Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-7x8c8-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_f16_qc4w_gemm_minmax_ukernel_7x8c8__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 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 <= 7);
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
83 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
84 XNN_FORCE_REALIZATION(vsign_mask);
85 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
86 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
87 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
88 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
89 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
90 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point + 128);
91 const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point + 128);
92 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
93 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
94 // XNN_FORCE_REALIZATION(voutput_min);
95 // XNN_FORCE_REALIZATION(voutput_max);
96 const __m256i vmask = _mm256_set1_epi8(0xF0);
97 XNN_FORCE_REALIZATION(vmask);
98 do {
99 const __m256i vksum01234567 = _mm256_load_si256(w);
100 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
101 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
102 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
103 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
104 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
105 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
106 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
107 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
108 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
109 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
110 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
111 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
112 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
113 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
114 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
115 __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
116 __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
117 __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
118 __m256i vsum6x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point6);
119 __m256i vacc6x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 0));
120 __m256i vacc6x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 1));
121 w = (const int32_t*) w + 8;
122
123 size_t k = kc;
124 while (k >= 16 * sizeof(int8_t)) {
125 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
126 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
127 a0 += 16;
128 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
129 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
130 a1 += 16;
131 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
132 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
133 a2 += 16;
134 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
135 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
136 a3 += 16;
137 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
138 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
139 a4 += 16;
140 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
141 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
142 a5 += 16;
143 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
144 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
145 a6 += 16;
146
147 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
148 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
149 const __m256i vbs01234567x0123 = _mm256_slli_epi32(vbb01234567x01234567, 4);
150 const __m256i vbs89ABCDEFx0123 = _mm256_slli_epi32(vbb89ABCDEFx01234567, 4);
151 const __m256i vb01234567x4567 = _mm256_and_si256(vbb01234567x01234567, vmask);
152 const __m256i vb89ABCDEFx4567 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
153 const __m256i vb01234567x0123 = _mm256_and_si256(vbs01234567x0123, vmask);
154 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbs89ABCDEFx0123, vmask);
155
156 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
157 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
158 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
159 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
160 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
161 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
162 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
163 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
164 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
165 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
166 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
167 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
168 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
169 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
170 xnn_prefetch_to_l1((const int8_t*) w + 960);
171 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
172 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
173 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
174 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
175 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
176 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
177 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
178 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
179 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
180 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
181 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
182 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
183 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
184 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
185
186 w = (const int8_t*) w + 64;
187 k -= 16 * sizeof(int8_t);
188 }
189
190 if (k != 0) {
191 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
192 a0 += 8;
193 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
194 a1 += 8;
195 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
196 a2 += 8;
197 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
198 a3 += 8;
199 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
200 a4 += 8;
201 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
202 a5 += 8;
203 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
204 a6 += 8;
205
206 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
207 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
208 const __m256i vb01234567x0123 = _mm256_slli_epi32(vbb01234567x01234567, 4);
209 const __m256i vb89ABCDEFx0123 = _mm256_slli_epi32(vbb89ABCDEFx01234567, 4);
210
211 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
212 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
213 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
214 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
215 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
216 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
217 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
218 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
219 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
220 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
221 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
222 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
223 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
224 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
225 xnn_prefetch_to_l1((const int8_t*) w + 960);
226
227 w = (const int8_t*) w + 64;
228 k -= 8 * sizeof(int8_t);
229 }
230
231 // Add adjacent pairs
232 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
233 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
234 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
235 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
236 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
237 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
238 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
239 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
240 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
241 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
242 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
243 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
244 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
245 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
246
247 vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4);
248 vacc1x01234567 = _mm256_srai_epi32(vacc1x01234567, 4);
249 vacc2x01234567 = _mm256_srai_epi32(vacc2x01234567, 4);
250 vacc3x01234567 = _mm256_srai_epi32(vacc3x01234567, 4);
251 vacc4x01234567 = _mm256_srai_epi32(vacc4x01234567, 4);
252 vacc5x01234567 = _mm256_srai_epi32(vacc5x01234567, 4);
253 vacc6x01234567 = _mm256_srai_epi32(vacc6x01234567, 4);
254 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
255 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
256 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
257 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
258 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
259 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
260 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
261
262 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
263 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
264 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
265 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
266 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
267 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
268 vout6x01234567 = _mm256_mul_ps(vout6x01234567, _mm256_set1_ps(quantization_params[6].inv_scale));
269
270 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
271 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
272 w = (const float*) w + 16;
273
274 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
275 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
276 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
277 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
278 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
279 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
280 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
281
282 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
283 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
284 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
285 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
286 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
287 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
288 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
289
290 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
291 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
292 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
293 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
294 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
295 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
296 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
297
298 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
299 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
300 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
301 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
302 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
303 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
304 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
305 if XNN_LIKELY(nc >= 8) {
306 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
307 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
308 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
309 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
310 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
311 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
312 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
313 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
314 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
315 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
316 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
317 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
318 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
319 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
320 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
321 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
322 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
323 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
324 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
325 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
326 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
327 nc -= 8;
328 } else {
329 if (nc & 4) {
330 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
331 c0 += 4;
332 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
333 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
334 c1 += 4;
335 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
336 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
337 c2 += 4;
338 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
339 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
340 c3 += 4;
341 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
342 _mm_storel_epi64((__m128i*) c4, vfp16out4x01234567);
343 c4 += 4;
344 vfp16out4x01234567 = _mm_unpackhi_epi64(vfp16out4x01234567, vfp16out4x01234567);
345 _mm_storel_epi64((__m128i*) c5, vfp16out5x01234567);
346 c5 += 4;
347 vfp16out5x01234567 = _mm_unpackhi_epi64(vfp16out5x01234567, vfp16out5x01234567);
348 _mm_storel_epi64((__m128i*) c6, vfp16out6x01234567);
349 c6 += 4;
350 vfp16out6x01234567 = _mm_unpackhi_epi64(vfp16out6x01234567, vfp16out6x01234567);
351 }
352 if (nc & 2) {
353 _mm_storeu_si32(c0, vfp16out0x01234567);
354 c0 += 2;
355 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
356 _mm_storeu_si32(c1, vfp16out1x01234567);
357 c1 += 2;
358 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
359 _mm_storeu_si32(c2, vfp16out2x01234567);
360 c2 += 2;
361 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
362 _mm_storeu_si32(c3, vfp16out3x01234567);
363 c3 += 2;
364 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
365 _mm_storeu_si32(c4, vfp16out4x01234567);
366 c4 += 2;
367 vfp16out4x01234567 = _mm_srli_epi64(vfp16out4x01234567, 32);
368 _mm_storeu_si32(c5, vfp16out5x01234567);
369 c5 += 2;
370 vfp16out5x01234567 = _mm_srli_epi64(vfp16out5x01234567, 32);
371 _mm_storeu_si32(c6, vfp16out6x01234567);
372 c6 += 2;
373 vfp16out6x01234567 = _mm_srli_epi64(vfp16out6x01234567, 32);
374 }
375 if (nc & 1) {
376 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
377 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
378 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
379 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
380 *c4 = (uint16_t) _mm_extract_epi16(vfp16out4x01234567, 0);
381 *c5 = (uint16_t) _mm_extract_epi16(vfp16out5x01234567, 0);
382 *c6 = (uint16_t) _mm_extract_epi16(vfp16out6x01234567, 0);
383 }
384 nc = 0;
385 }
386 } while (nc != 0);
387 }