Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-8x8c8-minmax-avxvnni-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx8c8-avxvnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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/gemm.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18 #include "xnnpack/prefetch.h"
19
20
21 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_8x8c8__avxvnni_prfm(
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 union xnn_f32_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 <= 8);
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 const int8_t* a0 = a;
45 float* c0 = c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 a1 = a0;
50 c1 = c0;
51 }
52 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
53 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 a2 = a1;
56 c2 = c1;
57 }
58 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
59 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
60 if XNN_UNPREDICTABLE(mr < 4) {
61 a3 = a2;
62 c3 = c2;
63 }
64 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
65 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
66 if XNN_UNPREDICTABLE(mr <= 4) {
67 a4 = a3;
68 c4 = c3;
69 }
70 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
71 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
72 if XNN_UNPREDICTABLE(mr < 6) {
73 a5 = a4;
74 c5 = c4;
75 }
76 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
77 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
78 if XNN_UNPREDICTABLE(mr <= 6) {
79 a6 = a5;
80 c6 = c5;
81 }
82 const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
83 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
84 if XNN_UNPREDICTABLE(mr != 8) {
85 a7 = a6;
86 c7 = c6;
87 }
88
89 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
90 XNN_FORCE_REALIZATION(vsign_mask);
91 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
92 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
93 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
94 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
95 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
96 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point + 128);
97 const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point + 128);
98 const __m256i vinput_zero_point7 = _mm256_set1_epi32((int) quantization_params[7].zero_point + 128);
99 const __m256 voutput_min = _mm256_set1_ps(params->scalar.min);
100 const __m256 voutput_max = _mm256_set1_ps(params->scalar.max);
101 // XNN_FORCE_REALIZATION(voutput_min);
102 // XNN_FORCE_REALIZATION(voutput_max);
103 do {
104 const __m256i vksum01234567 = _mm256_load_si256(w);
105 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
106 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
107 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
108 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
109 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
110 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
111 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
112 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
113 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
114 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
115 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
116 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
117 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
118 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
119 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
120 __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
121 __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
122 __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
123 __m256i vsum6x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point6);
124 __m256i vacc6x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 0));
125 __m256i vacc6x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 1));
126 __m256i vsum7x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point7);
127 __m256i vacc7x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 0));
128 __m256i vacc7x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 1));
129 w = (const int32_t*) w + 8;
130
131 size_t k = kc;
132 while (k >= 16 * sizeof(int8_t)) {
133 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
134 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
135 a0 += 16;
136 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
137 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
138 a1 += 16;
139 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
140 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
141 a2 += 16;
142 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
143 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
144 a3 += 16;
145 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
146 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
147 a4 += 16;
148 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
149 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
150 a5 += 16;
151 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
152 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
153 a6 += 16;
154 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
155 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
156 a7 += 16;
157
158 const __m256i vb01234567x0123 = _mm256_load_si256(w);
159 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
160 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
161 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
162 xnn_prefetch_to_l1((const int8_t*) w + 896);
163
164 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
165 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
166 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
167 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
168 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
169 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
170 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
171 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
172 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
173 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
174 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
175 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
176 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
177 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
178 vacc7x0123 = _mm256_dpbusd_avx_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
179 vacc7x4567 = _mm256_dpbusd_avx_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
180 xnn_prefetch_to_l1((const int8_t*) w + 960);
181 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
182 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
183 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
184 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
185 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
186 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
187 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
188 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
189 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
190 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
191 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
192 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
193 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
194 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
195 vacc7x0123 = _mm256_dpbusd_avx_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
196 vacc7x4567 = _mm256_dpbusd_avx_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
197
198 w = (const int8_t*) w + 128;
199 k -= 16 * sizeof(int8_t);
200 }
201
202 if (k != 0) {
203 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
204 a0 += 8;
205 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
206 a1 += 8;
207 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
208 a2 += 8;
209 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
210 a3 += 8;
211 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
212 a4 += 8;
213 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
214 a5 += 8;
215 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
216 a6 += 8;
217 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
218 a7 += 8;
219
220 const __m256i vb01234567x0123 = _mm256_load_si256(w);
221 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
222
223 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
224 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
225 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
226 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
227 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
228 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
229 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
230 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
231 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
232 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
233 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
234 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
235 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
236 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
237 vacc7x0123 = _mm256_dpbusd_avx_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
238 vacc7x4567 = _mm256_dpbusd_avx_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
239 xnn_prefetch_to_l1((const int8_t*) w + 960);
240
241 w = (const int8_t*) w + 64;
242 k -= 8 * sizeof(int8_t);
243 }
244
245 // Add adjacent pairs
246 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
247 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
248 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
249 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
250 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
251 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
252 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
253 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
254 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
255 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
256 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
257 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
258 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
259 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
260 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
261 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
262
263 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
264 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
265 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
266 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
267 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
268 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
269 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
270 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
271
272 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
273 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
274 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
275 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
276 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
277 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
278 vout6x01234567 = _mm256_mul_ps(vout6x01234567, _mm256_set1_ps(quantization_params[6].inv_scale));
279 vout7x01234567 = _mm256_mul_ps(vout7x01234567, _mm256_set1_ps(quantization_params[7].inv_scale));
280
281 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
282 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
283 w = (const float*) w + 16;
284
285 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
286 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
287 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
288 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
289 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
290 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
291 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
292 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
293
294 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
295 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
296 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
297 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
298 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
299 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
300 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
301 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
302
303 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
304 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
305 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
306 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
307 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
308 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
309 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
310 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
311
312 if XNN_LIKELY(nc >= 8) {
313 _mm256_storeu_ps(c0, vout0x01234567);
314 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
315 c0 = (float*) ((uintptr_t) c0 + cn_stride);
316 _mm256_storeu_ps(c1, vout1x01234567);
317 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
318 c1 = (float*) ((uintptr_t) c1 + cn_stride);
319 _mm256_storeu_ps(c2, vout2x01234567);
320 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
321 c2 = (float*) ((uintptr_t) c2 + cn_stride);
322 _mm256_storeu_ps(c3, vout3x01234567);
323 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
324 c3 = (float*) ((uintptr_t) c3 + cn_stride);
325 _mm256_storeu_ps(c4, vout4x01234567);
326 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
327 c4 = (float*) ((uintptr_t) c4 + cn_stride);
328 _mm256_storeu_ps(c5, vout5x01234567);
329 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
330 c5 = (float*) ((uintptr_t) c5 + cn_stride);
331 _mm256_storeu_ps(c6, vout6x01234567);
332 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
333 c6 = (float*) ((uintptr_t) c6 + cn_stride);
334 _mm256_storeu_ps(c7, vout7x01234567);
335 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
336 c7 = (float*) ((uintptr_t) c7 + cn_stride);
337 nc -= 8;
338 } else {
339 __m128 vout0x0123 = _mm256_castps256_ps128(vout0x01234567);
340 __m128 vout1x0123 = _mm256_castps256_ps128(vout1x01234567);
341 __m128 vout2x0123 = _mm256_castps256_ps128(vout2x01234567);
342 __m128 vout3x0123 = _mm256_castps256_ps128(vout3x01234567);
343 __m128 vout4x0123 = _mm256_castps256_ps128(vout4x01234567);
344 __m128 vout5x0123 = _mm256_castps256_ps128(vout5x01234567);
345 __m128 vout6x0123 = _mm256_castps256_ps128(vout6x01234567);
346 __m128 vout7x0123 = _mm256_castps256_ps128(vout7x01234567);
347 if (nc & 4) {
348 _mm_storeu_ps(c0, vout0x0123);
349 c0 += 4;
350 vout0x0123 = _mm256_extractf128_ps(vout0x01234567, 1);
351 _mm_storeu_ps(c1, vout1x0123);
352 c1 += 4;
353 vout1x0123 = _mm256_extractf128_ps(vout1x01234567, 1);
354 _mm_storeu_ps(c2, vout2x0123);
355 c2 += 4;
356 vout2x0123 = _mm256_extractf128_ps(vout2x01234567, 1);
357 _mm_storeu_ps(c3, vout3x0123);
358 c3 += 4;
359 vout3x0123 = _mm256_extractf128_ps(vout3x01234567, 1);
360 _mm_storeu_ps(c4, vout4x0123);
361 c4 += 4;
362 vout4x0123 = _mm256_extractf128_ps(vout4x01234567, 1);
363 _mm_storeu_ps(c5, vout5x0123);
364 c5 += 4;
365 vout5x0123 = _mm256_extractf128_ps(vout5x01234567, 1);
366 _mm_storeu_ps(c6, vout6x0123);
367 c6 += 4;
368 vout6x0123 = _mm256_extractf128_ps(vout6x01234567, 1);
369 _mm_storeu_ps(c7, vout7x0123);
370 c7 += 4;
371 vout7x0123 = _mm256_extractf128_ps(vout7x01234567, 1);
372 }
373 if (nc & 2) {
374 _mm_storel_pi((__m64*) c0, vout0x0123);
375 c0 += 2;
376 vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123);
377 _mm_storel_pi((__m64*) c1, vout1x0123);
378 c1 += 2;
379 vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123);
380 _mm_storel_pi((__m64*) c2, vout2x0123);
381 c2 += 2;
382 vout2x0123 = _mm_movehl_ps(vout2x0123, vout2x0123);
383 _mm_storel_pi((__m64*) c3, vout3x0123);
384 c3 += 2;
385 vout3x0123 = _mm_movehl_ps(vout3x0123, vout3x0123);
386 _mm_storel_pi((__m64*) c4, vout4x0123);
387 c4 += 2;
388 vout4x0123 = _mm_movehl_ps(vout4x0123, vout4x0123);
389 _mm_storel_pi((__m64*) c5, vout5x0123);
390 c5 += 2;
391 vout5x0123 = _mm_movehl_ps(vout5x0123, vout5x0123);
392 _mm_storel_pi((__m64*) c6, vout6x0123);
393 c6 += 2;
394 vout6x0123 = _mm_movehl_ps(vout6x0123, vout6x0123);
395 _mm_storel_pi((__m64*) c7, vout7x0123);
396 c7 += 2;
397 vout7x0123 = _mm_movehl_ps(vout7x0123, vout7x0123);
398 }
399 if (nc & 1) {
400 _mm_store_ss(c0, vout0x0123);
401 _mm_store_ss(c1, vout1x0123);
402 _mm_store_ss(c2, vout2x0123);
403 _mm_store_ss(c3, vout3x0123);
404 _mm_store_ss(c4, vout4x0123);
405 _mm_store_ss(c5, vout5x0123);
406 _mm_store_ss(c6, vout6x0123);
407 _mm_store_ss(c7, vout7x0123);
408 }
409 nc = 0;
410 }
411 } while (nc != 0);
412 }