Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-8x16c8-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_8x16c8__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 <= 8);
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 const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
81 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
82 if XNN_UNPREDICTABLE(mr != 8) {
83 a7 = a6;
84 c7 = c6;
85 }
86
87 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
88 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point);
89 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point);
90 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point);
91 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point);
92 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point);
93 const __m512i vinput_zero_point5 = _mm512_set1_epi32((int) quantization_params[5].zero_point);
94 const __m512i vinput_zero_point6 = _mm512_set1_epi32((int) quantization_params[6].zero_point);
95 const __m512i vinput_zero_point7 = _mm512_set1_epi32((int) quantization_params[7].zero_point);
96 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
97 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
98 // XNN_FORCE_REALIZATION(voutput_min);
99 // XNN_FORCE_REALIZATION(voutput_max);
100 const __m256i vmask = _mm256_set1_epi8(0xF0);
101 XNN_FORCE_REALIZATION(vmask);
102
103 do {
104 const __m512i vksum0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
105 const __m512i vksum4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
106 const __m512i vksum89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
107 const __m512i vksumCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
108
109 __m512i vacc0x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point0);
110 __m512i vacc0x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point0);
111 __m512i vacc0x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point0);
112 __m512i vacc0xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point0);
113 __m512i vacc1x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point1);
114 __m512i vacc1x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point1);
115 __m512i vacc1x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point1);
116 __m512i vacc1xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point1);
117 __m512i vacc2x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point2);
118 __m512i vacc2x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point2);
119 __m512i vacc2x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point2);
120 __m512i vacc2xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point2);
121 __m512i vacc3x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point3);
122 __m512i vacc3x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point3);
123 __m512i vacc3x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point3);
124 __m512i vacc3xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point3);
125 __m512i vacc4x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point4);
126 __m512i vacc4x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point4);
127 __m512i vacc4x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point4);
128 __m512i vacc4xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point4);
129 __m512i vacc5x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point5);
130 __m512i vacc5x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point5);
131 __m512i vacc5x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point5);
132 __m512i vacc5xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point5);
133 __m512i vacc6x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point6);
134 __m512i vacc6x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point6);
135 __m512i vacc6x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point6);
136 __m512i vacc6xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point6);
137 __m512i vacc7x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point7);
138 __m512i vacc7x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point7);
139 __m512i vacc7x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point7);
140 __m512i vacc7xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point7);
141 w = (const int32_t*) w + 16;
142
143 size_t k = kc;
144 while (k >= 16 * sizeof(int8_t)) {
145 __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
146 a0 += 8;
147 __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
148 a1 += 8;
149 __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
150 a2 += 8;
151 __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
152 a3 += 8;
153 __m512i va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
154 a4 += 8;
155 __m512i va5 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a5)));
156 a5 += 8;
157 __m512i va6 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a6)));
158 a6 += 8;
159 __m512i va7 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a7)));
160 a7 += 8;
161
162 __m256i vbb0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 0));
163 __m256i vbs0123 = _mm256_slli_epi32(vbb0123, 4);
164 __m256i vbm0123 = _mm256_and_si256(vbs0123, vmask);
165 __m512i vb0123 = _mm512_cvtepi8_epi16(vbm0123);
166 xnn_prefetch_to_l1((const int8_t*) w + 896);
167
168 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
169 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
170 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
171 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
172 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
173 vacc5x0123 = _mm512_add_epi32(vacc5x0123, _mm512_madd_epi16(va5, vb0123));
174 vacc6x0123 = _mm512_add_epi32(vacc6x0123, _mm512_madd_epi16(va6, vb0123));
175 vacc7x0123 = _mm512_add_epi32(vacc7x0123, _mm512_madd_epi16(va7, vb0123));
176 __m256i vbb4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
177 __m256i vbs4567 = _mm256_slli_epi32(vbb4567, 4);
178 __m256i vbm4567 = _mm256_and_si256(vbs4567, vmask);
179 __m512i vb4567 = _mm512_cvtepi8_epi16(vbm4567);
180
181 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
182 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
183 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
184 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
185 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
186 vacc5x4567 = _mm512_add_epi32(vacc5x4567, _mm512_madd_epi16(va5, vb4567));
187 vacc6x4567 = _mm512_add_epi32(vacc6x4567, _mm512_madd_epi16(va6, vb4567));
188 vacc7x4567 = _mm512_add_epi32(vacc7x4567, _mm512_madd_epi16(va7, vb4567));
189 __m256i vbb89AB = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
190 __m256i vbs89AB = _mm256_slli_epi32(vbb89AB, 4);
191 __m256i vbm89AB = _mm256_and_si256(vbs89AB, vmask);
192 __m512i vb89AB = _mm512_cvtepi8_epi16(vbm89AB);
193 xnn_prefetch_to_l1((const int8_t*) w + 960);
194
195 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
196 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
197 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
198 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
199 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
200 vacc5x89AB = _mm512_add_epi32(vacc5x89AB, _mm512_madd_epi16(va5, vb89AB));
201 vacc6x89AB = _mm512_add_epi32(vacc6x89AB, _mm512_madd_epi16(va6, vb89AB));
202 vacc7x89AB = _mm512_add_epi32(vacc7x89AB, _mm512_madd_epi16(va7, vb89AB));
203 __m256i vbbCDEF = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
204 __m256i vbsCDEF = _mm256_slli_epi32(vbbCDEF, 4);
205 __m256i vbmCDEF = _mm256_and_si256(vbsCDEF, vmask);
206 __m512i vbCDEF = _mm512_cvtepi8_epi16(vbmCDEF);
207
208 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
209 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
210 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
211 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
212 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
213 vacc5xCDEF = _mm512_add_epi32(vacc5xCDEF, _mm512_madd_epi16(va5, vbCDEF));
214 vacc6xCDEF = _mm512_add_epi32(vacc6xCDEF, _mm512_madd_epi16(va6, vbCDEF));
215 vacc7xCDEF = _mm512_add_epi32(vacc7xCDEF, _mm512_madd_epi16(va7, vbCDEF));
216
217 va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
218 a0 += 8;
219 va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
220 a1 += 8;
221 va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
222 a2 += 8;
223 va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
224 a3 += 8;
225 va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
226 a4 += 8;
227 va5 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a5)));
228 a5 += 8;
229 va6 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a6)));
230 a6 += 8;
231 va7 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a7)));
232 a7 += 8;
233
234 vbm0123 = _mm256_and_si256(vbb0123, vmask);
235 vb0123 = _mm512_cvtepi8_epi16(vbm0123);
236
237 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
238 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
239 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
240 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
241 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
242 vacc5x0123 = _mm512_add_epi32(vacc5x0123, _mm512_madd_epi16(va5, vb0123));
243 vacc6x0123 = _mm512_add_epi32(vacc6x0123, _mm512_madd_epi16(va6, vb0123));
244 vacc7x0123 = _mm512_add_epi32(vacc7x0123, _mm512_madd_epi16(va7, vb0123));
245 vbm4567 = _mm256_and_si256(vbb4567, vmask);
246 vb4567 = _mm512_cvtepi8_epi16(vbm4567);
247
248 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
249 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
250 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
251 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
252 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
253 vacc5x4567 = _mm512_add_epi32(vacc5x4567, _mm512_madd_epi16(va5, vb4567));
254 vacc6x4567 = _mm512_add_epi32(vacc6x4567, _mm512_madd_epi16(va6, vb4567));
255 vacc7x4567 = _mm512_add_epi32(vacc7x4567, _mm512_madd_epi16(va7, vb4567));
256 vbm89AB = _mm256_and_si256(vbb89AB, vmask);
257 vb89AB = _mm512_cvtepi8_epi16(vbm89AB);
258
259 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
260 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
261 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
262 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
263 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
264 vacc5x89AB = _mm512_add_epi32(vacc5x89AB, _mm512_madd_epi16(va5, vb89AB));
265 vacc6x89AB = _mm512_add_epi32(vacc6x89AB, _mm512_madd_epi16(va6, vb89AB));
266 vacc7x89AB = _mm512_add_epi32(vacc7x89AB, _mm512_madd_epi16(va7, vb89AB));
267 vbmCDEF = _mm256_and_si256(vbbCDEF, vmask);
268 vbCDEF = _mm512_cvtepi8_epi16(vbmCDEF);
269
270 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
271 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
272 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
273 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
274 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
275 vacc5xCDEF = _mm512_add_epi32(vacc5xCDEF, _mm512_madd_epi16(va5, vbCDEF));
276 vacc6xCDEF = _mm512_add_epi32(vacc6xCDEF, _mm512_madd_epi16(va6, vbCDEF));
277 vacc7xCDEF = _mm512_add_epi32(vacc7xCDEF, _mm512_madd_epi16(va7, vbCDEF));
278
279 w = (const int8_t*) w + 128;
280 k -= 16 * sizeof(int8_t);
281 }
282
283 while (k >= 8 * sizeof(int8_t)) {
284 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
285 a0 += 8;
286 const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
287 a1 += 8;
288 const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
289 a2 += 8;
290 const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
291 a3 += 8;
292 const __m512i va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
293 a4 += 8;
294 const __m512i va5 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a5)));
295 a5 += 8;
296 const __m512i va6 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a6)));
297 a6 += 8;
298 const __m512i va7 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a7)));
299 a7 += 8;
300
301 const __m256i vbb0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 0));
302 const __m256i vbs0123 = _mm256_slli_epi32(vbb0123, 4);
303 const __m256i vbm0123 = _mm256_and_si256(vbs0123, vmask);
304 const __m512i vb0123 = _mm512_cvtepi8_epi16(vbm0123);
305 xnn_prefetch_to_l1((const int8_t*) w + 896);
306
307 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
308 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
309 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
310 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
311 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
312 vacc5x0123 = _mm512_add_epi32(vacc5x0123, _mm512_madd_epi16(va5, vb0123));
313 vacc6x0123 = _mm512_add_epi32(vacc6x0123, _mm512_madd_epi16(va6, vb0123));
314 vacc7x0123 = _mm512_add_epi32(vacc7x0123, _mm512_madd_epi16(va7, vb0123));
315 const __m256i vbb4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
316 const __m256i vbs4567 = _mm256_slli_epi32(vbb4567, 4);
317 const __m256i vbm4567 = _mm256_and_si256(vbs4567, vmask);
318 const __m512i vb4567 = _mm512_cvtepi8_epi16(vbm4567);
319
320 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
321 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
322 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
323 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
324 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
325 vacc5x4567 = _mm512_add_epi32(vacc5x4567, _mm512_madd_epi16(va5, vb4567));
326 vacc6x4567 = _mm512_add_epi32(vacc6x4567, _mm512_madd_epi16(va6, vb4567));
327 vacc7x4567 = _mm512_add_epi32(vacc7x4567, _mm512_madd_epi16(va7, vb4567));
328 const __m256i vbb89AB = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
329 const __m256i vbs89AB = _mm256_slli_epi32(vbb89AB, 4);
330 const __m256i vbm89AB = _mm256_and_si256(vbs89AB, vmask);
331 const __m512i vb89AB = _mm512_cvtepi8_epi16(vbm89AB);
332 xnn_prefetch_to_l1((const int8_t*) w + 960);
333
334 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
335 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
336 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
337 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
338 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
339 vacc5x89AB = _mm512_add_epi32(vacc5x89AB, _mm512_madd_epi16(va5, vb89AB));
340 vacc6x89AB = _mm512_add_epi32(vacc6x89AB, _mm512_madd_epi16(va6, vb89AB));
341 vacc7x89AB = _mm512_add_epi32(vacc7x89AB, _mm512_madd_epi16(va7, vb89AB));
342 const __m256i vbbCDEF = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
343 const __m256i vbsCDEF = _mm256_slli_epi32(vbbCDEF, 4);
344 const __m256i vbmCDEF = _mm256_and_si256(vbsCDEF, vmask);
345 const __m512i vbCDEF = _mm512_cvtepi8_epi16(vbmCDEF);
346
347 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
348 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
349 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
350 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
351 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
352 vacc5xCDEF = _mm512_add_epi32(vacc5xCDEF, _mm512_madd_epi16(va5, vbCDEF));
353 vacc6xCDEF = _mm512_add_epi32(vacc6xCDEF, _mm512_madd_epi16(va6, vbCDEF));
354 vacc7xCDEF = _mm512_add_epi32(vacc7xCDEF, _mm512_madd_epi16(va7, vbCDEF));
355
356 w = (const int8_t*) w + 128;
357 k -= 8 * sizeof(int8_t);
358 }
359
360 // Add 4 adjacent sums
361 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
362 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
363 const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
364 const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
365 const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm512_unpackhi_epi32(vacc2x0123, vacc2x4567));
366 const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm512_unpackhi_epi32(vacc2x89AB, vacc2xCDEF));
367 const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm512_unpackhi_epi32(vacc3x0123, vacc3x4567));
368 const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm512_unpackhi_epi32(vacc3x89AB, vacc3xCDEF));
369 const __m512i vacc4x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x0123, vacc4x4567), _mm512_unpackhi_epi32(vacc4x0123, vacc4x4567));
370 const __m512i vacc4x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x89AB, vacc4xCDEF), _mm512_unpackhi_epi32(vacc4x89AB, vacc4xCDEF));
371 const __m512i vacc5x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x0123, vacc5x4567), _mm512_unpackhi_epi32(vacc5x0123, vacc5x4567));
372 const __m512i vacc5x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x89AB, vacc5xCDEF), _mm512_unpackhi_epi32(vacc5x89AB, vacc5xCDEF));
373 const __m512i vacc6x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x0123, vacc6x4567), _mm512_unpackhi_epi32(vacc6x0123, vacc6x4567));
374 const __m512i vacc6x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x89AB, vacc6xCDEF), _mm512_unpackhi_epi32(vacc6x89AB, vacc6xCDEF));
375 const __m512i vacc7x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc7x0123, vacc7x4567), _mm512_unpackhi_epi32(vacc7x0123, vacc7x4567));
376 const __m512i vacc7x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc7x89AB, vacc7xCDEF), _mm512_unpackhi_epi32(vacc7x89AB, vacc7xCDEF));
377
378 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
379 const __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
380 const __m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9DAEBF), _mm512_unpackhi_epi32(vacc2x04152637, vacc2x8C9DAEBF));
381 const __m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9DAEBF), _mm512_unpackhi_epi32(vacc3x04152637, vacc3x8C9DAEBF));
382 const __m512i vacc4x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x04152637, vacc4x8C9DAEBF), _mm512_unpackhi_epi32(vacc4x04152637, vacc4x8C9DAEBF));
383 const __m512i vacc5x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x04152637, vacc5x8C9DAEBF), _mm512_unpackhi_epi32(vacc5x04152637, vacc5x8C9DAEBF));
384 const __m512i vacc6x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x04152637, vacc6x8C9DAEBF), _mm512_unpackhi_epi32(vacc6x04152637, vacc6x8C9DAEBF));
385 const __m512i vacc7x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc7x04152637, vacc7x8C9DAEBF), _mm512_unpackhi_epi32(vacc7x04152637, vacc7x8C9DAEBF));
386
387 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
388 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
389 __m512i vacc1x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc1x084C195D2A6E3B7F);
390 __m512i vacc2x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc2x084C195D2A6E3B7F);
391 __m512i vacc3x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc3x084C195D2A6E3B7F);
392 __m512i vacc4x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc4x084C195D2A6E3B7F);
393 __m512i vacc5x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc5x084C195D2A6E3B7F);
394 __m512i vacc6x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc6x084C195D2A6E3B7F);
395 __m512i vacc7x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc7x084C195D2A6E3B7F);
396
397 vacc0x0123456789ABCDEF = _mm512_srai_epi32(vacc0x0123456789ABCDEF, 4);
398 vacc1x0123456789ABCDEF = _mm512_srai_epi32(vacc1x0123456789ABCDEF, 4);
399 vacc2x0123456789ABCDEF = _mm512_srai_epi32(vacc2x0123456789ABCDEF, 4);
400 vacc3x0123456789ABCDEF = _mm512_srai_epi32(vacc3x0123456789ABCDEF, 4);
401 vacc4x0123456789ABCDEF = _mm512_srai_epi32(vacc4x0123456789ABCDEF, 4);
402 vacc5x0123456789ABCDEF = _mm512_srai_epi32(vacc5x0123456789ABCDEF, 4);
403 vacc6x0123456789ABCDEF = _mm512_srai_epi32(vacc6x0123456789ABCDEF, 4);
404 vacc7x0123456789ABCDEF = _mm512_srai_epi32(vacc7x0123456789ABCDEF, 4);
405 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
406 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
407 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
408 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
409 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
410 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
411 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
412 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
413
414 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
415 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
416 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
417 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
418 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
419 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
420 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
421 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
422
423 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
424 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
425 w = (const float*) w + 32;
426
427 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
428 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
429 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
430 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
431 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
432 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
433 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
434 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
435
436 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
437 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
438 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
439 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
440 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
441 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
442 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
443 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
444
445 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
446 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
447 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
448 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
449 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
450 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
451 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
452 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
453
454 if XNN_LIKELY(nc >= 16) {
455 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
456 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
457 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
458 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
459 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
460 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
461 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
462 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
463
464 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
465 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
466 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
467 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
468 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
469 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
470 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
471 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
472
473 c0 = (float*) ((uintptr_t) c0 + cn_stride);
474 c1 = (float*) ((uintptr_t) c1 + cn_stride);
475 c2 = (float*) ((uintptr_t) c2 + cn_stride);
476 c3 = (float*) ((uintptr_t) c3 + cn_stride);
477 c4 = (float*) ((uintptr_t) c4 + cn_stride);
478 c5 = (float*) ((uintptr_t) c5 + cn_stride);
479 c6 = (float*) ((uintptr_t) c6 + cn_stride);
480 c7 = (float*) ((uintptr_t) c7 + cn_stride);
481
482 nc -= 16;
483 } else {
484 // Prepare mask for valid 32-bit elements (depends on nc).
485 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
486 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
487 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
488 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
489 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
490 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
491 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
492 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
493 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
494 nc = 0;
495 }
496 } while (nc != 0);
497 }