Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-7x8c8-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_7x8c8__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 <= 7);
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
72 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
73 XNN_FORCE_REALIZATION(vsign_mask);
74 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);
75 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
76 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
77 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
78 // XNN_FORCE_REALIZATION(voutput_zero_point);
79 // XNN_FORCE_REALIZATION(voutput_min);
80 do {
81 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
82 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
83 __m256i vacc1x0123 = vacc0x0123;
84 __m256i vacc1x4567 = vacc0x4567;
85 __m256i vacc2x0123 = vacc0x0123;
86 __m256i vacc2x4567 = vacc0x4567;
87 __m256i vacc3x0123 = vacc0x0123;
88 __m256i vacc3x4567 = vacc0x4567;
89 __m256i vacc4x0123 = vacc0x0123;
90 __m256i vacc4x4567 = vacc0x4567;
91 __m256i vacc5x0123 = vacc0x0123;
92 __m256i vacc5x4567 = vacc0x4567;
93 __m256i vacc6x0123 = vacc0x0123;
94 __m256i vacc6x4567 = vacc0x4567;
95 w = (const int32_t*) w + 8;
96
97 size_t p = ks;
98 do {
99 const int8_t* restrict a0 = a[0];
100 if XNN_UNPREDICTABLE(a0 != zero) {
101 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
102 }
103 const int8_t* restrict a1 = a[1];
104 if XNN_UNPREDICTABLE(a1 != zero) {
105 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
106 }
107 const int8_t* restrict a2 = a[2];
108 if XNN_UNPREDICTABLE(a2 != zero) {
109 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
110 }
111 const int8_t* restrict a3 = a[3];
112 if XNN_UNPREDICTABLE(a3 != zero) {
113 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
114 }
115 const int8_t* restrict a4 = a[4];
116 if XNN_UNPREDICTABLE(a4 != zero) {
117 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
118 }
119 const int8_t* restrict a5 = a[5];
120 if XNN_UNPREDICTABLE(a5 != zero) {
121 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
122 }
123 const int8_t* restrict a6 = a[6];
124 if XNN_UNPREDICTABLE(a6 != zero) {
125 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
126 }
127 a += 7;
128
129 size_t k = kc;
130 while (k >= 16 * sizeof(int8_t)) {
131 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
132 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
133 a0 += 16;
134 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
135 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
136 a1 += 16;
137 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
138 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
139 a2 += 16;
140 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
141 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
142 a3 += 16;
143 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
144 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
145 a4 += 16;
146 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
147 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
148 a5 += 16;
149 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
150 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
151 a6 += 16;
152
153 const __m256i vb01234567x0123 = _mm256_load_si256(w);
154 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
155 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
156 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
157 xnn_prefetch_to_l1((const int8_t*) w + 896);
158
159 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
160 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
161 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
162 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
163 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
164 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
165 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
166 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
167 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
168 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
169 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
170 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
171 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
172 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
173 xnn_prefetch_to_l1((const int8_t*) w + 960);
174 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
175 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
176 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
177 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
178 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
179 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
180 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
181 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
182 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
183 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
184 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
185 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
186 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
187 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
188
189 w = (const int8_t*) w + 128;
190 k -= 16 * sizeof(int8_t);
191 }
192
193 if (k != 0) {
194 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
195 a0 += 8;
196 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
197 a1 += 8;
198 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
199 a2 += 8;
200 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
201 a3 += 8;
202 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
203 a4 += 8;
204 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
205 a5 += 8;
206 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
207 a6 += 8;
208
209 const __m256i vb01234567x0123 = _mm256_load_si256(w);
210 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
211
212 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
213 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
214 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
215 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
216 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
217 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
218 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
219 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
220 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
221 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
222 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
223 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
224 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
225 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
226 xnn_prefetch_to_l1((const int8_t*) w + 960);
227
228 w = (const int8_t*) w + 64;
229 k -= 8 * sizeof(int8_t);
230 }
231
232 p -= 7 * sizeof(void*);
233 } while (p != 0);
234
235
236 // Add adjacent pairs
237 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
238 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
239 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
240 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
241 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
242 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
243 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
244 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
245 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
246 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
247 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
248 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
249 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
250 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
251
252 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
253 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
254 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
255 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
256 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
257 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
258 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
259
260 const __m256 vscale01234567 = _mm256_load_ps(w);
261 w = (const float*) w + 8;
262 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
263 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
264 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);
265 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);
266 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vscale01234567);
267 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vscale01234567);
268 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vscale01234567);
269
270 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
271 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
272 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);
273 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);
274 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max_less_zero_point);
275 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max_less_zero_point);
276 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max_less_zero_point);
277
278 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
279 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
280 vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);
281 vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);
282 vacc4x01234567 = _mm256_cvtps_epi32(vout4x01234567);
283 vacc5x01234567 = _mm256_cvtps_epi32(vout5x01234567);
284 vacc6x01234567 = _mm256_cvtps_epi32(vout6x01234567);
285
286 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
287 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
288 vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);
289 vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);
290 vacc4x01234567 = _mm256_add_epi32(vacc4x01234567, voutput_zero_point);
291 vacc5x01234567 = _mm256_add_epi32(vacc5x01234567, voutput_zero_point);
292 vacc6x01234567 = _mm256_add_epi32(vacc6x01234567, voutput_zero_point);
293
294 vacc0x01234567 = _mm256_packs_epi32(vacc0x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc0x01234567, 1)));
295 __m128i voutb0x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc0x01234567, vacc0x01234567));
296 vacc1x01234567 = _mm256_packs_epi32(vacc1x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc1x01234567, 1)));
297 __m128i voutb1x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc1x01234567, vacc1x01234567));
298 vacc2x01234567 = _mm256_packs_epi32(vacc2x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc2x01234567, 1)));
299 __m128i voutb2x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc2x01234567, vacc2x01234567));
300 vacc3x01234567 = _mm256_packs_epi32(vacc3x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc3x01234567, 1)));
301 __m128i voutb3x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc3x01234567, vacc3x01234567));
302 vacc4x01234567 = _mm256_packs_epi32(vacc4x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc4x01234567, 1)));
303 __m128i voutb4x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc4x01234567, vacc4x01234567));
304 vacc5x01234567 = _mm256_packs_epi32(vacc5x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc5x01234567, 1)));
305 __m128i voutb5x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc5x01234567, vacc5x01234567));
306 vacc6x01234567 = _mm256_packs_epi32(vacc6x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc6x01234567, 1)));
307 __m128i voutb6x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc6x01234567, vacc6x01234567));
308
309 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
310 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
311 voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);
312 voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);
313 voutb4x01234567 = _mm_max_epi8(voutb4x01234567, voutput_min);
314 voutb5x01234567 = _mm_max_epi8(voutb5x01234567, voutput_min);
315 voutb6x01234567 = _mm_max_epi8(voutb6x01234567, voutput_min);
316
317 if (nc >= 8) {
318 _mm_storel_epi64((__m128i*) c6, voutb6x01234567);
319 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
320 _mm_storel_epi64((__m128i*) c5, voutb5x01234567);
321 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
322 _mm_storel_epi64((__m128i*) c4, voutb4x01234567);
323 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
324 _mm_storel_epi64((__m128i*) c3, voutb3x01234567);
325 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
326 _mm_storel_epi64((__m128i*) c2, voutb2x01234567);
327 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
328 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
329 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
330 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
331 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
332 a = (const int8_t**restrict) ((uintptr_t) a - ks);
333 nc -= 8;
334 } else {
335 if (nc & 4) {
336 _mm_storeu_si32(c6, voutb6x01234567);
337 c6 += 4;
338 _mm_storeu_si32(c5, voutb5x01234567);
339 c5 += 4;
340 _mm_storeu_si32(c4, voutb4x01234567);
341 c4 += 4;
342 _mm_storeu_si32(c3, voutb3x01234567);
343 c3 += 4;
344 _mm_storeu_si32(c2, voutb2x01234567);
345 c2 += 4;
346 _mm_storeu_si32(c1, voutb1x01234567);
347 c1 += 4;
348 _mm_storeu_si32(c0, voutb0x01234567);
349 c0 += 4;
350 voutb6x01234567 = _mm_srli_epi64(voutb6x01234567, 32);
351 voutb5x01234567 = _mm_srli_epi64(voutb5x01234567, 32);
352 voutb4x01234567 = _mm_srli_epi64(voutb4x01234567, 32);
353 voutb3x01234567 = _mm_srli_epi64(voutb3x01234567, 32);
354 voutb2x01234567 = _mm_srli_epi64(voutb2x01234567, 32);
355 voutb1x01234567 = _mm_srli_epi64(voutb1x01234567, 32);
356 voutb0x01234567 = _mm_srli_epi64(voutb0x01234567, 32);
357 }
358 if (nc & 2) {
359 unaligned_store_u16(c6, (uint16_t) _mm_extract_epi16(voutb6x01234567, 0));
360 c6 += 2;
361 unaligned_store_u16(c5, (uint16_t) _mm_extract_epi16(voutb5x01234567, 0));
362 c5 += 2;
363 unaligned_store_u16(c4, (uint16_t) _mm_extract_epi16(voutb4x01234567, 0));
364 c4 += 2;
365 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(voutb3x01234567, 0));
366 c3 += 2;
367 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(voutb2x01234567, 0));
368 c2 += 2;
369 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(voutb1x01234567, 0));
370 c1 += 2;
371 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(voutb0x01234567, 0));
372 c0 += 2;
373 voutb6x01234567 = _mm_srli_epi32(voutb6x01234567, 16);
374 voutb5x01234567 = _mm_srli_epi32(voutb5x01234567, 16);
375 voutb4x01234567 = _mm_srli_epi32(voutb4x01234567, 16);
376 voutb3x01234567 = _mm_srli_epi32(voutb3x01234567, 16);
377 voutb2x01234567 = _mm_srli_epi32(voutb2x01234567, 16);
378 voutb1x01234567 = _mm_srli_epi32(voutb1x01234567, 16);
379 voutb0x01234567 = _mm_srli_epi32(voutb0x01234567, 16);
380 }
381 if (nc & 1) {
382 *c6 = (int8_t) _mm_extract_epi8(voutb6x01234567, 0);
383 *c5 = (int8_t) _mm_extract_epi8(voutb5x01234567, 0);
384 *c4 = (int8_t) _mm_extract_epi8(voutb4x01234567, 0);
385 *c3 = (int8_t) _mm_extract_epi8(voutb3x01234567, 0);
386 *c2 = (int8_t) _mm_extract_epi8(voutb2x01234567, 0);
387 *c1 = (int8_t) _mm_extract_epi8(voutb1x01234567, 0);
388 *c0 = (int8_t) _mm_extract_epi8(voutb0x01234567, 0);
389 }
390 nc = 0;
391 }
392 } while (nc != 0);
393 }