summaryrefslogtreecommitdiff
path: root/src/post/deinterlace/plugins/greedyh.asm
diff options
context:
space:
mode:
Diffstat (limited to 'src/post/deinterlace/plugins/greedyh.asm')
-rw-r--r--src/post/deinterlace/plugins/greedyh.asm66
1 files changed, 36 insertions, 30 deletions
diff --git a/src/post/deinterlace/plugins/greedyh.asm b/src/post/deinterlace/plugins/greedyh.asm
index 0bbd745aa..11b28ca76 100644
--- a/src/post/deinterlace/plugins/greedyh.asm
+++ b/src/post/deinterlace/plugins/greedyh.asm
@@ -43,7 +43,6 @@ static void FUNCT_NAME(uint8_t *output, int outstride,
int Line;
long LoopCtr;
- long oldbx;
unsigned int Pitch = stride*2;
int FieldHeight = height / 2;
@@ -52,6 +51,7 @@ static void FUNCT_NAME(uint8_t *output, int outstride,
unsigned char* L3; // ptr to Line3
unsigned char* L2P; // ptr to prev Line2
+ unsigned char* temp;
unsigned char* Dest = output;
int64_t LastAvg=0; //interp value from left qword
@@ -121,25 +121,21 @@ static void FUNCT_NAME(uint8_t *output, int outstride,
#define asmLastAvg "%0"
#define asmL1 "%1"
#define asmL3 "%2"
-#define asmL2P "%3"
+#define asmtemp "%3"
#define asmL2 "%4"
#define asmDest "%5"
#define asmLoopCtr "%6"
-#define asmoldbx "%7"
#endif
// For ease of reading, the comments below assume that we're operating on an odd
// field (i.e., that InfoIsOdd is true). Assume the obvious for even lines..
+ temp = L2P;
__asm__ __volatile__
(
- // save ebx (-fPIC)
- MOVX" %%"XBX", "asmoldbx"\n\t"
-
MOVX" "asmL1", %%"XAX"\n\t"
- LEAX" 8(%%"XAX"), %%"XBX"\n\t" // next qword needed by DJR
+ LEAX" 8(%%"XAX"), %%"XDX"\n\t" // next qword needed by DJR
MOVX" "asmL3", %%"XCX"\n\t"
SUBX" %%"XAX", %%"XCX"\n\t" // carry L3 addr as an offset
- MOVX" "asmL2P", %%"XDX"\n\t"
MOVX" "asmL2", %%"XSI"\n\t"
MOVX" "asmDest", %%"XDI"\n\t" // DL1 if Odd or DL2 if Even
@@ -148,11 +144,14 @@ static void FUNCT_NAME(uint8_t *output, int outstride,
"movq (%%"XSI"), %%mm0\n\t" // L2 - the newest weave pixel value
"movq (%%"XAX"), %%mm1\n\t" // L1 - the top pixel
+ PUSHX" %%"XDX "\n\t"
+ MOVX" "asmtemp", %%"XDX"\n\t"
"movq (%%"XDX"), %%mm2\n\t" // L2P - the prev weave pixel
+ POPX" %%"XDX "\n\t"
"movq (%%"XAX", %%"XCX"), %%mm3\n\t" // L3, next odd row
"movq %%mm1, %%mm6\n\t" // L1 - get simple single pixel interp
// pavgb mm6, mm3 // use macro below
- V_PAVGB ("%%mm6", "%%mm3", "%%mm4", MANGLE(ShiftMask))
+ V_PAVGB ("%%mm6", "%%mm3", "%%mm4", "%8")
// DJR - Diagonal Jaggie Reduction
// In the event that we are going to use an average (Bob) pixel we do not want a jagged
@@ -166,24 +165,24 @@ static void FUNCT_NAME(uint8_t *output, int outstride,
"psllq $16, %%mm7\n\t" // left justify 3 pixels
"por %%mm7, %%mm4\n\t" // and combine
- "movq (%%"XBX"), %%mm5\n\t" // next horiz qword from L1
+ "movq (%%"XDX"), %%mm5\n\t" // next horiz qword from L1
// pavgb mm5, qword ptr[ebx+ecx] // next horiz qword from L3, use macro below
- V_PAVGB ("%%mm5", "(%%"XBX",%%"XCX")", "%%mm7", MANGLE(ShiftMask))
+ V_PAVGB ("%%mm5", "(%%"XDX",%%"XCX")", "%%mm7", "%8")
"psllq $48, %%mm5\n\t" // left just 1 pixel
"movq %%mm6, %%mm7\n\t" // another copy of simple bob pixel
"psrlq $16, %%mm7\n\t" // right just 3 pixels
"por %%mm7, %%mm5\n\t" // combine
// pavgb mm4, mm5 // avg of forward and prev by 1 pixel, use macro
- V_PAVGB ("%%mm4", "%%mm5", "%%mm5", MANGLE(ShiftMask)) // mm5 gets modified if MMX
+ V_PAVGB ("%%mm4", "%%mm5", "%%mm5", "%8") // mm5 gets modified if MMX
// pavgb mm6, mm4 // avg of center and surround interp vals, use macro
- V_PAVGB ("%%mm6", "%%mm4", "%%mm7", MANGLE(ShiftMask))
+ V_PAVGB ("%%mm6", "%%mm4", "%%mm7", "%8")
// Don't do any more averaging than needed for mmx. It hurts performance and causes rounding errors.
#ifndef IS_MMX
// pavgb mm4, mm6 // 1/4 center, 3/4 adjacent
- V_PAVGB ("%%mm4", "%%mm6", "%%mm7", MANGLE(ShiftMask))
+ V_PAVGB ("%%mm4", "%%mm6", "%%mm7", "%8")
// pavgb mm6, mm4 // 3/8 center, 5/8 adjacent
- V_PAVGB ("%%mm6", "%%mm4", "%%mm7", MANGLE(ShiftMask))
+ V_PAVGB ("%%mm6", "%%mm4", "%%mm7", "%8")
#endif
// get abs value of possible L2 comb
@@ -236,64 +235,71 @@ static void FUNCT_NAME(uint8_t *output, int outstride,
// pminub mm5, mm3 // now = Min(L1,L3), use macro
V_PMINUB ("%%mm5", "%%mm3", "%%mm7")
// allow the value to be above the high or below the low by amt of MaxComb
- "psubusb "MANGLE(MaxComb)", %%mm5\n\t" // lower min by diff
- "paddusb "MANGLE(MaxComb)", %%mm2\n\t" // increase max by diff
+ "psubusb %9, %%mm5\n\t" // lower min by diff
+ "paddusb %9, %%mm2\n\t" // increase max by diff
// pmaxub mm4, mm5 // now = Max(best,Min(L1,L3) use macro
V_PMAXUB ("%%mm4", "%%mm5")
// pminub mm4, mm2 // now = Min( Max(best, Min(L1,L3), L2 )=L2 clipped
V_PMINUB ("%%mm4", "%%mm2", "%%mm7")
// Blend weave pixel with bob pixel, depending on motion val in mm0
- "psubusb "MANGLE(MotionThreshold)", %%mm0\n\t"// test Threshold, clear chroma change >>>??
- "pmullw "MANGLE(MotionSense)", %%mm0\n\t" // mul by user factor, keep low 16 bits
- "movq "MANGLE(QW256)", %%mm7\n\t"
+ "psubusb %10, %%mm0\n\t"// test Threshold, clear chroma change >>>??
+ "pmullw %11, %%mm0\n\t" // mul by user factor, keep low 16 bits
+ "movq %12, %%mm7\n\t"
#ifdef IS_SSE
"pminsw %%mm7, %%mm0\n\t" // max = 256
#else
- "paddusw "MANGLE(QW256B)", %%mm0\n\t" // add, may sat at fff..
- "psubusw "MANGLE(QW256B)", %%mm0\n\t" // now = Min(L1,256)
+ "paddusw %13, %%mm0\n\t" // add, may sat at fff..
+ "psubusw %13, %%mm0\n\t" // now = Min(L1,256)
#endif
"psubusw %%mm0, %%mm7\n\t" // so the 2 sum to 256, weighted avg
"movq %%mm4, %%mm2\n\t" // save weave chroma info before trashing
- "pand "MANGLE(YMask)", %%mm4\n\t" // keep only luma from calc'd value
+ "pand %14, %%mm4\n\t" // keep only luma from calc'd value
"pmullw %%mm7, %%mm4\n\t" // use more weave for less motion
- "pand "MANGLE(YMask)", %%mm6\n\t" // keep only luma from calc'd value
+ "pand %14, %%mm6\n\t" // keep only luma from calc'd value
"pmullw %%mm0, %%mm6\n\t" // use more bob for large motion
"paddusw %%mm6, %%mm4\n\t" // combine
"psrlw $8, %%mm4\n\t" // div by 256 to get weighted avg
// chroma comes from weave pixel
- "pand "MANGLE(UVMask)", %%mm2\n\t" // keep chroma
+ "pand %15, %%mm2\n\t" // keep chroma
"por %%mm4, %%mm2\n\t" // and combine
V_MOVNTQ ("(%%"XDI")", "%%mm2") // move in our clipped best, use macro
// bump ptrs and loop
LEAX" 8(%%"XAX"), %%"XAX"\n\t"
- LEAX" 8(%%"XBX"), %%"XBX"\n\t"
LEAX" 8(%%"XDX"), %%"XDX"\n\t"
+ ADDX" $8, "asmtemp"\n\t"
LEAX" 8(%%"XDI"), %%"XDI"\n\t"
LEAX" 8(%%"XSI"), %%"XSI"\n\t"
DECX" "asmLoopCtr"\n\t"
"jg 1b\n\t" // loop if not to last line
// note P-III default assumes backward branches taken
"jl 1f\n\t" // done
- MOVX" %%"XAX", %%"XBX"\n\t" // sharpness lookahead 1 byte only, be wrong on 1
+ MOVX" %%"XAX", %%"XDX"\n\t" // sharpness lookahead 1 byte only, be wrong on 1
"jmp 1b\n\t"
"1:\n\t"
- MOVX" "asmoldbx", %%"XBX"\n\t"
: /* no outputs */
: "m"(LastAvg),
"m"(L1),
"m"(L3),
- "m"(L2P),
+ "m"(temp),
"m"(L2),
"m"(Dest),
"m"(LoopCtr),
- "m"(oldbx)
+ "m"(temp),
+ "m"(ShiftMask),
+ "m"(MaxComb),
+ "m"(MotionThreshold),
+ "m"(MotionSense),
+ "m"(QW256),
+ "m"(QW256B),
+ "m"(YMask),
+ "m"(UVMask)
: XAX, XCX, XDX, XSI, XDI,
#ifdef ARCH_X86
generated by cgit v1.2.3 (git 2.25.1) at 2025-04-26 08:13:17 +0000