Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-9x8c8-minmax-avx256skx-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_9x8c8__avx256skx_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 <= 9);
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 const int8_t* a8 = (const int8_t*) ((uintptr_t) a7 + a_stride);
89 uint16_t* c8 = (uint16_t*) ((uintptr_t) c7 + cm_stride);
90 if XNN_UNPREDICTABLE(mr <= 8) {
91 a8 = a7;
92 c8 = c7;
93 }
94
95 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
96 XNN_FORCE_REALIZATION(vsign_mask);
97 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
98 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
99 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
100 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
101 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
102 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point + 128);
103 const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point + 128);
104 const __m256i vinput_zero_point7 = _mm256_set1_epi32((int) quantization_params[7].zero_point + 128);
105 const __m256i vinput_zero_point8 = _mm256_set1_epi32((int) quantization_params[8].zero_point + 128);
106 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
107 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
108 // XNN_FORCE_REALIZATION(voutput_min);
109 // XNN_FORCE_REALIZATION(voutput_max);
110 const __m256i vmask = _mm256_set1_epi8(0x0F);
111 XNN_FORCE_REALIZATION(vmask);
112 do {
113 const __m256i vksum01234567 = _mm256_load_si256(w);
114 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
115 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
116 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
117 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
118 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
119 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
120 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
121 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
122 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
123 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
124 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
125 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
126 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
127 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
128 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
129 __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
130 __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
131 __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
132 __m256i vsum6x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point6);
133 __m256i vacc6x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 0));
134 __m256i vacc6x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 1));
135 __m256i vsum7x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point7);
136 __m256i vacc7x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 0));
137 __m256i vacc7x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 1));
138 __m256i vsum8x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point8);
139 __m256i vacc8x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum8x01234567, 0));
140 __m256i vacc8x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum8x01234567, 1));
141 w = (const int32_t*) w + 8;
142
143 size_t k = kc;
144 while (k >= 16 * sizeof(int8_t)) {
145 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
146 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
147 a0 += 16;
148 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
149 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
150 a1 += 16;
151 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
152 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
153 a2 += 16;
154 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
155 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
156 a3 += 16;
157 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
158 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
159 a4 += 16;
160 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
161 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
162 a5 += 16;
163 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
164 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
165 a6 += 16;
166 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
167 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
168 a7 += 16;
169 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
170 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
171 a8 += 16;
172
173 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
174 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
175 const __m256i vbs01234567x0123 = _mm256_srli_epi32(vbb01234567x01234567, 4);
176 const __m256i vbs89ABCDEFx0123 = _mm256_srli_epi32(vbb89ABCDEFx01234567, 4);
177 const __m256i vb01234567x0123 = _mm256_and_si256(vbb01234567x01234567, vmask);
178 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
179 const __m256i vb01234567x4567 = _mm256_and_si256(vbs01234567x0123, vmask);
180 const __m256i vb89ABCDEFx4567 = _mm256_and_si256(vbs89ABCDEFx0123, vmask);
181
182 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x01234567, vb01234567x0123);
183 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
184 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x01234567, vb01234567x0123);
185 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
186 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x01234567, vb01234567x0123);
187 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
188 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x01234567, vb01234567x0123);
189 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
190 vacc4x0123 = _mm256_dpbusd_epi32_madd(vacc4x0123, va4x01234567, vb01234567x0123);
191 vacc4x4567 = _mm256_dpbusd_epi32_madd(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
192 vacc5x0123 = _mm256_dpbusd_epi32_madd(vacc5x0123, va5x01234567, vb01234567x0123);
193 vacc5x4567 = _mm256_dpbusd_epi32_madd(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
194 vacc6x0123 = _mm256_dpbusd_epi32_madd(vacc6x0123, va6x01234567, vb01234567x0123);
195 vacc6x4567 = _mm256_dpbusd_epi32_madd(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
196 vacc7x0123 = _mm256_dpbusd_epi32_madd(vacc7x0123, va7x01234567, vb01234567x0123);
197 vacc7x4567 = _mm256_dpbusd_epi32_madd(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
198 vacc8x0123 = _mm256_dpbusd_epi32_madd(vacc8x0123, va8x01234567, vb01234567x0123);
199 vacc8x4567 = _mm256_dpbusd_epi32_madd(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
200 xnn_prefetch_to_l1((const int8_t*) w + 960);
201 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
202 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
203 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
204 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
205 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
206 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
207 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
208 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
209 vacc4x0123 = _mm256_dpbusd_epi32_madd(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
210 vacc4x4567 = _mm256_dpbusd_epi32_madd(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
211 vacc5x0123 = _mm256_dpbusd_epi32_madd(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
212 vacc5x4567 = _mm256_dpbusd_epi32_madd(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
213 vacc6x0123 = _mm256_dpbusd_epi32_madd(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
214 vacc6x4567 = _mm256_dpbusd_epi32_madd(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
215 vacc7x0123 = _mm256_dpbusd_epi32_madd(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
216 vacc7x4567 = _mm256_dpbusd_epi32_madd(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
217 vacc8x0123 = _mm256_dpbusd_epi32_madd(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
218 vacc8x4567 = _mm256_dpbusd_epi32_madd(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
219
220 w = (const int8_t*) w + 64;
221 k -= 16 * sizeof(int8_t);
222 }
223
224 if (k != 0) {
225 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
226 a0 += 8;
227 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
228 a1 += 8;
229 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
230 a2 += 8;
231 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
232 a3 += 8;
233 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
234 a4 += 8;
235 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
236 a5 += 8;
237 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
238 a6 += 8;
239 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
240 a7 += 8;
241 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
242 a8 += 8;
243
244 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
245 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
246 const __m256i vb01234567x0123 = _mm256_and_si256(vbb01234567x01234567, vmask);
247 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
248
249 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x01234567, vb01234567x0123);
250 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
251 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x01234567, vb01234567x0123);
252 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
253 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x01234567, vb01234567x0123);
254 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
255 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x01234567, vb01234567x0123);
256 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
257 vacc4x0123 = _mm256_dpbusd_epi32_madd(vacc4x0123, va4x01234567, vb01234567x0123);
258 vacc4x4567 = _mm256_dpbusd_epi32_madd(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
259 vacc5x0123 = _mm256_dpbusd_epi32_madd(vacc5x0123, va5x01234567, vb01234567x0123);
260 vacc5x4567 = _mm256_dpbusd_epi32_madd(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
261 vacc6x0123 = _mm256_dpbusd_epi32_madd(vacc6x0123, va6x01234567, vb01234567x0123);
262 vacc6x4567 = _mm256_dpbusd_epi32_madd(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
263 vacc7x0123 = _mm256_dpbusd_epi32_madd(vacc7x0123, va7x01234567, vb01234567x0123);
264 vacc7x4567 = _mm256_dpbusd_epi32_madd(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
265 vacc8x0123 = _mm256_dpbusd_epi32_madd(vacc8x0123, va8x01234567, vb01234567x0123);
266 vacc8x4567 = _mm256_dpbusd_epi32_madd(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
267 xnn_prefetch_to_l1((const int8_t*) w + 960);
268
269 w = (const int8_t*) w + 64;
270 k -= 8 * sizeof(int8_t);
271 }
272
273 // Add adjacent pairs
274 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
275 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
276 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
277 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
278 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
279 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
280 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
281 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
282 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
283 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
284 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
285 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
286 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
287 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
288 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
289 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
290 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
291 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
292
293 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
294 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
295 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
296 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
297 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
298 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
299 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
300 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
301 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
302
303 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
304 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
305 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
306 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
307 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
308 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
309 vout6x01234567 = _mm256_mul_ps(vout6x01234567, _mm256_set1_ps(quantization_params[6].inv_scale));
310 vout7x01234567 = _mm256_mul_ps(vout7x01234567, _mm256_set1_ps(quantization_params[7].inv_scale));
311 vout8x01234567 = _mm256_mul_ps(vout8x01234567, _mm256_set1_ps(quantization_params[8].inv_scale));
312
313 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
314 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
315 w = (const float*) w + 16;
316
317 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
318 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
319 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
320 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
321 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
322 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
323 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
324 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
325 vout8x01234567 = _mm256_fmadd_ps(vout8x01234567, vfilter_output_scale01234567, vbias01234567);
326
327 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
328 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
329 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
330 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
331 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
332 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
333 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
334 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
335 vout8x01234567 = _mm256_max_ps(vout8x01234567, voutput_min);
336
337 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
338 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
339 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
340 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
341 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
342 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
343 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
344 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
345 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max);
346
347 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
348 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
349 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
350 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
351 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
352 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
353 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
354 __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT);
355 __m128i vfp16out8x01234567 = _mm256_cvtps_ph(vout8x01234567, _MM_FROUND_TO_NEAREST_INT);
356 if XNN_LIKELY(nc >= 8) {
357 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
358 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
359 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
360 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
361 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
362 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
363 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
364 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
365 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
366 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
367 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
368 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
369 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
370 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
371 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
372 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
373 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
374 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
375 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
376 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
377 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
378 _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567);
379 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
380 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
381 _mm_storeu_si128((__m128i*) c8, vfp16out8x01234567);
382 a8 = (const int8_t*) ((uintptr_t) a8 - kc);
383 c8 = (uint16_t*) ((uintptr_t) c8 + cn_stride);
384 nc -= 8;
385 } else {
386 // Prepare mask for valid 16-bit elements (depends on nc).
387 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
388 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
389 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
390 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
391 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
392 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
393 _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567);
394 _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567);
395 _mm_mask_storeu_epi16(c7, vmask, vfp16out7x01234567);
396 _mm_mask_storeu_epi16(c8, vmask, vfp16out8x01234567);
397 nc = 0;
398 }
399 } while (nc != 0);
400 }