Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-7x8c8-minmax-avx256skx.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx8c8-avx2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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_f16_qc4w_gemm_minmax_ukernel_7x8c8__avx256skx(
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 xnn_float16* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const struct xnn_f16_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 <= 7);
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, 8 * sizeof(int8_t));
44 const int8_t* a0 = a;
45 uint16_t* c0 = (uint16_t*) c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 uint16_t* c1 = (uint16_t*) ((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 uint16_t* c2 = (uint16_t*) ((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 uint16_t* c3 = (uint16_t*) ((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 uint16_t* c4 = (uint16_t*) ((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 uint16_t* c5 = (uint16_t*) ((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 uint16_t* c6 = (uint16_t*) ((uintptr_t) c5 + cm_stride);
78 if XNN_UNPREDICTABLE(mr <= 6) {
79 a6 = a5;
80 c6 = c5;
81 }
82
83 const __m128i vmask = _mm_set1_epi8(0xF0);
84 XNN_FORCE_REALIZATION(vmask);
85 const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
86 const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
87 XNN_FORCE_REALIZATION(vmin);
88 XNN_FORCE_REALIZATION(vmax);
89
90 do {
91 const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
92 const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
93 const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
94 const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
95 const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
96 const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
97 const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
98 const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
99 const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
100 const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
101 const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
102 const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
103 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point);
104 __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point0);
105 __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point0);
106 __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point0);
107 __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point0);
108 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point);
109 __m256i vacc1x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point1);
110 __m256i vacc1x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point1);
111 __m256i vacc1x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point1);
112 __m256i vacc1x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point1);
113 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point);
114 __m256i vacc2x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point2);
115 __m256i vacc2x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point2);
116 __m256i vacc2x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point2);
117 __m256i vacc2x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point2);
118 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point);
119 __m256i vacc3x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point3);
120 __m256i vacc3x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point3);
121 __m256i vacc3x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point3);
122 __m256i vacc3x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point3);
123 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point);
124 __m256i vacc4x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point4);
125 __m256i vacc4x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point4);
126 __m256i vacc4x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point4);
127 __m256i vacc4x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point4);
128 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point);
129 __m256i vacc5x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point5);
130 __m256i vacc5x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point5);
131 __m256i vacc5x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point5);
132 __m256i vacc5x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point5);
133 const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point);
134 __m256i vacc6x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point6);
135 __m256i vacc6x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point6);
136 __m256i vacc6x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point6);
137 __m256i vacc6x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point6);
138 w = (const int32_t*) w + 8;
139
140 size_t k = kc;
141 while (k >= 16 * sizeof(int8_t)) {
142 __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
143 __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
144 a0 += 8;
145 __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
146 __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
147 a1 += 8;
148 __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
149 __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
150 a2 += 8;
151 __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
152 __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
153 a3 += 8;
154 __m128i va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
155 __m256i vxa4 = _mm256_cvtepi8_epi16(va4);
156 a4 += 8;
157 __m128i va5 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a5));
158 __m256i vxa5 = _mm256_cvtepi8_epi16(va5);
159 a5 += 8;
160 __m128i va6 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a6));
161 __m256i vxa6 = _mm256_cvtepi8_epi16(va6);
162 a6 += 8;
163
164 __m128i vb01 = _mm_load_si128((const __m128i*) w);
165 __m128i vbs01 = _mm_slli_epi32(vb01, 4);
166 __m128i vbm01 = _mm_and_si128(vbs01, vmask);
167 __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
168
169 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
170 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
171 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
172 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
173 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
174 vacc5x01 = _mm256_add_epi32(vacc5x01, _mm256_madd_epi16(vxa5, vxb01));
175 vacc6x01 = _mm256_add_epi32(vacc6x01, _mm256_madd_epi16(vxa6, vxb01));
176 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
177 __m128i vbs23 = _mm_slli_epi32(vb23, 4);
178 __m128i vbm23 = _mm_and_si128(vbs23, vmask);
179 __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
180
181 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
182 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
183 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
184 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
185 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
186 vacc5x23 = _mm256_add_epi32(vacc5x23, _mm256_madd_epi16(vxa5, vxb23));
187 vacc6x23 = _mm256_add_epi32(vacc6x23, _mm256_madd_epi16(vxa6, vxb23));
188 __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
189 __m128i vbs45 = _mm_slli_epi32(vb45, 4);
190 __m128i vbm45 = _mm_and_si128(vbs45, vmask);
191 __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
192
193 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
194 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
195 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
196 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
197 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
198 vacc5x45 = _mm256_add_epi32(vacc5x45, _mm256_madd_epi16(vxa5, vxb45));
199 vacc6x45 = _mm256_add_epi32(vacc6x45, _mm256_madd_epi16(vxa6, vxb45));
200 __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
201 __m128i vbs67 = _mm_slli_epi32(vb67, 4);
202 __m128i vbm67 = _mm_and_si128(vbs67, vmask);
203 __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
204
205 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
206 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
207 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
208 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
209 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
210 vacc5x67 = _mm256_add_epi32(vacc5x67, _mm256_madd_epi16(vxa5, vxb67));
211 vacc6x67 = _mm256_add_epi32(vacc6x67, _mm256_madd_epi16(vxa6, vxb67));
212
213 va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
214 vxa0 = _mm256_cvtepi8_epi16(va0);
215 a0 += 8;
216 va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
217 vxa1 = _mm256_cvtepi8_epi16(va1);
218 a1 += 8;
219 va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
220 vxa2 = _mm256_cvtepi8_epi16(va2);
221 a2 += 8;
222 va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
223 vxa3 = _mm256_cvtepi8_epi16(va3);
224 a3 += 8;
225 va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
226 vxa4 = _mm256_cvtepi8_epi16(va4);
227 a4 += 8;
228 va5 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a5));
229 vxa5 = _mm256_cvtepi8_epi16(va5);
230 a5 += 8;
231 va6 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a6));
232 vxa6 = _mm256_cvtepi8_epi16(va6);
233 a6 += 8;
234
235 vbm01 = _mm_and_si128(vb01, vmask);
236 vxb01 = _mm256_cvtepi8_epi16(vbm01);
237
238 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
239 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
240 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
241 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
242 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
243 vacc5x01 = _mm256_add_epi32(vacc5x01, _mm256_madd_epi16(vxa5, vxb01));
244 vacc6x01 = _mm256_add_epi32(vacc6x01, _mm256_madd_epi16(vxa6, vxb01));
245 vbm23 = _mm_and_si128(vb23, vmask);
246 vxb23 = _mm256_cvtepi8_epi16(vbm23);
247
248 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
249 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
250 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
251 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
252 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
253 vacc5x23 = _mm256_add_epi32(vacc5x23, _mm256_madd_epi16(vxa5, vxb23));
254 vacc6x23 = _mm256_add_epi32(vacc6x23, _mm256_madd_epi16(vxa6, vxb23));
255 vbm45 = _mm_and_si128(vb45, vmask);
256 vxb45 = _mm256_cvtepi8_epi16(vbm45);
257
258 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
259 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
260 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
261 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
262 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
263 vacc5x45 = _mm256_add_epi32(vacc5x45, _mm256_madd_epi16(vxa5, vxb45));
264 vacc6x45 = _mm256_add_epi32(vacc6x45, _mm256_madd_epi16(vxa6, vxb45));
265 vbm67 = _mm_and_si128(vb67, vmask);
266 vxb67 = _mm256_cvtepi8_epi16(vbm67);
267
268 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
269 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
270 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
271 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
272 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
273 vacc5x67 = _mm256_add_epi32(vacc5x67, _mm256_madd_epi16(vxa5, vxb67));
274 vacc6x67 = _mm256_add_epi32(vacc6x67, _mm256_madd_epi16(vxa6, vxb67));
275
276 w = (const int8_t*) w + 64;
277 k -= 16 * sizeof(int8_t);
278 }
279
280 while (k >= 8 * sizeof(int8_t)) {
281 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
282 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
283 a0 += 8;
284 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
285 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
286 a1 += 8;
287 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
288 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
289 a2 += 8;
290 const __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
291 const __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
292 a3 += 8;
293 const __m128i va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
294 const __m256i vxa4 = _mm256_cvtepi8_epi16(va4);
295 a4 += 8;
296 const __m128i va5 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a5));
297 const __m256i vxa5 = _mm256_cvtepi8_epi16(va5);
298 a5 += 8;
299 const __m128i va6 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a6));
300 const __m256i vxa6 = _mm256_cvtepi8_epi16(va6);
301 a6 += 8;
302
303 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
304 const __m128i vbs01 = _mm_slli_epi32(vb01, 4);
305 const __m128i vbm01 = _mm_and_si128(vbs01, vmask);
306 const __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
307
308 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
309 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
310 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
311 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
312 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
313 vacc5x01 = _mm256_add_epi32(vacc5x01, _mm256_madd_epi16(vxa5, vxb01));
314 vacc6x01 = _mm256_add_epi32(vacc6x01, _mm256_madd_epi16(vxa6, vxb01));
315 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
316 const __m128i vbs23 = _mm_slli_epi32(vb23, 4);
317 const __m128i vbm23 = _mm_and_si128(vbs23, vmask);
318 const __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
319
320 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
321 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
322 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
323 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
324 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
325 vacc5x23 = _mm256_add_epi32(vacc5x23, _mm256_madd_epi16(vxa5, vxb23));
326 vacc6x23 = _mm256_add_epi32(vacc6x23, _mm256_madd_epi16(vxa6, vxb23));
327 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
328 const __m128i vbs45 = _mm_slli_epi32(vb45, 4);
329 const __m128i vbm45 = _mm_and_si128(vbs45, vmask);
330 const __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
331
332 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
333 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
334 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
335 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
336 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
337 vacc5x45 = _mm256_add_epi32(vacc5x45, _mm256_madd_epi16(vxa5, vxb45));
338 vacc6x45 = _mm256_add_epi32(vacc6x45, _mm256_madd_epi16(vxa6, vxb45));
339 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
340 const __m128i vbs67 = _mm_slli_epi32(vb67, 4);
341 const __m128i vbm67 = _mm_and_si128(vbs67, vmask);
342 const __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
343
344 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
345 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
346 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
347 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
348 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
349 vacc5x67 = _mm256_add_epi32(vacc5x67, _mm256_madd_epi16(vxa5, vxb67));
350 vacc6x67 = _mm256_add_epi32(vacc6x67, _mm256_madd_epi16(vxa6, vxb67));
351
352 w = (const int8_t*) w + 64;
353 k -= 8 * sizeof(int8_t);
354 }
355
356 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
357 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
358 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
359 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
360 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
361 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
362 const __m256i vacc3x0213 = _mm256_hadd_epi32(vacc3x01, vacc3x23);
363 const __m256i vacc3x4657 = _mm256_hadd_epi32(vacc3x45, vacc3x67);
364 const __m256i vacc4x0213 = _mm256_hadd_epi32(vacc4x01, vacc4x23);
365 const __m256i vacc4x4657 = _mm256_hadd_epi32(vacc4x45, vacc4x67);
366 const __m256i vacc5x0213 = _mm256_hadd_epi32(vacc5x01, vacc5x23);
367 const __m256i vacc5x4657 = _mm256_hadd_epi32(vacc5x45, vacc5x67);
368 const __m256i vacc6x0213 = _mm256_hadd_epi32(vacc6x01, vacc6x23);
369 const __m256i vacc6x4657 = _mm256_hadd_epi32(vacc6x45, vacc6x67);
370
371 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
372 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
373 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
374 const __m256i vacc3x02461357 = _mm256_hadd_epi32(vacc3x0213, vacc3x4657);
375 const __m256i vacc4x02461357 = _mm256_hadd_epi32(vacc4x0213, vacc4x4657);
376 const __m256i vacc5x02461357 = _mm256_hadd_epi32(vacc5x0213, vacc5x4657);
377 const __m256i vacc6x02461357 = _mm256_hadd_epi32(vacc6x0213, vacc6x4657);
378
379 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
380 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
381 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
382 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
383 __m256i vacc3x01234567 = _mm256_permutevar8x32_epi32(vacc3x02461357, vpermute_mask);
384 __m256i vacc4x01234567 = _mm256_permutevar8x32_epi32(vacc4x02461357, vpermute_mask);
385 __m256i vacc5x01234567 = _mm256_permutevar8x32_epi32(vacc5x02461357, vpermute_mask);
386 __m256i vacc6x01234567 = _mm256_permutevar8x32_epi32(vacc6x02461357, vpermute_mask);
387
388 vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4);
389 vacc1x01234567 = _mm256_srai_epi32(vacc1x01234567, 4);
390 vacc2x01234567 = _mm256_srai_epi32(vacc2x01234567, 4);
391 vacc3x01234567 = _mm256_srai_epi32(vacc3x01234567, 4);
392 vacc4x01234567 = _mm256_srai_epi32(vacc4x01234567, 4);
393 vacc5x01234567 = _mm256_srai_epi32(vacc5x01234567, 4);
394 vacc6x01234567 = _mm256_srai_epi32(vacc6x01234567, 4);
395 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
396 const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale);
397 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
398 const __m256 vinput_scale1 = _mm256_broadcast_ss(&quantization_params[1].inv_scale);
399 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
400 const __m256 vinput_scale2 = _mm256_broadcast_ss(&quantization_params[2].inv_scale);
401 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
402 const __m256 vinput_scale3 = _mm256_broadcast_ss(&quantization_params[3].inv_scale);
403 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
404 const __m256 vinput_scale4 = _mm256_broadcast_ss(&quantization_params[4].inv_scale);
405 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
406 const __m256 vinput_scale5 = _mm256_broadcast_ss(&quantization_params[5].inv_scale);
407 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
408 const __m256 vinput_scale6 = _mm256_broadcast_ss(&quantization_params[6].inv_scale);
409
410 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale0);
411 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_scale1);
412 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_scale2);
413 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_scale3);
414 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_scale4);
415 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_scale5);
416 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_scale6);
417
418 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
419 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
420 w = (const float*) w + 16;
421 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
422 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
423 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
424 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
425 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
426 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
427 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
428
429 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
430 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
431 vout2x01234567 = _mm256_max_ps(vout2x01234567, vmin);
432 vout3x01234567 = _mm256_max_ps(vout3x01234567, vmin);
433 vout4x01234567 = _mm256_max_ps(vout4x01234567, vmin);
434 vout5x01234567 = _mm256_max_ps(vout5x01234567, vmin);
435 vout6x01234567 = _mm256_max_ps(vout6x01234567, vmin);
436
437 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
438 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
439 vout2x01234567 = _mm256_min_ps(vout2x01234567, vmax);
440 vout3x01234567 = _mm256_min_ps(vout3x01234567, vmax);
441 vout4x01234567 = _mm256_min_ps(vout4x01234567, vmax);
442 vout5x01234567 = _mm256_min_ps(vout5x01234567, vmax);
443 vout6x01234567 = _mm256_min_ps(vout6x01234567, vmax);
444 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
445 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
446 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
447 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
448 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
449 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
450 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
451 if XNN_LIKELY(nc >= 8) {
452 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
453 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
454 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
455 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
456 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
457 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
458 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
459 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
460 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
461 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
462 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
463 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
464 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
465 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
466
467 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
468 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
469 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
470 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
471 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
472 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
473 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
474
475 nc -= 8;
476 } else {
477 // Prepare mask for valid 16-bit elements (depends on nc).
478 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
479 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
480 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
481 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
482 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
483 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
484 _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567);
485 _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567);
486 nc = 0;
487 }
488 } while (nc != 0);
489 }