Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-14x16c4-minmax-avx512skx-madd.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_14x16c4__avx512skx_madd(
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 <= 14);
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 const int8_t* a12 = (const int8_t*) ((uintptr_t) a11 + a_stride);
113 float* c12 = (float*) ((uintptr_t) c11 + cm_stride);
114 if XNN_UNPREDICTABLE(mr <= 12) {
115 a12 = a11;
116 c12 = c11;
117 }
118 const int8_t* a13 = (const int8_t*) ((uintptr_t) a12 + a_stride);
119 float* c13 = (float*) ((uintptr_t) c12 + cm_stride);
120 if XNN_UNPREDICTABLE(mr != 14) {
121 a13 = a12;
122 c13 = c12;
123 }
124
125 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
126 XNN_FORCE_REALIZATION(vsign_mask);
127 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point + 128);
128 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point + 128);
129 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point + 128);
130 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point + 128);
131 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point + 128);
132 const __m512i vinput_zero_point5 = _mm512_set1_epi32((int) quantization_params[5].zero_point + 128);
133 const __m512i vinput_zero_point6 = _mm512_set1_epi32((int) quantization_params[6].zero_point + 128);
134 const __m512i vinput_zero_point7 = _mm512_set1_epi32((int) quantization_params[7].zero_point + 128);
135 const __m512i vinput_zero_point8 = _mm512_set1_epi32((int) quantization_params[8].zero_point + 128);
136 const __m512i vinput_zero_point9 = _mm512_set1_epi32((int) quantization_params[9].zero_point + 128);
137 const __m512i vinput_zero_point10 = _mm512_set1_epi32((int) quantization_params[10].zero_point + 128);
138 const __m512i vinput_zero_point11 = _mm512_set1_epi32((int) quantization_params[11].zero_point + 128);
139 const __m512i vinput_zero_point12 = _mm512_set1_epi32((int) quantization_params[12].zero_point + 128);
140 const __m512i vinput_zero_point13 = _mm512_set1_epi32((int) quantization_params[13].zero_point + 128);
141 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
142 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
143 // XNN_FORCE_REALIZATION(voutput_min);
144 // XNN_FORCE_REALIZATION(voutput_max);
145 const __m512i vmask = _mm512_set1_epi8(0x0F);
146 XNN_FORCE_REALIZATION(vmask);
147 do {
148 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
149 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point0);
150 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
151 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point1);
152 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
153 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point2);
154 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
155 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point3);
156 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
157 __m512i vacc4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point4);
158 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
159 __m512i vacc5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point5);
160 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
161 __m512i vacc6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point6);
162 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
163 __m512i vacc7x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point7);
164 __m512i vacc1x7x0123456789ABCDEF = _mm512_setzero_epi32();
165 __m512i vacc8x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point8);
166 __m512i vacc1x8x0123456789ABCDEF = _mm512_setzero_epi32();
167 __m512i vacc9x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point9);
168 __m512i vacc1x9x0123456789ABCDEF = _mm512_setzero_epi32();
169 __m512i vacc10x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point10);
170 __m512i vacc1x10x0123456789ABCDEF = _mm512_setzero_epi32();
171 __m512i vacc11x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point11);
172 __m512i vacc1x11x0123456789ABCDEF = _mm512_setzero_epi32();
173 __m512i vacc12x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point12);
174 __m512i vacc1x12x0123456789ABCDEF = _mm512_setzero_epi32();
175 __m512i vacc13x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point13);
176 __m512i vacc1x13x0123456789ABCDEF = _mm512_setzero_epi32();
177 w = (const int32_t*) w + 16;
178
179 size_t k = kc;
180 while (k >= 8 * sizeof(int8_t)) {
181 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
182 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
183 a0 += 8;
184 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
185 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
186 a1 += 8;
187 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
188 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
189 a2 += 8;
190 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
191 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
192 a3 += 8;
193 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
194 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
195 a4 += 8;
196 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
197 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
198 a5 += 8;
199 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
200 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
201 a6 += 8;
202 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
203 const __m512i va7x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7 + 4)), vsign_mask);
204 a7 += 8;
205 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
206 const __m512i va8x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8 + 4)), vsign_mask);
207 a8 += 8;
208 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
209 const __m512i va9x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9 + 4)), vsign_mask);
210 a9 += 8;
211 const __m512i va10x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10)), vsign_mask);
212 const __m512i va10x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10 + 4)), vsign_mask);
213 a10 += 8;
214 const __m512i va11x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11)), vsign_mask);
215 const __m512i va11x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11 + 4)), vsign_mask);
216 a11 += 8;
217 const __m512i va12x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a12)), vsign_mask);
218 const __m512i va12x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a12 + 4)), vsign_mask);
219 a12 += 8;
220 const __m512i va13x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a13)), vsign_mask);
221 const __m512i va13x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a13 + 4)), vsign_mask);
222 a13 += 8;
223
224 const __m512i vbb0123456789ABCDEFx01234567 = _mm512_load_si512(w);
225 const __m512i vbs0123456789ABCDEFx01234567 = _mm512_srli_epi32(vbb0123456789ABCDEFx01234567, 4);
226 const __m512i vb0123456789ABCDEFx0123 = _mm512_and_si512(vbb0123456789ABCDEFx01234567, vmask);
227 const __m512i vb0123456789ABCDEFx4567 = _mm512_and_si512(vbs0123456789ABCDEFx01234567, vmask);
228
229 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
230 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
231 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
232 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
233 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
234 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
235 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
236 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEFx0123);
237 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEFx0123);
238 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEFx0123);
239 vacc10x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc10x0123456789ABCDEF, va10x0123, vb0123456789ABCDEFx0123);
240 vacc11x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc11x0123456789ABCDEF, va11x0123, vb0123456789ABCDEFx0123);
241 vacc12x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc12x0123456789ABCDEF, va12x0123, vb0123456789ABCDEFx0123);
242 vacc13x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc13x0123456789ABCDEF, va13x0123, vb0123456789ABCDEFx0123);
243 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
244 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
245 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
246 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
247 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
248 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
249 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
250 vacc1x7x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x7x0123456789ABCDEF, va7x4567, vb0123456789ABCDEFx4567);
251 vacc1x8x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x8x0123456789ABCDEF, va8x4567, vb0123456789ABCDEFx4567);
252 vacc1x9x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x9x0123456789ABCDEF, va9x4567, vb0123456789ABCDEFx4567);
253 vacc1x10x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x10x0123456789ABCDEF, va10x4567, vb0123456789ABCDEFx4567);
254 vacc1x11x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x11x0123456789ABCDEF, va11x4567, vb0123456789ABCDEFx4567);
255 vacc1x12x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x12x0123456789ABCDEF, va12x4567, vb0123456789ABCDEFx4567);
256 vacc1x13x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x13x0123456789ABCDEF, va13x4567, vb0123456789ABCDEFx4567);
257
258 w = (const int8_t*) w + 64;
259 k -= 8 * sizeof(int8_t);
260 }
261 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
262 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
263 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
264 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
265 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
266 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
267 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
268 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, vacc1x7x0123456789ABCDEF);
269 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, vacc1x8x0123456789ABCDEF);
270 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, vacc1x9x0123456789ABCDEF);
271 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, vacc1x10x0123456789ABCDEF);
272 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, vacc1x11x0123456789ABCDEF);
273 vacc12x0123456789ABCDEF = _mm512_add_epi32(vacc12x0123456789ABCDEF, vacc1x12x0123456789ABCDEF);
274 vacc13x0123456789ABCDEF = _mm512_add_epi32(vacc13x0123456789ABCDEF, vacc1x13x0123456789ABCDEF);
275
276 if (k != 0) {
277 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
278 a0 += 4;
279 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
280 a1 += 4;
281 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
282 a2 += 4;
283 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
284 a3 += 4;
285 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
286 a4 += 4;
287 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
288 a5 += 4;
289 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
290 a6 += 4;
291 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
292 a7 += 4;
293 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
294 a8 += 4;
295 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
296 a9 += 4;
297 const __m512i va10x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10)), vsign_mask);
298 a10 += 4;
299 const __m512i va11x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11)), vsign_mask);
300 a11 += 4;
301 const __m512i va12x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a12)), vsign_mask);
302 a12 += 4;
303 const __m512i va13x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a13)), vsign_mask);
304 a13 += 4;
305
306 const __m512i vbb0123456789ABCDEF = _mm512_load_si512(w);
307 const __m512i vb0123456789ABCDEF = _mm512_and_si512(vbb0123456789ABCDEF, vmask);
308
309 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
310 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
311 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
312 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
313 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
314 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
315 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
316 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEF);
317 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEF);
318 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEF);
319 vacc10x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc10x0123456789ABCDEF, va10x0123, vb0123456789ABCDEF);
320 vacc11x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc11x0123456789ABCDEF, va11x0123, vb0123456789ABCDEF);
321 vacc12x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc12x0123456789ABCDEF, va12x0123, vb0123456789ABCDEF);
322 vacc13x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc13x0123456789ABCDEF, va13x0123, vb0123456789ABCDEF);
323
324 w = (const int8_t*) w + 64;
325 k -= 4 * sizeof(int8_t);
326 }
327
328 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
329 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
330 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
331 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
332 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
333 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
334 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
335 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
336 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
337 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
338 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
339 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
340 __m512 vscaled12x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc12x0123456789ABCDEF);
341 __m512 vscaled13x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc13x0123456789ABCDEF);
342
343 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
344 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
345 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
346 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
347 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
348 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
349 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
350 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
351 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, _mm512_set1_ps(quantization_params[8].inv_scale));
352 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, _mm512_set1_ps(quantization_params[9].inv_scale));
353 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, _mm512_set1_ps(quantization_params[10].inv_scale));
354 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, _mm512_set1_ps(quantization_params[11].inv_scale));
355 vscaled12x0123456789ABCDEF = _mm512_mul_ps(vscaled12x0123456789ABCDEF, _mm512_set1_ps(quantization_params[12].inv_scale));
356 vscaled13x0123456789ABCDEF = _mm512_mul_ps(vscaled13x0123456789ABCDEF, _mm512_set1_ps(quantization_params[13].inv_scale));
357
358 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
359 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
360 w = (const float*) w + 32;
361
362 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
363 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
364 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
365 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
366 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
367 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
368 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
369 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
370 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
371 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vscaled9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
372 vscaled10x0123456789ABCDEF = _mm512_fmadd_ps(vscaled10x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
373 vscaled11x0123456789ABCDEF = _mm512_fmadd_ps(vscaled11x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
374 vscaled12x0123456789ABCDEF = _mm512_fmadd_ps(vscaled12x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
375 vscaled13x0123456789ABCDEF = _mm512_fmadd_ps(vscaled13x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
376
377 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
378 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
379 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
380 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
381 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
382 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
383 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
384 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
385 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
386 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
387 vscaled10x0123456789ABCDEF = _mm512_max_ps(vscaled10x0123456789ABCDEF, voutput_min);
388 vscaled11x0123456789ABCDEF = _mm512_max_ps(vscaled11x0123456789ABCDEF, voutput_min);
389 vscaled12x0123456789ABCDEF = _mm512_max_ps(vscaled12x0123456789ABCDEF, voutput_min);
390 vscaled13x0123456789ABCDEF = _mm512_max_ps(vscaled13x0123456789ABCDEF, voutput_min);
391
392 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
393 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
394 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
395 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
396 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
397 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
398 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
399 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
400 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
401 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
402 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max);
403 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max);
404 vscaled12x0123456789ABCDEF = _mm512_min_ps(vscaled12x0123456789ABCDEF, voutput_max);
405 vscaled13x0123456789ABCDEF = _mm512_min_ps(vscaled13x0123456789ABCDEF, voutput_max);
406
407 if XNN_LIKELY(nc >= 16) {
408 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
409 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
410 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
411 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
412 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
413 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
414 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
415 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
416 _mm512_storeu_ps(c8, vscaled8x0123456789ABCDEF);
417 _mm512_storeu_ps(c9, vscaled9x0123456789ABCDEF);
418 _mm512_storeu_ps(c10, vscaled10x0123456789ABCDEF);
419 _mm512_storeu_ps(c11, vscaled11x0123456789ABCDEF);
420 _mm512_storeu_ps(c12, vscaled12x0123456789ABCDEF);
421 _mm512_storeu_ps(c13, vscaled13x0123456789ABCDEF);
422
423 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
424 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
425 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
426 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
427 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
428 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
429 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
430 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
431 a8 = (const int8_t*) ((uintptr_t) a8 - kc);
432 a9 = (const int8_t*) ((uintptr_t) a9 - kc);
433 a10 = (const int8_t*) ((uintptr_t) a10 - kc);
434 a11 = (const int8_t*) ((uintptr_t) a11 - kc);
435 a12 = (const int8_t*) ((uintptr_t) a12 - kc);
436 a13 = (const int8_t*) ((uintptr_t) a13 - kc);
437
438 c0 = (float*) ((uintptr_t) c0 + cn_stride);
439 c1 = (float*) ((uintptr_t) c1 + cn_stride);
440 c2 = (float*) ((uintptr_t) c2 + cn_stride);
441 c3 = (float*) ((uintptr_t) c3 + cn_stride);
442 c4 = (float*) ((uintptr_t) c4 + cn_stride);
443 c5 = (float*) ((uintptr_t) c5 + cn_stride);
444 c6 = (float*) ((uintptr_t) c6 + cn_stride);
445 c7 = (float*) ((uintptr_t) c7 + cn_stride);
446 c8 = (float*) ((uintptr_t) c8 + cn_stride);
447 c9 = (float*) ((uintptr_t) c9 + cn_stride);
448 c10 = (float*) ((uintptr_t) c10 + cn_stride);
449 c11 = (float*) ((uintptr_t) c11 + cn_stride);
450 c12 = (float*) ((uintptr_t) c12 + cn_stride);
451 c13 = (float*) ((uintptr_t) c13 + cn_stride);
452
453 nc -= 16;
454 } else {
455 // Prepare mask for valid 32-bit elements (depends on nc).
456 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
457 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
458 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
459 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
460 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
461 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
462 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
463 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
464 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
465 _mm512_mask_storeu_ps(c8, vmask, vscaled8x0123456789ABCDEF);
466 _mm512_mask_storeu_ps(c9, vmask, vscaled9x0123456789ABCDEF);
467 _mm512_mask_storeu_ps(c10, vmask, vscaled10x0123456789ABCDEF);
468 _mm512_mask_storeu_ps(c11, vmask, vscaled11x0123456789ABCDEF);
469 _mm512_mask_storeu_ps(c12, vmask, vscaled12x0123456789ABCDEF);
470 _mm512_mask_storeu_ps(c13, vmask, vscaled13x0123456789ABCDEF);
471 nc = 0;
472 }
473 } while (nc != 0);
474 }