daiduo/TurboPFor-Integer-Compression
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

add loongarch64 support

Browse Source
This commit is contained in:
huchangqi
2025-05-28 10:04:23 +08:00
parent 06d6aad98b
commit e4f6853116
16 changed files with 401 additions and 66 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
lib/transpose.c
View File

@ -45,6 +45,8 @@
#include <emmintrin.h>
#elif defined(__ARM_NEON)
#include <arm_neon.h>
#include "include_/sse_neon.h"
#elif defined(__loongarch_lp64)
#include "include_/sse_neon.h"
#endif
@ -728,7 +730,7 @@ void T2(TPDEC256V, ESIZE)(unsigned char *in, unsigned n, unsigned char *out) {
#endif //TPDEC256V
#else //__AVX2__
#if (defined(__SSE3__) || defined(__ARM_NEON)) && (ESIZE == 2 || ESIZE == 4 || ESIZE == 8)
#if (defined(__SSE3__) || defined(__ARM_NEON) || defined(__loongarch_lp64)) && (ESIZE == 2 || ESIZE == 4 || ESIZE == 8)
#define ST(_p_,_v_,_i_) _mm_storeu_si128((__m128i *)SIE(_p_,_i_), _v_)
#define ST0(_p_,_v_) _mm_storeu_si128((__m128i *)(_p_), _v_)
@ -737,7 +739,7 @@ void T2(TPENC128V, ESIZE)(unsigned char *in, unsigned n, unsigned char *out) {
unsigned stride = v/STRIDE;
unsigned char *op,*ip;
#if defined(__SSE3__) || defined(__ARM_NEON)
#if defined(__SSE3__) || defined(__ARM_NEON) || defined(__loongarch_lp64)
#if ESIZE == 2
__m128i sf = _mm_set_epi8(15, 13, 11, 9, 7, 5, 3, 1,
14, 12, 10, 8, 6, 4, 2, 0);
@ -764,7 +766,7 @@ void T2(TPENC128V, ESIZE)(unsigned char *in, unsigned n, unsigned char *out) {
for(ip = in, op = out; ip != in+v; ip+=ESIZE*16,op += ESIZE*16/STRIDE) { unsigned char *p = op; PREFETCH(ip+(ESIZE*16)*ESIZE,0);
__m128i iv[ESIZE], ov[ESIZE == 2 ? ESIZE + 2 : ESIZE];
#if defined(__SSSE3__) || defined(__ARM_NEON)
#if defined(__SSSE3__) || defined(__ARM_NEON) || defined(__loongarch_lp64)
#if ESIZE == 2
#ifdef __ARM_NEON
uint8x16x2_t w = vld2q_u8(ip);