Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-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_qb4w_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_qb4w_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 size_t bl = params->scalar.blocksize;
102 assert(bl != 0);
103 assert(bl <= kc);
104 assert(kc % bl == 0);
105 assert(bl % 32 == 0);
106
107 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
108 XNN_FORCE_REALIZATION(vsign_mask);
109 const __m512 vinput_zero_point0 = _mm512_set1_ps((float) quantization_params[0].zero_point + 128);
110 const __m512 vinput_zero_point1 = _mm512_set1_ps((float) quantization_params[1].zero_point + 128);
111 const __m512 vinput_zero_point2 = _mm512_set1_ps((float) quantization_params[2].zero_point + 128);
112 const __m512 vinput_zero_point3 = _mm512_set1_ps((float) quantization_params[3].zero_point + 128);
113 const __m512 vinput_zero_point4 = _mm512_set1_ps((float) quantization_params[4].zero_point + 128);
114 const __m512 vinput_zero_point5 = _mm512_set1_ps((float) quantization_params[5].zero_point + 128);
115 const __m512 vinput_zero_point6 = _mm512_set1_ps((float) quantization_params[6].zero_point + 128);
116 const __m512 vinput_zero_point7 = _mm512_set1_ps((float) quantization_params[7].zero_point + 128);
117 const __m512 vinput_zero_point8 = _mm512_set1_ps((float) quantization_params[8].zero_point + 128);
118 const __m512 vinput_zero_point9 = _mm512_set1_ps((float) quantization_params[9].zero_point + 128);
119 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
120 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
121 // XNN_FORCE_REALIZATION(voutput_min);
122 // XNN_FORCE_REALIZATION(voutput_max);
123 const __m512i vmask = _mm512_set1_epi8(0xF0);
124 XNN_FORCE_REALIZATION(vmask);
125 do {
126 const __m512 vksum0123456789ABCDEF = _mm512_loadu_ps(w);
127 __m512 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point0);
128 __m512 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point1);
129 __m512 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point2);
130 __m512 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point3);
131 __m512 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point4);
132 __m512 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point5);
133 __m512 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point6);
134 __m512 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point7);
135 __m512 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point8);
136 __m512 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point9);
137 w = (const int32_t*) w + 16;
138
139 for (size_t kb=0; kb < kc; kb+=bl) {
140 __m512i vacc0x01234567 = _mm512_setzero_epi32();
141 __m512i vacc0x89ABCDEF = _mm512_setzero_epi32();
142 __m512i vacc1x01234567 = _mm512_setzero_epi32();
143 __m512i vacc1x89ABCDEF = _mm512_setzero_epi32();
144 __m512i vacc2x01234567 = _mm512_setzero_epi32();
145 __m512i vacc2x89ABCDEF = _mm512_setzero_epi32();
146 __m512i vacc3x01234567 = _mm512_setzero_epi32();
147 __m512i vacc3x89ABCDEF = _mm512_setzero_epi32();
148 __m512i vacc4x01234567 = _mm512_setzero_epi32();
149 __m512i vacc4x89ABCDEF = _mm512_setzero_epi32();
150 __m512i vacc5x01234567 = _mm512_setzero_epi32();
151 __m512i vacc5x89ABCDEF = _mm512_setzero_epi32();
152 __m512i vacc6x01234567 = _mm512_setzero_epi32();
153 __m512i vacc6x89ABCDEF = _mm512_setzero_epi32();
154 __m512i vacc7x01234567 = _mm512_setzero_epi32();
155 __m512i vacc7x89ABCDEF = _mm512_setzero_epi32();
156 __m512i vacc8x01234567 = _mm512_setzero_epi32();
157 __m512i vacc8x89ABCDEF = _mm512_setzero_epi32();
158 __m512i vacc9x01234567 = _mm512_setzero_epi32();
159 __m512i vacc9x89ABCDEF = _mm512_setzero_epi32();
160 size_t k = bl;
161 while (k >= 16 * sizeof(int8_t)) {
162 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
163 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
164 a0 += 16;
165 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
166 const __m512i va1x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
167 a1 += 16;
168 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
169 const __m512i va2x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
170 a2 += 16;
171 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
172 const __m512i va3x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
173 a3 += 16;
174 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
175 const __m512i va4x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
176 a4 += 16;
177 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
178 const __m512i va5x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
179 a5 += 16;
180 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
181 const __m512i va6x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
182 a6 += 16;
183 const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask);
184 const __m512i va7x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
185 a7 += 16;
186 const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask);
187 const __m512i va8x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
188 a8 += 16;
189 const __m512i va9x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9)), vsign_mask);
190 const __m512i va9x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
191 a9 += 16;
192
193 const __m512i vbb01234567x01234567 = _mm512_loadu_si512(w);
194 const __m512i vbb89ABCDEFx01234567 = _mm512_loadu_si512((const int8_t*) w + 64);
195 const __m512i vbs01234567x01234567 = _mm512_slli_epi32(vbb01234567x01234567, 4);
196 const __m512i vbs89ABCDEFx01234567 = _mm512_slli_epi32(vbb89ABCDEFx01234567, 4);
197 const __m512i vb01234567x89ABCDEF = _mm512_and_si512(vbb01234567x01234567, vmask);
198 const __m512i vb89ABCDEFx89ABCDEF = _mm512_and_si512(vbb89ABCDEFx01234567, vmask);
199 const __m512i vb01234567x01234567 = _mm512_and_si512(vbs01234567x01234567, vmask);
200 const __m512i vb89ABCDEFx01234567 = _mm512_and_si512(vbs89ABCDEFx01234567, vmask);
201
202 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
203 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
204 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
205 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
206 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
207 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
208 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
209 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
210 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
211 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
212 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
213 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
214 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
215 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
216 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567);
217 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567);
218 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567);
219 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567);
220 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x01234567, vb01234567x01234567);
221 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x01234567, vb89ABCDEFx01234567);
222 xnn_prefetch_to_l1((const int8_t*) w + 896);
223 xnn_prefetch_to_l1((const int8_t*) w + 960);
224 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
225 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
226 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x89ABCDEF, vb01234567x89ABCDEF);
227 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x89ABCDEF, vb89ABCDEFx89ABCDEF);
228 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x89ABCDEF, vb01234567x89ABCDEF);
229 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x89ABCDEF, vb89ABCDEFx89ABCDEF);
230 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x89ABCDEF, vb01234567x89ABCDEF);
231 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x89ABCDEF, vb89ABCDEFx89ABCDEF);
232 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x89ABCDEF, vb01234567x89ABCDEF);
233 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x89ABCDEF, vb89ABCDEFx89ABCDEF);
234 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x89ABCDEF, vb01234567x89ABCDEF);
235 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x89ABCDEF, vb89ABCDEFx89ABCDEF);
236 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x89ABCDEF, vb01234567x89ABCDEF);
237 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x89ABCDEF, vb89ABCDEFx89ABCDEF);
238 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x89ABCDEF, vb01234567x89ABCDEF);
239 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x89ABCDEF, vb89ABCDEFx89ABCDEF);
240 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x89ABCDEF, vb01234567x89ABCDEF);
241 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x89ABCDEF, vb89ABCDEFx89ABCDEF);
242 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x89ABCDEF, vb01234567x89ABCDEF);
243 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x89ABCDEF, vb89ABCDEFx89ABCDEF);
244
245 w = (const int8_t*) w + 128;
246 k -= 16 * sizeof(int8_t);
247 }
248
249 if (k != 0) {
250 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
251 a0 += 8;
252 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
253 a1 += 8;
254 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
255 a2 += 8;
256 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
257 a3 += 8;
258 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
259 a4 += 8;
260 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
261 a5 += 8;
262 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
263 a6 += 8;
264 const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask);
265 a7 += 8;
266 const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask);
267 a8 += 8;
268 const __m512i va9x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9)), vsign_mask);
269 a9 += 8;
270
271 const __m512i vbb01234567x01234567 = _mm512_load_si512(w);
272 const __m512i vbb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
273 const __m512i vb01234567x01234567 = _mm512_slli_epi32(vbb01234567x01234567, 4);
274 const __m512i vb89ABCDEFx01234567 = _mm512_slli_epi32(vbb89ABCDEFx01234567, 4);
275
276 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
277 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
278 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
279 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
280 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
281 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
282 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
283 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
284 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
285 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
286 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
287 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
288 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
289 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
290 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567);
291 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567);
292 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567);
293 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567);
294 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x01234567, vb01234567x01234567);
295 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x01234567, vb89ABCDEFx01234567);
296 xnn_prefetch_to_l1((const int8_t*) w + 896);
297 xnn_prefetch_to_l1((const int8_t*) w + 960);
298
299 w = (const int8_t*) w + 128;
300 k -= 8 * sizeof(int8_t);
301 }
302 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_castsi512_ps(_mm512_slli_epi32(
303 _mm512_cvtepu16_epi32(_mm256_load_si256((const __m256i*) w)), 16));
304 w = (const uint16_t*) w + 16;
305
306 // Add adjacent pairs
307 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
308 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
309 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
310 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
311 const __m512i vsum1x01234567 = _mm512_add_epi32(vacc1x01234567, _mm512_srli_epi64(vacc1x01234567, 32));
312 const __m512i vsum1x89ABCDEF = _mm512_add_epi32(vacc1x89ABCDEF, _mm512_srli_epi64(vacc1x89ABCDEF, 32));
313 __m512i vacc1x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum1x01234567, vidx, vsum1x89ABCDEF);
314 const __m512i vsum2x01234567 = _mm512_add_epi32(vacc2x01234567, _mm512_srli_epi64(vacc2x01234567, 32));
315 const __m512i vsum2x89ABCDEF = _mm512_add_epi32(vacc2x89ABCDEF, _mm512_srli_epi64(vacc2x89ABCDEF, 32));
316 __m512i vacc2x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum2x01234567, vidx, vsum2x89ABCDEF);
317 const __m512i vsum3x01234567 = _mm512_add_epi32(vacc3x01234567, _mm512_srli_epi64(vacc3x01234567, 32));
318 const __m512i vsum3x89ABCDEF = _mm512_add_epi32(vacc3x89ABCDEF, _mm512_srli_epi64(vacc3x89ABCDEF, 32));
319 __m512i vacc3x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum3x01234567, vidx, vsum3x89ABCDEF);
320 const __m512i vsum4x01234567 = _mm512_add_epi32(vacc4x01234567, _mm512_srli_epi64(vacc4x01234567, 32));
321 const __m512i vsum4x89ABCDEF = _mm512_add_epi32(vacc4x89ABCDEF, _mm512_srli_epi64(vacc4x89ABCDEF, 32));
322 __m512i vacc4x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum4x01234567, vidx, vsum4x89ABCDEF);
323 const __m512i vsum5x01234567 = _mm512_add_epi32(vacc5x01234567, _mm512_srli_epi64(vacc5x01234567, 32));
324 const __m512i vsum5x89ABCDEF = _mm512_add_epi32(vacc5x89ABCDEF, _mm512_srli_epi64(vacc5x89ABCDEF, 32));
325 __m512i vacc5x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum5x01234567, vidx, vsum5x89ABCDEF);
326 const __m512i vsum6x01234567 = _mm512_add_epi32(vacc6x01234567, _mm512_srli_epi64(vacc6x01234567, 32));
327 const __m512i vsum6x89ABCDEF = _mm512_add_epi32(vacc6x89ABCDEF, _mm512_srli_epi64(vacc6x89ABCDEF, 32));
328 __m512i vacc6x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum6x01234567, vidx, vsum6x89ABCDEF);
329 const __m512i vsum7x01234567 = _mm512_add_epi32(vacc7x01234567, _mm512_srli_epi64(vacc7x01234567, 32));
330 const __m512i vsum7x89ABCDEF = _mm512_add_epi32(vacc7x89ABCDEF, _mm512_srli_epi64(vacc7x89ABCDEF, 32));
331 __m512i vacc7x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum7x01234567, vidx, vsum7x89ABCDEF);
332 const __m512i vsum8x01234567 = _mm512_add_epi32(vacc8x01234567, _mm512_srli_epi64(vacc8x01234567, 32));
333 const __m512i vsum8x89ABCDEF = _mm512_add_epi32(vacc8x89ABCDEF, _mm512_srli_epi64(vacc8x89ABCDEF, 32));
334 __m512i vacc8x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum8x01234567, vidx, vsum8x89ABCDEF);
335 const __m512i vsum9x01234567 = _mm512_add_epi32(vacc9x01234567, _mm512_srli_epi64(vacc9x01234567, 32));
336 const __m512i vsum9x89ABCDEF = _mm512_add_epi32(vacc9x89ABCDEF, _mm512_srli_epi64(vacc9x89ABCDEF, 32));
337 __m512i vacc9x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum9x01234567, vidx, vsum9x89ABCDEF);
338 __m512 vf0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
339 __m512 vf1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
340 __m512 vf2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
341 __m512 vf3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
342 __m512 vf4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
343 __m512 vf5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
344 __m512 vf6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
345 __m512 vf7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
346 __m512 vf8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
347 __m512 vf9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
348
349 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vf0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled0x0123456789ABCDEF);
350 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vf1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled1x0123456789ABCDEF);
351 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vf2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled2x0123456789ABCDEF);
352 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vf3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled3x0123456789ABCDEF);
353 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vf4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled4x0123456789ABCDEF);
354 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vf5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled5x0123456789ABCDEF);
355 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vf6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled6x0123456789ABCDEF);
356 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vf7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled7x0123456789ABCDEF);
357 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vf8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled8x0123456789ABCDEF);
358 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vf9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled9x0123456789ABCDEF);
359 }
360
361 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
362 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
363 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
364 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
365 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
366 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
367 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
368 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
369 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, _mm512_set1_ps(quantization_params[8].inv_scale));
370 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, _mm512_set1_ps(quantization_params[9].inv_scale));
371
372 const __m512 vbias0123456789ABCDEF = _mm512_loadu_ps((const float*) w);
373 w = (const float*) w + 16;
374
375 vscaled0x0123456789ABCDEF = _mm512_add_ps(vscaled0x0123456789ABCDEF, vbias0123456789ABCDEF);
376 vscaled1x0123456789ABCDEF = _mm512_add_ps(vscaled1x0123456789ABCDEF, vbias0123456789ABCDEF);
377 vscaled2x0123456789ABCDEF = _mm512_add_ps(vscaled2x0123456789ABCDEF, vbias0123456789ABCDEF);
378 vscaled3x0123456789ABCDEF = _mm512_add_ps(vscaled3x0123456789ABCDEF, vbias0123456789ABCDEF);
379 vscaled4x0123456789ABCDEF = _mm512_add_ps(vscaled4x0123456789ABCDEF, vbias0123456789ABCDEF);
380 vscaled5x0123456789ABCDEF = _mm512_add_ps(vscaled5x0123456789ABCDEF, vbias0123456789ABCDEF);
381 vscaled6x0123456789ABCDEF = _mm512_add_ps(vscaled6x0123456789ABCDEF, vbias0123456789ABCDEF);
382 vscaled7x0123456789ABCDEF = _mm512_add_ps(vscaled7x0123456789ABCDEF, vbias0123456789ABCDEF);
383 vscaled8x0123456789ABCDEF = _mm512_add_ps(vscaled8x0123456789ABCDEF, vbias0123456789ABCDEF);
384 vscaled9x0123456789ABCDEF = _mm512_add_ps(vscaled9x0123456789ABCDEF, vbias0123456789ABCDEF);
385
386 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
387 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
388 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
389 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
390 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
391 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
392 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
393 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
394 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
395 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
396
397 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
398 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
399 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
400 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
401 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
402 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
403 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
404 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
405 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
406 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
407
408 if XNN_LIKELY(nc >= 16) {
409 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
410 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
411 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
412 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
413 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
414 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
415 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
416 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
417 _mm512_storeu_ps(c8, vscaled8x0123456789ABCDEF);
418 _mm512_storeu_ps(c9, vscaled9x0123456789ABCDEF);
419
420 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
421 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
422 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
423 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
424 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
425 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
426 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
427 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
428 a8 = (const int8_t*) ((uintptr_t) a8 - kc);
429 a9 = (const int8_t*) ((uintptr_t) a9 - kc);
430
431 c0 = (float*) ((uintptr_t) c0 + cn_stride);
432 c1 = (float*) ((uintptr_t) c1 + cn_stride);
433 c2 = (float*) ((uintptr_t) c2 + cn_stride);
434 c3 = (float*) ((uintptr_t) c3 + cn_stride);
435 c4 = (float*) ((uintptr_t) c4 + cn_stride);
436 c5 = (float*) ((uintptr_t) c5 + cn_stride);
437 c6 = (float*) ((uintptr_t) c6 + cn_stride);
438 c7 = (float*) ((uintptr_t) c7 + cn_stride);
439 c8 = (float*) ((uintptr_t) c8 + cn_stride);
440 c9 = (float*) ((uintptr_t) c9 + cn_stride);
441
442 nc -= 16;
443 } else {
444 // Prepare mask for valid 32-bit elements (depends on nc).
445 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
446 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
447 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
448 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
449 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
450 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
451 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
452 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
453 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
454 _mm512_mask_storeu_ps(c8, vmask, vscaled8x0123456789ABCDEF);
455 _mm512_mask_storeu_ps(c9, vmask, vscaled9x0123456789ABCDEF);
456 nc = 0;
457 }
458 } while (nc != 0);
459 }