Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-5x8c8-minmax-avx2.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_5x8c8__avx2(
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 <= 5);
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
71 const __m128i vmask = _mm_set1_epi8(0xF0);
72 XNN_FORCE_REALIZATION(vmask);
73 const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
74 const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
75 XNN_FORCE_REALIZATION(vmin);
76 XNN_FORCE_REALIZATION(vmax);
77
78 do {
79 const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
80 const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
81 const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
82 const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
83 const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
84 const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
85 const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
86 const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
87 const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
88 const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
89 const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
90 const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
91 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point);
92 __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point0);
93 __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point0);
94 __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point0);
95 __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point0);
96 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point);
97 __m256i vacc1x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point1);
98 __m256i vacc1x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point1);
99 __m256i vacc1x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point1);
100 __m256i vacc1x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point1);
101 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point);
102 __m256i vacc2x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point2);
103 __m256i vacc2x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point2);
104 __m256i vacc2x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point2);
105 __m256i vacc2x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point2);
106 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point);
107 __m256i vacc3x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point3);
108 __m256i vacc3x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point3);
109 __m256i vacc3x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point3);
110 __m256i vacc3x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point3);
111 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point);
112 __m256i vacc4x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point4);
113 __m256i vacc4x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point4);
114 __m256i vacc4x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point4);
115 __m256i vacc4x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point4);
116 w = (const int32_t*) w + 8;
117
118 size_t k = kc;
119 while (k >= 16 * sizeof(int8_t)) {
120 __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
121 __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
122 a0 += 8;
123 __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
124 __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
125 a1 += 8;
126 __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
127 __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
128 a2 += 8;
129 __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
130 __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
131 a3 += 8;
132 __m128i va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
133 __m256i vxa4 = _mm256_cvtepi8_epi16(va4);
134 a4 += 8;
135
136 __m128i vb01 = _mm_load_si128((const __m128i*) w);
137 __m128i vbs01 = _mm_slli_epi32(vb01, 4);
138 __m128i vbm01 = _mm_and_si128(vbs01, vmask);
139 __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
140
141 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
142 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
143 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
144 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
145 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
146 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
147 __m128i vbs23 = _mm_slli_epi32(vb23, 4);
148 __m128i vbm23 = _mm_and_si128(vbs23, vmask);
149 __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
150
151 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
152 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
153 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
154 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
155 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
156 __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
157 __m128i vbs45 = _mm_slli_epi32(vb45, 4);
158 __m128i vbm45 = _mm_and_si128(vbs45, vmask);
159 __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
160
161 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
162 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
163 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
164 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
165 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
166 __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
167 __m128i vbs67 = _mm_slli_epi32(vb67, 4);
168 __m128i vbm67 = _mm_and_si128(vbs67, vmask);
169 __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
170
171 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
172 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
173 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
174 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
175 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
176
177 va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
178 vxa0 = _mm256_cvtepi8_epi16(va0);
179 a0 += 8;
180 va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
181 vxa1 = _mm256_cvtepi8_epi16(va1);
182 a1 += 8;
183 va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
184 vxa2 = _mm256_cvtepi8_epi16(va2);
185 a2 += 8;
186 va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
187 vxa3 = _mm256_cvtepi8_epi16(va3);
188 a3 += 8;
189 va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
190 vxa4 = _mm256_cvtepi8_epi16(va4);
191 a4 += 8;
192
193 vbm01 = _mm_and_si128(vb01, vmask);
194 vxb01 = _mm256_cvtepi8_epi16(vbm01);
195
196 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
197 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
198 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
199 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
200 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
201 vbm23 = _mm_and_si128(vb23, vmask);
202 vxb23 = _mm256_cvtepi8_epi16(vbm23);
203
204 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
205 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
206 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
207 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
208 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
209 vbm45 = _mm_and_si128(vb45, vmask);
210 vxb45 = _mm256_cvtepi8_epi16(vbm45);
211
212 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
213 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
214 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
215 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
216 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
217 vbm67 = _mm_and_si128(vb67, vmask);
218 vxb67 = _mm256_cvtepi8_epi16(vbm67);
219
220 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
221 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
222 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
223 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
224 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
225
226 w = (const int8_t*) w + 64;
227 k -= 16 * sizeof(int8_t);
228 }
229
230 while (k >= 8 * sizeof(int8_t)) {
231 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
232 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
233 a0 += 8;
234 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
235 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
236 a1 += 8;
237 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
238 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
239 a2 += 8;
240 const __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
241 const __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
242 a3 += 8;
243 const __m128i va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
244 const __m256i vxa4 = _mm256_cvtepi8_epi16(va4);
245 a4 += 8;
246
247 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
248 const __m128i vbs01 = _mm_slli_epi32(vb01, 4);
249 const __m128i vbm01 = _mm_and_si128(vbs01, vmask);
250 const __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
251
252 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
253 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
254 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
255 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
256 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
257 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
258 const __m128i vbs23 = _mm_slli_epi32(vb23, 4);
259 const __m128i vbm23 = _mm_and_si128(vbs23, vmask);
260 const __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
261
262 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
263 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
264 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
265 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
266 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
267 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
268 const __m128i vbs45 = _mm_slli_epi32(vb45, 4);
269 const __m128i vbm45 = _mm_and_si128(vbs45, vmask);
270 const __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
271
272 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
273 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
274 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
275 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
276 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
277 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
278 const __m128i vbs67 = _mm_slli_epi32(vb67, 4);
279 const __m128i vbm67 = _mm_and_si128(vbs67, vmask);
280 const __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
281
282 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
283 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
284 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
285 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
286 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
287
288 w = (const int8_t*) w + 64;
289 k -= 8 * sizeof(int8_t);
290 }
291
292 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
293 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
294 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
295 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
296 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
297 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
298 const __m256i vacc3x0213 = _mm256_hadd_epi32(vacc3x01, vacc3x23);
299 const __m256i vacc3x4657 = _mm256_hadd_epi32(vacc3x45, vacc3x67);
300 const __m256i vacc4x0213 = _mm256_hadd_epi32(vacc4x01, vacc4x23);
301 const __m256i vacc4x4657 = _mm256_hadd_epi32(vacc4x45, vacc4x67);
302
303 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
304 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
305 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
306 const __m256i vacc3x02461357 = _mm256_hadd_epi32(vacc3x0213, vacc3x4657);
307 const __m256i vacc4x02461357 = _mm256_hadd_epi32(vacc4x0213, vacc4x4657);
308
309 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
310 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
311 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
312 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
313 __m256i vacc3x01234567 = _mm256_permutevar8x32_epi32(vacc3x02461357, vpermute_mask);
314 __m256i vacc4x01234567 = _mm256_permutevar8x32_epi32(vacc4x02461357, vpermute_mask);
315
316 vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4);
317 vacc1x01234567 = _mm256_srai_epi32(vacc1x01234567, 4);
318 vacc2x01234567 = _mm256_srai_epi32(vacc2x01234567, 4);
319 vacc3x01234567 = _mm256_srai_epi32(vacc3x01234567, 4);
320 vacc4x01234567 = _mm256_srai_epi32(vacc4x01234567, 4);
321 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
322 const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale);
323 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
324 const __m256 vinput_scale1 = _mm256_broadcast_ss(&quantization_params[1].inv_scale);
325 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
326 const __m256 vinput_scale2 = _mm256_broadcast_ss(&quantization_params[2].inv_scale);
327 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
328 const __m256 vinput_scale3 = _mm256_broadcast_ss(&quantization_params[3].inv_scale);
329 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
330 const __m256 vinput_scale4 = _mm256_broadcast_ss(&quantization_params[4].inv_scale);
331
332 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale0);
333 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_scale1);
334 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_scale2);
335 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_scale3);
336 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_scale4);
337
338 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
339 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
340 w = (const float*) w + 16;
341 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
342 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
343 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
344 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
345 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
346
347 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
348 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
349 vout2x01234567 = _mm256_max_ps(vout2x01234567, vmin);
350 vout3x01234567 = _mm256_max_ps(vout3x01234567, vmin);
351 vout4x01234567 = _mm256_max_ps(vout4x01234567, vmin);
352
353 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
354 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
355 vout2x01234567 = _mm256_min_ps(vout2x01234567, vmax);
356 vout3x01234567 = _mm256_min_ps(vout3x01234567, vmax);
357 vout4x01234567 = _mm256_min_ps(vout4x01234567, vmax);
358 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
359 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
360 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
361 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
362 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
363 if XNN_LIKELY(nc >= 8) {
364 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
365 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
366 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
367 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
368 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
369 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
370 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
371 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
372 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
373 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
374
375 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
376 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
377 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
378 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
379 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
380
381 nc -= 8;
382 } else {
383 if (nc & 4) {
384 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
385 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
386 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
387 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
388 _mm_storel_epi64((__m128i*) c4, vfp16out4x01234567);
389
390 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
391 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
392 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
393 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
394 vfp16out4x01234567 = _mm_unpackhi_epi64(vfp16out4x01234567, vfp16out4x01234567);
395
396 c0 += 4;
397 c1 += 4;
398 c2 += 4;
399 c3 += 4;
400 c4 += 4;
401 }
402 if (nc & 2) {
403 _mm_storeu_si32(c0, vfp16out0x01234567);
404 _mm_storeu_si32(c1, vfp16out1x01234567);
405 _mm_storeu_si32(c2, vfp16out2x01234567);
406 _mm_storeu_si32(c3, vfp16out3x01234567);
407 _mm_storeu_si32(c4, vfp16out4x01234567);
408
409 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
410 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
411 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
412 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
413 vfp16out4x01234567 = _mm_srli_epi64(vfp16out4x01234567, 32);
414
415 c0 += 2;
416 c1 += 2;
417 c2 += 2;
418 c3 += 2;
419 c4 += 2;
420 }
421 if (nc & 1) {
422 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
423 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
424 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
425 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
426 *c4 = (uint16_t) _mm_extract_epi16(vfp16out4x01234567, 0);
427 }
428 nc = 0;
429 }
430 } while (nc != 0);
431 }