Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-7x32c4-minmax-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_qd8_f32_qc8w_igemm_minmax_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 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const int8_t* zero_data,
33 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
34 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)])
35 {
36 assert(mr != 0);
37 assert(mr <= 7);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 // TODO: amxintrin.h only provide intrinsics for __x86_64__
46 // Update if amxintrin changes
47 #if defined(__x86_64__)
48 __attribute__((aligned(64))) int32_t vintile[7 * 16];
49 __attribute__((aligned(64))) int32_t res0[7 * 16];
50 __attribute__((aligned(64))) int32_t res1[7 * 16];
51
52 kc = round_up_po2(kc, 4 * sizeof(int8_t));
53 const size_t kremainder = (kc & 63) ? (kc & 63) : 64;
54 const __mmask16 kremainder_mask = _cvtu32_mask16((UINT32_C(1) << (kremainder >> 2)) - 1);
55
56 // Define tile config data structure
57 struct __tile_config {
58 uint8_t palette_id;
59 uint8_t start_row;
60 uint8_t reserved_0[14];
61 uint16_t colsb[8];
62 uint16_t reserved_1[8];
63 uint8_t rows[8];
64 uint8_t reserved_2[8];
65 };
66
67 // Load tile configuration
68 __attribute__((aligned(64))) struct __tile_config tile_data = {0};
69 tile_data.palette_id = 1;
70 tile_data.rows[0] = mr; // tmm0 = res 0
71 tile_data.rows[1] = mr; // tmm1 = res 1
72 tile_data.rows[2] = mr; // tmm2 = res 2
73 tile_data.rows[3] = mr; // tmm3 = res 3
74 tile_data.rows[4] = mr; // tmm4 = input
75 tile_data.rows[5] = 16; // tmm5 = weights
76 tile_data.rows[6] = mr; // tmm6 = input remainder
77 tile_data.rows[7] = kremainder >> 2; // tmm7 = weights remainder
78
79 tile_data.colsb[0] = 64; // tmm0 = res 0
80 tile_data.colsb[1] = 64; // tmm1 = res 1
81 tile_data.colsb[2] = 64; // tmm2 = res 1
82 tile_data.colsb[3] = 64; // tmm3 = res 1
83 tile_data.colsb[4] = 64; // tmm4 = input
84 tile_data.colsb[5] = 64; // tmm5 = weights
85 tile_data.colsb[6] = kremainder; // tmm6 = input remainder
86 tile_data.colsb[7] = 64; // tmm7 = weights remainder
87
88 //_tile_loadconfig(&tile_data);
89 __asm__ volatile ("ldtilecfg %0" :: "m" (tile_data));
90
91 float* c0 = c;
92 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
93 if XNN_UNPREDICTABLE(mr < 2) {
94 c1 = c0;
95 }
96 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
97 if XNN_UNPREDICTABLE(mr <= 2) {
98 c2 = c1;
99 }
100 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
101 if XNN_UNPREDICTABLE(mr < 4) {
102 c3 = c2;
103 }
104 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
105 if XNN_UNPREDICTABLE(mr <= 4) {
106 c4 = c3;
107 }
108 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
109 if XNN_UNPREDICTABLE(mr < 6) {
110 c5 = c4;
111 }
112 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
113 if XNN_UNPREDICTABLE(mr <= 6) {
114 c6 = c5;
115 }
116
117 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
118 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
119 // XNN_FORCE_REALIZATION(voutput_min);
120 // XNN_FORCE_REALIZATION(voutput_max);
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 } else {
139 a0 = zero_data;
140 }
141 const int8_t* restrict a1 = a[1];
142 if XNN_UNPREDICTABLE(a1 != zero) {
143 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
144 } else {
145 a1 = zero_data;
146 }
147 const int8_t* restrict a2 = a[2];
148 if XNN_UNPREDICTABLE(a2 != zero) {
149 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
150 } else {
151 a2 = zero_data;
152 }
153 const int8_t* restrict a3 = a[3];
154 if XNN_UNPREDICTABLE(a3 != zero) {
155 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
156 } else {
157 a3 = zero_data;
158 }
159 const int8_t* restrict a4 = a[4];
160 if XNN_UNPREDICTABLE(a4 != zero) {
161 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
162 } else {
163 a4 = zero_data;
164 }
165 const int8_t* restrict a5 = a[5];
166 if XNN_UNPREDICTABLE(a5 != zero) {
167 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
168 } else {
169 a5 = zero_data;
170 }
171 const int8_t* restrict a6 = a[6];
172 if XNN_UNPREDICTABLE(a6 != zero) {
173 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
174 } else {
175 a6 = zero_data;
176 }
177 a += 7;
178
179 size_t k = kc;
180 if (mr == 1)
181 {
182 while (k >= 64 * sizeof(int8_t)) {
183 _tile_loadd(4, a0, 64); // Directly load input for mr=1
184 a6 += 64;
185 _tile_loadd(5, (const int8_t*) w + 0, 128);
186 _tile_dpbssd(0, 4, 5);
187 _tile_loadd(5, (const int8_t*) w + 64, 128);
188 _tile_dpbssd(1, 4, 5);
189
190 w = (const int8_t*) w + 2048;
191 k -= 64 * sizeof(int8_t);
192 }
193 }
194 else {
195 while (k >= 64 * sizeof(int8_t)) {
196 const __m512i vin0 = _mm512_loadu_epi32(a0);
197 a0 += 64;
198 _mm512_store_epi32(vintile + 0, vin0);
199 const __m512i vin1 = _mm512_loadu_epi32(a1);
200 a1 += 64;
201 _mm512_store_epi32(vintile + 16, vin1);
202 const __m512i vin2 = _mm512_loadu_epi32(a2);
203 a2 += 64;
204 _mm512_store_epi32(vintile + 32, vin2);
205 const __m512i vin3 = _mm512_loadu_epi32(a3);
206 a3 += 64;
207 _mm512_store_epi32(vintile + 48, vin3);
208 const __m512i vin4 = _mm512_loadu_epi32(a4);
209 a4 += 64;
210 _mm512_store_epi32(vintile + 64, vin4);
211 const __m512i vin5 = _mm512_loadu_epi32(a5);
212 a5 += 64;
213 _mm512_store_epi32(vintile + 80, vin5);
214 const __m512i vin6 = _mm512_loadu_epi32(a6);
215 a6 += 64;
216 _mm512_store_epi32(vintile + 96, vin6);
217 _tile_loadd(4, vintile, 64);
218 _tile_loadd(5, (const int8_t*) w + 0, 128);
219 _tile_dpbssd(0, 4, 5);
220 _tile_loadd(5, (const int8_t*) w + 64, 128);
221 _tile_dpbssd(1, 4, 5);
222
223 w = (const int8_t*) w + 2048;
224 k -= 64 * sizeof(int8_t);
225 }
226 }
227
228 if XNN_UNLIKELY(k != 0) {
229 const __m512i vin0 = _mm512_maskz_loadu_epi32(kremainder_mask, a0);
230 a0 += kremainder;
231 _mm512_store_epi32(vintile + 0, vin0);
232 const __m512i vin1 = _mm512_maskz_loadu_epi32(kremainder_mask, a1);
233 a1 += kremainder;
234 _mm512_store_epi32(vintile + 16, vin1);
235 const __m512i vin2 = _mm512_maskz_loadu_epi32(kremainder_mask, a2);
236 a2 += kremainder;
237 _mm512_store_epi32(vintile + 32, vin2);
238 const __m512i vin3 = _mm512_maskz_loadu_epi32(kremainder_mask, a3);
239 a3 += kremainder;
240 _mm512_store_epi32(vintile + 48, vin3);
241 const __m512i vin4 = _mm512_maskz_loadu_epi32(kremainder_mask, a4);
242 a4 += kremainder;
243 _mm512_store_epi32(vintile + 64, vin4);
244 const __m512i vin5 = _mm512_maskz_loadu_epi32(kremainder_mask, a5);
245 a5 += kremainder;
246 _mm512_store_epi32(vintile + 80, vin5);
247 const __m512i vin6 = _mm512_maskz_loadu_epi32(kremainder_mask, a6);
248 a6 += kremainder;
249 _mm512_store_epi32(vintile + 96, vin6);
250 _tile_loadd(6, vintile, 64);
251 _tile_loadd(7, (const int8_t*) w + 0, 128);
252 _tile_dpbssd(0, 6, 7);
253 _tile_loadd(7, (const int8_t*) w + 64, 128);
254 _tile_dpbssd(1, 6, 7);
255
256 w = (const int8_t*) w + kremainder * 32;
257 k -= kremainder * sizeof(int8_t);
258 }
259
260 p -= 7 * sizeof(void*);
261 } while (p != 0);
262
263 // Add tile to bias
264 _tile_stored(0, res0, 64);
265 _tile_stored(1, res1, 64);
266
267 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
268 __m512i vacc0xGHIJKLMNOPQRSTUV = _mm512_mullo_epi32(vksumGHIJKLMNOPQRSTUV, _mm512_set1_epi32((int) quantization_params->zero_point));
269 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
270 __m512i vacc1xGHIJKLMNOPQRSTUV = _mm512_mullo_epi32(vksumGHIJKLMNOPQRSTUV, _mm512_set1_epi32((int) quantization_params->zero_point));
271 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
272 __m512i vacc2xGHIJKLMNOPQRSTUV = _mm512_mullo_epi32(vksumGHIJKLMNOPQRSTUV, _mm512_set1_epi32((int) quantization_params->zero_point));
273 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
274 __m512i vacc3xGHIJKLMNOPQRSTUV = _mm512_mullo_epi32(vksumGHIJKLMNOPQRSTUV, _mm512_set1_epi32((int) quantization_params->zero_point));
275 __m512i vacc4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
276 __m512i vacc4xGHIJKLMNOPQRSTUV = _mm512_mullo_epi32(vksumGHIJKLMNOPQRSTUV, _mm512_set1_epi32((int) quantization_params->zero_point));
277 __m512i vacc5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
278 __m512i vacc5xGHIJKLMNOPQRSTUV = _mm512_mullo_epi32(vksumGHIJKLMNOPQRSTUV, _mm512_set1_epi32((int) quantization_params->zero_point));
279 __m512i vacc6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
280 __m512i vacc6xGHIJKLMNOPQRSTUV = _mm512_mullo_epi32(vksumGHIJKLMNOPQRSTUV, _mm512_set1_epi32((int) quantization_params->zero_point));
281 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
282 vacc0xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vacc0xGHIJKLMNOPQRSTUV, _mm512_load_epi32(res1 + 0));
283 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, _mm512_load_epi32(res0 + 16));
284 vacc1xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vacc1xGHIJKLMNOPQRSTUV, _mm512_load_epi32(res1 + 16));
285 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, _mm512_load_epi32(res0 + 32));
286 vacc2xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vacc2xGHIJKLMNOPQRSTUV, _mm512_load_epi32(res1 + 32));
287 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, _mm512_load_epi32(res0 + 48));
288 vacc3xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vacc3xGHIJKLMNOPQRSTUV, _mm512_load_epi32(res1 + 48));
289 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, _mm512_load_epi32(res0 + 64));
290 vacc4xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vacc4xGHIJKLMNOPQRSTUV, _mm512_load_epi32(res1 + 64));
291 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, _mm512_load_epi32(res0 + 80));
292 vacc5xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vacc5xGHIJKLMNOPQRSTUV, _mm512_load_epi32(res1 + 80));
293 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, _mm512_load_epi32(res0 + 96));
294 vacc6xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vacc6xGHIJKLMNOPQRSTUV, _mm512_load_epi32(res1 + 96));
295
296 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
297 __m512 vscaled0xGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vacc0xGHIJKLMNOPQRSTUV);
298 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
299 __m512 vscaled1xGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vacc1xGHIJKLMNOPQRSTUV);
300 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
301 __m512 vscaled2xGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vacc2xGHIJKLMNOPQRSTUV);
302 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
303 __m512 vscaled3xGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vacc3xGHIJKLMNOPQRSTUV);
304 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
305 __m512 vscaled4xGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vacc4xGHIJKLMNOPQRSTUV);
306 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
307 __m512 vscaled5xGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vacc5xGHIJKLMNOPQRSTUV);
308 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
309 __m512 vscaled6xGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vacc6xGHIJKLMNOPQRSTUV);
310
311 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
312 vscaled0xGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaled0xGHIJKLMNOPQRSTUV, _mm512_set1_ps(quantization_params->inv_scale));
313 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
314 vscaled1xGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaled1xGHIJKLMNOPQRSTUV, _mm512_set1_ps(quantization_params->inv_scale));
315 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
316 vscaled2xGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaled2xGHIJKLMNOPQRSTUV, _mm512_set1_ps(quantization_params->inv_scale));
317 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
318 vscaled3xGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaled3xGHIJKLMNOPQRSTUV, _mm512_set1_ps(quantization_params->inv_scale));
319 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
320 vscaled4xGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaled4xGHIJKLMNOPQRSTUV, _mm512_set1_ps(quantization_params->inv_scale));
321 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
322 vscaled5xGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaled5xGHIJKLMNOPQRSTUV, _mm512_set1_ps(quantization_params->inv_scale));
323 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
324 vscaled6xGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaled6xGHIJKLMNOPQRSTUV, _mm512_set1_ps(quantization_params->inv_scale));
325
326 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_loadu_ps((const float*) w + 0);
327 const __m512 vfilter_output_scaleGHIJKLMNOPQRSTUV = _mm512_loadu_ps((const float*) w + 16);
328 w = (const int32_t*) w + 32;
329 const __m512 vbias0123456789ABCDEF = _mm512_loadu_ps((const float*) w + 0);
330 const __m512 vbiasGHIJKLMNOPQRSTUV = _mm512_loadu_ps((const float*) w + 16);
331 w = (const int32_t*) w + 32;
332
333 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
334 vscaled0xGHIJKLMNOPQRSTUV = _mm512_fmadd_ps(vscaled0xGHIJKLMNOPQRSTUV, vfilter_output_scaleGHIJKLMNOPQRSTUV, vbiasGHIJKLMNOPQRSTUV);
335 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
336 vscaled1xGHIJKLMNOPQRSTUV = _mm512_fmadd_ps(vscaled1xGHIJKLMNOPQRSTUV, vfilter_output_scaleGHIJKLMNOPQRSTUV, vbiasGHIJKLMNOPQRSTUV);
337 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
338 vscaled2xGHIJKLMNOPQRSTUV = _mm512_fmadd_ps(vscaled2xGHIJKLMNOPQRSTUV, vfilter_output_scaleGHIJKLMNOPQRSTUV, vbiasGHIJKLMNOPQRSTUV);
339 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
340 vscaled3xGHIJKLMNOPQRSTUV = _mm512_fmadd_ps(vscaled3xGHIJKLMNOPQRSTUV, vfilter_output_scaleGHIJKLMNOPQRSTUV, vbiasGHIJKLMNOPQRSTUV);
341 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
342 vscaled4xGHIJKLMNOPQRSTUV = _mm512_fmadd_ps(vscaled4xGHIJKLMNOPQRSTUV, vfilter_output_scaleGHIJKLMNOPQRSTUV, vbiasGHIJKLMNOPQRSTUV);
343 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
344 vscaled5xGHIJKLMNOPQRSTUV = _mm512_fmadd_ps(vscaled5xGHIJKLMNOPQRSTUV, vfilter_output_scaleGHIJKLMNOPQRSTUV, vbiasGHIJKLMNOPQRSTUV);
345 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
346 vscaled6xGHIJKLMNOPQRSTUV = _mm512_fmadd_ps(vscaled6xGHIJKLMNOPQRSTUV, vfilter_output_scaleGHIJKLMNOPQRSTUV, vbiasGHIJKLMNOPQRSTUV);
347
348 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
349 vscaled0xGHIJKLMNOPQRSTUV = _mm512_max_ps(vscaled0xGHIJKLMNOPQRSTUV, voutput_min);
350 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
351 vscaled1xGHIJKLMNOPQRSTUV = _mm512_max_ps(vscaled1xGHIJKLMNOPQRSTUV, voutput_min);
352 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
353 vscaled2xGHIJKLMNOPQRSTUV = _mm512_max_ps(vscaled2xGHIJKLMNOPQRSTUV, voutput_min);
354 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
355 vscaled3xGHIJKLMNOPQRSTUV = _mm512_max_ps(vscaled3xGHIJKLMNOPQRSTUV, voutput_min);
356 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
357 vscaled4xGHIJKLMNOPQRSTUV = _mm512_max_ps(vscaled4xGHIJKLMNOPQRSTUV, voutput_min);
358 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
359 vscaled5xGHIJKLMNOPQRSTUV = _mm512_max_ps(vscaled5xGHIJKLMNOPQRSTUV, voutput_min);
360 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
361 vscaled6xGHIJKLMNOPQRSTUV = _mm512_max_ps(vscaled6xGHIJKLMNOPQRSTUV, voutput_min);
362
363 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
364 vscaled0xGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaled0xGHIJKLMNOPQRSTUV, voutput_max);
365 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
366 vscaled1xGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaled1xGHIJKLMNOPQRSTUV, voutput_max);
367 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
368 vscaled2xGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaled2xGHIJKLMNOPQRSTUV, voutput_max);
369 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
370 vscaled3xGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaled3xGHIJKLMNOPQRSTUV, voutput_max);
371 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
372 vscaled4xGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaled4xGHIJKLMNOPQRSTUV, voutput_max);
373 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
374 vscaled5xGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaled5xGHIJKLMNOPQRSTUV, voutput_max);
375 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
376 vscaled6xGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaled6xGHIJKLMNOPQRSTUV, voutput_max);
377
378 if XNN_LIKELY(nc >= 32) {
379 _mm512_storeu_ps(c6 + 0, vscaled6x0123456789ABCDEF);
380 _mm512_storeu_ps(c6 + 16, vscaled6xGHIJKLMNOPQRSTUV);
381 c6 = (float*) ((uintptr_t) c6 + cn_stride);
382 _mm512_storeu_ps(c5 + 0, vscaled5x0123456789ABCDEF);
383 _mm512_storeu_ps(c5 + 16, vscaled5xGHIJKLMNOPQRSTUV);
384 c5 = (float*) ((uintptr_t) c5 + cn_stride);
385 _mm512_storeu_ps(c4 + 0, vscaled4x0123456789ABCDEF);
386 _mm512_storeu_ps(c4 + 16, vscaled4xGHIJKLMNOPQRSTUV);
387 c4 = (float*) ((uintptr_t) c4 + cn_stride);
388 _mm512_storeu_ps(c3 + 0, vscaled3x0123456789ABCDEF);
389 _mm512_storeu_ps(c3 + 16, vscaled3xGHIJKLMNOPQRSTUV);
390 c3 = (float*) ((uintptr_t) c3 + cn_stride);
391 _mm512_storeu_ps(c2 + 0, vscaled2x0123456789ABCDEF);
392 _mm512_storeu_ps(c2 + 16, vscaled2xGHIJKLMNOPQRSTUV);
393 c2 = (float*) ((uintptr_t) c2 + cn_stride);
394 _mm512_storeu_ps(c1 + 0, vscaled1x0123456789ABCDEF);
395 _mm512_storeu_ps(c1 + 16, vscaled1xGHIJKLMNOPQRSTUV);
396 c1 = (float*) ((uintptr_t) c1 + cn_stride);
397 _mm512_storeu_ps(c0 + 0, vscaled0x0123456789ABCDEF);
398 _mm512_storeu_ps(c0 + 16, vscaled0xGHIJKLMNOPQRSTUV);
399 c0 = (float*) ((uintptr_t) c0 + cn_stride);
400
401 a = (const int8_t**restrict) ((uintptr_t) a - ks);
402 nc -= 32;
403 } else {
404 // Prepare mask for valid 32-bit elements (depends on nc).
405 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
406 const __mmask16 vmask1 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 16) & 0xFFFF));
407 _mm512_mask_storeu_ps(c6 + 0, vmask0, vscaled6x0123456789ABCDEF);
408 _mm512_mask_storeu_ps(c6 + 16, vmask1, vscaled6xGHIJKLMNOPQRSTUV);
409 _mm512_mask_storeu_ps(c5 + 0, vmask0, vscaled5x0123456789ABCDEF);
410 _mm512_mask_storeu_ps(c5 + 16, vmask1, vscaled5xGHIJKLMNOPQRSTUV);
411 _mm512_mask_storeu_ps(c4 + 0, vmask0, vscaled4x0123456789ABCDEF);
412 _mm512_mask_storeu_ps(c4 + 16, vmask1, vscaled4xGHIJKLMNOPQRSTUV);
413 _mm512_mask_storeu_ps(c3 + 0, vmask0, vscaled3x0123456789ABCDEF);
414 _mm512_mask_storeu_ps(c3 + 16, vmask1, vscaled3xGHIJKLMNOPQRSTUV);
415 _mm512_mask_storeu_ps(c2 + 0, vmask0, vscaled2x0123456789ABCDEF);
416 _mm512_mask_storeu_ps(c2 + 16, vmask1, vscaled2xGHIJKLMNOPQRSTUV);
417 _mm512_mask_storeu_ps(c1 + 0, vmask0, vscaled1x0123456789ABCDEF);
418 _mm512_mask_storeu_ps(c1 + 16, vmask1, vscaled1xGHIJKLMNOPQRSTUV);
419 _mm512_mask_storeu_ps(c0 + 0, vmask0, vscaled0x0123456789ABCDEF);
420 _mm512_mask_storeu_ps(c0 + 16, vmask1, vscaled0xGHIJKLMNOPQRSTUV);
421 nc = 0;
422 }
423 } while (nc != 0);
424
425 // Release tile config
426 // _tile_release();
427 __asm__ volatile ("tilerelease" ::);
428 #endif // defined(__x86_64__)
429 }