thirdparty/ppsspp
2
0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

GPU: Remove neon xxhash implementation.

Browse source 
It's typically around the same speed now with modern compilers, and much
slower than XXH3.
This commit is contained in:
Unknown W. Brackets 2020-08-27 20:31:09 -07:00
parent 184d4a1fc0
commit 30625225b0
5 changed files with 5 additions and 116 deletions
Download patch file Download diff file Expand all files Collapse all files

101
GPU/Common/TextureDecoderNEON.cpp
View file

@ -162,107 +162,6 @@ void DoUnswizzleTex16NEON(const u8 *texptr, u32 *ydestp, int bxc, int byc, u32 p
}
}
// NOTE: This is just a NEON version of xxhash.
// GCC sucks at making things NEON and can't seem to handle it.
#if defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L // C99
# include <stdint.h>
typedef uint8_t BYTE;
typedef uint16_t U16;
typedef uint32_t U32;
typedef int32_t S32;
typedef uint64_t U64;
#else
typedef unsigned char BYTE;
typedef unsigned short U16;
typedef unsigned int U32;
typedef signed int S32;
typedef unsigned long long U64;
#endif
#define PRIME32_1 2654435761U
#define PRIME32_2 2246822519U
#define PRIME32_3 3266489917U
#define PRIME32_4 668265263U
#define PRIME32_5 374761393U
#if defined(_MSC_VER)
# define XXH_rotl32(x,r) _rotl(x,r)
#else
# define XXH_rotl32(x,r) ((x << r) | (x >> (32 - r)))
#endif
u32 ReliableHash32NEON(const void *input, size_t len, u32 seed) {
if (((uintptr_t)input & 3) != 0) {
// Cannot handle misaligned data. Fall back to XXH32.
return XXH32(input, len, seed);
}
const u8 *p = (const u8 *)input;
const u8 *const bEnd = p + len;
U32 h32;
#ifdef XXH_ACCEPT_NULL_INPUT_POINTER
if (p==NULL) { len=0; p=(const BYTE*)(size_t)16; }
#endif
if (len>=16)
{
const BYTE* const limit = bEnd - 16;
U32 v1 = seed + PRIME32_1 + PRIME32_2;
U32 v2 = seed + PRIME32_2;
U32 v3 = seed + 0;
U32 v4 = seed - PRIME32_1;
uint32x4_t prime32_1q = vdupq_n_u32(PRIME32_1);
uint32x4_t prime32_2q = vdupq_n_u32(PRIME32_2);
uint32x4_t vq = vcombine_u32(vcreate_u32(v1 | ((U64)v2 << 32)), vcreate_u32(v3 | ((U64)v4 << 32)));
do {
__builtin_prefetch(p + 0xc0, 0, 0);
vq = vmlaq_u32(vq, vld1q_u32((const U32*)p), prime32_2q);
vq = vorrq_u32(vshlq_n_u32(vq, 13), vshrq_n_u32(vq, 32 - 13));
p += 16;
vq = vmulq_u32(vq, prime32_1q);
} while (p<=limit);
v1 = vgetq_lane_u32(vq, 0);
v2 = vgetq_lane_u32(vq, 1);
v3 = vgetq_lane_u32(vq, 2);
v4 = vgetq_lane_u32(vq, 3);
h32 = XXH_rotl32(v1, 1) + XXH_rotl32(v2, 7) + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18);
}
else
{
h32 = seed + PRIME32_5;
}
h32 += (U32) len;
while (p<=bEnd-4)
{
h32 += *(const U32*)p * PRIME32_3;
h32 = XXH_rotl32(h32, 17) * PRIME32_4 ;
p+=4;
}
while (p<bEnd)
{
h32 += (*p) * PRIME32_5;
h32 = XXH_rotl32(h32, 11) * PRIME32_1 ;
p++;
}
h32 ^= h32 >> 15;
h32 *= PRIME32_2;
h32 ^= h32 >> 13;
h32 *= PRIME32_3;
h32 ^= h32 >> 16;
return h32;
}
static inline bool VectorIsNonZeroNEON(const uint32x4_t &v) {
u64 low = vgetq_lane_u64(vreinterpretq_u64_u32(v), 0);
u64 high = vgetq_lane_u64(vreinterpretq_u64_u32(v), 1);