Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-16x16c4-minmax-avx512amx.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
20
21 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_16x16c4__avx512amx(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 const int8_t* restrict a,
26 size_t a_stride,
27 const void* restrict w,
28 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)])
33 {
34 assert(mr != 0);
35 assert(mr <= 16);
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 res0[16 * 16];
47
48 kc = round_up_po2(kc, 4 * sizeof(int8_t));
49 const size_t kremainder = (kc & 63) ? (kc & 63) : 64;
50
51 // Define tile config data structure
52 struct __tile_config {
53 uint8_t palette_id;
54 uint8_t start_row;
55 uint8_t reserved_0[14];
56 uint16_t colsb[8];
57 uint16_t reserved_1[8];
58 uint8_t rows[8];
59 uint8_t reserved_2[8];
60 };
61
62 // Load tile configuration
63 __attribute__((aligned(64))) struct __tile_config tile_data = {0};
64 tile_data.palette_id = 1;
65 tile_data.rows[0] = mr; // tmm0 = res 0
66 tile_data.rows[1] = mr; // tmm1 = res 1
67 tile_data.rows[2] = mr; // tmm2 = res 2
68 tile_data.rows[3] = mr; // tmm3 = res 3
69 tile_data.rows[4] = mr; // tmm4 = input
70 tile_data.rows[5] = 16; // tmm5 = weights
71 tile_data.rows[6] = mr; // tmm6 = input remainder
72 tile_data.rows[7] = kremainder >> 2; // tmm7 = weights remainder
73
74 tile_data.colsb[0] = 64; // tmm0 = res 0
75 tile_data.colsb[1] = 64; // tmm1 = res 1
76 tile_data.colsb[2] = 64; // tmm2 = res 1
77 tile_data.colsb[3] = 64; // tmm3 = res 1
78 tile_data.colsb[4] = 64; // tmm4 = input
79 tile_data.colsb[5] = 64; // tmm5 = weights
80 tile_data.colsb[6] = kremainder; // tmm6 = input remainder
81 tile_data.colsb[7] = 64; // tmm7 = weights remainder
82
83 //_tile_loadconfig(&tile_data);
84 __asm__ volatile ("ldtilecfg %0" :: "m" (tile_data));
85
86 float* c0 = c;
87 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
88 if XNN_UNPREDICTABLE(mr < 2) {
89 c1 = c0;
90 }
91 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
92 if XNN_UNPREDICTABLE(mr <= 2) {
93 c2 = c1;
94 }
95 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
96 if XNN_UNPREDICTABLE(mr < 4) {
97 c3 = c2;
98 }
99 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
100 if XNN_UNPREDICTABLE(mr <= 4) {
101 c4 = c3;
102 }
103 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
104 if XNN_UNPREDICTABLE(mr < 6) {
105 c5 = c4;
106 }
107 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
108 if XNN_UNPREDICTABLE(mr <= 6) {
109 c6 = c5;
110 }
111 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
112 if XNN_UNPREDICTABLE(mr < 8) {
113 c7 = c6;
114 }
115 float* c8 = (float*) ((uintptr_t) c7 + cm_stride);
116 if XNN_UNPREDICTABLE(mr <= 8) {
117 c8 = c7;
118 }
119 float* c9 = (float*) ((uintptr_t) c8 + cm_stride);
120 if XNN_UNPREDICTABLE(mr < 10) {
121 c9 = c8;
122 }
123 float* c10 = (float*) ((uintptr_t) c9 + cm_stride);
124 if XNN_UNPREDICTABLE(mr <= 10) {
125 c10 = c9;
126 }
127 float* c11 = (float*) ((uintptr_t) c10 + cm_stride);
128 if XNN_UNPREDICTABLE(mr < 12) {
129 c11 = c10;
130 }
131 float* c12 = (float*) ((uintptr_t) c11 + cm_stride);
132 if XNN_UNPREDICTABLE(mr <= 12) {
133 c12 = c11;
134 }
135 float* c13 = (float*) ((uintptr_t) c12 + cm_stride);
136 if XNN_UNPREDICTABLE(mr < 14) {
137 c13 = c12;
138 }
139 float* c14 = (float*) ((uintptr_t) c13 + cm_stride);
140 if XNN_UNPREDICTABLE(mr <= 14) {
141 c14 = c13;
142 }
143 float* c15 = (float*) ((uintptr_t) c14 + cm_stride);
144 if XNN_UNPREDICTABLE(mr != 16) {
145 c15 = c14;
146 }
147
148 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
149 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
150 // XNN_FORCE_REALIZATION(voutput_min);
151 // XNN_FORCE_REALIZATION(voutput_max);
152
153 do {
154 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32((const int32_t*) w + 0);
155 w = (const int32_t*) w + 16;
156
157 // Zero tile accumulator
158 __asm__ volatile (
159 "tilezero %%tmm0\n"
160 ::);
161
162 size_t k = kc;
163 while (k >= 64 * sizeof(int8_t)) {
164 _tile_loadd(4, a, a_stride);
165 a += 64;
166 _tile_loadd(5, (const int8_t*) w + 0, 64);
167 _tile_dpbssd(0, 4, 5);
168
169 w = (const int8_t*) w + 1024;
170 k -= 64 * sizeof(int8_t);
171 }
172
173 if XNN_UNLIKELY(k != 0) {
174 _tile_loadd(6, a, a_stride);
175 a += kremainder;
176 _tile_loadd(7, (const int8_t*) w + 0, 64);
177 _tile_dpbssd(0, 6, 7);
178
179 w = (const int8_t*) w + kremainder * 16;
180 k -= kremainder * sizeof(int8_t);
181 }
182
183 // Add tile to bias
184 _tile_stored(0, res0, 64);
185
186 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[0].zero_point));
187 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[1].zero_point));
188 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[2].zero_point));
189 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[3].zero_point));
190 __m512i vacc4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[4].zero_point));
191 __m512i vacc5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[5].zero_point));
192 __m512i vacc6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[6].zero_point));
193 __m512i vacc7x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[7].zero_point));
194 __m512i vacc8x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[8].zero_point));
195 __m512i vacc9x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[9].zero_point));
196 __m512i vacc10x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[10].zero_point));
197 __m512i vacc11x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[11].zero_point));
198 __m512i vacc12x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[12].zero_point));
199 __m512i vacc13x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[13].zero_point));
200 __m512i vacc14x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[14].zero_point));
201 __m512i vacc15x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[15].zero_point));
202 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
203 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, _mm512_load_epi32(res0 + 16));
204 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, _mm512_load_epi32(res0 + 32));
205 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, _mm512_load_epi32(res0 + 48));
206 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, _mm512_load_epi32(res0 + 64));
207 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, _mm512_load_epi32(res0 + 80));
208 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, _mm512_load_epi32(res0 + 96));
209 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, _mm512_load_epi32(res0 + 112));
210 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, _mm512_load_epi32(res0 + 128));
211 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, _mm512_load_epi32(res0 + 144));
212 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, _mm512_load_epi32(res0 + 160));
213 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, _mm512_load_epi32(res0 + 176));
214 vacc12x0123456789ABCDEF = _mm512_add_epi32(vacc12x0123456789ABCDEF, _mm512_load_epi32(res0 + 192));
215 vacc13x0123456789ABCDEF = _mm512_add_epi32(vacc13x0123456789ABCDEF, _mm512_load_epi32(res0 + 208));
216 vacc14x0123456789ABCDEF = _mm512_add_epi32(vacc14x0123456789ABCDEF, _mm512_load_epi32(res0 + 224));
217 vacc15x0123456789ABCDEF = _mm512_add_epi32(vacc15x0123456789ABCDEF, _mm512_load_epi32(res0 + 240));
218
219 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
220 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
221 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
222 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
223 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
224 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
225 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
226 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
227 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
228 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
229 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
230 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
231 __m512 vscaled12x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc12x0123456789ABCDEF);
232 __m512 vscaled13x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc13x0123456789ABCDEF);
233 __m512 vscaled14x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc14x0123456789ABCDEF);
234 __m512 vscaled15x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc15x0123456789ABCDEF);
235
236 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
237 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
238 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
239 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
240 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
241 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
242 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
243 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
244 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, _mm512_set1_ps(quantization_params[8].inv_scale));
245 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, _mm512_set1_ps(quantization_params[9].inv_scale));
246 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, _mm512_set1_ps(quantization_params[10].inv_scale));
247 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, _mm512_set1_ps(quantization_params[11].inv_scale));
248 vscaled12x0123456789ABCDEF = _mm512_mul_ps(vscaled12x0123456789ABCDEF, _mm512_set1_ps(quantization_params[12].inv_scale));
249 vscaled13x0123456789ABCDEF = _mm512_mul_ps(vscaled13x0123456789ABCDEF, _mm512_set1_ps(quantization_params[13].inv_scale));
250 vscaled14x0123456789ABCDEF = _mm512_mul_ps(vscaled14x0123456789ABCDEF, _mm512_set1_ps(quantization_params[14].inv_scale));
251 vscaled15x0123456789ABCDEF = _mm512_mul_ps(vscaled15x0123456789ABCDEF, _mm512_set1_ps(quantization_params[15].inv_scale));
252
253 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w + 0);
254 w = (const int32_t*) w + 16;
255 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 0);
256 w = (const int32_t*) w + 16;
257
258 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
259 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
260 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
261 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
262 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
263 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
264 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
265 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
266 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
267 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vscaled9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
268 vscaled10x0123456789ABCDEF = _mm512_fmadd_ps(vscaled10x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
269 vscaled11x0123456789ABCDEF = _mm512_fmadd_ps(vscaled11x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
270 vscaled12x0123456789ABCDEF = _mm512_fmadd_ps(vscaled12x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
271 vscaled13x0123456789ABCDEF = _mm512_fmadd_ps(vscaled13x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
272 vscaled14x0123456789ABCDEF = _mm512_fmadd_ps(vscaled14x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
273 vscaled15x0123456789ABCDEF = _mm512_fmadd_ps(vscaled15x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
274
275 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
276 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
277 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
278 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
279 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
280 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
281 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
282 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
283 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
284 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
285 vscaled10x0123456789ABCDEF = _mm512_max_ps(vscaled10x0123456789ABCDEF, voutput_min);
286 vscaled11x0123456789ABCDEF = _mm512_max_ps(vscaled11x0123456789ABCDEF, voutput_min);
287 vscaled12x0123456789ABCDEF = _mm512_max_ps(vscaled12x0123456789ABCDEF, voutput_min);
288 vscaled13x0123456789ABCDEF = _mm512_max_ps(vscaled13x0123456789ABCDEF, voutput_min);
289 vscaled14x0123456789ABCDEF = _mm512_max_ps(vscaled14x0123456789ABCDEF, voutput_min);
290 vscaled15x0123456789ABCDEF = _mm512_max_ps(vscaled15x0123456789ABCDEF, voutput_min);
291
292 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
293 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
294 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
295 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
296 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
297 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
298 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
299 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
300 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
301 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
302 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max);
303 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max);
304 vscaled12x0123456789ABCDEF = _mm512_min_ps(vscaled12x0123456789ABCDEF, voutput_max);
305 vscaled13x0123456789ABCDEF = _mm512_min_ps(vscaled13x0123456789ABCDEF, voutput_max);
306 vscaled14x0123456789ABCDEF = _mm512_min_ps(vscaled14x0123456789ABCDEF, voutput_max);
307 vscaled15x0123456789ABCDEF = _mm512_min_ps(vscaled15x0123456789ABCDEF, voutput_max);
308
309 if XNN_LIKELY(nc >= 16) {
310 _mm512_storeu_ps(c15 + 0, vscaled15x0123456789ABCDEF);
311 c15 = (float*) ((uintptr_t) c15 + cn_stride);
312 _mm512_storeu_ps(c14 + 0, vscaled14x0123456789ABCDEF);
313 c14 = (float*) ((uintptr_t) c14 + cn_stride);
314 _mm512_storeu_ps(c13 + 0, vscaled13x0123456789ABCDEF);
315 c13 = (float*) ((uintptr_t) c13 + cn_stride);
316 _mm512_storeu_ps(c12 + 0, vscaled12x0123456789ABCDEF);
317 c12 = (float*) ((uintptr_t) c12 + cn_stride);
318 _mm512_storeu_ps(c11 + 0, vscaled11x0123456789ABCDEF);
319 c11 = (float*) ((uintptr_t) c11 + cn_stride);
320 _mm512_storeu_ps(c10 + 0, vscaled10x0123456789ABCDEF);
321 c10 = (float*) ((uintptr_t) c10 + cn_stride);
322 _mm512_storeu_ps(c9 + 0, vscaled9x0123456789ABCDEF);
323 c9 = (float*) ((uintptr_t) c9 + cn_stride);
324 _mm512_storeu_ps(c8 + 0, vscaled8x0123456789ABCDEF);
325 c8 = (float*) ((uintptr_t) c8 + cn_stride);
326 _mm512_storeu_ps(c7 + 0, vscaled7x0123456789ABCDEF);
327 c7 = (float*) ((uintptr_t) c7 + cn_stride);
328 _mm512_storeu_ps(c6 + 0, vscaled6x0123456789ABCDEF);
329 c6 = (float*) ((uintptr_t) c6 + cn_stride);
330 _mm512_storeu_ps(c5 + 0, vscaled5x0123456789ABCDEF);
331 c5 = (float*) ((uintptr_t) c5 + cn_stride);
332 _mm512_storeu_ps(c4 + 0, vscaled4x0123456789ABCDEF);
333 c4 = (float*) ((uintptr_t) c4 + cn_stride);
334 _mm512_storeu_ps(c3 + 0, vscaled3x0123456789ABCDEF);
335 c3 = (float*) ((uintptr_t) c3 + cn_stride);
336 _mm512_storeu_ps(c2 + 0, vscaled2x0123456789ABCDEF);
337 c2 = (float*) ((uintptr_t) c2 + cn_stride);
338 _mm512_storeu_ps(c1 + 0, vscaled1x0123456789ABCDEF);
339 c1 = (float*) ((uintptr_t) c1 + cn_stride);
340 _mm512_storeu_ps(c0 + 0, vscaled0x0123456789ABCDEF);
341 c0 = (float*) ((uintptr_t) c0 + cn_stride);
342
343 a -= kc;
344 nc -= 16;
345 } else {
346 // Prepare mask for valid 32-bit elements (depends on nc).
347 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
348 _mm512_mask_storeu_ps(c15 + 0, vmask0, vscaled15x0123456789ABCDEF);
349 _mm512_mask_storeu_ps(c14 + 0, vmask0, vscaled14x0123456789ABCDEF);
350 _mm512_mask_storeu_ps(c13 + 0, vmask0, vscaled13x0123456789ABCDEF);
351 _mm512_mask_storeu_ps(c12 + 0, vmask0, vscaled12x0123456789ABCDEF);
352 _mm512_mask_storeu_ps(c11 + 0, vmask0, vscaled11x0123456789ABCDEF);
353 _mm512_mask_storeu_ps(c10 + 0, vmask0, vscaled10x0123456789ABCDEF);
354 _mm512_mask_storeu_ps(c9 + 0, vmask0, vscaled9x0123456789ABCDEF);
355 _mm512_mask_storeu_ps(c8 + 0, vmask0, vscaled8x0123456789ABCDEF);
356 _mm512_mask_storeu_ps(c7 + 0, vmask0, vscaled7x0123456789ABCDEF);
357 _mm512_mask_storeu_ps(c6 + 0, vmask0, vscaled6x0123456789ABCDEF);
358 _mm512_mask_storeu_ps(c5 + 0, vmask0, vscaled5x0123456789ABCDEF);
359 _mm512_mask_storeu_ps(c4 + 0, vmask0, vscaled4x0123456789ABCDEF);
360 _mm512_mask_storeu_ps(c3 + 0, vmask0, vscaled3x0123456789ABCDEF);
361 _mm512_mask_storeu_ps(c2 + 0, vmask0, vscaled2x0123456789ABCDEF);
362 _mm512_mask_storeu_ps(c1 + 0, vmask0, vscaled1x0123456789ABCDEF);
363 _mm512_mask_storeu_ps(c0 + 0, vmask0, vscaled0x0123456789ABCDEF);
364 nc = 0;
365 }
366 } while (nc != 0);
367 // Release tile config
368 // _tile_release();
369 __asm__ volatile ("tilerelease" ::);
370 #endif // defined(__x86_64__)
371 }