Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-7x16c8-minmax-avx512skx-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx16c8-avx512skx.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/prefetch.h"
18
19 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_7x16c8__avx512skx_prfm(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
30 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 7);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(int8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 8 * sizeof(int8_t));
42 const int8_t* a0 = a;
43 float* c0 = c;
44 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
45 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 a1 = a0;
48 c1 = c0;
49 }
50 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 a2 = a1;
54 c2 = c1;
55 }
56 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
57 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
58 if XNN_UNPREDICTABLE(mr < 4) {
59 a3 = a2;
60 c3 = c2;
61 }
62 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
63 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
64 if XNN_UNPREDICTABLE(mr <= 4) {
65 a4 = a3;
66 c4 = c3;
67 }
68 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
69 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
70 if XNN_UNPREDICTABLE(mr < 6) {
71 a5 = a4;
72 c5 = c4;
73 }
74 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
75 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
76 if XNN_UNPREDICTABLE(mr <= 6) {
77 a6 = a5;
78 c6 = c5;
79 }
80
81 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
82 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point);
83 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point);
84 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point);
85 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point);
86 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point);
87 const __m512i vinput_zero_point5 = _mm512_set1_epi32((int) quantization_params[5].zero_point);
88 const __m512i vinput_zero_point6 = _mm512_set1_epi32((int) quantization_params[6].zero_point);
89 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
90 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
91 // XNN_FORCE_REALIZATION(voutput_min);
92 // XNN_FORCE_REALIZATION(voutput_max);
93 const __m256i vmask = _mm256_set1_epi8(0xF0);
94 XNN_FORCE_REALIZATION(vmask);
95
96 do {
97 const __m512i vksum0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
98 const __m512i vksum4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
99 const __m512i vksum89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
100 const __m512i vksumCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
101
102 __m512i vacc0x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point0);
103 __m512i vacc0x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point0);
104 __m512i vacc0x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point0);
105 __m512i vacc0xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point0);
106 __m512i vacc1x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point1);
107 __m512i vacc1x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point1);
108 __m512i vacc1x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point1);
109 __m512i vacc1xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point1);
110 __m512i vacc2x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point2);
111 __m512i vacc2x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point2);
112 __m512i vacc2x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point2);
113 __m512i vacc2xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point2);
114 __m512i vacc3x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point3);
115 __m512i vacc3x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point3);
116 __m512i vacc3x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point3);
117 __m512i vacc3xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point3);
118 __m512i vacc4x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point4);
119 __m512i vacc4x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point4);
120 __m512i vacc4x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point4);
121 __m512i vacc4xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point4);
122 __m512i vacc5x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point5);
123 __m512i vacc5x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point5);
124 __m512i vacc5x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point5);
125 __m512i vacc5xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point5);
126 __m512i vacc6x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point6);
127 __m512i vacc6x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point6);
128 __m512i vacc6x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point6);
129 __m512i vacc6xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point6);
130 w = (const int32_t*) w + 16;
131
132 size_t k = kc;
133 while (k >= 16 * sizeof(int8_t)) {
134 __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
135 a0 += 8;
136 __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
137 a1 += 8;
138 __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
139 a2 += 8;
140 __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
141 a3 += 8;
142 __m512i va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
143 a4 += 8;
144 __m512i va5 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a5)));
145 a5 += 8;
146 __m512i va6 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a6)));
147 a6 += 8;
148
149 __m256i vbb0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 0));
150 __m256i vbs0123 = _mm256_slli_epi32(vbb0123, 4);
151 __m256i vbm0123 = _mm256_and_si256(vbs0123, vmask);
152 __m512i vb0123 = _mm512_cvtepi8_epi16(vbm0123);
153 xnn_prefetch_to_l1((const int8_t*) w + 896);
154
155 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
156 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
157 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
158 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
159 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
160 vacc5x0123 = _mm512_add_epi32(vacc5x0123, _mm512_madd_epi16(va5, vb0123));
161 vacc6x0123 = _mm512_add_epi32(vacc6x0123, _mm512_madd_epi16(va6, vb0123));
162 __m256i vbb4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
163 __m256i vbs4567 = _mm256_slli_epi32(vbb4567, 4);
164 __m256i vbm4567 = _mm256_and_si256(vbs4567, vmask);
165 __m512i vb4567 = _mm512_cvtepi8_epi16(vbm4567);
166
167 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
168 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
169 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
170 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
171 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
172 vacc5x4567 = _mm512_add_epi32(vacc5x4567, _mm512_madd_epi16(va5, vb4567));
173 vacc6x4567 = _mm512_add_epi32(vacc6x4567, _mm512_madd_epi16(va6, vb4567));
174 __m256i vbb89AB = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
175 __m256i vbs89AB = _mm256_slli_epi32(vbb89AB, 4);
176 __m256i vbm89AB = _mm256_and_si256(vbs89AB, vmask);
177 __m512i vb89AB = _mm512_cvtepi8_epi16(vbm89AB);
178 xnn_prefetch_to_l1((const int8_t*) w + 960);
179
180 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
181 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
182 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
183 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
184 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
185 vacc5x89AB = _mm512_add_epi32(vacc5x89AB, _mm512_madd_epi16(va5, vb89AB));
186 vacc6x89AB = _mm512_add_epi32(vacc6x89AB, _mm512_madd_epi16(va6, vb89AB));
187 __m256i vbbCDEF = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
188 __m256i vbsCDEF = _mm256_slli_epi32(vbbCDEF, 4);
189 __m256i vbmCDEF = _mm256_and_si256(vbsCDEF, vmask);
190 __m512i vbCDEF = _mm512_cvtepi8_epi16(vbmCDEF);
191
192 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
193 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
194 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
195 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
196 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
197 vacc5xCDEF = _mm512_add_epi32(vacc5xCDEF, _mm512_madd_epi16(va5, vbCDEF));
198 vacc6xCDEF = _mm512_add_epi32(vacc6xCDEF, _mm512_madd_epi16(va6, vbCDEF));
199
200 va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
201 a0 += 8;
202 va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
203 a1 += 8;
204 va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
205 a2 += 8;
206 va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
207 a3 += 8;
208 va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
209 a4 += 8;
210 va5 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a5)));
211 a5 += 8;
212 va6 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a6)));
213 a6 += 8;
214
215 vbm0123 = _mm256_and_si256(vbb0123, vmask);
216 vb0123 = _mm512_cvtepi8_epi16(vbm0123);
217
218 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
219 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
220 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
221 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
222 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
223 vacc5x0123 = _mm512_add_epi32(vacc5x0123, _mm512_madd_epi16(va5, vb0123));
224 vacc6x0123 = _mm512_add_epi32(vacc6x0123, _mm512_madd_epi16(va6, vb0123));
225 vbm4567 = _mm256_and_si256(vbb4567, vmask);
226 vb4567 = _mm512_cvtepi8_epi16(vbm4567);
227
228 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
229 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
230 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
231 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
232 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
233 vacc5x4567 = _mm512_add_epi32(vacc5x4567, _mm512_madd_epi16(va5, vb4567));
234 vacc6x4567 = _mm512_add_epi32(vacc6x4567, _mm512_madd_epi16(va6, vb4567));
235 vbm89AB = _mm256_and_si256(vbb89AB, vmask);
236 vb89AB = _mm512_cvtepi8_epi16(vbm89AB);
237
238 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
239 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
240 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
241 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
242 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
243 vacc5x89AB = _mm512_add_epi32(vacc5x89AB, _mm512_madd_epi16(va5, vb89AB));
244 vacc6x89AB = _mm512_add_epi32(vacc6x89AB, _mm512_madd_epi16(va6, vb89AB));
245 vbmCDEF = _mm256_and_si256(vbbCDEF, vmask);
246 vbCDEF = _mm512_cvtepi8_epi16(vbmCDEF);
247
248 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
249 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
250 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
251 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
252 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
253 vacc5xCDEF = _mm512_add_epi32(vacc5xCDEF, _mm512_madd_epi16(va5, vbCDEF));
254 vacc6xCDEF = _mm512_add_epi32(vacc6xCDEF, _mm512_madd_epi16(va6, vbCDEF));
255
256 w = (const int8_t*) w + 128;
257 k -= 16 * sizeof(int8_t);
258 }
259
260 while (k >= 8 * sizeof(int8_t)) {
261 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
262 a0 += 8;
263 const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
264 a1 += 8;
265 const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
266 a2 += 8;
267 const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
268 a3 += 8;
269 const __m512i va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
270 a4 += 8;
271 const __m512i va5 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a5)));
272 a5 += 8;
273 const __m512i va6 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a6)));
274 a6 += 8;
275
276 const __m256i vbb0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 0));
277 const __m256i vbs0123 = _mm256_slli_epi32(vbb0123, 4);
278 const __m256i vbm0123 = _mm256_and_si256(vbs0123, vmask);
279 const __m512i vb0123 = _mm512_cvtepi8_epi16(vbm0123);
280 xnn_prefetch_to_l1((const int8_t*) w + 896);
281
282 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
283 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
284 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
285 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
286 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
287 vacc5x0123 = _mm512_add_epi32(vacc5x0123, _mm512_madd_epi16(va5, vb0123));
288 vacc6x0123 = _mm512_add_epi32(vacc6x0123, _mm512_madd_epi16(va6, vb0123));
289 const __m256i vbb4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
290 const __m256i vbs4567 = _mm256_slli_epi32(vbb4567, 4);
291 const __m256i vbm4567 = _mm256_and_si256(vbs4567, vmask);
292 const __m512i vb4567 = _mm512_cvtepi8_epi16(vbm4567);
293
294 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
295 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
296 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
297 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
298 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
299 vacc5x4567 = _mm512_add_epi32(vacc5x4567, _mm512_madd_epi16(va5, vb4567));
300 vacc6x4567 = _mm512_add_epi32(vacc6x4567, _mm512_madd_epi16(va6, vb4567));
301 const __m256i vbb89AB = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
302 const __m256i vbs89AB = _mm256_slli_epi32(vbb89AB, 4);
303 const __m256i vbm89AB = _mm256_and_si256(vbs89AB, vmask);
304 const __m512i vb89AB = _mm512_cvtepi8_epi16(vbm89AB);
305 xnn_prefetch_to_l1((const int8_t*) w + 960);
306
307 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
308 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
309 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
310 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
311 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
312 vacc5x89AB = _mm512_add_epi32(vacc5x89AB, _mm512_madd_epi16(va5, vb89AB));
313 vacc6x89AB = _mm512_add_epi32(vacc6x89AB, _mm512_madd_epi16(va6, vb89AB));
314 const __m256i vbbCDEF = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
315 const __m256i vbsCDEF = _mm256_slli_epi32(vbbCDEF, 4);
316 const __m256i vbmCDEF = _mm256_and_si256(vbsCDEF, vmask);
317 const __m512i vbCDEF = _mm512_cvtepi8_epi16(vbmCDEF);
318
319 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
320 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
321 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
322 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
323 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
324 vacc5xCDEF = _mm512_add_epi32(vacc5xCDEF, _mm512_madd_epi16(va5, vbCDEF));
325 vacc6xCDEF = _mm512_add_epi32(vacc6xCDEF, _mm512_madd_epi16(va6, vbCDEF));
326
327 w = (const int8_t*) w + 128;
328 k -= 8 * sizeof(int8_t);
329 }
330
331 // Add 4 adjacent sums
332 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
333 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
334 const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
335 const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
336 const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm512_unpackhi_epi32(vacc2x0123, vacc2x4567));
337 const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm512_unpackhi_epi32(vacc2x89AB, vacc2xCDEF));
338 const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm512_unpackhi_epi32(vacc3x0123, vacc3x4567));
339 const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm512_unpackhi_epi32(vacc3x89AB, vacc3xCDEF));
340 const __m512i vacc4x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x0123, vacc4x4567), _mm512_unpackhi_epi32(vacc4x0123, vacc4x4567));
341 const __m512i vacc4x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x89AB, vacc4xCDEF), _mm512_unpackhi_epi32(vacc4x89AB, vacc4xCDEF));
342 const __m512i vacc5x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x0123, vacc5x4567), _mm512_unpackhi_epi32(vacc5x0123, vacc5x4567));
343 const __m512i vacc5x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x89AB, vacc5xCDEF), _mm512_unpackhi_epi32(vacc5x89AB, vacc5xCDEF));
344 const __m512i vacc6x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x0123, vacc6x4567), _mm512_unpackhi_epi32(vacc6x0123, vacc6x4567));
345 const __m512i vacc6x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x89AB, vacc6xCDEF), _mm512_unpackhi_epi32(vacc6x89AB, vacc6xCDEF));
346
347 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
348 const __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
349 const __m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9DAEBF), _mm512_unpackhi_epi32(vacc2x04152637, vacc2x8C9DAEBF));
350 const __m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9DAEBF), _mm512_unpackhi_epi32(vacc3x04152637, vacc3x8C9DAEBF));
351 const __m512i vacc4x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x04152637, vacc4x8C9DAEBF), _mm512_unpackhi_epi32(vacc4x04152637, vacc4x8C9DAEBF));
352 const __m512i vacc5x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x04152637, vacc5x8C9DAEBF), _mm512_unpackhi_epi32(vacc5x04152637, vacc5x8C9DAEBF));
353 const __m512i vacc6x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x04152637, vacc6x8C9DAEBF), _mm512_unpackhi_epi32(vacc6x04152637, vacc6x8C9DAEBF));
354
355 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
356 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
357 __m512i vacc1x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc1x084C195D2A6E3B7F);
358 __m512i vacc2x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc2x084C195D2A6E3B7F);
359 __m512i vacc3x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc3x084C195D2A6E3B7F);
360 __m512i vacc4x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc4x084C195D2A6E3B7F);
361 __m512i vacc5x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc5x084C195D2A6E3B7F);
362 __m512i vacc6x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc6x084C195D2A6E3B7F);
363
364 vacc0x0123456789ABCDEF = _mm512_srai_epi32(vacc0x0123456789ABCDEF, 4);
365 vacc1x0123456789ABCDEF = _mm512_srai_epi32(vacc1x0123456789ABCDEF, 4);
366 vacc2x0123456789ABCDEF = _mm512_srai_epi32(vacc2x0123456789ABCDEF, 4);
367 vacc3x0123456789ABCDEF = _mm512_srai_epi32(vacc3x0123456789ABCDEF, 4);
368 vacc4x0123456789ABCDEF = _mm512_srai_epi32(vacc4x0123456789ABCDEF, 4);
369 vacc5x0123456789ABCDEF = _mm512_srai_epi32(vacc5x0123456789ABCDEF, 4);
370 vacc6x0123456789ABCDEF = _mm512_srai_epi32(vacc6x0123456789ABCDEF, 4);
371 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
372 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
373 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
374 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
375 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
376 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
377 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
378
379 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
380 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
381 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
382 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
383 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
384 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
385 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
386
387 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
388 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
389 w = (const float*) w + 32;
390
391 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
392 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
393 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
394 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
395 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
396 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
397 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
398
399 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
400 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
401 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
402 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
403 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
404 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
405 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
406
407 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
408 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
409 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
410 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
411 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
412 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
413 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
414
415 if XNN_LIKELY(nc >= 16) {
416 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
417 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
418 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
419 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
420 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
421 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
422 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
423
424 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
425 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
426 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
427 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
428 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
429 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
430 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
431
432 c0 = (float*) ((uintptr_t) c0 + cn_stride);
433 c1 = (float*) ((uintptr_t) c1 + cn_stride);
434 c2 = (float*) ((uintptr_t) c2 + cn_stride);
435 c3 = (float*) ((uintptr_t) c3 + cn_stride);
436 c4 = (float*) ((uintptr_t) c4 + cn_stride);
437 c5 = (float*) ((uintptr_t) c5 + cn_stride);
438 c6 = (float*) ((uintptr_t) c6 + cn_stride);
439
440 nc -= 16;
441 } else {
442 // Prepare mask for valid 32-bit elements (depends on nc).
443 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
444 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
445 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
446 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
447 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
448 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
449 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
450 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
451 nc = 0;
452 }
453 } while (nc != 0);
454 }