Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-7x16c8-minmax-avx512skx-madd-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx16c8-avx512vnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/unaligned.h"
19 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_7x16c8__avx512skx_madd_prfm(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 const int8_t* restrict a,
27 size_t a_stride,
28 const void* restrict w,
29 float* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
33 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 7);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 const int8_t* a0 = a;
46 float* c0 = c;
47 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
48 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 a1 = a0;
51 c1 = c0;
52 }
53 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
54 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
55 if XNN_UNPREDICTABLE(mr <= 2) {
56 a2 = a1;
57 c2 = c1;
58 }
59 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
60 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
61 if XNN_UNPREDICTABLE(mr < 4) {
62 a3 = a2;
63 c3 = c2;
64 }
65 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
66 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
67 if XNN_UNPREDICTABLE(mr <= 4) {
68 a4 = a3;
69 c4 = c3;
70 }
71 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
72 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
73 if XNN_UNPREDICTABLE(mr < 6) {
74 a5 = a4;
75 c5 = c4;
76 }
77 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
78 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
79 if XNN_UNPREDICTABLE(mr <= 6) {
80 a6 = a5;
81 c6 = c5;
82 }
83
84 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
85 XNN_FORCE_REALIZATION(vsign_mask);
86 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point + 128);
87 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point + 128);
88 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point + 128);
89 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point + 128);
90 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point + 128);
91 const __m512i vinput_zero_point5 = _mm512_set1_epi32((int) quantization_params[5].zero_point + 128);
92 const __m512i vinput_zero_point6 = _mm512_set1_epi32((int) quantization_params[6].zero_point + 128);
93 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
94 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
95 // XNN_FORCE_REALIZATION(voutput_min);
96 // XNN_FORCE_REALIZATION(voutput_max);
97 const __m512i vmask = _mm512_set1_epi8(0x0F);
98 XNN_FORCE_REALIZATION(vmask);
99 do {
100 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
101 __m512i vsum0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point0);
102 __m512i vacc0x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 0));
103 __m512i vacc0x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 1));
104 __m512i vsum1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point1);
105 __m512i vacc1x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum1x0123456789ABCDEF, 0));
106 __m512i vacc1x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum1x0123456789ABCDEF, 1));
107 __m512i vsum2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point2);
108 __m512i vacc2x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum2x0123456789ABCDEF, 0));
109 __m512i vacc2x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum2x0123456789ABCDEF, 1));
110 __m512i vsum3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point3);
111 __m512i vacc3x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum3x0123456789ABCDEF, 0));
112 __m512i vacc3x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum3x0123456789ABCDEF, 1));
113 __m512i vsum4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point4);
114 __m512i vacc4x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum4x0123456789ABCDEF, 0));
115 __m512i vacc4x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum4x0123456789ABCDEF, 1));
116 __m512i vsum5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point5);
117 __m512i vacc5x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum5x0123456789ABCDEF, 0));
118 __m512i vacc5x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum5x0123456789ABCDEF, 1));
119 __m512i vsum6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point6);
120 __m512i vacc6x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum6x0123456789ABCDEF, 0));
121 __m512i vacc6x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum6x0123456789ABCDEF, 1));
122 w = (const int32_t*) w + 16;
123
124 size_t k = kc;
125 while (k >= 16 * sizeof(int8_t)) {
126 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
127 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
128 a0 += 16;
129 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
130 const __m512i va1x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
131 a1 += 16;
132 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
133 const __m512i va2x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
134 a2 += 16;
135 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
136 const __m512i va3x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
137 a3 += 16;
138 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
139 const __m512i va4x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
140 a4 += 16;
141 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
142 const __m512i va5x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
143 a5 += 16;
144 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
145 const __m512i va6x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
146 a6 += 16;
147
148 const __m512i vbb01234567x01234567 = _mm512_load_si512(w);
149 const __m512i vbb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
150 const __m512i vbs01234567x89ABCDEF = _mm512_srli_epi32(vbb01234567x01234567, 4);
151 const __m512i vbs89ABCDEFx89ABCDEF = _mm512_srli_epi32(vbb89ABCDEFx01234567, 4);
152 const __m512i vb01234567x01234567 = _mm512_and_si512(vbb01234567x01234567, vmask);
153 const __m512i vb89ABCDEFx01234567 = _mm512_and_si512(vbb89ABCDEFx01234567, vmask);
154 const __m512i vb01234567x89ABCDEF = _mm512_and_si512(vbs01234567x89ABCDEF, vmask);
155 const __m512i vb89ABCDEFx89ABCDEF = _mm512_and_si512(vbs89ABCDEFx89ABCDEF, vmask);
156
157 vacc0x01234567 = _mm512_dpbusd_epi32_madd(vacc0x01234567, va0x01234567, vb01234567x01234567);
158 vacc0x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
159 vacc1x01234567 = _mm512_dpbusd_epi32_madd(vacc1x01234567, va1x01234567, vb01234567x01234567);
160 vacc1x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
161 vacc2x01234567 = _mm512_dpbusd_epi32_madd(vacc2x01234567, va2x01234567, vb01234567x01234567);
162 vacc2x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
163 vacc3x01234567 = _mm512_dpbusd_epi32_madd(vacc3x01234567, va3x01234567, vb01234567x01234567);
164 vacc3x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
165 vacc4x01234567 = _mm512_dpbusd_epi32_madd(vacc4x01234567, va4x01234567, vb01234567x01234567);
166 vacc4x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
167 vacc5x01234567 = _mm512_dpbusd_epi32_madd(vacc5x01234567, va5x01234567, vb01234567x01234567);
168 vacc5x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
169 vacc6x01234567 = _mm512_dpbusd_epi32_madd(vacc6x01234567, va6x01234567, vb01234567x01234567);
170 vacc6x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
171 xnn_prefetch_to_l1((const int8_t*) w + 896);
172 xnn_prefetch_to_l1((const int8_t*) w + 960);
173 vacc0x01234567 = _mm512_dpbusd_epi32_madd(vacc0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
174 vacc0x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
175 vacc1x01234567 = _mm512_dpbusd_epi32_madd(vacc1x01234567, va1x89ABCDEF, vb01234567x89ABCDEF);
176 vacc1x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x89ABCDEF, va1x89ABCDEF, vb89ABCDEFx89ABCDEF);
177 vacc2x01234567 = _mm512_dpbusd_epi32_madd(vacc2x01234567, va2x89ABCDEF, vb01234567x89ABCDEF);
178 vacc2x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc2x89ABCDEF, va2x89ABCDEF, vb89ABCDEFx89ABCDEF);
179 vacc3x01234567 = _mm512_dpbusd_epi32_madd(vacc3x01234567, va3x89ABCDEF, vb01234567x89ABCDEF);
180 vacc3x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc3x89ABCDEF, va3x89ABCDEF, vb89ABCDEFx89ABCDEF);
181 vacc4x01234567 = _mm512_dpbusd_epi32_madd(vacc4x01234567, va4x89ABCDEF, vb01234567x89ABCDEF);
182 vacc4x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc4x89ABCDEF, va4x89ABCDEF, vb89ABCDEFx89ABCDEF);
183 vacc5x01234567 = _mm512_dpbusd_epi32_madd(vacc5x01234567, va5x89ABCDEF, vb01234567x89ABCDEF);
184 vacc5x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc5x89ABCDEF, va5x89ABCDEF, vb89ABCDEFx89ABCDEF);
185 vacc6x01234567 = _mm512_dpbusd_epi32_madd(vacc6x01234567, va6x89ABCDEF, vb01234567x89ABCDEF);
186 vacc6x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc6x89ABCDEF, va6x89ABCDEF, vb89ABCDEFx89ABCDEF);
187
188 w = (const int8_t*) w + 128;
189 k -= 16 * sizeof(int8_t);
190 }
191
192 if (k != 0) {
193 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
194 a0 += 8;
195 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
196 a1 += 8;
197 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
198 a2 += 8;
199 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
200 a3 += 8;
201 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
202 a4 += 8;
203 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
204 a5 += 8;
205 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
206 a6 += 8;
207
208 const __m512i vbb01234567x01234567 = _mm512_load_si512(w);
209 const __m512i vbb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
210 const __m512i vb01234567x01234567 = _mm512_and_si512(vbb01234567x01234567, vmask);
211 const __m512i vb89ABCDEFx01234567 = _mm512_and_si512(vbb89ABCDEFx01234567, vmask);
212
213 vacc0x01234567 = _mm512_dpbusd_epi32_madd(vacc0x01234567, va0x01234567, vb01234567x01234567);
214 vacc0x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
215 vacc1x01234567 = _mm512_dpbusd_epi32_madd(vacc1x01234567, va1x01234567, vb01234567x01234567);
216 vacc1x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
217 vacc2x01234567 = _mm512_dpbusd_epi32_madd(vacc2x01234567, va2x01234567, vb01234567x01234567);
218 vacc2x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
219 vacc3x01234567 = _mm512_dpbusd_epi32_madd(vacc3x01234567, va3x01234567, vb01234567x01234567);
220 vacc3x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
221 vacc4x01234567 = _mm512_dpbusd_epi32_madd(vacc4x01234567, va4x01234567, vb01234567x01234567);
222 vacc4x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
223 vacc5x01234567 = _mm512_dpbusd_epi32_madd(vacc5x01234567, va5x01234567, vb01234567x01234567);
224 vacc5x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
225 vacc6x01234567 = _mm512_dpbusd_epi32_madd(vacc6x01234567, va6x01234567, vb01234567x01234567);
226 vacc6x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
227 xnn_prefetch_to_l1((const int8_t*) w + 896);
228 xnn_prefetch_to_l1((const int8_t*) w + 960);
229
230 w = (const int8_t*) w + 128;
231 k -= 8 * sizeof(int8_t);
232 }
233
234 // Add adjacent pairs
235 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
236 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
237 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
238 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
239 const __m512i vsum1x01234567 = _mm512_add_epi32(vacc1x01234567, _mm512_srli_epi64(vacc1x01234567, 32));
240 const __m512i vsum1x89ABCDEF = _mm512_add_epi32(vacc1x89ABCDEF, _mm512_srli_epi64(vacc1x89ABCDEF, 32));
241 __m512i vacc1x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum1x01234567, vidx, vsum1x89ABCDEF);
242 const __m512i vsum2x01234567 = _mm512_add_epi32(vacc2x01234567, _mm512_srli_epi64(vacc2x01234567, 32));
243 const __m512i vsum2x89ABCDEF = _mm512_add_epi32(vacc2x89ABCDEF, _mm512_srli_epi64(vacc2x89ABCDEF, 32));
244 __m512i vacc2x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum2x01234567, vidx, vsum2x89ABCDEF);
245 const __m512i vsum3x01234567 = _mm512_add_epi32(vacc3x01234567, _mm512_srli_epi64(vacc3x01234567, 32));
246 const __m512i vsum3x89ABCDEF = _mm512_add_epi32(vacc3x89ABCDEF, _mm512_srli_epi64(vacc3x89ABCDEF, 32));
247 __m512i vacc3x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum3x01234567, vidx, vsum3x89ABCDEF);
248 const __m512i vsum4x01234567 = _mm512_add_epi32(vacc4x01234567, _mm512_srli_epi64(vacc4x01234567, 32));
249 const __m512i vsum4x89ABCDEF = _mm512_add_epi32(vacc4x89ABCDEF, _mm512_srli_epi64(vacc4x89ABCDEF, 32));
250 __m512i vacc4x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum4x01234567, vidx, vsum4x89ABCDEF);
251 const __m512i vsum5x01234567 = _mm512_add_epi32(vacc5x01234567, _mm512_srli_epi64(vacc5x01234567, 32));
252 const __m512i vsum5x89ABCDEF = _mm512_add_epi32(vacc5x89ABCDEF, _mm512_srli_epi64(vacc5x89ABCDEF, 32));
253 __m512i vacc5x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum5x01234567, vidx, vsum5x89ABCDEF);
254 const __m512i vsum6x01234567 = _mm512_add_epi32(vacc6x01234567, _mm512_srli_epi64(vacc6x01234567, 32));
255 const __m512i vsum6x89ABCDEF = _mm512_add_epi32(vacc6x89ABCDEF, _mm512_srli_epi64(vacc6x89ABCDEF, 32));
256 __m512i vacc6x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum6x01234567, vidx, vsum6x89ABCDEF);
257
258 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
259 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
260 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
261 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
262 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
263 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
264 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
265
266 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
267 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
268 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
269 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
270 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
271 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
272 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
273
274 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
275 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
276 w = (const float*) w + 32;
277
278 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
279 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
280 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
281 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
282 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
283 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
284 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
285
286 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
287 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
288 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
289 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
290 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
291 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
292 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
293
294 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
295 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
296 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
297 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
298 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
299 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
300 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
301
302 if XNN_LIKELY(nc >= 16) {
303 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
304 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
305 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
306 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
307 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
308 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
309 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
310
311 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
312 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
313 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
314 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
315 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
316 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
317 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
318
319 c0 = (float*) ((uintptr_t) c0 + cn_stride);
320 c1 = (float*) ((uintptr_t) c1 + cn_stride);
321 c2 = (float*) ((uintptr_t) c2 + cn_stride);
322 c3 = (float*) ((uintptr_t) c3 + cn_stride);
323 c4 = (float*) ((uintptr_t) c4 + cn_stride);
324 c5 = (float*) ((uintptr_t) c5 + cn_stride);
325 c6 = (float*) ((uintptr_t) c6 + cn_stride);
326
327 nc -= 16;
328 } else {
329 // Prepare mask for valid 32-bit elements (depends on nc).
330 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
331 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
332 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
333 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
334 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
335 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
336 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
337 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
338 nc = 0;
339 }
340 } while (nc != 0);
341 }