Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-8x8c8-minmax-avxvnni.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx8c8-avxvnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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_qc8w_igemm_minmax_ukernel_8x8c8__avxvnni(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 xnn_float16* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const int8_t* zero_data,
34 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
35 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 8);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 8 * sizeof(int8_t));
47 uint16_t* c0 = (uint16_t*) c;
48 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 c1 = c0;
51 }
52 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 c2 = c1;
55 }
56 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 c3 = c2;
59 }
60 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
61 if XNN_UNPREDICTABLE(mr <= 4) {
62 c4 = c3;
63 }
64 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
65 if XNN_UNPREDICTABLE(mr < 6) {
66 c5 = c4;
67 }
68 uint16_t* c6 = (uint16_t*) ((uintptr_t) c5 + cm_stride);
69 if XNN_UNPREDICTABLE(mr <= 6) {
70 c6 = c5;
71 }
72 uint16_t* c7 = (uint16_t*) ((uintptr_t) c6 + cm_stride);
73 if XNN_UNPREDICTABLE(mr != 8) {
74 c7 = c6;
75 }
76
77 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
78 XNN_FORCE_REALIZATION(vsign_mask);
79 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
80 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
81 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
82 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
83 // XNN_FORCE_REALIZATION(vinput_zero_point);
84 // XNN_FORCE_REALIZATION(vinput_inv_scale);
85 // XNN_FORCE_REALIZATION(voutput_min);
86 // XNN_FORCE_REALIZATION(voutput_max);
87 do {
88 const __m256i vksum01234567 = _mm256_load_si256(w);
89 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
90 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
91 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
92 __m256i vacc1x0123 = vacc0x0123;
93 __m256i vacc1x4567 = vacc0x4567;
94 __m256i vacc2x0123 = vacc0x0123;
95 __m256i vacc2x4567 = vacc0x4567;
96 __m256i vacc3x0123 = vacc0x0123;
97 __m256i vacc3x4567 = vacc0x4567;
98 __m256i vacc4x0123 = vacc0x0123;
99 __m256i vacc4x4567 = vacc0x4567;
100 __m256i vacc5x0123 = vacc0x0123;
101 __m256i vacc5x4567 = vacc0x4567;
102 __m256i vacc6x0123 = vacc0x0123;
103 __m256i vacc6x4567 = vacc0x4567;
104 __m256i vacc7x0123 = vacc0x0123;
105 __m256i vacc7x4567 = vacc0x4567;
106 w = (const int32_t*) w + 8;
107
108 size_t p = ks;
109 do {
110 const int8_t* restrict a0 = a[0];
111 if XNN_UNPREDICTABLE(a0 != zero) {
112 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
113 } else {
114 a0 = zero_data;
115 }
116 const int8_t* restrict a1 = a[1];
117 if XNN_UNPREDICTABLE(a1 != zero) {
118 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
119 } else {
120 a1 = zero_data;
121 }
122 const int8_t* restrict a2 = a[2];
123 if XNN_UNPREDICTABLE(a2 != zero) {
124 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
125 } else {
126 a2 = zero_data;
127 }
128 const int8_t* restrict a3 = a[3];
129 if XNN_UNPREDICTABLE(a3 != zero) {
130 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
131 } else {
132 a3 = zero_data;
133 }
134 const int8_t* restrict a4 = a[4];
135 if XNN_UNPREDICTABLE(a4 != zero) {
136 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
137 } else {
138 a4 = zero_data;
139 }
140 const int8_t* restrict a5 = a[5];
141 if XNN_UNPREDICTABLE(a5 != zero) {
142 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
143 } else {
144 a5 = zero_data;
145 }
146 const int8_t* restrict a6 = a[6];
147 if XNN_UNPREDICTABLE(a6 != zero) {
148 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
149 } else {
150 a6 = zero_data;
151 }
152 const int8_t* restrict a7 = a[7];
153 if XNN_UNPREDICTABLE(a7 != zero) {
154 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
155 } else {
156 a7 = zero_data;
157 }
158 a += 8;
159
160 size_t k = kc;
161 while (k >= 16 * sizeof(int8_t)) {
162 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
163 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
164 a0 += 16;
165 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
166 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
167 a1 += 16;
168 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
169 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
170 a2 += 16;
171 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
172 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
173 a3 += 16;
174 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
175 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
176 a4 += 16;
177 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
178 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
179 a5 += 16;
180 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
181 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
182 a6 += 16;
183 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
184 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
185 a7 += 16;
186
187 const __m256i vb01234567x0123 = _mm256_load_si256(w);
188 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
189 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
190 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
191
192 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
193 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
194 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
195 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
196 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
197 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
198 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
199 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
200 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
201 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
202 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
203 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
204 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
205 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
206 vacc7x0123 = _mm256_dpbusd_avx_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
207 vacc7x4567 = _mm256_dpbusd_avx_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
208 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
209 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
210 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
211 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
212 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
213 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
214 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
215 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
216 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
217 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
218 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
219 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
220 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
221 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
222 vacc7x0123 = _mm256_dpbusd_avx_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
223 vacc7x4567 = _mm256_dpbusd_avx_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
224
225 w = (const int8_t*) w + 128;
226 k -= 16 * sizeof(int8_t);
227 }
228
229 if (k != 0) {
230 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
231 a0 += 8;
232 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
233 a1 += 8;
234 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
235 a2 += 8;
236 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
237 a3 += 8;
238 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
239 a4 += 8;
240 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
241 a5 += 8;
242 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
243 a6 += 8;
244 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
245 a7 += 8;
246
247 const __m256i vb01234567x0123 = _mm256_load_si256(w);
248 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
249
250 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
251 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
252 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
253 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
254 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
255 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
256 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
257 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
258 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
259 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
260 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
261 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
262 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
263 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
264 vacc7x0123 = _mm256_dpbusd_avx_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
265 vacc7x4567 = _mm256_dpbusd_avx_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
266
267 w = (const int8_t*) w + 64;
268 k -= 8 * sizeof(int8_t);
269 }
270
271 p -= 8 * sizeof(void*);
272 } while (p != 0);
273
274
275 // Add adjacent pairs
276 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
277 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
278 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
279 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
280 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
281 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
282 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
283 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
284 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
285 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
286 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
287 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
288 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
289 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
290 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
291 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
292
293 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
294 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
295 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
296 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
297 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
298 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
299 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
300 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
301
302 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
303 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
304 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
305 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_inv_scale);
306 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_inv_scale);
307 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_inv_scale);
308 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_inv_scale);
309 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vinput_inv_scale);
310
311 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
312 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
313 w = (const float*) w + 16;
314 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
315 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
316 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
317 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
318 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
319 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
320 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
321 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
322
323 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
324 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
325 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
326 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
327 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
328 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
329 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
330 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
331
332 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
333 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
334 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
335 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
336 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
337 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
338 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
339 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
340
341 __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT);
342 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
343 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
344 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
345 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
346 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
347 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
348 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
349 if XNN_LIKELY(nc >= 8) {
350 _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567);
351 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
352 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
353 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
354 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
355 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
356 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
357 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
358 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
359 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
360 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
361 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
362 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
363 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
364 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
365 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
366
367 a = (const int8_t**restrict) ((uintptr_t) a - ks);
368 nc -= 8;
369 } else {
370 if (nc & 4) {
371 _mm_storel_epi64((__m128i*) c7, vfp16out7x01234567);
372 c7 += 4;
373 vfp16out7x01234567 = _mm_unpackhi_epi64(vfp16out7x01234567, vfp16out7x01234567);
374 _mm_storel_epi64((__m128i*) c6, vfp16out6x01234567);
375 c6 += 4;
376 vfp16out6x01234567 = _mm_unpackhi_epi64(vfp16out6x01234567, vfp16out6x01234567);
377 _mm_storel_epi64((__m128i*) c5, vfp16out5x01234567);
378 c5 += 4;
379 vfp16out5x01234567 = _mm_unpackhi_epi64(vfp16out5x01234567, vfp16out5x01234567);
380 _mm_storel_epi64((__m128i*) c4, vfp16out4x01234567);
381 c4 += 4;
382 vfp16out4x01234567 = _mm_unpackhi_epi64(vfp16out4x01234567, vfp16out4x01234567);
383 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
384 c3 += 4;
385 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
386 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
387 c2 += 4;
388 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
389 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
390 c1 += 4;
391 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
392 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
393 c0 += 4;
394 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
395 }
396 if (nc & 2) {
397 _mm_storeu_si32(c7, vfp16out7x01234567);
398 vfp16out7x01234567 = _mm_srli_epi64(vfp16out7x01234567, 32);
399 c7 += 2;
400 _mm_storeu_si32(c6, vfp16out6x01234567);
401 vfp16out6x01234567 = _mm_srli_epi64(vfp16out6x01234567, 32);
402 c6 += 2;
403 _mm_storeu_si32(c5, vfp16out5x01234567);
404 vfp16out5x01234567 = _mm_srli_epi64(vfp16out5x01234567, 32);
405 c5 += 2;
406 _mm_storeu_si32(c4, vfp16out4x01234567);
407 vfp16out4x01234567 = _mm_srli_epi64(vfp16out4x01234567, 32);
408 c4 += 2;
409 _mm_storeu_si32(c3, vfp16out3x01234567);
410 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
411 c3 += 2;
412 _mm_storeu_si32(c2, vfp16out2x01234567);
413 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
414 c2 += 2;
415 _mm_storeu_si32(c1, vfp16out1x01234567);
416 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
417 c1 += 2;
418 _mm_storeu_si32(c0, vfp16out0x01234567);
419 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
420 c0 += 2;
421 }
422 if (nc & 1) {
423 *c7 = (uint16_t) _mm_extract_epi16(vfp16out7x01234567, 0);
424 *c6 = (uint16_t) _mm_extract_epi16(vfp16out6x01234567, 0);
425 *c5 = (uint16_t) _mm_extract_epi16(vfp16out5x01234567, 0);
426 *c4 = (uint16_t) _mm_extract_epi16(vfp16out4x01234567, 0);
427 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
428 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
429 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
430 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
431 }
432 nc = 0;
433 }
434 } while (nc != 0);
435 }