Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x8c8-minmax-avx2.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx8c8-avx2.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 <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
20
21 void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_4x8c8__avx2(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 const int8_t* restrict a,
26 size_t a_stride,
27 const void* restrict w,
28 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const struct xnn_f32_qb4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 4);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 size_t bl = params->scalar.blocksize;
45 assert(bl <= round_up_po2(kc, 16));
46 assert(bl != 0);
47 assert(bl % 32 == 0);
48 const int8_t* a0 = a;
49 float* c0 = c;
50 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
51 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
52 if XNN_UNPREDICTABLE(mr < 2) {
53 a1 = a0;
54 c1 = c0;
55 }
56 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
57 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
58 if XNN_UNPREDICTABLE(mr <= 2) {
59 a2 = a1;
60 c2 = c1;
61 }
62 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
63 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
64 if XNN_UNPREDICTABLE(mr != 4) {
65 a3 = a2;
66 c3 = c2;
67 }
68
69 const __m128i vmask = _mm_set1_epi8(0xF0);
70 XNN_FORCE_REALIZATION(vmask);
71 const __m256 vmin = _mm256_set1_ps(params->scalar.min);
72 const __m256 vmax = _mm256_set1_ps(params->scalar.max);
73 XNN_FORCE_REALIZATION(vmin);
74 XNN_FORCE_REALIZATION(vmax);
75
76 do {
77 const __m128 vinit0 = _mm_load_ss(&((const float*) w)[0]);
78 const __m128 vinit1 = _mm_load_ss(&((const float*) w)[1]);
79 const __m256 vinit01 = _mm256_insertf128_ps(_mm256_castps128_ps256(vinit0), vinit1, 1);
80 const __m128 vinit2 = _mm_load_ss(&((const float*) w)[2]);
81 const __m128 vinit3 = _mm_load_ss(&((const float*) w)[3]);
82 const __m256 vinit23 = _mm256_insertf128_ps(_mm256_castps128_ps256(vinit2), vinit3, 1);
83 const __m128 vinit4 = _mm_load_ss(&((const float*) w)[4]);
84 const __m128 vinit5 = _mm_load_ss(&((const float*) w)[5]);
85 const __m256 vinit45 = _mm256_insertf128_ps(_mm256_castps128_ps256(vinit4), vinit5, 1);
86 const __m128 vinit6 = _mm_load_ss(&((const float*) w)[6]);
87 const __m128 vinit7 = _mm_load_ss(&((const float*) w)[7]);
88 const __m256 vinit67 = _mm256_insertf128_ps(_mm256_castps128_ps256(vinit6), vinit7, 1);
89 const __m256 vinput_zero_point0 = _mm256_set1_ps((float) quantization_params[0].zero_point);
90 __m256 vout0x01 = _mm256_mul_ps(vinit01, vinput_zero_point0);
91 __m256 vout0x23 = _mm256_mul_ps(vinit23, vinput_zero_point0);
92 __m256 vout0x45 = _mm256_mul_ps(vinit45, vinput_zero_point0);
93 __m256 vout0x67 = _mm256_mul_ps(vinit67, vinput_zero_point0);
94 const __m256 vinput_zero_point1 = _mm256_set1_ps((float) quantization_params[1].zero_point);
95 __m256 vout1x01 = _mm256_mul_ps(vinit01, vinput_zero_point1);
96 __m256 vout1x23 = _mm256_mul_ps(vinit23, vinput_zero_point1);
97 __m256 vout1x45 = _mm256_mul_ps(vinit45, vinput_zero_point1);
98 __m256 vout1x67 = _mm256_mul_ps(vinit67, vinput_zero_point1);
99 const __m256 vinput_zero_point2 = _mm256_set1_ps((float) quantization_params[2].zero_point);
100 __m256 vout2x01 = _mm256_mul_ps(vinit01, vinput_zero_point2);
101 __m256 vout2x23 = _mm256_mul_ps(vinit23, vinput_zero_point2);
102 __m256 vout2x45 = _mm256_mul_ps(vinit45, vinput_zero_point2);
103 __m256 vout2x67 = _mm256_mul_ps(vinit67, vinput_zero_point2);
104 const __m256 vinput_zero_point3 = _mm256_set1_ps((float) quantization_params[3].zero_point);
105 __m256 vout3x01 = _mm256_mul_ps(vinit01, vinput_zero_point3);
106 __m256 vout3x23 = _mm256_mul_ps(vinit23, vinput_zero_point3);
107 __m256 vout3x45 = _mm256_mul_ps(vinit45, vinput_zero_point3);
108 __m256 vout3x67 = _mm256_mul_ps(vinit67, vinput_zero_point3);
109 w = (const int32_t*) w + 8;
110
111 for (size_t kb=0; kb < kc; kb += bl) {
112 __m256i vacc0x01 = _mm256_setzero_si256();
113 __m256i vacc0x23 = _mm256_setzero_si256();
114 __m256i vacc0x45 = _mm256_setzero_si256();
115 __m256i vacc0x67 = _mm256_setzero_si256();
116 __m256i vacc1x01 = _mm256_setzero_si256();
117 __m256i vacc1x23 = _mm256_setzero_si256();
118 __m256i vacc1x45 = _mm256_setzero_si256();
119 __m256i vacc1x67 = _mm256_setzero_si256();
120 __m256i vacc2x01 = _mm256_setzero_si256();
121 __m256i vacc2x23 = _mm256_setzero_si256();
122 __m256i vacc2x45 = _mm256_setzero_si256();
123 __m256i vacc2x67 = _mm256_setzero_si256();
124 __m256i vacc3x01 = _mm256_setzero_si256();
125 __m256i vacc3x23 = _mm256_setzero_si256();
126 __m256i vacc3x45 = _mm256_setzero_si256();
127 __m256i vacc3x67 = _mm256_setzero_si256();
128
129 size_t k = bl;
130 while (k >= 16 * sizeof(int8_t)) {
131 __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
132 __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
133 a0 += 8;
134 __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
135 __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
136 a1 += 8;
137 __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
138 __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
139 a2 += 8;
140 __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
141 __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
142 a3 += 8;
143
144 __m128i vb01 = _mm_load_si128((const __m128i*) w);
145 __m128i vbs01 = _mm_slli_epi32(vb01, 4);
146 __m128i vbm01 = _mm_and_si128(vbs01, vmask);
147 __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
148
149 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
150 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
151 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
152 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
153 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
154 __m128i vbs23 = _mm_slli_epi32(vb23, 4);
155 __m128i vbm23 = _mm_and_si128(vbs23, vmask);
156 __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
157
158 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
159 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
160 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
161 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
162 __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
163 __m128i vbs45 = _mm_slli_epi32(vb45, 4);
164 __m128i vbm45 = _mm_and_si128(vbs45, vmask);
165 __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
166
167 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
168 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
169 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
170 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
171 __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
172 __m128i vbs67 = _mm_slli_epi32(vb67, 4);
173 __m128i vbm67 = _mm_and_si128(vbs67, vmask);
174 __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
175
176 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
177 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
178 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
179 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
180
181 va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
182 vxa0 = _mm256_cvtepi8_epi16(va0);
183 a0 += 8;
184 va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
185 vxa1 = _mm256_cvtepi8_epi16(va1);
186 a1 += 8;
187 va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
188 vxa2 = _mm256_cvtepi8_epi16(va2);
189 a2 += 8;
190 va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
191 vxa3 = _mm256_cvtepi8_epi16(va3);
192 a3 += 8;
193
194 vbm01 = _mm_and_si128(vb01, vmask);
195 vxb01 = _mm256_cvtepi8_epi16(vbm01);
196
197 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
198 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
199 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
200 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
201 vbm23 = _mm_and_si128(vb23, vmask);
202 vxb23 = _mm256_cvtepi8_epi16(vbm23);
203
204 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
205 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
206 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
207 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
208 vbm45 = _mm_and_si128(vb45, vmask);
209 vxb45 = _mm256_cvtepi8_epi16(vbm45);
210
211 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
212 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
213 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
214 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
215 vbm67 = _mm_and_si128(vb67, vmask);
216 vxb67 = _mm256_cvtepi8_epi16(vbm67);
217
218 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
219 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
220 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
221 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
222
223 w = (const int8_t*) w + 64;
224 k -= 16 * sizeof(int8_t);
225 }
226
227 while (k >= 8 * sizeof(int8_t)) {
228 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
229 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
230 a0 += 8;
231 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
232 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
233 a1 += 8;
234 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
235 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
236 a2 += 8;
237 const __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
238 const __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
239 a3 += 8;
240
241 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
242 const __m128i vbs01 = _mm_slli_epi32(vb01, 4);
243 const __m128i vbm01 = _mm_and_si128(vbs01, vmask);
244 const __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
245
246 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
247 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
248 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
249 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
250 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
251 const __m128i vbs23 = _mm_slli_epi32(vb23, 4);
252 const __m128i vbm23 = _mm_and_si128(vbs23, vmask);
253 const __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
254
255 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
256 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
257 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
258 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
259 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
260 const __m128i vbs45 = _mm_slli_epi32(vb45, 4);
261 const __m128i vbm45 = _mm_and_si128(vbs45, vmask);
262 const __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
263
264 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
265 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
266 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
267 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
268 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
269 const __m128i vbs67 = _mm_slli_epi32(vb67, 4);
270 const __m128i vbm67 = _mm_and_si128(vbs67, vmask);
271 const __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
272
273 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
274 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
275 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
276 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
277
278 w = (const int8_t*) w + 64;
279 k -= 8 * sizeof(int8_t);
280 }
281
282 const __m128 vfilter_output_scale0 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[0] << 16));
283 const __m128 vfilter_output_scale1 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[1] << 16));
284 const __m256 vfilter_output_scale01 = _mm256_insertf128_ps(
285 _mm256_castps128_ps256(vfilter_output_scale0), vfilter_output_scale1, 1);
286 vout0x01 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc0x01), vfilter_output_scale01, vout0x01);
287 vout1x01 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc1x01), vfilter_output_scale01, vout1x01);
288 vout2x01 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc2x01), vfilter_output_scale01, vout2x01);
289 vout3x01 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc3x01), vfilter_output_scale01, vout3x01);
290 const __m128 vfilter_output_scale2 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[2] << 16));
291 const __m128 vfilter_output_scale3 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[3] << 16));
292 const __m256 vfilter_output_scale23 = _mm256_insertf128_ps(
293 _mm256_castps128_ps256(vfilter_output_scale2), vfilter_output_scale3, 1);
294 vout0x23 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc0x23), vfilter_output_scale23, vout0x23);
295 vout1x23 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc1x23), vfilter_output_scale23, vout1x23);
296 vout2x23 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc2x23), vfilter_output_scale23, vout2x23);
297 vout3x23 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc3x23), vfilter_output_scale23, vout3x23);
298 const __m128 vfilter_output_scale4 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[4] << 16));
299 const __m128 vfilter_output_scale5 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[5] << 16));
300 const __m256 vfilter_output_scale45 = _mm256_insertf128_ps(
301 _mm256_castps128_ps256(vfilter_output_scale4), vfilter_output_scale5, 1);
302 vout0x45 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc0x45), vfilter_output_scale45, vout0x45);
303 vout1x45 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc1x45), vfilter_output_scale45, vout1x45);
304 vout2x45 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc2x45), vfilter_output_scale45, vout2x45);
305 vout3x45 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc3x45), vfilter_output_scale45, vout3x45);
306 const __m128 vfilter_output_scale6 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[6] << 16));
307 const __m128 vfilter_output_scale7 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[7] << 16));
308 const __m256 vfilter_output_scale67 = _mm256_insertf128_ps(
309 _mm256_castps128_ps256(vfilter_output_scale6), vfilter_output_scale7, 1);
310 vout0x67 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc0x67), vfilter_output_scale67, vout0x67);
311 vout1x67 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc1x67), vfilter_output_scale67, vout1x67);
312 vout2x67 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc2x67), vfilter_output_scale67, vout2x67);
313 vout3x67 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc3x67), vfilter_output_scale67, vout3x67);
314
315 w = (const uint16_t*) w + 8;
316 }
317
318 const __m256 vout0x0213 = _mm256_hadd_ps(vout0x01, vout0x23);
319 const __m256 vout0x4657 = _mm256_hadd_ps(vout0x45, vout0x67);
320 const __m256 vout1x0213 = _mm256_hadd_ps(vout1x01, vout1x23);
321 const __m256 vout1x4657 = _mm256_hadd_ps(vout1x45, vout1x67);
322 const __m256 vout2x0213 = _mm256_hadd_ps(vout2x01, vout2x23);
323 const __m256 vout2x4657 = _mm256_hadd_ps(vout2x45, vout2x67);
324 const __m256 vout3x0213 = _mm256_hadd_ps(vout3x01, vout3x23);
325 const __m256 vout3x4657 = _mm256_hadd_ps(vout3x45, vout3x67);
326
327 const __m256 vout0x02461357 = _mm256_hadd_ps(vout0x0213, vout0x4657);
328 const __m256 vout1x02461357 = _mm256_hadd_ps(vout1x0213, vout1x4657);
329 const __m256 vout2x02461357 = _mm256_hadd_ps(vout2x0213, vout2x4657);
330 const __m256 vout3x02461357 = _mm256_hadd_ps(vout3x0213, vout3x4657);
331
332 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
333 __m256 vout0x01234567 = _mm256_permutevar8x32_ps(vout0x02461357, vpermute_mask);
334 __m256 vout1x01234567 = _mm256_permutevar8x32_ps(vout1x02461357, vpermute_mask);
335 __m256 vout2x01234567 = _mm256_permutevar8x32_ps(vout2x02461357, vpermute_mask);
336 __m256 vout3x01234567 = _mm256_permutevar8x32_ps(vout3x02461357, vpermute_mask);
337
338 const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale);
339 const __m256 vinput_scale1 = _mm256_broadcast_ss(&quantization_params[1].inv_scale);
340 const __m256 vinput_scale2 = _mm256_broadcast_ss(&quantization_params[2].inv_scale);
341 const __m256 vinput_scale3 = _mm256_broadcast_ss(&quantization_params[3].inv_scale);
342
343 const __m256 vbias01234567 = _mm256_loadu_ps((const float*) w);
344 w = (const float*) w + 8;
345 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vinput_scale0, vbias01234567);
346 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vinput_scale1, vbias01234567);
347 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vinput_scale2, vbias01234567);
348 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vinput_scale3, vbias01234567);
349
350
351
352 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
353 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
354 vout2x01234567 = _mm256_max_ps(vout2x01234567, vmin);
355 vout3x01234567 = _mm256_max_ps(vout3x01234567, vmin);
356
357 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
358 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
359 vout2x01234567 = _mm256_min_ps(vout2x01234567, vmax);
360 vout3x01234567 = _mm256_min_ps(vout3x01234567, vmax);
361
362 if XNN_LIKELY(nc >= 8) {
363 _mm256_storeu_ps(c0, vout0x01234567);
364 c0 = (float*) ((uintptr_t) c0 + cn_stride);
365 _mm256_storeu_ps(c1, vout1x01234567);
366 c1 = (float*) ((uintptr_t) c1 + cn_stride);
367 _mm256_storeu_ps(c2, vout2x01234567);
368 c2 = (float*) ((uintptr_t) c2 + cn_stride);
369 _mm256_storeu_ps(c3, vout3x01234567);
370 c3 = (float*) ((uintptr_t) c3 + cn_stride);
371
372 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
373 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
374 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
375 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
376
377 nc -= 8;
378 } else {
379 __m128 vout0x0123 = _mm256_castps256_ps128(vout0x01234567);
380 __m128 vout1x0123 = _mm256_castps256_ps128(vout1x01234567);
381 __m128 vout2x0123 = _mm256_castps256_ps128(vout2x01234567);
382 __m128 vout3x0123 = _mm256_castps256_ps128(vout3x01234567);
383 if (nc & 4) {
384 _mm_storeu_ps(c0, vout0x0123);
385 _mm_storeu_ps(c1, vout1x0123);
386 _mm_storeu_ps(c2, vout2x0123);
387 _mm_storeu_ps(c3, vout3x0123);
388
389 vout0x0123 = _mm256_extractf128_ps(vout0x01234567, 1);
390 vout1x0123 = _mm256_extractf128_ps(vout1x01234567, 1);
391 vout2x0123 = _mm256_extractf128_ps(vout2x01234567, 1);
392 vout3x0123 = _mm256_extractf128_ps(vout3x01234567, 1);
393
394 c0 += 4;
395 c1 += 4;
396 c2 += 4;
397 c3 += 4;
398 }
399 if (nc & 2) {
400 _mm_storel_pi((__m64*) c0, vout0x0123);
401 _mm_storel_pi((__m64*) c1, vout1x0123);
402 _mm_storel_pi((__m64*) c2, vout2x0123);
403 _mm_storel_pi((__m64*) c3, vout3x0123);
404
405 vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123);
406 vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123);
407 vout2x0123 = _mm_movehl_ps(vout2x0123, vout2x0123);
408 vout3x0123 = _mm_movehl_ps(vout3x0123, vout3x0123);
409
410 c0 += 2;
411 c1 += 2;
412 c2 += 2;
413 c3 += 2;
414 }
415 if (nc & 1) {
416 _mm_store_ss(c0, vout0x0123);
417 _mm_store_ss(c1, vout1x0123);
418 _mm_store_ss(c2, vout2x0123);
419 _mm_store_ss(c3, vout3x0123);
420 }
421 nc = 0;
422 }
423 } while (nc != 0);
424 }