Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-7x16c8-minmax-avx512skx.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
18 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_7x16c8__avx512skx(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
29 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 7);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 8 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 float* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
62 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
63 if XNN_UNPREDICTABLE(mr <= 4) {
64 a4 = a3;
65 c4 = c3;
66 }
67 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
68 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
69 if XNN_UNPREDICTABLE(mr < 6) {
70 a5 = a4;
71 c5 = c4;
72 }
73 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
74 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
75 if XNN_UNPREDICTABLE(mr <= 6) {
76 a6 = a5;
77 c6 = c5;
78 }
79
80 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
81 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point);
82 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point);
83 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point);
84 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point);
85 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point);
86 const __m512i vinput_zero_point5 = _mm512_set1_epi32((int) quantization_params[5].zero_point);
87 const __m512i vinput_zero_point6 = _mm512_set1_epi32((int) quantization_params[6].zero_point);
88 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
89 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
90 // XNN_FORCE_REALIZATION(voutput_min);
91 // XNN_FORCE_REALIZATION(voutput_max);
92
93 do {
94 const __m512i vksum0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
95 const __m512i vksum4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
96 const __m512i vksum89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
97 const __m512i vksumCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
98
99 __m512i vacc0x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point0);
100 __m512i vacc0x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point0);
101 __m512i vacc0x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point0);
102 __m512i vacc0xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point0);
103 __m512i vacc1x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point1);
104 __m512i vacc1x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point1);
105 __m512i vacc1x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point1);
106 __m512i vacc1xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point1);
107 __m512i vacc2x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point2);
108 __m512i vacc2x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point2);
109 __m512i vacc2x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point2);
110 __m512i vacc2xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point2);
111 __m512i vacc3x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point3);
112 __m512i vacc3x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point3);
113 __m512i vacc3x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point3);
114 __m512i vacc3xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point3);
115 __m512i vacc4x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point4);
116 __m512i vacc4x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point4);
117 __m512i vacc4x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point4);
118 __m512i vacc4xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point4);
119 __m512i vacc5x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point5);
120 __m512i vacc5x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point5);
121 __m512i vacc5x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point5);
122 __m512i vacc5xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point5);
123 __m512i vacc6x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point6);
124 __m512i vacc6x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point6);
125 __m512i vacc6x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point6);
126 __m512i vacc6xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point6);
127 w = (const int32_t*) w + 16;
128
129 size_t k = kc;
130
131 while (k >= 8 * sizeof(int8_t)) {
132 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
133 a0 += 8;
134 const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
135 a1 += 8;
136 const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
137 a2 += 8;
138 const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
139 a3 += 8;
140 const __m512i va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
141 a4 += 8;
142 const __m512i va5 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a5)));
143 a5 += 8;
144 const __m512i va6 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a6)));
145 a6 += 8;
146
147 const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
148
149 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
150 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
151 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
152 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
153 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
154 vacc5x0123 = _mm512_add_epi32(vacc5x0123, _mm512_madd_epi16(va5, vb0123));
155 vacc6x0123 = _mm512_add_epi32(vacc6x0123, _mm512_madd_epi16(va6, vb0123));
156 const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)));
157
158 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
159 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
160 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
161 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
162 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
163 vacc5x4567 = _mm512_add_epi32(vacc5x4567, _mm512_madd_epi16(va5, vb4567));
164 vacc6x4567 = _mm512_add_epi32(vacc6x4567, _mm512_madd_epi16(va6, vb4567));
165 const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)));
166
167 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
168 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
169 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
170 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
171 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
172 vacc5x89AB = _mm512_add_epi32(vacc5x89AB, _mm512_madd_epi16(va5, vb89AB));
173 vacc6x89AB = _mm512_add_epi32(vacc6x89AB, _mm512_madd_epi16(va6, vb89AB));
174 const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)));
175
176 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
177 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
178 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
179 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
180 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
181 vacc5xCDEF = _mm512_add_epi32(vacc5xCDEF, _mm512_madd_epi16(va5, vbCDEF));
182 vacc6xCDEF = _mm512_add_epi32(vacc6xCDEF, _mm512_madd_epi16(va6, vbCDEF));
183
184 w = (const int8_t*) w + 128;
185 k -= 8 * sizeof(int8_t);
186 }
187
188 // Add 4 adjacent sums
189 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
190 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
191 const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
192 const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
193 const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm512_unpackhi_epi32(vacc2x0123, vacc2x4567));
194 const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm512_unpackhi_epi32(vacc2x89AB, vacc2xCDEF));
195 const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm512_unpackhi_epi32(vacc3x0123, vacc3x4567));
196 const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm512_unpackhi_epi32(vacc3x89AB, vacc3xCDEF));
197 const __m512i vacc4x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x0123, vacc4x4567), _mm512_unpackhi_epi32(vacc4x0123, vacc4x4567));
198 const __m512i vacc4x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x89AB, vacc4xCDEF), _mm512_unpackhi_epi32(vacc4x89AB, vacc4xCDEF));
199 const __m512i vacc5x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x0123, vacc5x4567), _mm512_unpackhi_epi32(vacc5x0123, vacc5x4567));
200 const __m512i vacc5x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x89AB, vacc5xCDEF), _mm512_unpackhi_epi32(vacc5x89AB, vacc5xCDEF));
201 const __m512i vacc6x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x0123, vacc6x4567), _mm512_unpackhi_epi32(vacc6x0123, vacc6x4567));
202 const __m512i vacc6x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x89AB, vacc6xCDEF), _mm512_unpackhi_epi32(vacc6x89AB, vacc6xCDEF));
203
204 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
205 const __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
206 const __m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9DAEBF), _mm512_unpackhi_epi32(vacc2x04152637, vacc2x8C9DAEBF));
207 const __m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9DAEBF), _mm512_unpackhi_epi32(vacc3x04152637, vacc3x8C9DAEBF));
208 const __m512i vacc4x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x04152637, vacc4x8C9DAEBF), _mm512_unpackhi_epi32(vacc4x04152637, vacc4x8C9DAEBF));
209 const __m512i vacc5x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x04152637, vacc5x8C9DAEBF), _mm512_unpackhi_epi32(vacc5x04152637, vacc5x8C9DAEBF));
210 const __m512i vacc6x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x04152637, vacc6x8C9DAEBF), _mm512_unpackhi_epi32(vacc6x04152637, vacc6x8C9DAEBF));
211
212 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
213 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
214 __m512i vacc1x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc1x084C195D2A6E3B7F);
215 __m512i vacc2x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc2x084C195D2A6E3B7F);
216 __m512i vacc3x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc3x084C195D2A6E3B7F);
217 __m512i vacc4x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc4x084C195D2A6E3B7F);
218 __m512i vacc5x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc5x084C195D2A6E3B7F);
219 __m512i vacc6x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc6x084C195D2A6E3B7F);
220
221 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
222 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
223 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
224 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
225 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
226 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
227 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
228
229 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
230 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
231 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
232 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
233 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
234 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
235 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
236
237 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
238 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
239 w = (const float*) w + 32;
240
241 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
242 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
243 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
244 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
245 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
246 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
247 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
248
249 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
250 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
251 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
252 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
253 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
254 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
255 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
256
257 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
258 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
259 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
260 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
261 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
262 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
263 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
264
265 if XNN_LIKELY(nc >= 16) {
266 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
267 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
268 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
269 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
270 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
271 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
272 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
273
274 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
275 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
276 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
277 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
278 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
279 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
280 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
281
282 c0 = (float*) ((uintptr_t) c0 + cn_stride);
283 c1 = (float*) ((uintptr_t) c1 + cn_stride);
284 c2 = (float*) ((uintptr_t) c2 + cn_stride);
285 c3 = (float*) ((uintptr_t) c3 + cn_stride);
286 c4 = (float*) ((uintptr_t) c4 + cn_stride);
287 c5 = (float*) ((uintptr_t) c5 + cn_stride);
288 c6 = (float*) ((uintptr_t) c6 + cn_stride);
289
290 nc -= 16;
291 } else {
292 // Prepare mask for valid 32-bit elements (depends on nc).
293 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
294 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
295 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
296 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
297 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
298 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
299 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
300 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
301 nc = 0;
302 }
303 } while (nc != 0);
304 }