Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-12x16c8-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_12x16c8__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 <= 12);
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 const int8_t* a10 = (const int8_t*) ((uintptr_t) a9 + a_stride);
102 float* c10 = (float*) ((uintptr_t) c9 + cm_stride);
103 if XNN_UNPREDICTABLE(mr <= 10) {
104 a10 = a9;
105 c10 = c9;
106 }
107 const int8_t* a11 = (const int8_t*) ((uintptr_t) a10 + a_stride);
108 float* c11 = (float*) ((uintptr_t) c10 + cm_stride);
109 if XNN_UNPREDICTABLE(mr != 12) {
110 a11 = a10;
111 c11 = c10;
112 }
113 size_t bl = params->scalar.blocksize;
114 assert(bl != 0);
115 assert(bl <= kc);
116 assert(kc % bl == 0);
117 assert(bl % 32 == 0);
118
119 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
120 XNN_FORCE_REALIZATION(vsign_mask);
121 const __m512 vinput_zero_point0 = _mm512_set1_ps((float) quantization_params[0].zero_point + 128);
122 const __m512 vinput_zero_point1 = _mm512_set1_ps((float) quantization_params[1].zero_point + 128);
123 const __m512 vinput_zero_point2 = _mm512_set1_ps((float) quantization_params[2].zero_point + 128);
124 const __m512 vinput_zero_point3 = _mm512_set1_ps((float) quantization_params[3].zero_point + 128);
125 const __m512 vinput_zero_point4 = _mm512_set1_ps((float) quantization_params[4].zero_point + 128);
126 const __m512 vinput_zero_point5 = _mm512_set1_ps((float) quantization_params[5].zero_point + 128);
127 const __m512 vinput_zero_point6 = _mm512_set1_ps((float) quantization_params[6].zero_point + 128);
128 const __m512 vinput_zero_point7 = _mm512_set1_ps((float) quantization_params[7].zero_point + 128);
129 const __m512 vinput_zero_point8 = _mm512_set1_ps((float) quantization_params[8].zero_point + 128);
130 const __m512 vinput_zero_point9 = _mm512_set1_ps((float) quantization_params[9].zero_point + 128);
131 const __m512 vinput_zero_point10 = _mm512_set1_ps((float) quantization_params[10].zero_point + 128);
132 const __m512 vinput_zero_point11 = _mm512_set1_ps((float) quantization_params[11].zero_point + 128);
133 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
134 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
135 // XNN_FORCE_REALIZATION(voutput_min);
136 // XNN_FORCE_REALIZATION(voutput_max);
137 const __m512i vmask = _mm512_set1_epi8(0xF0);
138 XNN_FORCE_REALIZATION(vmask);
139 do {
140 const __m512 vksum0123456789ABCDEF = _mm512_loadu_ps(w);
141 __m512 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point0);
142 __m512 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point1);
143 __m512 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point2);
144 __m512 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point3);
145 __m512 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point4);
146 __m512 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point5);
147 __m512 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point6);
148 __m512 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point7);
149 __m512 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point8);
150 __m512 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point9);
151 __m512 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point10);
152 __m512 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point11);
153 w = (const int32_t*) w + 16;
154
155 for (size_t kb=0; kb < kc; kb+=bl) {
156 __m512i vacc0x01234567 = _mm512_setzero_epi32();
157 __m512i vacc0x89ABCDEF = _mm512_setzero_epi32();
158 __m512i vacc1x01234567 = _mm512_setzero_epi32();
159 __m512i vacc1x89ABCDEF = _mm512_setzero_epi32();
160 __m512i vacc2x01234567 = _mm512_setzero_epi32();
161 __m512i vacc2x89ABCDEF = _mm512_setzero_epi32();
162 __m512i vacc3x01234567 = _mm512_setzero_epi32();
163 __m512i vacc3x89ABCDEF = _mm512_setzero_epi32();
164 __m512i vacc4x01234567 = _mm512_setzero_epi32();
165 __m512i vacc4x89ABCDEF = _mm512_setzero_epi32();
166 __m512i vacc5x01234567 = _mm512_setzero_epi32();
167 __m512i vacc5x89ABCDEF = _mm512_setzero_epi32();
168 __m512i vacc6x01234567 = _mm512_setzero_epi32();
169 __m512i vacc6x89ABCDEF = _mm512_setzero_epi32();
170 __m512i vacc7x01234567 = _mm512_setzero_epi32();
171 __m512i vacc7x89ABCDEF = _mm512_setzero_epi32();
172 __m512i vacc8x01234567 = _mm512_setzero_epi32();
173 __m512i vacc8x89ABCDEF = _mm512_setzero_epi32();
174 __m512i vacc9x01234567 = _mm512_setzero_epi32();
175 __m512i vacc9x89ABCDEF = _mm512_setzero_epi32();
176 __m512i vacc10x01234567 = _mm512_setzero_epi32();
177 __m512i vacc10x89ABCDEF = _mm512_setzero_epi32();
178 __m512i vacc11x01234567 = _mm512_setzero_epi32();
179 __m512i vacc11x89ABCDEF = _mm512_setzero_epi32();
180 size_t k = bl;
181 while (k >= 16 * sizeof(int8_t)) {
182 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
183 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
184 a0 += 16;
185 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
186 const __m512i va1x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
187 a1 += 16;
188 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
189 const __m512i va2x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
190 a2 += 16;
191 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
192 const __m512i va3x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
193 a3 += 16;
194 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
195 const __m512i va4x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
196 a4 += 16;
197 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
198 const __m512i va5x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
199 a5 += 16;
200 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
201 const __m512i va6x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
202 a6 += 16;
203 const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask);
204 const __m512i va7x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
205 a7 += 16;
206 const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask);
207 const __m512i va8x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
208 a8 += 16;
209 const __m512i va9x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9)), vsign_mask);
210 const __m512i va9x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
211 a9 += 16;
212 const __m512i va10x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a10)), vsign_mask);
213 const __m512i va10x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a10 + 8)), vsign_mask);
214 a10 += 16;
215 const __m512i va11x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a11)), vsign_mask);
216 const __m512i va11x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a11 + 8)), vsign_mask);
217 a11 += 16;
218
219 const __m512i vbb01234567x01234567 = _mm512_loadu_si512(w);
220 const __m512i vbb89ABCDEFx01234567 = _mm512_loadu_si512((const int8_t*) w + 64);
221 const __m512i vbs01234567x01234567 = _mm512_slli_epi32(vbb01234567x01234567, 4);
222 const __m512i vbs89ABCDEFx01234567 = _mm512_slli_epi32(vbb89ABCDEFx01234567, 4);
223 const __m512i vb01234567x89ABCDEF = _mm512_and_si512(vbb01234567x01234567, vmask);
224 const __m512i vb89ABCDEFx89ABCDEF = _mm512_and_si512(vbb89ABCDEFx01234567, vmask);
225 const __m512i vb01234567x01234567 = _mm512_and_si512(vbs01234567x01234567, vmask);
226 const __m512i vb89ABCDEFx01234567 = _mm512_and_si512(vbs89ABCDEFx01234567, vmask);
227
228 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
229 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
230 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
231 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
232 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
233 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
234 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
235 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
236 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
237 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
238 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
239 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
240 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
241 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
242 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567);
243 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567);
244 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567);
245 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567);
246 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x01234567, vb01234567x01234567);
247 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x01234567, vb89ABCDEFx01234567);
248 vacc10x01234567 = _mm512_dpbusd_epi32(vacc10x01234567, va10x01234567, vb01234567x01234567);
249 vacc10x89ABCDEF = _mm512_dpbusd_epi32(vacc10x89ABCDEF, va10x01234567, vb89ABCDEFx01234567);
250 vacc11x01234567 = _mm512_dpbusd_epi32(vacc11x01234567, va11x01234567, vb01234567x01234567);
251 vacc11x89ABCDEF = _mm512_dpbusd_epi32(vacc11x89ABCDEF, va11x01234567, vb89ABCDEFx01234567);
252 xnn_prefetch_to_l1((const int8_t*) w + 896);
253 xnn_prefetch_to_l1((const int8_t*) w + 960);
254 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
255 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
256 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x89ABCDEF, vb01234567x89ABCDEF);
257 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x89ABCDEF, vb89ABCDEFx89ABCDEF);
258 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x89ABCDEF, vb01234567x89ABCDEF);
259 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x89ABCDEF, vb89ABCDEFx89ABCDEF);
260 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x89ABCDEF, vb01234567x89ABCDEF);
261 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x89ABCDEF, vb89ABCDEFx89ABCDEF);
262 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x89ABCDEF, vb01234567x89ABCDEF);
263 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x89ABCDEF, vb89ABCDEFx89ABCDEF);
264 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x89ABCDEF, vb01234567x89ABCDEF);
265 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x89ABCDEF, vb89ABCDEFx89ABCDEF);
266 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x89ABCDEF, vb01234567x89ABCDEF);
267 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x89ABCDEF, vb89ABCDEFx89ABCDEF);
268 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x89ABCDEF, vb01234567x89ABCDEF);
269 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x89ABCDEF, vb89ABCDEFx89ABCDEF);
270 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x89ABCDEF, vb01234567x89ABCDEF);
271 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x89ABCDEF, vb89ABCDEFx89ABCDEF);
272 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x89ABCDEF, vb01234567x89ABCDEF);
273 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x89ABCDEF, vb89ABCDEFx89ABCDEF);
274 vacc10x01234567 = _mm512_dpbusd_epi32(vacc10x01234567, va10x89ABCDEF, vb01234567x89ABCDEF);
275 vacc10x89ABCDEF = _mm512_dpbusd_epi32(vacc10x89ABCDEF, va10x89ABCDEF, vb89ABCDEFx89ABCDEF);
276 vacc11x01234567 = _mm512_dpbusd_epi32(vacc11x01234567, va11x89ABCDEF, vb01234567x89ABCDEF);
277 vacc11x89ABCDEF = _mm512_dpbusd_epi32(vacc11x89ABCDEF, va11x89ABCDEF, vb89ABCDEFx89ABCDEF);
278
279 w = (const int8_t*) w + 128;
280 k -= 16 * sizeof(int8_t);
281 }
282
283 if (k != 0) {
284 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
285 a0 += 8;
286 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
287 a1 += 8;
288 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
289 a2 += 8;
290 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
291 a3 += 8;
292 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
293 a4 += 8;
294 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
295 a5 += 8;
296 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
297 a6 += 8;
298 const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask);
299 a7 += 8;
300 const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask);
301 a8 += 8;
302 const __m512i va9x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9)), vsign_mask);
303 a9 += 8;
304 const __m512i va10x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a10)), vsign_mask);
305 a10 += 8;
306 const __m512i va11x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a11)), vsign_mask);
307 a11 += 8;
308
309 const __m512i vbb01234567x01234567 = _mm512_load_si512(w);
310 const __m512i vbb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
311 const __m512i vb01234567x01234567 = _mm512_slli_epi32(vbb01234567x01234567, 4);
312 const __m512i vb89ABCDEFx01234567 = _mm512_slli_epi32(vbb89ABCDEFx01234567, 4);
313
314 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
315 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
316 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
317 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
318 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
319 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
320 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
321 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
322 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
323 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
324 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
325 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
326 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
327 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
328 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567);
329 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567);
330 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567);
331 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567);
332 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x01234567, vb01234567x01234567);
333 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x01234567, vb89ABCDEFx01234567);
334 vacc10x01234567 = _mm512_dpbusd_epi32(vacc10x01234567, va10x01234567, vb01234567x01234567);
335 vacc10x89ABCDEF = _mm512_dpbusd_epi32(vacc10x89ABCDEF, va10x01234567, vb89ABCDEFx01234567);
336 vacc11x01234567 = _mm512_dpbusd_epi32(vacc11x01234567, va11x01234567, vb01234567x01234567);
337 vacc11x89ABCDEF = _mm512_dpbusd_epi32(vacc11x89ABCDEF, va11x01234567, vb89ABCDEFx01234567);
338 xnn_prefetch_to_l1((const int8_t*) w + 896);
339 xnn_prefetch_to_l1((const int8_t*) w + 960);
340
341 w = (const int8_t*) w + 128;
342 k -= 8 * sizeof(int8_t);
343 }
344 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_castsi512_ps(_mm512_slli_epi32(
345 _mm512_cvtepu16_epi32(_mm256_load_si256((const __m256i*) w)), 16));
346 w = (const uint16_t*) w + 16;
347
348 // Add adjacent pairs
349 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
350 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
351 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
352 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
353 const __m512i vsum1x01234567 = _mm512_add_epi32(vacc1x01234567, _mm512_srli_epi64(vacc1x01234567, 32));
354 const __m512i vsum1x89ABCDEF = _mm512_add_epi32(vacc1x89ABCDEF, _mm512_srli_epi64(vacc1x89ABCDEF, 32));
355 __m512i vacc1x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum1x01234567, vidx, vsum1x89ABCDEF);
356 const __m512i vsum2x01234567 = _mm512_add_epi32(vacc2x01234567, _mm512_srli_epi64(vacc2x01234567, 32));
357 const __m512i vsum2x89ABCDEF = _mm512_add_epi32(vacc2x89ABCDEF, _mm512_srli_epi64(vacc2x89ABCDEF, 32));
358 __m512i vacc2x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum2x01234567, vidx, vsum2x89ABCDEF);
359 const __m512i vsum3x01234567 = _mm512_add_epi32(vacc3x01234567, _mm512_srli_epi64(vacc3x01234567, 32));
360 const __m512i vsum3x89ABCDEF = _mm512_add_epi32(vacc3x89ABCDEF, _mm512_srli_epi64(vacc3x89ABCDEF, 32));
361 __m512i vacc3x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum3x01234567, vidx, vsum3x89ABCDEF);
362 const __m512i vsum4x01234567 = _mm512_add_epi32(vacc4x01234567, _mm512_srli_epi64(vacc4x01234567, 32));
363 const __m512i vsum4x89ABCDEF = _mm512_add_epi32(vacc4x89ABCDEF, _mm512_srli_epi64(vacc4x89ABCDEF, 32));
364 __m512i vacc4x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum4x01234567, vidx, vsum4x89ABCDEF);
365 const __m512i vsum5x01234567 = _mm512_add_epi32(vacc5x01234567, _mm512_srli_epi64(vacc5x01234567, 32));
366 const __m512i vsum5x89ABCDEF = _mm512_add_epi32(vacc5x89ABCDEF, _mm512_srli_epi64(vacc5x89ABCDEF, 32));
367 __m512i vacc5x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum5x01234567, vidx, vsum5x89ABCDEF);
368 const __m512i vsum6x01234567 = _mm512_add_epi32(vacc6x01234567, _mm512_srli_epi64(vacc6x01234567, 32));
369 const __m512i vsum6x89ABCDEF = _mm512_add_epi32(vacc6x89ABCDEF, _mm512_srli_epi64(vacc6x89ABCDEF, 32));
370 __m512i vacc6x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum6x01234567, vidx, vsum6x89ABCDEF);
371 const __m512i vsum7x01234567 = _mm512_add_epi32(vacc7x01234567, _mm512_srli_epi64(vacc7x01234567, 32));
372 const __m512i vsum7x89ABCDEF = _mm512_add_epi32(vacc7x89ABCDEF, _mm512_srli_epi64(vacc7x89ABCDEF, 32));
373 __m512i vacc7x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum7x01234567, vidx, vsum7x89ABCDEF);
374 const __m512i vsum8x01234567 = _mm512_add_epi32(vacc8x01234567, _mm512_srli_epi64(vacc8x01234567, 32));
375 const __m512i vsum8x89ABCDEF = _mm512_add_epi32(vacc8x89ABCDEF, _mm512_srli_epi64(vacc8x89ABCDEF, 32));
376 __m512i vacc8x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum8x01234567, vidx, vsum8x89ABCDEF);
377 const __m512i vsum9x01234567 = _mm512_add_epi32(vacc9x01234567, _mm512_srli_epi64(vacc9x01234567, 32));
378 const __m512i vsum9x89ABCDEF = _mm512_add_epi32(vacc9x89ABCDEF, _mm512_srli_epi64(vacc9x89ABCDEF, 32));
379 __m512i vacc9x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum9x01234567, vidx, vsum9x89ABCDEF);
380 const __m512i vsum10x01234567 = _mm512_add_epi32(vacc10x01234567, _mm512_srli_epi64(vacc10x01234567, 32));
381 const __m512i vsum10x89ABCDEF = _mm512_add_epi32(vacc10x89ABCDEF, _mm512_srli_epi64(vacc10x89ABCDEF, 32));
382 __m512i vacc10x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum10x01234567, vidx, vsum10x89ABCDEF);
383 const __m512i vsum11x01234567 = _mm512_add_epi32(vacc11x01234567, _mm512_srli_epi64(vacc11x01234567, 32));
384 const __m512i vsum11x89ABCDEF = _mm512_add_epi32(vacc11x89ABCDEF, _mm512_srli_epi64(vacc11x89ABCDEF, 32));
385 __m512i vacc11x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum11x01234567, vidx, vsum11x89ABCDEF);
386 __m512 vf0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
387 __m512 vf1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
388 __m512 vf2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
389 __m512 vf3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
390 __m512 vf4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
391 __m512 vf5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
392 __m512 vf6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
393 __m512 vf7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
394 __m512 vf8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
395 __m512 vf9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
396 __m512 vf10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
397 __m512 vf11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
398
399 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vf0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled0x0123456789ABCDEF);
400 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vf1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled1x0123456789ABCDEF);
401 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vf2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled2x0123456789ABCDEF);
402 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vf3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled3x0123456789ABCDEF);
403 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vf4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled4x0123456789ABCDEF);
404 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vf5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled5x0123456789ABCDEF);
405 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vf6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled6x0123456789ABCDEF);
406 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vf7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled7x0123456789ABCDEF);
407 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vf8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled8x0123456789ABCDEF);
408 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vf9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled9x0123456789ABCDEF);
409 vscaled10x0123456789ABCDEF = _mm512_fmadd_ps(vf10x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled10x0123456789ABCDEF);
410 vscaled11x0123456789ABCDEF = _mm512_fmadd_ps(vf11x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled11x0123456789ABCDEF);
411 }
412
413 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
414 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
415 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
416 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
417 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
418 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
419 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
420 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
421 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, _mm512_set1_ps(quantization_params[8].inv_scale));
422 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, _mm512_set1_ps(quantization_params[9].inv_scale));
423 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, _mm512_set1_ps(quantization_params[10].inv_scale));
424 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, _mm512_set1_ps(quantization_params[11].inv_scale));
425
426 const __m512 vbias0123456789ABCDEF = _mm512_loadu_ps((const float*) w);
427 w = (const float*) w + 16;
428
429 vscaled0x0123456789ABCDEF = _mm512_add_ps(vscaled0x0123456789ABCDEF, vbias0123456789ABCDEF);
430 vscaled1x0123456789ABCDEF = _mm512_add_ps(vscaled1x0123456789ABCDEF, vbias0123456789ABCDEF);
431 vscaled2x0123456789ABCDEF = _mm512_add_ps(vscaled2x0123456789ABCDEF, vbias0123456789ABCDEF);
432 vscaled3x0123456789ABCDEF = _mm512_add_ps(vscaled3x0123456789ABCDEF, vbias0123456789ABCDEF);
433 vscaled4x0123456789ABCDEF = _mm512_add_ps(vscaled4x0123456789ABCDEF, vbias0123456789ABCDEF);
434 vscaled5x0123456789ABCDEF = _mm512_add_ps(vscaled5x0123456789ABCDEF, vbias0123456789ABCDEF);
435 vscaled6x0123456789ABCDEF = _mm512_add_ps(vscaled6x0123456789ABCDEF, vbias0123456789ABCDEF);
436 vscaled7x0123456789ABCDEF = _mm512_add_ps(vscaled7x0123456789ABCDEF, vbias0123456789ABCDEF);
437 vscaled8x0123456789ABCDEF = _mm512_add_ps(vscaled8x0123456789ABCDEF, vbias0123456789ABCDEF);
438 vscaled9x0123456789ABCDEF = _mm512_add_ps(vscaled9x0123456789ABCDEF, vbias0123456789ABCDEF);
439 vscaled10x0123456789ABCDEF = _mm512_add_ps(vscaled10x0123456789ABCDEF, vbias0123456789ABCDEF);
440 vscaled11x0123456789ABCDEF = _mm512_add_ps(vscaled11x0123456789ABCDEF, vbias0123456789ABCDEF);
441
442 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
443 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
444 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
445 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
446 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
447 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
448 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
449 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
450 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
451 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
452 vscaled10x0123456789ABCDEF = _mm512_max_ps(vscaled10x0123456789ABCDEF, voutput_min);
453 vscaled11x0123456789ABCDEF = _mm512_max_ps(vscaled11x0123456789ABCDEF, voutput_min);
454
455 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
456 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
457 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
458 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
459 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
460 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
461 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
462 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
463 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
464 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
465 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max);
466 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max);
467
468 if XNN_LIKELY(nc >= 16) {
469 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
470 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
471 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
472 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
473 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
474 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
475 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
476 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
477 _mm512_storeu_ps(c8, vscaled8x0123456789ABCDEF);
478 _mm512_storeu_ps(c9, vscaled9x0123456789ABCDEF);
479 _mm512_storeu_ps(c10, vscaled10x0123456789ABCDEF);
480 _mm512_storeu_ps(c11, vscaled11x0123456789ABCDEF);
481
482 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
483 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
484 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
485 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
486 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
487 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
488 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
489 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
490 a8 = (const int8_t*) ((uintptr_t) a8 - kc);
491 a9 = (const int8_t*) ((uintptr_t) a9 - kc);
492 a10 = (const int8_t*) ((uintptr_t) a10 - kc);
493 a11 = (const int8_t*) ((uintptr_t) a11 - kc);
494
495 c0 = (float*) ((uintptr_t) c0 + cn_stride);
496 c1 = (float*) ((uintptr_t) c1 + cn_stride);
497 c2 = (float*) ((uintptr_t) c2 + cn_stride);
498 c3 = (float*) ((uintptr_t) c3 + cn_stride);
499 c4 = (float*) ((uintptr_t) c4 + cn_stride);
500 c5 = (float*) ((uintptr_t) c5 + cn_stride);
501 c6 = (float*) ((uintptr_t) c6 + cn_stride);
502 c7 = (float*) ((uintptr_t) c7 + cn_stride);
503 c8 = (float*) ((uintptr_t) c8 + cn_stride);
504 c9 = (float*) ((uintptr_t) c9 + cn_stride);
505 c10 = (float*) ((uintptr_t) c10 + cn_stride);
506 c11 = (float*) ((uintptr_t) c11 + cn_stride);
507
508 nc -= 16;
509 } else {
510 // Prepare mask for valid 32-bit elements (depends on nc).
511 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
512 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
513 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
514 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
515 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
516 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
517 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
518 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
519 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
520 _mm512_mask_storeu_ps(c8, vmask, vscaled8x0123456789ABCDEF);
521 _mm512_mask_storeu_ps(c9, vmask, vscaled9x0123456789ABCDEF);
522 _mm512_mask_storeu_ps(c10, vmask, vscaled10x0123456789ABCDEF);
523 _mm512_mask_storeu_ps(c11, vmask, vscaled11x0123456789ABCDEF);
524 nc = 0;
525 }
526 } while (nc != 0);
527 }