Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x16c8-minmax-avx512vnni.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx16c8-avx512vnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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_1x16c8__avx512vnni(
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 <= 1);
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
47 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
48 XNN_FORCE_REALIZATION(vsign_mask);
49 const __m512 voutput_max_less_zero_point = _mm512_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
50 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
51 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
52 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
53 // XNN_FORCE_REALIZATION(voutput_zero_point);
54 XNN_FORCE_REALIZATION(voutput_min);
55 do {
56 __m512i vacc0x01234567 = _mm512_cvtepu32_epi64(_mm256_load_si256((const __m256i*) w));
57 __m512i vacc0x89ABCDEF = _mm512_cvtepu32_epi64(_mm256_load_si256((const __m256i*) ((const int32_t*) w + 8)));
58 __m512i vacc1x0x01234567 = _mm512_setzero_epi32();
59 __m512i vacc1x0x89ABCDEF = _mm512_setzero_epi32();
60 w = (const int32_t*) w + 16;
61
62 size_t p = ks;
63 do {
64 const int8_t* restrict a0 = a[0];
65 if XNN_UNPREDICTABLE(a0 != zero) {
66 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
67 }
68 a += 1;
69
70 size_t k = kc;
71 while (k >= 16 * sizeof(int8_t)) {
72 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
73 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
74 a0 += 16;
75
76 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
77 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
78 const __m512i vb01234567x89ABCDEF = _mm512_load_si512((const int8_t*) w + 128);
79 const __m512i vb89ABCDEFx89ABCDEF = _mm512_load_si512((const int8_t*) w + 192);
80
81 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
82 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
83 vacc1x0x01234567 = _mm512_dpbusd_epi32(vacc1x0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
84 vacc1x0x89ABCDEF = _mm512_dpbusd_epi32(vacc1x0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
85
86 w = (const int8_t*) w + 256;
87 k -= 16 * sizeof(int8_t);
88 }
89
90 if (k != 0) {
91 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
92 a0 += 8;
93
94 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
95 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
96
97 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
98 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
99
100 w = (const int8_t*) w + 128;
101 k -= 8 * sizeof(int8_t);
102 }
103
104 p -= 1 * sizeof(void*);
105 } while (p != 0);
106
107 vacc0x01234567 = _mm512_add_epi32(vacc0x01234567, vacc1x0x01234567);
108 vacc0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, vacc1x0x89ABCDEF);
109
110 // Add adjacent pairs
111 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
112 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
113 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
114 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
115
116 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
117
118 const __m512 vscale012345678ABCDEF = _mm512_load_ps(w);
119 w = (const float*) w + 16;
120 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale012345678ABCDEF);
121
122 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
123
124 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
125
126 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
127
128 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
129
130 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
131
132 if (nc >= 16) {
133 _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);
134 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
135 a = (const int8_t**restrict) ((uintptr_t) a - ks);
136 nc -= 16;
137 } else {
138 // Prepare mask for valid 8-bit elements (depends on nc).
139 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
140 _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);
141 nc = 0;
142 }
143 } while (nc != 0);
144 }