Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-12x8c8-minmax-avx256vnni.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_12x8c8__avx256vnni(
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 <= 12);
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 const int8_t* a8 = (const int8_t*) ((uintptr_t) a7 + a_stride);
88 uint16_t* c8 = (uint16_t*) ((uintptr_t) c7 + cm_stride);
89 if XNN_UNPREDICTABLE(mr <= 8) {
90 a8 = a7;
91 c8 = c7;
92 }
93 const int8_t* a9 = (const int8_t*) ((uintptr_t) a8 + a_stride);
94 uint16_t* c9 = (uint16_t*) ((uintptr_t) c8 + cm_stride);
95 if XNN_UNPREDICTABLE(mr < 10) {
96 a9 = a8;
97 c9 = c8;
98 }
99 const int8_t* a10 = (const int8_t*) ((uintptr_t) a9 + a_stride);
100 uint16_t* c10 = (uint16_t*) ((uintptr_t) c9 + cm_stride);
101 if XNN_UNPREDICTABLE(mr <= 10) {
102 a10 = a9;
103 c10 = c9;
104 }
105 const int8_t* a11 = (const int8_t*) ((uintptr_t) a10 + a_stride);
106 uint16_t* c11 = (uint16_t*) ((uintptr_t) c10 + cm_stride);
107 if XNN_UNPREDICTABLE(mr != 12) {
108 a11 = a10;
109 c11 = c10;
110 }
111
112 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
113 XNN_FORCE_REALIZATION(vsign_mask);
114 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
115 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
116 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
117 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
118 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
119 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point + 128);
120 const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point + 128);
121 const __m256i vinput_zero_point7 = _mm256_set1_epi32((int) quantization_params[7].zero_point + 128);
122 const __m256i vinput_zero_point8 = _mm256_set1_epi32((int) quantization_params[8].zero_point + 128);
123 const __m256i vinput_zero_point9 = _mm256_set1_epi32((int) quantization_params[9].zero_point + 128);
124 const __m256i vinput_zero_point10 = _mm256_set1_epi32((int) quantization_params[10].zero_point + 128);
125 const __m256i vinput_zero_point11 = _mm256_set1_epi32((int) quantization_params[11].zero_point + 128);
126 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
127 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
128 // XNN_FORCE_REALIZATION(voutput_min);
129 // XNN_FORCE_REALIZATION(voutput_max);
130 const __m256i vmask = _mm256_set1_epi8(0xF0);
131 XNN_FORCE_REALIZATION(vmask);
132 do {
133 const __m256i vksum01234567 = _mm256_load_si256(w);
134 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
135 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
136 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
137 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
138 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
139 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
140 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
141 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
142 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
143 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
144 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
145 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
146 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
147 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
148 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
149 __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
150 __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
151 __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
152 __m256i vsum6x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point6);
153 __m256i vacc6x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 0));
154 __m256i vacc6x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 1));
155 __m256i vsum7x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point7);
156 __m256i vacc7x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 0));
157 __m256i vacc7x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 1));
158 __m256i vsum8x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point8);
159 __m256i vacc8x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum8x01234567, 0));
160 __m256i vacc8x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum8x01234567, 1));
161 __m256i vsum9x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point9);
162 __m256i vacc9x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum9x01234567, 0));
163 __m256i vacc9x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum9x01234567, 1));
164 __m256i vsum10x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point10);
165 __m256i vacc10x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum10x01234567, 0));
166 __m256i vacc10x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum10x01234567, 1));
167 __m256i vsum11x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point11);
168 __m256i vacc11x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum11x01234567, 0));
169 __m256i vacc11x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum11x01234567, 1));
170 w = (const int32_t*) w + 8;
171
172 size_t k = kc;
173 while (k >= 16 * sizeof(int8_t)) {
174 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
175 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
176 a0 += 16;
177 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
178 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
179 a1 += 16;
180 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
181 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
182 a2 += 16;
183 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
184 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
185 a3 += 16;
186 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
187 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
188 a4 += 16;
189 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
190 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
191 a5 += 16;
192 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
193 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
194 a6 += 16;
195 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
196 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
197 a7 += 16;
198 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
199 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
200 a8 += 16;
201 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
202 const __m256i va9x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
203 a9 += 16;
204 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
205 const __m256i va10x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10 + 8)), vsign_mask);
206 a10 += 16;
207 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
208 const __m256i va11x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11 + 8)), vsign_mask);
209 a11 += 16;
210
211 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
212 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
213 const __m256i vbs01234567x0123 = _mm256_slli_epi32(vbb01234567x01234567, 4);
214 const __m256i vbs89ABCDEFx0123 = _mm256_slli_epi32(vbb89ABCDEFx01234567, 4);
215 const __m256i vb01234567x4567 = _mm256_and_si256(vbb01234567x01234567, vmask);
216 const __m256i vb89ABCDEFx4567 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
217 const __m256i vb01234567x0123 = _mm256_and_si256(vbs01234567x0123, vmask);
218 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbs89ABCDEFx0123, vmask);
219
220 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
221 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
222 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
223 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
224 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
225 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
226 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
227 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
228 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
229 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
230 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
231 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
232 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
233 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
234 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
235 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
236 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
237 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
238 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
239 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
240 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
241 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
242 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
243 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
244 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
245 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
246 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
247 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
248 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
249 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
250 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
251 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
252 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
253 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
254 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
255 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
256 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
257 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
258 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
259 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
260 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
261 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
262 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x89ABCDEF, vb01234567x4567);
263 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x89ABCDEF, vb89ABCDEFx4567);
264 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x89ABCDEF, vb01234567x4567);
265 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x89ABCDEF, vb89ABCDEFx4567);
266 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x89ABCDEF, vb01234567x4567);
267 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x89ABCDEF, vb89ABCDEFx4567);
268
269 w = (const int8_t*) w + 64;
270 k -= 16 * sizeof(int8_t);
271 }
272
273 if (k != 0) {
274 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
275 a0 += 8;
276 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
277 a1 += 8;
278 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
279 a2 += 8;
280 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
281 a3 += 8;
282 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
283 a4 += 8;
284 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
285 a5 += 8;
286 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
287 a6 += 8;
288 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
289 a7 += 8;
290 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
291 a8 += 8;
292 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
293 a9 += 8;
294 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
295 a10 += 8;
296 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
297 a11 += 8;
298
299 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
300 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
301 const __m256i vb01234567x0123 = _mm256_slli_epi32(vbb01234567x01234567, 4);
302 const __m256i vb89ABCDEFx0123 = _mm256_slli_epi32(vbb89ABCDEFx01234567, 4);
303
304 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
305 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
306 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
307 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
308 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
309 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
310 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
311 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
312 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
313 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
314 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
315 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
316 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
317 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
318 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
319 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
320 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
321 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
322 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
323 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
324 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
325 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
326 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
327 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
328
329 w = (const int8_t*) w + 64;
330 k -= 8 * sizeof(int8_t);
331 }
332
333 // Add adjacent pairs
334 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
335 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
336 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
337 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
338 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
339 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
340 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
341 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
342 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
343 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
344 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
345 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
346 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
347 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
348 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
349 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
350 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
351 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
352 const __m256i vsum9x02134657 = _mm256_hadd_epi32(vacc9x0123, vacc9x4567);
353 __m256i vacc9x01234567 = _mm256_permute4x64_epi64(vsum9x02134657, _MM_SHUFFLE(3, 1, 2, 0));
354 const __m256i vsum10x02134657 = _mm256_hadd_epi32(vacc10x0123, vacc10x4567);
355 __m256i vacc10x01234567 = _mm256_permute4x64_epi64(vsum10x02134657, _MM_SHUFFLE(3, 1, 2, 0));
356 const __m256i vsum11x02134657 = _mm256_hadd_epi32(vacc11x0123, vacc11x4567);
357 __m256i vacc11x01234567 = _mm256_permute4x64_epi64(vsum11x02134657, _MM_SHUFFLE(3, 1, 2, 0));
358
359 vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4);
360 vacc1x01234567 = _mm256_srai_epi32(vacc1x01234567, 4);
361 vacc2x01234567 = _mm256_srai_epi32(vacc2x01234567, 4);
362 vacc3x01234567 = _mm256_srai_epi32(vacc3x01234567, 4);
363 vacc4x01234567 = _mm256_srai_epi32(vacc4x01234567, 4);
364 vacc5x01234567 = _mm256_srai_epi32(vacc5x01234567, 4);
365 vacc6x01234567 = _mm256_srai_epi32(vacc6x01234567, 4);
366 vacc7x01234567 = _mm256_srai_epi32(vacc7x01234567, 4);
367 vacc8x01234567 = _mm256_srai_epi32(vacc8x01234567, 4);
368 vacc9x01234567 = _mm256_srai_epi32(vacc9x01234567, 4);
369 vacc10x01234567 = _mm256_srai_epi32(vacc10x01234567, 4);
370 vacc11x01234567 = _mm256_srai_epi32(vacc11x01234567, 4);
371 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
372 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
373 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
374 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
375 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
376 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
377 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
378 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
379 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
380 __m256 vout9x01234567 = _mm256_cvtepi32_ps(vacc9x01234567);
381 __m256 vout10x01234567 = _mm256_cvtepi32_ps(vacc10x01234567);
382 __m256 vout11x01234567 = _mm256_cvtepi32_ps(vacc11x01234567);
383
384 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
385 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
386 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
387 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
388 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
389 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
390 vout6x01234567 = _mm256_mul_ps(vout6x01234567, _mm256_set1_ps(quantization_params[6].inv_scale));
391 vout7x01234567 = _mm256_mul_ps(vout7x01234567, _mm256_set1_ps(quantization_params[7].inv_scale));
392 vout8x01234567 = _mm256_mul_ps(vout8x01234567, _mm256_set1_ps(quantization_params[8].inv_scale));
393 vout9x01234567 = _mm256_mul_ps(vout9x01234567, _mm256_set1_ps(quantization_params[9].inv_scale));
394 vout10x01234567 = _mm256_mul_ps(vout10x01234567, _mm256_set1_ps(quantization_params[10].inv_scale));
395 vout11x01234567 = _mm256_mul_ps(vout11x01234567, _mm256_set1_ps(quantization_params[11].inv_scale));
396
397 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
398 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
399 w = (const float*) w + 16;
400
401 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
402 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
403 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
404 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
405 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
406 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
407 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
408 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
409 vout8x01234567 = _mm256_fmadd_ps(vout8x01234567, vfilter_output_scale01234567, vbias01234567);
410 vout9x01234567 = _mm256_fmadd_ps(vout9x01234567, vfilter_output_scale01234567, vbias01234567);
411 vout10x01234567 = _mm256_fmadd_ps(vout10x01234567, vfilter_output_scale01234567, vbias01234567);
412 vout11x01234567 = _mm256_fmadd_ps(vout11x01234567, vfilter_output_scale01234567, vbias01234567);
413
414 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
415 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
416 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
417 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
418 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
419 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
420 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
421 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
422 vout8x01234567 = _mm256_max_ps(vout8x01234567, voutput_min);
423 vout9x01234567 = _mm256_max_ps(vout9x01234567, voutput_min);
424 vout10x01234567 = _mm256_max_ps(vout10x01234567, voutput_min);
425 vout11x01234567 = _mm256_max_ps(vout11x01234567, voutput_min);
426
427 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
428 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
429 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
430 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
431 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
432 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
433 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
434 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
435 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max);
436 vout9x01234567 = _mm256_min_ps(vout9x01234567, voutput_max);
437 vout10x01234567 = _mm256_min_ps(vout10x01234567, voutput_max);
438 vout11x01234567 = _mm256_min_ps(vout11x01234567, voutput_max);
439
440 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
441 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
442 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
443 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
444 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
445 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
446 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
447 __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT);
448 __m128i vfp16out8x01234567 = _mm256_cvtps_ph(vout8x01234567, _MM_FROUND_TO_NEAREST_INT);
449 __m128i vfp16out9x01234567 = _mm256_cvtps_ph(vout9x01234567, _MM_FROUND_TO_NEAREST_INT);
450 __m128i vfp16out10x01234567 = _mm256_cvtps_ph(vout10x01234567, _MM_FROUND_TO_NEAREST_INT);
451 __m128i vfp16out11x01234567 = _mm256_cvtps_ph(vout11x01234567, _MM_FROUND_TO_NEAREST_INT);
452 if XNN_LIKELY(nc >= 8) {
453 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
454 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
455 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
456 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
457 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
458 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
459 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
460 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
461 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
462 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
463 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
464 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
465 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
466 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
467 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
468 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
469 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
470 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
471 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
472 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
473 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
474 _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567);
475 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
476 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
477 _mm_storeu_si128((__m128i*) c8, vfp16out8x01234567);
478 a8 = (const int8_t*) ((uintptr_t) a8 - kc);
479 c8 = (uint16_t*) ((uintptr_t) c8 + cn_stride);
480 _mm_storeu_si128((__m128i*) c9, vfp16out9x01234567);
481 a9 = (const int8_t*) ((uintptr_t) a9 - kc);
482 c9 = (uint16_t*) ((uintptr_t) c9 + cn_stride);
483 _mm_storeu_si128((__m128i*) c10, vfp16out10x01234567);
484 a10 = (const int8_t*) ((uintptr_t) a10 - kc);
485 c10 = (uint16_t*) ((uintptr_t) c10 + cn_stride);
486 _mm_storeu_si128((__m128i*) c11, vfp16out11x01234567);
487 a11 = (const int8_t*) ((uintptr_t) a11 - kc);
488 c11 = (uint16_t*) ((uintptr_t) c11 + cn_stride);
489 nc -= 8;
490 } else {
491 // Prepare mask for valid 16-bit elements (depends on nc).
492 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
493 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
494 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
495 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
496 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
497 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
498 _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567);
499 _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567);
500 _mm_mask_storeu_epi16(c7, vmask, vfp16out7x01234567);
501 _mm_mask_storeu_epi16(c8, vmask, vfp16out8x01234567);
502 _mm_mask_storeu_epi16(c9, vmask, vfp16out9x01234567);
503 _mm_mask_storeu_epi16(c10, vmask, vfp16out10x01234567);
504 _mm_mask_storeu_epi16(c11, vmask, vfp16out11x01234567);
505 nc = 0;
506 }
507 } while (nc != 0);
508 }