 * sizeof(int8_t)) {
156 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
157 a0 += 8;
158 const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
159 a1 += 8;
160 const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
161 a2 += 8;
162 const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
163 a3 += 8;
164 const __m512i va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
165 a4 += 8;
166 const __m512i va5 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a5)));
167 a5 += 8;
168 const __m512i va6 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a6)));
169 a6 += 8;
170 const __m512i va7 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a7)));
171 a7 += 8;
172
173 const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
174 xnn_prefetch_to_l1((const int8_t*) w + 896);
175
176 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
177 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
178 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
179 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
180 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
181 vacc5x0123 = _mm512_add_epi32(vacc5x0123, _mm512_madd_epi16(va5, vb0123));
182 vacc6x0123 = _mm512_add_epi32(vacc6x0123, _mm512_madd_epi16(va6, vb0123));
183 vacc7x0123 = _mm512_add_epi32(vacc7x0123, _mm512_madd_epi16(va7, vb0123));
184 const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)));
185
186 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
187 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
188 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
189 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
190 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
191 vacc5x4567 = _mm512_add_epi32(vacc5x4567, _mm512_madd_epi16(va5, vb4567));
192 vacc6x4567 = _mm512_add_epi32(vacc6x4567, _mm512_madd_epi16(va6, vb4567));
193 vacc7x4567 = _mm512_add_epi32(vacc7x4567, _mm512_madd_epi16(va7, vb4567));
194 const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)));
195 xnn_prefetch_to_l1((const int8_t*) w + 960);
196
197 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
198 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
199 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
200 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
201 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
202 vacc5x89AB = _mm512_add_epi32(vacc5x89AB, _mm512_madd_epi16(va5, vb89AB));
203 vacc6x89AB = _mm512_add_epi32(vacc6x89AB, _mm512_madd_epi16(va6, vb89AB));
204 vacc7x89AB = _mm512_add_epi32(vacc7x89AB, _mm512_madd_epi16(va7, vb89AB));
205 const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)));
206
207 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
208 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
209 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
210 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
211 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
212 vacc5xCDEF = _mm512_add_epi32(vacc5xCDEF, _mm512_madd_epi16(va5, vbCDEF));
213 vacc6xCDEF = _mm512_add_epi32(vacc6xCDEF, _mm512_madd_epi16(va6, vbCDEF));
214 vacc7xCDEF = _mm512_add_epi32(vacc7xCDEF, _mm512_madd_epi16(va7, vbCDEF));
215
216 w = (const int8_t*) w + 128;
217 k -= 8 * sizeof(int8_t);
218 }
219 p -= 8 * sizeof(void*);
220 } while (p != 0);
221
222 // Add 4 adjacent sums
223 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
224 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
225 const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
226 const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
227 const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm512_unpackhi_epi32(vacc2x0123, vacc2x4567));
228 const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm512_unpackhi_epi32(vacc2x89AB, vacc2xCDEF));
229 const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm512_unpackhi_epi32(vacc3x0123, vacc3x4567));
230 const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm512_unpackhi_epi32(vacc3x89AB, vacc3xCDEF));
231 const __m512i vacc4x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x0123, vacc4x4567), _mm512_unpackhi_epi32(vacc4x0123, vacc4x4567));
232 const __m512i vacc4x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x89AB, vacc4xCDEF), _mm512_unpackhi_epi32(vacc4x89AB, vacc4xCDEF));
233 const __m512i vacc5x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x0123, vacc5x4567), _mm512_unpackhi_epi32(vacc5x0123, vacc5x4567));
234 const __m512i vacc5x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x89AB, vacc5xCDEF), _mm512_unpackhi_epi32(vacc5x89AB, vacc5xCDEF));
235 const __m512i vacc6x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x0123, vacc6x4567), _mm512_unpackhi_epi32(vacc6x0123, vacc6x4567));
236 const __m512i vacc6x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x89AB, vacc6xCDEF), _mm512_unpackhi_epi32(vacc6x89AB, vacc6xCDEF));
237 const __m512i vacc7x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc7x0123, vacc7x4567), _mm512_unpackhi_epi32(vacc7x0123, vacc7x4567));
238 const __m512i vacc7x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc7x89AB, vacc7xCDEF), _mm512_unpackhi_epi32(vacc7x89AB, vacc7xCDEF));
239 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
240 const __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
241 const __m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9DAEBF), _mm512_unpackhi_epi32(vacc2x04152637, vacc2x8C9DAEBF));
242 const __m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9DAEBF), _mm512_unpackhi_epi32(vacc3x04152637, vacc3x8C9DAEBF));
243 const __m512i vacc4x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x04152637, vacc4x8C9DAEBF), _mm512_unpackhi_epi32(vacc4x04152637, vacc4x8C9DAEBF));
244 const __m512i vacc5x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x04152637, vacc5x8C9DAEBF), _mm512_unpackhi_epi32(vacc5x04152637, vacc5x8C9DAEBF));
245 const __m512i vacc6x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x04152637, vacc6x8C9DAEBF), _mm512_unpackhi_epi32(vacc6x04152637, vacc6x8C9DAEBF));
246 const __m512i vacc7x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc7x04152637, vacc7x8C9DAEBF), _mm512_unpackhi_epi32(vacc7x04152637, vacc7x8C9DAEBF));
247 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
248 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
249 __m512i vacc1x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc1x084C195D2A6E3B7F);
250 __m512i vacc2x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc2x084C195D2A6E3B7F);
251 __m512i vacc3x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc3x084C195D2A6E3B7F);
252 __m512i vacc4x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc4x084C195D2A6E3B7F);
253 __m512i vacc5x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc5x084C195D2A6E3B7F);
254 __m512i vacc6x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc6x084C195D2A6E3B7F);
255 __m512i vacc7x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc7x084C195D2A6E3B7F);
256
257 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
258 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
259 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
260 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
261 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
262 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
263 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
264 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
265
266 const __m512 vscale012345678ABCDEF = _mm512_load_ps(w);
267 w = (const float*) w + 16;
268 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale012345678ABCDEF);
269 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vscale012345678ABCDEF);
270 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vscale012345678ABCDEF);
271 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vscale012345678ABCDEF);
272 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vscale012345678ABCDEF);
273 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vscale012345678ABCDEF);
274 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vscale012345678ABCDEF);
275 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, vscale012345678ABCDEF);
276
277 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
278 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max_less_zero_point);
279 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max_less_zero_point);
280 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max_less_zero_point);
281 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max_less_zero_point);
282 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max_less_zero_point);
283 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max_less_zero_point);
284 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max_less_zero_point);
285
286 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
287 vacc1x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled1x0123456789ABCDEF);
288 vacc2x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled2x0123456789ABCDEF);
289 vacc3x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled3x0123456789ABCDEF);
290 vacc4x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled4x0123456789ABCDEF);
291 vacc5x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled5x0123456789ABCDEF);
292 vacc6x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled6x0123456789ABCDEF);
293 vacc7x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled7x0123456789ABCDEF);
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
304 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
305 __m128i vout1x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc1x0123456789ABCDEF);
306 __m128i vout2x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc2x0123456789ABCDEF);
307 __m128i vout3x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc3x0123456789ABCDEF);
308 __m128i vout4x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc4x0123456789ABCDEF);
309 __m128i vout5x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc5x0123456789ABCDEF);
310 __m128i vout6x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc6x0123456789ABCDEF);
311 __m128i vout7x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc7x0123456789ABCDEF);
312
313 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
314 vout1x0123456789ABCDEF = _mm_max_epi8(vout1x0123456789ABCDEF, voutput_min);
315 vout2x0123456789ABCDEF = _mm_max_epi8(vout2x0123456789ABCDEF, voutput_min);
316 vout3x0123456789ABCDEF = _mm_max_epi8(vout3x0123456789ABCDEF, voutput_min);
317 vout4x0123456789ABCDEF = _mm_max_epi8(vout4x0123456789ABCDEF, voutput_min);
318 vout5x0123456789ABCDEF = _mm_max_epi8(vout5x0123456789ABCDEF, voutput_min);
319 vout6x0123456789ABCDEF = _mm_max_epi8(vout6x0123456789ABCDEF, voutput_min);
320 vout7x0123456789ABCDEF = _mm_max_epi8(vout7x0123456789ABCDEF, voutput_min);
321
322 if (nc >= 16) {
323 _mm_storeu_si128((__m128i*) c7, vout7x0123456789ABCDEF);
324 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
325 _mm_storeu_si128((__m128i*) c6, vout6x0123456789ABCDEF);
326 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
327 _mm_storeu_si128((__m128i*) c5, vout5x0123456789ABCDEF);
328 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
329 _mm_storeu_si128((__m128i*) c4, vout4x0123456789ABCDEF);
330 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
331 _mm_storeu_si128((__m128i*) c3, vout3x0123456789ABCDEF);
332 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
333 _mm_storeu_si128((__m128i*) c2, vout2x0123456789ABCDEF);
334 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
335 _mm_storeu_si128((__m128i*) c1, vout1x0123456789ABCDEF);
336 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
337 _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);
338 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
339 a = (const int8_t**restrict) ((uintptr_t) a - ks);
340 nc -= 16;
341 } else {
342 // Prepare mask for valid 8-bit elements (depends on nc).
343 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
344 _mm_mask_storeu_epi8(c7, vmask, vout7x0123456789ABCDEF);
345 _mm_mask_storeu_epi8(c6, vmask, vout6x0123456789ABCDEF);
346 _mm_mask_storeu_epi8(c5, vmask, vout5x0123456789ABCDEF);
347 _mm_mask_storeu_epi8(c4, vmask, vout4x0123456789ABCDEF);
348 _mm_mask_storeu_epi8(c3, vmask, vout3x0123456789ABCDEF);
349 _mm_mask_storeu_epi8(c2, vmask, vout2x0123456789ABCDEF);
350 _mm_mask_storeu_epi8(c1, vmask, vout1x0123456789ABCDEF);
351 _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);
352 nc = 0;
353 }
354 } while (nc != 0);
355 }