Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-1x16c8-minmax-avx512vnni-prfm.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 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x16c8__avx512vnni_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 float* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const int8_t* zero_data,
35 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
36 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38 assert(mr != 0);
39 assert(mr <= 1);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(kc % sizeof(int8_t) == 0);
43 assert(a != NULL);
44 assert(w != NULL);
45 assert(c != NULL);
46
47 kc = round_up_po2(kc, 8 * sizeof(int8_t));
48 float* c0 = c;
49
50 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
51 XNN_FORCE_REALIZATION(vsign_mask);
52 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point + 128);
53 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
54 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
55 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
56 // XNN_FORCE_REALIZATION(vinput_zero_point);
57 // XNN_FORCE_REALIZATION(vinput_inv_scale);
58 // XNN_FORCE_REALIZATION(voutput_min);
59 // XNN_FORCE_REALIZATION(voutput_max);
60 do {
61 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
62 const __m512i vsum0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
63 __m512i vacc0x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 0));
64 __m512i vacc0x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 1));
65 __m512i vacc1x0x01234567 = _mm512_setzero_epi32();
66 __m512i vacc1x0x89ABCDEF = _mm512_setzero_epi32();
67 w = (const int32_t*) w + 16;
68
69 size_t p = ks;
70 do {
71 const int8_t* restrict a0 = a[0];
72 if XNN_UNPREDICTABLE(a0 != zero) {
73 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
74 } else {
75 a0 = zero_data;
76 }
77 a += 1;
78
79 size_t k = kc;
80 while (k >= 16 * sizeof(int8_t)) {
81 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
82 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
83 a0 += 16;
84
85 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
86 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
87 const __m512i vb01234567x89ABCDEF = _mm512_load_si512((const int8_t*) w + 128);
88 const __m512i vb89ABCDEFx89ABCDEF = _mm512_load_si512((const int8_t*) w + 192);
89 xnn_prefetch_to_l1((const int8_t*) w + 768);
90 xnn_prefetch_to_l1((const int8_t*) w + 832);
91
92 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
93 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
94 xnn_prefetch_to_l1((const int8_t*) w + 896);
95 xnn_prefetch_to_l1((const int8_t*) w + 960);
96 vacc1x0x01234567 = _mm512_dpbusd_epi32(vacc1x0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
97 vacc1x0x89ABCDEF = _mm512_dpbusd_epi32(vacc1x0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
98
99 w = (const int8_t*) w + 256;
100 k -= 16 * sizeof(int8_t);
101 }
102
103 if (k != 0) {
104 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
105 a0 += 8;
106
107 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
108 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
109
110 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
111 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
112 xnn_prefetch_to_l1((const int8_t*) w + 896);
113 xnn_prefetch_to_l1((const int8_t*) w + 960);
114
115 w = (const int8_t*) w + 128;
116 k -= 8 * sizeof(int8_t);
117 }
118
119 p -= 1 * sizeof(void*);
120 } while (p != 0);
121
122 vacc0x01234567 = _mm512_add_epi32(vacc0x01234567, vacc1x0x01234567);
123 vacc0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, vacc1x0x89ABCDEF);
124
125 // Add adjacent pairs
126 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
127 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
128 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
129 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
130
131 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
132
133 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
134
135 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
136 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
137 w = (const float*) w + 32;
138
139 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
140
141 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
142
143 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
144
145 if XNN_LIKELY(nc >= 16) {
146 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
147 c0 = (float*) ((uintptr_t) c0 + cn_stride);
148 a = (const int8_t**restrict) ((uintptr_t) a - ks);
149 nc -= 16;
150 } else {
151 // Prepare mask for valid 32-bit elements (depends on nc).
152 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
153 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
154 nc = 0;
155 }
156 } while (nc != 0);
157 }