1 // Copyright (c) Facebook, Inc. and its affiliates.
2 // All rights reserved.
3 //
4 // Copyright 2019 Google LLC
5 //
6 // This source code is licensed under the BSD-style license found in the
7 // LICENSE file in the root directory of this source tree.
8
9 #include <assert.h>
10
11 #include <immintrin.h>
12
13 #include <xnnpack/igemm.h>
14 #include <xnnpack/math.h>
15
16
xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const uint8_t** restrict a,
23 const void* restrict w,
24 uint8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const uint8_t* zero,
29 const union xnn_qu8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
30 {
31 assert(mr != 0);
32 assert(mr <= 4);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (4 * sizeof(void*)) == 0);
37 assert(a_offset % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 2);
43 uint8_t* c0 = c;
44 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
53 if (mr != 4) {
54 c3 = c2;
55 }
56
57 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->sse2.kernel_zero_point);
58
59 do {
60 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
61 __m128i vacc1x0123 = vacc0x0123;
62 __m128i vacc2x0123 = vacc0x0123;
63 __m128i vacc3x0123 = vacc0x0123;
64 w = (const void*) ((uintptr_t) w + 16);
65
66 const __m128i vzero = _mm_setzero_si128();
67 size_t p = ks;
68 do {
69 const uint8_t* restrict a0 = a[0];
70 if XNN_UNPREDICTABLE(a0 != zero) {
71 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
72 }
73 const uint8_t* restrict a1 = a[1];
74 if XNN_UNPREDICTABLE(a1 != zero) {
75 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
76 }
77 const uint8_t* restrict a2 = a[2];
78 if XNN_UNPREDICTABLE(a2 != zero) {
79 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
80 }
81 const uint8_t* restrict a3 = a[3];
82 if XNN_UNPREDICTABLE(a3 != zero) {
83 a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
84 }
85 a += 4;
86
87 size_t k = kc;
88 while (k >= 8 * sizeof(uint8_t)) {
89 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
90 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
91 a0 += 8;
92 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
93 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
94 a1 += 8;
95 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
96 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
97 a2 += 8;
98 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
99 const __m128i vxa3 = _mm_unpacklo_epi8(va3, vzero);
100 a3 += 8;
101
102 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
103 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
104 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
105 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
106 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
107 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
108
109 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
110 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
111 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
112 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
113 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
114 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
115
116 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
117 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
118 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
119 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
120 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
121 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
122
123 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
124 const __m128i vxb3 = _mm_sub_epi16(_mm_unpacklo_epi8(vb3, vzero), vb_zero_point);
125 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
126 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
127 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
128 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
129
130 w = (void*) ((uintptr_t) w + 32);
131
132 k -= 8 * sizeof(uint8_t);
133 }
134 if (k != 0) {
135 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
136 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
137 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
138 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
139 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
140 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
141 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
142 const __m128i vxa3 = _mm_unpacklo_epi8(va3, vzero);
143
144 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
145 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
146 w = (void*) ((uintptr_t) w + 8);
147
148 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
149 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
150 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
151 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
152
153 if (k > 2 * sizeof(uint8_t)) {
154 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
155 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
156 w = (void*) ((uintptr_t) w + 8);
157
158 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
159 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
160 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
161 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
162
163 if (k > 4 * sizeof(uint8_t)) {
164 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
165 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
166 w = (void*) ((uintptr_t) w + 8);
167
168 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
169 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
170 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
171 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
172 }
173 }
174 }
175 p -= 4 * sizeof(void*);
176 } while (p != 0);
177
178 const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
179 const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
180
181 const __m128i vnmask0x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0x0123);
182 const __m128i vnmask1x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc1x0123);
183 const __m128i vnmask2x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc2x0123);
184 const __m128i vnmask3x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc3x0123);
185
186 const __m128i vabsacc0x0123 = _mm_sub_epi32(_mm_xor_si128(vacc0x0123, vnmask0x0123), vnmask0x0123);
187 const __m128i vabsacc1x0123 = _mm_sub_epi32(_mm_xor_si128(vacc1x0123, vnmask1x0123), vnmask1x0123);
188 const __m128i vabsacc2x0123 = _mm_sub_epi32(_mm_xor_si128(vacc2x0123, vnmask2x0123), vnmask2x0123);
189 const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123);
190
191 const __m128i vabsacc0x1032 = _mm_shuffle_epi32(vabsacc0x0123, _MM_SHUFFLE(2, 3, 0, 1));
192 const __m128i vabsacc1x1032 = _mm_shuffle_epi32(vabsacc1x0123, _MM_SHUFFLE(2, 3, 0, 1));
193 const __m128i vabsacc2x1032 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(2, 3, 0, 1));
194 const __m128i vabsacc3x1032 = _mm_shuffle_epi32(vabsacc3x0123, _MM_SHUFFLE(2, 3, 0, 1));
195
196 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier);
197 const __m128i vabsprod1x02 = _mm_mul_epu32(vabsacc1x0123, vmultiplier);
198 const __m128i vabsprod2x02 = _mm_mul_epu32(vabsacc2x0123, vmultiplier);
199 const __m128i vabsprod3x02 = _mm_mul_epu32(vabsacc3x0123, vmultiplier);
200
201 const __m128i vnmask0x02 = _mm_shuffle_epi32(vnmask0x0123, _MM_SHUFFLE(2, 2, 0, 0));
202 const __m128i vnmask1x02 = _mm_shuffle_epi32(vnmask1x0123, _MM_SHUFFLE(2, 2, 0, 0));
203 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0));
204 const __m128i vnmask3x02 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(2, 2, 0, 0));
205
206 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02);
207 const __m128i vprod1x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod1x02, vnmask1x02), vnmask1x02);
208 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02);
209 const __m128i vprod3x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod3x02, vnmask3x02), vnmask3x02);
210
211 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31);
212 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31);
213 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31);
214 const __m128i vq31prod3x02 = _mm_srli_epi64(_mm_add_epi64(vprod3x02, vrounding), 31);
215
216 const __m128i vabsprod0x13 = _mm_mul_epu32(vabsacc0x1032, vmultiplier);
217 const __m128i vabsprod1x13 = _mm_mul_epu32(vabsacc1x1032, vmultiplier);
218 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1032, vmultiplier);
219 const __m128i vabsprod3x13 = _mm_mul_epu32(vabsacc3x1032, vmultiplier);
220
221 const __m128i vnmask0x13 = _mm_shuffle_epi32(vnmask0x0123, _MM_SHUFFLE(3, 3, 1, 1));
222 const __m128i vnmask1x13 = _mm_shuffle_epi32(vnmask1x0123, _MM_SHUFFLE(3, 3, 1, 1));
223 const __m128i vnmask2x13 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(3, 3, 1, 1));
224 const __m128i vnmask3x13 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(3, 3, 1, 1));
225
226 const __m128i vprod0x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x13, vnmask0x13), vnmask0x13);
227 const __m128i vprod1x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod1x13, vnmask1x13), vnmask1x13);
228 const __m128i vprod2x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x13, vnmask2x13), vnmask2x13);
229 const __m128i vprod3x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod3x13, vnmask3x13), vnmask3x13);
230
231 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31);
232 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31);
233 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31);
234 const __m128i vq31prod3x13 = _mm_srli_epi64(_mm_add_epi64(vprod3x13, vrounding), 31);
235
236 const __m128i vq31prod0x0213 = _mm_castps_si128(_mm_shuffle_ps(
237 _mm_castsi128_ps(vq31prod0x02), _mm_castsi128_ps(vq31prod0x13), _MM_SHUFFLE(2, 0, 2, 0)));
238 const __m128i vq31prod1x0213 = _mm_castps_si128(_mm_shuffle_ps(
239 _mm_castsi128_ps(vq31prod1x02), _mm_castsi128_ps(vq31prod1x13), _MM_SHUFFLE(2, 0, 2, 0)));
240 const __m128i vq31prod2x0213 = _mm_castps_si128(_mm_shuffle_ps(
241 _mm_castsi128_ps(vq31prod2x02), _mm_castsi128_ps(vq31prod2x13), _MM_SHUFFLE(2, 0, 2, 0)));
242 const __m128i vq31prod3x0213 = _mm_castps_si128(_mm_shuffle_ps(
243 _mm_castsi128_ps(vq31prod3x02), _mm_castsi128_ps(vq31prod3x13), _MM_SHUFFLE(2, 0, 2, 0)));
244
245 const __m128i vq31prod0x0123 = _mm_shuffle_epi32(vq31prod0x0213, _MM_SHUFFLE(3, 1, 2, 0));
246 const __m128i vq31prod1x0123 = _mm_shuffle_epi32(vq31prod1x0213, _MM_SHUFFLE(3, 1, 2, 0));
247 const __m128i vq31prod2x0123 = _mm_shuffle_epi32(vq31prod2x0213, _MM_SHUFFLE(3, 1, 2, 0));
248 const __m128i vq31prod3x0123 = _mm_shuffle_epi32(vq31prod3x0213, _MM_SHUFFLE(3, 1, 2, 0));
249
250 const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
251
252 const __m128i vrem0x0123 =
253 _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
254 const __m128i vrem1x0123 =
255 _mm_add_epi32(_mm_and_si128(vq31prod1x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod1x0123));
256 const __m128i vrem2x0123 =
257 _mm_add_epi32(_mm_and_si128(vq31prod2x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod2x0123));
258 const __m128i vrem3x0123 =
259 _mm_add_epi32(_mm_and_si128(vq31prod3x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod3x0123));
260
261 const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
262 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
263
264 vacc0x0123 = _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
265 vacc1x0123 = _mm_sub_epi32(_mm_sra_epi32(vq31prod1x0123, vshift), _mm_cmpgt_epi32(vrem1x0123, vremainder_threshold));
266 vacc2x0123 = _mm_sub_epi32(_mm_sra_epi32(vq31prod2x0123, vshift), _mm_cmpgt_epi32(vrem2x0123, vremainder_threshold));
267 vacc3x0123 = _mm_sub_epi32(_mm_sra_epi32(vq31prod3x0123, vshift), _mm_cmpgt_epi32(vrem3x0123, vremainder_threshold));
268
269 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
270 const __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
271 const __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
272 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc23x0123);
273 vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) params->sse2.output_max));
274 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->sse2.output_min));
275
276 if XNN_LIKELY(nc >= 4) {
277 *((uint32_t*) c3) = (uint32_t) _mm_cvtsi128_si32(_mm_srli_si128(vout, 12));
278 c3 += cn_stride;
279 *((uint32_t*) c2) = (uint32_t) _mm_cvtsi128_si32(_mm_unpackhi_epi32(vout, vout));
280 c2 += cn_stride;
281 *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(_mm_srli_epi64(vout, 32));
282 c1 += cn_stride;
283 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
284 c0 += cn_stride;
285
286 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
287
288 nc -= 4;
289 } else {
290 if (nc & 2) {
291 *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6); c3 += 2;
292 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4); c2 += 2;
293 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2); c1 += 2;
294 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0); c0 += 2;
295 vout = _mm_srli_epi32(vout, 16);
296 }
297 if (nc & 1) {
298 *((uint8_t*) c3) = (uint8_t) _mm_extract_epi16(vout, 6);
299 *((uint8_t*) c2) = (uint8_t) _mm_extract_epi16(vout, 4);
300 *((uint8_t*) c1) = (uint8_t) _mm_extract_epi16(vout, 2);
301 *((uint8_t*) c0) = (uint8_t) _mm_cvtsi128_si32(vout);
302 }
303
304 nc = 0;
305 }
306 } while (nc != 0);
307 }
308