Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-6x8c8-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_6x8c8__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 <= 6);
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
77 const __m128i vmask = _mm_set1_epi8(0xF0);
78 XNN_FORCE_REALIZATION(vmask);
79 const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
80 const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
81 XNN_FORCE_REALIZATION(vmin);
82 XNN_FORCE_REALIZATION(vmax);
83
84 do {
85 const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
86 const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
87 const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
88 const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
89 const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
90 const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
91 const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
92 const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
93 const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
94 const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
95 const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
96 const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
97 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point);
98 __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point0);
99 __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point0);
100 __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point0);
101 __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point0);
102 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point);
103 __m256i vacc1x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point1);
104 __m256i vacc1x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point1);
105 __m256i vacc1x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point1);
106 __m256i vacc1x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point1);
107 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point);
108 __m256i vacc2x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point2);
109 __m256i vacc2x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point2);
110 __m256i vacc2x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point2);
111 __m256i vacc2x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point2);
112 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point);
113 __m256i vacc3x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point3);
114 __m256i vacc3x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point3);
115 __m256i vacc3x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point3);
116 __m256i vacc3x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point3);
117 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point);
118 __m256i vacc4x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point4);
119 __m256i vacc4x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point4);
120 __m256i vacc4x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point4);
121 __m256i vacc4x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point4);
122 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point);
123 __m256i vacc5x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point5);
124 __m256i vacc5x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point5);
125 __m256i vacc5x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point5);
126 __m256i vacc5x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point5);
127 w = (const int32_t*) w + 8;
128
129 size_t k = kc;
130 while (k >= 16 * sizeof(int8_t)) {
131 __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
132 __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
133 a0 += 8;
134 __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
135 __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
136 a1 += 8;
137 __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
138 __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
139 a2 += 8;
140 __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
141 __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
142 a3 += 8;
143 __m128i va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
144 __m256i vxa4 = _mm256_cvtepi8_epi16(va4);
145 a4 += 8;
146 __m128i va5 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a5));
147 __m256i vxa5 = _mm256_cvtepi8_epi16(va5);
148 a5 += 8;
149
150 __m128i vb01 = _mm_load_si128((const __m128i*) w);
151 __m128i vbs01 = _mm_slli_epi32(vb01, 4);
152 __m128i vbm01 = _mm_and_si128(vbs01, vmask);
153 __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
154
155 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
156 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
157 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
158 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
159 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
160 vacc5x01 = _mm256_add_epi32(vacc5x01, _mm256_madd_epi16(vxa5, vxb01));
161 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
162 __m128i vbs23 = _mm_slli_epi32(vb23, 4);
163 __m128i vbm23 = _mm_and_si128(vbs23, vmask);
164 __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
165
166 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
167 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
168 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
169 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
170 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
171 vacc5x23 = _mm256_add_epi32(vacc5x23, _mm256_madd_epi16(vxa5, vxb23));
172 __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
173 __m128i vbs45 = _mm_slli_epi32(vb45, 4);
174 __m128i vbm45 = _mm_and_si128(vbs45, vmask);
175 __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
176
177 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
178 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
179 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
180 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
181 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
182 vacc5x45 = _mm256_add_epi32(vacc5x45, _mm256_madd_epi16(vxa5, vxb45));
183 __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
184 __m128i vbs67 = _mm_slli_epi32(vb67, 4);
185 __m128i vbm67 = _mm_and_si128(vbs67, vmask);
186 __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
187
188 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
189 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
190 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
191 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
192 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
193 vacc5x67 = _mm256_add_epi32(vacc5x67, _mm256_madd_epi16(vxa5, vxb67));
194
195 va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
196 vxa0 = _mm256_cvtepi8_epi16(va0);
197 a0 += 8;
198 va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
199 vxa1 = _mm256_cvtepi8_epi16(va1);
200 a1 += 8;
201 va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
202 vxa2 = _mm256_cvtepi8_epi16(va2);
203 a2 += 8;
204 va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
205 vxa3 = _mm256_cvtepi8_epi16(va3);
206 a3 += 8;
207 va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
208 vxa4 = _mm256_cvtepi8_epi16(va4);
209 a4 += 8;
210 va5 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a5));
211 vxa5 = _mm256_cvtepi8_epi16(va5);
212 a5 += 8;
213
214 vbm01 = _mm_and_si128(vb01, vmask);
215 vxb01 = _mm256_cvtepi8_epi16(vbm01);
216
217 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
218 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
219 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
220 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
221 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
222 vacc5x01 = _mm256_add_epi32(vacc5x01, _mm256_madd_epi16(vxa5, vxb01));
223 vbm23 = _mm_and_si128(vb23, vmask);
224 vxb23 = _mm256_cvtepi8_epi16(vbm23);
225
226 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
227 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
228 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
229 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
230 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
231 vacc5x23 = _mm256_add_epi32(vacc5x23, _mm256_madd_epi16(vxa5, vxb23));
232 vbm45 = _mm_and_si128(vb45, vmask);
233 vxb45 = _mm256_cvtepi8_epi16(vbm45);
234
235 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
236 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
237 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
238 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
239 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
240 vacc5x45 = _mm256_add_epi32(vacc5x45, _mm256_madd_epi16(vxa5, vxb45));
241 vbm67 = _mm_and_si128(vb67, vmask);
242 vxb67 = _mm256_cvtepi8_epi16(vbm67);
243
244 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
245 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
246 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
247 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
248 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
249 vacc5x67 = _mm256_add_epi32(vacc5x67, _mm256_madd_epi16(vxa5, vxb67));
250
251 w = (const int8_t*) w + 64;
252 k -= 16 * sizeof(int8_t);
253 }
254
255 while (k >= 8 * sizeof(int8_t)) {
256 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
257 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
258 a0 += 8;
259 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
260 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
261 a1 += 8;
262 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
263 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
264 a2 += 8;
265 const __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
266 const __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
267 a3 += 8;
268 const __m128i va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
269 const __m256i vxa4 = _mm256_cvtepi8_epi16(va4);
270 a4 += 8;
271 const __m128i va5 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a5));
272 const __m256i vxa5 = _mm256_cvtepi8_epi16(va5);
273 a5 += 8;
274
275 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
276 const __m128i vbs01 = _mm_slli_epi32(vb01, 4);
277 const __m128i vbm01 = _mm_and_si128(vbs01, vmask);
278 const __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
279
280 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
281 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
282 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
283 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
284 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
285 vacc5x01 = _mm256_add_epi32(vacc5x01, _mm256_madd_epi16(vxa5, vxb01));
286 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
287 const __m128i vbs23 = _mm_slli_epi32(vb23, 4);
288 const __m128i vbm23 = _mm_and_si128(vbs23, vmask);
289 const __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
290
291 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
292 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
293 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
294 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
295 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
296 vacc5x23 = _mm256_add_epi32(vacc5x23, _mm256_madd_epi16(vxa5, vxb23));
297 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
298 const __m128i vbs45 = _mm_slli_epi32(vb45, 4);
299 const __m128i vbm45 = _mm_and_si128(vbs45, vmask);
300 const __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
301
302 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
303 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
304 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
305 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
306 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
307 vacc5x45 = _mm256_add_epi32(vacc5x45, _mm256_madd_epi16(vxa5, vxb45));
308 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
309 const __m128i vbs67 = _mm_slli_epi32(vb67, 4);
310 const __m128i vbm67 = _mm_and_si128(vbs67, vmask);
311 const __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
312
313 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
314 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
315 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
316 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
317 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
318 vacc5x67 = _mm256_add_epi32(vacc5x67, _mm256_madd_epi16(vxa5, vxb67));
319
320 w = (const int8_t*) w + 64;
321 k -= 8 * sizeof(int8_t);
322 }
323
324 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
325 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
326 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
327 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
328 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
329 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
330 const __m256i vacc3x0213 = _mm256_hadd_epi32(vacc3x01, vacc3x23);
331 const __m256i vacc3x4657 = _mm256_hadd_epi32(vacc3x45, vacc3x67);
332 const __m256i vacc4x0213 = _mm256_hadd_epi32(vacc4x01, vacc4x23);
333 const __m256i vacc4x4657 = _mm256_hadd_epi32(vacc4x45, vacc4x67);
334 const __m256i vacc5x0213 = _mm256_hadd_epi32(vacc5x01, vacc5x23);
335 const __m256i vacc5x4657 = _mm256_hadd_epi32(vacc5x45, vacc5x67);
336
337 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
338 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
339 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
340 const __m256i vacc3x02461357 = _mm256_hadd_epi32(vacc3x0213, vacc3x4657);
341 const __m256i vacc4x02461357 = _mm256_hadd_epi32(vacc4x0213, vacc4x4657);
342 const __m256i vacc5x02461357 = _mm256_hadd_epi32(vacc5x0213, vacc5x4657);
343
344 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
345 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
346 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
347 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
348 __m256i vacc3x01234567 = _mm256_permutevar8x32_epi32(vacc3x02461357, vpermute_mask);
349 __m256i vacc4x01234567 = _mm256_permutevar8x32_epi32(vacc4x02461357, vpermute_mask);
350 __m256i vacc5x01234567 = _mm256_permutevar8x32_epi32(vacc5x02461357, vpermute_mask);
351
352 vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4);
353 vacc1x01234567 = _mm256_srai_epi32(vacc1x01234567, 4);
354 vacc2x01234567 = _mm256_srai_epi32(vacc2x01234567, 4);
355 vacc3x01234567 = _mm256_srai_epi32(vacc3x01234567, 4);
356 vacc4x01234567 = _mm256_srai_epi32(vacc4x01234567, 4);
357 vacc5x01234567 = _mm256_srai_epi32(vacc5x01234567, 4);
358 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
359 const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale);
360 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
361 const __m256 vinput_scale1 = _mm256_broadcast_ss(&quantization_params[1].inv_scale);
362 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
363 const __m256 vinput_scale2 = _mm256_broadcast_ss(&quantization_params[2].inv_scale);
364 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
365 const __m256 vinput_scale3 = _mm256_broadcast_ss(&quantization_params[3].inv_scale);
366 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
367 const __m256 vinput_scale4 = _mm256_broadcast_ss(&quantization_params[4].inv_scale);
368 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
369 const __m256 vinput_scale5 = _mm256_broadcast_ss(&quantization_params[5].inv_scale);
370
371 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale0);
372 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_scale1);
373 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_scale2);
374 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_scale3);
375 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_scale4);
376 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_scale5);
377
378 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
379 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
380 w = (const float*) w + 16;
381 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
382 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
383 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
384 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
385 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
386 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
387
388 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
389 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
390 vout2x01234567 = _mm256_max_ps(vout2x01234567, vmin);
391 vout3x01234567 = _mm256_max_ps(vout3x01234567, vmin);
392 vout4x01234567 = _mm256_max_ps(vout4x01234567, vmin);
393 vout5x01234567 = _mm256_max_ps(vout5x01234567, vmin);
394
395 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
396 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
397 vout2x01234567 = _mm256_min_ps(vout2x01234567, vmax);
398 vout3x01234567 = _mm256_min_ps(vout3x01234567, vmax);
399 vout4x01234567 = _mm256_min_ps(vout4x01234567, vmax);
400 vout5x01234567 = _mm256_min_ps(vout5x01234567, vmax);
401 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
402 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
403 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
404 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
405 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
406 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
407 if XNN_LIKELY(nc >= 8) {
408 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
409 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
410 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
411 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
412 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
413 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
414 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
415 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
416 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
417 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
418 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
419 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
420
421 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
422 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
423 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
424 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
425 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
426 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
427
428 nc -= 8;
429 } else {
430 if (nc & 4) {
431 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
432 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
433 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
434 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
435 _mm_storel_epi64((__m128i*) c4, vfp16out4x01234567);
436 _mm_storel_epi64((__m128i*) c5, vfp16out5x01234567);
437
438 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
439 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
440 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
441 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
442 vfp16out4x01234567 = _mm_unpackhi_epi64(vfp16out4x01234567, vfp16out4x01234567);
443 vfp16out5x01234567 = _mm_unpackhi_epi64(vfp16out5x01234567, vfp16out5x01234567);
444
445 c0 += 4;
446 c1 += 4;
447 c2 += 4;
448 c3 += 4;
449 c4 += 4;
450 c5 += 4;
451 }
452 if (nc & 2) {
453 _mm_storeu_si32(c0, vfp16out0x01234567);
454 _mm_storeu_si32(c1, vfp16out1x01234567);
455 _mm_storeu_si32(c2, vfp16out2x01234567);
456 _mm_storeu_si32(c3, vfp16out3x01234567);
457 _mm_storeu_si32(c4, vfp16out4x01234567);
458 _mm_storeu_si32(c5, vfp16out5x01234567);
459
460 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
461 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
462 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
463 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
464 vfp16out4x01234567 = _mm_srli_epi64(vfp16out4x01234567, 32);
465 vfp16out5x01234567 = _mm_srli_epi64(vfp16out5x01234567, 32);
466
467 c0 += 2;
468 c1 += 2;
469 c2 += 2;
470 c3 += 2;
471 c4 += 2;
472 c5 += 2;
473 }
474 if (nc & 1) {
475 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
476 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
477 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
478 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
479 *c4 = (uint16_t) _mm_extract_epi16(vfp16out4x01234567, 0);
480 *c5 = (uint16_t) _mm_extract_epi16(vfp16out5x01234567, 0);
481 }
482 nc = 0;
483 }
484 } while (nc != 0);
485 }