Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-14x16c4-minmax-avx512vnni-prfm.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 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_14x16c4__avx512vnni_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 union xnn_f32_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 <= 14);
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, 4 * 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 const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
84 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
85 if XNN_UNPREDICTABLE(mr < 8) {
86 a7 = a6;
87 c7 = c6;
88 }
89 const int8_t* a8 = (const int8_t*) ((uintptr_t) a7 + a_stride);
90 float* c8 = (float*) ((uintptr_t) c7 + cm_stride);
91 if XNN_UNPREDICTABLE(mr <= 8) {
92 a8 = a7;
93 c8 = c7;
94 }
95 const int8_t* a9 = (const int8_t*) ((uintptr_t) a8 + a_stride);
96 float* c9 = (float*) ((uintptr_t) c8 + cm_stride);
97 if XNN_UNPREDICTABLE(mr < 10) {
98 a9 = a8;
99 c9 = c8;
100 }
101 const int8_t* a10 = (const int8_t*) ((uintptr_t) a9 + a_stride);
102 float* c10 = (float*) ((uintptr_t) c9 + cm_stride);
103 if XNN_UNPREDICTABLE(mr <= 10) {
104 a10 = a9;
105 c10 = c9;
106 }
107 const int8_t* a11 = (const int8_t*) ((uintptr_t) a10 + a_stride);
108 float* c11 = (float*) ((uintptr_t) c10 + cm_stride);
109 if XNN_UNPREDICTABLE(mr < 12) {
110 a11 = a10;
111 c11 = c10;
112 }
113 const int8_t* a12 = (const int8_t*) ((uintptr_t) a11 + a_stride);
114 float* c12 = (float*) ((uintptr_t) c11 + cm_stride);
115 if XNN_UNPREDICTABLE(mr <= 12) {
116 a12 = a11;
117 c12 = c11;
118 }
119 const int8_t* a13 = (const int8_t*) ((uintptr_t) a12 + a_stride);
120 float* c13 = (float*) ((uintptr_t) c12 + cm_stride);
121 if XNN_UNPREDICTABLE(mr != 14) {
122 a13 = a12;
123 c13 = c12;
124 }
125
126 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
127 XNN_FORCE_REALIZATION(vsign_mask);
128 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point + 128);
129 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point + 128);
130 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point + 128);
131 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point + 128);
132 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point + 128);
133 const __m512i vinput_zero_point5 = _mm512_set1_epi32((int) quantization_params[5].zero_point + 128);
134 const __m512i vinput_zero_point6 = _mm512_set1_epi32((int) quantization_params[6].zero_point + 128);
135 const __m512i vinput_zero_point7 = _mm512_set1_epi32((int) quantization_params[7].zero_point + 128);
136 const __m512i vinput_zero_point8 = _mm512_set1_epi32((int) quantization_params[8].zero_point + 128);
137 const __m512i vinput_zero_point9 = _mm512_set1_epi32((int) quantization_params[9].zero_point + 128);
138 const __m512i vinput_zero_point10 = _mm512_set1_epi32((int) quantization_params[10].zero_point + 128);
139 const __m512i vinput_zero_point11 = _mm512_set1_epi32((int) quantization_params[11].zero_point + 128);
140 const __m512i vinput_zero_point12 = _mm512_set1_epi32((int) quantization_params[12].zero_point + 128);
141 const __m512i vinput_zero_point13 = _mm512_set1_epi32((int) quantization_params[13].zero_point + 128);
142 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
143 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
144 // XNN_FORCE_REALIZATION(voutput_min);
145 // XNN_FORCE_REALIZATION(voutput_max);
146 do {
147 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
148 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point0);
149 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
150 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point1);
151 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
152 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point2);
153 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
154 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point3);
155 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
156 __m512i vacc4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point4);
157 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
158 __m512i vacc5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point5);
159 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
160 __m512i vacc6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point6);
161 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
162 __m512i vacc7x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point7);
163 __m512i vacc1x7x0123456789ABCDEF = _mm512_setzero_epi32();
164 __m512i vacc8x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point8);
165 __m512i vacc1x8x0123456789ABCDEF = _mm512_setzero_epi32();
166 __m512i vacc9x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point9);
167 __m512i vacc1x9x0123456789ABCDEF = _mm512_setzero_epi32();
168 __m512i vacc10x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point10);
169 __m512i vacc1x10x0123456789ABCDEF = _mm512_setzero_epi32();
170 __m512i vacc11x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point11);
171 __m512i vacc1x11x0123456789ABCDEF = _mm512_setzero_epi32();
172 __m512i vacc12x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point12);
173 __m512i vacc1x12x0123456789ABCDEF = _mm512_setzero_epi32();
174 __m512i vacc13x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point13);
175 __m512i vacc1x13x0123456789ABCDEF = _mm512_setzero_epi32();
176 w = (const int32_t*) w + 16;
177
178 size_t k = kc;
179 while (k >= 8 * sizeof(int8_t)) {
180 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
181 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
182 a0 += 8;
183 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
184 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
185 a1 += 8;
186 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
187 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
188 a2 += 8;
189 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
190 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
191 a3 += 8;
192 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
193 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
194 a4 += 8;
195 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
196 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
197 a5 += 8;
198 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
199 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
200 a6 += 8;
201 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
202 const __m512i va7x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7 + 4)), vsign_mask);
203 a7 += 8;
204 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
205 const __m512i va8x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8 + 4)), vsign_mask);
206 a8 += 8;
207 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
208 const __m512i va9x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9 + 4)), vsign_mask);
209 a9 += 8;
210 const __m512i va10x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10)), vsign_mask);
211 const __m512i va10x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10 + 4)), vsign_mask);
212 a10 += 8;
213 const __m512i va11x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11)), vsign_mask);
214 const __m512i va11x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11 + 4)), vsign_mask);
215 a11 += 8;
216 const __m512i va12x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a12)), vsign_mask);
217 const __m512i va12x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a12 + 4)), vsign_mask);
218 a12 += 8;
219 const __m512i va13x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a13)), vsign_mask);
220 const __m512i va13x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a13 + 4)), vsign_mask);
221 a13 += 8;
222
223 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
224 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
225 xnn_prefetch_to_l1((const int8_t*) w + 896);
226
227 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
228 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
229 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
230 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
231 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
232 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
233 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
234 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEFx0123);
235 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEFx0123);
236 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEFx0123);
237 vacc10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc10x0123456789ABCDEF, va10x0123, vb0123456789ABCDEFx0123);
238 vacc11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc11x0123456789ABCDEF, va11x0123, vb0123456789ABCDEFx0123);
239 vacc12x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc12x0123456789ABCDEF, va12x0123, vb0123456789ABCDEFx0123);
240 vacc13x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc13x0123456789ABCDEF, va13x0123, vb0123456789ABCDEFx0123);
241 xnn_prefetch_to_l1((const int8_t*) w + 960);
242 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
243 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
244 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
245 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
246 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
247 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
248 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
249 vacc1x7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x7x0123456789ABCDEF, va7x4567, vb0123456789ABCDEFx4567);
250 vacc1x8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x8x0123456789ABCDEF, va8x4567, vb0123456789ABCDEFx4567);
251 vacc1x9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x9x0123456789ABCDEF, va9x4567, vb0123456789ABCDEFx4567);
252 vacc1x10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x10x0123456789ABCDEF, va10x4567, vb0123456789ABCDEFx4567);
253 vacc1x11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x11x0123456789ABCDEF, va11x4567, vb0123456789ABCDEFx4567);
254 vacc1x12x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x12x0123456789ABCDEF, va12x4567, vb0123456789ABCDEFx4567);
255 vacc1x13x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x13x0123456789ABCDEF, va13x4567, vb0123456789ABCDEFx4567);
256
257 w = (const int8_t*) w + 128;
258 k -= 8 * sizeof(int8_t);
259 }
260 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
261 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
262 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
263 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
264 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
265 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
266 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
267 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, vacc1x7x0123456789ABCDEF);
268 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, vacc1x8x0123456789ABCDEF);
269 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, vacc1x9x0123456789ABCDEF);
270 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, vacc1x10x0123456789ABCDEF);
271 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, vacc1x11x0123456789ABCDEF);
272 vacc12x0123456789ABCDEF = _mm512_add_epi32(vacc12x0123456789ABCDEF, vacc1x12x0123456789ABCDEF);
273 vacc13x0123456789ABCDEF = _mm512_add_epi32(vacc13x0123456789ABCDEF, vacc1x13x0123456789ABCDEF);
274
275 if (k != 0) {
276 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
277 a0 += 4;
278 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
279 a1 += 4;
280 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
281 a2 += 4;
282 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
283 a3 += 4;
284 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
285 a4 += 4;
286 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
287 a5 += 4;
288 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
289 a6 += 4;
290 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
291 a7 += 4;
292 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
293 a8 += 4;
294 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
295 a9 += 4;
296 const __m512i va10x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10)), vsign_mask);
297 a10 += 4;
298 const __m512i va11x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11)), vsign_mask);
299 a11 += 4;
300 const __m512i va12x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a12)), vsign_mask);
301 a12 += 4;
302 const __m512i va13x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a13)), vsign_mask);
303 a13 += 4;
304
305 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
306 xnn_prefetch_to_l1((const int8_t*) w + 960);
307
308 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
309 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
310 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
311 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
312 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
313 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
314 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
315 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEF);
316 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEF);
317 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEF);
318 vacc10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc10x0123456789ABCDEF, va10x0123, vb0123456789ABCDEF);
319 vacc11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc11x0123456789ABCDEF, va11x0123, vb0123456789ABCDEF);
320 vacc12x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc12x0123456789ABCDEF, va12x0123, vb0123456789ABCDEF);
321 vacc13x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc13x0123456789ABCDEF, va13x0123, vb0123456789ABCDEF);
322
323 w = (const int8_t*) w + 64;
324 k -= 4 * sizeof(int8_t);
325 }
326
327 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
328 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
329 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
330 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
331 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
332 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
333 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
334 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
335 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
336 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
337 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
338 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
339 __m512 vscaled12x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc12x0123456789ABCDEF);
340 __m512 vscaled13x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc13x0123456789ABCDEF);
341
342 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
343 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
344 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
345 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
346 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
347 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
348 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
349 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
350 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, _mm512_set1_ps(quantization_params[8].inv_scale));
351 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, _mm512_set1_ps(quantization_params[9].inv_scale));
352 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, _mm512_set1_ps(quantization_params[10].inv_scale));
353 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, _mm512_set1_ps(quantization_params[11].inv_scale));
354 vscaled12x0123456789ABCDEF = _mm512_mul_ps(vscaled12x0123456789ABCDEF, _mm512_set1_ps(quantization_params[12].inv_scale));
355 vscaled13x0123456789ABCDEF = _mm512_mul_ps(vscaled13x0123456789ABCDEF, _mm512_set1_ps(quantization_params[13].inv_scale));
356
357 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
358 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
359 w = (const float*) w + 32;
360
361 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
362 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
363 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
364 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
365 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
366 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
367 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
368 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
369 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
370 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vscaled9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
371 vscaled10x0123456789ABCDEF = _mm512_fmadd_ps(vscaled10x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
372 vscaled11x0123456789ABCDEF = _mm512_fmadd_ps(vscaled11x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
373 vscaled12x0123456789ABCDEF = _mm512_fmadd_ps(vscaled12x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
374 vscaled13x0123456789ABCDEF = _mm512_fmadd_ps(vscaled13x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
375
376 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
377 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
378 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
379 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
380 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
381 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
382 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
383 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
384 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
385 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
386 vscaled10x0123456789ABCDEF = _mm512_max_ps(vscaled10x0123456789ABCDEF, voutput_min);
387 vscaled11x0123456789ABCDEF = _mm512_max_ps(vscaled11x0123456789ABCDEF, voutput_min);
388 vscaled12x0123456789ABCDEF = _mm512_max_ps(vscaled12x0123456789ABCDEF, voutput_min);
389 vscaled13x0123456789ABCDEF = _mm512_max_ps(vscaled13x0123456789ABCDEF, voutput_min);
390
391 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
392 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
393 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
394 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
395 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
396 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
397 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
398 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
399 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
400 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
401 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max);
402 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max);
403 vscaled12x0123456789ABCDEF = _mm512_min_ps(vscaled12x0123456789ABCDEF, voutput_max);
404 vscaled13x0123456789ABCDEF = _mm512_min_ps(vscaled13x0123456789ABCDEF, voutput_max);
405
406 if XNN_LIKELY(nc >= 16) {
407 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
408 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
409 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
410 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
411 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
412 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
413 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
414 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
415 _mm512_storeu_ps(c8, vscaled8x0123456789ABCDEF);
416 _mm512_storeu_ps(c9, vscaled9x0123456789ABCDEF);
417 _mm512_storeu_ps(c10, vscaled10x0123456789ABCDEF);
418 _mm512_storeu_ps(c11, vscaled11x0123456789ABCDEF);
419 _mm512_storeu_ps(c12, vscaled12x0123456789ABCDEF);
420 _mm512_storeu_ps(c13, vscaled13x0123456789ABCDEF);
421
422 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
423 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
424 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
425 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
426 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
427 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
428 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
429 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
430 a8 = (const int8_t*) ((uintptr_t) a8 - kc);
431 a9 = (const int8_t*) ((uintptr_t) a9 - kc);
432 a10 = (const int8_t*) ((uintptr_t) a10 - kc);
433 a11 = (const int8_t*) ((uintptr_t) a11 - kc);
434 a12 = (const int8_t*) ((uintptr_t) a12 - kc);
435 a13 = (const int8_t*) ((uintptr_t) a13 - kc);
436
437 c0 = (float*) ((uintptr_t) c0 + cn_stride);
438 c1 = (float*) ((uintptr_t) c1 + cn_stride);
439 c2 = (float*) ((uintptr_t) c2 + cn_stride);
440 c3 = (float*) ((uintptr_t) c3 + cn_stride);
441 c4 = (float*) ((uintptr_t) c4 + cn_stride);
442 c5 = (float*) ((uintptr_t) c5 + cn_stride);
443 c6 = (float*) ((uintptr_t) c6 + cn_stride);
444 c7 = (float*) ((uintptr_t) c7 + cn_stride);
445 c8 = (float*) ((uintptr_t) c8 + cn_stride);
446 c9 = (float*) ((uintptr_t) c9 + cn_stride);
447 c10 = (float*) ((uintptr_t) c10 + cn_stride);
448 c11 = (float*) ((uintptr_t) c11 + cn_stride);
449 c12 = (float*) ((uintptr_t) c12 + cn_stride);
450 c13 = (float*) ((uintptr_t) c13 + cn_stride);
451
452 nc -= 16;
453 } else {
454 // Prepare mask for valid 32-bit elements (depends on nc).
455 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
456 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
457 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
458 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
459 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
460 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
461 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
462 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
463 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
464 _mm512_mask_storeu_ps(c8, vmask, vscaled8x0123456789ABCDEF);
465 _mm512_mask_storeu_ps(c9, vmask, vscaled9x0123456789ABCDEF);
466 _mm512_mask_storeu_ps(c10, vmask, vscaled10x0123456789ABCDEF);
467 _mm512_mask_storeu_ps(c11, vmask, vscaled11x0123456789ABCDEF);
468 _mm512_mask_storeu_ps(c12, vmask, vscaled12x0123456789ABCDEF);
469 _mm512_mask_storeu_ps(c13, vmask, vscaled13x0123456789ABCDEF);
470 nc = 0;
471 }
472 } while (nc != 0);
473 }