Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x8c8-minmax-fp32-avxvnni.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/unaligned.h"
19
20
21 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_4x8c8__avxvnni(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 int8_t* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 4);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr != 4) {
56 c3 = c2;
57 }
58
59 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
60 XNN_FORCE_REALIZATION(vsign_mask);
61 const __m256 voutput_max_less_zero_point = _mm256_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
62 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
63 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
64 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
65 // XNN_FORCE_REALIZATION(voutput_zero_point);
66 // XNN_FORCE_REALIZATION(voutput_min);
67 do {
68 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
69 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
70 __m256i vacc1x0123 = vacc0x0123;
71 __m256i vacc1x4567 = vacc0x4567;
72 __m256i vacc2x0123 = vacc0x0123;
73 __m256i vacc2x4567 = vacc0x4567;
74 __m256i vacc3x0123 = vacc0x0123;
75 __m256i vacc3x4567 = vacc0x4567;
76 w = (const int32_t*) w + 8;
77
78 size_t p = ks;
79 do {
80 const int8_t* restrict a0 = a[0];
81 if XNN_UNPREDICTABLE(a0 != zero) {
82 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
83 }
84 const int8_t* restrict a1 = a[1];
85 if XNN_UNPREDICTABLE(a1 != zero) {
86 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
87 }
88 const int8_t* restrict a2 = a[2];
89 if XNN_UNPREDICTABLE(a2 != zero) {
90 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
91 }
92 const int8_t* restrict a3 = a[3];
93 if XNN_UNPREDICTABLE(a3 != zero) {
94 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
95 }
96 a += 4;
97
98 size_t k = kc;
99 while (k >= 16 * sizeof(int8_t)) {
100 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
101 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
102 a0 += 16;
103 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
104 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
105 a1 += 16;
106 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
107 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
108 a2 += 16;
109 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
110 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
111 a3 += 16;
112
113 const __m256i vb01234567x0123 = _mm256_load_si256(w);
114 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
115 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
116 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
117
118 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
119 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
120 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
121 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
122 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
123 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
124 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
125 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
126 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
127 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
128 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
129 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
130 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
131 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
132 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
133 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
134
135 w = (const int8_t*) w + 128;
136 k -= 16 * sizeof(int8_t);
137 }
138
139 if (k != 0) {
140 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
141 a0 += 8;
142 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
143 a1 += 8;
144 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
145 a2 += 8;
146 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
147 a3 += 8;
148
149 const __m256i vb01234567x0123 = _mm256_load_si256(w);
150 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
151
152 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
153 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
154 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
155 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
156 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
157 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
158 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
159 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
160
161 w = (const int8_t*) w + 64;
162 k -= 8 * sizeof(int8_t);
163 }
164
165 p -= 4 * sizeof(void*);
166 } while (p != 0);
167
168
169 // Add adjacent pairs
170 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
171 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
172 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
173 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
174 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
175 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
176 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
177 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
178
179 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
180 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
181 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
182 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
183
184 const __m256 vscale01234567 = _mm256_load_ps(w);
185 w = (const float*) w + 8;
186 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
187 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
188 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);
189 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);
190
191 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
192 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
193 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);
194 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);
195
196 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
197 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
198 vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);
199 vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);
200
201 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
202 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
203 vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);
204 vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);
205
206 vacc0x01234567 = _mm256_packs_epi32(vacc0x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc0x01234567, 1)));
207 __m128i voutb0x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc0x01234567, vacc0x01234567));
208 vacc1x01234567 = _mm256_packs_epi32(vacc1x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc1x01234567, 1)));
209 __m128i voutb1x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc1x01234567, vacc1x01234567));
210 vacc2x01234567 = _mm256_packs_epi32(vacc2x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc2x01234567, 1)));
211 __m128i voutb2x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc2x01234567, vacc2x01234567));
212 vacc3x01234567 = _mm256_packs_epi32(vacc3x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc3x01234567, 1)));
213 __m128i voutb3x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc3x01234567, vacc3x01234567));
214
215 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
216 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
217 voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);
218 voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);
219
220 if (nc >= 8) {
221 _mm_storel_epi64((__m128i*) c3, voutb3x01234567);
222 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
223 _mm_storel_epi64((__m128i*) c2, voutb2x01234567);
224 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
225 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
226 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
227 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
228 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
229 a = (const int8_t**restrict) ((uintptr_t) a - ks);
230 nc -= 8;
231 } else {
232 if (nc & 4) {
233 _mm_storeu_si32(c3, voutb3x01234567);
234 c3 += 4;
235 _mm_storeu_si32(c2, voutb2x01234567);
236 c2 += 4;
237 _mm_storeu_si32(c1, voutb1x01234567);
238 c1 += 4;
239 _mm_storeu_si32(c0, voutb0x01234567);
240 c0 += 4;
241 voutb3x01234567 = _mm_srli_epi64(voutb3x01234567, 32);
242 voutb2x01234567 = _mm_srli_epi64(voutb2x01234567, 32);
243 voutb1x01234567 = _mm_srli_epi64(voutb1x01234567, 32);
244 voutb0x01234567 = _mm_srli_epi64(voutb0x01234567, 32);
245 }
246 if (nc & 2) {
247 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(voutb3x01234567, 0));
248 c3 += 2;
249 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(voutb2x01234567, 0));
250 c2 += 2;
251 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(voutb1x01234567, 0));
252 c1 += 2;
253 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(voutb0x01234567, 0));
254 c0 += 2;
255 voutb3x01234567 = _mm_srli_epi32(voutb3x01234567, 16);
256 voutb2x01234567 = _mm_srli_epi32(voutb2x01234567, 16);
257 voutb1x01234567 = _mm_srli_epi32(voutb1x01234567, 16);
258 voutb0x01234567 = _mm_srli_epi32(voutb0x01234567, 16);
259 }
260 if (nc & 1) {
261 *c3 = (int8_t) _mm_extract_epi8(voutb3x01234567, 0);
262 *c2 = (int8_t) _mm_extract_epi8(voutb2x01234567, 0);
263 *c1 = (int8_t) _mm_extract_epi8(voutb1x01234567, 0);
264 *c0 = (int8_t) _mm_extract_epi8(voutb0x01234567, 0);
265 }
266 nc = 0;
267 }
268 } while (nc != 0);
269 }