Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-6x8c8-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_6x8c8__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 <= 6);
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
77 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
78 XNN_FORCE_REALIZATION(vsign_mask);
79 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
80 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
81 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
82 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
83 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
84 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point + 128);
85 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
86 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
87 // XNN_FORCE_REALIZATION(voutput_min);
88 // XNN_FORCE_REALIZATION(voutput_max);
89 const __m256i vmask = _mm256_set1_epi8(0xF0);
90 XNN_FORCE_REALIZATION(vmask);
91 do {
92 const __m256i vksum01234567 = _mm256_load_si256(w);
93 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
94 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
95 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
96 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
97 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
98 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
99 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
100 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
101 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
102 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
103 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
104 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
105 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
106 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
107 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
108 __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
109 __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
110 __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
111 w = (const int32_t*) w + 8;
112
113 size_t k = kc;
114 while (k >= 16 * sizeof(int8_t)) {
115 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
116 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
117 a0 += 16;
118 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
119 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
120 a1 += 16;
121 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
122 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
123 a2 += 16;
124 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
125 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
126 a3 += 16;
127 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
128 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
129 a4 += 16;
130 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
131 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
132 a5 += 16;
133
134 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
135 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
136 const __m256i vbs01234567x0123 = _mm256_slli_epi32(vbb01234567x01234567, 4);
137 const __m256i vbs89ABCDEFx0123 = _mm256_slli_epi32(vbb89ABCDEFx01234567, 4);
138 const __m256i vb01234567x4567 = _mm256_and_si256(vbb01234567x01234567, vmask);
139 const __m256i vb89ABCDEFx4567 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
140 const __m256i vb01234567x0123 = _mm256_and_si256(vbs01234567x0123, vmask);
141 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbs89ABCDEFx0123, vmask);
142
143 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
144 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
145 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
146 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
147 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
148 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
149 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
150 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
151 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
152 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
153 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
154 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
155 xnn_prefetch_to_l1((const int8_t*) w + 960);
156 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
157 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
158 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
159 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
160 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
161 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
162 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
163 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
164 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
165 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
166 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
167 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
168
169 w = (const int8_t*) w + 64;
170 k -= 16 * sizeof(int8_t);
171 }
172
173 if (k != 0) {
174 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
175 a0 += 8;
176 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
177 a1 += 8;
178 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
179 a2 += 8;
180 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
181 a3 += 8;
182 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
183 a4 += 8;
184 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
185 a5 += 8;
186
187 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
188 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
189 const __m256i vb01234567x0123 = _mm256_slli_epi32(vbb01234567x01234567, 4);
190 const __m256i vb89ABCDEFx0123 = _mm256_slli_epi32(vbb89ABCDEFx01234567, 4);
191
192 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
193 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
194 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
195 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
196 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
197 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
198 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
199 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
200 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
201 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
202 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
203 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
204 xnn_prefetch_to_l1((const int8_t*) w + 960);
205
206 w = (const int8_t*) w + 64;
207 k -= 8 * sizeof(int8_t);
208 }
209
210 // Add adjacent pairs
211 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
212 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
213 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
214 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
215 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
216 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
217 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
218 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
219 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
220 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
221 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
222 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
223
224 vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4);
225 vacc1x01234567 = _mm256_srai_epi32(vacc1x01234567, 4);
226 vacc2x01234567 = _mm256_srai_epi32(vacc2x01234567, 4);
227 vacc3x01234567 = _mm256_srai_epi32(vacc3x01234567, 4);
228 vacc4x01234567 = _mm256_srai_epi32(vacc4x01234567, 4);
229 vacc5x01234567 = _mm256_srai_epi32(vacc5x01234567, 4);
230 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
231 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
232 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
233 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
234 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
235 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
236
237 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
238 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
239 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
240 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
241 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
242 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
243
244 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
245 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
246 w = (const float*) w + 16;
247
248 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
249 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
250 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
251 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
252 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
253 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
254
255 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
256 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
257 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
258 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
259 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
260 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
261
262 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
263 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
264 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
265 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
266 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
267 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
268
269 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
270 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
271 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
272 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
273 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
274 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
275 if XNN_LIKELY(nc >= 8) {
276 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
277 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
278 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
279 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
280 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
281 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
282 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
283 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
284 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
285 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
286 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
287 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
288 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
289 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
290 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
291 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
292 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
293 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
294 nc -= 8;
295 } else {
296 if (nc & 4) {
297 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
298 c0 += 4;
299 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
300 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
301 c1 += 4;
302 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
303 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
304 c2 += 4;
305 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
306 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
307 c3 += 4;
308 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
309 _mm_storel_epi64((__m128i*) c4, vfp16out4x01234567);
310 c4 += 4;
311 vfp16out4x01234567 = _mm_unpackhi_epi64(vfp16out4x01234567, vfp16out4x01234567);
312 _mm_storel_epi64((__m128i*) c5, vfp16out5x01234567);
313 c5 += 4;
314 vfp16out5x01234567 = _mm_unpackhi_epi64(vfp16out5x01234567, vfp16out5x01234567);
315 }
316 if (nc & 2) {
317 _mm_storeu_si32(c0, vfp16out0x01234567);
318 c0 += 2;
319 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
320 _mm_storeu_si32(c1, vfp16out1x01234567);
321 c1 += 2;
322 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
323 _mm_storeu_si32(c2, vfp16out2x01234567);
324 c2 += 2;
325 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
326 _mm_storeu_si32(c3, vfp16out3x01234567);
327 c3 += 2;
328 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
329 _mm_storeu_si32(c4, vfp16out4x01234567);
330 c4 += 2;
331 vfp16out4x01234567 = _mm_srli_epi64(vfp16out4x01234567, 32);
332 _mm_storeu_si32(c5, vfp16out5x01234567);
333 c5 += 2;
334 vfp16out5x01234567 = _mm_srli_epi64(vfp16out5x01234567, 32);
335 }
336 if (nc & 1) {
337 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
338 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
339 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
340 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
341 *c4 = (uint16_t) _mm_extract_epi16(vfp16out4x01234567, 0);
342 *c5 = (uint16_t) _mm_extract_epi16(vfp16out5x01234567, 0);
343 }
344 nc = 0;
345 }
346 } while (nc != 0);
347 }