打印编程

开发平台：

Visual C++

pngvcrd.c：源码内容

inc ebx
shr al, 1 // divide by 2
add al, [edi+ebx-1] // Add Avg(x); -1 to offset inc ebx
cmp ebx, bpp
mov [edi+ebx-1], al // Write back Raw(x);
// mov does not affect flags; -1 to offset inc ebx
jb davgrlp
// get # of bytes to alignment
mov diff, edi // take start of row
add diff, ebx // add bpp
add diff, 0xf // add 7 + 8 to incr past alignment boundary
and diff, 0xfffffff8 // mask to alignment boundary
sub diff, edi // subtract from start ==> value ebx at alignment
jz davggo
// fix alignment
// Compute the Raw value for the bytes upto the alignment boundary
// Raw(x) = Avg(x) + ((Raw(x-bpp) + Prior(x))/2)
xor ecx, ecx
davglp1:
xor eax, eax
mov cl, [esi + ebx] // load cl with Prior(x)
mov al, [edx + ebx] // load al with Raw(x-bpp)
add ax, cx
inc ebx
shr ax, 1 // divide by 2
add al, [edi+ebx-1] // Add Avg(x); -1 to offset inc ebx
cmp ebx, diff // Check if at alignment boundary
mov [edi+ebx-1], al // Write back Raw(x);
// mov does not affect flags; -1 to offset inc ebx
jb davglp1 // Repeat until at alignment boundary
davggo:
mov eax, FullLength
mov ecx, eax
sub eax, ebx // subtract alignment fix
and eax, 0x00000007 // calc bytes over mult of 8
sub ecx, eax // drop over bytes from original length
mov MMXLength, ecx
} // end _asm block
// Now do the math for the rest of the row
switch ( bpp )
{
case 3:
{
ActiveMask.use = 0x0000000000ffffff;
ShiftBpp.use = 24; // == 3 * 8
ShiftRem.use = 40; // == 64 - 24
_asm {
// Re-init address pointers and offset
movq mm7, ActiveMask
mov ebx, diff // ebx ==> x = offset to alignment boundary
movq mm5, LBCarryMask
mov edi, row // edi ==> Avg(x)
movq mm4, HBClearMask
mov esi, prev_row // esi ==> Prior(x)
// PRIME the pump (load the first Raw(x-bpp) data set
movq mm2, [edi + ebx - 8] // Load previous aligned 8 bytes
// (we correct position in loop below)
davg3lp:
movq mm0, [edi + ebx] // Load mm0 with Avg(x)
// Add (Prev_row/2) to Average
movq mm3, mm5
psrlq mm2, ShiftRem // Correct position Raw(x-bpp) data
movq mm1, [esi + ebx] // Load mm1 with Prior(x)
movq mm6, mm7
pand mm3, mm1 // get lsb for each prev_row byte
psrlq mm1, 1 // divide prev_row bytes by 2
pand mm1, mm4 // clear invalid bit 7 of each byte
paddb mm0, mm1 // add (Prev_row/2) to Avg for each byte
// Add 1st active group (Raw(x-bpp)/2) to Average with LBCarry
movq mm1, mm3 // now use mm1 for getting LBCarrys
pand mm1, mm2 // get LBCarrys for each byte where both
// lsb's were == 1 (Only valid for active group)
psrlq mm2, 1 // divide raw bytes by 2
pand mm2, mm4 // clear invalid bit 7 of each byte
paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
pand mm2, mm6 // Leave only Active Group 1 bytes to add to Avg
paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active
// byte
// Add 2nd active group (Raw(x-bpp)/2) to Average with LBCarry
psllq mm6, ShiftBpp // shift the mm6 mask to cover bytes 3-5
movq mm2, mm0 // mov updated Raws to mm2
psllq mm2, ShiftBpp // shift data to position correctly
movq mm1, mm3 // now use mm1 for getting LBCarrys
pand mm1, mm2 // get LBCarrys for each byte where both
// lsb's were == 1 (Only valid for active group)
psrlq mm2, 1 // divide raw bytes by 2
pand mm2, mm4 // clear invalid bit 7 of each byte
paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
pand mm2, mm6 // Leave only Active Group 2 bytes to add to Avg
paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active
// byte
// Add 3rd active group (Raw(x-bpp)/2) to Average with LBCarry
psllq mm6, ShiftBpp // shift the mm6 mask to cover the last two
// bytes
movq mm2, mm0 // mov updated Raws to mm2
psllq mm2, ShiftBpp // shift data to position correctly
// Data only needs to be shifted once here to
// get the correct x-bpp offset.
movq mm1, mm3 // now use mm1 for getting LBCarrys
pand mm1, mm2 // get LBCarrys for each byte where both
// lsb's were == 1 (Only valid for active group)
psrlq mm2, 1 // divide raw bytes by 2
pand mm2, mm4 // clear invalid bit 7 of each byte
paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
pand mm2, mm6 // Leave only Active Group 2 bytes to add to Avg
add ebx, 8
paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active
// byte
// Now ready to write back to memory
movq [edi + ebx - 8], mm0
// Move updated Raw(x) to use as Raw(x-bpp) for next loop
cmp ebx, MMXLength
movq mm2, mm0 // mov updated Raw(x) to mm2
jb davg3lp
} // end _asm block
}
break;
case 6:
case 4:
case 7:
case 5:
{
ActiveMask.use = 0xffffffffffffffff; // use shift below to clear
// appropriate inactive bytes
ShiftBpp.use = bpp << 3;
ShiftRem.use = 64 - ShiftBpp.use;
_asm {
movq mm4, HBClearMask
// Re-init address pointers and offset
mov ebx, diff // ebx ==> x = offset to alignment boundary
// Load ActiveMask and clear all bytes except for 1st active group
movq mm7, ActiveMask
mov edi, row // edi ==> Avg(x)
psrlq mm7, ShiftRem
mov esi, prev_row // esi ==> Prior(x)
movq mm6, mm7
movq mm5, LBCarryMask
psllq mm6, ShiftBpp // Create mask for 2nd active group
// PRIME the pump (load the first Raw(x-bpp) data set
movq mm2, [edi + ebx - 8] // Load previous aligned 8 bytes
// (we correct position in loop below)
davg4lp:
movq mm0, [edi + ebx]
psrlq mm2, ShiftRem // shift data to position correctly
movq mm1, [esi + ebx]
// Add (Prev_row/2) to Average
movq mm3, mm5
pand mm3, mm1 // get lsb for each prev_row byte
psrlq mm1, 1 // divide prev_row bytes by 2
pand mm1, mm4 // clear invalid bit 7 of each byte
paddb mm0, mm1 // add (Prev_row/2) to Avg for each byte
// Add 1st active group (Raw(x-bpp)/2) to Average with LBCarry
movq mm1, mm3 // now use mm1 for getting LBCarrys
pand mm1, mm2 // get LBCarrys for each byte where both
// lsb's were == 1 (Only valid for active group)
psrlq mm2, 1 // divide raw bytes by 2
pand mm2, mm4 // clear invalid bit 7 of each byte
paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
pand mm2, mm7 // Leave only Active Group 1 bytes to add to Avg
paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active
// byte
// Add 2nd active group (Raw(x-bpp)/2) to Average with LBCarry
movq mm2, mm0 // mov updated Raws to mm2
psllq mm2, ShiftBpp // shift data to position correctly
add ebx, 8
movq mm1, mm3 // now use mm1 for getting LBCarrys
pand mm1, mm2 // get LBCarrys for each byte where both
// lsb's were == 1 (Only valid for active group)
psrlq mm2, 1 // divide raw bytes by 2
pand mm2, mm4 // clear invalid bit 7 of each byte
paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
pand mm2, mm6 // Leave only Active Group 2 bytes to add to Avg
paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active
// byte
cmp ebx, MMXLength
// Now ready to write back to memory
movq [edi + ebx - 8], mm0
// Prep Raw(x-bpp) for next loop
movq mm2, mm0 // mov updated Raws to mm2
jb davg4lp
} // end _asm block
}
break;
case 2:
{
ActiveMask.use = 0x000000000000ffff;
ShiftBpp.use = 16; // == 2 * 8 [BUGFIX]
ShiftRem.use = 48; // == 64 - 16 [BUGFIX]
_asm {
// Load ActiveMask
movq mm7, ActiveMask
// Re-init address pointers and offset
mov ebx, diff // ebx ==> x = offset to alignment boundary
movq mm5, LBCarryMask
mov edi, row // edi ==> Avg(x)
movq mm4, HBClearMask
mov esi, prev_row // esi ==> Prior(x)
// PRIME the pump (load the first Raw(x-bpp) data set
movq mm2, [edi + ebx - 8] // Load previous aligned 8 bytes
// (we correct position in loop below)
davg2lp:
movq mm0, [edi + ebx]
psrlq mm2, ShiftRem // shift data to position correctly [BUGFIX]
movq mm1, [esi + ebx]
// Add (Prev_row/2) to Average
movq mm3, mm5
pand mm3, mm1 // get lsb for each prev_row byte
psrlq mm1, 1 // divide prev_row bytes by 2
pand mm1, mm4 // clear invalid bit 7 of each byte
movq mm6, mm7
paddb mm0, mm1 // add (Prev_row/2) to Avg for each byte
// Add 1st active group (Raw(x-bpp)/2) to Average with LBCarry
movq mm1, mm3 // now use mm1 for getting LBCarrys
pand mm1, mm2 // get LBCarrys for each byte where both
// lsb's were == 1 (Only valid for active group)
psrlq mm2, 1 // divide raw bytes by 2
pand mm2, mm4 // clear invalid bit 7 of each byte
paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
pand mm2, mm6 // Leave only Active Group 1 bytes to add to Avg
paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active byte
// Add 2nd active group (Raw(x-bpp)/2) to Average with LBCarry
psllq mm6, ShiftBpp // shift the mm6 mask to cover bytes 2 & 3
movq mm2, mm0 // mov updated Raws to mm2
psllq mm2, ShiftBpp // shift data to position correctly
movq mm1, mm3 // now use mm1 for getting LBCarrys
pand mm1, mm2 // get LBCarrys for each byte where both
// lsb's were == 1 (Only valid for active group)
psrlq mm2, 1 // divide raw bytes by 2
pand mm2, mm4 // clear invalid bit 7 of each byte
paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
pand mm2, mm6 // Leave only Active Group 2 bytes to add to Avg
paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active byte
// Add rdd active group (Raw(x-bpp)/2) to Average with LBCarry
psllq mm6, ShiftBpp // shift the mm6 mask to cover bytes 4 & 5
movq mm2, mm0 // mov updated Raws to mm2
psllq mm2, ShiftBpp // shift data to position correctly
// Data only needs to be shifted once here to
// get the correct x-bpp offset.
movq mm1, mm3 // now use mm1 for getting LBCarrys
pand mm1, mm2 // get LBCarrys for each byte where both
// lsb's were == 1 (Only valid for active group)
psrlq mm2, 1 // divide raw bytes by 2
pand mm2, mm4 // clear invalid bit 7 of each byte
paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
pand mm2, mm6 // Leave only Active Group 2 bytes to add to Avg
paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active byte
// Add 4th active group (Raw(x-bpp)/2) to Average with LBCarry
psllq mm6, ShiftBpp // shift the mm6 mask to cover bytes 6 & 7
movq mm2, mm0 // mov updated Raws to mm2
psllq mm2, ShiftBpp // shift data to position correctly
// Data only needs to be shifted once here to
// get the correct x-bpp offset.
add ebx, 8
movq mm1, mm3 // now use mm1 for getting LBCarrys
pand mm1, mm2 // get LBCarrys for each byte where both
// lsb's were == 1 (Only valid for active group)
psrlq mm2, 1 // divide raw bytes by 2
pand mm2, mm4 // clear invalid bit 7 of each byte
paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
pand mm2, mm6 // Leave only Active Group 2 bytes to add to Avg
paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active byte
cmp ebx, MMXLength
// Now ready to write back to memory
movq [edi + ebx - 8], mm0
// Prep Raw(x-bpp) for next loop
movq mm2, mm0 // mov updated Raws to mm2
jb davg2lp
} // end _asm block
}
break;
case 1: // bpp == 1
{
_asm {
// Re-init address pointers and offset
mov ebx, diff // ebx ==> x = offset to alignment boundary
mov edi, row // edi ==> Avg(x)
cmp ebx, FullLength // Test if offset at end of array
jnb davg1end
// Do Paeth decode for remaining bytes
mov esi, prev_row // esi ==> Prior(x)
mov edx, edi
xor ecx, ecx // zero ecx before using cl & cx in loop below
sub edx, bpp // edx ==> Raw(x-bpp)
davg1lp:
// Raw(x) = Avg(x) + ((Raw(x-bpp) + Prior(x))/2)
xor eax, eax
mov cl, [esi + ebx] // load cl with Prior(x)
mov al, [edx + ebx] // load al with Raw(x-bpp)
add ax, cx
inc ebx
shr ax, 1 // divide by 2
add al, [edi+ebx-1] // Add Avg(x); -1 to offset inc ebx
cmp ebx, FullLength // Check if at end of array
mov [edi+ebx-1], al // Write back Raw(x);
// mov does not affect flags; -1 to offset inc ebx
jb davg1lp
davg1end:
} // end _asm block
}
return;
case 8: // bpp == 8
{
_asm {
// Re-init address pointers and offset
mov ebx, diff // ebx ==> x = offset to alignment boundary
movq mm5, LBCarryMask
mov edi, row // edi ==> Avg(x)
movq mm4, HBClearMask
mov esi, prev_row // esi ==> Prior(x)
// PRIME the pump (load the first Raw(x-bpp) data set
movq mm2, [edi + ebx - 8] // Load previous aligned 8 bytes
// (NO NEED to correct position in loop below)
davg8lp:
movq mm0, [edi + ebx]
movq mm3, mm5
movq mm1, [esi + ebx]
add ebx, 8
pand mm3, mm1 // get lsb for each prev_row byte
psrlq mm1, 1 // divide prev_row bytes by 2
pand mm3, mm2 // get LBCarrys for each byte where both
// lsb's were == 1
psrlq mm2, 1 // divide raw bytes by 2
pand mm1, mm4 // clear invalid bit 7 of each byte
paddb mm0, mm3 // add LBCarrys to Avg for each byte
pand mm2, mm4 // clear invalid bit 7 of each byte
paddb mm0, mm1 // add (Prev_row/2) to Avg for each byte
paddb mm0, mm2 // add (Raw/2) to Avg for each byte
cmp ebx, MMXLength
movq [edi + ebx - 8], mm0
movq mm2, mm0 // reuse as Raw(x-bpp)
jb davg8lp
} // end _asm block
}
break;
default: // bpp greater than 8
{
_asm {
movq mm5, LBCarryMask
// Re-init address pointers and offset
mov ebx, diff // ebx ==> x = offset to alignment boundary
mov edi, row // edi ==> Avg(x)
movq mm4, HBClearMask
mov edx, edi
mov esi, prev_row // esi ==> Prior(x)
sub edx, bpp // edx ==> Raw(x-bpp)
davgAlp:
movq mm0, [edi + ebx]
movq mm3, mm5
movq mm1, [esi + ebx]
pand mm3, mm1 // get lsb for each prev_row byte
movq mm2, [edx + ebx]
psrlq mm1, 1 // divide prev_row bytes by 2
pand mm3, mm2 // get LBCarrys for each byte where both
// lsb's were == 1
psrlq mm2, 1 // divide raw bytes by 2
pand mm1, mm4 // clear invalid bit 7 of each byte
paddb mm0, mm3 // add LBCarrys to Avg for each byte
pand mm2, mm4 // clear invalid bit 7 of each byte
paddb mm0, mm1 // add (Prev_row/2) to Avg for each byte
add ebx, 8
paddb mm0, mm2 // add (Raw/2) to Avg for each byte
cmp ebx, MMXLength
movq [edi + ebx - 8], mm0
jb davgAlp
} // end _asm block
}
break;
} // end switch ( bpp )
_asm {
// MMX acceleration complete now do clean-up
// Check if any remaining bytes left to decode
mov ebx, MMXLength // ebx ==> x = offset bytes remaining after MMX
mov edi, row // edi ==> Avg(x)
cmp ebx, FullLength // Test if offset at end of array
jnb davgend
// Do Paeth decode for remaining bytes
mov esi, prev_row // esi ==> Prior(x)
mov edx, edi
xor ecx, ecx // zero ecx before using cl & cx in loop below
sub edx, bpp // edx ==> Raw(x-bpp)
davglp2:
// Raw(x) = Avg(x) + ((Raw(x-bpp) + Prior(x))/2)
xor eax, eax
mov cl, [esi + ebx] // load cl with Prior(x)
mov al, [edx + ebx] // load al with Raw(x-bpp)
add ax, cx
inc ebx
shr ax, 1 // divide by 2
add al, [edi+ebx-1] // Add Avg(x); -1 to offset inc ebx
cmp ebx, FullLength // Check if at end of array
mov [edi+ebx-1], al // Write back Raw(x);
// mov does not affect flags; -1 to offset inc ebx
jb davglp2
davgend:
emms // End MMX instructions; prep for possible FP instrs.
} // end _asm block
}
// Optimized code for PNG Paeth filter decoder
void /* PRIVATE */
png_read_filter_row_mmx_paeth(png_row_infop row_info, png_bytep row,
png_bytep prev_row)
{
png_uint_32 FullLength;
png_uint_32 MMXLength;
//png_uint_32 len;
int bpp;
int diff;
//int ptemp;
int patemp, pbtemp, pctemp;
bpp = (row_info->pixel_depth + 7) >> 3; // Get # bytes per pixel
FullLength = row_info->rowbytes; // # of bytes to filter
_asm
{
xor ebx, ebx // ebx ==> x offset
mov edi, row
xor edx, edx // edx ==> x-bpp offset
mov esi, prev_row
xor eax, eax
// Compute the Raw value for the first bpp bytes
// Note: the formula works out to be always
// Paeth(x) = Raw(x) + Prior(x) where x < bpp
dpthrlp:
mov al, [edi + ebx]
add al, [esi + ebx]
inc ebx
cmp ebx, bpp
mov [edi + ebx - 1], al
jb dpthrlp
// get # of bytes to alignment
mov diff, edi // take start of row
add diff, ebx // add bpp
xor ecx, ecx
add diff, 0xf // add 7 + 8 to incr past alignment boundary
and diff, 0xfffffff8 // mask to alignment boundary
sub diff, edi // subtract from start ==> value ebx at alignment
jz dpthgo
// fix alignment
dpthlp1:
xor eax, eax
// pav = p - a = (a + b - c) - a = b - c
mov al, [esi + ebx] // load Prior(x) into al
mov cl, [esi + edx] // load Prior(x-bpp) into cl
sub eax, ecx // subtract Prior(x-bpp)
mov patemp, eax // Save pav for later use
xor eax, eax
// pbv = p - b = (a + b - c) - b = a - c
mov al, [edi + edx] // load Raw(x-bpp) into al
sub eax, ecx // subtract Prior(x-bpp)
mov ecx, eax
// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
add eax, patemp // pcv = pav + pbv
// pc = abs(pcv)
test eax, 0x80000000
jz dpthpca
neg eax // reverse sign of neg values
dpthpca:
mov pctemp, eax // save pc for later use
// pb = abs(pbv)
test ecx, 0x80000000
jz dpthpba
neg ecx // reverse sign of neg values
dpthpba:
mov pbtemp, ecx // save pb for later use
// pa = abs(pav)
mov eax, patemp
test eax, 0x80000000
jz dpthpaa
neg eax // reverse sign of neg values
dpthpaa:
mov patemp, eax // save pa for later use
// test if pa <= pb
cmp eax, ecx
jna dpthabb
// pa > pb; now test if pb <= pc
cmp ecx, pctemp
jna dpthbbc
// pb > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
mov cl, [esi + edx] // load Prior(x-bpp) into cl
jmp dpthpaeth
dpthbbc:
// pb <= pc; Raw(x) = Paeth(x) + Prior(x)
mov cl, [esi + ebx] // load Prior(x) into cl
jmp dpthpaeth
dpthabb:
// pa <= pb; now test if pa <= pc
cmp eax, pctemp
jna dpthabc
// pa > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
mov cl, [esi + edx] // load Prior(x-bpp) into cl
jmp dpthpaeth
dpthabc:
// pa <= pc; Raw(x) = Paeth(x) + Raw(x-bpp)
mov cl, [edi + edx] // load Raw(x-bpp) into cl
dpthpaeth:
inc ebx
inc edx
// Raw(x) = (Paeth(x) + Paeth_Predictor( a, b, c )) mod 256
add [edi + ebx - 1], cl
cmp ebx, diff
jb dpthlp1
dpthgo:
mov ecx, FullLength
mov eax, ecx
sub eax, ebx // subtract alignment fix
and eax, 0x00000007 // calc bytes over mult of 8
sub ecx, eax // drop over bytes from original length
mov MMXLength, ecx
} // end _asm block
// Now do the math for the rest of the row
switch ( bpp )
{
case 3:
{
ActiveMask.use = 0x0000000000ffffff;
ActiveMaskEnd.use = 0xffff000000000000;
ShiftBpp.use = 24; // == bpp(3) * 8
ShiftRem.use = 40; // == 64 - 24
_asm
{
mov ebx, diff
mov edi, row
mov esi, prev_row
pxor mm0, mm0
// PRIME the pump (load the first Raw(x-bpp) data set
movq mm1, [edi+ebx-8]
dpth3lp:
psrlq mm1, ShiftRem // shift last 3 bytes to 1st 3 bytes
movq mm2, [esi + ebx] // load b=Prior(x)
punpcklbw mm1, mm0 // Unpack High bytes of a
movq mm3, [esi+ebx-8] // Prep c=Prior(x-bpp) bytes
punpcklbw mm2, mm0 // Unpack High bytes of b
psrlq mm3, ShiftRem // shift last 3 bytes to 1st 3 bytes
// pav = p - a = (a + b - c) - a = b - c
movq mm4, mm2
punpcklbw mm3, mm0 // Unpack High bytes of c
// pbv = p - b = (a + b - c) - b = a - c
movq mm5, mm1
psubw mm4, mm3
pxor mm7, mm7
// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
movq mm6, mm4
psubw mm5, mm3
// pa = abs(p-a) = abs(pav)
// pb = abs(p-b) = abs(pbv)
// pc = abs(p-c) = abs(pcv)
pcmpgtw mm0, mm4 // Create mask pav bytes < 0
paddw mm6, mm5
pand mm0, mm4 // Only pav bytes < 0 in mm7
pcmpgtw mm7, mm5 // Create mask pbv bytes < 0
psubw mm4, mm0
pand mm7, mm5 // Only pbv bytes < 0 in mm0
psubw mm4, mm0
psubw mm5, mm7
pxor mm0, mm0
pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
pand mm0, mm6 // Only pav bytes < 0 in mm7
psubw mm5, mm7
psubw mm6, mm0
// test pa <= pb
movq mm7, mm4
psubw mm6, mm0
pcmpgtw mm7, mm5 // pa > pb?
movq mm0, mm7
// use mm7 mask to merge pa & pb
pand mm5, mm7
// use mm0 mask copy to merge a & b
pand mm2, mm0
pandn mm7, mm4
pandn mm0, mm1
paddw mm7, mm5
paddw mm0, mm2
// test ((pa <= pb)? pa:pb) <= pc
pcmpgtw mm7, mm6 // pab > pc?
pxor mm1, mm1
pand mm3, mm7
pandn mm7, mm0
paddw mm7, mm3
pxor mm0, mm0
packuswb mm7, mm1
movq mm3, [esi + ebx] // load c=Prior(x-bpp)
pand mm7, ActiveMask
movq mm2, mm3 // load b=Prior(x) step 1
paddb mm7, [edi + ebx] // add Paeth predictor with Raw(x)
punpcklbw mm3, mm0 // Unpack High bytes of c
movq [edi + ebx], mm7 // write back updated value
movq mm1, mm7 // Now mm1 will be used as Raw(x-bpp)
// Now do Paeth for 2nd set of bytes (3-5)
psrlq mm2, ShiftBpp // load b=Prior(x) step 2
punpcklbw mm1, mm0 // Unpack High bytes of a
pxor mm7, mm7
punpcklbw mm2, mm0 // Unpack High bytes of b
// pbv = p - b = (a + b - c) - b = a - c
movq mm5, mm1
// pav = p - a = (a + b - c) - a = b - c
movq mm4, mm2
psubw mm5, mm3
psubw mm4, mm3
// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) =
// pav + pbv = pbv + pav
movq mm6, mm5
paddw mm6, mm4
// pa = abs(p-a) = abs(pav)
// pb = abs(p-b) = abs(pbv)
// pc = abs(p-c) = abs(pcv)
pcmpgtw mm0, mm5 // Create mask pbv bytes < 0
pcmpgtw mm7, mm4 // Create mask pav bytes < 0
pand mm0, mm5 // Only pbv bytes < 0 in mm0
pand mm7, mm4 // Only pav bytes < 0 in mm7
psubw mm5, mm0
psubw mm4, mm7
psubw mm5, mm0
psubw mm4, mm7
pxor mm0, mm0
pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
pand mm0, mm6 // Only pav bytes < 0 in mm7
psubw mm6, mm0
// test pa <= pb
movq mm7, mm4
psubw mm6, mm0
pcmpgtw mm7, mm5 // pa > pb?
movq mm0, mm7
// use mm7 mask to merge pa & pb
pand mm5, mm7
// use mm0 mask copy to merge a & b
pand mm2, mm0
pandn mm7, mm4
pandn mm0, mm1
paddw mm7, mm5
paddw mm0, mm2
// test ((pa <= pb)? pa:pb) <= pc
pcmpgtw mm7, mm6 // pab > pc?
movq mm2, [esi + ebx] // load b=Prior(x)
pand mm3, mm7
pandn mm7, mm0
pxor mm1, mm1
paddw mm7, mm3
pxor mm0, mm0
packuswb mm7, mm1
movq mm3, mm2 // load c=Prior(x-bpp) step 1
pand mm7, ActiveMask
punpckhbw mm2, mm0 // Unpack High bytes of b
psllq mm7, ShiftBpp // Shift bytes to 2nd group of 3 bytes
// pav = p - a = (a + b - c) - a = b - c
movq mm4, mm2
paddb mm7, [edi + ebx] // add Paeth predictor with Raw(x)
psllq mm3, ShiftBpp // load c=Prior(x-bpp) step 2
movq [edi + ebx], mm7 // write back updated value
movq mm1, mm7
punpckhbw mm3, mm0 // Unpack High bytes of c
psllq mm1, ShiftBpp // Shift bytes
// Now mm1 will be used as Raw(x-bpp)
// Now do Paeth for 3rd, and final, set of bytes (6-7)
pxor mm7, mm7
punpckhbw mm1, mm0 // Unpack High bytes of a
psubw mm4, mm3
// pbv = p - b = (a + b - c) - b = a - c
movq mm5, mm1
// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
movq mm6, mm4
psubw mm5, mm3
pxor mm0, mm0
paddw mm6, mm5
// pa = abs(p-a) = abs(pav)
// pb = abs(p-b) = abs(pbv)
// pc = abs(p-c) = abs(pcv)
pcmpgtw mm0, mm4 // Create mask pav bytes < 0
pcmpgtw mm7, mm5 // Create mask pbv bytes < 0
pand mm0, mm4 // Only pav bytes < 0 in mm7
pand mm7, mm5 // Only pbv bytes < 0 in mm0
psubw mm4, mm0
psubw mm5, mm7
psubw mm4, mm0
psubw mm5, mm7
pxor mm0, mm0
pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
pand mm0, mm6 // Only pav bytes < 0 in mm7
psubw mm6, mm0
// test pa <= pb
movq mm7, mm4
psubw mm6, mm0
pcmpgtw mm7, mm5 // pa > pb?
movq mm0, mm7
// use mm0 mask copy to merge a & b
pand mm2, mm0
// use mm7 mask to merge pa & pb
pand mm5, mm7
pandn mm0, mm1
pandn mm7, mm4
paddw mm0, mm2
paddw mm7, mm5
// test ((pa <= pb)? pa:pb) <= pc
pcmpgtw mm7, mm6 // pab > pc?
pand mm3, mm7
pandn mm7, mm0
paddw mm7, mm3
pxor mm1, mm1
packuswb mm1, mm7
// Step ebx to next set of 8 bytes and repeat loop til done
add ebx, 8
pand mm1, ActiveMaskEnd
paddb mm1, [edi + ebx - 8] // add Paeth predictor with Raw(x)
cmp ebx, MMXLength
pxor mm0, mm0 // pxor does not affect flags
movq [edi + ebx - 8], mm1 // write back updated value
// mm1 will be used as Raw(x-bpp) next loop
// mm3 ready to be used as Prior(x-bpp) next loop
jb dpth3lp
} // end _asm block
}
break;
case 6:
case 7:
case 5:
{
ActiveMask.use = 0x00000000ffffffff;
ActiveMask2.use = 0xffffffff00000000;
ShiftBpp.use = bpp << 3; // == bpp * 8
ShiftRem.use = 64 - ShiftBpp.use;
_asm
{
mov ebx, diff
mov edi, row
mov esi, prev_row
// PRIME the pump (load the first Raw(x-bpp) data set
movq mm1, [edi+ebx-8]
pxor mm0, mm0
dpth6lp:
// Must shift to position Raw(x-bpp) data
psrlq mm1, ShiftRem
// Do first set of 4 bytes
movq mm3, [esi+ebx-8] // read c=Prior(x-bpp) bytes
punpcklbw mm1, mm0 // Unpack Low bytes of a
movq mm2, [esi + ebx] // load b=Prior(x)
punpcklbw mm2, mm0 // Unpack Low bytes of b
// Must shift to position Prior(x-bpp) data
psrlq mm3, ShiftRem
// pav = p - a = (a + b - c) - a = b - c
movq mm4, mm2
punpcklbw mm3, mm0 // Unpack Low bytes of c
// pbv = p - b = (a + b - c) - b = a - c
movq mm5, mm1
psubw mm4, mm3
pxor mm7, mm7
// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
movq mm6, mm4
psubw mm5, mm3
// pa = abs(p-a) = abs(pav)
// pb = abs(p-b) = abs(pbv)
// pc = abs(p-c) = abs(pcv)
pcmpgtw mm0, mm4 // Create mask pav bytes < 0
paddw mm6, mm5
pand mm0, mm4 // Only pav bytes < 0 in mm7
pcmpgtw mm7, mm5 // Create mask pbv bytes < 0
psubw mm4, mm0
pand mm7, mm5 // Only pbv bytes < 0 in mm0
psubw mm4, mm0
psubw mm5, mm7
pxor mm0, mm0
pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
pand mm0, mm6 // Only pav bytes < 0 in mm7
psubw mm5, mm7
psubw mm6, mm0
// test pa <= pb
movq mm7, mm4
psubw mm6, mm0
pcmpgtw mm7, mm5 // pa > pb?
movq mm0, mm7
// use mm7 mask to merge pa & pb
pand mm5, mm7
// use mm0 mask copy to merge a & b
pand mm2, mm0
pandn mm7, mm4
pandn mm0, mm1
paddw mm7, mm5
paddw mm0, mm2
// test ((pa <= pb)? pa:pb) <= pc
pcmpgtw mm7, mm6 // pab > pc?
pxor mm1, mm1
pand mm3, mm7
pandn mm7, mm0
paddw mm7, mm3
pxor mm0, mm0
packuswb mm7, mm1
movq mm3, [esi + ebx - 8] // load c=Prior(x-bpp)
pand mm7, ActiveMask
psrlq mm3, ShiftRem
movq mm2, [esi + ebx] // load b=Prior(x) step 1
paddb mm7, [edi + ebx] // add Paeth predictor with Raw(x)
movq mm6, mm2
movq [edi + ebx], mm7 // write back updated value
movq mm1, [edi+ebx-8]
psllq mm6, ShiftBpp
movq mm5, mm7
psrlq mm1, ShiftRem
por mm3, mm6
psllq mm5, ShiftBpp
punpckhbw mm3, mm0 // Unpack High bytes of c
por mm1, mm5
// Do second set of 4 bytes
punpckhbw mm2, mm0 // Unpack High bytes of b
punpckhbw mm1, mm0 // Unpack High bytes of a
// pav = p - a = (a + b - c) - a = b - c
movq mm4, mm2
// pbv = p - b = (a + b - c) - b = a - c
movq mm5, mm1
psubw mm4, mm3
pxor mm7, mm7
// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
movq mm6, mm4
psubw mm5, mm3
// pa = abs(p-a) = abs(pav)
// pb = abs(p-b) = abs(pbv)
// pc = abs(p-c) = abs(pcv)
pcmpgtw mm0, mm4 // Create mask pav bytes < 0
paddw mm6, mm5
pand mm0, mm4 // Only pav bytes < 0 in mm7
pcmpgtw mm7, mm5 // Create mask pbv bytes < 0
psubw mm4, mm0
pand mm7, mm5 // Only pbv bytes < 0 in mm0
psubw mm4, mm0
psubw mm5, mm7
pxor mm0, mm0
pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
pand mm0, mm6 // Only pav bytes < 0 in mm7
psubw mm5, mm7
psubw mm6, mm0
// test pa <= pb
movq mm7, mm4
psubw mm6, mm0
pcmpgtw mm7, mm5 // pa > pb?
movq mm0, mm7
// use mm7 mask to merge pa & pb
pand mm5, mm7
// use mm0 mask copy to merge a & b
pand mm2, mm0
pandn mm7, mm4
pandn mm0, mm1
paddw mm7, mm5
paddw mm0, mm2
// test ((pa <= pb)? pa:pb) <= pc
pcmpgtw mm7, mm6 // pab > pc?
pxor mm1, mm1
pand mm3, mm7
pandn mm7, mm0
pxor mm1, mm1
paddw mm7, mm3
pxor mm0, mm0
// Step ex to next set of 8 bytes and repeat loop til done
add ebx, 8
packuswb mm1, mm7
paddb mm1, [edi + ebx - 8] // add Paeth predictor with Raw(x)
cmp ebx, MMXLength
movq [edi + ebx - 8], mm1 // write back updated value
// mm1 will be used as Raw(x-bpp) next loop
jb dpth6lp
} // end _asm block
}
break;
case 4:
{
ActiveMask.use = 0x00000000ffffffff;
_asm {
mov ebx, diff
mov edi, row
mov esi, prev_row
pxor mm0, mm0
// PRIME the pump (load the first Raw(x-bpp) data set
movq mm1, [edi+ebx-8] // Only time should need to read
// a=Raw(x-bpp) bytes
dpth4lp:
// Do first set of 4 bytes
movq mm3, [esi+ebx-8] // read c=Prior(x-bpp) bytes
punpckhbw mm1, mm0 // Unpack Low bytes of a
movq mm2, [esi + ebx] // load b=Prior(x)
punpcklbw mm2, mm0 // Unpack High bytes of b
// pav = p - a = (a + b - c) - a = b - c
movq mm4, mm2
punpckhbw mm3, mm0 // Unpack High bytes of c
// pbv = p - b = (a + b - c) - b = a - c
movq mm5, mm1
psubw mm4, mm3
pxor mm7, mm7
// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
movq mm6, mm4
psubw mm5, mm3
// pa = abs(p-a) = abs(pav)
// pb = abs(p-b) = abs(pbv)
// pc = abs(p-c) = abs(pcv)
pcmpgtw mm0, mm4 // Create mask pav bytes < 0
paddw mm6, mm5
pand mm0, mm4 // Only pav bytes < 0 in mm7
pcmpgtw mm7, mm5 // Create mask pbv bytes < 0
psubw mm4, mm0
pand mm7, mm5 // Only pbv bytes < 0 in mm0
psubw mm4, mm0
psubw mm5, mm7
pxor mm0, mm0
pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
pand mm0, mm6 // Only pav bytes < 0 in mm7
psubw mm5, mm7
psubw mm6, mm0
// test pa <= pb
movq mm7, mm4
psubw mm6, mm0
pcmpgtw mm7, mm5 // pa > pb?
movq mm0, mm7
// use mm7 mask to merge pa & pb
pand mm5, mm7
// use mm0 mask copy to merge a & b
pand mm2, mm0
pandn mm7, mm4
pandn mm0, mm1
paddw mm7, mm5
paddw mm0, mm2
// test ((pa <= pb)? pa:pb) <= pc
pcmpgtw mm7, mm6 // pab > pc?
pxor mm1, mm1
pand mm3, mm7
pandn mm7, mm0
paddw mm7, mm3
pxor mm0, mm0
packuswb mm7, mm1
movq mm3, [esi + ebx] // load c=Prior(x-bpp)
pand mm7, ActiveMask
movq mm2, mm3 // load b=Prior(x) step 1
paddb mm7, [edi + ebx] // add Paeth predictor with Raw(x)
punpcklbw mm3, mm0 // Unpack High bytes of c
movq [edi + ebx], mm7 // write back updated value
movq mm1, mm7 // Now mm1 will be used as Raw(x-bpp)
// Do second set of 4 bytes
punpckhbw mm2, mm0 // Unpack Low bytes of b
punpcklbw mm1, mm0 // Unpack Low bytes of a
// pav = p - a = (a + b - c) - a = b - c
movq mm4, mm2
// pbv = p - b = (a + b - c) - b = a - c
movq mm5, mm1
psubw mm4, mm3
pxor mm7, mm7
// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
movq mm6, mm4
psubw mm5, mm3
// pa = abs(p-a) = abs(pav)
// pb = abs(p-b) = abs(pbv)
// pc = abs(p-c) = abs(pcv)
pcmpgtw mm0, mm4 // Create mask pav bytes < 0
paddw mm6, mm5
pand mm0, mm4 // Only pav bytes < 0 in mm7
pcmpgtw mm7, mm5 // Create mask pbv bytes < 0
psubw mm4, mm0
pand mm7, mm5 // Only pbv bytes < 0 in mm0
psubw mm4, mm0
psubw mm5, mm7
pxor mm0, mm0
pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
pand mm0, mm6 // Only pav bytes < 0 in mm7
psubw mm5, mm7
psubw mm6, mm0
// test pa <= pb
movq mm7, mm4
psubw mm6, mm0
pcmpgtw mm7, mm5 // pa > pb?
movq mm0, mm7
// use mm7 mask to merge pa & pb
pand mm5, mm7
// use mm0 mask copy to merge a & b
pand mm2, mm0
pandn mm7, mm4
pandn mm0, mm1
paddw mm7, mm5
paddw mm0, mm2
// test ((pa <= pb)? pa:pb) <= pc
pcmpgtw mm7, mm6 // pab > pc?
pxor mm1, mm1
pand mm3, mm7
pandn mm7, mm0
pxor mm1, mm1
paddw mm7, mm3
pxor mm0, mm0
// Step ex to next set of 8 bytes and repeat loop til done
add ebx, 8
packuswb mm1, mm7
paddb mm1, [edi + ebx - 8] // add Paeth predictor with Raw(x)
cmp ebx, MMXLength
movq [edi + ebx - 8], mm1 // write back updated value
// mm1 will be used as Raw(x-bpp) next loop
jb dpth4lp
} // end _asm block
}
break;
case 8: // bpp == 8
{
ActiveMask.use = 0x00000000ffffffff;
_asm {
mov ebx, diff
mov edi, row
mov esi, prev_row
pxor mm0, mm0
// PRIME the pump (load the first Raw(x-bpp) data set
movq mm1, [edi+ebx-8] // Only time should need to read
// a=Raw(x-bpp) bytes
dpth8lp:
// Do first set of 4 bytes
movq mm3, [esi+ebx-8] // read c=Prior(x-bpp) bytes
punpcklbw mm1, mm0 // Unpack Low bytes of a
movq mm2, [esi + ebx] // load b=Prior(x)
punpcklbw mm2, mm0 // Unpack Low bytes of b
// pav = p - a = (a + b - c) - a = b - c
movq mm4, mm2
punpcklbw mm3, mm0 // Unpack Low bytes of c
// pbv = p - b = (a + b - c) - b = a - c
movq mm5, mm1
psubw mm4, mm3
pxor mm7, mm7
// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
movq mm6, mm4
psubw mm5, mm3
// pa = abs(p-a) = abs(pav)
// pb = abs(p-b) = abs(pbv)
// pc = abs(p-c) = abs(pcv)
pcmpgtw mm0, mm4 // Create mask pav bytes < 0
paddw mm6, mm5
pand mm0, mm4 // Only pav bytes < 0 in mm7
pcmpgtw mm7, mm5 // Create mask pbv bytes < 0
psubw mm4, mm0
pand mm7, mm5 // Only pbv bytes < 0 in mm0
psubw mm4, mm0
psubw mm5, mm7
pxor mm0, mm0
pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
pand mm0, mm6 // Only pav bytes < 0 in mm7
psubw mm5, mm7
psubw mm6, mm0
// test pa <= pb
movq mm7, mm4
psubw mm6, mm0
pcmpgtw mm7, mm5 // pa > pb?
movq mm0, mm7
// use mm7 mask to merge pa & pb
pand mm5, mm7
// use mm0 mask copy to merge a & b
pand mm2, mm0
pandn mm7, mm4
pandn mm0, mm1
paddw mm7, mm5
paddw mm0, mm2
// test ((pa <= pb)? pa:pb) <= pc
pcmpgtw mm7, mm6 // pab > pc?
pxor mm1, mm1
pand mm3, mm7
pandn mm7, mm0
paddw mm7, mm3
pxor mm0, mm0
packuswb mm7, mm1
movq mm3, [esi+ebx-8] // read c=Prior(x-bpp) bytes
pand mm7, ActiveMask
movq mm2, [esi + ebx] // load b=Prior(x)
paddb mm7, [edi + ebx] // add Paeth predictor with Raw(x)
punpckhbw mm3, mm0 // Unpack High bytes of c
movq [edi + ebx], mm7 // write back updated value
movq mm1, [edi+ebx-8] // read a=Raw(x-bpp) bytes
// Do second set of 4 bytes
punpckhbw mm2, mm0 // Unpack High bytes of b
punpckhbw mm1, mm0 // Unpack High bytes of a
// pav = p - a = (a + b - c) - a = b - c
movq mm4, mm2
// pbv = p - b = (a + b - c) - b = a - c
movq mm5, mm1
psubw mm4, mm3
pxor mm7, mm7
// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
movq mm6, mm4
psubw mm5, mm3
// pa = abs(p-a) = abs(pav)
// pb = abs(p-b) = abs(pbv)
// pc = abs(p-c) = abs(pcv)
pcmpgtw mm0, mm4 // Create mask pav bytes < 0
paddw mm6, mm5
pand mm0, mm4 // Only pav bytes < 0 in mm7
pcmpgtw mm7, mm5 // Create mask pbv bytes < 0
psubw mm4, mm0
pand mm7, mm5 // Only pbv bytes < 0 in mm0
psubw mm4, mm0
psubw mm5, mm7
pxor mm0, mm0
pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
pand mm0, mm6 // Only pav bytes < 0 in mm7
psubw mm5, mm7
psubw mm6, mm0
// test pa <= pb
movq mm7, mm4
psubw mm6, mm0
pcmpgtw mm7, mm5 // pa > pb?
movq mm0, mm7
// use mm7 mask to merge pa & pb
pand mm5, mm7
// use mm0 mask copy to merge a & b
pand mm2, mm0
pandn mm7, mm4
pandn mm0, mm1
paddw mm7, mm5
paddw mm0, mm2
// test ((pa <= pb)? pa:pb) <= pc
pcmpgtw mm7, mm6 // pab > pc?
pxor mm1, mm1
pand mm3, mm7
pandn mm7, mm0
pxor mm1, mm1
paddw mm7, mm3
pxor mm0, mm0
// Step ex to next set of 8 bytes and repeat loop til done
add ebx, 8
packuswb mm1, mm7
paddb mm1, [edi + ebx - 8] // add Paeth predictor with Raw(x)
cmp ebx, MMXLength
movq [edi + ebx - 8], mm1 // write back updated value
// mm1 will be used as Raw(x-bpp) next loop
jb dpth8lp
} // end _asm block
}
break;
case 1: // bpp = 1
case 2: // bpp = 2
default: // bpp > 8
{
_asm {
mov ebx, diff
cmp ebx, FullLength
jnb dpthdend
mov edi, row
mov esi, prev_row
// Do Paeth decode for remaining bytes
mov edx, ebx
xor ecx, ecx // zero ecx before using cl & cx in loop below
sub edx, bpp // Set edx = ebx - bpp
dpthdlp:
xor eax, eax
// pav = p - a = (a + b - c) - a = b - c
mov al, [esi + ebx] // load Prior(x) into al
mov cl, [esi + edx] // load Prior(x-bpp) into cl
sub eax, ecx // subtract Prior(x-bpp)
mov patemp, eax // Save pav for later use
xor eax, eax
// pbv = p - b = (a + b - c) - b = a - c
mov al, [edi + edx] // load Raw(x-bpp) into al
sub eax, ecx // subtract Prior(x-bpp)
mov ecx, eax
// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
add eax, patemp // pcv = pav + pbv
// pc = abs(pcv)
test eax, 0x80000000
jz dpthdpca
neg eax // reverse sign of neg values
dpthdpca:
mov pctemp, eax // save pc for later use
// pb = abs(pbv)
test ecx, 0x80000000
jz dpthdpba
neg ecx // reverse sign of neg values
dpthdpba:
mov pbtemp, ecx // save pb for later use
// pa = abs(pav)
mov eax, patemp
test eax, 0x80000000
jz dpthdpaa
neg eax // reverse sign of neg values
dpthdpaa:
mov patemp, eax // save pa for later use
// test if pa <= pb
cmp eax, ecx
jna dpthdabb
// pa > pb; now test if pb <= pc
cmp ecx, pctemp
jna dpthdbbc
// pb > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
mov cl, [esi + edx] // load Prior(x-bpp) into cl
jmp dpthdpaeth
dpthdbbc:
// pb <= pc; Raw(x) = Paeth(x) + Prior(x)
mov cl, [esi + ebx] // load Prior(x) into cl
jmp dpthdpaeth
dpthdabb:
// pa <= pb; now test if pa <= pc
cmp eax, pctemp
jna dpthdabc
// pa > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
mov cl, [esi + edx] // load Prior(x-bpp) into cl
jmp dpthdpaeth
dpthdabc:
// pa <= pc; Raw(x) = Paeth(x) + Raw(x-bpp)
mov cl, [edi + edx] // load Raw(x-bpp) into cl
dpthdpaeth:
inc ebx
inc edx
// Raw(x) = (Paeth(x) + Paeth_Predictor( a, b, c )) mod 256
add [edi + ebx - 1], cl
cmp ebx, FullLength
jb dpthdlp
dpthdend:
} // end _asm block
}
return; // No need to go further with this one
} // end switch ( bpp )
_asm
{
// MMX acceleration complete now do clean-up
// Check if any remaining bytes left to decode
mov ebx, MMXLength
cmp ebx, FullLength
jnb dpthend
mov edi, row
mov esi, prev_row
// Do Paeth decode for remaining bytes
mov edx, ebx
xor ecx, ecx // zero ecx before using cl & cx in loop below
sub edx, bpp // Set edx = ebx - bpp
dpthlp2:
xor eax, eax
// pav = p - a = (a + b - c) - a = b - c
mov al, [esi + ebx] // load Prior(x) into al
mov cl, [esi + edx] // load Prior(x-bpp) into cl
sub eax, ecx // subtract Prior(x-bpp)
mov patemp, eax // Save pav for later use
xor eax, eax
// pbv = p - b = (a + b - c) - b = a - c
mov al, [edi + edx] // load Raw(x-bpp) into al
sub eax, ecx // subtract Prior(x-bpp)
mov ecx, eax
// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
add eax, patemp // pcv = pav + pbv
// pc = abs(pcv)
test eax, 0x80000000
jz dpthpca2
neg eax // reverse sign of neg values
dpthpca2:
mov pctemp, eax // save pc for later use
// pb = abs(pbv)
test ecx, 0x80000000
jz dpthpba2
neg ecx // reverse sign of neg values
dpthpba2:
mov pbtemp, ecx // save pb for later use
// pa = abs(pav)
mov eax, patemp
test eax, 0x80000000
jz dpthpaa2
neg eax // reverse sign of neg values
dpthpaa2:
mov patemp, eax // save pa for later use
// test if pa <= pb
cmp eax, ecx
jna dpthabb2
// pa > pb; now test if pb <= pc
cmp ecx, pctemp
jna dpthbbc2
// pb > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
mov cl, [esi + edx] // load Prior(x-bpp) into cl
jmp dpthpaeth2
dpthbbc2:
// pb <= pc; Raw(x) = Paeth(x) + Prior(x)
mov cl, [esi + ebx] // load Prior(x) into cl
jmp dpthpaeth2
dpthabb2:
// pa <= pb; now test if pa <= pc
cmp eax, pctemp
jna dpthabc2
// pa > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
mov cl, [esi + edx] // load Prior(x-bpp) into cl
jmp dpthpaeth2
dpthabc2:
// pa <= pc; Raw(x) = Paeth(x) + Raw(x-bpp)
mov cl, [edi + edx] // load Raw(x-bpp) into cl
dpthpaeth2:
inc ebx
inc edx
// Raw(x) = (Paeth(x) + Paeth_Predictor( a, b, c )) mod 256
add [edi + ebx - 1], cl
cmp ebx, FullLength
jb dpthlp2
dpthend:
emms // End MMX instructions; prep for possible FP instrs.
} // end _asm block
}
// Optimized code for PNG Sub filter decoder
void /* PRIVATE */
png_read_filter_row_mmx_sub(png_row_infop row_info, png_bytep row)
{
//int test;
int bpp;
png_uint_32 FullLength;
png_uint_32 MMXLength;
int diff;
bpp = (row_info->pixel_depth + 7) >> 3; // Get # bytes per pixel
FullLength = row_info->rowbytes - bpp; // # of bytes to filter
_asm {
mov edi, row
mov esi, edi // lp = row
add edi, bpp // rp = row + bpp
xor eax, eax
// get # of bytes to alignment
mov diff, edi // take start of row
add diff, 0xf // add 7 + 8 to incr past
// alignment boundary
xor ebx, ebx
and diff, 0xfffffff8 // mask to alignment boundary
sub diff, edi // subtract from start ==> value
// ebx at alignment
jz dsubgo
// fix alignment
dsublp1:
mov al, [esi+ebx]
add [edi+ebx], al
inc ebx
cmp ebx, diff
jb dsublp1
dsubgo:
mov ecx, FullLength
mov edx, ecx
sub edx, ebx // subtract alignment fix
and edx, 0x00000007 // calc bytes over mult of 8
sub ecx, edx // drop over bytes from length
mov MMXLength, ecx
} // end _asm block
// Now do the math for the rest of the row
switch ( bpp )
{
case 3:
{
ActiveMask.use = 0x0000ffffff000000;
ShiftBpp.use = 24; // == 3 * 8
ShiftRem.use = 40; // == 64 - 24
_asm {
mov edi, row
movq mm7, ActiveMask // Load ActiveMask for 2nd active byte group
mov esi, edi // lp = row
add edi, bpp // rp = row + bpp
movq mm6, mm7
mov ebx, diff
psllq mm6, ShiftBpp // Move mask in mm6 to cover 3rd active
// byte group
// PRIME the pump (load the first Raw(x-bpp) data set
movq mm1, [edi+ebx-8]
dsub3lp:
psrlq mm1, ShiftRem // Shift data for adding 1st bpp bytes
// no need for mask; shift clears inactive bytes
// Add 1st active group
movq mm0, [edi+ebx]
paddb mm0, mm1
// Add 2nd active group
movq mm1, mm0 // mov updated Raws to mm1
psllq mm1, ShiftBpp // shift data to position correctly
pand mm1, mm7 // mask to use only 2nd active group
paddb mm0, mm1
// Add 3rd active group
movq mm1, mm0 // mov updated Raws to mm1
psllq mm1, ShiftBpp // shift data to position correctly
pand mm1, mm6 // mask to use only 3rd active group
add ebx, 8
paddb mm0, mm1
cmp ebx, MMXLength
movq [edi+ebx-8], mm0 // Write updated Raws back to array
// Prep for doing 1st add at top of loop
movq mm1, mm0
jb dsub3lp
} // end _asm block
}
break;
case 1:
{
// Placed here just in case this is a duplicate of the
// non-MMX code for the SUB filter in png_read_filter_row below
//
// png_bytep rp;
// png_bytep lp;
// png_uint_32 i;
// bpp = (row_info->pixel_depth + 7) >> 3;
// for (i = (png_uint_32)bpp, rp = row + bpp, lp = row;
// i < row_info->rowbytes; i++, rp++, lp++)
// {
// *rp = (png_byte)(((int)(*rp) + (int)(*lp)) & 0xff);
// }
_asm {
mov ebx, diff
mov edi, row
cmp ebx, FullLength
jnb dsub1end
mov esi, edi // lp = row
xor eax, eax
add edi, bpp // rp = row + bpp
dsub1lp:
mov al, [esi+ebx]
add [edi+ebx], al
inc ebx
cmp ebx, FullLength
jb dsub1lp
dsub1end:
} // end _asm block
}
return;
case 6:
case 7:
case 4:
case 5:
{
ShiftBpp.use = bpp << 3;
ShiftRem.use = 64 - ShiftBpp.use;
_asm {
mov edi, row
mov ebx, diff
mov esi, edi // lp = row
add edi, bpp // rp = row + bpp
// PRIME the pump (load the first Raw(x-bpp) data set
movq mm1, [edi+ebx-8]
dsub4lp:
psrlq mm1, ShiftRem // Shift data for adding 1st bpp bytes
// no need for mask; shift clears inactive bytes
movq mm0, [edi+ebx]
paddb mm0, mm1
// Add 2nd active group
movq mm1, mm0 // mov updated Raws to mm1
psllq mm1, ShiftBpp // shift data to position correctly
// there is no need for any mask
// since shift clears inactive bits/bytes
add ebx, 8
paddb mm0, mm1
cmp ebx, MMXLength
movq [edi+ebx-8], mm0
movq mm1, mm0 // Prep for doing 1st add at top of loop
jb dsub4lp
} // end _asm block
}
break;
case 2:
{
ActiveMask.use = 0x00000000ffff0000;
ShiftBpp.use = 16; // == 2 * 8
ShiftRem.use = 48; // == 64 - 16
_asm {
movq mm7, ActiveMask // Load ActiveMask for 2nd active byte group
mov ebx, diff
movq mm6, mm7
mov edi, row
psllq mm6, ShiftBpp // Move mask in mm6 to cover 3rd active
mov esi, edi // lp = row
movq mm5, mm6
add edi, bpp // rp = row + bpp
psllq mm5, ShiftBpp // Move mask in mm5 to cover 4th active
// byte group
// PRIME the pump (load the first Raw(x-bpp) data set
movq mm1, [edi+ebx-8]
dsub2lp:
// Add 1st active group
psrlq mm1, ShiftRem // Shift data for adding 1st bpp bytes
// no need for mask; shift clears inactive
// bytes
movq mm0, [edi+ebx]
paddb mm0, mm1
// Add 2nd active group
movq mm1, mm0 // mov updated Raws to mm1
psllq mm1, ShiftBpp // shift data to position correctly
pand mm1, mm7 // mask to use only 2nd active group
paddb mm0, mm1
// Add 3rd active group
movq mm1, mm0 // mov updated Raws to mm1
psllq mm1, ShiftBpp // shift data to position correctly
pand mm1, mm6 // mask to use only 3rd active group
paddb mm0, mm1
// Add 4th active group
movq mm1, mm0 // mov updated Raws to mm1
psllq mm1, ShiftBpp // shift data to position correctly
pand mm1, mm5 // mask to use only 4th active group
add ebx, 8
paddb mm0, mm1
cmp ebx, MMXLength
movq [edi+ebx-8], mm0 // Write updated Raws back to array
movq mm1, mm0 // Prep for doing 1st add at top of loop
jb dsub2lp
} // end _asm block
}
break;
case 8:
{
_asm {
mov edi, row
mov ebx, diff
mov esi, edi // lp = row
add edi, bpp // rp = row + bpp
mov ecx, MMXLength
movq mm7, [edi+ebx-8] // PRIME the pump (load the first
// Raw(x-bpp) data set
and ecx, 0x0000003f // calc bytes over mult of 64
dsub8lp:
movq mm0, [edi+ebx] // Load Sub(x) for 1st 8 bytes
paddb mm0, mm7
movq mm1, [edi+ebx+8] // Load Sub(x) for 2nd 8 bytes
movq [edi+ebx], mm0 // Write Raw(x) for 1st 8 bytes
// Now mm0 will be used as Raw(x-bpp) for
// the 2nd group of 8 bytes. This will be
// repeated for each group of 8 bytes with
// the 8th group being used as the Raw(x-bpp)
// for the 1st group of the next loop.
paddb mm1, mm0
movq mm2, [edi+ebx+16] // Load Sub(x) for 3rd 8 bytes
movq [edi+ebx+8], mm1 // Write Raw(x) for 2nd 8 bytes
paddb mm2, mm1
movq mm3, [edi+ebx+24] // Load Sub(x) for 4th 8 bytes
movq [edi+ebx+16], mm2 // Write Raw(x) for 3rd 8 bytes
paddb mm3, mm2
movq mm4, [edi+ebx+32] // Load Sub(x) for 5th 8 bytes
movq [edi+ebx+24], mm3 // Write Raw(x) for 4th 8 bytes
paddb mm4, mm3
movq mm5, [edi+ebx+40] // Load Sub(x) for 6th 8 bytes
movq [edi+ebx+32], mm4 // Write Raw(x) for 5th 8 bytes
paddb mm5, mm4
movq mm6, [edi+ebx+48] // Load Sub(x) for 7th 8 bytes
movq [edi+ebx+40], mm5 // Write Raw(x) for 6th 8 bytes
paddb mm6, mm5
movq mm7, [edi+ebx+56] // Load Sub(x) for 8th 8 bytes
movq [edi+ebx+48], mm6 // Write Raw(x) for 7th 8 bytes
add ebx, 64
paddb mm7, mm6
cmp ebx, ecx
movq [edi+ebx-8], mm7 // Write Raw(x) for 8th 8 bytes
jb dsub8lp
cmp ebx, MMXLength
jnb dsub8lt8
dsub8lpA:
movq mm0, [edi+ebx]
add ebx, 8
paddb mm0, mm7
cmp ebx, MMXLength
movq [edi+ebx-8], mm0 // use -8 to offset early add to ebx
movq mm7, mm0 // Move calculated Raw(x) data to mm1 to
// be the new Raw(x-bpp) for the next loop
jb dsub8lpA
dsub8lt8:
} // end _asm block
}
break;
default: // bpp greater than 8 bytes
{
_asm {
mov ebx, diff
mov edi, row
mov esi, edi // lp = row
add edi, bpp // rp = row + bpp
dsubAlp:
movq mm0, [edi+ebx]
movq mm1, [esi+ebx]
add ebx, 8
paddb mm0, mm1
cmp ebx, MMXLength
movq [edi+ebx-8], mm0 // mov does not affect flags; -8 to offset
// add ebx
jb dsubAlp
} // end _asm block
}
break;
} // end switch ( bpp )
_asm {
mov ebx, MMXLength
mov edi, row
cmp ebx, FullLength
jnb dsubend
mov esi, edi // lp = row
xor eax, eax
add edi, bpp // rp = row + bpp
dsublp2:
mov al, [esi+ebx]
add [edi+ebx], al
inc ebx
cmp ebx, FullLength
jb dsublp2
dsubend:
emms // End MMX instructions; prep for possible FP instrs.
} // end _asm block
}
// Optimized code for PNG Up filter decoder
void /* PRIVATE */
png_read_filter_row_mmx_up(png_row_infop row_info, png_bytep row,
png_bytep prev_row)
{
png_uint_32 len;
len = row_info->rowbytes; // # of bytes to filter
_asm {
mov edi, row
// get # of bytes to alignment
mov ecx, edi
xor ebx, ebx
add ecx, 0x7
xor eax, eax
and ecx, 0xfffffff8
mov esi, prev_row
sub ecx, edi
jz dupgo
// fix alignment
duplp1:
mov al, [edi+ebx]
add al, [esi+ebx]
inc ebx
cmp ebx, ecx
mov [edi + ebx-1], al // mov does not affect flags; -1 to offset inc ebx
jb duplp1
dupgo:
mov ecx, len
mov edx, ecx
sub edx, ebx // subtract alignment fix
and edx, 0x0000003f // calc bytes over mult of 64
sub ecx, edx // drop over bytes from length
// Unrolled loop - use all MMX registers and interleave to reduce
// number of branch instructions (loops) and reduce partial stalls
duploop:
movq mm1, [esi+ebx]
movq mm0, [edi+ebx]
movq mm3, [esi+ebx+8]
paddb mm0, mm1
movq mm2, [edi+ebx+8]
movq [edi+ebx], mm0
paddb mm2, mm3
movq mm5, [esi+ebx+16]
movq [edi+ebx+8], mm2
movq mm4, [edi+ebx+16]
movq mm7, [esi+ebx+24]
paddb mm4, mm5
movq mm6, [edi+ebx+24]
movq [edi+ebx+16], mm4
paddb mm6, mm7
movq mm1, [esi+ebx+32]
movq [edi+ebx+24], mm6
movq mm0, [edi+ebx+32]
movq mm3, [esi+ebx+40]
paddb mm0, mm1
movq mm2, [edi+ebx+40]
movq [edi+ebx+32], mm0
paddb mm2, mm3
movq mm5, [esi+ebx+48]
movq [edi+ebx+40], mm2
movq mm4, [edi+ebx+48]
movq mm7, [esi+ebx+56]
paddb mm4, mm5
movq mm6, [edi+ebx+56]
movq [edi+ebx+48], mm4
add ebx, 64
paddb mm6, mm7
cmp ebx, ecx
movq [edi+ebx-8], mm6 // (+56)movq does not affect flags;
// -8 to offset add ebx
jb duploop
cmp edx, 0 // Test for bytes over mult of 64
jz dupend
// 2 lines added by lcreeve at netins.net
// (mail 11 Jul 98 in png-implement list)
cmp edx, 8 //test for less than 8 bytes
jb duplt8
add ecx, edx
and edx, 0x00000007 // calc bytes over mult of 8
sub ecx, edx // drop over bytes from length
jz duplt8
// Loop using MMX registers mm0 & mm1 to update 8 bytes simultaneously
duplpA:
movq mm1, [esi+ebx]
movq mm0, [edi+ebx]
add ebx, 8
paddb mm0, mm1
cmp ebx, ecx
movq [edi+ebx-8], mm0 // movq does not affect flags; -8 to offset add ebx
jb duplpA
cmp edx, 0 // Test for bytes over mult of 8
jz dupend
duplt8:
xor eax, eax
add ecx, edx // move over byte count into counter
// Loop using x86 registers to update remaining bytes
duplp2:
mov al, [edi + ebx]
add al, [esi + ebx]
inc ebx
cmp ebx, ecx
mov [edi + ebx-1], al // mov does not affect flags; -1 to offset inc ebx
jb duplp2
dupend:
// Conversion of filtered row completed
emms // End MMX instructions; prep for possible FP instrs.
} // end _asm block
}
// Optimized png_read_filter_row routines
void /* PRIVATE */
png_read_filter_row(png_structp png_ptr, png_row_infop row_info, png_bytep
row, png_bytep prev_row, int filter)
{
#ifdef PNG_DEBUG
char filnm[10];
#endif
if (mmx_supported == 2) {
#if !defined(PNG_1_0_X)
/* this should have happened in png_init_mmx_flags() already */
png_warning(png_ptr, "asm_flags may not have been initialized");
#endif
png_mmx_support();
}
#ifdef PNG_DEBUG
png_debug(1, "in png_read_filter_rown");
switch (filter)
{
case 0: sprintf(filnm, "none");
break;
#if !defined(PNG_1_0_X)
case 1: sprintf(filnm, "sub-%s",
(png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_SUB)? "MMX" : "x86");
break;
case 2: sprintf(filnm, "up-%s",
(png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_UP)? "MMX" : "x86");
break;
case 3: sprintf(filnm, "avg-%s",
(png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_AVG)? "MMX" : "x86");
break;
case 4: sprintf(filnm, "Paeth-%s",
(png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_PAETH)? "MMX":"x86");
break;
#else
case 1: sprintf(filnm, "sub");
break;
case 2: sprintf(filnm, "up");
break;
case 3: sprintf(filnm, "avg");
break;
case 4: sprintf(filnm, "Paeth");
break;
#endif
default: sprintf(filnm, "unknw");
break;
}
png_debug2(0,"row=%5d, %s, ", png_ptr->row_number, filnm);
png_debug2(0, "pd=%2d, b=%d, ", (int)row_info->pixel_depth,
(int)((row_info->pixel_depth + 7) >> 3));
png_debug1(0,"len=%8d, ", row_info->rowbytes);
#endif /* PNG_DEBUG */
switch (filter)
{
case PNG_FILTER_VALUE_NONE:
break;
case PNG_FILTER_VALUE_SUB:
{
#if !defined(PNG_1_0_X)
if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_SUB) &&
(row_info->pixel_depth >= png_ptr->mmx_bitdepth_threshold) &&
(row_info->rowbytes >= png_ptr->mmx_rowbytes_threshold))
#else
if (mmx_supported)
#endif
{
png_read_filter_row_mmx_sub(row_info, row);
}
else
{
png_uint_32 i;
png_uint_32 istop = row_info->rowbytes;
png_uint_32 bpp = (row_info->pixel_depth + 7) >> 3;
png_bytep rp = row + bpp;
png_bytep lp = row;
for (i = bpp; i < istop; i++)
{
*rp = (png_byte)(((int)(*rp) + (int)(*lp++)) & 0xff);
rp++;
}
}
break;
}
case PNG_FILTER_VALUE_UP:
{
#if !defined(PNG_1_0_X)
if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_UP) &&
(row_info->pixel_depth >= png_ptr->mmx_bitdepth_threshold) &&
(row_info->rowbytes >= png_ptr->mmx_rowbytes_threshold))
#else
if (mmx_supported)
#endif
{
png_read_filter_row_mmx_up(row_info, row, prev_row);
}
else
{
png_uint_32 i;
png_uint_32 istop = row_info->rowbytes;
png_bytep rp = row;
png_bytep pp = prev_row;
for (i = 0; i < istop; ++i)
{
*rp = (png_byte)(((int)(*rp) + (int)(*pp++)) & 0xff);
rp++;
}
}
break;
}
case PNG_FILTER_VALUE_AVG:
{
#if !defined(PNG_1_0_X)
if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_AVG) &&
(row_info->pixel_depth >= png_ptr->mmx_bitdepth_threshold) &&
(row_info->rowbytes >= png_ptr->mmx_rowbytes_threshold))
#else
if (mmx_supported)
#endif
{
png_read_filter_row_mmx_avg(row_info, row, prev_row);
}
else
{
png_uint_32 i;
png_bytep rp = row;
png_bytep pp = prev_row;
png_bytep lp = row;
png_uint_32 bpp = (row_info->pixel_depth + 7) >> 3;
png_uint_32 istop = row_info->rowbytes - bpp;
for (i = 0; i < bpp; i++)
{
*rp = (png_byte)(((int)(*rp) +
((int)(*pp++) >> 1)) & 0xff);
rp++;
}
for (i = 0; i < istop; i++)
{
*rp = (png_byte)(((int)(*rp) +
((int)(*pp++ + *lp++) >> 1)) & 0xff);
rp++;
}
}
break;
}
case PNG_FILTER_VALUE_PAETH:
{
#if !defined(PNG_1_0_X)
if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_PAETH) &&
(row_info->pixel_depth >= png_ptr->mmx_bitdepth_threshold) &&
(row_info->rowbytes >= png_ptr->mmx_rowbytes_threshold))
#else
if (mmx_supported)
#endif
{
png_read_filter_row_mmx_paeth(row_info, row, prev_row);
}
else
{
png_uint_32 i;
png_bytep rp = row;
png_bytep pp = prev_row;
png_bytep lp = row;
png_bytep cp = prev_row;
png_uint_32 bpp = (row_info->pixel_depth + 7) >> 3;
png_uint_32 istop=row_info->rowbytes - bpp;
for (i = 0; i < bpp; i++)
{
*rp = (png_byte)(((int)(*rp) + (int)(*pp++)) & 0xff);
rp++;
}
for (i = 0; i < istop; i++) // use leftover rp,pp
{
int a, b, c, pa, pb, pc, p;
a = *lp++;
b = *pp++;
c = *cp++;
p = b - c;
pc = a - c;
#ifdef PNG_USE_ABS
pa = abs(p);
pb = abs(pc);
pc = abs(p + pc);
#else
pa = p < 0 ? -p : p;
pb = pc < 0 ? -pc : pc;
pc = (p + pc) < 0 ? -(p + pc) : p + pc;
#endif
/*
if (pa <= pb && pa <= pc)
p = a;
else if (pb <= pc)
p = b;
else
p = c;
*/
p = (pa <= pb && pa <=pc) ? a : (pb <= pc) ? b : c;
*rp = (png_byte)(((int)(*rp) + p) & 0xff);
rp++;
}
}
break;
}
default:
png_warning(png_ptr, "Ignoring bad row filter type");
*row=0;
break;
}
}
#endif /* PNG_ASSEMBLER_CODE_SUPPORTED && PNG_USE_PNGVCRD */