Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-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_qc4w_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 struct xnn_f32_qc4w_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 __attribute__((aligned(64))) int8_t weight_buffer[16 * 64];
49
50 kc = round_up_po2(kc, 4 * sizeof(int8_t));
51 const size_t kremainder = (kc & 63) ? (kc & 63) : 64;
52
53 // Define tile config data structure
54 struct __tile_config {
55 uint8_t palette_id;
56 uint8_t start_row;
57 uint8_t reserved_0[14];
58 uint16_t colsb[8];
59 uint16_t reserved_1[8];
60 uint8_t rows[8];
61 uint8_t reserved_2[8];
62 };
63
64 // Load tile configuration
65 __attribute__((aligned(64))) struct __tile_config tile_data = {0};
66 tile_data.palette_id = 1;
67 tile_data.rows[0] = mr; // tmm0 = res 0
68 tile_data.rows[1] = mr; // tmm1 = res 1
69 tile_data.rows[2] = mr; // tmm2 = res 2
70 tile_data.rows[3] = mr; // tmm3 = res 3
71 tile_data.rows[4] = mr; // tmm4 = input
72 tile_data.rows[5] = 16; // tmm5 = weights
73 tile_data.rows[6] = mr; // tmm6 = input remainder
74 tile_data.rows[7] = kremainder >> 2; // tmm7 = weights remainder
75
76 tile_data.colsb[0] = 64; // tmm0 = res 0
77 tile_data.colsb[1] = 64; // tmm1 = res 1
78 tile_data.colsb[2] = 64; // tmm2 = res 1
79 tile_data.colsb[3] = 64; // tmm3 = res 1
80 tile_data.colsb[4] = 64; // tmm4 = input
81 tile_data.colsb[5] = 64; // tmm5 = weights
82 tile_data.colsb[6] = kremainder; // tmm6 = input remainder
83 tile_data.colsb[7] = 64; // tmm7 = weights remainder
84
85 //_tile_loadconfig(&tile_data);
86 __asm__ volatile ("ldtilecfg %0" :: "m" (tile_data));
87
88 float* c0 = c;
89 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
90 if XNN_UNPREDICTABLE(mr < 2) {
91 c1 = c0;
92 }
93 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
94 if XNN_UNPREDICTABLE(mr <= 2) {
95 c2 = c1;
96 }
97 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
98 if XNN_UNPREDICTABLE(mr < 4) {
99 c3 = c2;
100 }
101 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
102 if XNN_UNPREDICTABLE(mr <= 4) {
103 c4 = c3;
104 }
105 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
106 if XNN_UNPREDICTABLE(mr < 6) {
107 c5 = c4;
108 }
109 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
110 if XNN_UNPREDICTABLE(mr <= 6) {
111 c6 = c5;
112 }
113 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
114 if XNN_UNPREDICTABLE(mr < 8) {
115 c7 = c6;
116 }
117 float* c8 = (float*) ((uintptr_t) c7 + cm_stride);
118 if XNN_UNPREDICTABLE(mr <= 8) {
119 c8 = c7;
120 }
121 float* c9 = (float*) ((uintptr_t) c8 + cm_stride);
122 if XNN_UNPREDICTABLE(mr < 10) {
123 c9 = c8;
124 }
125 float* c10 = (float*) ((uintptr_t) c9 + cm_stride);
126 if XNN_UNPREDICTABLE(mr <= 10) {
127 c10 = c9;
128 }
129 float* c11 = (float*) ((uintptr_t) c10 + cm_stride);
130 if XNN_UNPREDICTABLE(mr < 12) {
131 c11 = c10;
132 }
133 float* c12 = (float*) ((uintptr_t) c11 + cm_stride);
134 if XNN_UNPREDICTABLE(mr <= 12) {
135 c12 = c11;
136 }
137 float* c13 = (float*) ((uintptr_t) c12 + cm_stride);
138 if XNN_UNPREDICTABLE(mr < 14) {
139 c13 = c12;
140 }
141 float* c14 = (float*) ((uintptr_t) c13 + cm_stride);
142 if XNN_UNPREDICTABLE(mr <= 14) {
143 c14 = c13;
144 }
145 float* c15 = (float*) ((uintptr_t) c14 + cm_stride);
146 if XNN_UNPREDICTABLE(mr != 16) {
147 c15 = c14;
148 }
149
150 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
151 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
152 // XNN_FORCE_REALIZATION(voutput_min);
153 // XNN_FORCE_REALIZATION(voutput_max);
154 const __m512i vmask = _mm512_set1_epi8(0xF0);
155 const __m512i vshl4 = _mm512_set1_epi64(0x01020408);
156 XNN_FORCE_REALIZATION(vmask);
157 XNN_FORCE_REALIZATION(vshl4);
158
159 do {
160 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32((const int32_t*) w + 0);
161 w = (const int32_t*) w + 16;
162
163 // Zero tile accumulator
164 __asm__ volatile (
165 "tilezero %%tmm0\n"
166 ::);
167
168 size_t k = kc;
169 while (k >= 64 * sizeof(int8_t)) {
170 _tile_loadd(4, a, a_stride);
171 a += 64;
172 const __m512i vb0x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 0);
173 const __m512i vb1x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 64);
174 const __m512i vb2x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 128);
175 const __m512i vb3x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 192);
176 const __m512i vb4x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 256);
177 const __m512i vb5x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 320);
178 const __m512i vb6x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 384);
179 const __m512i vb7x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 448);
180 const __m512i vl0x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb0x0123456789ABCDEF, vshl4, 0);
181 const __m512i vh0x0123456789ABCDEF = _mm512_and_si512(vb0x0123456789ABCDEF, vmask);
182 const __m512i vl1x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb1x0123456789ABCDEF, vshl4, 0);
183 const __m512i vh1x0123456789ABCDEF = _mm512_and_si512(vb1x0123456789ABCDEF, vmask);
184 const __m512i vl2x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb2x0123456789ABCDEF, vshl4, 0);
185 const __m512i vh2x0123456789ABCDEF = _mm512_and_si512(vb2x0123456789ABCDEF, vmask);
186 const __m512i vl3x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb3x0123456789ABCDEF, vshl4, 0);
187 const __m512i vh3x0123456789ABCDEF = _mm512_and_si512(vb3x0123456789ABCDEF, vmask);
188 const __m512i vl4x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb4x0123456789ABCDEF, vshl4, 0);
189 const __m512i vh4x0123456789ABCDEF = _mm512_and_si512(vb4x0123456789ABCDEF, vmask);
190 const __m512i vl5x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb5x0123456789ABCDEF, vshl4, 0);
191 const __m512i vh5x0123456789ABCDEF = _mm512_and_si512(vb5x0123456789ABCDEF, vmask);
192 const __m512i vl6x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb6x0123456789ABCDEF, vshl4, 0);
193 const __m512i vh6x0123456789ABCDEF = _mm512_and_si512(vb6x0123456789ABCDEF, vmask);
194 const __m512i vl7x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb7x0123456789ABCDEF, vshl4, 0);
195 const __m512i vh7x0123456789ABCDEF = _mm512_and_si512(vb7x0123456789ABCDEF, vmask);
196 _mm512_store_epi32(weight_buffer + 0, vl0x0123456789ABCDEF);
197 _mm512_store_epi32(weight_buffer + 64, vh0x0123456789ABCDEF);
198 _mm512_store_epi32(weight_buffer + 128, vl1x0123456789ABCDEF);
199 _mm512_store_epi32(weight_buffer + 192, vh1x0123456789ABCDEF);
200 _mm512_store_epi32(weight_buffer + 256, vl2x0123456789ABCDEF);
201 _mm512_store_epi32(weight_buffer + 320, vh2x0123456789ABCDEF);
202 _mm512_store_epi32(weight_buffer + 384, vl3x0123456789ABCDEF);
203 _mm512_store_epi32(weight_buffer + 448, vh3x0123456789ABCDEF);
204 _mm512_store_epi32(weight_buffer + 512, vl4x0123456789ABCDEF);
205 _mm512_store_epi32(weight_buffer + 576, vh4x0123456789ABCDEF);
206 _mm512_store_epi32(weight_buffer + 640, vl5x0123456789ABCDEF);
207 _mm512_store_epi32(weight_buffer + 704, vh5x0123456789ABCDEF);
208 _mm512_store_epi32(weight_buffer + 768, vl6x0123456789ABCDEF);
209 _mm512_store_epi32(weight_buffer + 832, vh6x0123456789ABCDEF);
210 _mm512_store_epi32(weight_buffer + 896, vl7x0123456789ABCDEF);
211 _mm512_store_epi32(weight_buffer + 960, vh7x0123456789ABCDEF);
212 _tile_loadd(5, weight_buffer, 64);
213 _tile_dpbssd(0, 4, 5);
214 xnn_prefetch_to_l1((const int8_t*) w + 4096);
215 xnn_prefetch_to_l1((const int8_t*) w + 4160);
216 xnn_prefetch_to_l1((const int8_t*) w + 4224);
217 xnn_prefetch_to_l1((const int8_t*) w + 4288);
218 xnn_prefetch_to_l1((const int8_t*) w + 4352);
219 xnn_prefetch_to_l1((const int8_t*) w + 4416);
220 xnn_prefetch_to_l1((const int8_t*) w + 4480);
221 xnn_prefetch_to_l1((const int8_t*) w + 4544);
222 xnn_prefetch_to_l1((const int8_t*) w + 4608);
223 xnn_prefetch_to_l1((const int8_t*) w + 4672);
224 xnn_prefetch_to_l1((const int8_t*) w + 4736);
225 xnn_prefetch_to_l1((const int8_t*) w + 4800);
226 xnn_prefetch_to_l1((const int8_t*) w + 4864);
227 xnn_prefetch_to_l1((const int8_t*) w + 4928);
228 xnn_prefetch_to_l1((const int8_t*) w + 4992);
229 xnn_prefetch_to_l1((const int8_t*) w + 5056);
230 xnn_prefetch_to_l1((const int8_t*) w + 5120);
231 xnn_prefetch_to_l1((const int8_t*) w + 5184);
232 xnn_prefetch_to_l1((const int8_t*) w + 5248);
233 xnn_prefetch_to_l1((const int8_t*) w + 5312);
234 xnn_prefetch_to_l1((const int8_t*) w + 5376);
235 xnn_prefetch_to_l1((const int8_t*) w + 5440);
236 xnn_prefetch_to_l1((const int8_t*) w + 5504);
237 xnn_prefetch_to_l1((const int8_t*) w + 5568);
238 xnn_prefetch_to_l1((const int8_t*) w + 5632);
239 xnn_prefetch_to_l1((const int8_t*) w + 5696);
240 xnn_prefetch_to_l1((const int8_t*) w + 5760);
241 xnn_prefetch_to_l1((const int8_t*) w + 5824);
242 xnn_prefetch_to_l1((const int8_t*) w + 5888);
243 xnn_prefetch_to_l1((const int8_t*) w + 5952);
244 xnn_prefetch_to_l1((const int8_t*) w + 6016);
245 xnn_prefetch_to_l1((const int8_t*) w + 6080);
246
247 w = (const int8_t*) w + 512;
248 k -= 64 * sizeof(int8_t);
249 }
250
251 if XNN_UNLIKELY(k != 0) {
252 _tile_loadd(6, a, a_stride);
253 a += kremainder;
254 for (size_t k = 0; k < ((kremainder + 7) >> 3); ++k) {
255 const __m512i vb0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 0 + 64 * k);
256 const __m512i vl0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb0123456789ABCDEF, vshl4, 0);
257 const __m512i vh0123456789ABCDEF = _mm512_and_si512(vb0123456789ABCDEF, vmask);
258 _mm512_store_epi32(weight_buffer + 128 * k + 0, vl0123456789ABCDEF);
259 _mm512_store_epi32(weight_buffer + 128 * k + 64, vh0123456789ABCDEF);
260 }
261 _tile_loadd(7, weight_buffer, 64);
262 _tile_dpbssd(0, 6, 7);
263
264 w = (const int8_t*) w + ((kremainder + 7) >> 3) * 64;
265 k -= kremainder * sizeof(int8_t);
266 }
267
268 // Add tile to bias
269 _tile_stored(0, res0, 64);
270
271 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[0].zero_point));
272 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[1].zero_point));
273 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[2].zero_point));
274 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[3].zero_point));
275 __m512i vacc4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[4].zero_point));
276 __m512i vacc5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[5].zero_point));
277 __m512i vacc6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[6].zero_point));
278 __m512i vacc7x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[7].zero_point));
279 __m512i vacc8x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[8].zero_point));
280 __m512i vacc9x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[9].zero_point));
281 __m512i vacc10x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[10].zero_point));
282 __m512i vacc11x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[11].zero_point));
283 __m512i vacc12x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[12].zero_point));
284 __m512i vacc13x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[13].zero_point));
285 __m512i vacc14x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[14].zero_point));
286 __m512i vacc15x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[15].zero_point));
287 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
288 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, _mm512_load_epi32(res0 + 16));
289 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, _mm512_load_epi32(res0 + 32));
290 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, _mm512_load_epi32(res0 + 48));
291 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, _mm512_load_epi32(res0 + 64));
292 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, _mm512_load_epi32(res0 + 80));
293 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, _mm512_load_epi32(res0 + 96));
294 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, _mm512_load_epi32(res0 + 112));
295 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, _mm512_load_epi32(res0 + 128));
296 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, _mm512_load_epi32(res0 + 144));
297 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, _mm512_load_epi32(res0 + 160));
298 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, _mm512_load_epi32(res0 + 176));
299 vacc12x0123456789ABCDEF = _mm512_add_epi32(vacc12x0123456789ABCDEF, _mm512_load_epi32(res0 + 192));
300 vacc13x0123456789ABCDEF = _mm512_add_epi32(vacc13x0123456789ABCDEF, _mm512_load_epi32(res0 + 208));
301 vacc14x0123456789ABCDEF = _mm512_add_epi32(vacc14x0123456789ABCDEF, _mm512_load_epi32(res0 + 224));
302 vacc15x0123456789ABCDEF = _mm512_add_epi32(vacc15x0123456789ABCDEF, _mm512_load_epi32(res0 + 240));
303
304 vacc0x0123456789ABCDEF = _mm512_srai_epi32(vacc0x0123456789ABCDEF, 4);
305 vacc1x0123456789ABCDEF = _mm512_srai_epi32(vacc1x0123456789ABCDEF, 4);
306 vacc2x0123456789ABCDEF = _mm512_srai_epi32(vacc2x0123456789ABCDEF, 4);
307 vacc3x0123456789ABCDEF = _mm512_srai_epi32(vacc3x0123456789ABCDEF, 4);
308 vacc4x0123456789ABCDEF = _mm512_srai_epi32(vacc4x0123456789ABCDEF, 4);
309 vacc5x0123456789ABCDEF = _mm512_srai_epi32(vacc5x0123456789ABCDEF, 4);
310 vacc6x0123456789ABCDEF = _mm512_srai_epi32(vacc6x0123456789ABCDEF, 4);
311 vacc7x0123456789ABCDEF = _mm512_srai_epi32(vacc7x0123456789ABCDEF, 4);
312 vacc8x0123456789ABCDEF = _mm512_srai_epi32(vacc8x0123456789ABCDEF, 4);
313 vacc9x0123456789ABCDEF = _mm512_srai_epi32(vacc9x0123456789ABCDEF, 4);
314 vacc10x0123456789ABCDEF = _mm512_srai_epi32(vacc10x0123456789ABCDEF, 4);
315 vacc11x0123456789ABCDEF = _mm512_srai_epi32(vacc11x0123456789ABCDEF, 4);
316 vacc12x0123456789ABCDEF = _mm512_srai_epi32(vacc12x0123456789ABCDEF, 4);
317 vacc13x0123456789ABCDEF = _mm512_srai_epi32(vacc13x0123456789ABCDEF, 4);
318 vacc14x0123456789ABCDEF = _mm512_srai_epi32(vacc14x0123456789ABCDEF, 4);
319 vacc15x0123456789ABCDEF = _mm512_srai_epi32(vacc15x0123456789ABCDEF, 4);
320 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
321 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
322 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
323 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
324 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
325 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
326 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
327 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
328 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
329 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
330 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
331 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
332 __m512 vscaled12x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc12x0123456789ABCDEF);
333 __m512 vscaled13x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc13x0123456789ABCDEF);
334 __m512 vscaled14x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc14x0123456789ABCDEF);
335 __m512 vscaled15x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc15x0123456789ABCDEF);
336
337 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
338 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
339 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
340 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
341 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
342 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
343 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
344 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
345 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, _mm512_set1_ps(quantization_params[8].inv_scale));
346 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, _mm512_set1_ps(quantization_params[9].inv_scale));
347 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, _mm512_set1_ps(quantization_params[10].inv_scale));
348 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, _mm512_set1_ps(quantization_params[11].inv_scale));
349 vscaled12x0123456789ABCDEF = _mm512_mul_ps(vscaled12x0123456789ABCDEF, _mm512_set1_ps(quantization_params[12].inv_scale));
350 vscaled13x0123456789ABCDEF = _mm512_mul_ps(vscaled13x0123456789ABCDEF, _mm512_set1_ps(quantization_params[13].inv_scale));
351 vscaled14x0123456789ABCDEF = _mm512_mul_ps(vscaled14x0123456789ABCDEF, _mm512_set1_ps(quantization_params[14].inv_scale));
352 vscaled15x0123456789ABCDEF = _mm512_mul_ps(vscaled15x0123456789ABCDEF, _mm512_set1_ps(quantization_params[15].inv_scale));
353
354 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w + 0);
355 w = (const int32_t*) w + 16;
356 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 0);
357 w = (const int32_t*) w + 16;
358
359 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
360 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
361 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
362 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
363 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
364 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
365 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
366 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
367 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
368 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vscaled9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
369 vscaled10x0123456789ABCDEF = _mm512_fmadd_ps(vscaled10x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
370 vscaled11x0123456789ABCDEF = _mm512_fmadd_ps(vscaled11x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
371 vscaled12x0123456789ABCDEF = _mm512_fmadd_ps(vscaled12x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
372 vscaled13x0123456789ABCDEF = _mm512_fmadd_ps(vscaled13x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
373 vscaled14x0123456789ABCDEF = _mm512_fmadd_ps(vscaled14x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
374 vscaled15x0123456789ABCDEF = _mm512_fmadd_ps(vscaled15x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
375
376 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
377 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
378 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
379 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
380 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
381 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
382 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
383 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
384 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
385 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
386 vscaled10x0123456789ABCDEF = _mm512_max_ps(vscaled10x0123456789ABCDEF, voutput_min);
387 vscaled11x0123456789ABCDEF = _mm512_max_ps(vscaled11x0123456789ABCDEF, voutput_min);
388 vscaled12x0123456789ABCDEF = _mm512_max_ps(vscaled12x0123456789ABCDEF, voutput_min);
389 vscaled13x0123456789ABCDEF = _mm512_max_ps(vscaled13x0123456789ABCDEF, voutput_min);
390 vscaled14x0123456789ABCDEF = _mm512_max_ps(vscaled14x0123456789ABCDEF, voutput_min);
391 vscaled15x0123456789ABCDEF = _mm512_max_ps(vscaled15x0123456789ABCDEF, voutput_min);
392
393 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
394 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
395 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
396 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
397 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
398 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
399 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
400 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
401 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
402 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
403 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max);
404 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max);
405 vscaled12x0123456789ABCDEF = _mm512_min_ps(vscaled12x0123456789ABCDEF, voutput_max);
406 vscaled13x0123456789ABCDEF = _mm512_min_ps(vscaled13x0123456789ABCDEF, voutput_max);
407 vscaled14x0123456789ABCDEF = _mm512_min_ps(vscaled14x0123456789ABCDEF, voutput_max);
408 vscaled15x0123456789ABCDEF = _mm512_min_ps(vscaled15x0123456789ABCDEF, voutput_max);
409
410 if XNN_LIKELY(nc >= 16) {
411 _mm512_storeu_ps(c15 + 0, vscaled15x0123456789ABCDEF);
412 c15 = (float*) ((uintptr_t) c15 + cn_stride);
413 _mm512_storeu_ps(c14 + 0, vscaled14x0123456789ABCDEF);
414 c14 = (float*) ((uintptr_t) c14 + cn_stride);
415 _mm512_storeu_ps(c13 + 0, vscaled13x0123456789ABCDEF);
416 c13 = (float*) ((uintptr_t) c13 + cn_stride);
417 _mm512_storeu_ps(c12 + 0, vscaled12x0123456789ABCDEF);
418 c12 = (float*) ((uintptr_t) c12 + cn_stride);
419 _mm512_storeu_ps(c11 + 0, vscaled11x0123456789ABCDEF);
420 c11 = (float*) ((uintptr_t) c11 + cn_stride);
421 _mm512_storeu_ps(c10 + 0, vscaled10x0123456789ABCDEF);
422 c10 = (float*) ((uintptr_t) c10 + cn_stride);
423 _mm512_storeu_ps(c9 + 0, vscaled9x0123456789ABCDEF);
424 c9 = (float*) ((uintptr_t) c9 + cn_stride);
425 _mm512_storeu_ps(c8 + 0, vscaled8x0123456789ABCDEF);
426 c8 = (float*) ((uintptr_t) c8 + cn_stride);
427 _mm512_storeu_ps(c7 + 0, vscaled7x0123456789ABCDEF);
428 c7 = (float*) ((uintptr_t) c7 + cn_stride);
429 _mm512_storeu_ps(c6 + 0, vscaled6x0123456789ABCDEF);
430 c6 = (float*) ((uintptr_t) c6 + cn_stride);
431 _mm512_storeu_ps(c5 + 0, vscaled5x0123456789ABCDEF);
432 c5 = (float*) ((uintptr_t) c5 + cn_stride);
433 _mm512_storeu_ps(c4 + 0, vscaled4x0123456789ABCDEF);
434 c4 = (float*) ((uintptr_t) c4 + cn_stride);
435 _mm512_storeu_ps(c3 + 0, vscaled3x0123456789ABCDEF);
436 c3 = (float*) ((uintptr_t) c3 + cn_stride);
437 _mm512_storeu_ps(c2 + 0, vscaled2x0123456789ABCDEF);
438 c2 = (float*) ((uintptr_t) c2 + cn_stride);
439 _mm512_storeu_ps(c1 + 0, vscaled1x0123456789ABCDEF);
440 c1 = (float*) ((uintptr_t) c1 + cn_stride);
441 _mm512_storeu_ps(c0 + 0, vscaled0x0123456789ABCDEF);
442 c0 = (float*) ((uintptr_t) c0 + cn_stride);
443
444 a -= kc;
445 nc -= 16;
446 } else {
447 // Prepare mask for valid 32-bit elements (depends on nc).
448 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
449 _mm512_mask_storeu_ps(c15 + 0, vmask0, vscaled15x0123456789ABCDEF);
450 _mm512_mask_storeu_ps(c14 + 0, vmask0, vscaled14x0123456789ABCDEF);
451 _mm512_mask_storeu_ps(c13 + 0, vmask0, vscaled13x0123456789ABCDEF);
452 _mm512_mask_storeu_ps(c12 + 0, vmask0, vscaled12x0123456789ABCDEF);
453 _mm512_mask_storeu_ps(c11 + 0, vmask0, vscaled11x0123456789ABCDEF);
454 _mm512_mask_storeu_ps(c10 + 0, vmask0, vscaled10x0123456789ABCDEF);
455 _mm512_mask_storeu_ps(c9 + 0, vmask0, vscaled9x0123456789ABCDEF);
456 _mm512_mask_storeu_ps(c8 + 0, vmask0, vscaled8x0123456789ABCDEF);
457 _mm512_mask_storeu_ps(c7 + 0, vmask0, vscaled7x0123456789ABCDEF);
458 _mm512_mask_storeu_ps(c6 + 0, vmask0, vscaled6x0123456789ABCDEF);
459 _mm512_mask_storeu_ps(c5 + 0, vmask0, vscaled5x0123456789ABCDEF);
460 _mm512_mask_storeu_ps(c4 + 0, vmask0, vscaled4x0123456789ABCDEF);
461 _mm512_mask_storeu_ps(c3 + 0, vmask0, vscaled3x0123456789ABCDEF);
462 _mm512_mask_storeu_ps(c2 + 0, vmask0, vscaled2x0123456789ABCDEF);
463 _mm512_mask_storeu_ps(c1 + 0, vmask0, vscaled1x0123456789ABCDEF);
464 _mm512_mask_storeu_ps(c0 + 0, vmask0, vscaled0x0123456789ABCDEF);
465 nc = 0;
466 }
467 } while (nc != 0);
468 // Release tile config
469 // _tile_release();
470 __asm__ volatile ("tilerelease" ::);
471 #endif // defined(__x86_64__)
472 }