Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x8c8-minmax-fp32-avx2.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx8c8-avx2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/igemm.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x8c8__avx2(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const int8_t** restrict a,
26 const void* restrict w,
27 int8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 1);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (1 * sizeof(void*)) == 0);
40 assert(a_offset % 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
48 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);
49 const __m256i voutput_zero_point = _mm256_set1_epi16(params->fp32_scalar.output_zero_point);
50 const __m256i voutput_min = _mm256_set1_epi8(params->fp32_scalar.output_min);
51 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
52 // XNN_FORCE_REALIZATION(voutput_zero_point);
53 // XNN_FORCE_REALIZATION(voutput_min);
54
55 do {
56 const __m128i vbias0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
57 const __m128i vbias0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
58 __m256i vacc0x01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x0), vbias0x1, 1);
59 const __m128i vbias0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
60 const __m128i vbias0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
61 __m256i vacc0x23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x2), vbias0x3, 1);
62 const __m128i vbias0x4 = _mm_cvtsi32_si128(((const int*) w)[4]);
63 const __m128i vbias0x5 = _mm_cvtsi32_si128(((const int*) w)[5]);
64 __m256i vacc0x45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x4), vbias0x5, 1);
65 const __m128i vbias0x6 = _mm_cvtsi32_si128(((const int*) w)[6]);
66 const __m128i vbias0x7 = _mm_cvtsi32_si128(((const int*) w)[7]);
67 __m256i vacc0x67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x6), vbias0x7, 1);
68 w = (const int32_t*) w + 8;
69
70 size_t p = ks;
71 do {
72 const int8_t* restrict a0 = a[0];
73 if XNN_UNPREDICTABLE(a0 != zero) {
74 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
75 }
76 a += 1;
77
78 size_t k = 0;
79 while (k < kc) {
80 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
81 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
82 a0 += 8;
83
84 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
85 const __m256i vxb01 = _mm256_cvtepi8_epi16(vb01);
86
87 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
88 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
89 const __m256i vxb23 = _mm256_cvtepi8_epi16(vb23);
90
91 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
92 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
93 const __m256i vxb45 = _mm256_cvtepi8_epi16(vb45);
94
95 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
96 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
97 const __m256i vxb67 = _mm256_cvtepi8_epi16(vb67);
98
99 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
100
101 w = (const void*) ((const int8_t*) w + 64);
102 k += 8 * sizeof(int8_t);
103 }
104 p -= 1 * sizeof(void*);
105 } while (p != 0);
106
107 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
108 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
109
110 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
111
112 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
113 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
114
115 __m256 vscaled0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
116
117 const __m256 vscale01234567 = _mm256_load_ps(w);
118 w = (const void*) ((const float*) w + 8);
119 vscaled0x01234567 = _mm256_mul_ps(vscaled0x01234567, vscale01234567);
120
121 vscaled0x01234567 = _mm256_min_ps(vscaled0x01234567, voutput_max_less_zero_point);
122
123 vacc0x01234567 = _mm256_cvtps_epi32(vscaled0x01234567);
124
125 __m256i vacc00x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc0x01234567), voutput_zero_point);
126
127 vacc00x01234567 = _mm256_permute4x64_epi64(vacc00x01234567, _MM_SHUFFLE(3, 1, 2, 0));
128
129 __m256i vout = _mm256_packs_epi16(vacc00x01234567, vacc00x01234567);
130
131 vout = _mm256_max_epi8(vout, voutput_min);
132
133 __m128i vout_lo = _mm256_castsi256_si128(vout);
134 __m128i vout_hi = _mm256_extracti128_si256(vout, 1);
135
136 if (nc >= 8) {
137 _mm_storel_epi64((__m128i*) c0, vout_lo);
138
139 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
140
141 a = (const int8_t**restrict) ((uintptr_t) a - ks);
142
143 nc -= 8;
144 } else {
145 if (nc & 4) {
146 _mm_storeu_si32(c0, vout_lo);
147
148 c0 += 4;
149
150 vout_lo = _mm_srli_epi64(vout_lo, 32);
151 vout_hi = _mm_srli_epi64(vout_hi, 32);
152 }
153 if (nc & 2) {
154 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout_lo, 0));
155
156 c0 += 2;
157
158 vout_lo = _mm_srli_epi32(vout_lo, 16);
159 vout_hi = _mm_srli_epi32(vout_hi, 16);
160 }
161 if (nc & 1) {
162 *c0 = (int8_t) _mm_extract_epi8(vout_lo, 0);
163 }
164
165 nc = 0;
166 }
167 } while (nc != 0);
168 }