Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-8x8c8-minmax-avxvnni.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
19
20 void xnn_qd8_f16_qc4w_gemm_minmax_ukernel_8x8c8__avxvnni(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 xnn_float16* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const struct xnn_f16_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
31 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 8);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 const int8_t* a0 = a;
44 uint16_t* c0 = (uint16_t*) c;
45 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
46 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 a1 = a0;
49 c1 = c0;
50 }
51 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
52 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 a2 = a1;
55 c2 = c1;
56 }
57 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
58 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
59 if XNN_UNPREDICTABLE(mr < 4) {
60 a3 = a2;
61 c3 = c2;
62 }
63 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
64 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
65 if XNN_UNPREDICTABLE(mr <= 4) {
66 a4 = a3;
67 c4 = c3;
68 }
69 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
70 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
71 if XNN_UNPREDICTABLE(mr < 6) {
72 a5 = a4;
73 c5 = c4;
74 }
75 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
76 uint16_t* c6 = (uint16_t*) ((uintptr_t) c5 + cm_stride);
77 if XNN_UNPREDICTABLE(mr <= 6) {
78 a6 = a5;
79 c6 = c5;
80 }
81 const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
82 uint16_t* c7 = (uint16_t*) ((uintptr_t) c6 + cm_stride);
83 if XNN_UNPREDICTABLE(mr != 8) {
84 a7 = a6;
85 c7 = c6;
86 }
87
88 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
89 XNN_FORCE_REALIZATION(vsign_mask);
90 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
91 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
92 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
93 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
94 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
95 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point + 128);
96 const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point + 128);
97 const __m256i vinput_zero_point7 = _mm256_set1_epi32((int) quantization_params[7].zero_point + 128);
98 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
99 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
100 // XNN_FORCE_REALIZATION(voutput_min);
101 // XNN_FORCE_REALIZATION(voutput_max);
102 const __m256i vmask = _mm256_set1_epi8(0xF0);
103 XNN_FORCE_REALIZATION(vmask);
104 do {
105 const __m256i vksum01234567 = _mm256_load_si256(w);
106 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
107 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
108 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
109 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
110 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
111 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
112 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
113 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
114 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
115 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
116 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
117 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
118 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
119 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
120 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
121 __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
122 __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
123 __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
124 __m256i vsum6x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point6);
125 __m256i vacc6x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 0));
126 __m256i vacc6x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 1));
127 __m256i vsum7x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point7);
128 __m256i vacc7x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 0));
129 __m256i vacc7x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 1));
130 w = (const int32_t*) w + 8;
131
132 size_t k = kc;
133 while (k >= 16 * sizeof(int8_t)) {
134 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
135 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
136 a0 += 16;
137 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
138 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
139 a1 += 16;
140 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
141 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
142 a2 += 16;
143 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
144 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
145 a3 += 16;
146 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
147 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
148 a4 += 16;
149 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
150 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
151 a5 += 16;
152 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
153 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
154 a6 += 16;
155 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
156 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
157 a7 += 16;
158
159 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
160 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
161 const __m256i vbs01234567x0123 = _mm256_slli_epi32(vbb01234567x01234567, 4);
162 const __m256i vbs89ABCDEFx0123 = _mm256_slli_epi32(vbb89ABCDEFx01234567, 4);
163 const __m256i vb01234567x4567 = _mm256_and_si256(vbb01234567x01234567, vmask);
164 const __m256i vb89ABCDEFx4567 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
165 const __m256i vb01234567x0123 = _mm256_and_si256(vbs01234567x0123, vmask);
166 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbs89ABCDEFx0123, vmask);
167
168 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
169 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
170 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
171 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
172 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
173 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
174 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
175 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
176 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
177 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
178 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
179 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
180 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
181 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
182 vacc7x0123 = _mm256_dpbusd_avx_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
183 vacc7x4567 = _mm256_dpbusd_avx_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
184 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
185 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
186 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
187 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
188 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
189 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
190 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
191 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
192 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
193 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
194 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
195 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
196 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
197 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
198 vacc7x0123 = _mm256_dpbusd_avx_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
199 vacc7x4567 = _mm256_dpbusd_avx_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
200
201 w = (const int8_t*) w + 64;
202 k -= 16 * sizeof(int8_t);
203 }
204
205 if (k != 0) {
206 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
207 a0 += 8;
208 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
209 a1 += 8;
210 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
211 a2 += 8;
212 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
213 a3 += 8;
214 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
215 a4 += 8;
216 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
217 a5 += 8;
218 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
219 a6 += 8;
220 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
221 a7 += 8;
222
223 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
224 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
225 const __m256i vb01234567x0123 = _mm256_slli_epi32(vbb01234567x01234567, 4);
226 const __m256i vb89ABCDEFx0123 = _mm256_slli_epi32(vbb89ABCDEFx01234567, 4);
227
228 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
229 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
230 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
231 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
232 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
233 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
234 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
235 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
236 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
237 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
238 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
239 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
240 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
241 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
242 vacc7x0123 = _mm256_dpbusd_avx_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
243 vacc7x4567 = _mm256_dpbusd_avx_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
244
245 w = (const int8_t*) w + 64;
246 k -= 8 * sizeof(int8_t);
247 }
248
249 // Add adjacent pairs
250 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
251 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
252 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
253 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
254 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
255 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
256 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
257 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
258 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
259 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
260 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
261 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
262 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
263 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
264 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
265 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
266
267 vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4);
268 vacc1x01234567 = _mm256_srai_epi32(vacc1x01234567, 4);
269 vacc2x01234567 = _mm256_srai_epi32(vacc2x01234567, 4);
270 vacc3x01234567 = _mm256_srai_epi32(vacc3x01234567, 4);
271 vacc4x01234567 = _mm256_srai_epi32(vacc4x01234567, 4);
272 vacc5x01234567 = _mm256_srai_epi32(vacc5x01234567, 4);
273 vacc6x01234567 = _mm256_srai_epi32(vacc6x01234567, 4);
274 vacc7x01234567 = _mm256_srai_epi32(vacc7x01234567, 4);
275 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
276 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
277 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
278 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
279 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
280 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
281 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
282 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
283
284 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
285 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
286 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
287 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
288 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
289 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
290 vout6x01234567 = _mm256_mul_ps(vout6x01234567, _mm256_set1_ps(quantization_params[6].inv_scale));
291 vout7x01234567 = _mm256_mul_ps(vout7x01234567, _mm256_set1_ps(quantization_params[7].inv_scale));
292
293 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
294 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
295 w = (const float*) w + 16;
296
297 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
298 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
299 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
300 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
301 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
302 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
303 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
304 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
305
306 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
307 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
308 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
309 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
310 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
311 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
312 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
313 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
314
315 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
316 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
317 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
318 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
319 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
320 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
321 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
322 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
323
324 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
325 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
326 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
327 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
328 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
329 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
330 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
331 __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT);
332 if XNN_LIKELY(nc >= 8) {
333 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
334 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
335 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
336 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
337 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
338 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
339 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
340 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
341 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
342 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
343 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
344 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
345 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
346 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
347 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
348 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
349 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
350 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
351 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
352 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
353 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
354 _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567);
355 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
356 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
357 nc -= 8;
358 } else {
359 if (nc & 4) {
360 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
361 c0 += 4;
362 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
363 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
364 c1 += 4;
365 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
366 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
367 c2 += 4;
368 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
369 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
370 c3 += 4;
371 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
372 _mm_storel_epi64((__m128i*) c4, vfp16out4x01234567);
373 c4 += 4;
374 vfp16out4x01234567 = _mm_unpackhi_epi64(vfp16out4x01234567, vfp16out4x01234567);
375 _mm_storel_epi64((__m128i*) c5, vfp16out5x01234567);
376 c5 += 4;
377 vfp16out5x01234567 = _mm_unpackhi_epi64(vfp16out5x01234567, vfp16out5x01234567);
378 _mm_storel_epi64((__m128i*) c6, vfp16out6x01234567);
379 c6 += 4;
380 vfp16out6x01234567 = _mm_unpackhi_epi64(vfp16out6x01234567, vfp16out6x01234567);
381 _mm_storel_epi64((__m128i*) c7, vfp16out7x01234567);
382 c7 += 4;
383 vfp16out7x01234567 = _mm_unpackhi_epi64(vfp16out7x01234567, vfp16out7x01234567);
384 }
385 if (nc & 2) {
386 _mm_storeu_si32(c0, vfp16out0x01234567);
387 c0 += 2;
388 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
389 _mm_storeu_si32(c1, vfp16out1x01234567);
390 c1 += 2;
391 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
392 _mm_storeu_si32(c2, vfp16out2x01234567);
393 c2 += 2;
394 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
395 _mm_storeu_si32(c3, vfp16out3x01234567);
396 c3 += 2;
397 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
398 _mm_storeu_si32(c4, vfp16out4x01234567);
399 c4 += 2;
400 vfp16out4x01234567 = _mm_srli_epi64(vfp16out4x01234567, 32);
401 _mm_storeu_si32(c5, vfp16out5x01234567);
402 c5 += 2;
403 vfp16out5x01234567 = _mm_srli_epi64(vfp16out5x01234567, 32);
404 _mm_storeu_si32(c6, vfp16out6x01234567);
405 c6 += 2;
406 vfp16out6x01234567 = _mm_srli_epi64(vfp16out6x01234567, 32);
407 _mm_storeu_si32(c7, vfp16out7x01234567);
408 c7 += 2;
409 vfp16out7x01234567 = _mm_srli_epi64(vfp16out7x01234567, 32);
410 }
411 if (nc & 1) {
412 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
413 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
414 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
415 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
416 *c4 = (uint16_t) _mm_extract_epi16(vfp16out4x01234567, 0);
417 *c5 = (uint16_t) _mm_extract_epi16(vfp16out5x01234567, 0);
418 *c6 = (uint16_t) _mm_extract_epi16(vfp16out6x01234567, 0);
419 *c7 = (uint16_t) _mm_extract_epi16(vfp16out7x01234567, 0);
420 }
421 nc = 0;
422 }
423 } while (nc != 0);
424 }