Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-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_qd8_f32_qc8w_igemm_minmax_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 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const int8_t* zero_data,
34 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
35 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 1);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 8 * sizeof(int8_t));
47 float* c0 = c;
48
49 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
50 XNN_FORCE_REALIZATION(vsign_mask);
51 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point + 128);
52 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
53 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
54 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
55 // XNN_FORCE_REALIZATION(vinput_zero_point);
56 // XNN_FORCE_REALIZATION(vinput_inv_scale);
57 // XNN_FORCE_REALIZATION(voutput_min);
58 // XNN_FORCE_REALIZATION(voutput_max);
59 do {
60 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
61 const __m512i vsum0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
62 __m512i vacc0x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 0));
63 __m512i vacc0x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 1));
64 __m512i vacc1x0x01234567 = _mm512_setzero_epi32();
65 __m512i vacc1x0x89ABCDEF = _mm512_setzero_epi32();
66 w = (const int32_t*) w + 16;
67
68 size_t p = ks;
69 do {
70 const int8_t* restrict a0 = a[0];
71 if XNN_UNPREDICTABLE(a0 != zero) {
72 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
73 } else {
74 a0 = zero_data;
75 }
76 a += 1;
77
78 size_t k = kc;
79 while (k >= 16 * sizeof(int8_t)) {
80 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
81 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
82 a0 += 16;
83
84 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
85 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
86 const __m512i vb01234567x89ABCDEF = _mm512_load_si512((const int8_t*) w + 128);
87 const __m512i vb89ABCDEFx89ABCDEF = _mm512_load_si512((const int8_t*) w + 192);
88
89 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
90 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
91 vacc1x0x01234567 = _mm512_dpbusd_epi32(vacc1x0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
92 vacc1x0x89ABCDEF = _mm512_dpbusd_epi32(vacc1x0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
93
94 w = (const int8_t*) w + 256;
95 k -= 16 * sizeof(int8_t);
96 }
97
98 if (k != 0) {
99 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
100 a0 += 8;
101
102 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
103 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
104
105 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
106 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
107
108 w = (const int8_t*) w + 128;
109 k -= 8 * sizeof(int8_t);
110 }
111
112 p -= 1 * sizeof(void*);
113 } while (p != 0);
114
115 vacc0x01234567 = _mm512_add_epi32(vacc0x01234567, vacc1x0x01234567);
116 vacc0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, vacc1x0x89ABCDEF);
117
118 // Add adjacent pairs
119 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
120 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
121 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
122 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
123
124 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
125
126 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
127
128 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
129 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
130 w = (const float*) w + 32;
131
132 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
133
134 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
135
136 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
137
138 if XNN_LIKELY(nc >= 16) {
139 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
140 c0 = (float*) ((uintptr_t) c0 + cn_stride);
141 a = (const int8_t**restrict) ((uintptr_t) a - ks);
142 nc -= 16;
143 } else {
144 // Prepare mask for valid 32-bit elements (depends on nc).
145 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
146 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
147 nc = 0;
148 }
149 } while (nc != 0);
150 }