Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-7x16c8-minmax-avx512skx.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
18 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_7x16c8__avx512skx(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
29 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 7);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 8 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 float* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
62 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
63 if XNN_UNPREDICTABLE(mr <= 4) {
64 a4 = a3;
65 c4 = c3;
66 }
67 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
68 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
69 if XNN_UNPREDICTABLE(mr < 6) {
70 a5 = a4;
71 c5 = c4;
72 }
73 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
74 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
75 if XNN_UNPREDICTABLE(mr <= 6) {
76 a6 = a5;
77 c6 = c5;
78 }
79
80 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
81 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point);
82 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point);
83 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point);
84 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point);
85 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point);
86 const __m512i vinput_zero_point5 = _mm512_set1_epi32((int) quantization_params[5].zero_point);
87 const __m512i vinput_zero_point6 = _mm512_set1_epi32((int) quantization_params[6].zero_point);
88 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
89 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
90 // XNN_FORCE_REALIZATION(voutput_min);
91 // XNN_FORCE_REALIZATION(voutput_max);
92 const __m256i vmask = _mm256_set1_epi8(0xF0);
93 XNN_FORCE_REALIZATION(vmask);
94
95 do {
96 const __m512i vksum0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
97 const __m512i vksum4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
98 const __m512i vksum89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
99 const __m512i vksumCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
100
101 __m512i vacc0x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point0);
102 __m512i vacc0x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point0);
103 __m512i vacc0x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point0);
104 __m512i vacc0xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point0);
105 __m512i vacc1x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point1);
106 __m512i vacc1x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point1);
107 __m512i vacc1x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point1);
108 __m512i vacc1xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point1);
109 __m512i vacc2x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point2);
110 __m512i vacc2x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point2);
111 __m512i vacc2x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point2);
112 __m512i vacc2xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point2);
113 __m512i vacc3x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point3);
114 __m512i vacc3x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point3);
115 __m512i vacc3x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point3);
116 __m512i vacc3xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point3);
117 __m512i vacc4x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point4);
118 __m512i vacc4x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point4);
119 __m512i vacc4x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point4);
120 __m512i vacc4xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point4);
121 __m512i vacc5x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point5);
122 __m512i vacc5x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point5);
123 __m512i vacc5x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point5);
124 __m512i vacc5xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point5);
125 __m512i vacc6x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point6);
126 __m512i vacc6x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point6);
127 __m512i vacc6x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point6);
128 __m512i vacc6xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point6);
129 w = (const int32_t*) w + 16;
130
131 size_t k = kc;
132 while (k >= 16 * sizeof(int8_t)) {
133 __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
134 a0 += 8;
135 __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
136 a1 += 8;
137 __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
138 a2 += 8;
139 __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
140 a3 += 8;
141 __m512i va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
142 a4 += 8;
143 __m512i va5 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a5)));
144 a5 += 8;
145 __m512i va6 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a6)));
146 a6 += 8;
147
148 __m256i vbb0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 0));
149 __m256i vbs0123 = _mm256_slli_epi32(vbb0123, 4);
150 __m256i vbm0123 = _mm256_and_si256(vbs0123, vmask);
151 __m512i vb0123 = _mm512_cvtepi8_epi16(vbm0123);
152
153 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
154 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
155 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
156 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
157 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
158 vacc5x0123 = _mm512_add_epi32(vacc5x0123, _mm512_madd_epi16(va5, vb0123));
159 vacc6x0123 = _mm512_add_epi32(vacc6x0123, _mm512_madd_epi16(va6, vb0123));
160 __m256i vbb4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
161 __m256i vbs4567 = _mm256_slli_epi32(vbb4567, 4);
162 __m256i vbm4567 = _mm256_and_si256(vbs4567, vmask);
163 __m512i vb4567 = _mm512_cvtepi8_epi16(vbm4567);
164
165 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
166 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
167 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
168 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
169 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
170 vacc5x4567 = _mm512_add_epi32(vacc5x4567, _mm512_madd_epi16(va5, vb4567));
171 vacc6x4567 = _mm512_add_epi32(vacc6x4567, _mm512_madd_epi16(va6, vb4567));
172 __m256i vbb89AB = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
173 __m256i vbs89AB = _mm256_slli_epi32(vbb89AB, 4);
174 __m256i vbm89AB = _mm256_and_si256(vbs89AB, vmask);
175 __m512i vb89AB = _mm512_cvtepi8_epi16(vbm89AB);
176
177 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
178 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
179 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
180 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
181 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
182 vacc5x89AB = _mm512_add_epi32(vacc5x89AB, _mm512_madd_epi16(va5, vb89AB));
183 vacc6x89AB = _mm512_add_epi32(vacc6x89AB, _mm512_madd_epi16(va6, vb89AB));
184 __m256i vbbCDEF = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
185 __m256i vbsCDEF = _mm256_slli_epi32(vbbCDEF, 4);
186 __m256i vbmCDEF = _mm256_and_si256(vbsCDEF, vmask);
187 __m512i vbCDEF = _mm512_cvtepi8_epi16(vbmCDEF);
188
189 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
190 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
191 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
192 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
193 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
194 vacc5xCDEF = _mm512_add_epi32(vacc5xCDEF, _mm512_madd_epi16(va5, vbCDEF));
195 vacc6xCDEF = _mm512_add_epi32(vacc6xCDEF, _mm512_madd_epi16(va6, vbCDEF));
196
197 va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
198 a0 += 8;
199 va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
200 a1 += 8;
201 va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
202 a2 += 8;
203 va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
204 a3 += 8;
205 va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
206 a4 += 8;
207 va5 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a5)));
208 a5 += 8;
209 va6 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a6)));
210 a6 += 8;
211
212 vbm0123 = _mm256_and_si256(vbb0123, vmask);
213 vb0123 = _mm512_cvtepi8_epi16(vbm0123);
214
215 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
216 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
217 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
218 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
219 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
220 vacc5x0123 = _mm512_add_epi32(vacc5x0123, _mm512_madd_epi16(va5, vb0123));
221 vacc6x0123 = _mm512_add_epi32(vacc6x0123, _mm512_madd_epi16(va6, vb0123));
222 vbm4567 = _mm256_and_si256(vbb4567, vmask);
223 vb4567 = _mm512_cvtepi8_epi16(vbm4567);
224
225 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
226 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
227 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
228 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
229 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
230 vacc5x4567 = _mm512_add_epi32(vacc5x4567, _mm512_madd_epi16(va5, vb4567));
231 vacc6x4567 = _mm512_add_epi32(vacc6x4567, _mm512_madd_epi16(va6, vb4567));
232 vbm89AB = _mm256_and_si256(vbb89AB, vmask);
233 vb89AB = _mm512_cvtepi8_epi16(vbm89AB);
234
235 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
236 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
237 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
238 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
239 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
240 vacc5x89AB = _mm512_add_epi32(vacc5x89AB, _mm512_madd_epi16(va5, vb89AB));
241 vacc6x89AB = _mm512_add_epi32(vacc6x89AB, _mm512_madd_epi16(va6, vb89AB));
242 vbmCDEF = _mm256_and_si256(vbbCDEF, vmask);
243 vbCDEF = _mm512_cvtepi8_epi16(vbmCDEF);
244
245 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
246 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
247 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
248 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
249 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
250 vacc5xCDEF = _mm512_add_epi32(vacc5xCDEF, _mm512_madd_epi16(va5, vbCDEF));
251 vacc6xCDEF = _mm512_add_epi32(vacc6xCDEF, _mm512_madd_epi16(va6, vbCDEF));
252
253 w = (const int8_t*) w + 128;
254 k -= 16 * sizeof(int8_t);
255 }
256
257 while (k >= 8 * sizeof(int8_t)) {
258 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
259 a0 += 8;
260 const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
261 a1 += 8;
262 const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
263 a2 += 8;
264 const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
265 a3 += 8;
266 const __m512i va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
267 a4 += 8;
268 const __m512i va5 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a5)));
269 a5 += 8;
270 const __m512i va6 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a6)));
271 a6 += 8;
272
273 const __m256i vbb0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 0));
274 const __m256i vbs0123 = _mm256_slli_epi32(vbb0123, 4);
275 const __m256i vbm0123 = _mm256_and_si256(vbs0123, vmask);
276 const __m512i vb0123 = _mm512_cvtepi8_epi16(vbm0123);
277
278 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
279 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
280 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
281 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
282 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
283 vacc5x0123 = _mm512_add_epi32(vacc5x0123, _mm512_madd_epi16(va5, vb0123));
284 vacc6x0123 = _mm512_add_epi32(vacc6x0123, _mm512_madd_epi16(va6, vb0123));
285 const __m256i vbb4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
286 const __m256i vbs4567 = _mm256_slli_epi32(vbb4567, 4);
287 const __m256i vbm4567 = _mm256_and_si256(vbs4567, vmask);
288 const __m512i vb4567 = _mm512_cvtepi8_epi16(vbm4567);
289
290 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
291 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
292 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
293 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
294 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
295 vacc5x4567 = _mm512_add_epi32(vacc5x4567, _mm512_madd_epi16(va5, vb4567));
296 vacc6x4567 = _mm512_add_epi32(vacc6x4567, _mm512_madd_epi16(va6, vb4567));
297 const __m256i vbb89AB = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
298 const __m256i vbs89AB = _mm256_slli_epi32(vbb89AB, 4);
299 const __m256i vbm89AB = _mm256_and_si256(vbs89AB, vmask);
300 const __m512i vb89AB = _mm512_cvtepi8_epi16(vbm89AB);
301
302 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
303 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
304 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
305 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
306 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
307 vacc5x89AB = _mm512_add_epi32(vacc5x89AB, _mm512_madd_epi16(va5, vb89AB));
308 vacc6x89AB = _mm512_add_epi32(vacc6x89AB, _mm512_madd_epi16(va6, vb89AB));
309 const __m256i vbbCDEF = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
310 const __m256i vbsCDEF = _mm256_slli_epi32(vbbCDEF, 4);
311 const __m256i vbmCDEF = _mm256_and_si256(vbsCDEF, vmask);
312 const __m512i vbCDEF = _mm512_cvtepi8_epi16(vbmCDEF);
313
314 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
315 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
316 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
317 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
318 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
319 vacc5xCDEF = _mm512_add_epi32(vacc5xCDEF, _mm512_madd_epi16(va5, vbCDEF));
320 vacc6xCDEF = _mm512_add_epi32(vacc6xCDEF, _mm512_madd_epi16(va6, vbCDEF));
321
322 w = (const int8_t*) w + 128;
323 k -= 8 * sizeof(int8_t);
324 }
325
326 // Add 4 adjacent sums
327 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
328 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
329 const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
330 const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
331 const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm512_unpackhi_epi32(vacc2x0123, vacc2x4567));
332 const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm512_unpackhi_epi32(vacc2x89AB, vacc2xCDEF));
333 const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm512_unpackhi_epi32(vacc3x0123, vacc3x4567));
334 const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm512_unpackhi_epi32(vacc3x89AB, vacc3xCDEF));
335 const __m512i vacc4x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x0123, vacc4x4567), _mm512_unpackhi_epi32(vacc4x0123, vacc4x4567));
336 const __m512i vacc4x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x89AB, vacc4xCDEF), _mm512_unpackhi_epi32(vacc4x89AB, vacc4xCDEF));
337 const __m512i vacc5x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x0123, vacc5x4567), _mm512_unpackhi_epi32(vacc5x0123, vacc5x4567));
338 const __m512i vacc5x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x89AB, vacc5xCDEF), _mm512_unpackhi_epi32(vacc5x89AB, vacc5xCDEF));
339 const __m512i vacc6x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x0123, vacc6x4567), _mm512_unpackhi_epi32(vacc6x0123, vacc6x4567));
340 const __m512i vacc6x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x89AB, vacc6xCDEF), _mm512_unpackhi_epi32(vacc6x89AB, vacc6xCDEF));
341
342 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
343 const __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
344 const __m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9DAEBF), _mm512_unpackhi_epi32(vacc2x04152637, vacc2x8C9DAEBF));
345 const __m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9DAEBF), _mm512_unpackhi_epi32(vacc3x04152637, vacc3x8C9DAEBF));
346 const __m512i vacc4x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x04152637, vacc4x8C9DAEBF), _mm512_unpackhi_epi32(vacc4x04152637, vacc4x8C9DAEBF));
347 const __m512i vacc5x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x04152637, vacc5x8C9DAEBF), _mm512_unpackhi_epi32(vacc5x04152637, vacc5x8C9DAEBF));
348 const __m512i vacc6x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x04152637, vacc6x8C9DAEBF), _mm512_unpackhi_epi32(vacc6x04152637, vacc6x8C9DAEBF));
349
350 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
351 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
352 __m512i vacc1x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc1x084C195D2A6E3B7F);
353 __m512i vacc2x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc2x084C195D2A6E3B7F);
354 __m512i vacc3x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc3x084C195D2A6E3B7F);
355 __m512i vacc4x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc4x084C195D2A6E3B7F);
356 __m512i vacc5x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc5x084C195D2A6E3B7F);
357 __m512i vacc6x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc6x084C195D2A6E3B7F);
358
359 vacc0x0123456789ABCDEF = _mm512_srai_epi32(vacc0x0123456789ABCDEF, 4);
360 vacc1x0123456789ABCDEF = _mm512_srai_epi32(vacc1x0123456789ABCDEF, 4);
361 vacc2x0123456789ABCDEF = _mm512_srai_epi32(vacc2x0123456789ABCDEF, 4);
362 vacc3x0123456789ABCDEF = _mm512_srai_epi32(vacc3x0123456789ABCDEF, 4);
363 vacc4x0123456789ABCDEF = _mm512_srai_epi32(vacc4x0123456789ABCDEF, 4);
364 vacc5x0123456789ABCDEF = _mm512_srai_epi32(vacc5x0123456789ABCDEF, 4);
365 vacc6x0123456789ABCDEF = _mm512_srai_epi32(vacc6x0123456789ABCDEF, 4);
366 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
367 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
368 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
369 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
370 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
371 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
372 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
373
374 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
375 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
376 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
377 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
378 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
379 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
380 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
381
382 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
383 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
384 w = (const float*) w + 32;
385
386 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
387 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
388 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
389 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
390 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
391 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
392 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
393
394 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
395 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
396 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
397 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
398 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
399 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
400 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
401
402 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
403 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
404 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
405 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
406 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
407 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
408 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
409
410 if XNN_LIKELY(nc >= 16) {
411 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
412 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
413 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
414 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
415 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
416 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
417 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
418
419 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
420 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
421 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
422 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
423 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
424 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
425 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
426
427 c0 = (float*) ((uintptr_t) c0 + cn_stride);
428 c1 = (float*) ((uintptr_t) c1 + cn_stride);
429 c2 = (float*) ((uintptr_t) c2 + cn_stride);
430 c3 = (float*) ((uintptr_t) c3 + cn_stride);
431 c4 = (float*) ((uintptr_t) c4 + cn_stride);
432 c5 = (float*) ((uintptr_t) c5 + cn_stride);
433 c6 = (float*) ((uintptr_t) c6 + cn_stride);
434
435 nc -= 16;
436 } else {
437 // Prepare mask for valid 32-bit elements (depends on nc).
438 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
439 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
440 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
441 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
442 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
443 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
444 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
445 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
446 nc = 0;
447 }
448 } while (nc != 0);
449 }