Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-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_qc4w_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_qc4w_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 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 const int8_t* a0 = a;
44 float* c0 = c;
45 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
46 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 a1 = a0;
49 c1 = c0;
50 }
51 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
52 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 a2 = a1;
55 c2 = c1;
56 }
57 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
58 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
59 if XNN_UNPREDICTABLE(mr != 4) {
60 a3 = a2;
61 c3 = c2;
62 }
63
64 const __m128 vmin = _mm_set1_ps(params->scalar.min);
65 const __m128 vmax = _mm_set1_ps(params->scalar.max);
66 XNN_FORCE_REALIZATION(vmin);
67 XNN_FORCE_REALIZATION(vmax);
68
69 const __m128i vmask = _mm_set1_epi8(0xF0);
70 XNN_FORCE_REALIZATION(vmask);
71
72 do {
73 const __m128i vksum = _mm_load_si128((const __m128i*) w);
74 const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
75 const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
76 const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));
77 const __m128i vinput_zero_point23 = _mm_loadu_si128((const __m128i*) &quantization_params[2]);
78 const __m128i vinput_zero_point2 = _mm_shuffle_epi32(vinput_zero_point23, _MM_SHUFFLE(0, 0, 0, 0));
79 const __m128i vinput_zero_point3 = _mm_shuffle_epi32(vinput_zero_point23, _MM_SHUFFLE(2, 2, 2, 2));
80 const __m128i vzero = _mm_setzero_si128();
81 const __m128i vksum_lo = _mm_srli_epi32(_mm_slli_epi32(vksum, 16), 16);
82 const __m128i vksum_hi = _mm_srli_epi32(vksum, 16);
83
84 __m128i vzpprodksumhi0 = _mm_mulhi_epu16(vinput_zero_point0, vksum_lo);
85 const __m128i vzpprodksumlo0 = _mm_mullo_epi16(vinput_zero_point0, vksum_lo);
86 __m128i vzpprodksumhi1 = _mm_mulhi_epu16(vinput_zero_point1, vksum_lo);
87 const __m128i vzpprodksumlo1 = _mm_mullo_epi16(vinput_zero_point1, vksum_lo);
88 __m128i vzpprodksumhi2 = _mm_mulhi_epu16(vinput_zero_point2, vksum_lo);
89 const __m128i vzpprodksumlo2 = _mm_mullo_epi16(vinput_zero_point2, vksum_lo);
90 __m128i vzpprodksumhi3 = _mm_mulhi_epu16(vinput_zero_point3, vksum_lo);
91 const __m128i vzpprodksumlo3 = _mm_mullo_epi16(vinput_zero_point3, vksum_lo);
92
93 vzpprodksumhi0 = _mm_add_epi16(vzpprodksumhi0, _mm_mullo_epi16(vinput_zero_point0, vksum_hi));
94 vzpprodksumhi0 = _mm_sub_epi16(vzpprodksumhi0, _mm_and_si128(_mm_srai_epi16(vinput_zero_point0, 15), vksum_lo));
95 vzpprodksumhi1 = _mm_add_epi16(vzpprodksumhi1, _mm_mullo_epi16(vinput_zero_point1, vksum_hi));
96 vzpprodksumhi1 = _mm_sub_epi16(vzpprodksumhi1, _mm_and_si128(_mm_srai_epi16(vinput_zero_point1, 15), vksum_lo));
97 vzpprodksumhi2 = _mm_add_epi16(vzpprodksumhi2, _mm_mullo_epi16(vinput_zero_point2, vksum_hi));
98 vzpprodksumhi2 = _mm_sub_epi16(vzpprodksumhi2, _mm_and_si128(_mm_srai_epi16(vinput_zero_point2, 15), vksum_lo));
99 vzpprodksumhi3 = _mm_add_epi16(vzpprodksumhi3, _mm_mullo_epi16(vinput_zero_point3, vksum_hi));
100 vzpprodksumhi3 = _mm_sub_epi16(vzpprodksumhi3, _mm_and_si128(_mm_srai_epi16(vinput_zero_point3, 15), vksum_lo));
101
102 vzpprodksumhi0 = _mm_slli_si128(vzpprodksumhi0, 2);
103 vzpprodksumhi1 = _mm_slli_si128(vzpprodksumhi1, 2);
104 vzpprodksumhi2 = _mm_slli_si128(vzpprodksumhi2, 2);
105 vzpprodksumhi3 = _mm_slli_si128(vzpprodksumhi3, 2);
106
107 const __m128i vksumzp0 = _mm_or_si128(vzpprodksumhi0, vzpprodksumlo0);
108 const __m128i vksumzp1 = _mm_or_si128(vzpprodksumhi1, vzpprodksumlo1);
109 const __m128i vksumzp2 = _mm_or_si128(vzpprodksumhi2, vzpprodksumlo2);
110 const __m128i vksumzp3 = _mm_or_si128(vzpprodksumhi3, vzpprodksumlo3);
111
112 const __m128i vksum010 = _mm_unpacklo_epi32(vksumzp0, vzero);
113 const __m128i vksum230 = _mm_unpackhi_epi32(vksumzp0, vzero);
114 const __m128i vksum011 = _mm_unpacklo_epi32(vksumzp1, vzero);
115 const __m128i vksum231 = _mm_unpackhi_epi32(vksumzp1, vzero);
116 const __m128i vksum012 = _mm_unpacklo_epi32(vksumzp2, vzero);
117 const __m128i vksum232 = _mm_unpackhi_epi32(vksumzp2, vzero);
118 const __m128i vksum013 = _mm_unpacklo_epi32(vksumzp3, vzero);
119 const __m128i vksum233 = _mm_unpackhi_epi32(vksumzp3, vzero);
120
121 __m128i vacc0x0 = _mm_unpacklo_epi64(vksum010, vzero);
122 __m128i vacc0x1 = _mm_unpackhi_epi64(vksum010, vzero);
123 __m128i vacc1x0 = _mm_unpacklo_epi64(vksum011, vzero);
124 __m128i vacc1x1 = _mm_unpackhi_epi64(vksum011, vzero);
125 __m128i vacc2x0 = _mm_unpacklo_epi64(vksum012, vzero);
126 __m128i vacc2x1 = _mm_unpackhi_epi64(vksum012, vzero);
127 __m128i vacc3x0 = _mm_unpacklo_epi64(vksum013, vzero);
128 __m128i vacc3x1 = _mm_unpackhi_epi64(vksum013, vzero);
129
130 __m128i vacc0x2 = _mm_unpacklo_epi64(vksum230, vzero);
131 __m128i vacc0x3 = _mm_unpackhi_epi64(vksum230, vzero);
132 __m128i vacc1x2 = _mm_unpacklo_epi64(vksum231, vzero);
133 __m128i vacc1x3 = _mm_unpackhi_epi64(vksum231, vzero);
134 __m128i vacc2x2 = _mm_unpacklo_epi64(vksum232, vzero);
135 __m128i vacc2x3 = _mm_unpackhi_epi64(vksum232, vzero);
136 __m128i vacc3x2 = _mm_unpacklo_epi64(vksum233, vzero);
137 __m128i vacc3x3 = _mm_unpackhi_epi64(vksum233, vzero);
138 w = (const int32_t*) w + 4;
139
140 size_t k = kc;
141
142 while (k >= 16 * sizeof(int8_t)) {
143 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
144 const __m128i vxa0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c0, va0c0), 8);
145 a0 += 8;
146 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
147 const __m128i vxa1c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va1c0, va1c0), 8);
148 a1 += 8;
149 const __m128i va2c0 = _mm_loadl_epi64((const __m128i*) a2);
150 const __m128i vxa2c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va2c0, va2c0), 8);
151 a2 += 8;
152 const __m128i va3c0 = _mm_loadl_epi64((const __m128i*) a3);
153 const __m128i vxa3c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va3c0, va3c0), 8);
154 a3 += 8;
155
156 const __m128i vb0c01 = _mm_loadl_epi64((const __m128i*) w);
157 const __m128i vbs0c0 = _mm_slli_epi32(vb0c01, 4);
158 const __m128i vb0c0 = _mm_and_si128(vbs0c0, vmask);
159 const __m128i vxb0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0c0, vb0c0), 8);
160
161 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
162 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
163 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c0, vxb0c0));
164 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3c0, vxb0c0));
165 const __m128i vb1c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
166 const __m128i vbs1c0 = _mm_slli_epi32(vb1c01, 4);
167 const __m128i vb1c0 = _mm_and_si128(vbs1c0, vmask);
168 const __m128i vxb1c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1c0, vb1c0), 8);
169
170 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
171 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
172 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c0, vxb1c0));
173 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3c0, vxb1c0));
174 const __m128i vb2c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
175 const __m128i vbs2c0 = _mm_slli_epi32(vb2c01, 4);
176 const __m128i vb2c0 = _mm_and_si128(vbs2c0, vmask);
177 const __m128i vxb2c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2c0, vb2c0), 8);
178
179 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
180 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
181 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c0, vxb2c0));
182 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3c0, vxb2c0));
183 const __m128i vb3c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
184 const __m128i vbs3c0 = _mm_slli_epi32(vb3c01, 4);
185 const __m128i vb3c0 = _mm_and_si128(vbs3c0, vmask);
186 const __m128i vxb3c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3c0, vb3c0), 8);
187
188 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
189 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
190 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c0, vxb3c0));
191 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3c0, vxb3c0));
192
193 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
194 const __m128i vxa0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c1, va0c1), 8);
195 a0 += 8;
196 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
197 const __m128i vxa1c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1c1, va1c1), 8);
198 a1 += 8;
199 const __m128i va2c1 = _mm_loadl_epi64((const __m128i*) a2);
200 const __m128i vxa2c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va2c1, va2c1), 8);
201 a2 += 8;
202 const __m128i va3c1 = _mm_loadl_epi64((const __m128i*) a3);
203 const __m128i vxa3c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va3c1, va3c1), 8);
204 a3 += 8;
205
206 const __m128i vb0c1 = _mm_and_si128(vb0c01, vmask);
207 const __m128i vxb0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0c1, vb0c1), 8);
208
209 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
210 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
211 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c1, vxb0c1));
212 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3c1, vxb0c1));
213 const __m128i vb1c1 = _mm_and_si128(vb1c01, vmask);
214 const __m128i vxb1c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1c1, vb1c1), 8);
215
216 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
217 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
218 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c1, vxb1c1));
219 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3c1, vxb1c1));
220 const __m128i vb2c1 = _mm_and_si128(vb2c01, vmask);
221 const __m128i vxb2c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2c1, vb2c1), 8);
222
223 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
224 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
225 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c1, vxb2c1));
226 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3c1, vxb2c1));
227 const __m128i vb3c1 = _mm_and_si128(vb3c01, vmask);
228 const __m128i vxb3c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3c1, vb3c1), 8);
229
230 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
231 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
232 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c1, vxb3c1));
233 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3c1, vxb3c1));
234
235
236 w = (const int8_t*) w + 32;
237 k -= 16 * sizeof(int8_t);
238 }
239
240 while (k >= 8 * sizeof(int8_t)) {
241 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
242 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
243 a0 += 8;
244 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
245 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
246 a1 += 8;
247 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
248 const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
249 a2 += 8;
250 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
251 const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
252 a3 += 8;
253
254 __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
255 vb0 = _mm_slli_epi32(vb0, 4);
256 vb0 = _mm_and_si128(vb0, vmask);
257
258 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
259
260 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
261 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
262 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
263 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3, vxb0));
264 __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
265 vb1 = _mm_slli_epi32(vb1, 4);
266 vb1 = _mm_and_si128(vb1, vmask);
267
268 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
269
270 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
271 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
272 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
273 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3, vxb1));
274 __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
275 vb2 = _mm_slli_epi32(vb2, 4);
276 vb2 = _mm_and_si128(vb2, vmask);
277
278 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
279
280 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
281 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
282 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
283 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3, vxb2));
284 __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
285 vb3 = _mm_slli_epi32(vb3, 4);
286 vb3 = _mm_and_si128(vb3, vmask);
287
288 const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
289
290 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
291 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
292 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
293 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3, vxb3));
294
295 w = (const int8_t*) w + 32;
296 k -= 8 * sizeof(int8_t);
297 }
298
299 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
300 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
301 const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
302 const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
303 const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2));
304 const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3));
305 const __m128i vacc3x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc3x0, vacc3x2), _mm_unpackhi_epi32(vacc3x0, vacc3x2));
306 const __m128i vacc3x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc3x1, vacc3x3), _mm_unpackhi_epi32(vacc3x1, vacc3x3));
307
308 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
309 __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
310 __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13));
311 __m128i vacc3x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc3x02, vacc3x13), _mm_unpackhi_epi32(vacc3x02, vacc3x13));
312
313 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
314 vacc1x0123 = _mm_srai_epi32(vacc1x0123, 4);
315 vacc2x0123 = _mm_srai_epi32(vacc2x0123, 4);
316 vacc3x0123 = _mm_srai_epi32(vacc3x0123, 4);
317 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
318 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
319 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
320 __m128 vout3x0123 = _mm_cvtepi32_ps(vacc3x0123);
321
322 const __m128i vinput_scale01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
323 const __m128 vinput_scale0 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(1, 1, 1, 1)));
324 const __m128 vinput_scale1 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(3, 3, 3, 3)));
325 const __m128i vinput_scale23 = _mm_loadu_si128((const __m128i*) &quantization_params[2]);
326 const __m128 vinput_scale2 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale23, _MM_SHUFFLE(1, 1, 1, 1)));
327 const __m128 vinput_scale3 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale23, _MM_SHUFFLE(3, 3, 3, 3)));
328
329 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
330 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
331 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
332 vout3x0123 = _mm_mul_ps(vout3x0123, vinput_scale3);
333
334 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
335 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
336 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
337 vout2x0123 = _mm_mul_ps(vout2x0123, vfilter_output_scale0123);
338 vout3x0123 = _mm_mul_ps(vout3x0123, vfilter_output_scale0123);
339
340 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
341 w = (const float*) w + 8;
342 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
343 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
344 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
345 vout3x0123 = _mm_add_ps(vout3x0123, vbias0123);
346
347 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
348 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
349 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
350 vout3x0123 = _mm_max_ps(vout3x0123, vmin);
351
352 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
353 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
354 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
355 vout3x0123 = _mm_min_ps(vout3x0123, vmax);
356
357 if XNN_LIKELY(nc >= 4) {
358 _mm_storeu_ps(c0, vout0x0123);
359 _mm_storeu_ps(c1, vout1x0123);
360 _mm_storeu_ps(c2, vout2x0123);
361 _mm_storeu_ps(c3, vout3x0123);
362
363 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
364 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
365 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
366 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
367
368 c0 = (float*) ((uintptr_t) c0 + cn_stride);
369 c1 = (float*) ((uintptr_t) c1 + cn_stride);
370 c2 = (float*) ((uintptr_t) c2 + cn_stride);
371 c3 = (float*) ((uintptr_t) c3 + cn_stride);
372
373 nc -= 4;
374 } else {
375 if (nc & 2) {
376 _mm_storel_pi((__m64*) c0, vout0x0123);
377 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
378 c0 += 2;
379 _mm_storel_pi((__m64*) c1, vout1x0123);
380 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
381 c1 += 2;
382 _mm_storel_pi((__m64*) c2, vout2x0123);
383 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
384 c2 += 2;
385 _mm_storel_pi((__m64*) c3, vout3x0123);
386 vout3x0123 = _mm_unpackhi_ps(vout3x0123, vout3x0123);
387 c3 += 2;
388 }
389 if (nc & 1) {
390 _mm_store_ss(c0, vout0x0123);
391 _mm_store_ss(c1, vout1x0123);
392 _mm_store_ss(c2, vout2x0123);
393 _mm_store_ss(c3, vout3x0123);
394 }
395 nc = 0;
396 }
397 } while (nc != 0);
398 }