Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-8x8c8-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_8x8c8__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 <= 8);
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 const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
83 uint16_t* c7 = (uint16_t*) ((uintptr_t) c6 + cm_stride);
84 if XNN_UNPREDICTABLE(mr != 8) {
85 a7 = a6;
86 c7 = c6;
87 }
88
89 const __m128i vmask = _mm_set1_epi8(0xF0);
90 XNN_FORCE_REALIZATION(vmask);
91 const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
92 const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
93 XNN_FORCE_REALIZATION(vmin);
94 XNN_FORCE_REALIZATION(vmax);
95
96 do {
97 const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
98 const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
99 const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
100 const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
101 const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
102 const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
103 const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
104 const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
105 const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
106 const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
107 const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
108 const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
109 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point);
110 __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point0);
111 __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point0);
112 __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point0);
113 __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point0);
114 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point);
115 __m256i vacc1x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point1);
116 __m256i vacc1x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point1);
117 __m256i vacc1x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point1);
118 __m256i vacc1x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point1);
119 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point);
120 __m256i vacc2x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point2);
121 __m256i vacc2x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point2);
122 __m256i vacc2x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point2);
123 __m256i vacc2x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point2);
124 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point);
125 __m256i vacc3x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point3);
126 __m256i vacc3x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point3);
127 __m256i vacc3x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point3);
128 __m256i vacc3x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point3);
129 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point);
130 __m256i vacc4x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point4);
131 __m256i vacc4x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point4);
132 __m256i vacc4x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point4);
133 __m256i vacc4x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point4);
134 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point);
135 __m256i vacc5x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point5);
136 __m256i vacc5x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point5);
137 __m256i vacc5x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point5);
138 __m256i vacc5x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point5);
139 const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point);
140 __m256i vacc6x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point6);
141 __m256i vacc6x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point6);
142 __m256i vacc6x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point6);
143 __m256i vacc6x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point6);
144 const __m256i vinput_zero_point7 = _mm256_set1_epi32((int) quantization_params[7].zero_point);
145 __m256i vacc7x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point7);
146 __m256i vacc7x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point7);
147 __m256i vacc7x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point7);
148 __m256i vacc7x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point7);
149 w = (const int32_t*) w + 8;
150
151 size_t k = kc;
152 while (k >= 16 * sizeof(int8_t)) {
153 __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
154 __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
155 a0 += 8;
156 __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
157 __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
158 a1 += 8;
159 __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
160 __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
161 a2 += 8;
162 __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
163 __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
164 a3 += 8;
165 __m128i va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
166 __m256i vxa4 = _mm256_cvtepi8_epi16(va4);
167 a4 += 8;
168 __m128i va5 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a5));
169 __m256i vxa5 = _mm256_cvtepi8_epi16(va5);
170 a5 += 8;
171 __m128i va6 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a6));
172 __m256i vxa6 = _mm256_cvtepi8_epi16(va6);
173 a6 += 8;
174 __m128i va7 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a7));
175 __m256i vxa7 = _mm256_cvtepi8_epi16(va7);
176 a7 += 8;
177
178 __m128i vb01 = _mm_load_si128((const __m128i*) w);
179 __m128i vbs01 = _mm_slli_epi32(vb01, 4);
180 __m128i vbm01 = _mm_and_si128(vbs01, vmask);
181 __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
182
183 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
184 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
185 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
186 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
187 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
188 vacc5x01 = _mm256_add_epi32(vacc5x01, _mm256_madd_epi16(vxa5, vxb01));
189 vacc6x01 = _mm256_add_epi32(vacc6x01, _mm256_madd_epi16(vxa6, vxb01));
190 vacc7x01 = _mm256_add_epi32(vacc7x01, _mm256_madd_epi16(vxa7, vxb01));
191 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
192 __m128i vbs23 = _mm_slli_epi32(vb23, 4);
193 __m128i vbm23 = _mm_and_si128(vbs23, vmask);
194 __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
195
196 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
197 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
198 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
199 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
200 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
201 vacc5x23 = _mm256_add_epi32(vacc5x23, _mm256_madd_epi16(vxa5, vxb23));
202 vacc6x23 = _mm256_add_epi32(vacc6x23, _mm256_madd_epi16(vxa6, vxb23));
203 vacc7x23 = _mm256_add_epi32(vacc7x23, _mm256_madd_epi16(vxa7, vxb23));
204 __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
205 __m128i vbs45 = _mm_slli_epi32(vb45, 4);
206 __m128i vbm45 = _mm_and_si128(vbs45, vmask);
207 __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
208
209 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
210 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
211 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
212 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
213 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
214 vacc5x45 = _mm256_add_epi32(vacc5x45, _mm256_madd_epi16(vxa5, vxb45));
215 vacc6x45 = _mm256_add_epi32(vacc6x45, _mm256_madd_epi16(vxa6, vxb45));
216 vacc7x45 = _mm256_add_epi32(vacc7x45, _mm256_madd_epi16(vxa7, vxb45));
217 __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
218 __m128i vbs67 = _mm_slli_epi32(vb67, 4);
219 __m128i vbm67 = _mm_and_si128(vbs67, vmask);
220 __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
221
222 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
223 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
224 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
225 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
226 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
227 vacc5x67 = _mm256_add_epi32(vacc5x67, _mm256_madd_epi16(vxa5, vxb67));
228 vacc6x67 = _mm256_add_epi32(vacc6x67, _mm256_madd_epi16(vxa6, vxb67));
229 vacc7x67 = _mm256_add_epi32(vacc7x67, _mm256_madd_epi16(vxa7, vxb67));
230
231 va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
232 vxa0 = _mm256_cvtepi8_epi16(va0);
233 a0 += 8;
234 va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
235 vxa1 = _mm256_cvtepi8_epi16(va1);
236 a1 += 8;
237 va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
238 vxa2 = _mm256_cvtepi8_epi16(va2);
239 a2 += 8;
240 va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
241 vxa3 = _mm256_cvtepi8_epi16(va3);
242 a3 += 8;
243 va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
244 vxa4 = _mm256_cvtepi8_epi16(va4);
245 a4 += 8;
246 va5 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a5));
247 vxa5 = _mm256_cvtepi8_epi16(va5);
248 a5 += 8;
249 va6 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a6));
250 vxa6 = _mm256_cvtepi8_epi16(va6);
251 a6 += 8;
252 va7 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a7));
253 vxa7 = _mm256_cvtepi8_epi16(va7);
254 a7 += 8;
255
256 vbm01 = _mm_and_si128(vb01, vmask);
257 vxb01 = _mm256_cvtepi8_epi16(vbm01);
258
259 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
260 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
261 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
262 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
263 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
264 vacc5x01 = _mm256_add_epi32(vacc5x01, _mm256_madd_epi16(vxa5, vxb01));
265 vacc6x01 = _mm256_add_epi32(vacc6x01, _mm256_madd_epi16(vxa6, vxb01));
266 vacc7x01 = _mm256_add_epi32(vacc7x01, _mm256_madd_epi16(vxa7, vxb01));
267 vbm23 = _mm_and_si128(vb23, vmask);
268 vxb23 = _mm256_cvtepi8_epi16(vbm23);
269
270 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
271 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
272 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
273 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
274 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
275 vacc5x23 = _mm256_add_epi32(vacc5x23, _mm256_madd_epi16(vxa5, vxb23));
276 vacc6x23 = _mm256_add_epi32(vacc6x23, _mm256_madd_epi16(vxa6, vxb23));
277 vacc7x23 = _mm256_add_epi32(vacc7x23, _mm256_madd_epi16(vxa7, vxb23));
278 vbm45 = _mm_and_si128(vb45, vmask);
279 vxb45 = _mm256_cvtepi8_epi16(vbm45);
280
281 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
282 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
283 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
284 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
285 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
286 vacc5x45 = _mm256_add_epi32(vacc5x45, _mm256_madd_epi16(vxa5, vxb45));
287 vacc6x45 = _mm256_add_epi32(vacc6x45, _mm256_madd_epi16(vxa6, vxb45));
288 vacc7x45 = _mm256_add_epi32(vacc7x45, _mm256_madd_epi16(vxa7, vxb45));
289 vbm67 = _mm_and_si128(vb67, vmask);
290 vxb67 = _mm256_cvtepi8_epi16(vbm67);
291
292 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
293 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
294 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
295 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
296 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
297 vacc5x67 = _mm256_add_epi32(vacc5x67, _mm256_madd_epi16(vxa5, vxb67));
298 vacc6x67 = _mm256_add_epi32(vacc6x67, _mm256_madd_epi16(vxa6, vxb67));
299 vacc7x67 = _mm256_add_epi32(vacc7x67, _mm256_madd_epi16(vxa7, vxb67));
300
301 w = (const int8_t*) w + 64;
302 k -= 16 * sizeof(int8_t);
303 }
304
305 while (k >= 8 * sizeof(int8_t)) {
306 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
307 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
308 a0 += 8;
309 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
310 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
311 a1 += 8;
312 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
313 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
314 a2 += 8;
315 const __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
316 const __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
317 a3 += 8;
318 const __m128i va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
319 const __m256i vxa4 = _mm256_cvtepi8_epi16(va4);
320 a4 += 8;
321 const __m128i va5 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a5));
322 const __m256i vxa5 = _mm256_cvtepi8_epi16(va5);
323 a5 += 8;
324 const __m128i va6 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a6));
325 const __m256i vxa6 = _mm256_cvtepi8_epi16(va6);
326 a6 += 8;
327 const __m128i va7 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a7));
328 const __m256i vxa7 = _mm256_cvtepi8_epi16(va7);
329 a7 += 8;
330
331 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
332 const __m128i vbs01 = _mm_slli_epi32(vb01, 4);
333 const __m128i vbm01 = _mm_and_si128(vbs01, vmask);
334 const __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
335
336 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
337 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
338 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
339 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
340 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
341 vacc5x01 = _mm256_add_epi32(vacc5x01, _mm256_madd_epi16(vxa5, vxb01));
342 vacc6x01 = _mm256_add_epi32(vacc6x01, _mm256_madd_epi16(vxa6, vxb01));
343 vacc7x01 = _mm256_add_epi32(vacc7x01, _mm256_madd_epi16(vxa7, vxb01));
344 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
345 const __m128i vbs23 = _mm_slli_epi32(vb23, 4);
346 const __m128i vbm23 = _mm_and_si128(vbs23, vmask);
347 const __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
348
349 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
350 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
351 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
352 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
353 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
354 vacc5x23 = _mm256_add_epi32(vacc5x23, _mm256_madd_epi16(vxa5, vxb23));
355 vacc6x23 = _mm256_add_epi32(vacc6x23, _mm256_madd_epi16(vxa6, vxb23));
356 vacc7x23 = _mm256_add_epi32(vacc7x23, _mm256_madd_epi16(vxa7, vxb23));
357 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
358 const __m128i vbs45 = _mm_slli_epi32(vb45, 4);
359 const __m128i vbm45 = _mm_and_si128(vbs45, vmask);
360 const __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
361
362 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
363 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
364 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
365 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
366 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
367 vacc5x45 = _mm256_add_epi32(vacc5x45, _mm256_madd_epi16(vxa5, vxb45));
368 vacc6x45 = _mm256_add_epi32(vacc6x45, _mm256_madd_epi16(vxa6, vxb45));
369 vacc7x45 = _mm256_add_epi32(vacc7x45, _mm256_madd_epi16(vxa7, vxb45));
370 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
371 const __m128i vbs67 = _mm_slli_epi32(vb67, 4);
372 const __m128i vbm67 = _mm_and_si128(vbs67, vmask);
373 const __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
374
375 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
376 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
377 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
378 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
379 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
380 vacc5x67 = _mm256_add_epi32(vacc5x67, _mm256_madd_epi16(vxa5, vxb67));
381 vacc6x67 = _mm256_add_epi32(vacc6x67, _mm256_madd_epi16(vxa6, vxb67));
382 vacc7x67 = _mm256_add_epi32(vacc7x67, _mm256_madd_epi16(vxa7, vxb67));
383
384 w = (const int8_t*) w + 64;
385 k -= 8 * sizeof(int8_t);
386 }
387
388 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
389 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
390 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
391 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
392 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
393 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
394 const __m256i vacc3x0213 = _mm256_hadd_epi32(vacc3x01, vacc3x23);
395 const __m256i vacc3x4657 = _mm256_hadd_epi32(vacc3x45, vacc3x67);
396 const __m256i vacc4x0213 = _mm256_hadd_epi32(vacc4x01, vacc4x23);
397 const __m256i vacc4x4657 = _mm256_hadd_epi32(vacc4x45, vacc4x67);
398 const __m256i vacc5x0213 = _mm256_hadd_epi32(vacc5x01, vacc5x23);
399 const __m256i vacc5x4657 = _mm256_hadd_epi32(vacc5x45, vacc5x67);
400 const __m256i vacc6x0213 = _mm256_hadd_epi32(vacc6x01, vacc6x23);
401 const __m256i vacc6x4657 = _mm256_hadd_epi32(vacc6x45, vacc6x67);
402 const __m256i vacc7x0213 = _mm256_hadd_epi32(vacc7x01, vacc7x23);
403 const __m256i vacc7x4657 = _mm256_hadd_epi32(vacc7x45, vacc7x67);
404
405 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
406 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
407 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
408 const __m256i vacc3x02461357 = _mm256_hadd_epi32(vacc3x0213, vacc3x4657);
409 const __m256i vacc4x02461357 = _mm256_hadd_epi32(vacc4x0213, vacc4x4657);
410 const __m256i vacc5x02461357 = _mm256_hadd_epi32(vacc5x0213, vacc5x4657);
411 const __m256i vacc6x02461357 = _mm256_hadd_epi32(vacc6x0213, vacc6x4657);
412 const __m256i vacc7x02461357 = _mm256_hadd_epi32(vacc7x0213, vacc7x4657);
413
414 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
415 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
416 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
417 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
418 __m256i vacc3x01234567 = _mm256_permutevar8x32_epi32(vacc3x02461357, vpermute_mask);
419 __m256i vacc4x01234567 = _mm256_permutevar8x32_epi32(vacc4x02461357, vpermute_mask);
420 __m256i vacc5x01234567 = _mm256_permutevar8x32_epi32(vacc5x02461357, vpermute_mask);
421 __m256i vacc6x01234567 = _mm256_permutevar8x32_epi32(vacc6x02461357, vpermute_mask);
422 __m256i vacc7x01234567 = _mm256_permutevar8x32_epi32(vacc7x02461357, vpermute_mask);
423
424 vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4);
425 vacc1x01234567 = _mm256_srai_epi32(vacc1x01234567, 4);
426 vacc2x01234567 = _mm256_srai_epi32(vacc2x01234567, 4);
427 vacc3x01234567 = _mm256_srai_epi32(vacc3x01234567, 4);
428 vacc4x01234567 = _mm256_srai_epi32(vacc4x01234567, 4);
429 vacc5x01234567 = _mm256_srai_epi32(vacc5x01234567, 4);
430 vacc6x01234567 = _mm256_srai_epi32(vacc6x01234567, 4);
431 vacc7x01234567 = _mm256_srai_epi32(vacc7x01234567, 4);
432 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
433 const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale);
434 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
435 const __m256 vinput_scale1 = _mm256_broadcast_ss(&quantization_params[1].inv_scale);
436 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
437 const __m256 vinput_scale2 = _mm256_broadcast_ss(&quantization_params[2].inv_scale);
438 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
439 const __m256 vinput_scale3 = _mm256_broadcast_ss(&quantization_params[3].inv_scale);
440 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
441 const __m256 vinput_scale4 = _mm256_broadcast_ss(&quantization_params[4].inv_scale);
442 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
443 const __m256 vinput_scale5 = _mm256_broadcast_ss(&quantization_params[5].inv_scale);
444 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
445 const __m256 vinput_scale6 = _mm256_broadcast_ss(&quantization_params[6].inv_scale);
446 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
447 const __m256 vinput_scale7 = _mm256_broadcast_ss(&quantization_params[7].inv_scale);
448
449 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale0);
450 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_scale1);
451 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_scale2);
452 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_scale3);
453 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_scale4);
454 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_scale5);
455 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_scale6);
456 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vinput_scale7);
457
458 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
459 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
460 w = (const float*) w + 16;
461 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
462 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
463 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
464 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
465 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
466 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
467 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
468 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
469
470 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
471 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
472 vout2x01234567 = _mm256_max_ps(vout2x01234567, vmin);
473 vout3x01234567 = _mm256_max_ps(vout3x01234567, vmin);
474 vout4x01234567 = _mm256_max_ps(vout4x01234567, vmin);
475 vout5x01234567 = _mm256_max_ps(vout5x01234567, vmin);
476 vout6x01234567 = _mm256_max_ps(vout6x01234567, vmin);
477 vout7x01234567 = _mm256_max_ps(vout7x01234567, vmin);
478
479 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
480 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
481 vout2x01234567 = _mm256_min_ps(vout2x01234567, vmax);
482 vout3x01234567 = _mm256_min_ps(vout3x01234567, vmax);
483 vout4x01234567 = _mm256_min_ps(vout4x01234567, vmax);
484 vout5x01234567 = _mm256_min_ps(vout5x01234567, vmax);
485 vout6x01234567 = _mm256_min_ps(vout6x01234567, vmax);
486 vout7x01234567 = _mm256_min_ps(vout7x01234567, vmax);
487 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
488 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
489 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
490 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
491 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
492 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
493 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
494 __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT);
495 if XNN_LIKELY(nc >= 8) {
496 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
497 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
498 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
499 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
500 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
501 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
502 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
503 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
504 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
505 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
506 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
507 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
508 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
509 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
510 _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567);
511 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
512
513 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
514 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
515 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
516 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
517 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
518 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
519 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
520 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
521
522 nc -= 8;
523 } else {
524 // Prepare mask for valid 16-bit elements (depends on nc).
525 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
526 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
527 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
528 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
529 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
530 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
531 _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567);
532 _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567);
533 _mm_mask_storeu_epi16(c7, vmask, vfp16out7x01234567);
534 nc = 0;
535 }
536 } while (nc != 0);
537 }