Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-12x16c8-minmax-avx512vnnigfni.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
20
21 void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_12x16c8__avx512vnnigfni(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 const int8_t* restrict a,
26 size_t a_stride,
27 const void* restrict w,
28 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const struct xnn_f32_qb4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 12);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 const int8_t* a0 = a;
45 float* c0 = c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 a1 = a0;
50 c1 = c0;
51 }
52 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
53 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 a2 = a1;
56 c2 = c1;
57 }
58 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
59 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
60 if XNN_UNPREDICTABLE(mr < 4) {
61 a3 = a2;
62 c3 = c2;
63 }
64 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
65 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
66 if XNN_UNPREDICTABLE(mr <= 4) {
67 a4 = a3;
68 c4 = c3;
69 }
70 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
71 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
72 if XNN_UNPREDICTABLE(mr < 6) {
73 a5 = a4;
74 c5 = c4;
75 }
76 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
77 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
78 if XNN_UNPREDICTABLE(mr <= 6) {
79 a6 = a5;
80 c6 = c5;
81 }
82 const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
83 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
84 if XNN_UNPREDICTABLE(mr < 8) {
85 a7 = a6;
86 c7 = c6;
87 }
88 const int8_t* a8 = (const int8_t*) ((uintptr_t) a7 + a_stride);
89 float* c8 = (float*) ((uintptr_t) c7 + cm_stride);
90 if XNN_UNPREDICTABLE(mr <= 8) {
91 a8 = a7;
92 c8 = c7;
93 }
94 const int8_t* a9 = (const int8_t*) ((uintptr_t) a8 + a_stride);
95 float* c9 = (float*) ((uintptr_t) c8 + cm_stride);
96 if XNN_UNPREDICTABLE(mr < 10) {
97 a9 = a8;
98 c9 = c8;
99 }
100 const int8_t* a10 = (const int8_t*) ((uintptr_t) a9 + a_stride);
101 float* c10 = (float*) ((uintptr_t) c9 + cm_stride);
102 if XNN_UNPREDICTABLE(mr <= 10) {
103 a10 = a9;
104 c10 = c9;
105 }
106 const int8_t* a11 = (const int8_t*) ((uintptr_t) a10 + a_stride);
107 float* c11 = (float*) ((uintptr_t) c10 + cm_stride);
108 if XNN_UNPREDICTABLE(mr != 12) {
109 a11 = a10;
110 c11 = c10;
111 }
112 size_t bl = params->scalar.blocksize;
113 assert(bl != 0);
114 assert(bl <= kc);
115 assert(kc % bl == 0);
116 assert(bl % 32 == 0);
117
118 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
119 XNN_FORCE_REALIZATION(vsign_mask);
120 const __m512 vinput_zero_point0 = _mm512_set1_ps((float) quantization_params[0].zero_point + 128);
121 const __m512 vinput_zero_point1 = _mm512_set1_ps((float) quantization_params[1].zero_point + 128);
122 const __m512 vinput_zero_point2 = _mm512_set1_ps((float) quantization_params[2].zero_point + 128);
123 const __m512 vinput_zero_point3 = _mm512_set1_ps((float) quantization_params[3].zero_point + 128);
124 const __m512 vinput_zero_point4 = _mm512_set1_ps((float) quantization_params[4].zero_point + 128);
125 const __m512 vinput_zero_point5 = _mm512_set1_ps((float) quantization_params[5].zero_point + 128);
126 const __m512 vinput_zero_point6 = _mm512_set1_ps((float) quantization_params[6].zero_point + 128);
127 const __m512 vinput_zero_point7 = _mm512_set1_ps((float) quantization_params[7].zero_point + 128);
128 const __m512 vinput_zero_point8 = _mm512_set1_ps((float) quantization_params[8].zero_point + 128);
129 const __m512 vinput_zero_point9 = _mm512_set1_ps((float) quantization_params[9].zero_point + 128);
130 const __m512 vinput_zero_point10 = _mm512_set1_ps((float) quantization_params[10].zero_point + 128);
131 const __m512 vinput_zero_point11 = _mm512_set1_ps((float) quantization_params[11].zero_point + 128);
132 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
133 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
134 // XNN_FORCE_REALIZATION(voutput_min);
135 // XNN_FORCE_REALIZATION(voutput_max);
136 const __m512i vmask = _mm512_set1_epi8(0xF0);
137 XNN_FORCE_REALIZATION(vmask);
138 const __m512i vshl4 = _mm512_set1_epi64(0x01020408);
139 XNN_FORCE_REALIZATION(vshl4);
140 do {
141 const __m512 vksum0123456789ABCDEF = _mm512_loadu_ps(w);
142 __m512 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point0);
143 __m512 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point1);
144 __m512 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point2);
145 __m512 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point3);
146 __m512 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point4);
147 __m512 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point5);
148 __m512 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point6);
149 __m512 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point7);
150 __m512 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point8);
151 __m512 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point9);
152 __m512 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point10);
153 __m512 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point11);
154 w = (const int32_t*) w + 16;
155
156 for (size_t kb=0; kb < kc; kb+=bl) {
157 __m512i vacc0x01234567 = _mm512_setzero_epi32();
158 __m512i vacc0x89ABCDEF = _mm512_setzero_epi32();
159 __m512i vacc1x01234567 = _mm512_setzero_epi32();
160 __m512i vacc1x89ABCDEF = _mm512_setzero_epi32();
161 __m512i vacc2x01234567 = _mm512_setzero_epi32();
162 __m512i vacc2x89ABCDEF = _mm512_setzero_epi32();
163 __m512i vacc3x01234567 = _mm512_setzero_epi32();
164 __m512i vacc3x89ABCDEF = _mm512_setzero_epi32();
165 __m512i vacc4x01234567 = _mm512_setzero_epi32();
166 __m512i vacc4x89ABCDEF = _mm512_setzero_epi32();
167 __m512i vacc5x01234567 = _mm512_setzero_epi32();
168 __m512i vacc5x89ABCDEF = _mm512_setzero_epi32();
169 __m512i vacc6x01234567 = _mm512_setzero_epi32();
170 __m512i vacc6x89ABCDEF = _mm512_setzero_epi32();
171 __m512i vacc7x01234567 = _mm512_setzero_epi32();
172 __m512i vacc7x89ABCDEF = _mm512_setzero_epi32();
173 __m512i vacc8x01234567 = _mm512_setzero_epi32();
174 __m512i vacc8x89ABCDEF = _mm512_setzero_epi32();
175 __m512i vacc9x01234567 = _mm512_setzero_epi32();
176 __m512i vacc9x89ABCDEF = _mm512_setzero_epi32();
177 __m512i vacc10x01234567 = _mm512_setzero_epi32();
178 __m512i vacc10x89ABCDEF = _mm512_setzero_epi32();
179 __m512i vacc11x01234567 = _mm512_setzero_epi32();
180 __m512i vacc11x89ABCDEF = _mm512_setzero_epi32();
181 size_t k = bl;
182 while (k >= 16 * sizeof(int8_t)) {
183 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
184 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
185 a0 += 16;
186 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
187 const __m512i va1x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
188 a1 += 16;
189 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
190 const __m512i va2x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
191 a2 += 16;
192 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
193 const __m512i va3x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
194 a3 += 16;
195 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
196 const __m512i va4x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
197 a4 += 16;
198 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
199 const __m512i va5x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
200 a5 += 16;
201 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
202 const __m512i va6x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
203 a6 += 16;
204 const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask);
205 const __m512i va7x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
206 a7 += 16;
207 const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask);
208 const __m512i va8x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
209 a8 += 16;
210 const __m512i va9x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9)), vsign_mask);
211 const __m512i va9x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
212 a9 += 16;
213 const __m512i va10x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a10)), vsign_mask);
214 const __m512i va10x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a10 + 8)), vsign_mask);
215 a10 += 16;
216 const __m512i va11x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a11)), vsign_mask);
217 const __m512i va11x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a11 + 8)), vsign_mask);
218 a11 += 16;
219
220 const __m512i vbb01234567x01234567 = _mm512_loadu_si512(w);
221 const __m512i vbb89ABCDEFx01234567 = _mm512_loadu_si512((const int8_t*) w + 64);
222 const __m512i vb01234567x01234567 = _mm512_gf2p8affine_epi64_epi8(vbb01234567x01234567, vshl4, 0);
223 const __m512i vb89ABCDEFx01234567 = _mm512_gf2p8affine_epi64_epi8(vbb89ABCDEFx01234567, vshl4, 0);
224 const __m512i vb01234567x89ABCDEF = _mm512_and_si512(vbb01234567x01234567, vmask);
225 const __m512i vb89ABCDEFx89ABCDEF = _mm512_and_si512(vbb89ABCDEFx01234567, vmask);
226
227 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
228 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
229 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
230 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
231 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
232 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
233 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
234 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
235 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
236 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
237 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
238 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
239 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
240 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
241 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567);
242 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567);
243 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567);
244 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567);
245 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x01234567, vb01234567x01234567);
246 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x01234567, vb89ABCDEFx01234567);
247 vacc10x01234567 = _mm512_dpbusd_epi32(vacc10x01234567, va10x01234567, vb01234567x01234567);
248 vacc10x89ABCDEF = _mm512_dpbusd_epi32(vacc10x89ABCDEF, va10x01234567, vb89ABCDEFx01234567);
249 vacc11x01234567 = _mm512_dpbusd_epi32(vacc11x01234567, va11x01234567, vb01234567x01234567);
250 vacc11x89ABCDEF = _mm512_dpbusd_epi32(vacc11x89ABCDEF, va11x01234567, vb89ABCDEFx01234567);
251 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
252 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
253 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x89ABCDEF, vb01234567x89ABCDEF);
254 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x89ABCDEF, vb89ABCDEFx89ABCDEF);
255 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x89ABCDEF, vb01234567x89ABCDEF);
256 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x89ABCDEF, vb89ABCDEFx89ABCDEF);
257 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x89ABCDEF, vb01234567x89ABCDEF);
258 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x89ABCDEF, vb89ABCDEFx89ABCDEF);
259 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x89ABCDEF, vb01234567x89ABCDEF);
260 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x89ABCDEF, vb89ABCDEFx89ABCDEF);
261 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x89ABCDEF, vb01234567x89ABCDEF);
262 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x89ABCDEF, vb89ABCDEFx89ABCDEF);
263 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x89ABCDEF, vb01234567x89ABCDEF);
264 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x89ABCDEF, vb89ABCDEFx89ABCDEF);
265 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x89ABCDEF, vb01234567x89ABCDEF);
266 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x89ABCDEF, vb89ABCDEFx89ABCDEF);
267 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x89ABCDEF, vb01234567x89ABCDEF);
268 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x89ABCDEF, vb89ABCDEFx89ABCDEF);
269 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x89ABCDEF, vb01234567x89ABCDEF);
270 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x89ABCDEF, vb89ABCDEFx89ABCDEF);
271 vacc10x01234567 = _mm512_dpbusd_epi32(vacc10x01234567, va10x89ABCDEF, vb01234567x89ABCDEF);
272 vacc10x89ABCDEF = _mm512_dpbusd_epi32(vacc10x89ABCDEF, va10x89ABCDEF, vb89ABCDEFx89ABCDEF);
273 vacc11x01234567 = _mm512_dpbusd_epi32(vacc11x01234567, va11x89ABCDEF, vb01234567x89ABCDEF);
274 vacc11x89ABCDEF = _mm512_dpbusd_epi32(vacc11x89ABCDEF, va11x89ABCDEF, vb89ABCDEFx89ABCDEF);
275
276 w = (const int8_t*) w + 128;
277 k -= 16 * sizeof(int8_t);
278 }
279
280 if (k != 0) {
281 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
282 a0 += 8;
283 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
284 a1 += 8;
285 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
286 a2 += 8;
287 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
288 a3 += 8;
289 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
290 a4 += 8;
291 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
292 a5 += 8;
293 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
294 a6 += 8;
295 const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask);
296 a7 += 8;
297 const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask);
298 a8 += 8;
299 const __m512i va9x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9)), vsign_mask);
300 a9 += 8;
301 const __m512i va10x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a10)), vsign_mask);
302 a10 += 8;
303 const __m512i va11x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a11)), vsign_mask);
304 a11 += 8;
305
306 const __m512i vbb01234567x01234567 = _mm512_load_si512(w);
307 const __m512i vbb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
308 const __m512i vb01234567x01234567 = _mm512_gf2p8affine_epi64_epi8(vbb01234567x01234567, vshl4, 0);
309 const __m512i vb89ABCDEFx01234567 = _mm512_gf2p8affine_epi64_epi8(vbb89ABCDEFx01234567, vshl4, 0);
310
311 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
312 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
313 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
314 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
315 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
316 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
317 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
318 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
319 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
320 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
321 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
322 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
323 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
324 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
325 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567);
326 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567);
327 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567);
328 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567);
329 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x01234567, vb01234567x01234567);
330 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x01234567, vb89ABCDEFx01234567);
331 vacc10x01234567 = _mm512_dpbusd_epi32(vacc10x01234567, va10x01234567, vb01234567x01234567);
332 vacc10x89ABCDEF = _mm512_dpbusd_epi32(vacc10x89ABCDEF, va10x01234567, vb89ABCDEFx01234567);
333 vacc11x01234567 = _mm512_dpbusd_epi32(vacc11x01234567, va11x01234567, vb01234567x01234567);
334 vacc11x89ABCDEF = _mm512_dpbusd_epi32(vacc11x89ABCDEF, va11x01234567, vb89ABCDEFx01234567);
335
336 w = (const int8_t*) w + 128;
337 k -= 8 * sizeof(int8_t);
338 }
339 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_castsi512_ps(_mm512_slli_epi32(
340 _mm512_cvtepu16_epi32(_mm256_load_si256((const __m256i*) w)), 16));
341 w = (const uint16_t*) w + 16;
342
343 // Add adjacent pairs
344 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
345 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
346 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
347 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
348 const __m512i vsum1x01234567 = _mm512_add_epi32(vacc1x01234567, _mm512_srli_epi64(vacc1x01234567, 32));
349 const __m512i vsum1x89ABCDEF = _mm512_add_epi32(vacc1x89ABCDEF, _mm512_srli_epi64(vacc1x89ABCDEF, 32));
350 __m512i vacc1x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum1x01234567, vidx, vsum1x89ABCDEF);
351 const __m512i vsum2x01234567 = _mm512_add_epi32(vacc2x01234567, _mm512_srli_epi64(vacc2x01234567, 32));
352 const __m512i vsum2x89ABCDEF = _mm512_add_epi32(vacc2x89ABCDEF, _mm512_srli_epi64(vacc2x89ABCDEF, 32));
353 __m512i vacc2x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum2x01234567, vidx, vsum2x89ABCDEF);
354 const __m512i vsum3x01234567 = _mm512_add_epi32(vacc3x01234567, _mm512_srli_epi64(vacc3x01234567, 32));
355 const __m512i vsum3x89ABCDEF = _mm512_add_epi32(vacc3x89ABCDEF, _mm512_srli_epi64(vacc3x89ABCDEF, 32));
356 __m512i vacc3x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum3x01234567, vidx, vsum3x89ABCDEF);
357 const __m512i vsum4x01234567 = _mm512_add_epi32(vacc4x01234567, _mm512_srli_epi64(vacc4x01234567, 32));
358 const __m512i vsum4x89ABCDEF = _mm512_add_epi32(vacc4x89ABCDEF, _mm512_srli_epi64(vacc4x89ABCDEF, 32));
359 __m512i vacc4x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum4x01234567, vidx, vsum4x89ABCDEF);
360 const __m512i vsum5x01234567 = _mm512_add_epi32(vacc5x01234567, _mm512_srli_epi64(vacc5x01234567, 32));
361 const __m512i vsum5x89ABCDEF = _mm512_add_epi32(vacc5x89ABCDEF, _mm512_srli_epi64(vacc5x89ABCDEF, 32));
362 __m512i vacc5x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum5x01234567, vidx, vsum5x89ABCDEF);
363 const __m512i vsum6x01234567 = _mm512_add_epi32(vacc6x01234567, _mm512_srli_epi64(vacc6x01234567, 32));
364 const __m512i vsum6x89ABCDEF = _mm512_add_epi32(vacc6x89ABCDEF, _mm512_srli_epi64(vacc6x89ABCDEF, 32));
365 __m512i vacc6x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum6x01234567, vidx, vsum6x89ABCDEF);
366 const __m512i vsum7x01234567 = _mm512_add_epi32(vacc7x01234567, _mm512_srli_epi64(vacc7x01234567, 32));
367 const __m512i vsum7x89ABCDEF = _mm512_add_epi32(vacc7x89ABCDEF, _mm512_srli_epi64(vacc7x89ABCDEF, 32));
368 __m512i vacc7x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum7x01234567, vidx, vsum7x89ABCDEF);
369 const __m512i vsum8x01234567 = _mm512_add_epi32(vacc8x01234567, _mm512_srli_epi64(vacc8x01234567, 32));
370 const __m512i vsum8x89ABCDEF = _mm512_add_epi32(vacc8x89ABCDEF, _mm512_srli_epi64(vacc8x89ABCDEF, 32));
371 __m512i vacc8x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum8x01234567, vidx, vsum8x89ABCDEF);
372 const __m512i vsum9x01234567 = _mm512_add_epi32(vacc9x01234567, _mm512_srli_epi64(vacc9x01234567, 32));
373 const __m512i vsum9x89ABCDEF = _mm512_add_epi32(vacc9x89ABCDEF, _mm512_srli_epi64(vacc9x89ABCDEF, 32));
374 __m512i vacc9x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum9x01234567, vidx, vsum9x89ABCDEF);
375 const __m512i vsum10x01234567 = _mm512_add_epi32(vacc10x01234567, _mm512_srli_epi64(vacc10x01234567, 32));
376 const __m512i vsum10x89ABCDEF = _mm512_add_epi32(vacc10x89ABCDEF, _mm512_srli_epi64(vacc10x89ABCDEF, 32));
377 __m512i vacc10x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum10x01234567, vidx, vsum10x89ABCDEF);
378 const __m512i vsum11x01234567 = _mm512_add_epi32(vacc11x01234567, _mm512_srli_epi64(vacc11x01234567, 32));
379 const __m512i vsum11x89ABCDEF = _mm512_add_epi32(vacc11x89ABCDEF, _mm512_srli_epi64(vacc11x89ABCDEF, 32));
380 __m512i vacc11x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum11x01234567, vidx, vsum11x89ABCDEF);
381 __m512 vf0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
382 __m512 vf1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
383 __m512 vf2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
384 __m512 vf3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
385 __m512 vf4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
386 __m512 vf5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
387 __m512 vf6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
388 __m512 vf7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
389 __m512 vf8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
390 __m512 vf9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
391 __m512 vf10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
392 __m512 vf11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
393
394 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vf0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled0x0123456789ABCDEF);
395 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vf1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled1x0123456789ABCDEF);
396 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vf2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled2x0123456789ABCDEF);
397 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vf3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled3x0123456789ABCDEF);
398 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vf4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled4x0123456789ABCDEF);
399 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vf5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled5x0123456789ABCDEF);
400 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vf6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled6x0123456789ABCDEF);
401 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vf7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled7x0123456789ABCDEF);
402 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vf8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled8x0123456789ABCDEF);
403 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vf9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled9x0123456789ABCDEF);
404 vscaled10x0123456789ABCDEF = _mm512_fmadd_ps(vf10x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled10x0123456789ABCDEF);
405 vscaled11x0123456789ABCDEF = _mm512_fmadd_ps(vf11x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled11x0123456789ABCDEF);
406 }
407
408 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
409 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
410 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
411 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
412 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
413 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
414 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
415 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
416 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, _mm512_set1_ps(quantization_params[8].inv_scale));
417 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, _mm512_set1_ps(quantization_params[9].inv_scale));
418 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, _mm512_set1_ps(quantization_params[10].inv_scale));
419 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, _mm512_set1_ps(quantization_params[11].inv_scale));
420
421 const __m512 vbias0123456789ABCDEF = _mm512_loadu_ps((const float*) w);
422 w = (const float*) w + 16;
423
424 vscaled0x0123456789ABCDEF = _mm512_add_ps(vscaled0x0123456789ABCDEF, vbias0123456789ABCDEF);
425 vscaled1x0123456789ABCDEF = _mm512_add_ps(vscaled1x0123456789ABCDEF, vbias0123456789ABCDEF);
426 vscaled2x0123456789ABCDEF = _mm512_add_ps(vscaled2x0123456789ABCDEF, vbias0123456789ABCDEF);
427 vscaled3x0123456789ABCDEF = _mm512_add_ps(vscaled3x0123456789ABCDEF, vbias0123456789ABCDEF);
428 vscaled4x0123456789ABCDEF = _mm512_add_ps(vscaled4x0123456789ABCDEF, vbias0123456789ABCDEF);
429 vscaled5x0123456789ABCDEF = _mm512_add_ps(vscaled5x0123456789ABCDEF, vbias0123456789ABCDEF);
430 vscaled6x0123456789ABCDEF = _mm512_add_ps(vscaled6x0123456789ABCDEF, vbias0123456789ABCDEF);
431 vscaled7x0123456789ABCDEF = _mm512_add_ps(vscaled7x0123456789ABCDEF, vbias0123456789ABCDEF);
432 vscaled8x0123456789ABCDEF = _mm512_add_ps(vscaled8x0123456789ABCDEF, vbias0123456789ABCDEF);
433 vscaled9x0123456789ABCDEF = _mm512_add_ps(vscaled9x0123456789ABCDEF, vbias0123456789ABCDEF);
434 vscaled10x0123456789ABCDEF = _mm512_add_ps(vscaled10x0123456789ABCDEF, vbias0123456789ABCDEF);
435 vscaled11x0123456789ABCDEF = _mm512_add_ps(vscaled11x0123456789ABCDEF, vbias0123456789ABCDEF);
436
437 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
438 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
439 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
440 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
441 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
442 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
443 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
444 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
445 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
446 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
447 vscaled10x0123456789ABCDEF = _mm512_max_ps(vscaled10x0123456789ABCDEF, voutput_min);
448 vscaled11x0123456789ABCDEF = _mm512_max_ps(vscaled11x0123456789ABCDEF, voutput_min);
449
450 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
451 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
452 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
453 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
454 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
455 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
456 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
457 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
458 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
459 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
460 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max);
461 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max);
462
463 if XNN_LIKELY(nc >= 16) {
464 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
465 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
466 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
467 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
468 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
469 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
470 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
471 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
472 _mm512_storeu_ps(c8, vscaled8x0123456789ABCDEF);
473 _mm512_storeu_ps(c9, vscaled9x0123456789ABCDEF);
474 _mm512_storeu_ps(c10, vscaled10x0123456789ABCDEF);
475 _mm512_storeu_ps(c11, vscaled11x0123456789ABCDEF);
476
477 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
478 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
479 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
480 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
481 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
482 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
483 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
484 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
485 a8 = (const int8_t*) ((uintptr_t) a8 - kc);
486 a9 = (const int8_t*) ((uintptr_t) a9 - kc);
487 a10 = (const int8_t*) ((uintptr_t) a10 - kc);
488 a11 = (const int8_t*) ((uintptr_t) a11 - kc);
489
490 c0 = (float*) ((uintptr_t) c0 + cn_stride);
491 c1 = (float*) ((uintptr_t) c1 + cn_stride);
492 c2 = (float*) ((uintptr_t) c2 + cn_stride);
493 c3 = (float*) ((uintptr_t) c3 + cn_stride);
494 c4 = (float*) ((uintptr_t) c4 + cn_stride);
495 c5 = (float*) ((uintptr_t) c5 + cn_stride);
496 c6 = (float*) ((uintptr_t) c6 + cn_stride);
497 c7 = (float*) ((uintptr_t) c7 + cn_stride);
498 c8 = (float*) ((uintptr_t) c8 + cn_stride);
499 c9 = (float*) ((uintptr_t) c9 + cn_stride);
500 c10 = (float*) ((uintptr_t) c10 + cn_stride);
501 c11 = (float*) ((uintptr_t) c11 + cn_stride);
502
503 nc -= 16;
504 } else {
505 // Prepare mask for valid 32-bit elements (depends on nc).
506 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
507 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
508 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
509 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
510 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
511 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
512 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
513 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
514 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
515 _mm512_mask_storeu_ps(c8, vmask, vscaled8x0123456789ABCDEF);
516 _mm512_mask_storeu_ps(c9, vmask, vscaled9x0123456789ABCDEF);
517 _mm512_mask_storeu_ps(c10, vmask, vscaled10x0123456789ABCDEF);
518 _mm512_mask_storeu_ps(c11, vmask, vscaled11x0123456789ABCDEF);
519 nc = 0;
520 }
521 } while (nc != 0);
522 }