1// Copyright 2022 Google LLC
2//
3// This source code is licensed under the BSD-style license found in the
4// LICENSE file in the root directory of this source tree.
5
6#include <assert.h>
7#include <stddef.h>
8
9#ifdef _WIN32
10 #include <windows.h>
11#else
12 #include <pthread.h>
13#endif
14
15#include <xnnpack/common.h>
16#include <xnnpack/config.h>
17#include <xnnpack/lut.h>
18
19
20static struct xnn_x8_lut_config x8_lut_config = {0};
21
22#if XNN_PLATFORM_WINDOWS
23 static INIT_ONCE init_guard = INIT_ONCE_STATIC_INIT;
24#else
25 static pthread_once_t init_guard = PTHREAD_ONCE_INIT;
26#endif
27
28static void init_x8_lut_config(void) {
29 #if XNN_ARCH_ARM
30 x8_lut_config.microkernel = xnn_x8_lut_ukernel__scalar_x4;
31 x8_lut_config.tile_size = 4;
32 #elif XNN_ARCH_ARM64
33 x8_lut_config.microkernel = xnn_x8_lut_ukernel__aarch64_neon_tbx128x4_x64;
34 x8_lut_config.tile_size = 64;
35 #elif XNN_ARCH_X86 || XNN_ARCH_X86_64
36 const struct xnn_hardware_config* hardware_config = xnn_init_hardware_config();
37 assert(hardware_config != NULL);
38
39 if (!XNN_PLATFORM_MOBILE && hardware_config->use_x86_avx512skx) {
40 if (hardware_config->use_x86_avx512vbmi) {
41 x8_lut_config.microkernel = xnn_x8_lut_ukernel__avx512vbmi_vpermx2b_x128;
42 x8_lut_config.tile_size = 128;
43 } else {
44 x8_lut_config.microkernel = xnn_x8_lut_ukernel__avx512skx_vpshufb_x64;
45 x8_lut_config.tile_size = 64;
46 }
47 } else if (hardware_config->use_x86_avx2) {
48 x8_lut_config.microkernel = xnn_x8_lut_ukernel__avx2_x128;
49 x8_lut_config.tile_size = 128;
50 } else if (hardware_config->use_x86_avx) {
51 x8_lut_config.microkernel = xnn_x8_lut_ukernel__avx_x64;
52 x8_lut_config.tile_size = 64;
53 } else {
54 // Note: SSSE3 version is usually slower than scalar
55 x8_lut_config.microkernel = xnn_x8_lut_ukernel__scalar_x4;
56 x8_lut_config.tile_size = 4;
57 }
58 #elif XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
59 const struct xnn_hardware_config* hardware_config = xnn_init_hardware_config();
60 assert(hardware_config != NULL);
61
62 if (hardware_config->is_x86) {
63 #if XNN_ARCH_WASMRELAXEDSIMD
64 if (hardware_config->use_wasm_pshufb) {
65 x8_lut_config.microkernel = xnn_x8_lut_ukernel__wasmpshufb_x32;
66 x8_lut_config.tile_size = 32;
67 } else {
68 x8_lut_config.microkernel = xnn_x8_lut_ukernel__scalar_x1;
69 x8_lut_config.tile_size = 1;
70 }
71 #else
72 x8_lut_config.microkernel = xnn_x8_lut_ukernel__scalar_x1;
73 x8_lut_config.tile_size = 1;
74 #endif
75 } else {
76 x8_lut_config.microkernel = xnn_x8_lut_ukernel__wasmsimd_x32;
77 x8_lut_config.tile_size = 32;
78 }
79 #elif XNN_ARCH_WASM
80 x8_lut_config.microkernel = xnn_x8_lut_ukernel__scalar_x1;
81 x8_lut_config.tile_size = 1;
82 #elif XNN_ARCH_RISCV
83 x8_lut_config.microkernel = xnn_x8_lut_ukernel__scalar_x4;
84 x8_lut_config.tile_size = 4;
85 #else
86 #error "Unsupported architecture"
87 #endif
88}
89
90#if XNN_PLATFORM_WINDOWS
91 static BOOL CALLBACK init_x8_lut_config_windows(PINIT_ONCE init_once, PVOID parameter, PVOID* context) {
92 init_x8_lut_config();
93 return TRUE;
94 }
95#endif
96
97const struct xnn_x8_lut_config* xnn_init_x8_lut_config() {
98 const struct xnn_hardware_config* hardware_config = xnn_init_hardware_config();
99 if (hardware_config == NULL) {
100 return NULL;
101 }
102 #if XNN_PLATFORM_WINDOWS
103 InitOnceExecuteOnce(&init_guard, &init_x8_lut_config_windows, NULL, NULL);
104 #else
105 pthread_once(&init_guard, &init_x8_lut_config);
106 #endif
107 return &x8_lut_config;
108}
109