Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x8c8-minmax-fp32-avxvnni-prfm.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 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_4x8c8__avxvnni_prfm(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 size_t ks,
27 const int8_t** restrict a,
28 const void* restrict w,
29 int8_t* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 4);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 int8_t* c0 = c;
47 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 c3 = c2;
58 }
59
60 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
61 XNN_FORCE_REALIZATION(vsign_mask);
62 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);
63 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
64 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
65 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
66 // XNN_FORCE_REALIZATION(voutput_zero_point);
67 // XNN_FORCE_REALIZATION(voutput_min);
68 do {
69 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
70 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
71 __m256i vacc1x0123 = vacc0x0123;
72 __m256i vacc1x4567 = vacc0x4567;
73 __m256i vacc2x0123 = vacc0x0123;
74 __m256i vacc2x4567 = vacc0x4567;
75 __m256i vacc3x0123 = vacc0x0123;
76 __m256i vacc3x4567 = vacc0x4567;
77 w = (const int32_t*) w + 8;
78
79 size_t p = ks;
80 do {
81 const int8_t* restrict a0 = a[0];
82 if XNN_UNPREDICTABLE(a0 != zero) {
83 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
84 }
85 const int8_t* restrict a1 = a[1];
86 if XNN_UNPREDICTABLE(a1 != zero) {
87 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
88 }
89 const int8_t* restrict a2 = a[2];
90 if XNN_UNPREDICTABLE(a2 != zero) {
91 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
92 }
93 const int8_t* restrict a3 = a[3];
94 if XNN_UNPREDICTABLE(a3 != zero) {
95 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
96 }
97 a += 4;
98
99 size_t k = kc;
100 while (k >= 16 * sizeof(int8_t)) {
101 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
102 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
103 a0 += 16;
104 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
105 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
106 a1 += 16;
107 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
108 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
109 a2 += 16;
110 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
111 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
112 a3 += 16;
113
114 const __m256i vb01234567x0123 = _mm256_load_si256(w);
115 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
116 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
117 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
118 xnn_prefetch_to_l1((const int8_t*) w + 896);
119
120 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
121 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
122 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
123 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
124 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
125 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
126 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
127 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
128 xnn_prefetch_to_l1((const int8_t*) w + 960);
129 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
130 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
131 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
132 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
133 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
134 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
135 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
136 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
137
138 w = (const int8_t*) w + 128;
139 k -= 16 * sizeof(int8_t);
140 }
141
142 if (k != 0) {
143 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
144 a0 += 8;
145 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
146 a1 += 8;
147 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
148 a2 += 8;
149 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
150 a3 += 8;
151
152 const __m256i vb01234567x0123 = _mm256_load_si256(w);
153 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
154
155 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
156 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
157 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
158 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
159 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
160 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
161 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
162 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
163 xnn_prefetch_to_l1((const int8_t*) w + 960);
164
165 w = (const int8_t*) w + 64;
166 k -= 8 * sizeof(int8_t);
167 }
168
169 p -= 4 * sizeof(void*);
170 } while (p != 0);
171
172
173 // Add adjacent pairs
174 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
175 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
176 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
177 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
178 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
179 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
180 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
181 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
182
183 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
184 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
185 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
186 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
187
188 const __m256 vscale01234567 = _mm256_load_ps(w);
189 w = (const float*) w + 8;
190 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
191 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
192 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);
193 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);
194
195 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
196 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
197 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);
198 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);
199
200 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
201 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
202 vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);
203 vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);
204
205 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
206 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
207 vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);
208 vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);
209
210 vacc0x01234567 = _mm256_packs_epi32(vacc0x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc0x01234567, 1)));
211 __m128i voutb0x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc0x01234567, vacc0x01234567));
212 vacc1x01234567 = _mm256_packs_epi32(vacc1x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc1x01234567, 1)));
213 __m128i voutb1x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc1x01234567, vacc1x01234567));
214 vacc2x01234567 = _mm256_packs_epi32(vacc2x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc2x01234567, 1)));
215 __m128i voutb2x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc2x01234567, vacc2x01234567));
216 vacc3x01234567 = _mm256_packs_epi32(vacc3x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc3x01234567, 1)));
217 __m128i voutb3x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc3x01234567, vacc3x01234567));
218
219 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
220 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
221 voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);
222 voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);
223
224 if (nc >= 8) {
225 _mm_storel_epi64((__m128i*) c3, voutb3x01234567);
226 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
227 _mm_storel_epi64((__m128i*) c2, voutb2x01234567);
228 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
229 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
230 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
231 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
232 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
233 a = (const int8_t**restrict) ((uintptr_t) a - ks);
234 nc -= 8;
235 } else {
236 if (nc & 4) {
237 _mm_storeu_si32(c3, voutb3x01234567);
238 c3 += 4;
239 _mm_storeu_si32(c2, voutb2x01234567);
240 c2 += 4;
241 _mm_storeu_si32(c1, voutb1x01234567);
242 c1 += 4;
243 _mm_storeu_si32(c0, voutb0x01234567);
244 c0 += 4;
245 voutb3x01234567 = _mm_srli_epi64(voutb3x01234567, 32);
246 voutb2x01234567 = _mm_srli_epi64(voutb2x01234567, 32);
247 voutb1x01234567 = _mm_srli_epi64(voutb1x01234567, 32);
248 voutb0x01234567 = _mm_srli_epi64(voutb0x01234567, 32);
249 }
250 if (nc & 2) {
251 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(voutb3x01234567, 0));
252 c3 += 2;
253 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(voutb2x01234567, 0));
254 c2 += 2;
255 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(voutb1x01234567, 0));
256 c1 += 2;
257 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(voutb0x01234567, 0));
258 c0 += 2;
259 voutb3x01234567 = _mm_srli_epi32(voutb3x01234567, 16);
260 voutb2x01234567 = _mm_srli_epi32(voutb2x01234567, 16);
261 voutb1x01234567 = _mm_srli_epi32(voutb1x01234567, 16);
262 voutb0x01234567 = _mm_srli_epi32(voutb0x01234567, 16);
263 }
264 if (nc & 1) {
265 *c3 = (int8_t) _mm_extract_epi8(voutb3x01234567, 0);
266 *c2 = (int8_t) _mm_extract_epi8(voutb2x01234567, 0);
267 *c1 = (int8_t) _mm_extract_epi8(voutb1x01234567, 0);
268 *c0 = (int8_t) _mm_extract_epi8(voutb0x01234567, 0);
269 }
270 nc = 0;
271 }
272 } while (nc != 0);
273 }