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
15
xnn_q8_igemm_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_q8_gemm_params params[restrict static1])16 void xnn_q8_igemm_ukernel_4x4c2__sse2(
17 size_t mr,
18 size_t nc,
19 size_t kc,
20 size_t ks,
21 const uint8_t** restrict a,
22 const void* restrict w,
23 uint8_t* restrict c,
24 size_t cm_stride,
25 size_t cn_stride,
26 size_t a_offset,
27 const uint8_t* zero,
28 const union xnn_q8_gemm_params params[restrict static 1])
29 {
30 assert(mr != 0);
31 assert(mr <= 4);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(ks != 0);
35 assert(ks % (4 * sizeof(void*)) == 0);
36
37 uint8_t* c0 = c;
38 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
39 if XNN_UNPREDICTABLE(mr < 2) {
40 c1 = c0;
41 }
42 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
43 if XNN_UNPREDICTABLE(mr <= 2) {
44 c2 = c1;
45 }
46 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
47 if (mr != 4) {
48 c3 = c2;
49 }
50
51 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->sse2.kernel_zero_point);
52
53 do {
54 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
55 __m128i vacc1x0123 = vacc0x0123;
56 __m128i vacc2x0123 = vacc0x0123;
57 __m128i vacc3x0123 = vacc0x0123;
58 w = (const void*) ((uintptr_t) w + 16);
59
60 const __m128i vzero = _mm_setzero_si128();
61 size_t p = ks;
62 do {
63 const uint8_t* restrict a0 = a[0];
64 if XNN_UNPREDICTABLE(a0 != zero) {
65 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
66 }
67 const uint8_t* restrict a1 = a[1];
68 if XNN_UNPREDICTABLE(a1 != zero) {
69 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
70 }
71 const uint8_t* restrict a2 = a[2];
72 if XNN_UNPREDICTABLE(a2 != zero) {
73 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
74 }
75 const uint8_t* restrict a3 = a[3];
76 if XNN_UNPREDICTABLE(a3 != zero) {
77 a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
78 }
79 a += 4;
80
81 size_t k = kc;
82 while (k >= 8 * sizeof(uint8_t)) {
83 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
84 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
85 a0 += 8;
86 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
87 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
88 a1 += 8;
89 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
90 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
91 a2 += 8;
92 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
93 const __m128i vxa3 = _mm_unpacklo_epi8(va3, vzero);
94 a3 += 8;
95
96 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
97 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
98 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
99 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
100 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
101 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
102
103 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
104 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
105 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
106 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
107 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
108 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
109
110 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
111 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
112 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
113 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
114 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
115 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
116
117 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
118 const __m128i vxb3 = _mm_sub_epi16(_mm_unpacklo_epi8(vb3, vzero), vb_zero_point);
119 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
120 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
121 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
122 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
123
124 w = (void*) ((uintptr_t) w + 32);
125
126 k -= 8 * sizeof(uint8_t);
127 }
128 if (k != 0) {
129 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
130 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
131 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
132 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
133 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
134 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
135 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
136 const __m128i vxa3 = _mm_unpacklo_epi8(va3, vzero);
137
138 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
139 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
140 w = (void*) ((uintptr_t) w + 8);
141
142 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
143 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
144 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
145 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
146
147 if (k > 2 * sizeof(uint8_t)) {
148 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
149 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
150 w = (void*) ((uintptr_t) w + 8);
151
152 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
153 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
154 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
155 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
156
157 if (k > 4 * sizeof(uint8_t)) {
158 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
159 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
160 w = (void*) ((uintptr_t) w + 8);
161
162 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
163 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
164 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
165 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
166
167 if (k > 6 * sizeof(uint8_t)) {
168 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) w);
169 const __m128i vxb3 = _mm_sub_epi16(_mm_unpacklo_epi8(vb3, vzero), vb_zero_point);
170 w = (void*) ((uintptr_t) w + 8);
171
172 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
173 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
174 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
175 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
176 }
177 }
178 }
179 }
180 p -= 4 * sizeof(void*);
181 } while (p != 0);
182
183 const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
184 const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
185
186 const __m128i vnmask0x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0x0123);
187 const __m128i vnmask1x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc1x0123);
188 const __m128i vnmask2x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc2x0123);
189 const __m128i vnmask3x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc3x0123);
190
191 const __m128i vabsacc0x0123 = _mm_sub_epi32(_mm_xor_si128(vacc0x0123, vnmask0x0123), vnmask0x0123);
192 const __m128i vabsacc1x0123 = _mm_sub_epi32(_mm_xor_si128(vacc1x0123, vnmask1x0123), vnmask1x0123);
193 const __m128i vabsacc2x0123 = _mm_sub_epi32(_mm_xor_si128(vacc2x0123, vnmask2x0123), vnmask2x0123);
194 const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123);
195
196 const __m128i vabsacc0x1032 = _mm_shuffle_epi32(vabsacc0x0123, _MM_SHUFFLE(2, 3, 0, 1));
197 const __m128i vabsacc1x1032 = _mm_shuffle_epi32(vabsacc1x0123, _MM_SHUFFLE(2, 3, 0, 1));
198 const __m128i vabsacc2x1032 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(2, 3, 0, 1));
199 const __m128i vabsacc3x1032 = _mm_shuffle_epi32(vabsacc3x0123, _MM_SHUFFLE(2, 3, 0, 1));
200
201 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier);
202 const __m128i vabsprod1x02 = _mm_mul_epu32(vabsacc1x0123, vmultiplier);
203 const __m128i vabsprod2x02 = _mm_mul_epu32(vabsacc2x0123, vmultiplier);
204 const __m128i vabsprod3x02 = _mm_mul_epu32(vabsacc3x0123, vmultiplier);
205
206 const __m128i vnmask0x02 = _mm_shuffle_epi32(vnmask0x0123, _MM_SHUFFLE(2, 2, 0, 0));
207 const __m128i vnmask1x02 = _mm_shuffle_epi32(vnmask1x0123, _MM_SHUFFLE(2, 2, 0, 0));
208 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0));
209 const __m128i vnmask3x02 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(2, 2, 0, 0));
210
211 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02);
212 const __m128i vprod1x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod1x02, vnmask1x02), vnmask1x02);
213 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02);
214 const __m128i vprod3x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod3x02, vnmask3x02), vnmask3x02);
215
216 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31);
217 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31);
218 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31);
219 const __m128i vq31prod3x02 = _mm_srli_epi64(_mm_add_epi64(vprod3x02, vrounding), 31);
220
221 const __m128i vabsprod0x13 = _mm_mul_epu32(vabsacc0x1032, vmultiplier);
222 const __m128i vabsprod1x13 = _mm_mul_epu32(vabsacc1x1032, vmultiplier);
223 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1032, vmultiplier);
224 const __m128i vabsprod3x13 = _mm_mul_epu32(vabsacc3x1032, vmultiplier);
225
226 const __m128i vnmask0x13 = _mm_shuffle_epi32(vnmask0x0123, _MM_SHUFFLE(3, 3, 1, 1));
227 const __m128i vnmask1x13 = _mm_shuffle_epi32(vnmask1x0123, _MM_SHUFFLE(3, 3, 1, 1));
228 const __m128i vnmask2x13 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(3, 3, 1, 1));
229 const __m128i vnmask3x13 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(3, 3, 1, 1));
230
231 const __m128i vprod0x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x13, vnmask0x13), vnmask0x13);
232 const __m128i vprod1x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod1x13, vnmask1x13), vnmask1x13);
233 const __m128i vprod2x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x13, vnmask2x13), vnmask2x13);
234 const __m128i vprod3x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod3x13, vnmask3x13), vnmask3x13);
235
236 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31);
237 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31);
238 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31);
239 const __m128i vq31prod3x13 = _mm_srli_epi64(_mm_add_epi64(vprod3x13, vrounding), 31);
240
241 const __m128i vq31prod0x0213 = _mm_castps_si128(_mm_shuffle_ps(
242 _mm_castsi128_ps(vq31prod0x02), _mm_castsi128_ps(vq31prod0x13), _MM_SHUFFLE(2, 0, 2, 0)));
243 const __m128i vq31prod1x0213 = _mm_castps_si128(_mm_shuffle_ps(
244 _mm_castsi128_ps(vq31prod1x02), _mm_castsi128_ps(vq31prod1x13), _MM_SHUFFLE(2, 0, 2, 0)));
245 const __m128i vq31prod2x0213 = _mm_castps_si128(_mm_shuffle_ps(
246 _mm_castsi128_ps(vq31prod2x02), _mm_castsi128_ps(vq31prod2x13), _MM_SHUFFLE(2, 0, 2, 0)));
247 const __m128i vq31prod3x0213 = _mm_castps_si128(_mm_shuffle_ps(
248 _mm_castsi128_ps(vq31prod3x02), _mm_castsi128_ps(vq31prod3x13), _MM_SHUFFLE(2, 0, 2, 0)));
249
250 const __m128i vq31prod0x0123 = _mm_shuffle_epi32(vq31prod0x0213, _MM_SHUFFLE(3, 1, 2, 0));
251 const __m128i vq31prod1x0123 = _mm_shuffle_epi32(vq31prod1x0213, _MM_SHUFFLE(3, 1, 2, 0));
252 const __m128i vq31prod2x0123 = _mm_shuffle_epi32(vq31prod2x0213, _MM_SHUFFLE(3, 1, 2, 0));
253 const __m128i vq31prod3x0123 = _mm_shuffle_epi32(vq31prod3x0213, _MM_SHUFFLE(3, 1, 2, 0));
254
255 const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
256
257 const __m128i vrem0x0123 =
258 _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
259 const __m128i vrem1x0123 =
260 _mm_add_epi32(_mm_and_si128(vq31prod1x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod1x0123));
261 const __m128i vrem2x0123 =
262 _mm_add_epi32(_mm_and_si128(vq31prod2x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod2x0123));
263 const __m128i vrem3x0123 =
264 _mm_add_epi32(_mm_and_si128(vq31prod3x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod3x0123));
265
266 const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
267 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
268
269 vacc0x0123 = _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
270 vacc1x0123 = _mm_sub_epi32(_mm_sra_epi32(vq31prod1x0123, vshift), _mm_cmpgt_epi32(vrem1x0123, vremainder_threshold));
271 vacc2x0123 = _mm_sub_epi32(_mm_sra_epi32(vq31prod2x0123, vshift), _mm_cmpgt_epi32(vrem2x0123, vremainder_threshold));
272 vacc3x0123 = _mm_sub_epi32(_mm_sra_epi32(vq31prod3x0123, vshift), _mm_cmpgt_epi32(vrem3x0123, vremainder_threshold));
273
274 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
275 const __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
276 const __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
277 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc23x0123);
278 vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) params->sse2.output_max));
279 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->sse2.output_min));
280
281 if XNN_LIKELY(nc >= 4) {
282 *((uint32_t*) c3) = (uint32_t) _mm_cvtsi128_si32(_mm_srli_si128(vout, 12));
283 c3 += cn_stride;
284 *((uint32_t*) c2) = (uint32_t) _mm_cvtsi128_si32(_mm_unpackhi_epi32(vout, vout));
285 c2 += cn_stride;
286 *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(_mm_srli_epi64(vout, 32));
287 c1 += cn_stride;
288 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
289 c0 += cn_stride;
290
291 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
292
293 nc -= 4;
294 } else {
295 if (nc & 2) {
296 *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6); c3 += 2;
297 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4); c2 += 2;
298 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2); c1 += 2;
299 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0); c0 += 2;
300 vout = _mm_srli_epi32(vout, 16);
301 }
302 if (nc & 1) {
303 *((uint8_t*) c3) = (uint8_t) _mm_extract_epi16(vout, 6);
304 *((uint8_t*) c2) = (uint8_t) _mm_extract_epi16(vout, 4);
305 *((uint8_t*) c1) = (uint8_t) _mm_extract_epi16(vout, 2);
306 *((uint8_t*) c0) = (uint8_t) _mm_cvtsi128_si32(vout);
307 }
308
309 nc = 0;
310 }
311 } while (nc != 0);
312 }
313