Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-2x8c8-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_2x8c8__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 <= 2);
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
52 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
53 XNN_FORCE_REALIZATION(vsign_mask);
54 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);
55 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
56 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
57 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
58 // XNN_FORCE_REALIZATION(voutput_zero_point);
59 // XNN_FORCE_REALIZATION(voutput_min);
60 do {
61 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
62 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
63 __m256i vacc1x0123 = vacc0x0123;
64 __m256i vacc1x4567 = vacc0x4567;
65 __m256i vacc1x0x0123 = _mm256_setzero_si256();
66 __m256i vacc1x0x4567 = _mm256_setzero_si256();
67 __m256i vacc1x1x0123 = _mm256_setzero_si256();
68 __m256i vacc1x1x4567 = _mm256_setzero_si256();
69 w = (const int32_t*) w + 8;
70
71 size_t p = ks;
72 do {
73 const int8_t* restrict a0 = a[0];
74 if XNN_UNPREDICTABLE(a0 != zero) {
75 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
76 }
77 const int8_t* restrict a1 = a[1];
78 if XNN_UNPREDICTABLE(a1 != zero) {
79 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
80 }
81 a += 2;
82
83 size_t k = kc;
84 while (k >= 16 * sizeof(int8_t)) {
85 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
86 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
87 a0 += 16;
88 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
89 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
90 a1 += 16;
91
92 const __m256i vb01234567x0123 = _mm256_load_si256(w);
93 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
94 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
95 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
96 xnn_prefetch_to_l1((const int8_t*) w + 896);
97
98 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
99 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
100 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
101 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
102 xnn_prefetch_to_l1((const int8_t*) w + 960);
103 vacc1x0x0123 = _mm256_dpbusd_avx_epi32(vacc1x0x0123, va0x89ABCDEF, vb01234567x4567);
104 vacc1x0x4567 = _mm256_dpbusd_avx_epi32(vacc1x0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
105 vacc1x1x0123 = _mm256_dpbusd_avx_epi32(vacc1x1x0123, va1x89ABCDEF, vb01234567x4567);
106 vacc1x1x4567 = _mm256_dpbusd_avx_epi32(vacc1x1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
107
108 w = (const int8_t*) w + 128;
109 k -= 16 * sizeof(int8_t);
110 }
111
112 if (k != 0) {
113 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
114 a0 += 8;
115 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
116 a1 += 8;
117
118 const __m256i vb01234567x0123 = _mm256_load_si256(w);
119 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
120
121 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
122 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
123 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
124 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
125 xnn_prefetch_to_l1((const int8_t*) w + 960);
126
127 w = (const int8_t*) w + 64;
128 k -= 8 * sizeof(int8_t);
129 }
130
131 p -= 2 * sizeof(void*);
132 } while (p != 0);
133
134 vacc0x0123 = _mm256_add_epi32(vacc0x0123, vacc1x0x0123);
135 vacc0x4567 = _mm256_add_epi32(vacc0x4567, vacc1x0x4567);
136 vacc1x0123 = _mm256_add_epi32(vacc1x0123, vacc1x1x0123);
137 vacc1x4567 = _mm256_add_epi32(vacc1x4567, vacc1x1x4567);
138
139 // Add adjacent pairs
140 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
141 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
142 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
143 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
144
145 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
146 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
147
148 const __m256 vscale01234567 = _mm256_load_ps(w);
149 w = (const float*) w + 8;
150 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
151 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
152
153 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
154 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
155
156 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
157 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
158
159 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
160 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
161
162 vacc0x01234567 = _mm256_packs_epi32(vacc0x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc0x01234567, 1)));
163 __m128i voutb0x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc0x01234567, vacc0x01234567));
164 vacc1x01234567 = _mm256_packs_epi32(vacc1x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc1x01234567, 1)));
165 __m128i voutb1x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc1x01234567, vacc1x01234567));
166
167 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
168 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
169
170 if (nc >= 8) {
171 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
172 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
173 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
174 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
175 a = (const int8_t**restrict) ((uintptr_t) a - ks);
176 nc -= 8;
177 } else {
178 if (nc & 4) {
179 _mm_storeu_si32(c1, voutb1x01234567);
180 c1 += 4;
181 _mm_storeu_si32(c0, voutb0x01234567);
182 c0 += 4;
183 voutb1x01234567 = _mm_srli_epi64(voutb1x01234567, 32);
184 voutb0x01234567 = _mm_srli_epi64(voutb0x01234567, 32);
185 }
186 if (nc & 2) {
187 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(voutb1x01234567, 0));
188 c1 += 2;
189 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(voutb0x01234567, 0));
190 c0 += 2;
191 voutb1x01234567 = _mm_srli_epi32(voutb1x01234567, 16);
192 voutb0x01234567 = _mm_srli_epi32(voutb0x01234567, 16);
193 }
194 if (nc & 1) {
195 *c1 = (int8_t) _mm_extract_epi8(voutb1x01234567, 0);
196 *c0 = (int8_t) _mm_extract_epi8(voutb0x01234567, 0);
197 }
198 nc = 0;
199 }
200 } while (nc != 0);
201 }