Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-8x8c8-minmax-fp32-avxvnni-prfm.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 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_8x8c8__avxvnni_prfm(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 size_t ks,
27 const int8_t** restrict a,
28 const void* restrict w,
29 int8_t* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 8);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 int8_t* c0 = c;
47 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 c3 = c2;
58 }
59 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 c4 = c3;
62 }
63 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
64 if XNN_UNPREDICTABLE(mr < 6) {
65 c5 = c4;
66 }
67 int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
68 if XNN_UNPREDICTABLE(mr <= 6) {
69 c6 = c5;
70 }
71 int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride);
72 if XNN_UNPREDICTABLE(mr != 8) {
73 c7 = c6;
74 }
75
76 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
77 XNN_FORCE_REALIZATION(vsign_mask);
78 const __m256 voutput_max_less_zero_point = _mm256_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
79 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
80 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
81 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
82 // XNN_FORCE_REALIZATION(voutput_zero_point);
83 // XNN_FORCE_REALIZATION(voutput_min);
84 do {
85 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
86 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
87 __m256i vacc1x0123 = vacc0x0123;
88 __m256i vacc1x4567 = vacc0x4567;
89 __m256i vacc2x0123 = vacc0x0123;
90 __m256i vacc2x4567 = vacc0x4567;
91 __m256i vacc3x0123 = vacc0x0123;
92 __m256i vacc3x4567 = vacc0x4567;
93 __m256i vacc4x0123 = vacc0x0123;
94 __m256i vacc4x4567 = vacc0x4567;
95 __m256i vacc5x0123 = vacc0x0123;
96 __m256i vacc5x4567 = vacc0x4567;
97 __m256i vacc6x0123 = vacc0x0123;
98 __m256i vacc6x4567 = vacc0x4567;
99 __m256i vacc7x0123 = vacc0x0123;
100 __m256i vacc7x4567 = vacc0x4567;
101 w = (const int32_t*) w + 8;
102
103 size_t p = ks;
104 do {
105 const int8_t* restrict a0 = a[0];
106 if XNN_UNPREDICTABLE(a0 != zero) {
107 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
108 }
109 const int8_t* restrict a1 = a[1];
110 if XNN_UNPREDICTABLE(a1 != zero) {
111 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
112 }
113 const int8_t* restrict a2 = a[2];
114 if XNN_UNPREDICTABLE(a2 != zero) {
115 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
116 }
117 const int8_t* restrict a3 = a[3];
118 if XNN_UNPREDICTABLE(a3 != zero) {
119 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
120 }
121 const int8_t* restrict a4 = a[4];
122 if XNN_UNPREDICTABLE(a4 != zero) {
123 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
124 }
125 const int8_t* restrict a5 = a[5];
126 if XNN_UNPREDICTABLE(a5 != zero) {
127 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
128 }
129 const int8_t* restrict a6 = a[6];
130 if XNN_UNPREDICTABLE(a6 != zero) {
131 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
132 }
133 const int8_t* restrict a7 = a[7];
134 if XNN_UNPREDICTABLE(a7 != zero) {
135 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
136 }
137 a += 8;
138
139 size_t k = kc;
140 while (k >= 16 * sizeof(int8_t)) {
141 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
142 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
143 a0 += 16;
144 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
145 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
146 a1 += 16;
147 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
148 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
149 a2 += 16;
150 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
151 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
152 a3 += 16;
153 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
154 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
155 a4 += 16;
156 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
157 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
158 a5 += 16;
159 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
160 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
161 a6 += 16;
162 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
163 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
164 a7 += 16;
165
166 const __m256i vb01234567x0123 = _mm256_load_si256(w);
167 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
168 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
169 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
170 xnn_prefetch_to_l1((const int8_t*) w + 896);
171
172 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
173 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
174 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
175 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
176 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
177 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
178 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
179 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
180 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
181 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
182 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
183 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
184 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
185 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
186 vacc7x0123 = _mm256_dpbusd_avx_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
187 vacc7x4567 = _mm256_dpbusd_avx_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
188 xnn_prefetch_to_l1((const int8_t*) w + 960);
189 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
190 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
191 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
192 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
193 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
194 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
195 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
196 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
197 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
198 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
199 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
200 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
201 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
202 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
203 vacc7x0123 = _mm256_dpbusd_avx_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
204 vacc7x4567 = _mm256_dpbusd_avx_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
205
206 w = (const int8_t*) w + 128;
207 k -= 16 * sizeof(int8_t);
208 }
209
210 if (k != 0) {
211 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
212 a0 += 8;
213 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
214 a1 += 8;
215 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
216 a2 += 8;
217 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
218 a3 += 8;
219 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
220 a4 += 8;
221 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
222 a5 += 8;
223 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
224 a6 += 8;
225 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
226 a7 += 8;
227
228 const __m256i vb01234567x0123 = _mm256_load_si256(w);
229 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
230
231 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
232 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
233 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
234 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
235 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
236 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
237 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
238 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
239 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
240 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
241 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
242 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
243 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
244 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
245 vacc7x0123 = _mm256_dpbusd_avx_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
246 vacc7x4567 = _mm256_dpbusd_avx_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
247 xnn_prefetch_to_l1((const int8_t*) w + 960);
248
249 w = (const int8_t*) w + 64;
250 k -= 8 * sizeof(int8_t);
251 }
252
253 p -= 8 * sizeof(void*);
254 } while (p != 0);
255
256
257 // Add adjacent pairs
258 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
259 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
260 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
261 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
262 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
263 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
264 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
265 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
266 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
267 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
268 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
269 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
270 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
271 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
272 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
273 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
274
275 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
276 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
277 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
278 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
279 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
280 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
281 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
282 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
283
284 const __m256 vscale01234567 = _mm256_load_ps(w);
285 w = (const float*) w + 8;
286 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
287 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
288 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);
289 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);
290 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vscale01234567);
291 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vscale01234567);
292 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vscale01234567);
293 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vscale01234567);
294
295 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
296 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
297 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);
298 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);
299 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max_less_zero_point);
300 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max_less_zero_point);
301 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max_less_zero_point);
302 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max_less_zero_point);
303
304 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
305 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
306 vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);
307 vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);
308 vacc4x01234567 = _mm256_cvtps_epi32(vout4x01234567);
309 vacc5x01234567 = _mm256_cvtps_epi32(vout5x01234567);
310 vacc6x01234567 = _mm256_cvtps_epi32(vout6x01234567);
311 vacc7x01234567 = _mm256_cvtps_epi32(vout7x01234567);
312
313 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
314 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
315 vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);
316 vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);
317 vacc4x01234567 = _mm256_add_epi32(vacc4x01234567, voutput_zero_point);
318 vacc5x01234567 = _mm256_add_epi32(vacc5x01234567, voutput_zero_point);
319 vacc6x01234567 = _mm256_add_epi32(vacc6x01234567, voutput_zero_point);
320 vacc7x01234567 = _mm256_add_epi32(vacc7x01234567, voutput_zero_point);
321
322 vacc0x01234567 = _mm256_packs_epi32(vacc0x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc0x01234567, 1)));
323 __m128i voutb0x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc0x01234567, vacc0x01234567));
324 vacc1x01234567 = _mm256_packs_epi32(vacc1x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc1x01234567, 1)));
325 __m128i voutb1x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc1x01234567, vacc1x01234567));
326 vacc2x01234567 = _mm256_packs_epi32(vacc2x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc2x01234567, 1)));
327 __m128i voutb2x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc2x01234567, vacc2x01234567));
328 vacc3x01234567 = _mm256_packs_epi32(vacc3x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc3x01234567, 1)));
329 __m128i voutb3x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc3x01234567, vacc3x01234567));
330 vacc4x01234567 = _mm256_packs_epi32(vacc4x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc4x01234567, 1)));
331 __m128i voutb4x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc4x01234567, vacc4x01234567));
332 vacc5x01234567 = _mm256_packs_epi32(vacc5x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc5x01234567, 1)));
333 __m128i voutb5x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc5x01234567, vacc5x01234567));
334 vacc6x01234567 = _mm256_packs_epi32(vacc6x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc6x01234567, 1)));
335 __m128i voutb6x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc6x01234567, vacc6x01234567));
336 vacc7x01234567 = _mm256_packs_epi32(vacc7x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc7x01234567, 1)));
337 __m128i voutb7x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc7x01234567, vacc7x01234567));
338
339 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
340 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
341 voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);
342 voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);
343 voutb4x01234567 = _mm_max_epi8(voutb4x01234567, voutput_min);
344 voutb5x01234567 = _mm_max_epi8(voutb5x01234567, voutput_min);
345 voutb6x01234567 = _mm_max_epi8(voutb6x01234567, voutput_min);
346 voutb7x01234567 = _mm_max_epi8(voutb7x01234567, voutput_min);
347
348 if (nc >= 8) {
349 _mm_storel_epi64((__m128i*) c7, voutb7x01234567);
350 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
351 _mm_storel_epi64((__m128i*) c6, voutb6x01234567);
352 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
353 _mm_storel_epi64((__m128i*) c5, voutb5x01234567);
354 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
355 _mm_storel_epi64((__m128i*) c4, voutb4x01234567);
356 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
357 _mm_storel_epi64((__m128i*) c3, voutb3x01234567);
358 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
359 _mm_storel_epi64((__m128i*) c2, voutb2x01234567);
360 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
361 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
362 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
363 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
364 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
365 a = (const int8_t**restrict) ((uintptr_t) a - ks);
366 nc -= 8;
367 } else {
368 if (nc & 4) {
369 _mm_storeu_si32(c7, voutb7x01234567);
370 c7 += 4;
371 _mm_storeu_si32(c6, voutb6x01234567);
372 c6 += 4;
373 _mm_storeu_si32(c5, voutb5x01234567);
374 c5 += 4;
375 _mm_storeu_si32(c4, voutb4x01234567);
376 c4 += 4;
377 _mm_storeu_si32(c3, voutb3x01234567);
378 c3 += 4;
379 _mm_storeu_si32(c2, voutb2x01234567);
380 c2 += 4;
381 _mm_storeu_si32(c1, voutb1x01234567);
382 c1 += 4;
383 _mm_storeu_si32(c0, voutb0x01234567);
384 c0 += 4;
385 voutb7x01234567 = _mm_srli_epi64(voutb7x01234567, 32);
386 voutb6x01234567 = _mm_srli_epi64(voutb6x01234567, 32);
387 voutb5x01234567 = _mm_srli_epi64(voutb5x01234567, 32);
388 voutb4x01234567 = _mm_srli_epi64(voutb4x01234567, 32);
389 voutb3x01234567 = _mm_srli_epi64(voutb3x01234567, 32);
390 voutb2x01234567 = _mm_srli_epi64(voutb2x01234567, 32);
391 voutb1x01234567 = _mm_srli_epi64(voutb1x01234567, 32);
392 voutb0x01234567 = _mm_srli_epi64(voutb0x01234567, 32);
393 }
394 if (nc & 2) {
395 unaligned_store_u16(c7, (uint16_t) _mm_extract_epi16(voutb7x01234567, 0));
396 c7 += 2;
397 unaligned_store_u16(c6, (uint16_t) _mm_extract_epi16(voutb6x01234567, 0));
398 c6 += 2;
399 unaligned_store_u16(c5, (uint16_t) _mm_extract_epi16(voutb5x01234567, 0));
400 c5 += 2;
401 unaligned_store_u16(c4, (uint16_t) _mm_extract_epi16(voutb4x01234567, 0));
402 c4 += 2;
403 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(voutb3x01234567, 0));
404 c3 += 2;
405 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(voutb2x01234567, 0));
406 c2 += 2;
407 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(voutb1x01234567, 0));
408 c1 += 2;
409 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(voutb0x01234567, 0));
410 c0 += 2;
411 voutb7x01234567 = _mm_srli_epi32(voutb7x01234567, 16);
412 voutb6x01234567 = _mm_srli_epi32(voutb6x01234567, 16);
413 voutb5x01234567 = _mm_srli_epi32(voutb5x01234567, 16);
414 voutb4x01234567 = _mm_srli_epi32(voutb4x01234567, 16);
415 voutb3x01234567 = _mm_srli_epi32(voutb3x01234567, 16);
416 voutb2x01234567 = _mm_srli_epi32(voutb2x01234567, 16);
417 voutb1x01234567 = _mm_srli_epi32(voutb1x01234567, 16);
418 voutb0x01234567 = _mm_srli_epi32(voutb0x01234567, 16);
419 }
420 if (nc & 1) {
421 *c7 = (int8_t) _mm_extract_epi8(voutb7x01234567, 0);
422 *c6 = (int8_t) _mm_extract_epi8(voutb6x01234567, 0);
423 *c5 = (int8_t) _mm_extract_epi8(voutb5x01234567, 0);
424 *c4 = (int8_t) _mm_extract_epi8(voutb4x01234567, 0);
425 *c3 = (int8_t) _mm_extract_epi8(voutb3x01234567, 0);
426 *c2 = (int8_t) _mm_extract_epi8(voutb2x01234567, 0);
427 *c1 = (int8_t) _mm_extract_epi8(voutb1x01234567, 0);
428 *c0 = (int8_t) _mm_extract_epi8(voutb0x01234567, 0);
429 }
430 nc = 0;
431 }
432 } while (nc != 0);
433 }