Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-16x16c4-minmax-avx512amx-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/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_qd8_f32_qc8w_gemm_minmax_ukernel_16x16c4__avx512amx_prfm(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 const int8_t* restrict a,
27 size_t a_stride,
28 const void* restrict w,
29 float* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
33 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)])
34 {
35 assert(mr != 0);
36 assert(mr <= 16);
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 // TODO: amxintrin.h only provide intrinsics for __x86_64__
45 // Update if amxintrin changes
46 #if defined(__x86_64__)
47 __attribute__((aligned(64))) int32_t res0[16 * 16];
48
49 kc = round_up_po2(kc, 4 * sizeof(int8_t));
50 const size_t kremainder = (kc & 63) ? (kc & 63) : 64;
51
52 // Define tile config data structure
53 struct __tile_config {
54 uint8_t palette_id;
55 uint8_t start_row;
56 uint8_t reserved_0[14];
57 uint16_t colsb[8];
58 uint16_t reserved_1[8];
59 uint8_t rows[8];
60 uint8_t reserved_2[8];
61 };
62
63 // Load tile configuration
64 __attribute__((aligned(64))) struct __tile_config tile_data = {0};
65 tile_data.palette_id = 1;
66 tile_data.rows[0] = mr; // tmm0 = res 0
67 tile_data.rows[1] = mr; // tmm1 = res 1
68 tile_data.rows[2] = mr; // tmm2 = res 2
69 tile_data.rows[3] = mr; // tmm3 = res 3
70 tile_data.rows[4] = mr; // tmm4 = input
71 tile_data.rows[5] = 16; // tmm5 = weights
72 tile_data.rows[6] = mr; // tmm6 = input remainder
73 tile_data.rows[7] = kremainder >> 2; // tmm7 = weights remainder
74
75 tile_data.colsb[0] = 64; // tmm0 = res 0
76 tile_data.colsb[1] = 64; // tmm1 = res 1
77 tile_data.colsb[2] = 64; // tmm2 = res 1
78 tile_data.colsb[3] = 64; // tmm3 = res 1
79 tile_data.colsb[4] = 64; // tmm4 = input
80 tile_data.colsb[5] = 64; // tmm5 = weights
81 tile_data.colsb[6] = kremainder; // tmm6 = input remainder
82 tile_data.colsb[7] = 64; // tmm7 = weights remainder
83
84 //_tile_loadconfig(&tile_data);
85 __asm__ volatile ("ldtilecfg %0" :: "m" (tile_data));
86
87 float* c0 = c;
88 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
89 if XNN_UNPREDICTABLE(mr < 2) {
90 c1 = c0;
91 }
92 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
93 if XNN_UNPREDICTABLE(mr <= 2) {
94 c2 = c1;
95 }
96 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
97 if XNN_UNPREDICTABLE(mr < 4) {
98 c3 = c2;
99 }
100 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
101 if XNN_UNPREDICTABLE(mr <= 4) {
102 c4 = c3;
103 }
104 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
105 if XNN_UNPREDICTABLE(mr < 6) {
106 c5 = c4;
107 }
108 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
109 if XNN_UNPREDICTABLE(mr <= 6) {
110 c6 = c5;
111 }
112 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
113 if XNN_UNPREDICTABLE(mr < 8) {
114 c7 = c6;
115 }
116 float* c8 = (float*) ((uintptr_t) c7 + cm_stride);
117 if XNN_UNPREDICTABLE(mr <= 8) {
118 c8 = c7;
119 }
120 float* c9 = (float*) ((uintptr_t) c8 + cm_stride);
121 if XNN_UNPREDICTABLE(mr < 10) {
122 c9 = c8;
123 }
124 float* c10 = (float*) ((uintptr_t) c9 + cm_stride);
125 if XNN_UNPREDICTABLE(mr <= 10) {
126 c10 = c9;
127 }
128 float* c11 = (float*) ((uintptr_t) c10 + cm_stride);
129 if XNN_UNPREDICTABLE(mr < 12) {
130 c11 = c10;
131 }
132 float* c12 = (float*) ((uintptr_t) c11 + cm_stride);
133 if XNN_UNPREDICTABLE(mr <= 12) {
134 c12 = c11;
135 }
136 float* c13 = (float*) ((uintptr_t) c12 + cm_stride);
137 if XNN_UNPREDICTABLE(mr < 14) {
138 c13 = c12;
139 }
140 float* c14 = (float*) ((uintptr_t) c13 + cm_stride);
141 if XNN_UNPREDICTABLE(mr <= 14) {
142 c14 = c13;
143 }
144 float* c15 = (float*) ((uintptr_t) c14 + cm_stride);
145 if XNN_UNPREDICTABLE(mr != 16) {
146 c15 = c14;
147 }
148
149 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
150 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
151 // XNN_FORCE_REALIZATION(voutput_min);
152 // XNN_FORCE_REALIZATION(voutput_max);
153
154 do {
155 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32((const int32_t*) w + 0);
156 w = (const int32_t*) w + 16;
157
158 // Zero tile accumulator
159 __asm__ volatile (
160 "tilezero %%tmm0\n"
161 ::);
162
163 size_t k = kc;
164 while (k >= 64 * sizeof(int8_t)) {
165 _tile_loadd(4, a, a_stride);
166 a += 64;
167 _tile_loadd(5, (const int8_t*) w + 0, 64);
168 _tile_dpbssd(0, 4, 5);
169 xnn_prefetch_to_l1((const int8_t*) w + 4096);
170 xnn_prefetch_to_l1((const int8_t*) w + 4160);
171 xnn_prefetch_to_l1((const int8_t*) w + 4224);
172 xnn_prefetch_to_l1((const int8_t*) w + 4288);
173 xnn_prefetch_to_l1((const int8_t*) w + 4352);
174 xnn_prefetch_to_l1((const int8_t*) w + 4416);
175 xnn_prefetch_to_l1((const int8_t*) w + 4480);
176 xnn_prefetch_to_l1((const int8_t*) w + 4544);
177 xnn_prefetch_to_l1((const int8_t*) w + 4608);
178 xnn_prefetch_to_l1((const int8_t*) w + 4672);
179 xnn_prefetch_to_l1((const int8_t*) w + 4736);
180 xnn_prefetch_to_l1((const int8_t*) w + 4800);
181 xnn_prefetch_to_l1((const int8_t*) w + 4864);
182 xnn_prefetch_to_l1((const int8_t*) w + 4928);
183 xnn_prefetch_to_l1((const int8_t*) w + 4992);
184 xnn_prefetch_to_l1((const int8_t*) w + 5056);
185 xnn_prefetch_to_l1((const int8_t*) w + 5120);
186 xnn_prefetch_to_l1((const int8_t*) w + 5184);
187 xnn_prefetch_to_l1((const int8_t*) w + 5248);
188 xnn_prefetch_to_l1((const int8_t*) w + 5312);
189 xnn_prefetch_to_l1((const int8_t*) w + 5376);
190 xnn_prefetch_to_l1((const int8_t*) w + 5440);
191 xnn_prefetch_to_l1((const int8_t*) w + 5504);
192 xnn_prefetch_to_l1((const int8_t*) w + 5568);
193 xnn_prefetch_to_l1((const int8_t*) w + 5632);
194 xnn_prefetch_to_l1((const int8_t*) w + 5696);
195 xnn_prefetch_to_l1((const int8_t*) w + 5760);
196 xnn_prefetch_to_l1((const int8_t*) w + 5824);
197 xnn_prefetch_to_l1((const int8_t*) w + 5888);
198 xnn_prefetch_to_l1((const int8_t*) w + 5952);
199 xnn_prefetch_to_l1((const int8_t*) w + 6016);
200 xnn_prefetch_to_l1((const int8_t*) w + 6080);
201
202 w = (const int8_t*) w + 1024;
203 k -= 64 * sizeof(int8_t);
204 }
205
206 if XNN_UNLIKELY(k != 0) {
207 _tile_loadd(6, a, a_stride);
208 a += kremainder;
209 _tile_loadd(7, (const int8_t*) w + 0, 64);
210 _tile_dpbssd(0, 6, 7);
211
212 w = (const int8_t*) w + kremainder * 16;
213 k -= kremainder * sizeof(int8_t);
214 }
215
216 // Add tile to bias
217 _tile_stored(0, res0, 64);
218
219 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[0].zero_point));
220 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[1].zero_point));
221 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[2].zero_point));
222 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[3].zero_point));
223 __m512i vacc4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[4].zero_point));
224 __m512i vacc5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[5].zero_point));
225 __m512i vacc6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[6].zero_point));
226 __m512i vacc7x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[7].zero_point));
227 __m512i vacc8x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[8].zero_point));
228 __m512i vacc9x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[9].zero_point));
229 __m512i vacc10x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[10].zero_point));
230 __m512i vacc11x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[11].zero_point));
231 __m512i vacc12x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[12].zero_point));
232 __m512i vacc13x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[13].zero_point));
233 __m512i vacc14x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[14].zero_point));
234 __m512i vacc15x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[15].zero_point));
235 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
236 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, _mm512_load_epi32(res0 + 16));
237 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, _mm512_load_epi32(res0 + 32));
238 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, _mm512_load_epi32(res0 + 48));
239 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, _mm512_load_epi32(res0 + 64));
240 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, _mm512_load_epi32(res0 + 80));
241 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, _mm512_load_epi32(res0 + 96));
242 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, _mm512_load_epi32(res0 + 112));
243 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, _mm512_load_epi32(res0 + 128));
244 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, _mm512_load_epi32(res0 + 144));
245 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, _mm512_load_epi32(res0 + 160));
246 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, _mm512_load_epi32(res0 + 176));
247 vacc12x0123456789ABCDEF = _mm512_add_epi32(vacc12x0123456789ABCDEF, _mm512_load_epi32(res0 + 192));
248 vacc13x0123456789ABCDEF = _mm512_add_epi32(vacc13x0123456789ABCDEF, _mm512_load_epi32(res0 + 208));
249 vacc14x0123456789ABCDEF = _mm512_add_epi32(vacc14x0123456789ABCDEF, _mm512_load_epi32(res0 + 224));
250 vacc15x0123456789ABCDEF = _mm512_add_epi32(vacc15x0123456789ABCDEF, _mm512_load_epi32(res0 + 240));
251
252 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
253 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
254 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
255 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
256 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
257 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
258 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
259 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
260 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
261 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
262 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
263 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
264 __m512 vscaled12x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc12x0123456789ABCDEF);
265 __m512 vscaled13x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc13x0123456789ABCDEF);
266 __m512 vscaled14x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc14x0123456789ABCDEF);
267 __m512 vscaled15x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc15x0123456789ABCDEF);
268
269 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
270 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
271 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
272 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
273 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
274 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
275 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
276 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
277 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, _mm512_set1_ps(quantization_params[8].inv_scale));
278 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, _mm512_set1_ps(quantization_params[9].inv_scale));
279 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, _mm512_set1_ps(quantization_params[10].inv_scale));
280 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, _mm512_set1_ps(quantization_params[11].inv_scale));
281 vscaled12x0123456789ABCDEF = _mm512_mul_ps(vscaled12x0123456789ABCDEF, _mm512_set1_ps(quantization_params[12].inv_scale));
282 vscaled13x0123456789ABCDEF = _mm512_mul_ps(vscaled13x0123456789ABCDEF, _mm512_set1_ps(quantization_params[13].inv_scale));
283 vscaled14x0123456789ABCDEF = _mm512_mul_ps(vscaled14x0123456789ABCDEF, _mm512_set1_ps(quantization_params[14].inv_scale));
284 vscaled15x0123456789ABCDEF = _mm512_mul_ps(vscaled15x0123456789ABCDEF, _mm512_set1_ps(quantization_params[15].inv_scale));
285
286 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w + 0);
287 w = (const int32_t*) w + 16;
288 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 0);
289 w = (const int32_t*) w + 16;
290
291 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
292 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
293 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
294 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
295 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
296 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
297 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
298 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
299 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
300 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vscaled9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
301 vscaled10x0123456789ABCDEF = _mm512_fmadd_ps(vscaled10x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
302 vscaled11x0123456789ABCDEF = _mm512_fmadd_ps(vscaled11x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
303 vscaled12x0123456789ABCDEF = _mm512_fmadd_ps(vscaled12x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
304 vscaled13x0123456789ABCDEF = _mm512_fmadd_ps(vscaled13x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
305 vscaled14x0123456789ABCDEF = _mm512_fmadd_ps(vscaled14x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
306 vscaled15x0123456789ABCDEF = _mm512_fmadd_ps(vscaled15x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
307
308 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
309 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
310 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
311 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
312 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
313 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
314 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
315 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
316 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
317 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
318 vscaled10x0123456789ABCDEF = _mm512_max_ps(vscaled10x0123456789ABCDEF, voutput_min);
319 vscaled11x0123456789ABCDEF = _mm512_max_ps(vscaled11x0123456789ABCDEF, voutput_min);
320 vscaled12x0123456789ABCDEF = _mm512_max_ps(vscaled12x0123456789ABCDEF, voutput_min);
321 vscaled13x0123456789ABCDEF = _mm512_max_ps(vscaled13x0123456789ABCDEF, voutput_min);
322 vscaled14x0123456789ABCDEF = _mm512_max_ps(vscaled14x0123456789ABCDEF, voutput_min);
323 vscaled15x0123456789ABCDEF = _mm512_max_ps(vscaled15x0123456789ABCDEF, voutput_min);
324
325 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
326 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
327 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
328 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
329 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
330 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
331 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
332 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
333 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
334 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
335 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max);
336 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max);
337 vscaled12x0123456789ABCDEF = _mm512_min_ps(vscaled12x0123456789ABCDEF, voutput_max);
338 vscaled13x0123456789ABCDEF = _mm512_min_ps(vscaled13x0123456789ABCDEF, voutput_max);
339 vscaled14x0123456789ABCDEF = _mm512_min_ps(vscaled14x0123456789ABCDEF, voutput_max);
340 vscaled15x0123456789ABCDEF = _mm512_min_ps(vscaled15x0123456789ABCDEF, voutput_max);
341
342 if XNN_LIKELY(nc >= 16) {
343 _mm512_storeu_ps(c15 + 0, vscaled15x0123456789ABCDEF);
344 c15 = (float*) ((uintptr_t) c15 + cn_stride);
345 _mm512_storeu_ps(c14 + 0, vscaled14x0123456789ABCDEF);
346 c14 = (float*) ((uintptr_t) c14 + cn_stride);
347 _mm512_storeu_ps(c13 + 0, vscaled13x0123456789ABCDEF);
348 c13 = (float*) ((uintptr_t) c13 + cn_stride);
349 _mm512_storeu_ps(c12 + 0, vscaled12x0123456789ABCDEF);
350 c12 = (float*) ((uintptr_t) c12 + cn_stride);
351 _mm512_storeu_ps(c11 + 0, vscaled11x0123456789ABCDEF);
352 c11 = (float*) ((uintptr_t) c11 + cn_stride);
353 _mm512_storeu_ps(c10 + 0, vscaled10x0123456789ABCDEF);
354 c10 = (float*) ((uintptr_t) c10 + cn_stride);
355 _mm512_storeu_ps(c9 + 0, vscaled9x0123456789ABCDEF);
356 c9 = (float*) ((uintptr_t) c9 + cn_stride);
357 _mm512_storeu_ps(c8 + 0, vscaled8x0123456789ABCDEF);
358 c8 = (float*) ((uintptr_t) c8 + cn_stride);
359 _mm512_storeu_ps(c7 + 0, vscaled7x0123456789ABCDEF);
360 c7 = (float*) ((uintptr_t) c7 + cn_stride);
361 _mm512_storeu_ps(c6 + 0, vscaled6x0123456789ABCDEF);
362 c6 = (float*) ((uintptr_t) c6 + cn_stride);
363 _mm512_storeu_ps(c5 + 0, vscaled5x0123456789ABCDEF);
364 c5 = (float*) ((uintptr_t) c5 + cn_stride);
365 _mm512_storeu_ps(c4 + 0, vscaled4x0123456789ABCDEF);
366 c4 = (float*) ((uintptr_t) c4 + cn_stride);
367 _mm512_storeu_ps(c3 + 0, vscaled3x0123456789ABCDEF);
368 c3 = (float*) ((uintptr_t) c3 + cn_stride);
369 _mm512_storeu_ps(c2 + 0, vscaled2x0123456789ABCDEF);
370 c2 = (float*) ((uintptr_t) c2 + cn_stride);
371 _mm512_storeu_ps(c1 + 0, vscaled1x0123456789ABCDEF);
372 c1 = (float*) ((uintptr_t) c1 + cn_stride);
373 _mm512_storeu_ps(c0 + 0, vscaled0x0123456789ABCDEF);
374 c0 = (float*) ((uintptr_t) c0 + cn_stride);
375
376 a -= kc;
377 nc -= 16;
378 } else {
379 // Prepare mask for valid 32-bit elements (depends on nc).
380 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
381 _mm512_mask_storeu_ps(c15 + 0, vmask0, vscaled15x0123456789ABCDEF);
382 _mm512_mask_storeu_ps(c14 + 0, vmask0, vscaled14x0123456789ABCDEF);
383 _mm512_mask_storeu_ps(c13 + 0, vmask0, vscaled13x0123456789ABCDEF);
384 _mm512_mask_storeu_ps(c12 + 0, vmask0, vscaled12x0123456789ABCDEF);
385 _mm512_mask_storeu_ps(c11 + 0, vmask0, vscaled11x0123456789ABCDEF);
386 _mm512_mask_storeu_ps(c10 + 0, vmask0, vscaled10x0123456789ABCDEF);
387 _mm512_mask_storeu_ps(c9 + 0, vmask0, vscaled9x0123456789ABCDEF);
388 _mm512_mask_storeu_ps(c8 + 0, vmask0, vscaled8x0123456789ABCDEF);
389 _mm512_mask_storeu_ps(c7 + 0, vmask0, vscaled7x0123456789ABCDEF);
390 _mm512_mask_storeu_ps(c6 + 0, vmask0, vscaled6x0123456789ABCDEF);
391 _mm512_mask_storeu_ps(c5 + 0, vmask0, vscaled5x0123456789ABCDEF);
392 _mm512_mask_storeu_ps(c4 + 0, vmask0, vscaled4x0123456789ABCDEF);
393 _mm512_mask_storeu_ps(c3 + 0, vmask0, vscaled3x0123456789ABCDEF);
394 _mm512_mask_storeu_ps(c2 + 0, vmask0, vscaled2x0123456789ABCDEF);
395 _mm512_mask_storeu_ps(c1 + 0, vmask0, vscaled1x0123456789ABCDEF);
396 _mm512_mask_storeu_ps(c0 + 0, vmask0, vscaled0x0123456789ABCDEF);
397 nc = 0;
398 }
399 } while (nc != 0);
400 // Release tile config
401 // _tile_release();
402 __asm__ volatile ("tilerelease" ::);
403 #endif // defined(__x86_64__)
404 }