Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-8x8c8-minmax-fp32-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_qs8_qc8w_igemm_minmax_fp32_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 int8_t* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 8);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr < 4) {
56 c3 = c2;
57 }
58 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
59 if XNN_UNPREDICTABLE(mr <= 4) {
60 c4 = c3;
61 }
62 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
63 if XNN_UNPREDICTABLE(mr < 6) {
64 c5 = c4;
65 }
66 int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
67 if XNN_UNPREDICTABLE(mr <= 6) {
68 c6 = c5;
69 }
70 int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride);
71 if XNN_UNPREDICTABLE(mr != 8) {
72 c7 = c6;
73 }
74
75 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
76 XNN_FORCE_REALIZATION(vsign_mask);
77 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);
78 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
79 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
80 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
81 // XNN_FORCE_REALIZATION(voutput_zero_point);
82 // XNN_FORCE_REALIZATION(voutput_min);
83 do {
84 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
85 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
86 __m256i vacc1x0123 = vacc0x0123;
87 __m256i vacc1x4567 = vacc0x4567;
88 __m256i vacc2x0123 = vacc0x0123;
89 __m256i vacc2x4567 = vacc0x4567;
90 __m256i vacc3x0123 = vacc0x0123;
91 __m256i vacc3x4567 = vacc0x4567;
92 __m256i vacc4x0123 = vacc0x0123;
93 __m256i vacc4x4567 = vacc0x4567;
94 __m256i vacc5x0123 = vacc0x0123;
95 __m256i vacc5x4567 = vacc0x4567;
96 __m256i vacc6x0123 = vacc0x0123;
97 __m256i vacc6x4567 = vacc0x4567;
98 __m256i vacc7x0123 = vacc0x0123;
99 __m256i vacc7x4567 = vacc0x4567;
100 w = (const int32_t*) w + 8;
101
102 size_t p = ks;
103 do {
104 const int8_t* restrict a0 = a[0];
105 if XNN_UNPREDICTABLE(a0 != zero) {
106 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
107 }
108 const int8_t* restrict a1 = a[1];
109 if XNN_UNPREDICTABLE(a1 != zero) {
110 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
111 }
112 const int8_t* restrict a2 = a[2];
113 if XNN_UNPREDICTABLE(a2 != zero) {
114 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
115 }
116 const int8_t* restrict a3 = a[3];
117 if XNN_UNPREDICTABLE(a3 != zero) {
118 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
119 }
120 const int8_t* restrict a4 = a[4];
121 if XNN_UNPREDICTABLE(a4 != zero) {
122 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
123 }
124 const int8_t* restrict a5 = a[5];
125 if XNN_UNPREDICTABLE(a5 != zero) {
126 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
127 }
128 const int8_t* restrict a6 = a[6];
129 if XNN_UNPREDICTABLE(a6 != zero) {
130 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
131 }
132 const int8_t* restrict a7 = a[7];
133 if XNN_UNPREDICTABLE(a7 != zero) {
134 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
135 }
136 a += 8;
137
138 size_t k = kc;
139 while (k >= 16 * sizeof(int8_t)) {
140 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
141 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
142 a0 += 16;
143 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
144 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
145 a1 += 16;
146 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
147 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
148 a2 += 16;
149 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
150 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
151 a3 += 16;
152 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
153 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
154 a4 += 16;
155 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
156 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
157 a5 += 16;
158 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
159 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
160 a6 += 16;
161 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
162 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
163 a7 += 16;
164
165 const __m256i vb01234567x0123 = _mm256_load_si256(w);
166 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
167 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
168 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
169
170 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
171 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
172 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
173 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
174 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
175 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
176 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
177 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
178 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
179 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
180 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
181 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
182 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
183 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
184 vacc7x0123 = _mm256_dpbusd_avx_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
185 vacc7x4567 = _mm256_dpbusd_avx_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
186 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
187 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
188 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
189 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
190 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
191 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
192 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
193 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
194 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
195 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
196 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
197 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
198 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
199 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
200 vacc7x0123 = _mm256_dpbusd_avx_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
201 vacc7x4567 = _mm256_dpbusd_avx_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
202
203 w = (const int8_t*) w + 128;
204 k -= 16 * sizeof(int8_t);
205 }
206
207 if (k != 0) {
208 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
209 a0 += 8;
210 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
211 a1 += 8;
212 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
213 a2 += 8;
214 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
215 a3 += 8;
216 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
217 a4 += 8;
218 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
219 a5 += 8;
220 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
221 a6 += 8;
222 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
223 a7 += 8;
224
225 const __m256i vb01234567x0123 = _mm256_load_si256(w);
226 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
227
228 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
229 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
230 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
231 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
232 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
233 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
234 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
235 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
236 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
237 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
238 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
239 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
240 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
241 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
242 vacc7x0123 = _mm256_dpbusd_avx_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
243 vacc7x4567 = _mm256_dpbusd_avx_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
244
245 w = (const int8_t*) w + 64;
246 k -= 8 * sizeof(int8_t);
247 }
248
249 p -= 8 * sizeof(void*);
250 } while (p != 0);
251
252
253 // Add adjacent pairs
254 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
255 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
256 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
257 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
258 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
259 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
260 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
261 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
262 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
263 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
264 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
265 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
266 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
267 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
268 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
269 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
270
271 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
272 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
273 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
274 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
275 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
276 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
277 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
278 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
279
280 const __m256 vscale01234567 = _mm256_load_ps(w);
281 w = (const float*) w + 8;
282 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
283 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
284 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);
285 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);
286 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vscale01234567);
287 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vscale01234567);
288 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vscale01234567);
289 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vscale01234567);
290
291 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
292 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
293 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);
294 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);
295 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max_less_zero_point);
296 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max_less_zero_point);
297 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max_less_zero_point);
298 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max_less_zero_point);
299
300 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
301 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
302 vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);
303 vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);
304 vacc4x01234567 = _mm256_cvtps_epi32(vout4x01234567);
305 vacc5x01234567 = _mm256_cvtps_epi32(vout5x01234567);
306 vacc6x01234567 = _mm256_cvtps_epi32(vout6x01234567);
307 vacc7x01234567 = _mm256_cvtps_epi32(vout7x01234567);
308
309 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
310 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
311 vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);
312 vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);
313 vacc4x01234567 = _mm256_add_epi32(vacc4x01234567, voutput_zero_point);
314 vacc5x01234567 = _mm256_add_epi32(vacc5x01234567, voutput_zero_point);
315 vacc6x01234567 = _mm256_add_epi32(vacc6x01234567, voutput_zero_point);
316 vacc7x01234567 = _mm256_add_epi32(vacc7x01234567, voutput_zero_point);
317
318 vacc0x01234567 = _mm256_packs_epi32(vacc0x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc0x01234567, 1)));
319 __m128i voutb0x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc0x01234567, vacc0x01234567));
320 vacc1x01234567 = _mm256_packs_epi32(vacc1x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc1x01234567, 1)));
321 __m128i voutb1x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc1x01234567, vacc1x01234567));
322 vacc2x01234567 = _mm256_packs_epi32(vacc2x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc2x01234567, 1)));
323 __m128i voutb2x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc2x01234567, vacc2x01234567));
324 vacc3x01234567 = _mm256_packs_epi32(vacc3x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc3x01234567, 1)));
325 __m128i voutb3x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc3x01234567, vacc3x01234567));
326 vacc4x01234567 = _mm256_packs_epi32(vacc4x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc4x01234567, 1)));
327 __m128i voutb4x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc4x01234567, vacc4x01234567));
328 vacc5x01234567 = _mm256_packs_epi32(vacc5x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc5x01234567, 1)));
329 __m128i voutb5x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc5x01234567, vacc5x01234567));
330 vacc6x01234567 = _mm256_packs_epi32(vacc6x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc6x01234567, 1)));
331 __m128i voutb6x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc6x01234567, vacc6x01234567));
332 vacc7x01234567 = _mm256_packs_epi32(vacc7x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc7x01234567, 1)));
333 __m128i voutb7x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc7x01234567, vacc7x01234567));
334
335 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
336 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
337 voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);
338 voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);
339 voutb4x01234567 = _mm_max_epi8(voutb4x01234567, voutput_min);
340 voutb5x01234567 = _mm_max_epi8(voutb5x01234567, voutput_min);
341 voutb6x01234567 = _mm_max_epi8(voutb6x01234567, voutput_min);
342 voutb7x01234567 = _mm_max_epi8(voutb7x01234567, voutput_min);
343
344 if (nc >= 8) {
345 _mm_storel_epi64((__m128i*) c7, voutb7x01234567);
346 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
347 _mm_storel_epi64((__m128i*) c6, voutb6x01234567);
348 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
349 _mm_storel_epi64((__m128i*) c5, voutb5x01234567);
350 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
351 _mm_storel_epi64((__m128i*) c4, voutb4x01234567);
352 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
353 _mm_storel_epi64((__m128i*) c3, voutb3x01234567);
354 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
355 _mm_storel_epi64((__m128i*) c2, voutb2x01234567);
356 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
357 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
358 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
359 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
360 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
361 a = (const int8_t**restrict) ((uintptr_t) a - ks);
362 nc -= 8;
363 } else {
364 if (nc & 4) {
365 _mm_storeu_si32(c7, voutb7x01234567);
366 c7 += 4;
367 _mm_storeu_si32(c6, voutb6x01234567);
368 c6 += 4;
369 _mm_storeu_si32(c5, voutb5x01234567);
370 c5 += 4;
371 _mm_storeu_si32(c4, voutb4x01234567);
372 c4 += 4;
373 _mm_storeu_si32(c3, voutb3x01234567);
374 c3 += 4;
375 _mm_storeu_si32(c2, voutb2x01234567);
376 c2 += 4;
377 _mm_storeu_si32(c1, voutb1x01234567);
378 c1 += 4;
379 _mm_storeu_si32(c0, voutb0x01234567);
380 c0 += 4;
381 voutb7x01234567 = _mm_srli_epi64(voutb7x01234567, 32);
382 voutb6x01234567 = _mm_srli_epi64(voutb6x01234567, 32);
383 voutb5x01234567 = _mm_srli_epi64(voutb5x01234567, 32);
384 voutb4x01234567 = _mm_srli_epi64(voutb4x01234567, 32);
385 voutb3x01234567 = _mm_srli_epi64(voutb3x01234567, 32);
386 voutb2x01234567 = _mm_srli_epi64(voutb2x01234567, 32);
387 voutb1x01234567 = _mm_srli_epi64(voutb1x01234567, 32);
388 voutb0x01234567 = _mm_srli_epi64(voutb0x01234567, 32);
389 }
390 if (nc & 2) {
391 unaligned_store_u16(c7, (uint16_t) _mm_extract_epi16(voutb7x01234567, 0));
392 c7 += 2;
393 unaligned_store_u16(c6, (uint16_t) _mm_extract_epi16(voutb6x01234567, 0));
394 c6 += 2;
395 unaligned_store_u16(c5, (uint16_t) _mm_extract_epi16(voutb5x01234567, 0));
396 c5 += 2;
397 unaligned_store_u16(c4, (uint16_t) _mm_extract_epi16(voutb4x01234567, 0));
398 c4 += 2;
399 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(voutb3x01234567, 0));
400 c3 += 2;
401 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(voutb2x01234567, 0));
402 c2 += 2;
403 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(voutb1x01234567, 0));
404 c1 += 2;
405 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(voutb0x01234567, 0));
406 c0 += 2;
407 voutb7x01234567 = _mm_srli_epi32(voutb7x01234567, 16);
408 voutb6x01234567 = _mm_srli_epi32(voutb6x01234567, 16);
409 voutb5x01234567 = _mm_srli_epi32(voutb5x01234567, 16);
410 voutb4x01234567 = _mm_srli_epi32(voutb4x01234567, 16);
411 voutb3x01234567 = _mm_srli_epi32(voutb3x01234567, 16);
412 voutb2x01234567 = _mm_srli_epi32(voutb2x01234567, 16);
413 voutb1x01234567 = _mm_srli_epi32(voutb1x01234567, 16);
414 voutb0x01234567 = _mm_srli_epi32(voutb0x01234567, 16);
415 }
416 if (nc & 1) {
417 *c7 = (int8_t) _mm_extract_epi8(voutb7x01234567, 0);
418 *c6 = (int8_t) _mm_extract_epi8(voutb6x01234567, 0);
419 *c5 = (int8_t) _mm_extract_epi8(voutb5x01234567, 0);
420 *c4 = (int8_t) _mm_extract_epi8(voutb4x01234567, 0);
421 *c3 = (int8_t) _mm_extract_epi8(voutb3x01234567, 0);
422 *c2 = (int8_t) _mm_extract_epi8(voutb2x01234567, 0);
423 *c1 = (int8_t) _mm_extract_epi8(voutb1x01234567, 0);
424 *c0 = (int8_t) _mm_extract_epi8(voutb0x01234567, 0);
425 }
426 nc = 0;
427 }
428 } while (nc != 0);
429 }