diff options
Diffstat (limited to 'src/post/deinterlace/plugins/tomsmocomp/SearchLoopTop.inc')
| -rw-r--r-- | src/post/deinterlace/plugins/tomsmocomp/SearchLoopTop.inc | 54 |
1 files changed, 18 insertions, 36 deletions
diff --git a/src/post/deinterlace/plugins/tomsmocomp/SearchLoopTop.inc b/src/post/deinterlace/plugins/tomsmocomp/SearchLoopTop.inc index 17a3616ee..38140909a 100644 --- a/src/post/deinterlace/plugins/tomsmocomp/SearchLoopTop.inc +++ b/src/post/deinterlace/plugins/tomsmocomp/SearchLoopTop.inc @@ -6,17 +6,6 @@ const unsigned char* pSrc; const unsigned char* pBob; const unsigned char* pBobP; -int64_t Max_Mov = 0x0404040404040404ull; -int64_t DiffThres = 0x0f0f0f0f0f0f0f0full; -int64_t YMask = 0x00ff00ff00ff00ffull; // keeps only luma -int64_t UVMask = 0xff00ff00ff00ff00ull; // keeps only chroma -int64_t TENS = 0x0a0a0a0a0a0a0a0aull; -int64_t FOURS = 0x0404040404040404ull; -int64_t ONES = 0x0101010101010101ull; -int64_t Min_Vals = 0x0000000000000000ull; -int64_t Max_Vals = 0x0000000000000000ull; -int64_t ShiftMask = 0xfefffefffefffeffull; - // long is int32 on ARCH_368, int64 on ARCH_AMD64. Declaring it this way // saves a lot of xor's to delete 64bit garbage. @@ -71,25 +60,18 @@ long dst_pitchw = dst_pitch; // local stor so asm can ref #ifndef _pBob #define _pBob "%0" #define _src_pitch2 "%1" -#define _ShiftMask "%2" -#define _pDest "%3" -#define _dst_pitchw "%4" -#define _Last8 "%5" -#define _pSrc "%6" -#define _pSrcP "%7" -#define _pBobP "%8" -#define _DiffThres "%9" -#define _Min_Vals "%10" -#define _Max_Vals "%11" -#define _FOURS "%12" -#define _TENS "%13" -#define _ONES "%14" -#define _UVMask "%15" -#define _Max_Mov "%16" -#define _YMask "%17" -#define _oldbx "%18" +#define _pDest "%2" +#define _dst_pitchw "%3" +#define _Last8 "%4" +#define _pSrc "%5" +#define _pSrcP "%6" +#define _pBobP "%7" +#define _oldbx "%8" #endif + Min_Vals = 0x0000000000000000ull; + Max_Vals = 0x0000000000000000ull; + for (y=1; y < FldHeight-1; y++) { // pretend it's indented -->> @@ -120,9 +102,9 @@ long dst_pitchw = dst_pitch; // local stor so asm can ref "movq (%%"XBX"), %%mm0\n\t" "movq (%%"XBX", %%"XCX"), %%mm1\n\t" //, qword ptr["XBX"+"XCX"] "movq %%mm0, %%mm2\n\t" - V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // halfway between - V_PAVGB ("%%mm0", "%%mm2", "%%mm3", _ShiftMask) // 1/4 way - V_PAVGB ("%%mm1", "%%mm2", "%%mm3", _ShiftMask) // 3/4 way + V_PAVGB ("%%mm2", "%%mm1", "%%mm3", MANGLE(ShiftMask)) // halfway between + V_PAVGB ("%%mm0", "%%mm2", "%%mm3", MANGLE(ShiftMask)) // 1/4 way + V_PAVGB ("%%mm1", "%%mm2", "%%mm3", MANGLE(ShiftMask)) // 3/4 way MOVX" "_pDest", %%"XDI"\n\t" MOVX" "_dst_pitchw", %%"XAX"\n\t" V_MOVNTQ ("(%%"XDI")", "%%mm0") @@ -134,9 +116,9 @@ long dst_pitchw = dst_pitch; // local stor so asm can ref "movq (%%"XSI"), %%mm0\n\t" "movq (%%"XSI", %%"XCX"), %%mm1\n\t" // qword ptr["XSI"+"XCX"] "movq %%mm0, %%mm2\n\t" - V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // halfway between - V_PAVGB ("%%mm0", "%%mm2", "%%mm3", _ShiftMask) // 1/4 way - V_PAVGB ("%%mm1", "%%mm2", "%%mm3", _ShiftMask) // 3/4 way + V_PAVGB ("%%mm2", "%%mm1", "%%mm3", MANGLE(ShiftMask)) // halfway between + V_PAVGB ("%%mm0", "%%mm2", "%%mm3", MANGLE(ShiftMask)) // 1/4 way + V_PAVGB ("%%mm1", "%%mm2", "%%mm3", MANGLE(ShiftMask)) // 3/4 way ADDX" %%"XDX", %%"XDI"\n\t" // last 8 bytes of dest V_MOVNTQ ("%%"XDI"", "%%mm0") V_MOVNTQ ("(%%"XDI", %%"XAX")", "%%mm1") // qword ptr["XDI"+"XAX"], mm1) @@ -144,7 +126,7 @@ long dst_pitchw = dst_pitch; // local stor so asm can ref #else "movq (%%"XBX"), %%mm0\n\t" // pavgb mm0, qword ptr["XBX"+"XCX"] - V_PAVGB ("%%mm0", "(%%"XBX", %%"XCX")", "%%mm2", _ShiftMask) // qword ptr["XBX"+"XCX"], mm2, ShiftMask) + V_PAVGB ("%%mm0", "(%%"XBX", %%"XCX")", "%%mm2", MANGLE(ShiftMask)) // qword ptr["XBX"+"XCX"], mm2, ShiftMask) MOVX" "_pDest", %%"XDI"\n\t" V_MOVNTQ ("(%%"XDI")", "%%mm0") @@ -153,7 +135,7 @@ long dst_pitchw = dst_pitch; // local stor so asm can ref LEAX" (%%"XBX", %%"XDX"), %%"XSI"\n\t" //"XSI", ["XBX"+"XDX"] "movq (%%"XSI"), %%mm0\n\t" // pavgb mm0, qword ptr["XSI"+"XCX"] - V_PAVGB ("%%mm0", "(%%"XSI", %%"XCX")", "%%mm2", _ShiftMask) // qword ptr["XSI"+"XCX"], mm2, ShiftMask) + V_PAVGB ("%%mm0", "(%%"XSI", %%"XCX")", "%%mm2", MANGLE(ShiftMask)) // qword ptr["XSI"+"XCX"], mm2, ShiftMask) V_MOVNTQ ("(%%"XDI", %%"XDX")", "%%mm0") // qword ptr["XDI"+"XDX"], mm0) #endif // now loop and get the middle qwords |