Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-8x16c4-minmax-avx512vnni.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx16c4-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_qc4w_gemm_minmax_ukernel_8x16c4__avx512vnni(
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_qc4w_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 <= 8);
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, 4 * 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
89 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
90 XNN_FORCE_REALIZATION(vsign_mask);
91 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point + 128);
92 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point + 128);
93 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point + 128);
94 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point + 128);
95 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point + 128);
96 const __m512i vinput_zero_point5 = _mm512_set1_epi32((int) quantization_params[5].zero_point + 128);
97 const __m512i vinput_zero_point6 = _mm512_set1_epi32((int) quantization_params[6].zero_point + 128);
98 const __m512i vinput_zero_point7 = _mm512_set1_epi32((int) quantization_params[7].zero_point + 128);
99 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
100 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
101 // XNN_FORCE_REALIZATION(voutput_min);
102 // XNN_FORCE_REALIZATION(voutput_max);
103 const __m512i vmask = _mm512_set1_epi8(0xF0);
104 XNN_FORCE_REALIZATION(vmask);
105 do {
106 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
107 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point0);
108 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
109 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point1);
110 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
111 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point2);
112 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
113 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point3);
114 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
115 __m512i vacc4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point4);
116 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
117 __m512i vacc5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point5);
118 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
119 __m512i vacc6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point6);
120 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
121 __m512i vacc7x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point7);
122 __m512i vacc1x7x0123456789ABCDEF = _mm512_setzero_epi32();
123 w = (const int32_t*) w + 16;
124
125 size_t k = kc;
126 while (k >= 8 * sizeof(int8_t)) {
127 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
128 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
129 a0 += 8;
130 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
131 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
132 a1 += 8;
133 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
134 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
135 a2 += 8;
136 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
137 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
138 a3 += 8;
139 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
140 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
141 a4 += 8;
142 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
143 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
144 a5 += 8;
145 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
146 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
147 a6 += 8;
148 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
149 const __m512i va7x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7 + 4)), vsign_mask);
150 a7 += 8;
151
152 const __m512i vbb0123456789ABCDEFx01234567 = _mm512_load_si512(w);
153 const __m512i vbs0123456789ABCDEFx0123 = _mm512_slli_epi32(vbb0123456789ABCDEFx01234567, 4);
154 const __m512i vb0123456789ABCDEFx4567 = _mm512_and_si512(vbb0123456789ABCDEFx01234567, vmask);
155 const __m512i vb0123456789ABCDEFx0123 = _mm512_and_si512(vbs0123456789ABCDEFx0123, vmask);
156
157 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
158 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
159 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
160 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
161 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
162 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
163 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
164 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEFx0123);
165 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
166 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
167 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
168 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
169 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
170 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
171 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
172 vacc1x7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x7x0123456789ABCDEF, va7x4567, vb0123456789ABCDEFx4567);
173
174 w = (const int8_t*) w + 64;
175 k -= 8 * sizeof(int8_t);
176 }
177 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
178 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
179 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
180 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
181 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
182 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
183 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
184 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, vacc1x7x0123456789ABCDEF);
185
186 if (k != 0) {
187 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
188 a0 += 4;
189 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
190 a1 += 4;
191 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
192 a2 += 4;
193 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
194 a3 += 4;
195 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
196 a4 += 4;
197 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
198 a5 += 4;
199 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
200 a6 += 4;
201 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
202 a7 += 4;
203
204 const __m512i vbb0123456789ABCDEF = _mm512_load_si512(w);
205 const __m512i vb0123456789ABCDEF = _mm512_slli_epi32(vbb0123456789ABCDEF, 4);
206
207 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
208 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
209 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
210 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
211 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
212 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
213 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
214 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEF);
215
216 w = (const int8_t*) w + 64;
217 k -= 4 * sizeof(int8_t);
218 }
219
220 vacc0x0123456789ABCDEF = _mm512_srai_epi32(vacc0x0123456789ABCDEF, 4);
221 vacc1x0123456789ABCDEF = _mm512_srai_epi32(vacc1x0123456789ABCDEF, 4);
222 vacc2x0123456789ABCDEF = _mm512_srai_epi32(vacc2x0123456789ABCDEF, 4);
223 vacc3x0123456789ABCDEF = _mm512_srai_epi32(vacc3x0123456789ABCDEF, 4);
224 vacc4x0123456789ABCDEF = _mm512_srai_epi32(vacc4x0123456789ABCDEF, 4);
225 vacc5x0123456789ABCDEF = _mm512_srai_epi32(vacc5x0123456789ABCDEF, 4);
226 vacc6x0123456789ABCDEF = _mm512_srai_epi32(vacc6x0123456789ABCDEF, 4);
227 vacc7x0123456789ABCDEF = _mm512_srai_epi32(vacc7x0123456789ABCDEF, 4);
228 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
229 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
230 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
231 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
232 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
233 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
234 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
235 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
236
237 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
238 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
239 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
240 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
241 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
242 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
243 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
244 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
245
246 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
247 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
248 w = (const float*) w + 32;
249
250 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
251 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
252 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
253 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
254 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
255 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
256 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
257 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
258
259 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
260 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
261 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
262 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
263 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
264 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
265 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
266 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
267
268 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
269 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
270 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
271 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
272 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
273 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
274 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
275 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
276
277 if XNN_LIKELY(nc >= 16) {
278 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
279 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
280 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
281 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
282 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
283 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
284 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
285 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
286
287 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
288 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
289 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
290 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
291 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
292 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
293 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
294 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
295
296 c0 = (float*) ((uintptr_t) c0 + cn_stride);
297 c1 = (float*) ((uintptr_t) c1 + cn_stride);
298 c2 = (float*) ((uintptr_t) c2 + cn_stride);
299 c3 = (float*) ((uintptr_t) c3 + cn_stride);
300 c4 = (float*) ((uintptr_t) c4 + cn_stride);
301 c5 = (float*) ((uintptr_t) c5 + cn_stride);
302 c6 = (float*) ((uintptr_t) c6 + cn_stride);
303 c7 = (float*) ((uintptr_t) c7 + cn_stride);
304
305 nc -= 16;
306 } else {
307 // Prepare mask for valid 32-bit elements (depends on nc).
308 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
309 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
310 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
311 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
312 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
313 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
314 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
315 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
316 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
317 nc = 0;
318 }
319 } while (nc != 0);
320 }