Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-12x16c4-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_12x16c4__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 <= 12);
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
91 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
92 XNN_FORCE_REALIZATION(vsign_mask);
93 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);
94 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
95 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
96 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
97 // XNN_FORCE_REALIZATION(voutput_zero_point);
98 // XNN_FORCE_REALIZATION(voutput_min);
99 do {
100 __m512i vacc0x0123456789ABCDEF = _mm512_load_epi32(w);
101 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
102 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
103 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
104 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
105 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
106 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
107 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
108 __m512i vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
109 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
110 __m512i vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF;
111 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
112 __m512i vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF;
113 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
114 __m512i vacc7x0123456789ABCDEF = vacc0x0123456789ABCDEF;
115 __m512i vacc1x7x0123456789ABCDEF = _mm512_setzero_epi32();
116 __m512i vacc8x0123456789ABCDEF = vacc0x0123456789ABCDEF;
117 __m512i vacc1x8x0123456789ABCDEF = _mm512_setzero_epi32();
118 __m512i vacc9x0123456789ABCDEF = vacc0x0123456789ABCDEF;
119 __m512i vacc1x9x0123456789ABCDEF = _mm512_setzero_epi32();
120 __m512i vacc10x0123456789ABCDEF = vacc0x0123456789ABCDEF;
121 __m512i vacc1x10x0123456789ABCDEF = _mm512_setzero_epi32();
122 __m512i vacc11x0123456789ABCDEF = vacc0x0123456789ABCDEF;
123 __m512i vacc1x11x0123456789ABCDEF = _mm512_setzero_epi32();
124 w = (const int32_t*) w + 16;
125
126 size_t p = ks;
127 do {
128 const int8_t* restrict a0 = a[0];
129 if XNN_UNPREDICTABLE(a0 != zero) {
130 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
131 }
132 const int8_t* restrict a1 = a[1];
133 if XNN_UNPREDICTABLE(a1 != zero) {
134 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
135 }
136 const int8_t* restrict a2 = a[2];
137 if XNN_UNPREDICTABLE(a2 != zero) {
138 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
139 }
140 const int8_t* restrict a3 = a[3];
141 if XNN_UNPREDICTABLE(a3 != zero) {
142 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
143 }
144 const int8_t* restrict a4 = a[4];
145 if XNN_UNPREDICTABLE(a4 != zero) {
146 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
147 }
148 const int8_t* restrict a5 = a[5];
149 if XNN_UNPREDICTABLE(a5 != zero) {
150 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
151 }
152 const int8_t* restrict a6 = a[6];
153 if XNN_UNPREDICTABLE(a6 != zero) {
154 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
155 }
156 const int8_t* restrict a7 = a[7];
157 if XNN_UNPREDICTABLE(a7 != zero) {
158 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
159 }
160 const int8_t* restrict a8 = a[8];
161 if XNN_UNPREDICTABLE(a8 != zero) {
162 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
163 }
164 const int8_t* restrict a9 = a[9];
165 if XNN_UNPREDICTABLE(a9 != zero) {
166 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
167 }
168 const int8_t* restrict a10 = a[10];
169 if XNN_UNPREDICTABLE(a10 != zero) {
170 a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
171 }
172 const int8_t* restrict a11 = a[11];
173 if XNN_UNPREDICTABLE(a11 != zero) {
174 a11 = (const int8_t*) ((uintptr_t) a11 + a_offset);
175 }
176 a += 12;
177
178 size_t k = kc;
179 while (k >= 8 * sizeof(int8_t)) {
180 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
181 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
182 a0 += 8;
183 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
184 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
185 a1 += 8;
186 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
187 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
188 a2 += 8;
189 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
190 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
191 a3 += 8;
192 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
193 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
194 a4 += 8;
195 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
196 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
197 a5 += 8;
198 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
199 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
200 a6 += 8;
201 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
202 const __m512i va7x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7 + 4)), vsign_mask);
203 a7 += 8;
204 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
205 const __m512i va8x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8 + 4)), vsign_mask);
206 a8 += 8;
207 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
208 const __m512i va9x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9 + 4)), vsign_mask);
209 a9 += 8;
210 const __m512i va10x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10)), vsign_mask);
211 const __m512i va10x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10 + 4)), vsign_mask);
212 a10 += 8;
213 const __m512i va11x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11)), vsign_mask);
214 const __m512i va11x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11 + 4)), vsign_mask);
215 a11 += 8;
216
217 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
218 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
219
220 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
221 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
222 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
223 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
224 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
225 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
226 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
227 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEFx0123);
228 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEFx0123);
229 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEFx0123);
230 vacc10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc10x0123456789ABCDEF, va10x0123, vb0123456789ABCDEFx0123);
231 vacc11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc11x0123456789ABCDEF, va11x0123, vb0123456789ABCDEFx0123);
232 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
233 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
234 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
235 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
236 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
237 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
238 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
239 vacc1x7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x7x0123456789ABCDEF, va7x4567, vb0123456789ABCDEFx4567);
240 vacc1x8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x8x0123456789ABCDEF, va8x4567, vb0123456789ABCDEFx4567);
241 vacc1x9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x9x0123456789ABCDEF, va9x4567, vb0123456789ABCDEFx4567);
242 vacc1x10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x10x0123456789ABCDEF, va10x4567, vb0123456789ABCDEFx4567);
243 vacc1x11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x11x0123456789ABCDEF, va11x4567, vb0123456789ABCDEFx4567);
244
245 w = (const int8_t*) w + 128;
246 k -= 8 * sizeof(int8_t);
247 }
248
249 if (k != 0) {
250 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
251 a0 += 4;
252 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
253 a1 += 4;
254 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
255 a2 += 4;
256 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
257 a3 += 4;
258 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
259 a4 += 4;
260 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
261 a5 += 4;
262 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
263 a6 += 4;
264 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
265 a7 += 4;
266 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
267 a8 += 4;
268 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
269 a9 += 4;
270 const __m512i va10x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10)), vsign_mask);
271 a10 += 4;
272 const __m512i va11x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11)), vsign_mask);
273 a11 += 4;
274
275 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
276
277 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
278 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
279 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
280 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
281 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
282 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
283 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
284 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEF);
285 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEF);
286 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEF);
287 vacc10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc10x0123456789ABCDEF, va10x0123, vb0123456789ABCDEF);
288 vacc11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc11x0123456789ABCDEF, va11x0123, vb0123456789ABCDEF);
289
290 w = (const int8_t*) w + 64;
291 k -= 4 * sizeof(int8_t);
292 }
293
294 p -= 12 * sizeof(void*);
295 } while (p != 0);
296
297 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
298 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
299 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
300 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
301 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
302 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
303 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
304 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, vacc1x7x0123456789ABCDEF);
305 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, vacc1x8x0123456789ABCDEF);
306 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, vacc1x9x0123456789ABCDEF);
307 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, vacc1x10x0123456789ABCDEF);
308 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, vacc1x11x0123456789ABCDEF);
309
310 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
311 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
312 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
313 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
314 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
315 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
316 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
317 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
318 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
319 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
320 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
321 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
322
323 const __m512 vscale012345678ABCDEF = _mm512_load_ps(w);
324 w = (const float*) w + 16;
325 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale012345678ABCDEF);
326 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vscale012345678ABCDEF);
327 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vscale012345678ABCDEF);
328 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vscale012345678ABCDEF);
329 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vscale012345678ABCDEF);
330 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vscale012345678ABCDEF);
331 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vscale012345678ABCDEF);
332 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, vscale012345678ABCDEF);
333 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, vscale012345678ABCDEF);
334 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, vscale012345678ABCDEF);
335 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, vscale012345678ABCDEF);
336 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, vscale012345678ABCDEF);
337
338 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
339 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max_less_zero_point);
340 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max_less_zero_point);
341 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max_less_zero_point);
342 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max_less_zero_point);
343 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max_less_zero_point);
344 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max_less_zero_point);
345 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max_less_zero_point);
346 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max_less_zero_point);
347 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max_less_zero_point);
348 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max_less_zero_point);
349 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max_less_zero_point);
350
351 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
352 vacc1x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled1x0123456789ABCDEF);
353 vacc2x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled2x0123456789ABCDEF);
354 vacc3x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled3x0123456789ABCDEF);
355 vacc4x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled4x0123456789ABCDEF);
356 vacc5x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled5x0123456789ABCDEF);
357 vacc6x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled6x0123456789ABCDEF);
358 vacc7x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled7x0123456789ABCDEF);
359 vacc8x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled8x0123456789ABCDEF);
360 vacc9x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled9x0123456789ABCDEF);
361 vacc10x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled10x0123456789ABCDEF);
362 vacc11x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled11x0123456789ABCDEF);
363
364 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
365 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, voutput_zero_point);
366 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, voutput_zero_point);
367 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, voutput_zero_point);
368 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, voutput_zero_point);
369 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, voutput_zero_point);
370 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, voutput_zero_point);
371 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, voutput_zero_point);
372 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, voutput_zero_point);
373 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, voutput_zero_point);
374 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, voutput_zero_point);
375 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, voutput_zero_point);
376
377 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
378 __m128i vout1x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc1x0123456789ABCDEF);
379 __m128i vout2x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc2x0123456789ABCDEF);
380 __m128i vout3x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc3x0123456789ABCDEF);
381 __m128i vout4x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc4x0123456789ABCDEF);
382 __m128i vout5x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc5x0123456789ABCDEF);
383 __m128i vout6x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc6x0123456789ABCDEF);
384 __m128i vout7x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc7x0123456789ABCDEF);
385 __m128i vout8x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc8x0123456789ABCDEF);
386 __m128i vout9x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc9x0123456789ABCDEF);
387 __m128i vout10x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc10x0123456789ABCDEF);
388 __m128i vout11x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc11x0123456789ABCDEF);
389
390 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
391 vout1x0123456789ABCDEF = _mm_max_epi8(vout1x0123456789ABCDEF, voutput_min);
392 vout2x0123456789ABCDEF = _mm_max_epi8(vout2x0123456789ABCDEF, voutput_min);
393 vout3x0123456789ABCDEF = _mm_max_epi8(vout3x0123456789ABCDEF, voutput_min);
394 vout4x0123456789ABCDEF = _mm_max_epi8(vout4x0123456789ABCDEF, voutput_min);
395 vout5x0123456789ABCDEF = _mm_max_epi8(vout5x0123456789ABCDEF, voutput_min);
396 vout6x0123456789ABCDEF = _mm_max_epi8(vout6x0123456789ABCDEF, voutput_min);
397 vout7x0123456789ABCDEF = _mm_max_epi8(vout7x0123456789ABCDEF, voutput_min);
398 vout8x0123456789ABCDEF = _mm_max_epi8(vout8x0123456789ABCDEF, voutput_min);
399 vout9x0123456789ABCDEF = _mm_max_epi8(vout9x0123456789ABCDEF, voutput_min);
400 vout10x0123456789ABCDEF = _mm_max_epi8(vout10x0123456789ABCDEF, voutput_min);
401 vout11x0123456789ABCDEF = _mm_max_epi8(vout11x0123456789ABCDEF, voutput_min);
402
403 if (nc >= 16) {
404 _mm_storeu_si128((__m128i*) c11, vout11x0123456789ABCDEF);
405 c11 = (int8_t*) ((uintptr_t) c11 + cn_stride);
406 _mm_storeu_si128((__m128i*) c10, vout10x0123456789ABCDEF);
407 c10 = (int8_t*) ((uintptr_t) c10 + cn_stride);
408 _mm_storeu_si128((__m128i*) c9, vout9x0123456789ABCDEF);
409 c9 = (int8_t*) ((uintptr_t) c9 + cn_stride);
410 _mm_storeu_si128((__m128i*) c8, vout8x0123456789ABCDEF);
411 c8 = (int8_t*) ((uintptr_t) c8 + cn_stride);
412 _mm_storeu_si128((__m128i*) c7, vout7x0123456789ABCDEF);
413 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
414 _mm_storeu_si128((__m128i*) c6, vout6x0123456789ABCDEF);
415 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
416 _mm_storeu_si128((__m128i*) c5, vout5x0123456789ABCDEF);
417 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
418 _mm_storeu_si128((__m128i*) c4, vout4x0123456789ABCDEF);
419 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
420 _mm_storeu_si128((__m128i*) c3, vout3x0123456789ABCDEF);
421 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
422 _mm_storeu_si128((__m128i*) c2, vout2x0123456789ABCDEF);
423 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
424 _mm_storeu_si128((__m128i*) c1, vout1x0123456789ABCDEF);
425 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
426 _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);
427 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
428 a = (const int8_t**restrict) ((uintptr_t) a - ks);
429 nc -= 16;
430 } else {
431 // Prepare mask for valid 8-bit elements (depends on nc).
432 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
433 _mm_mask_storeu_epi8(c11, vmask, vout11x0123456789ABCDEF);
434 _mm_mask_storeu_epi8(c10, vmask, vout10x0123456789ABCDEF);
435 _mm_mask_storeu_epi8(c9, vmask, vout9x0123456789ABCDEF);
436 _mm_mask_storeu_epi8(c8, vmask, vout8x0123456789ABCDEF);
437 _mm_mask_storeu_epi8(c7, vmask, vout7x0123456789ABCDEF);
438 _mm_mask_storeu_epi8(c6, vmask, vout6x0123456789ABCDEF);
439 _mm_mask_storeu_epi8(c5, vmask, vout5x0123456789ABCDEF);
440 _mm_mask_storeu_epi8(c4, vmask, vout4x0123456789ABCDEF);
441 _mm_mask_storeu_epi8(c3, vmask, vout3x0123456789ABCDEF);
442 _mm_mask_storeu_epi8(c2, vmask, vout2x0123456789ABCDEF);
443 _mm_mask_storeu_epi8(c1, vmask, vout1x0123456789ABCDEF);
444 _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);
445 nc = 0;
446 }
447 } while (nc != 0);
448 }