ShuriZma
/
desmume
mirror of https://github.com/TASEmulators/desmume.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Colorspace Handler: Further optimize NEON code. 

- Commit 9ccc791 was, more or less, a straight port of the SSE2 code, making it less than ideal. This updated version uses more NEON-only instructions to further improve performance.
This commit is contained in:
rogerman 2022-04-06 23:51:29 -07:00
parent 27a32d6262
commit 28f12cf088
2 changed files with 1190 additions and 236 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1299
desmume/src/texcache.cpp
View File

File diff suppressed because it is too large Load Diff

53
desmume/src/utils/colorspacehandler/colorspacehandler_NEON.cpp
View File

@ -25,7 +25,7 @@
#define COLOR16_SWAPRB_NEON(src) vorrq_u16( vshlq_n_u16(vandq_u16(src,vdupq_n_u16(0x001F)),10), vorrq_u16( vandq_u16(src,vdupq_n_u16(0x03E0)), vorrq_u16(vshrq_n_u16(vandq_u16(src,vdupq_n_u16(0x7C00)),10), vandq_u16(src,vdupq_n_u16(0x8000))) ) )
#define COLOR32_SWAPRB_NEON(src) vcopyq_laneq_u32( vcopyq_laneq_u32(src, 2, src, 0), 0, src, 2 )
#define COLOR32_SWAPRB_NEON(src) vqtbl1q_u8( (src), ((v128u8){2,1,0,3, 6,5,4,7, 10,9,8,11, 14,13,12,15}) )
template <bool SWAP_RB>
FORCEINLINE void ColorspaceConvert555To8888_NEON(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi)
@ -47,12 +47,11 @@ FORCEINLINE void ColorspaceConvert555To8888_NEON(const v128u16 &srcColor, const
}
else
{
const v128u16 r = vandq_u16( vshlq_n_u16(srcColor, 3), vdupq_n_u16(0x00F8) );
v128u16 rg = vorrq_u16( r, vandq_u16(vshlq_n_u16(srcColor, 6), vdupq_n_u16(0xF800)) );
rg = vorrq_u16( rg, vandq_u16(vshrq_n_u16(rg, 5), vdupq_n_u16(0x0707)) );
v128u16 rg = vorrq_u16( vandq_u16( vshlq_n_u16(srcColor,3), vdupq_n_u16(0x00F8) ), vandq_u16( vshlq_n_u16(srcColor,6), vdupq_n_u16(0xF800) ) );
v128u16 ba = vandq_u16( vshrq_n_u16(srcColor,7), vdupq_n_u16(0x00F8) );
ba = vorrq_u16(ba, vshrq_n_u16(ba, 5));
rg = vorrq_u16( rg, vreinterpretq_u16_u8(vshrq_n_u8(vreinterpretq_u8_u16(rg), 5)) );
ba = vorrq_u16( ba, vreinterpretq_u16_u8(vshrq_n_u8(vreinterpretq_u8_u16(ba), 5)) );
ba = vorrq_u16( ba, srcAlphaBits );
dstLo = vzip1q_u16(rg, ba);
@ -79,12 +78,11 @@ FORCEINLINE void ColorspaceConvert555XTo888X_NEON(const v128u16 &srcColor, v128u
}
else
{
const v128u16 r = vandq_u16( vshlq_n_u16(srcColor, 3), vdupq_n_u16(0x00F8) );
v128u16 rg = vorrq_u16( r, vandq_u16(vshlq_n_u16(srcColor, 6), vdupq_n_u16(0xF800)) );
rg = vorrq_u16( rg, vandq_u16(vshrq_n_u16(rg, 5), vdupq_n_u16(0x0707)) );
v128u16 rg = vorrq_u16( vandq_u16( vshlq_n_u16(srcColor,3), vdupq_n_u16(0x00F8) ), vandq_u16( vshlq_n_u16(srcColor,6), vdupq_n_u16(0xF800) ) );
v128u16 b = vandq_u16( vshrq_n_u16(srcColor,7), vdupq_n_u16(0x00F8) );
b = vorrq_u16(b, vshrq_n_u16(b, 5));
rg = vorrq_u16( rg, vreinterpretq_u16_u8(vshrq_n_u8(vreinterpretq_u8_u16(rg), 5)) );
b = vorrq_u16( b, vreinterpretq_u16_u8(vshrq_n_u8(vreinterpretq_u8_u16( b), 5)) );
dstLo = vzip1q_u16(rg, b);
dstHi = vzip2q_u16(rg, b);
@ -111,13 +109,11 @@ FORCEINLINE void ColorspaceConvert555To6665_NEON(const v128u16 &srcColor, const
}
else
{
const v128u16 r = vandq_u16( vshlq_n_u16(srcColor, 1), vdupq_n_u16(0x003E) );
const v128u16 b = vandq_u16( vshrq_n_u16(srcColor, 9), vdupq_n_u16(0x003E) );
v128u16 rg = vorrq_u16( vandq_u16( vshlq_n_u16(srcColor,1), vdupq_n_u16(0x003E) ), vandq_u16( vshlq_n_u16(srcColor,4), vdupq_n_u16(0x3E00) ) );
v128u16 ba = vandq_u16( vshrq_n_u16(srcColor,9), vdupq_n_u16(0x003E) );
v128u16 rg = vorrq_u16( r, vandq_u16(vshlq_n_u16(srcColor, 4), vdupq_n_u16(0x3E00)) );
rg = vorrq_u16( rg, vandq_u16(vshrq_n_u16(rg, 5), vdupq_n_u16(0x0101)) );
v128u16 ba = vorrq_u16(b, vshrq_n_u16(b, 5));
rg = vorrq_u16( rg, vreinterpretq_u16_u8(vshrq_n_u8(vreinterpretq_u8_u16(rg), 5)) );
ba = vorrq_u16( ba, vreinterpretq_u16_u8(vshrq_n_u8(vreinterpretq_u8_u16(ba), 5)) );
ba = vorrq_u16( ba, srcAlphaBits );
dstLo = vzip1q_u16(rg, ba);
@ -144,12 +140,11 @@ FORCEINLINE void ColorspaceConvert555XTo666X_NEON(const v128u16 &srcColor, v128u
}
else
{
const v128u16 r = vandq_u16( vshlq_n_u16(srcColor, 1), vdupq_n_u16(0x003E) );
v128u16 rg = vorrq_u16( r, vandq_u16(vshlq_n_u16(srcColor, 4), vdupq_n_u16(0x3E00)) );
rg = vorrq_u16( rg, vandq_u16(vshrq_n_u16(rg, 5), vdupq_n_u16(0x0101)) );
v128u16 rg = vorrq_u16( vandq_u16( vshlq_n_u16(srcColor,1), vdupq_n_u16(0x003E) ), vandq_u16( vshlq_n_u16(srcColor,4), vdupq_n_u16(0x3E00) ) );
v128u16 b = vandq_u16( vshrq_n_u16(srcColor,9), vdupq_n_u16(0x003E) );
b = vorrq_u16(b, vshrq_n_u16(b, 5));
rg = vorrq_u16( rg, vreinterpretq_u16_u8(vshrq_n_u8(vreinterpretq_u8_u16(rg), 5)) );
b = vorrq_u16( b, vreinterpretq_u16_u8(vshrq_n_u8(vreinterpretq_u8_u16( b), 5)) );
dstLo = vzip1q_u16(rg, b);
dstHi = vzip2q_u16(rg, b);
@ -176,15 +171,14 @@ FORCEINLINE v128u32 ColorspaceConvert8888To6665_NEON(const v128u32 &src)
// Conversion algorithm:
// RGB 8-bit to 6-bit formula: dstRGB6 = (srcRGB8 >> 2)
// Alpha 8-bit to 6-bit formula: dstA5 = (srcA8 >> 3)
v128u32 rgb = vandq_u32( vshrq_n_u32(src, 2), vdupq_n_u32(0x003F3F3F) );
const v128u32 a = vandq_u32( vshrq_n_u32(src, 3), vdupq_n_u32(0x1F000000) );
v128u8 rgba = vshlq_u8(vreinterpretq_u8_u32(src), ((v128s8){-2,-2,-2,-3, -2,-2,-2,-3, -2,-2,-2,-3, -2,-2,-2,-3}));
if (SWAP_RB)
{
rgb = COLOR32_SWAPRB_NEON(rgb);
rgba = COLOR32_SWAPRB_NEON(rgba);
}
return vorrq_u32(rgb, a);
return vreinterpretq_u32_u8(rgba);
}
template <bool SWAP_RB>
@ -193,15 +187,14 @@ FORCEINLINE v128u32 ColorspaceConvert6665To8888_NEON(const v128u32 &src)
// Conversion algorithm:
// RGB 6-bit to 8-bit formula: dstRGB8 = (srcRGB6 << 2) | ((srcRGB6 >> 4) & 0x03)
// Alpha 5-bit to 8-bit formula: dstA8 = (srcA5 << 3) | ((srcA5 >> 2) & 0x07)
v128u32 rgb = vorrq_u32( vandq_u32(vshlq_n_u32(src, 2), vdupq_n_u32(0x00FCFCFC)), vandq_u32(vshrq_n_u32(src, 4), vdupq_n_u32(0x00030303)) );
const v128u32 a = vorrq_u32( vandq_u32(vshlq_n_u32(src, 3), vdupq_n_u32(0xF8000000)), vandq_u32(vshrq_n_u32(src, 2), vdupq_n_u32(0x07000000)) );
v128u8 rgba = vorrq_u8( vshlq_u8(vreinterpretq_u8_u32(src), ((v128s8){2,2,2,3, 2,2,2,3, 2,2,2,3, 2,2,2,3})), vshlq_u8(vreinterpretq_u8_u32(src), ((v128s8){-4,-4,-4,-2, -4,-4,-4,-2, -4,-4,-4,-2, -4,-4,-4,-2})) );
if (SWAP_RB)
{
rgb = COLOR32_SWAPRB_NEON(rgb);
rgba = COLOR32_SWAPRB_NEON(rgba);
}
return vorrq_u32(rgb, a);
return vreinterpretq_u32_u8(rgba);
}
template <NDSColorFormat COLORFORMAT, bool SWAP_RB>