Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-8x8c8-minmax-avx2-madd-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_f16_qc4w_gemm_minmax_ukernel_8x8c8__avx2_madd_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 xnn_float16* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const struct xnn_f16_qc4w_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 uint16_t* c0 = (uint16_t*) c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 uint16_t* c1 = (uint16_t*) ((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 uint16_t* c2 = (uint16_t*) ((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 uint16_t* c3 = (uint16_t*) ((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 uint16_t* c4 = (uint16_t*) ((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 uint16_t* c5 = (uint16_t*) ((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 uint16_t* c6 = (uint16_t*) ((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 uint16_t* c7 = (uint16_t*) ((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_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
100 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
101 // XNN_FORCE_REALIZATION(voutput_min);
102 // XNN_FORCE_REALIZATION(voutput_max);
103 const __m256i vmask = _mm256_set1_epi8(0x0F);
104 XNN_FORCE_REALIZATION(vmask);
105 do {
106 const __m256i vksum01234567 = _mm256_load_si256(w);
107 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
108 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
109 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
110 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
111 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
112 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
113 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
114 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
115 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
116 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
117 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
118 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
119 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
120 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
121 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
122 __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
123 __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
124 __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
125 __m256i vsum6x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point6);
126 __m256i vacc6x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 0));
127 __m256i vacc6x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 1));
128 __m256i vsum7x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point7);
129 __m256i vacc7x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 0));
130 __m256i vacc7x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 1));
131 w = (const int32_t*) w + 8;
132
133 size_t k = kc;
134 while (k >= 16 * sizeof(int8_t)) {
135 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
136 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
137 a0 += 16;
138 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
139 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
140 a1 += 16;
141 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
142 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
143 a2 += 16;
144 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
145 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
146 a3 += 16;
147 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
148 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
149 a4 += 16;
150 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
151 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
152 a5 += 16;
153 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
154 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
155 a6 += 16;
156 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
157 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
158 a7 += 16;
159
160 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
161 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
162 const __m256i vbs01234567x0123 = _mm256_srli_epi32(vbb01234567x01234567, 4);
163 const __m256i vbs89ABCDEFx0123 = _mm256_srli_epi32(vbb89ABCDEFx01234567, 4);
164 const __m256i vb01234567x0123 = _mm256_and_si256(vbb01234567x01234567, vmask);
165 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
166 const __m256i vb01234567x4567 = _mm256_and_si256(vbs01234567x0123, vmask);
167 const __m256i vb89ABCDEFx4567 = _mm256_and_si256(vbs89ABCDEFx0123, vmask);
168
169 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x01234567, vb01234567x0123);
170 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
171 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x01234567, vb01234567x0123);
172 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
173 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x01234567, vb01234567x0123);
174 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
175 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x01234567, vb01234567x0123);
176 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
177 vacc4x0123 = _mm256_dpbusd_epi32_madd(vacc4x0123, va4x01234567, vb01234567x0123);
178 vacc4x4567 = _mm256_dpbusd_epi32_madd(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
179 vacc5x0123 = _mm256_dpbusd_epi32_madd(vacc5x0123, va5x01234567, vb01234567x0123);
180 vacc5x4567 = _mm256_dpbusd_epi32_madd(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
181 vacc6x0123 = _mm256_dpbusd_epi32_madd(vacc6x0123, va6x01234567, vb01234567x0123);
182 vacc6x4567 = _mm256_dpbusd_epi32_madd(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
183 vacc7x0123 = _mm256_dpbusd_epi32_madd(vacc7x0123, va7x01234567, vb01234567x0123);
184 vacc7x4567 = _mm256_dpbusd_epi32_madd(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
185 xnn_prefetch_to_l1((const int8_t*) w + 960);
186 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
187 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
188 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
189 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
190 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
191 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
192 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
193 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
194 vacc4x0123 = _mm256_dpbusd_epi32_madd(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
195 vacc4x4567 = _mm256_dpbusd_epi32_madd(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
196 vacc5x0123 = _mm256_dpbusd_epi32_madd(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
197 vacc5x4567 = _mm256_dpbusd_epi32_madd(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
198 vacc6x0123 = _mm256_dpbusd_epi32_madd(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
199 vacc6x4567 = _mm256_dpbusd_epi32_madd(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
200 vacc7x0123 = _mm256_dpbusd_epi32_madd(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
201 vacc7x4567 = _mm256_dpbusd_epi32_madd(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
202
203 w = (const int8_t*) w + 64;
204 k -= 16 * sizeof(int8_t);
205 }
206
207 if (k != 0) {
208 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
209 a0 += 8;
210 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
211 a1 += 8;
212 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
213 a2 += 8;
214 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
215 a3 += 8;
216 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
217 a4 += 8;
218 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
219 a5 += 8;
220 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
221 a6 += 8;
222 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
223 a7 += 8;
224
225 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
226 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
227 const __m256i vb01234567x0123 = _mm256_and_si256(vbb01234567x01234567, vmask);
228 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
229
230 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x01234567, vb01234567x0123);
231 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
232 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x01234567, vb01234567x0123);
233 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
234 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x01234567, vb01234567x0123);
235 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
236 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x01234567, vb01234567x0123);
237 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
238 vacc4x0123 = _mm256_dpbusd_epi32_madd(vacc4x0123, va4x01234567, vb01234567x0123);
239 vacc4x4567 = _mm256_dpbusd_epi32_madd(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
240 vacc5x0123 = _mm256_dpbusd_epi32_madd(vacc5x0123, va5x01234567, vb01234567x0123);
241 vacc5x4567 = _mm256_dpbusd_epi32_madd(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
242 vacc6x0123 = _mm256_dpbusd_epi32_madd(vacc6x0123, va6x01234567, vb01234567x0123);
243 vacc6x4567 = _mm256_dpbusd_epi32_madd(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
244 vacc7x0123 = _mm256_dpbusd_epi32_madd(vacc7x0123, va7x01234567, vb01234567x0123);
245 vacc7x4567 = _mm256_dpbusd_epi32_madd(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
246 xnn_prefetch_to_l1((const int8_t*) w + 960);
247
248 w = (const int8_t*) w + 64;
249 k -= 8 * sizeof(int8_t);
250 }
251
252 // Add adjacent pairs
253 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
254 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
255 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
256 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
257 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
258 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
259 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
260 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
261 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
262 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
263 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
264 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
265 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
266 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
267 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
268 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
269
270 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
271 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
272 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
273 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
274 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
275 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
276 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
277 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
278
279 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
280 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
281 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
282 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
283 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
284 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
285 vout6x01234567 = _mm256_mul_ps(vout6x01234567, _mm256_set1_ps(quantization_params[6].inv_scale));
286 vout7x01234567 = _mm256_mul_ps(vout7x01234567, _mm256_set1_ps(quantization_params[7].inv_scale));
287
288 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
289 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
290 w = (const float*) w + 16;
291
292 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
293 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
294 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
295 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
296 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
297 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
298 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
299 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
300
301 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
302 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
303 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
304 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
305 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
306 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
307 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
308 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
309
310 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
311 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
312 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
313 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
314 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
315 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
316 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
317 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
318
319 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
320 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
321 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
322 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
323 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
324 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
325 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
326 __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT);
327 if XNN_LIKELY(nc >= 8) {
328 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
329 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
330 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
331 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
332 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
333 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
334 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
335 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
336 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
337 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
338 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
339 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
340 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
341 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
342 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
343 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
344 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
345 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
346 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
347 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
348 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
349 _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567);
350 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
351 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
352 nc -= 8;
353 } else {
354 if (nc & 4) {
355 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
356 c0 += 4;
357 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
358 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
359 c1 += 4;
360 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
361 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
362 c2 += 4;
363 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
364 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
365 c3 += 4;
366 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
367 _mm_storel_epi64((__m128i*) c4, vfp16out4x01234567);
368 c4 += 4;
369 vfp16out4x01234567 = _mm_unpackhi_epi64(vfp16out4x01234567, vfp16out4x01234567);
370 _mm_storel_epi64((__m128i*) c5, vfp16out5x01234567);
371 c5 += 4;
372 vfp16out5x01234567 = _mm_unpackhi_epi64(vfp16out5x01234567, vfp16out5x01234567);
373 _mm_storel_epi64((__m128i*) c6, vfp16out6x01234567);
374 c6 += 4;
375 vfp16out6x01234567 = _mm_unpackhi_epi64(vfp16out6x01234567, vfp16out6x01234567);
376 _mm_storel_epi64((__m128i*) c7, vfp16out7x01234567);
377 c7 += 4;
378 vfp16out7x01234567 = _mm_unpackhi_epi64(vfp16out7x01234567, vfp16out7x01234567);
379 }
380 if (nc & 2) {
381 _mm_storeu_si32(c0, vfp16out0x01234567);
382 c0 += 2;
383 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
384 _mm_storeu_si32(c1, vfp16out1x01234567);
385 c1 += 2;
386 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
387 _mm_storeu_si32(c2, vfp16out2x01234567);
388 c2 += 2;
389 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
390 _mm_storeu_si32(c3, vfp16out3x01234567);
391 c3 += 2;
392 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
393 _mm_storeu_si32(c4, vfp16out4x01234567);
394 c4 += 2;
395 vfp16out4x01234567 = _mm_srli_epi64(vfp16out4x01234567, 32);
396 _mm_storeu_si32(c5, vfp16out5x01234567);
397 c5 += 2;
398 vfp16out5x01234567 = _mm_srli_epi64(vfp16out5x01234567, 32);
399 _mm_storeu_si32(c6, vfp16out6x01234567);
400 c6 += 2;
401 vfp16out6x01234567 = _mm_srli_epi64(vfp16out6x01234567, 32);
402 _mm_storeu_si32(c7, vfp16out7x01234567);
403 c7 += 2;
404 vfp16out7x01234567 = _mm_srli_epi64(vfp16out7x01234567, 32);
405 }
406 if (nc & 1) {
407 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
408 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
409 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
410 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
411 *c4 = (uint16_t) _mm_extract_epi16(vfp16out4x01234567, 0);
412 *c5 = (uint16_t) _mm_extract_epi16(vfp16out5x01234567, 0);
413 *c6 = (uint16_t) _mm_extract_epi16(vfp16out6x01234567, 0);
414 *c7 = (uint16_t) _mm_extract_epi16(vfp16out7x01234567, 0);
415 }
416 nc = 0;
417 }
418 } while (nc != 0);
419 }