github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-06-28 16:41:01 +03:00
Code Issues Packages Projects Releases Wiki Activity

readd simd_onebit_masks for x86, needs more work

Browse Source
This commit is contained in:
Konstantinos Margaritis 2022-09-07 13:42:25 +03:00
parent 0af2ba8616
commit 1ae0d15181
2 changed files with 20 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/util/arch/common/simd_utils.h
View File

@ -88,6 +88,7 @@ static inline void print_m128_2x64(const char *label, m128 vec) {
#define print_m128_2x64(label, vec) ;
#endif
#if !defined(ARCH_IA32) && !defined(ARCH_X86_64)
#define ZEROES_8 0, 0, 0, 0, 0, 0, 0, 0
#define ZEROES_31 ZEROES_8, ZEROES_8, ZEROES_8, 0, 0, 0, 0, 0, 0, 0
#define ZEROES_32 ZEROES_8, ZEROES_8, ZEROES_8, ZEROES_8
@ -105,6 +106,7 @@ ALIGN_CL_DIRECTIVE static const u8 simd_onebit_masks[] = {
ZEROES_31, 0x80, ZEROES_32,
ZEROES_32, ZEROES_32,
};
#endif // !defined(ARCH_IA32) && !defined(ARCH_X86_64)
/****
**** 256-bit Primitives

26
src/util/arch/x86/simd_utils.h
View File

@ -42,6 +42,24 @@
#include <string.h> // for memcpy
#define ZEROES_8 0, 0, 0, 0, 0, 0, 0, 0
#define ZEROES_31 ZEROES_8, ZEROES_8, ZEROES_8, 0, 0, 0, 0, 0, 0, 0
#define ZEROES_32 ZEROES_8, ZEROES_8, ZEROES_8, ZEROES_8
/** \brief LUT for the mask1bit functions. */
ALIGN_CL_DIRECTIVE static const u8 simd_onebit_masks[] = {
ZEROES_32, ZEROES_32,
ZEROES_31, 0x01, ZEROES_32,
ZEROES_31, 0x02, ZEROES_32,
ZEROES_31, 0x04, ZEROES_32,
ZEROES_31, 0x08, ZEROES_32,
ZEROES_31, 0x10, ZEROES_32,
ZEROES_31, 0x20, ZEROES_32,
ZEROES_31, 0x40, ZEROES_32,
ZEROES_31, 0x80, ZEROES_32,
ZEROES_32, ZEROES_32,
};
static really_inline m128 ones128(void) {
#if defined(__GNUC__) || defined(__INTEL_COMPILER)
/* gcc gets this right */
@ -237,14 +255,6 @@ m128 loadbytes128(const void *ptr, unsigned int n) {
memcpy(&a, ptr, n);
return a;
}
/*
#ifdef __cplusplus
extern "C" {
#endif
extern const u8 simd_onebit_masks[];
#ifdef __cplusplus
}
#endif*/
static really_inline
m128 mask1bit128(unsigned int n) {