Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-12x16c4-minmax-avx512vnnigfni.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_12x16c4__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_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 <= 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, 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 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
113 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
114 XNN_FORCE_REALIZATION(vsign_mask);
115 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point + 128);
116 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point + 128);
117 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point + 128);
118 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point + 128);
119 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point + 128);
120 const __m512i vinput_zero_point5 = _mm512_set1_epi32((int) quantization_params[5].zero_point + 128);
121 const __m512i vinput_zero_point6 = _mm512_set1_epi32((int) quantization_params[6].zero_point + 128);
122 const __m512i vinput_zero_point7 = _mm512_set1_epi32((int) quantization_params[7].zero_point + 128);
123 const __m512i vinput_zero_point8 = _mm512_set1_epi32((int) quantization_params[8].zero_point + 128);
124 const __m512i vinput_zero_point9 = _mm512_set1_epi32((int) quantization_params[9].zero_point + 128);
125 const __m512i vinput_zero_point10 = _mm512_set1_epi32((int) quantization_params[10].zero_point + 128);
126 const __m512i vinput_zero_point11 = _mm512_set1_epi32((int) quantization_params[11].zero_point + 128);
127 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
128 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
129 // XNN_FORCE_REALIZATION(voutput_min);
130 // XNN_FORCE_REALIZATION(voutput_max);
131 const __m512i vmask = _mm512_set1_epi8(0xF0);
132 XNN_FORCE_REALIZATION(vmask);
133 const __m512i vshl4 = _mm512_set1_epi64(0x01020408);
134 XNN_FORCE_REALIZATION(vshl4);
135 do {
136 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
137 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point0);
138 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
139 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point1);
140 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
141 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point2);
142 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
143 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point3);
144 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
145 __m512i vacc4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point4);
146 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
147 __m512i vacc5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point5);
148 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
149 __m512i vacc6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point6);
150 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
151 __m512i vacc7x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point7);
152 __m512i vacc1x7x0123456789ABCDEF = _mm512_setzero_epi32();
153 __m512i vacc8x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point8);
154 __m512i vacc1x8x0123456789ABCDEF = _mm512_setzero_epi32();
155 __m512i vacc9x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point9);
156 __m512i vacc1x9x0123456789ABCDEF = _mm512_setzero_epi32();
157 __m512i vacc10x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point10);
158 __m512i vacc1x10x0123456789ABCDEF = _mm512_setzero_epi32();
159 __m512i vacc11x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point11);
160 __m512i vacc1x11x0123456789ABCDEF = _mm512_setzero_epi32();
161 w = (const int32_t*) w + 16;
162
163 size_t k = kc;
164 while (k >= 8 * sizeof(int8_t)) {
165 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
166 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
167 a0 += 8;
168 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
169 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
170 a1 += 8;
171 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
172 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
173 a2 += 8;
174 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
175 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
176 a3 += 8;
177 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
178 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
179 a4 += 8;
180 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
181 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
182 a5 += 8;
183 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
184 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
185 a6 += 8;
186 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
187 const __m512i va7x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7 + 4)), vsign_mask);
188 a7 += 8;
189 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
190 const __m512i va8x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8 + 4)), vsign_mask);
191 a8 += 8;
192 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
193 const __m512i va9x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9 + 4)), vsign_mask);
194 a9 += 8;
195 const __m512i va10x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10)), vsign_mask);
196 const __m512i va10x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10 + 4)), vsign_mask);
197 a10 += 8;
198 const __m512i va11x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11)), vsign_mask);
199 const __m512i va11x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11 + 4)), vsign_mask);
200 a11 += 8;
201
202 const __m512i vbb0123456789ABCDEFx01234567 = _mm512_load_si512(w);
203 const __m512i vb0123456789ABCDEFx0123 = _mm512_gf2p8affine_epi64_epi8(vbb0123456789ABCDEFx01234567, vshl4, 0);
204 const __m512i vb0123456789ABCDEFx4567 = _mm512_and_si512(vbb0123456789ABCDEFx01234567, vmask);
205
206 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
207 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
208 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
209 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
210 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
211 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
212 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
213 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEFx0123);
214 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEFx0123);
215 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEFx0123);
216 vacc10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc10x0123456789ABCDEF, va10x0123, vb0123456789ABCDEFx0123);
217 vacc11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc11x0123456789ABCDEF, va11x0123, vb0123456789ABCDEFx0123);
218 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
219 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
220 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
221 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
222 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
223 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
224 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
225 vacc1x7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x7x0123456789ABCDEF, va7x4567, vb0123456789ABCDEFx4567);
226 vacc1x8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x8x0123456789ABCDEF, va8x4567, vb0123456789ABCDEFx4567);
227 vacc1x9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x9x0123456789ABCDEF, va9x4567, vb0123456789ABCDEFx4567);
228 vacc1x10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x10x0123456789ABCDEF, va10x4567, vb0123456789ABCDEFx4567);
229 vacc1x11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x11x0123456789ABCDEF, va11x4567, vb0123456789ABCDEFx4567);
230
231 w = (const int8_t*) w + 64;
232 k -= 8 * sizeof(int8_t);
233 }
234 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
235 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
236 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
237 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
238 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
239 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
240 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
241 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, vacc1x7x0123456789ABCDEF);
242 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, vacc1x8x0123456789ABCDEF);
243 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, vacc1x9x0123456789ABCDEF);
244 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, vacc1x10x0123456789ABCDEF);
245 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, vacc1x11x0123456789ABCDEF);
246
247 if (k != 0) {
248 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
249 a0 += 4;
250 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
251 a1 += 4;
252 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
253 a2 += 4;
254 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
255 a3 += 4;
256 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
257 a4 += 4;
258 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
259 a5 += 4;
260 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
261 a6 += 4;
262 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
263 a7 += 4;
264 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
265 a8 += 4;
266 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
267 a9 += 4;
268 const __m512i va10x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10)), vsign_mask);
269 a10 += 4;
270 const __m512i va11x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11)), vsign_mask);
271 a11 += 4;
272
273 const __m512i vbb0123456789ABCDEF = _mm512_load_si512(w);
274 const __m512i vb0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vbb0123456789ABCDEF, vshl4, 0);
275
276 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
277 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
278 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
279 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
280 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
281 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
282 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
283 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEF);
284 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEF);
285 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEF);
286 vacc10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc10x0123456789ABCDEF, va10x0123, vb0123456789ABCDEF);
287 vacc11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc11x0123456789ABCDEF, va11x0123, vb0123456789ABCDEF);
288
289 w = (const int8_t*) w + 64;
290 k -= 4 * sizeof(int8_t);
291 }
292
293 vacc0x0123456789ABCDEF = _mm512_srai_epi32(vacc0x0123456789ABCDEF, 4);
294 vacc1x0123456789ABCDEF = _mm512_srai_epi32(vacc1x0123456789ABCDEF, 4);
295 vacc2x0123456789ABCDEF = _mm512_srai_epi32(vacc2x0123456789ABCDEF, 4);
296 vacc3x0123456789ABCDEF = _mm512_srai_epi32(vacc3x0123456789ABCDEF, 4);
297 vacc4x0123456789ABCDEF = _mm512_srai_epi32(vacc4x0123456789ABCDEF, 4);
298 vacc5x0123456789ABCDEF = _mm512_srai_epi32(vacc5x0123456789ABCDEF, 4);
299 vacc6x0123456789ABCDEF = _mm512_srai_epi32(vacc6x0123456789ABCDEF, 4);
300 vacc7x0123456789ABCDEF = _mm512_srai_epi32(vacc7x0123456789ABCDEF, 4);
301 vacc8x0123456789ABCDEF = _mm512_srai_epi32(vacc8x0123456789ABCDEF, 4);
302 vacc9x0123456789ABCDEF = _mm512_srai_epi32(vacc9x0123456789ABCDEF, 4);
303 vacc10x0123456789ABCDEF = _mm512_srai_epi32(vacc10x0123456789ABCDEF, 4);
304 vacc11x0123456789ABCDEF = _mm512_srai_epi32(vacc11x0123456789ABCDEF, 4);
305 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
306 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
307 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
308 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
309 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
310 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
311 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
312 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
313 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
314 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
315 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
316 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
317
318 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
319 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
320 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
321 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
322 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
323 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
324 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
325 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
326 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, _mm512_set1_ps(quantization_params[8].inv_scale));
327 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, _mm512_set1_ps(quantization_params[9].inv_scale));
328 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, _mm512_set1_ps(quantization_params[10].inv_scale));
329 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, _mm512_set1_ps(quantization_params[11].inv_scale));
330
331 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
332 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
333 w = (const float*) w + 32;
334
335 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
336 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
337 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
338 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
339 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
340 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
341 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
342 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
343 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
344 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vscaled9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
345 vscaled10x0123456789ABCDEF = _mm512_fmadd_ps(vscaled10x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
346 vscaled11x0123456789ABCDEF = _mm512_fmadd_ps(vscaled11x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
347
348 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
349 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
350 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
351 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
352 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
353 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
354 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
355 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
356 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
357 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
358 vscaled10x0123456789ABCDEF = _mm512_max_ps(vscaled10x0123456789ABCDEF, voutput_min);
359 vscaled11x0123456789ABCDEF = _mm512_max_ps(vscaled11x0123456789ABCDEF, voutput_min);
360
361 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
362 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
363 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
364 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
365 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
366 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
367 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
368 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
369 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
370 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
371 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max);
372 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max);
373
374 if XNN_LIKELY(nc >= 16) {
375 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
376 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
377 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
378 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
379 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
380 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
381 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
382 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
383 _mm512_storeu_ps(c8, vscaled8x0123456789ABCDEF);
384 _mm512_storeu_ps(c9, vscaled9x0123456789ABCDEF);
385 _mm512_storeu_ps(c10, vscaled10x0123456789ABCDEF);
386 _mm512_storeu_ps(c11, vscaled11x0123456789ABCDEF);
387
388 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
389 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
390 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
391 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
392 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
393 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
394 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
395 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
396 a8 = (const int8_t*) ((uintptr_t) a8 - kc);
397 a9 = (const int8_t*) ((uintptr_t) a9 - kc);
398 a10 = (const int8_t*) ((uintptr_t) a10 - kc);
399 a11 = (const int8_t*) ((uintptr_t) a11 - kc);
400
401 c0 = (float*) ((uintptr_t) c0 + cn_stride);
402 c1 = (float*) ((uintptr_t) c1 + cn_stride);
403 c2 = (float*) ((uintptr_t) c2 + cn_stride);
404 c3 = (float*) ((uintptr_t) c3 + cn_stride);
405 c4 = (float*) ((uintptr_t) c4 + cn_stride);
406 c5 = (float*) ((uintptr_t) c5 + cn_stride);
407 c6 = (float*) ((uintptr_t) c6 + cn_stride);
408 c7 = (float*) ((uintptr_t) c7 + cn_stride);
409 c8 = (float*) ((uintptr_t) c8 + cn_stride);
410 c9 = (float*) ((uintptr_t) c9 + cn_stride);
411 c10 = (float*) ((uintptr_t) c10 + cn_stride);
412 c11 = (float*) ((uintptr_t) c11 + cn_stride);
413
414 nc -= 16;
415 } else {
416 // Prepare mask for valid 32-bit elements (depends on nc).
417 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
418 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
419 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
420 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
421 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
422 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
423 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
424 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
425 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
426 _mm512_mask_storeu_ps(c8, vmask, vscaled8x0123456789ABCDEF);
427 _mm512_mask_storeu_ps(c9, vmask, vscaled9x0123456789ABCDEF);
428 _mm512_mask_storeu_ps(c10, vmask, vscaled10x0123456789ABCDEF);
429 _mm512_mask_storeu_ps(c11, vmask, vscaled11x0123456789ABCDEF);
430 nc = 0;
431 }
432 } while (nc != 0);
433 }