Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x4c8-minmax-sse2-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c8-sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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 <emmintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_4x4c8__sse2_ld64(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const struct xnn_f32_qb4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
31 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 4);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 size_t bl = params->scalar.blocksize;
43 assert(bl <= round_up_po2(kc, 2));
44 assert(bl != 0);
45 assert(bl % 32 == 0);
46 kc = round_up_po2(kc, 8 * sizeof(int8_t));
47 const int8_t* a0 = a;
48 float* c0 = c;
49 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
50 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
51 if XNN_UNPREDICTABLE(mr < 2) {
52 a1 = a0;
53 c1 = c0;
54 }
55 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
56 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
57 if XNN_UNPREDICTABLE(mr <= 2) {
58 a2 = a1;
59 c2 = c1;
60 }
61 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
62 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
63 if XNN_UNPREDICTABLE(mr != 4) {
64 a3 = a2;
65 c3 = c2;
66 }
67
68 const __m128 vmin = _mm_set1_ps(params->scalar.min);
69 const __m128 vmax = _mm_set1_ps(params->scalar.max);
70 XNN_FORCE_REALIZATION(vmin);
71 XNN_FORCE_REALIZATION(vmax);
72
73 const __m128i vmask = _mm_set1_epi8(0xF0);
74 XNN_FORCE_REALIZATION(vmask);
75
76 do {
77 const __m128 vksum = _mm_load_ps((const float*) w);
78 const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
79 const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
80 const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));
81 const __m128i vinput_zero_point23 = _mm_loadu_si128((const __m128i*) &quantization_params[2]);
82 const __m128i vinput_zero_point2 = _mm_shuffle_epi32(vinput_zero_point23, _MM_SHUFFLE(0, 0, 0, 0));
83 const __m128i vinput_zero_point3 = _mm_shuffle_epi32(vinput_zero_point23, _MM_SHUFFLE(2, 2, 2, 2));
84
85 __m128 vinput_zero_point0_float = _mm_cvtepi32_ps(vinput_zero_point0);
86 __m128 vout0x0123 = _mm_mul_ps(vksum, vinput_zero_point0_float);
87 __m128 vinput_zero_point1_float = _mm_cvtepi32_ps(vinput_zero_point1);
88 __m128 vout1x0123 = _mm_mul_ps(vksum, vinput_zero_point1_float);
89 __m128 vinput_zero_point2_float = _mm_cvtepi32_ps(vinput_zero_point2);
90 __m128 vout2x0123 = _mm_mul_ps(vksum, vinput_zero_point2_float);
91 __m128 vinput_zero_point3_float = _mm_cvtepi32_ps(vinput_zero_point3);
92 __m128 vout3x0123 = _mm_mul_ps(vksum, vinput_zero_point3_float);
93 w = (const int32_t*) w + 4;
94
95 for (size_t kb=0; kb < kc; kb += bl) {
96 __m128i vacc0x0 = _mm_setzero_si128();
97 __m128i vacc0x1 = _mm_setzero_si128();
98 __m128i vacc0x2 = _mm_setzero_si128();
99 __m128i vacc0x3 = _mm_setzero_si128();
100 __m128i vacc1x0 = _mm_setzero_si128();
101 __m128i vacc1x1 = _mm_setzero_si128();
102 __m128i vacc1x2 = _mm_setzero_si128();
103 __m128i vacc1x3 = _mm_setzero_si128();
104 __m128i vacc2x0 = _mm_setzero_si128();
105 __m128i vacc2x1 = _mm_setzero_si128();
106 __m128i vacc2x2 = _mm_setzero_si128();
107 __m128i vacc2x3 = _mm_setzero_si128();
108 __m128i vacc3x0 = _mm_setzero_si128();
109 __m128i vacc3x1 = _mm_setzero_si128();
110 __m128i vacc3x2 = _mm_setzero_si128();
111 __m128i vacc3x3 = _mm_setzero_si128();
112 size_t k = bl;
113
114 while (k >= 16 * sizeof(int8_t)) {
115 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
116 const __m128i vxa0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c0, va0c0), 8);
117 a0 += 8;
118 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
119 const __m128i vxa1c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va1c0, va1c0), 8);
120 a1 += 8;
121 const __m128i va2c0 = _mm_loadl_epi64((const __m128i*) a2);
122 const __m128i vxa2c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va2c0, va2c0), 8);
123 a2 += 8;
124 const __m128i va3c0 = _mm_loadl_epi64((const __m128i*) a3);
125 const __m128i vxa3c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va3c0, va3c0), 8);
126 a3 += 8;
127
128 const __m128i vb0c01 = _mm_loadl_epi64((const __m128i*) w);
129 const __m128i vbs0c0 = _mm_slli_epi32(vb0c01, 4);
130 const __m128i vb0c0 = _mm_and_si128(vbs0c0, vmask);
131 const __m128i vxb0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0c0, vb0c0), 8);
132
133 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
134 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
135 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c0, vxb0c0));
136 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3c0, vxb0c0));
137 const __m128i vb1c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
138 const __m128i vbs1c0 = _mm_slli_epi32(vb1c01, 4);
139 const __m128i vb1c0 = _mm_and_si128(vbs1c0, vmask);
140 const __m128i vxb1c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1c0, vb1c0), 8);
141
142 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
143 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
144 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c0, vxb1c0));
145 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3c0, vxb1c0));
146 const __m128i vb2c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
147 const __m128i vbs2c0 = _mm_slli_epi32(vb2c01, 4);
148 const __m128i vb2c0 = _mm_and_si128(vbs2c0, vmask);
149 const __m128i vxb2c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2c0, vb2c0), 8);
150
151 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
152 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
153 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c0, vxb2c0));
154 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3c0, vxb2c0));
155 const __m128i vb3c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
156 const __m128i vbs3c0 = _mm_slli_epi32(vb3c01, 4);
157 const __m128i vb3c0 = _mm_and_si128(vbs3c0, vmask);
158 const __m128i vxb3c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3c0, vb3c0), 8);
159
160 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
161 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
162 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c0, vxb3c0));
163 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3c0, vxb3c0));
164
165 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
166 const __m128i vxa0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c1, va0c1), 8);
167 a0 += 8;
168 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
169 const __m128i vxa1c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1c1, va1c1), 8);
170 a1 += 8;
171 const __m128i va2c1 = _mm_loadl_epi64((const __m128i*) a2);
172 const __m128i vxa2c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va2c1, va2c1), 8);
173 a2 += 8;
174 const __m128i va3c1 = _mm_loadl_epi64((const __m128i*) a3);
175 const __m128i vxa3c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va3c1, va3c1), 8);
176 a3 += 8;
177
178 const __m128i vb0c1 = _mm_and_si128(vb0c01, vmask);
179 const __m128i vxb0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0c1, vb0c1), 8);
180
181 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
182 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
183 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c1, vxb0c1));
184 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3c1, vxb0c1));
185 const __m128i vb1c1 = _mm_and_si128(vb1c01, vmask);
186 const __m128i vxb1c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1c1, vb1c1), 8);
187
188 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
189 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
190 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c1, vxb1c1));
191 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3c1, vxb1c1));
192 const __m128i vb2c1 = _mm_and_si128(vb2c01, vmask);
193 const __m128i vxb2c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2c1, vb2c1), 8);
194
195 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
196 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
197 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c1, vxb2c1));
198 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3c1, vxb2c1));
199 const __m128i vb3c1 = _mm_and_si128(vb3c01, vmask);
200 const __m128i vxb3c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3c1, vb3c1), 8);
201
202 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
203 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
204 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c1, vxb3c1));
205 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3c1, vxb3c1));
206
207
208 w = (const int8_t*) w + 32;
209 k -= 16 * sizeof(int8_t);
210 }
211
212 while (k >= 8 * sizeof(int8_t)) {
213 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
214 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
215 a0 += 8;
216 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
217 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
218 a1 += 8;
219 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
220 const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
221 a2 += 8;
222 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
223 const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
224 a3 += 8;
225
226 __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
227 vb0 = _mm_slli_epi32(vb0, 4);
228 vb0 = _mm_and_si128(vb0, vmask);
229
230 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
231
232 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
233 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
234 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
235 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3, vxb0));
236 __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
237 vb1 = _mm_slli_epi32(vb1, 4);
238 vb1 = _mm_and_si128(vb1, vmask);
239
240 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
241
242 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
243 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
244 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
245 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3, vxb1));
246 __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
247 vb2 = _mm_slli_epi32(vb2, 4);
248 vb2 = _mm_and_si128(vb2, vmask);
249
250 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
251
252 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
253 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
254 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
255 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3, vxb2));
256 __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
257 vb3 = _mm_slli_epi32(vb3, 4);
258 vb3 = _mm_and_si128(vb3, vmask);
259
260 const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
261
262 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
263 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
264 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
265 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3, vxb3));
266
267 w = (const int8_t*) w + 32;
268 k -= 8 * sizeof(int8_t);
269 }
270 // accumulate float
271 const __m128 vfilter_output_scale0123 = _mm_castsi128_ps(_mm_unpacklo_epi16(_mm_setzero_si128(), _mm_loadl_epi64((const __m128i*) w)));
272 w = (const uint16_t*) w + 4;
273
274 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
275 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
276 const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
277 const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
278 const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2));
279 const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3));
280 const __m128i vacc3x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc3x0, vacc3x2), _mm_unpackhi_epi32(vacc3x0, vacc3x2));
281 const __m128i vacc3x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc3x1, vacc3x3), _mm_unpackhi_epi32(vacc3x1, vacc3x3));
282
283 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
284 __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
285 __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13));
286 __m128i vacc3x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc3x02, vacc3x13), _mm_unpackhi_epi32(vacc3x02, vacc3x13));
287
288 vout0x0123 = _mm_add_ps(vout0x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc0x0123), vfilter_output_scale0123));
289 vout1x0123 = _mm_add_ps(vout1x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc1x0123), vfilter_output_scale0123));
290 vout2x0123 = _mm_add_ps(vout2x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc2x0123), vfilter_output_scale0123));
291 vout3x0123 = _mm_add_ps(vout3x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc3x0123), vfilter_output_scale0123));
292 }
293
294 const __m128i vinput_scale01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
295 const __m128 vinput_scale0 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(1, 1, 1, 1)));
296 const __m128 vinput_scale1 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(3, 3, 3, 3)));
297 const __m128i vinput_scale23 = _mm_loadu_si128((const __m128i*) &quantization_params[2]);
298 const __m128 vinput_scale2 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale23, _MM_SHUFFLE(1, 1, 1, 1)));
299 const __m128 vinput_scale3 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale23, _MM_SHUFFLE(3, 3, 3, 3)));
300
301 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
302 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
303 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
304 vout3x0123 = _mm_mul_ps(vout3x0123, vinput_scale3);
305
306
307 const __m128 vbias0123 = _mm_loadu_ps((const float*) w);
308 w = (const float*) w + 4;
309 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
310 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
311 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
312 vout3x0123 = _mm_add_ps(vout3x0123, vbias0123);
313
314 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
315 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
316 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
317 vout3x0123 = _mm_max_ps(vout3x0123, vmin);
318
319 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
320 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
321 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
322 vout3x0123 = _mm_min_ps(vout3x0123, vmax);
323
324 if XNN_LIKELY(nc >= 4) {
325 _mm_storeu_ps(c0, vout0x0123);
326 _mm_storeu_ps(c1, vout1x0123);
327 _mm_storeu_ps(c2, vout2x0123);
328 _mm_storeu_ps(c3, vout3x0123);
329
330 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
331 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
332 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
333 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
334
335 c0 = (float*) ((uintptr_t) c0 + cn_stride);
336 c1 = (float*) ((uintptr_t) c1 + cn_stride);
337 c2 = (float*) ((uintptr_t) c2 + cn_stride);
338 c3 = (float*) ((uintptr_t) c3 + cn_stride);
339
340 nc -= 4;
341 } else {
342 if (nc & 2) {
343 _mm_storel_pi((__m64*) c0, vout0x0123);
344 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
345 c0 += 2;
346 _mm_storel_pi((__m64*) c1, vout1x0123);
347 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
348 c1 += 2;
349 _mm_storel_pi((__m64*) c2, vout2x0123);
350 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
351 c2 += 2;
352 _mm_storel_pi((__m64*) c3, vout3x0123);
353 vout3x0123 = _mm_unpackhi_ps(vout3x0123, vout3x0123);
354 c3 += 2;
355 }
356 if (nc & 1) {
357 _mm_store_ss(c0, vout0x0123);
358 _mm_store_ss(c1, vout1x0123);
359 _mm_store_ss(c2, vout2x0123);
360 _mm_store_ss(c3, vout3x0123);
361 }
362 nc = 0;
363 }
364 } while (nc != 0);
365 }