Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-7x32c4-minmax-fp32-avx512amx.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-avx512amx.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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/gemm.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_7x32c4__avx512amx(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const int8_t** restrict a,
26 const void* restrict w,
27 int8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
33 {
34 assert(mr != 0);
35 assert(mr <= 7);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 // TODO: amxintrin.h only provide intrinsics for __x86_64__
44 // Update if amxintrin changes
45 #if defined(__x86_64__)
46 __attribute__((aligned(64))) int32_t vintile[7 * 16];
47 __attribute__((aligned(64))) int32_t res0[7 * 16];
48 __attribute__((aligned(64))) int32_t res1[7 * 16];
49
50 kc = round_up_po2(kc, 4 * sizeof(int8_t));
51 const size_t kremainder = (kc & 63) ? (kc & 63) : 64;
52 const __mmask16 kremainder_mask = _cvtu32_mask16((UINT32_C(1) << (kremainder >> 2)) - 1);
53
54 // Define tile config data structure
55 struct __tile_config {
56 uint8_t palette_id;
57 uint8_t start_row;
58 uint8_t reserved_0[14];
59 uint16_t colsb[8];
60 uint16_t reserved_1[8];
61 uint8_t rows[8];
62 uint8_t reserved_2[8];
63 };
64
65 // Load tile configuration
66 __attribute__((aligned(64))) struct __tile_config tile_data = {0};
67 tile_data.palette_id = 1;
68 tile_data.rows[0] = mr; // tmm0 = res 0
69 tile_data.rows[1] = mr; // tmm1 = res 1
70 tile_data.rows[2] = mr; // tmm2 = res 2
71 tile_data.rows[3] = mr; // tmm3 = res 3
72 tile_data.rows[4] = mr; // tmm4 = input
73 tile_data.rows[5] = 16; // tmm5 = weights
74 tile_data.rows[6] = mr; // tmm6 = input remainder
75 tile_data.rows[7] = kremainder >> 2; // tmm7 = weights remainder
76
77 tile_data.colsb[0] = 64; // tmm0 = res 0
78 tile_data.colsb[1] = 64; // tmm1 = res 1
79 tile_data.colsb[2] = 64; // tmm2 = res 1
80 tile_data.colsb[3] = 64; // tmm3 = res 1
81 tile_data.colsb[4] = 64; // tmm4 = input
82 tile_data.colsb[5] = 64; // tmm5 = weights
83 tile_data.colsb[6] = kremainder; // tmm6 = input remainder
84 tile_data.colsb[7] = 64; // tmm7 = weights remainder
85
86 //_tile_loadconfig(&tile_data);
87 __asm__ volatile ("ldtilecfg %0" :: "m" (tile_data));
88
89 int8_t* c0 = c;
90 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
91 if XNN_UNPREDICTABLE(mr < 2) {
92 c1 = c0;
93 }
94 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
95 if XNN_UNPREDICTABLE(mr <= 2) {
96 c2 = c1;
97 }
98 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
99 if XNN_UNPREDICTABLE(mr < 4) {
100 c3 = c2;
101 }
102 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
103 if XNN_UNPREDICTABLE(mr <= 4) {
104 c4 = c3;
105 }
106 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
107 if XNN_UNPREDICTABLE(mr < 6) {
108 c5 = c4;
109 }
110 int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
111 if XNN_UNPREDICTABLE(mr <= 6) {
112 c6 = c5;
113 }
114
115 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);
116 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
117 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
118 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
119 // XNN_FORCE_REALIZATION(voutput_zero_point);
120 // XNN_FORCE_REALIZATION(voutput_min);
121
122 do {
123 const __m512i vksum0123456789ABCDEF = _mm512_loadu_epi32((const int32_t*) w + 0);
124 const __m512i vksumGHIJKLMNOPQRSTUV = _mm512_loadu_epi32((const int32_t*) w + 16);
125 w = (const int32_t*) w + 32;
126
127 // Zero tile accumulator
128 __asm__ volatile (
129 "tilezero %%tmm0\n"
130 "tilezero %%tmm1\n"
131 ::);
132
133 size_t p = ks;
134 do {
135 const int8_t* restrict a0 = a[0];
136 if XNN_UNPREDICTABLE(a0 != zero) {
137 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
138 }
139 const int8_t* restrict a1 = a[1];
140 if XNN_UNPREDICTABLE(a1 != zero) {
141 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
142 }
143 const int8_t* restrict a2 = a[2];
144 if XNN_UNPREDICTABLE(a2 != zero) {
145 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
146 }
147 const int8_t* restrict a3 = a[3];
148 if XNN_UNPREDICTABLE(a3 != zero) {
149 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
150 }
151 const int8_t* restrict a4 = a[4];
152 if XNN_UNPREDICTABLE(a4 != zero) {
153 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
154 }
155 const int8_t* restrict a5 = a[5];
156 if XNN_UNPREDICTABLE(a5 != zero) {
157 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
158 }
159 const int8_t* restrict a6 = a[6];
160 if XNN_UNPREDICTABLE(a6 != zero) {
161 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
162 }
163 a += 7;
164
165 size_t k = kc;
166 if (mr == 1)
167 {
168 while (k >= 64 * sizeof(int8_t)) {
169 _tile_loadd(4, a0, 64); // Directly load input for mr=1
170 a6 += 64;
171 _tile_loadd(5, (const int8_t*) w + 0, 128);
172 _tile_dpbssd(0, 4, 5);
173 _tile_loadd(5, (const int8_t*) w + 64, 128);
174 _tile_dpbssd(1, 4, 5);
175
176 w = (const int8_t*) w + 2048;
177 k -= 64 * sizeof(int8_t);
178 }
179 }
180 else {
181 while (k >= 64 * sizeof(int8_t)) {
182 const __m512i vin0 = _mm512_loadu_epi32(a0);
183 a0 += 64;
184 _mm512_store_epi32(vintile + 0, vin0);
185 const __m512i vin1 = _mm512_loadu_epi32(a1);
186 a1 += 64;
187 _mm512_store_epi32(vintile + 16, vin1);
188 const __m512i vin2 = _mm512_loadu_epi32(a2);
189 a2 += 64;
190 _mm512_store_epi32(vintile + 32, vin2);
191 const __m512i vin3 = _mm512_loadu_epi32(a3);
192 a3 += 64;
193 _mm512_store_epi32(vintile + 48, vin3);
194 const __m512i vin4 = _mm512_loadu_epi32(a4);
195 a4 += 64;
196 _mm512_store_epi32(vintile + 64, vin4);
197 const __m512i vin5 = _mm512_loadu_epi32(a5);
198 a5 += 64;
199 _mm512_store_epi32(vintile + 80, vin5);
200 const __m512i vin6 = _mm512_loadu_epi32(a6);
201 a6 += 64;
202 _mm512_store_epi32(vintile + 96, vin6);
203 _tile_loadd(4, vintile, 64);
204 _tile_loadd(5, (const int8_t*) w + 0, 128);
205 _tile_dpbssd(0, 4, 5);
206 _tile_loadd(5, (const int8_t*) w + 64, 128);
207 _tile_dpbssd(1, 4, 5);
208
209 w = (const int8_t*) w + 2048;
210 k -= 64 * sizeof(int8_t);
211 }
212 }
213
214 if XNN_UNLIKELY(k != 0) {
215 const __m512i vin0 = _mm512_maskz_loadu_epi32(kremainder_mask, a0);
216 a0 += kremainder;
217 _mm512_store_epi32(vintile + 0, vin0);
218 const __m512i vin1 = _mm512_maskz_loadu_epi32(kremainder_mask, a1);
219 a1 += kremainder;
220 _mm512_store_epi32(vintile + 16, vin1);
221 const __m512i vin2 = _mm512_maskz_loadu_epi32(kremainder_mask, a2);
222 a2 += kremainder;
223 _mm512_store_epi32(vintile + 32, vin2);
224 const __m512i vin3 = _mm512_maskz_loadu_epi32(kremainder_mask, a3);
225 a3 += kremainder;
226 _mm512_store_epi32(vintile + 48, vin3);
227 const __m512i vin4 = _mm512_maskz_loadu_epi32(kremainder_mask, a4);
228 a4 += kremainder;
229 _mm512_store_epi32(vintile + 64, vin4);
230 const __m512i vin5 = _mm512_maskz_loadu_epi32(kremainder_mask, a5);
231 a5 += kremainder;
232 _mm512_store_epi32(vintile + 80, vin5);
233 const __m512i vin6 = _mm512_maskz_loadu_epi32(kremainder_mask, a6);
234 a6 += kremainder;
235 _mm512_store_epi32(vintile + 96, vin6);
236 _tile_loadd(6, vintile, 64);
237 _tile_loadd(7, (const int8_t*) w + 0, 128);
238 _tile_dpbssd(0, 6, 7);
239 _tile_loadd(7, (const int8_t*) w + 64, 128);
240 _tile_dpbssd(1, 6, 7);
241
242 w = (const int8_t*) w + kremainder * 32;
243 k -= kremainder * sizeof(int8_t);
244 }
245
246 p -= 7 * sizeof(void*);
247 } while (p != 0);
248
249 // Add tile to bias
250 _tile_stored(0, res0, 64);
251 _tile_stored(1, res1, 64);
252
253 __m512i vacc0x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
254 __m512i vacc0xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vksumGHIJKLMNOPQRSTUV, _mm512_load_epi32(res1 + 0));
255 __m512i vacc1x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 16));
256 __m512i vacc1xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vksumGHIJKLMNOPQRSTUV, _mm512_load_epi32(res1 + 16));
257 __m512i vacc2x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 32));
258 __m512i vacc2xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vksumGHIJKLMNOPQRSTUV, _mm512_load_epi32(res1 + 32));
259 __m512i vacc3x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 48));
260 __m512i vacc3xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vksumGHIJKLMNOPQRSTUV, _mm512_load_epi32(res1 + 48));
261 __m512i vacc4x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 64));
262 __m512i vacc4xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vksumGHIJKLMNOPQRSTUV, _mm512_load_epi32(res1 + 64));
263 __m512i vacc5x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 80));
264 __m512i vacc5xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vksumGHIJKLMNOPQRSTUV, _mm512_load_epi32(res1 + 80));
265 __m512i vacc6x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 96));
266 __m512i vacc6xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vksumGHIJKLMNOPQRSTUV, _mm512_load_epi32(res1 + 96));
267
268 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
269 __m512 vscaled0xGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vacc0xGHIJKLMNOPQRSTUV);
270 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
271 __m512 vscaled1xGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vacc1xGHIJKLMNOPQRSTUV);
272 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
273 __m512 vscaled2xGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vacc2xGHIJKLMNOPQRSTUV);
274 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
275 __m512 vscaled3xGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vacc3xGHIJKLMNOPQRSTUV);
276 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
277 __m512 vscaled4xGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vacc4xGHIJKLMNOPQRSTUV);
278 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
279 __m512 vscaled5xGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vacc5xGHIJKLMNOPQRSTUV);
280 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
281 __m512 vscaled6xGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vacc6xGHIJKLMNOPQRSTUV);
282
283 const __m512 vscale0123456789ABCDEF = _mm512_loadu_ps((const float*) w + 0);
284 const __m512 vscaleGHIJKLMNOPQRSTUV = _mm512_loadu_ps((const float*) w + 16);
285 w = (const int32_t*) w + 32;
286
287 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale0123456789ABCDEF);
288 vscaled0xGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaled0xGHIJKLMNOPQRSTUV, vscaleGHIJKLMNOPQRSTUV);
289 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vscale0123456789ABCDEF);
290 vscaled1xGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaled1xGHIJKLMNOPQRSTUV, vscaleGHIJKLMNOPQRSTUV);
291 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vscale0123456789ABCDEF);
292 vscaled2xGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaled2xGHIJKLMNOPQRSTUV, vscaleGHIJKLMNOPQRSTUV);
293 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vscale0123456789ABCDEF);
294 vscaled3xGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaled3xGHIJKLMNOPQRSTUV, vscaleGHIJKLMNOPQRSTUV);
295 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vscale0123456789ABCDEF);
296 vscaled4xGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaled4xGHIJKLMNOPQRSTUV, vscaleGHIJKLMNOPQRSTUV);
297 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vscale0123456789ABCDEF);
298 vscaled5xGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaled5xGHIJKLMNOPQRSTUV, vscaleGHIJKLMNOPQRSTUV);
299 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vscale0123456789ABCDEF);
300 vscaled6xGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaled6xGHIJKLMNOPQRSTUV, vscaleGHIJKLMNOPQRSTUV);
301
302 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
303 vscaled0xGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaled0xGHIJKLMNOPQRSTUV, voutput_max_less_zero_point);
304 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max_less_zero_point);
305 vscaled1xGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaled1xGHIJKLMNOPQRSTUV, voutput_max_less_zero_point);
306 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max_less_zero_point);
307 vscaled2xGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaled2xGHIJKLMNOPQRSTUV, voutput_max_less_zero_point);
308 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max_less_zero_point);
309 vscaled3xGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaled3xGHIJKLMNOPQRSTUV, voutput_max_less_zero_point);
310 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max_less_zero_point);
311 vscaled4xGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaled4xGHIJKLMNOPQRSTUV, voutput_max_less_zero_point);
312 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max_less_zero_point);
313 vscaled5xGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaled5xGHIJKLMNOPQRSTUV, voutput_max_less_zero_point);
314 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max_less_zero_point);
315 vscaled6xGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaled6xGHIJKLMNOPQRSTUV, voutput_max_less_zero_point);
316
317 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
318 vacc0xGHIJKLMNOPQRSTUV = _mm512_cvtps_epi32(vscaled0xGHIJKLMNOPQRSTUV);
319 vacc1x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled1x0123456789ABCDEF);
320 vacc1xGHIJKLMNOPQRSTUV = _mm512_cvtps_epi32(vscaled1xGHIJKLMNOPQRSTUV);
321 vacc2x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled2x0123456789ABCDEF);
322 vacc2xGHIJKLMNOPQRSTUV = _mm512_cvtps_epi32(vscaled2xGHIJKLMNOPQRSTUV);
323 vacc3x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled3x0123456789ABCDEF);
324 vacc3xGHIJKLMNOPQRSTUV = _mm512_cvtps_epi32(vscaled3xGHIJKLMNOPQRSTUV);
325 vacc4x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled4x0123456789ABCDEF);
326 vacc4xGHIJKLMNOPQRSTUV = _mm512_cvtps_epi32(vscaled4xGHIJKLMNOPQRSTUV);
327 vacc5x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled5x0123456789ABCDEF);
328 vacc5xGHIJKLMNOPQRSTUV = _mm512_cvtps_epi32(vscaled5xGHIJKLMNOPQRSTUV);
329 vacc6x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled6x0123456789ABCDEF);
330 vacc6xGHIJKLMNOPQRSTUV = _mm512_cvtps_epi32(vscaled6xGHIJKLMNOPQRSTUV);
331
332 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
333 vacc0xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vacc0xGHIJKLMNOPQRSTUV, voutput_zero_point);
334 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, voutput_zero_point);
335 vacc1xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vacc1xGHIJKLMNOPQRSTUV, voutput_zero_point);
336 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, voutput_zero_point);
337 vacc2xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vacc2xGHIJKLMNOPQRSTUV, voutput_zero_point);
338 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, voutput_zero_point);
339 vacc3xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vacc3xGHIJKLMNOPQRSTUV, voutput_zero_point);
340 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, voutput_zero_point);
341 vacc4xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vacc4xGHIJKLMNOPQRSTUV, voutput_zero_point);
342 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, voutput_zero_point);
343 vacc5xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vacc5xGHIJKLMNOPQRSTUV, voutput_zero_point);
344 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, voutput_zero_point);
345 vacc6xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vacc6xGHIJKLMNOPQRSTUV, voutput_zero_point);
346
347 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
348 __m128i vout0xGHIJKLMNOPQRSTUV = _mm512_cvtsepi32_epi8(vacc0xGHIJKLMNOPQRSTUV);
349 __m128i vout1x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc1x0123456789ABCDEF);
350 __m128i vout1xGHIJKLMNOPQRSTUV = _mm512_cvtsepi32_epi8(vacc1xGHIJKLMNOPQRSTUV);
351 __m128i vout2x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc2x0123456789ABCDEF);
352 __m128i vout2xGHIJKLMNOPQRSTUV = _mm512_cvtsepi32_epi8(vacc2xGHIJKLMNOPQRSTUV);
353 __m128i vout3x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc3x0123456789ABCDEF);
354 __m128i vout3xGHIJKLMNOPQRSTUV = _mm512_cvtsepi32_epi8(vacc3xGHIJKLMNOPQRSTUV);
355 __m128i vout4x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc4x0123456789ABCDEF);
356 __m128i vout4xGHIJKLMNOPQRSTUV = _mm512_cvtsepi32_epi8(vacc4xGHIJKLMNOPQRSTUV);
357 __m128i vout5x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc5x0123456789ABCDEF);
358 __m128i vout5xGHIJKLMNOPQRSTUV = _mm512_cvtsepi32_epi8(vacc5xGHIJKLMNOPQRSTUV);
359 __m128i vout6x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc6x0123456789ABCDEF);
360 __m128i vout6xGHIJKLMNOPQRSTUV = _mm512_cvtsepi32_epi8(vacc6xGHIJKLMNOPQRSTUV);
361
362 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
363 vout0xGHIJKLMNOPQRSTUV = _mm_max_epi8(vout0xGHIJKLMNOPQRSTUV, voutput_min);
364 vout1x0123456789ABCDEF = _mm_max_epi8(vout1x0123456789ABCDEF, voutput_min);
365 vout1xGHIJKLMNOPQRSTUV = _mm_max_epi8(vout1xGHIJKLMNOPQRSTUV, voutput_min);
366 vout2x0123456789ABCDEF = _mm_max_epi8(vout2x0123456789ABCDEF, voutput_min);
367 vout2xGHIJKLMNOPQRSTUV = _mm_max_epi8(vout2xGHIJKLMNOPQRSTUV, voutput_min);
368 vout3x0123456789ABCDEF = _mm_max_epi8(vout3x0123456789ABCDEF, voutput_min);
369 vout3xGHIJKLMNOPQRSTUV = _mm_max_epi8(vout3xGHIJKLMNOPQRSTUV, voutput_min);
370 vout4x0123456789ABCDEF = _mm_max_epi8(vout4x0123456789ABCDEF, voutput_min);
371 vout4xGHIJKLMNOPQRSTUV = _mm_max_epi8(vout4xGHIJKLMNOPQRSTUV, voutput_min);
372 vout5x0123456789ABCDEF = _mm_max_epi8(vout5x0123456789ABCDEF, voutput_min);
373 vout5xGHIJKLMNOPQRSTUV = _mm_max_epi8(vout5xGHIJKLMNOPQRSTUV, voutput_min);
374 vout6x0123456789ABCDEF = _mm_max_epi8(vout6x0123456789ABCDEF, voutput_min);
375 vout6xGHIJKLMNOPQRSTUV = _mm_max_epi8(vout6xGHIJKLMNOPQRSTUV, voutput_min);
376
377 if XNN_LIKELY(nc >= 32) {
378 _mm_storeu_si128((__m128i*) (c6 + 0), vout6x0123456789ABCDEF);
379 _mm_storeu_si128((__m128i*) (c6 + 16), vout6xGHIJKLMNOPQRSTUV);
380 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
381 _mm_storeu_si128((__m128i*) (c5 + 0), vout5x0123456789ABCDEF);
382 _mm_storeu_si128((__m128i*) (c5 + 16), vout5xGHIJKLMNOPQRSTUV);
383 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
384 _mm_storeu_si128((__m128i*) (c4 + 0), vout4x0123456789ABCDEF);
385 _mm_storeu_si128((__m128i*) (c4 + 16), vout4xGHIJKLMNOPQRSTUV);
386 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
387 _mm_storeu_si128((__m128i*) (c3 + 0), vout3x0123456789ABCDEF);
388 _mm_storeu_si128((__m128i*) (c3 + 16), vout3xGHIJKLMNOPQRSTUV);
389 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
390 _mm_storeu_si128((__m128i*) (c2 + 0), vout2x0123456789ABCDEF);
391 _mm_storeu_si128((__m128i*) (c2 + 16), vout2xGHIJKLMNOPQRSTUV);
392 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
393 _mm_storeu_si128((__m128i*) (c1 + 0), vout1x0123456789ABCDEF);
394 _mm_storeu_si128((__m128i*) (c1 + 16), vout1xGHIJKLMNOPQRSTUV);
395 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
396 _mm_storeu_si128((__m128i*) (c0 + 0), vout0x0123456789ABCDEF);
397 _mm_storeu_si128((__m128i*) (c0 + 16), vout0xGHIJKLMNOPQRSTUV);
398 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
399 a = (const int8_t**restrict) ((uintptr_t) a - ks);
400 nc -= 32;
401 } else {
402 // Prepare mask for valid 8-bit elements (depends on nc).
403 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
404 const __mmask16 vmask1 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 16) & 0xFFFF));
405
406 _mm_mask_storeu_epi8(c6 + 0, vmask0, vout6x0123456789ABCDEF);
407 _mm_mask_storeu_epi8(c6 + 16, vmask1, vout6xGHIJKLMNOPQRSTUV);
408 _mm_mask_storeu_epi8(c5 + 0, vmask0, vout5x0123456789ABCDEF);
409 _mm_mask_storeu_epi8(c5 + 16, vmask1, vout5xGHIJKLMNOPQRSTUV);
410 _mm_mask_storeu_epi8(c4 + 0, vmask0, vout4x0123456789ABCDEF);
411 _mm_mask_storeu_epi8(c4 + 16, vmask1, vout4xGHIJKLMNOPQRSTUV);
412 _mm_mask_storeu_epi8(c3 + 0, vmask0, vout3x0123456789ABCDEF);
413 _mm_mask_storeu_epi8(c3 + 16, vmask1, vout3xGHIJKLMNOPQRSTUV);
414 _mm_mask_storeu_epi8(c2 + 0, vmask0, vout2x0123456789ABCDEF);
415 _mm_mask_storeu_epi8(c2 + 16, vmask1, vout2xGHIJKLMNOPQRSTUV);
416 _mm_mask_storeu_epi8(c1 + 0, vmask0, vout1x0123456789ABCDEF);
417 _mm_mask_storeu_epi8(c1 + 16, vmask1, vout1xGHIJKLMNOPQRSTUV);
418 _mm_mask_storeu_epi8(c0 + 0, vmask0, vout0x0123456789ABCDEF);
419 _mm_mask_storeu_epi8(c0 + 16, vmask1, vout0xGHIJKLMNOPQRSTUV);
420 nc = 0;
421 }
422 } while (nc != 0);
423
424 // Release tile config
425 // _tile_release();
426 __asm__ volatile ("tilerelease" ::);
427 #endif // defined(__x86_64__)
428 }