Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-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_qc4w_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 struct xnn_f32_qc4w_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 __attribute__((aligned(64))) int8_t weight_buffer[16 * 64];
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 const __m512i vmask = _mm512_set1_epi8(0xF0);
154 const __m512i vshl4 = _mm512_set1_epi64(0x01020408);
155 XNN_FORCE_REALIZATION(vmask);
156 XNN_FORCE_REALIZATION(vshl4);
157
158 do {
159 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32((const int32_t*) w + 0);
160 w = (const int32_t*) w + 16;
161
162 // Zero tile accumulator
163 __asm__ volatile (
164 "tilezero %%tmm0\n"
165 ::);
166
167 size_t k = kc;
168 while (k >= 64 * sizeof(int8_t)) {
169 _tile_loadd(4, a, a_stride);
170 a += 64;
171 const __m512i vb0x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 0);
172 const __m512i vb1x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 64);
173 const __m512i vb2x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 128);
174 const __m512i vb3x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 192);
175 const __m512i vb4x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 256);
176 const __m512i vb5x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 320);
177 const __m512i vb6x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 384);
178 const __m512i vb7x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 448);
179 const __m512i vl0x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb0x0123456789ABCDEF, vshl4, 0);
180 const __m512i vh0x0123456789ABCDEF = _mm512_and_si512(vb0x0123456789ABCDEF, vmask);
181 const __m512i vl1x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb1x0123456789ABCDEF, vshl4, 0);
182 const __m512i vh1x0123456789ABCDEF = _mm512_and_si512(vb1x0123456789ABCDEF, vmask);
183 const __m512i vl2x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb2x0123456789ABCDEF, vshl4, 0);
184 const __m512i vh2x0123456789ABCDEF = _mm512_and_si512(vb2x0123456789ABCDEF, vmask);
185 const __m512i vl3x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb3x0123456789ABCDEF, vshl4, 0);
186 const __m512i vh3x0123456789ABCDEF = _mm512_and_si512(vb3x0123456789ABCDEF, vmask);
187 const __m512i vl4x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb4x0123456789ABCDEF, vshl4, 0);
188 const __m512i vh4x0123456789ABCDEF = _mm512_and_si512(vb4x0123456789ABCDEF, vmask);
189 const __m512i vl5x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb5x0123456789ABCDEF, vshl4, 0);
190 const __m512i vh5x0123456789ABCDEF = _mm512_and_si512(vb5x0123456789ABCDEF, vmask);
191 const __m512i vl6x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb6x0123456789ABCDEF, vshl4, 0);
192 const __m512i vh6x0123456789ABCDEF = _mm512_and_si512(vb6x0123456789ABCDEF, vmask);
193 const __m512i vl7x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb7x0123456789ABCDEF, vshl4, 0);
194 const __m512i vh7x0123456789ABCDEF = _mm512_and_si512(vb7x0123456789ABCDEF, vmask);
195 _mm512_store_epi32(weight_buffer + 0, vl0x0123456789ABCDEF);
196 _mm512_store_epi32(weight_buffer + 64, vh0x0123456789ABCDEF);
197 _mm512_store_epi32(weight_buffer + 128, vl1x0123456789ABCDEF);
198 _mm512_store_epi32(weight_buffer + 192, vh1x0123456789ABCDEF);
199 _mm512_store_epi32(weight_buffer + 256, vl2x0123456789ABCDEF);
200 _mm512_store_epi32(weight_buffer + 320, vh2x0123456789ABCDEF);
201 _mm512_store_epi32(weight_buffer + 384, vl3x0123456789ABCDEF);
202 _mm512_store_epi32(weight_buffer + 448, vh3x0123456789ABCDEF);
203 _mm512_store_epi32(weight_buffer + 512, vl4x0123456789ABCDEF);
204 _mm512_store_epi32(weight_buffer + 576, vh4x0123456789ABCDEF);
205 _mm512_store_epi32(weight_buffer + 640, vl5x0123456789ABCDEF);
206 _mm512_store_epi32(weight_buffer + 704, vh5x0123456789ABCDEF);
207 _mm512_store_epi32(weight_buffer + 768, vl6x0123456789ABCDEF);
208 _mm512_store_epi32(weight_buffer + 832, vh6x0123456789ABCDEF);
209 _mm512_store_epi32(weight_buffer + 896, vl7x0123456789ABCDEF);
210 _mm512_store_epi32(weight_buffer + 960, vh7x0123456789ABCDEF);
211 _tile_loadd(5, weight_buffer, 64);
212 _tile_dpbssd(0, 4, 5);
213
214 w = (const int8_t*) w + 512;
215 k -= 64 * sizeof(int8_t);
216 }
217
218 if XNN_UNLIKELY(k != 0) {
219 _tile_loadd(6, a, a_stride);
220 a += kremainder;
221 for (size_t k = 0; k < ((kremainder + 7) >> 3); ++k) {
222 const __m512i vb0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 0 + 64 * k);
223 const __m512i vl0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb0123456789ABCDEF, vshl4, 0);
224 const __m512i vh0123456789ABCDEF = _mm512_and_si512(vb0123456789ABCDEF, vmask);
225 _mm512_store_epi32(weight_buffer + 128 * k + 0, vl0123456789ABCDEF);
226 _mm512_store_epi32(weight_buffer + 128 * k + 64, vh0123456789ABCDEF);
227 }
228 _tile_loadd(7, weight_buffer, 64);
229 _tile_dpbssd(0, 6, 7);
230
231 w = (const int8_t*) w + ((kremainder + 7) >> 3) * 64;
232 k -= kremainder * sizeof(int8_t);
233 }
234
235 // Add tile to bias
236 _tile_stored(0, res0, 64);
237
238 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[0].zero_point));
239 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[1].zero_point));
240 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[2].zero_point));
241 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[3].zero_point));
242 __m512i vacc4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[4].zero_point));
243 __m512i vacc5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[5].zero_point));
244 __m512i vacc6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[6].zero_point));
245 __m512i vacc7x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[7].zero_point));
246 __m512i vacc8x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[8].zero_point));
247 __m512i vacc9x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[9].zero_point));
248 __m512i vacc10x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[10].zero_point));
249 __m512i vacc11x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[11].zero_point));
250 __m512i vacc12x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[12].zero_point));
251 __m512i vacc13x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[13].zero_point));
252 __m512i vacc14x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[14].zero_point));
253 __m512i vacc15x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[15].zero_point));
254 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
255 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, _mm512_load_epi32(res0 + 16));
256 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, _mm512_load_epi32(res0 + 32));
257 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, _mm512_load_epi32(res0 + 48));
258 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, _mm512_load_epi32(res0 + 64));
259 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, _mm512_load_epi32(res0 + 80));
260 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, _mm512_load_epi32(res0 + 96));
261 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, _mm512_load_epi32(res0 + 112));
262 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, _mm512_load_epi32(res0 + 128));
263 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, _mm512_load_epi32(res0 + 144));
264 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, _mm512_load_epi32(res0 + 160));
265 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, _mm512_load_epi32(res0 + 176));
266 vacc12x0123456789ABCDEF = _mm512_add_epi32(vacc12x0123456789ABCDEF, _mm512_load_epi32(res0 + 192));
267 vacc13x0123456789ABCDEF = _mm512_add_epi32(vacc13x0123456789ABCDEF, _mm512_load_epi32(res0 + 208));
268 vacc14x0123456789ABCDEF = _mm512_add_epi32(vacc14x0123456789ABCDEF, _mm512_load_epi32(res0 + 224));
269 vacc15x0123456789ABCDEF = _mm512_add_epi32(vacc15x0123456789ABCDEF, _mm512_load_epi32(res0 + 240));
270
271 vacc0x0123456789ABCDEF = _mm512_srai_epi32(vacc0x0123456789ABCDEF, 4);
272 vacc1x0123456789ABCDEF = _mm512_srai_epi32(vacc1x0123456789ABCDEF, 4);
273 vacc2x0123456789ABCDEF = _mm512_srai_epi32(vacc2x0123456789ABCDEF, 4);
274 vacc3x0123456789ABCDEF = _mm512_srai_epi32(vacc3x0123456789ABCDEF, 4);
275 vacc4x0123456789ABCDEF = _mm512_srai_epi32(vacc4x0123456789ABCDEF, 4);
276 vacc5x0123456789ABCDEF = _mm512_srai_epi32(vacc5x0123456789ABCDEF, 4);
277 vacc6x0123456789ABCDEF = _mm512_srai_epi32(vacc6x0123456789ABCDEF, 4);
278 vacc7x0123456789ABCDEF = _mm512_srai_epi32(vacc7x0123456789ABCDEF, 4);
279 vacc8x0123456789ABCDEF = _mm512_srai_epi32(vacc8x0123456789ABCDEF, 4);
280 vacc9x0123456789ABCDEF = _mm512_srai_epi32(vacc9x0123456789ABCDEF, 4);
281 vacc10x0123456789ABCDEF = _mm512_srai_epi32(vacc10x0123456789ABCDEF, 4);
282 vacc11x0123456789ABCDEF = _mm512_srai_epi32(vacc11x0123456789ABCDEF, 4);
283 vacc12x0123456789ABCDEF = _mm512_srai_epi32(vacc12x0123456789ABCDEF, 4);
284 vacc13x0123456789ABCDEF = _mm512_srai_epi32(vacc13x0123456789ABCDEF, 4);
285 vacc14x0123456789ABCDEF = _mm512_srai_epi32(vacc14x0123456789ABCDEF, 4);
286 vacc15x0123456789ABCDEF = _mm512_srai_epi32(vacc15x0123456789ABCDEF, 4);
287 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
288 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
289 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
290 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
291 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
292 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
293 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
294 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
295 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
296 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
297 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
298 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
299 __m512 vscaled12x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc12x0123456789ABCDEF);
300 __m512 vscaled13x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc13x0123456789ABCDEF);
301 __m512 vscaled14x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc14x0123456789ABCDEF);
302 __m512 vscaled15x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc15x0123456789ABCDEF);
303
304 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
305 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
306 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
307 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
308 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
309 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
310 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
311 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
312 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, _mm512_set1_ps(quantization_params[8].inv_scale));
313 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, _mm512_set1_ps(quantization_params[9].inv_scale));
314 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, _mm512_set1_ps(quantization_params[10].inv_scale));
315 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, _mm512_set1_ps(quantization_params[11].inv_scale));
316 vscaled12x0123456789ABCDEF = _mm512_mul_ps(vscaled12x0123456789ABCDEF, _mm512_set1_ps(quantization_params[12].inv_scale));
317 vscaled13x0123456789ABCDEF = _mm512_mul_ps(vscaled13x0123456789ABCDEF, _mm512_set1_ps(quantization_params[13].inv_scale));
318 vscaled14x0123456789ABCDEF = _mm512_mul_ps(vscaled14x0123456789ABCDEF, _mm512_set1_ps(quantization_params[14].inv_scale));
319 vscaled15x0123456789ABCDEF = _mm512_mul_ps(vscaled15x0123456789ABCDEF, _mm512_set1_ps(quantization_params[15].inv_scale));
320
321 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w + 0);
322 w = (const int32_t*) w + 16;
323 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 0);
324 w = (const int32_t*) w + 16;
325
326 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
327 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
328 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
329 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
330 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
331 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
332 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
333 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
334 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
335 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vscaled9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
336 vscaled10x0123456789ABCDEF = _mm512_fmadd_ps(vscaled10x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
337 vscaled11x0123456789ABCDEF = _mm512_fmadd_ps(vscaled11x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
338 vscaled12x0123456789ABCDEF = _mm512_fmadd_ps(vscaled12x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
339 vscaled13x0123456789ABCDEF = _mm512_fmadd_ps(vscaled13x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
340 vscaled14x0123456789ABCDEF = _mm512_fmadd_ps(vscaled14x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
341 vscaled15x0123456789ABCDEF = _mm512_fmadd_ps(vscaled15x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
342
343 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
344 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
345 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
346 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
347 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
348 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
349 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
350 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
351 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
352 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
353 vscaled10x0123456789ABCDEF = _mm512_max_ps(vscaled10x0123456789ABCDEF, voutput_min);
354 vscaled11x0123456789ABCDEF = _mm512_max_ps(vscaled11x0123456789ABCDEF, voutput_min);
355 vscaled12x0123456789ABCDEF = _mm512_max_ps(vscaled12x0123456789ABCDEF, voutput_min);
356 vscaled13x0123456789ABCDEF = _mm512_max_ps(vscaled13x0123456789ABCDEF, voutput_min);
357 vscaled14x0123456789ABCDEF = _mm512_max_ps(vscaled14x0123456789ABCDEF, voutput_min);
358 vscaled15x0123456789ABCDEF = _mm512_max_ps(vscaled15x0123456789ABCDEF, voutput_min);
359
360 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
361 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
362 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
363 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
364 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
365 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
366 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
367 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
368 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
369 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
370 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max);
371 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max);
372 vscaled12x0123456789ABCDEF = _mm512_min_ps(vscaled12x0123456789ABCDEF, voutput_max);
373 vscaled13x0123456789ABCDEF = _mm512_min_ps(vscaled13x0123456789ABCDEF, voutput_max);
374 vscaled14x0123456789ABCDEF = _mm512_min_ps(vscaled14x0123456789ABCDEF, voutput_max);
375 vscaled15x0123456789ABCDEF = _mm512_min_ps(vscaled15x0123456789ABCDEF, voutput_max);
376
377 if XNN_LIKELY(nc >= 16) {
378 _mm512_storeu_ps(c15 + 0, vscaled15x0123456789ABCDEF);
379 c15 = (float*) ((uintptr_t) c15 + cn_stride);
380 _mm512_storeu_ps(c14 + 0, vscaled14x0123456789ABCDEF);
381 c14 = (float*) ((uintptr_t) c14 + cn_stride);
382 _mm512_storeu_ps(c13 + 0, vscaled13x0123456789ABCDEF);
383 c13 = (float*) ((uintptr_t) c13 + cn_stride);
384 _mm512_storeu_ps(c12 + 0, vscaled12x0123456789ABCDEF);
385 c12 = (float*) ((uintptr_t) c12 + cn_stride);
386 _mm512_storeu_ps(c11 + 0, vscaled11x0123456789ABCDEF);
387 c11 = (float*) ((uintptr_t) c11 + cn_stride);
388 _mm512_storeu_ps(c10 + 0, vscaled10x0123456789ABCDEF);
389 c10 = (float*) ((uintptr_t) c10 + cn_stride);
390 _mm512_storeu_ps(c9 + 0, vscaled9x0123456789ABCDEF);
391 c9 = (float*) ((uintptr_t) c9 + cn_stride);
392 _mm512_storeu_ps(c8 + 0, vscaled8x0123456789ABCDEF);
393 c8 = (float*) ((uintptr_t) c8 + cn_stride);
394 _mm512_storeu_ps(c7 + 0, vscaled7x0123456789ABCDEF);
395 c7 = (float*) ((uintptr_t) c7 + cn_stride);
396 _mm512_storeu_ps(c6 + 0, vscaled6x0123456789ABCDEF);
397 c6 = (float*) ((uintptr_t) c6 + cn_stride);
398 _mm512_storeu_ps(c5 + 0, vscaled5x0123456789ABCDEF);
399 c5 = (float*) ((uintptr_t) c5 + cn_stride);
400 _mm512_storeu_ps(c4 + 0, vscaled4x0123456789ABCDEF);
401 c4 = (float*) ((uintptr_t) c4 + cn_stride);
402 _mm512_storeu_ps(c3 + 0, vscaled3x0123456789ABCDEF);
403 c3 = (float*) ((uintptr_t) c3 + cn_stride);
404 _mm512_storeu_ps(c2 + 0, vscaled2x0123456789ABCDEF);
405 c2 = (float*) ((uintptr_t) c2 + cn_stride);
406 _mm512_storeu_ps(c1 + 0, vscaled1x0123456789ABCDEF);
407 c1 = (float*) ((uintptr_t) c1 + cn_stride);
408 _mm512_storeu_ps(c0 + 0, vscaled0x0123456789ABCDEF);
409 c0 = (float*) ((uintptr_t) c0 + cn_stride);
410
411 a -= kc;
412 nc -= 16;
413 } else {
414 // Prepare mask for valid 32-bit elements (depends on nc).
415 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
416 _mm512_mask_storeu_ps(c15 + 0, vmask0, vscaled15x0123456789ABCDEF);
417 _mm512_mask_storeu_ps(c14 + 0, vmask0, vscaled14x0123456789ABCDEF);
418 _mm512_mask_storeu_ps(c13 + 0, vmask0, vscaled13x0123456789ABCDEF);
419 _mm512_mask_storeu_ps(c12 + 0, vmask0, vscaled12x0123456789ABCDEF);
420 _mm512_mask_storeu_ps(c11 + 0, vmask0, vscaled11x0123456789ABCDEF);
421 _mm512_mask_storeu_ps(c10 + 0, vmask0, vscaled10x0123456789ABCDEF);
422 _mm512_mask_storeu_ps(c9 + 0, vmask0, vscaled9x0123456789ABCDEF);
423 _mm512_mask_storeu_ps(c8 + 0, vmask0, vscaled8x0123456789ABCDEF);
424 _mm512_mask_storeu_ps(c7 + 0, vmask0, vscaled7x0123456789ABCDEF);
425 _mm512_mask_storeu_ps(c6 + 0, vmask0, vscaled6x0123456789ABCDEF);
426 _mm512_mask_storeu_ps(c5 + 0, vmask0, vscaled5x0123456789ABCDEF);
427 _mm512_mask_storeu_ps(c4 + 0, vmask0, vscaled4x0123456789ABCDEF);
428 _mm512_mask_storeu_ps(c3 + 0, vmask0, vscaled3x0123456789ABCDEF);
429 _mm512_mask_storeu_ps(c2 + 0, vmask0, vscaled2x0123456789ABCDEF);
430 _mm512_mask_storeu_ps(c1 + 0, vmask0, vscaled1x0123456789ABCDEF);
431 _mm512_mask_storeu_ps(c0 + 0, vmask0, vscaled0x0123456789ABCDEF);
432 nc = 0;
433 }
434 } while (nc != 0);
435 // Release tile config
436 // _tile_release();
437 __asm__ volatile ("tilerelease" ::);
438 #endif // defined(__x86_64__)
439 }