Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-10x16c8-minmax-avx512vnni-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx16c8-avx512vnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/unaligned.h"
19 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_10x16c8__avx512vnni_prfm(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 const int8_t* restrict a,
27 size_t a_stride,
28 const void* restrict w,
29 float* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
33 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 10);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 const int8_t* a0 = a;
46 float* c0 = c;
47 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
48 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 a1 = a0;
51 c1 = c0;
52 }
53 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
54 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
55 if XNN_UNPREDICTABLE(mr <= 2) {
56 a2 = a1;
57 c2 = c1;
58 }
59 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
60 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
61 if XNN_UNPREDICTABLE(mr < 4) {
62 a3 = a2;
63 c3 = c2;
64 }
65 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
66 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
67 if XNN_UNPREDICTABLE(mr <= 4) {
68 a4 = a3;
69 c4 = c3;
70 }
71 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
72 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
73 if XNN_UNPREDICTABLE(mr < 6) {
74 a5 = a4;
75 c5 = c4;
76 }
77 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
78 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
79 if XNN_UNPREDICTABLE(mr <= 6) {
80 a6 = a5;
81 c6 = c5;
82 }
83 const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
84 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
85 if XNN_UNPREDICTABLE(mr < 8) {
86 a7 = a6;
87 c7 = c6;
88 }
89 const int8_t* a8 = (const int8_t*) ((uintptr_t) a7 + a_stride);
90 float* c8 = (float*) ((uintptr_t) c7 + cm_stride);
91 if XNN_UNPREDICTABLE(mr <= 8) {
92 a8 = a7;
93 c8 = c7;
94 }
95 const int8_t* a9 = (const int8_t*) ((uintptr_t) a8 + a_stride);
96 float* c9 = (float*) ((uintptr_t) c8 + cm_stride);
97 if XNN_UNPREDICTABLE(mr != 10) {
98 a9 = a8;
99 c9 = c8;
100 }
101
102 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
103 XNN_FORCE_REALIZATION(vsign_mask);
104 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point + 128);
105 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point + 128);
106 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point + 128);
107 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point + 128);
108 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point + 128);
109 const __m512i vinput_zero_point5 = _mm512_set1_epi32((int) quantization_params[5].zero_point + 128);
110 const __m512i vinput_zero_point6 = _mm512_set1_epi32((int) quantization_params[6].zero_point + 128);
111 const __m512i vinput_zero_point7 = _mm512_set1_epi32((int) quantization_params[7].zero_point + 128);
112 const __m512i vinput_zero_point8 = _mm512_set1_epi32((int) quantization_params[8].zero_point + 128);
113 const __m512i vinput_zero_point9 = _mm512_set1_epi32((int) quantization_params[9].zero_point + 128);
114 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
115 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
116 // XNN_FORCE_REALIZATION(voutput_min);
117 // XNN_FORCE_REALIZATION(voutput_max);
118 const __m512i vmask = _mm512_set1_epi8(0xF0);
119 XNN_FORCE_REALIZATION(vmask);
120 do {
121 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
122 __m512i vsum0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point0);
123 __m512i vacc0x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 0));
124 __m512i vacc0x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 1));
125 __m512i vsum1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point1);
126 __m512i vacc1x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum1x0123456789ABCDEF, 0));
127 __m512i vacc1x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum1x0123456789ABCDEF, 1));
128 __m512i vsum2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point2);
129 __m512i vacc2x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum2x0123456789ABCDEF, 0));
130 __m512i vacc2x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum2x0123456789ABCDEF, 1));
131 __m512i vsum3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point3);
132 __m512i vacc3x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum3x0123456789ABCDEF, 0));
133 __m512i vacc3x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum3x0123456789ABCDEF, 1));
134 __m512i vsum4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point4);
135 __m512i vacc4x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum4x0123456789ABCDEF, 0));
136 __m512i vacc4x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum4x0123456789ABCDEF, 1));
137 __m512i vsum5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point5);
138 __m512i vacc5x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum5x0123456789ABCDEF, 0));
139 __m512i vacc5x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum5x0123456789ABCDEF, 1));
140 __m512i vsum6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point6);
141 __m512i vacc6x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum6x0123456789ABCDEF, 0));
142 __m512i vacc6x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum6x0123456789ABCDEF, 1));
143 __m512i vsum7x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point7);
144 __m512i vacc7x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum7x0123456789ABCDEF, 0));
145 __m512i vacc7x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum7x0123456789ABCDEF, 1));
146 __m512i vsum8x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point8);
147 __m512i vacc8x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum8x0123456789ABCDEF, 0));
148 __m512i vacc8x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum8x0123456789ABCDEF, 1));
149 __m512i vsum9x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point9);
150 __m512i vacc9x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum9x0123456789ABCDEF, 0));
151 __m512i vacc9x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum9x0123456789ABCDEF, 1));
152 w = (const int32_t*) w + 16;
153
154 size_t k = kc;
155 while (k >= 16 * sizeof(int8_t)) {
156 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
157 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
158 a0 += 16;
159 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
160 const __m512i va1x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
161 a1 += 16;
162 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
163 const __m512i va2x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
164 a2 += 16;
165 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
166 const __m512i va3x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
167 a3 += 16;
168 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
169 const __m512i va4x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
170 a4 += 16;
171 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
172 const __m512i va5x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
173 a5 += 16;
174 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
175 const __m512i va6x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
176 a6 += 16;
177 const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask);
178 const __m512i va7x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
179 a7 += 16;
180 const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask);
181 const __m512i va8x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
182 a8 += 16;
183 const __m512i va9x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9)), vsign_mask);
184 const __m512i va9x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
185 a9 += 16;
186
187 const __m512i vbb01234567x01234567 = _mm512_load_si512(w);
188 const __m512i vbb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
189 const __m512i vbs01234567x01234567 = _mm512_slli_epi32(vbb01234567x01234567, 4);
190 const __m512i vbs89ABCDEFx01234567 = _mm512_slli_epi32(vbb89ABCDEFx01234567, 4);
191 const __m512i vb01234567x89ABCDEF = _mm512_and_si512(vbb01234567x01234567, vmask);
192 const __m512i vb89ABCDEFx89ABCDEF = _mm512_and_si512(vbb89ABCDEFx01234567, vmask);
193 const __m512i vb01234567x01234567 = _mm512_and_si512(vbs01234567x01234567, vmask);
194 const __m512i vb89ABCDEFx01234567 = _mm512_and_si512(vbs89ABCDEFx01234567, vmask);
195
196 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
197 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
198 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
199 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
200 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
201 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
202 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
203 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
204 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
205 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
206 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
207 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
208 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
209 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
210 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567);
211 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567);
212 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567);
213 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567);
214 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x01234567, vb01234567x01234567);
215 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x01234567, vb89ABCDEFx01234567);
216 xnn_prefetch_to_l1((const int8_t*) w + 896);
217 xnn_prefetch_to_l1((const int8_t*) w + 960);
218 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
219 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
220 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x89ABCDEF, vb01234567x89ABCDEF);
221 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x89ABCDEF, vb89ABCDEFx89ABCDEF);
222 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x89ABCDEF, vb01234567x89ABCDEF);
223 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x89ABCDEF, vb89ABCDEFx89ABCDEF);
224 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x89ABCDEF, vb01234567x89ABCDEF);
225 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x89ABCDEF, vb89ABCDEFx89ABCDEF);
226 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x89ABCDEF, vb01234567x89ABCDEF);
227 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x89ABCDEF, vb89ABCDEFx89ABCDEF);
228 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x89ABCDEF, vb01234567x89ABCDEF);
229 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x89ABCDEF, vb89ABCDEFx89ABCDEF);
230 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x89ABCDEF, vb01234567x89ABCDEF);
231 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x89ABCDEF, vb89ABCDEFx89ABCDEF);
232 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x89ABCDEF, vb01234567x89ABCDEF);
233 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x89ABCDEF, vb89ABCDEFx89ABCDEF);
234 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x89ABCDEF, vb01234567x89ABCDEF);
235 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x89ABCDEF, vb89ABCDEFx89ABCDEF);
236 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x89ABCDEF, vb01234567x89ABCDEF);
237 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x89ABCDEF, vb89ABCDEFx89ABCDEF);
238
239 w = (const int8_t*) w + 128;
240 k -= 16 * sizeof(int8_t);
241 }
242
243 if (k != 0) {
244 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
245 a0 += 8;
246 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
247 a1 += 8;
248 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
249 a2 += 8;
250 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
251 a3 += 8;
252 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
253 a4 += 8;
254 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
255 a5 += 8;
256 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
257 a6 += 8;
258 const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask);
259 a7 += 8;
260 const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask);
261 a8 += 8;
262 const __m512i va9x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9)), vsign_mask);
263 a9 += 8;
264
265 const __m512i vbb01234567x01234567 = _mm512_load_si512(w);
266 const __m512i vbb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
267 const __m512i vb01234567x01234567 = _mm512_slli_epi32(vbb01234567x01234567, 4);
268 const __m512i vb89ABCDEFx01234567 = _mm512_slli_epi32(vbb89ABCDEFx01234567, 4);
269
270 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
271 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
272 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
273 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
274 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
275 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
276 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
277 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
278 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
279 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
280 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
281 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
282 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
283 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
284 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567);
285 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567);
286 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567);
287 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567);
288 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x01234567, vb01234567x01234567);
289 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x01234567, vb89ABCDEFx01234567);
290 xnn_prefetch_to_l1((const int8_t*) w + 896);
291 xnn_prefetch_to_l1((const int8_t*) w + 960);
292
293 w = (const int8_t*) w + 128;
294 k -= 8 * sizeof(int8_t);
295 }
296
297 // Add adjacent pairs
298 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
299 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
300 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
301 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
302 const __m512i vsum1x01234567 = _mm512_add_epi32(vacc1x01234567, _mm512_srli_epi64(vacc1x01234567, 32));
303 const __m512i vsum1x89ABCDEF = _mm512_add_epi32(vacc1x89ABCDEF, _mm512_srli_epi64(vacc1x89ABCDEF, 32));
304 __m512i vacc1x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum1x01234567, vidx, vsum1x89ABCDEF);
305 const __m512i vsum2x01234567 = _mm512_add_epi32(vacc2x01234567, _mm512_srli_epi64(vacc2x01234567, 32));
306 const __m512i vsum2x89ABCDEF = _mm512_add_epi32(vacc2x89ABCDEF, _mm512_srli_epi64(vacc2x89ABCDEF, 32));
307 __m512i vacc2x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum2x01234567, vidx, vsum2x89ABCDEF);
308 const __m512i vsum3x01234567 = _mm512_add_epi32(vacc3x01234567, _mm512_srli_epi64(vacc3x01234567, 32));
309 const __m512i vsum3x89ABCDEF = _mm512_add_epi32(vacc3x89ABCDEF, _mm512_srli_epi64(vacc3x89ABCDEF, 32));
310 __m512i vacc3x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum3x01234567, vidx, vsum3x89ABCDEF);
311 const __m512i vsum4x01234567 = _mm512_add_epi32(vacc4x01234567, _mm512_srli_epi64(vacc4x01234567, 32));
312 const __m512i vsum4x89ABCDEF = _mm512_add_epi32(vacc4x89ABCDEF, _mm512_srli_epi64(vacc4x89ABCDEF, 32));
313 __m512i vacc4x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum4x01234567, vidx, vsum4x89ABCDEF);
314 const __m512i vsum5x01234567 = _mm512_add_epi32(vacc5x01234567, _mm512_srli_epi64(vacc5x01234567, 32));
315 const __m512i vsum5x89ABCDEF = _mm512_add_epi32(vacc5x89ABCDEF, _mm512_srli_epi64(vacc5x89ABCDEF, 32));
316 __m512i vacc5x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum5x01234567, vidx, vsum5x89ABCDEF);
317 const __m512i vsum6x01234567 = _mm512_add_epi32(vacc6x01234567, _mm512_srli_epi64(vacc6x01234567, 32));
318 const __m512i vsum6x89ABCDEF = _mm512_add_epi32(vacc6x89ABCDEF, _mm512_srli_epi64(vacc6x89ABCDEF, 32));
319 __m512i vacc6x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum6x01234567, vidx, vsum6x89ABCDEF);
320 const __m512i vsum7x01234567 = _mm512_add_epi32(vacc7x01234567, _mm512_srli_epi64(vacc7x01234567, 32));
321 const __m512i vsum7x89ABCDEF = _mm512_add_epi32(vacc7x89ABCDEF, _mm512_srli_epi64(vacc7x89ABCDEF, 32));
322 __m512i vacc7x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum7x01234567, vidx, vsum7x89ABCDEF);
323 const __m512i vsum8x01234567 = _mm512_add_epi32(vacc8x01234567, _mm512_srli_epi64(vacc8x01234567, 32));
324 const __m512i vsum8x89ABCDEF = _mm512_add_epi32(vacc8x89ABCDEF, _mm512_srli_epi64(vacc8x89ABCDEF, 32));
325 __m512i vacc8x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum8x01234567, vidx, vsum8x89ABCDEF);
326 const __m512i vsum9x01234567 = _mm512_add_epi32(vacc9x01234567, _mm512_srli_epi64(vacc9x01234567, 32));
327 const __m512i vsum9x89ABCDEF = _mm512_add_epi32(vacc9x89ABCDEF, _mm512_srli_epi64(vacc9x89ABCDEF, 32));
328 __m512i vacc9x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum9x01234567, vidx, vsum9x89ABCDEF);
329
330 vacc0x0123456789ABCDEF = _mm512_srai_epi32(vacc0x0123456789ABCDEF, 4);
331 vacc1x0123456789ABCDEF = _mm512_srai_epi32(vacc1x0123456789ABCDEF, 4);
332 vacc2x0123456789ABCDEF = _mm512_srai_epi32(vacc2x0123456789ABCDEF, 4);
333 vacc3x0123456789ABCDEF = _mm512_srai_epi32(vacc3x0123456789ABCDEF, 4);
334 vacc4x0123456789ABCDEF = _mm512_srai_epi32(vacc4x0123456789ABCDEF, 4);
335 vacc5x0123456789ABCDEF = _mm512_srai_epi32(vacc5x0123456789ABCDEF, 4);
336 vacc6x0123456789ABCDEF = _mm512_srai_epi32(vacc6x0123456789ABCDEF, 4);
337 vacc7x0123456789ABCDEF = _mm512_srai_epi32(vacc7x0123456789ABCDEF, 4);
338 vacc8x0123456789ABCDEF = _mm512_srai_epi32(vacc8x0123456789ABCDEF, 4);
339 vacc9x0123456789ABCDEF = _mm512_srai_epi32(vacc9x0123456789ABCDEF, 4);
340 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
341 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
342 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
343 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
344 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
345 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
346 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
347 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
348 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
349 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
350
351 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
352 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
353 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
354 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
355 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
356 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
357 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
358 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
359 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, _mm512_set1_ps(quantization_params[8].inv_scale));
360 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, _mm512_set1_ps(quantization_params[9].inv_scale));
361
362 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
363 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
364 w = (const float*) w + 32;
365
366 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
367 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
368 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
369 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
370 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
371 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
372 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
373 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
374 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
375 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vscaled9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
376
377 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
378 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
379 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
380 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
381 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
382 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
383 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
384 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
385 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
386 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
387
388 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
389 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
390 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
391 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
392 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
393 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
394 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
395 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
396 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
397 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
398
399 if XNN_LIKELY(nc >= 16) {
400 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
401 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
402 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
403 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
404 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
405 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
406 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
407 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
408 _mm512_storeu_ps(c8, vscaled8x0123456789ABCDEF);
409 _mm512_storeu_ps(c9, vscaled9x0123456789ABCDEF);
410
411 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
412 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
413 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
414 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
415 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
416 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
417 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
418 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
419 a8 = (const int8_t*) ((uintptr_t) a8 - kc);
420 a9 = (const int8_t*) ((uintptr_t) a9 - kc);
421
422 c0 = (float*) ((uintptr_t) c0 + cn_stride);
423 c1 = (float*) ((uintptr_t) c1 + cn_stride);
424 c2 = (float*) ((uintptr_t) c2 + cn_stride);
425 c3 = (float*) ((uintptr_t) c3 + cn_stride);
426 c4 = (float*) ((uintptr_t) c4 + cn_stride);
427 c5 = (float*) ((uintptr_t) c5 + cn_stride);
428 c6 = (float*) ((uintptr_t) c6 + cn_stride);
429 c7 = (float*) ((uintptr_t) c7 + cn_stride);
430 c8 = (float*) ((uintptr_t) c8 + cn_stride);
431 c9 = (float*) ((uintptr_t) c9 + cn_stride);
432
433 nc -= 16;
434 } else {
435 // Prepare mask for valid 32-bit elements (depends on nc).
436 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
437 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
438 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
439 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
440 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
441 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
442 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
443 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
444 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
445 _mm512_mask_storeu_ps(c8, vmask, vscaled8x0123456789ABCDEF);
446 _mm512_mask_storeu_ps(c9, vmask, vscaled9x0123456789ABCDEF);
447 nc = 0;
448 }
449 } while (nc != 0);
450 }