Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-7x16c8-minmax-avx512skx-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx16c8-avx512skx.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include "xnnpack/gemm.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/prefetch.h"
18
19 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_7x16c8__avx512skx_prfm(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
30 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 7);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(int8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 8 * sizeof(int8_t));
42 const int8_t* a0 = a;
43 float* c0 = c;
44 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
45 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 a1 = a0;
48 c1 = c0;
49 }
50 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 a2 = a1;
54 c2 = c1;
55 }
56 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
57 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
58 if XNN_UNPREDICTABLE(mr < 4) {
59 a3 = a2;
60 c3 = c2;
61 }
62 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
63 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
64 if XNN_UNPREDICTABLE(mr <= 4) {
65 a4 = a3;
66 c4 = c3;
67 }
68 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
69 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
70 if XNN_UNPREDICTABLE(mr < 6) {
71 a5 = a4;
72 c5 = c4;
73 }
74 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
75 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
76 if XNN_UNPREDICTABLE(mr <= 6) {
77 a6 = a5;
78 c6 = c5;
79 }
80
81 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
82 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point);
83 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point);
84 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point);
85 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point);
86 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point);
87 const __m512i vinput_zero_point5 = _mm512_set1_epi32((int) quantization_params[5].zero_point);
88 const __m512i vinput_zero_point6 = _mm512_set1_epi32((int) quantization_params[6].zero_point);
89 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
90 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
91 // XNN_FORCE_REALIZATION(voutput_min);
92 // XNN_FORCE_REALIZATION(voutput_max);
93
94 do {
95 const __m512i vksum0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
96 const __m512i vksum4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
97 const __m512i vksum89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
98 const __m512i vksumCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
99
100 __m512i vacc0x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point0);
101 __m512i vacc0x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point0);
102 __m512i vacc0x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point0);
103 __m512i vacc0xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point0);
104 __m512i vacc1x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point1);
105 __m512i vacc1x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point1);
106 __m512i vacc1x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point1);
107 __m512i vacc1xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point1);
108 __m512i vacc2x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point2);
109 __m512i vacc2x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point2);
110 __m512i vacc2x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point2);
111 __m512i vacc2xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point2);
112 __m512i vacc3x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point3);
113 __m512i vacc3x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point3);
114 __m512i vacc3x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point3);
115 __m512i vacc3xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point3);
116 __m512i vacc4x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point4);
117 __m512i vacc4x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point4);
118 __m512i vacc4x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point4);
119 __m512i vacc4xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point4);
120 __m512i vacc5x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point5);
121 __m512i vacc5x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point5);
122 __m512i vacc5x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point5);
123 __m512i vacc5xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point5);
124 __m512i vacc6x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point6);
125 __m512i vacc6x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point6);
126 __m512i vacc6x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point6);
127 __m512i vacc6xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point6);
128 w = (const int32_t*) w + 16;
129
130 size_t k = kc;
131
132 while (k >= 8 * sizeof(int8_t)) {
133 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
134 a0 += 8;
135 const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
136 a1 += 8;
137 const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
138 a2 += 8;
139 const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
140 a3 += 8;
141 const __m512i va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
142 a4 += 8;
143 const __m512i va5 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a5)));
144 a5 += 8;
145 const __m512i va6 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a6)));
146 a6 += 8;
147
148 const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
149 xnn_prefetch_to_l1((const int8_t*) w + 896);
150
151 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
152 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
153 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
154 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
155 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
156 vacc5x0123 = _mm512_add_epi32(vacc5x0123, _mm512_madd_epi16(va5, vb0123));
157 vacc6x0123 = _mm512_add_epi32(vacc6x0123, _mm512_madd_epi16(va6, vb0123));
158 const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)));
159
160 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
161 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
162 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
163 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
164 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
165 vacc5x4567 = _mm512_add_epi32(vacc5x4567, _mm512_madd_epi16(va5, vb4567));
166 vacc6x4567 = _mm512_add_epi32(vacc6x4567, _mm512_madd_epi16(va6, vb4567));
167 const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)));
168 xnn_prefetch_to_l1((const int8_t*) w + 960);
169
170 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
171 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
172 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
173 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
174 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
175 vacc5x89AB = _mm512_add_epi32(vacc5x89AB, _mm512_madd_epi16(va5, vb89AB));
176 vacc6x89AB = _mm512_add_epi32(vacc6x89AB, _mm512_madd_epi16(va6, vb89AB));
177 const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)));
178
179 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
180 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
181 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
182 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
183 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
184 vacc5xCDEF = _mm512_add_epi32(vacc5xCDEF, _mm512_madd_epi16(va5, vbCDEF));
185 vacc6xCDEF = _mm512_add_epi32(vacc6xCDEF, _mm512_madd_epi16(va6, vbCDEF));
186
187 w = (const int8_t*) w + 128;
188 k -= 8 * sizeof(int8_t);
189 }
190
191 // Add 4 adjacent sums
192 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
193 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
194 const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
195 const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
196 const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm512_unpackhi_epi32(vacc2x0123, vacc2x4567));
197 const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm512_unpackhi_epi32(vacc2x89AB, vacc2xCDEF));
198 const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm512_unpackhi_epi32(vacc3x0123, vacc3x4567));
199 const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm512_unpackhi_epi32(vacc3x89AB, vacc3xCDEF));
200 const __m512i vacc4x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x0123, vacc4x4567), _mm512_unpackhi_epi32(vacc4x0123, vacc4x4567));
201 const __m512i vacc4x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x89AB, vacc4xCDEF), _mm512_unpackhi_epi32(vacc4x89AB, vacc4xCDEF));
202 const __m512i vacc5x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x0123, vacc5x4567), _mm512_unpackhi_epi32(vacc5x0123, vacc5x4567));
203 const __m512i vacc5x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x89AB, vacc5xCDEF), _mm512_unpackhi_epi32(vacc5x89AB, vacc5xCDEF));
204 const __m512i vacc6x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x0123, vacc6x4567), _mm512_unpackhi_epi32(vacc6x0123, vacc6x4567));
205 const __m512i vacc6x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x89AB, vacc6xCDEF), _mm512_unpackhi_epi32(vacc6x89AB, vacc6xCDEF));
206
207 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
208 const __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
209 const __m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9DAEBF), _mm512_unpackhi_epi32(vacc2x04152637, vacc2x8C9DAEBF));
210 const __m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9DAEBF), _mm512_unpackhi_epi32(vacc3x04152637, vacc3x8C9DAEBF));
211 const __m512i vacc4x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x04152637, vacc4x8C9DAEBF), _mm512_unpackhi_epi32(vacc4x04152637, vacc4x8C9DAEBF));
212 const __m512i vacc5x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x04152637, vacc5x8C9DAEBF), _mm512_unpackhi_epi32(vacc5x04152637, vacc5x8C9DAEBF));
213 const __m512i vacc6x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x04152637, vacc6x8C9DAEBF), _mm512_unpackhi_epi32(vacc6x04152637, vacc6x8C9DAEBF));
214
215 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
216 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
217 __m512i vacc1x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc1x084C195D2A6E3B7F);
218 __m512i vacc2x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc2x084C195D2A6E3B7F);
219 __m512i vacc3x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc3x084C195D2A6E3B7F);
220 __m512i vacc4x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc4x084C195D2A6E3B7F);
221 __m512i vacc5x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc5x084C195D2A6E3B7F);
222 __m512i vacc6x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc6x084C195D2A6E3B7F);
223
224 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
225 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
226 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
227 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
228 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
229 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
230 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
231
232 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
233 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
234 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
235 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
236 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
237 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
238 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
239
240 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
241 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
242 w = (const float*) w + 32;
243
244 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
245 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
246 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
247 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
248 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
249 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
250 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
251
252 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
253 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
254 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
255 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
256 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
257 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
258 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
259
260 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
261 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
262 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
263 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
264 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
265 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
266 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
267
268 if XNN_LIKELY(nc >= 16) {
269 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
270 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
271 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
272 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
273 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
274 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
275 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
276
277 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
278 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
279 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
280 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
281 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
282 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
283 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
284
285 c0 = (float*) ((uintptr_t) c0 + cn_stride);
286 c1 = (float*) ((uintptr_t) c1 + cn_stride);
287 c2 = (float*) ((uintptr_t) c2 + cn_stride);
288 c3 = (float*) ((uintptr_t) c3 + cn_stride);
289 c4 = (float*) ((uintptr_t) c4 + cn_stride);
290 c5 = (float*) ((uintptr_t) c5 + cn_stride);
291 c6 = (float*) ((uintptr_t) c6 + cn_stride);
292
293 nc -= 16;
294 } else {
295 // Prepare mask for valid 32-bit elements (depends on nc).
296 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
297 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
298 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
299 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
300 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
301 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
302 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
303 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
304 nc = 0;
305 }
306 } while (nc != 0);
307 }