jprdonnelly/entii-for-workcubes
1
0
Fork 0
mirror of https://github.com/Wack0/entii-for-workcubes.git synced 2025-12-19 18:06:24 -05:00
Code Issues Packages Projects Releases Wiki Activity

gdi: only skip single pixel or line on bounds checks fail, add slight optimisation for copying two consecutive pixels

Browse Source 
fixes #5
This commit is contained in:
Rairii
2025-02-28 12:07:27 +00:00
parent cca96f7f50
commit 3fca59ce16
3 changed files with 120 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
README.md
View File

@@ -111,6 +111,8 @@ You need devkitPPC. Additionally, a `libgcc.a` compiled for `powerpcle` must be
You need [peppc](https://github.com/Wack0/peppc). Additionally, the powerpc libs from the [NT4 DDK](https://archive.org/details/94396011997WindowsNTDDKForWinNT4.0WorkstationUS.iso.7z) (`ddk/lib/ppc/free/*.lib`) must be present in `lib`. The rest of the toolchain (VC6 PPC CE cross compiler used for the C preprocessor for asm, as multi-line defines are handled improperly by gcc cpp; assembler PASM.EXE with single branch patched to skip "dump statements"; resource compiler and linker from MSVC 4.2, and its dependencies; `SPLITSYM.EXE` from NT 3.51 DDK to split COFF debug symbols from executables) is present in `msvc-ppc` directory.
To build the NT 3.5x GDI driver `fpgx35dll` you also need the powerpc `winsrv.lib` from NT 3.51 DDK.
The headers are included and come from various places with slight modifications for working with this toolchain, or for backwards compatibility reasons:
* `nt4/sdk` - NT4 SDK
* `nt4/ddk` - NT4 DDK (including all the headers from the `src/*/inc` directories)

94
fpgx35dll/source/draw.c
View File

@@ -39,6 +39,15 @@ static void TexWriteRgb(PUCHAR pTex, ULONG offset, ULONG rgb) {
}
}
static void TexWriteRgb2Aligned(PUCHAR pTex, ULONG offset, ULONG rgb0, ULONG rgb1) {
UCHAR r0 = (rgb0 >> 16) & 0xFF;
UCHAR r1 = (rgb1 >> 16) & 0xFF;
ULONG value = 0xFF00FF00 | ((rgb1 >> 16) & 0xFF) | (((rgb0 >> 16) & 0xFF) << 16);
ULONG value2 = ((rgb0 & 0xFFFF) << 16) | (rgb1 & 0xFFFF);
NativeWriteBase32(pTex, offset, value);
NativeWriteBase32(pTex, offset + 0x20, value2);
}
BOOL CopyBitsSwap32(
SURFOBJ *psoDest,
SURFOBJ *psoSrc,
@@ -101,41 +110,56 @@ BOOL copyFromFb)
PULONG pulSrcTemp = pulSrc;
PULONG pulDstTemp = pulDst;
// Bounds check the pointers, we could be in here when drawing off the screen
if (pulSrc < pstartSrc) break;
if (pulDst < pstartDst) break;
ULONG TexOffset = 0;
ULONG cxTemp = cx;
ULONG cxIdx = 0;
if (copyFromFb) {
// Read swapped from source, write normally to dest.
while (cxTemp--) {
if (pulSrcTemp >= pendSrc) break;
if (pulDstTemp >= pendDst) break;
//ULONG sourceVal = EfbRead32(pulSrcTemp);
if ((xSrc + cxIdx) >= srcWidth) break;
if ((ySrc + cyIdx) >= srcHeight) break;
TexOffset = CalculateTexOffset(xSrc + cxIdx, ySrc + cyIdx, srcWidth);
//pulSrcTemp++;
*pulDstTemp = TexReadRgb(pSrcFbStart, TexOffset);
pulDstTemp++;
cxIdx++;
}
} else {
// Read normally from source, write swapped to dest.
while (cxTemp--) {
if (pulSrcTemp >= pendSrc) break;
if (pulDstTemp >= pendDst) break;
if ((xDst + cxIdx) >= destWidth) break;
if ((yDst + cyIdx) >= destHeight) break;
ULONG sourceVal = LoadToRegister32(*pulSrcTemp);
pulSrcTemp++;
//EfbWrite32(pulDstTemp, sourceVal);
//pulDstTemp++;
TexOffset = CalculateTexOffset(xDst + cxIdx, yDstStart + cyIdx, destWidth);
TexWriteRgb(pDestFbStart, TexOffset, sourceVal);
cxIdx++;
if (pulSrc >= pstartSrc && pulDst >= pstartDst) {
ULONG TexOffset = 0;
ULONG cxTemp = cx;
ULONG cxIdx = 0;
BOOLEAN validAccess;
if (copyFromFb) {
// Read swapped from source, write normally to dest.
while (cxTemp--) {
validAccess = pulSrcTemp >= pstartSrc && pulDstTemp >= pstartDst &&
pulSrcTemp < pendSrc && pulDstTemp < pendDst &&
(xSrc + cxIdx) < srcWidth &&
(ySrc + cyIdx) < srcHeight;
if (validAccess) {
TexOffset = CalculateTexOffset(xSrc + cxIdx, ySrc + cyIdx, srcWidth);
//pulSrcTemp++;
*pulDstTemp = TexReadRgb(pSrcFbStart, TexOffset);
}
pulDstTemp++;
cxIdx++;
}
} else {
// Read normally from source, write swapped to dest.
while (cxTemp--) {
validAccess = pulSrcTemp >= pstartSrc && pulDstTemp >= pstartDst &&
pulSrcTemp < pendSrc && pulDstTemp < pendDst &&
(xDst + cxIdx) < destWidth &&
(yDst + cyIdx) < destHeight;
if (validAccess) {
ULONG sourceVal = LoadToRegister32(*pulSrcTemp);
pulSrcTemp++;
//EfbWrite32(pulDstTemp, sourceVal);
//pulDstTemp++;
TexOffset = CalculateTexOffset(xDst + cxIdx, yDstStart + cyIdx, destWidth);
ULONG TexOffset2 = CalculateTexOffset(xDst + cxIdx + 1, yDstStart + cyIdx, destWidth);
if (((TexOffset & 3) == 0) && (TexOffset2 == (TexOffset + 2)) && (cxTemp > 1)) {
// Will be writing the next pixel too, we can optimise this to two writes from four reads and four writes.
ULONG sourceVal2 = LoadToRegister32(*pulSrcTemp);
pulSrcTemp++;
TexWriteRgb2Aligned(pDestFbStart, TexOffset, sourceVal, sourceVal2);
cxIdx++;
cxTemp--;
} else {
TexWriteRgb(pDestFbStart, TexOffset, sourceVal);
}
} else {
pulSrcTemp++;
}
cxIdx++;
}
}
}

94
fpvigxdll/source/draw.c
View File

@@ -40,6 +40,15 @@ static void TexWriteRgb(PUCHAR pTex, ULONG offset, ULONG rgb) {
}
}
static void TexWriteRgb2Aligned(PUCHAR pTex, ULONG offset, ULONG rgb0, ULONG rgb1) {
UCHAR r0 = (rgb0 >> 16) & 0xFF;
UCHAR r1 = (rgb1 >> 16) & 0xFF;
ULONG value = 0xFF00FF00 | ((rgb1 >> 16) & 0xFF) | (((rgb0 >> 16) & 0xFF) << 16);
ULONG value2 = ((rgb0 & 0xFFFF) << 16) | (rgb1 & 0xFFFF);
NativeWriteBase32(pTex, offset, value);
NativeWriteBase32(pTex, offset + 0x20, value2);
}
BOOL CopyBitsSwap32(
SURFOBJ *psoDest,
SURFOBJ *psoSrc,
@@ -102,41 +111,56 @@ BOOL copyFromFb)
PULONG pulSrcTemp = pulSrc;
PULONG pulDstTemp = pulDst;
// Bounds check the pointers, we could be in here when drawing off the screen
if (pulSrc < pstartSrc) break;
if (pulDst < pstartDst) break;
ULONG TexOffset = 0;
ULONG cxTemp = cx;
ULONG cxIdx = 0;
if (copyFromFb) {
// Read swapped from source, write normally to dest.
while (cxTemp--) {
if (pulSrcTemp >= pendSrc) break;
if (pulDstTemp >= pendDst) break;
//ULONG sourceVal = EfbRead32(pulSrcTemp);
if ((xSrc + cxIdx) >= srcWidth) break;
if ((ySrc + cyIdx) >= srcHeight) break;
TexOffset = CalculateTexOffset(xSrc + cxIdx, ySrc + cyIdx, srcWidth);
//pulSrcTemp++;
*pulDstTemp = TexReadRgb(pSrcFbStart, TexOffset);
pulDstTemp++;
cxIdx++;
}
} else {
// Read normally from source, write swapped to dest.
while (cxTemp--) {
if (pulSrcTemp >= pendSrc) break;
if (pulDstTemp >= pendDst) break;
if ((xDst + cxIdx) >= destWidth) break;
if ((yDst + cyIdx) >= destHeight) break;
ULONG sourceVal = LoadToRegister32(*pulSrcTemp);
pulSrcTemp++;
//EfbWrite32(pulDstTemp, sourceVal);
//pulDstTemp++;
TexOffset = CalculateTexOffset(xDst + cxIdx, yDstStart + cyIdx, destWidth);
TexWriteRgb(pDestFbStart, TexOffset, sourceVal);
cxIdx++;
if (pulSrc >= pstartSrc && pulDst >= pstartDst) {
ULONG TexOffset = 0;
ULONG cxTemp = cx;
ULONG cxIdx = 0;
BOOLEAN validAccess;
if (copyFromFb) {
// Read swapped from source, write normally to dest.
while (cxTemp--) {
validAccess = pulSrcTemp >= pstartSrc && pulDstTemp >= pstartDst &&
pulSrcTemp < pendSrc && pulDstTemp < pendDst &&
(xSrc + cxIdx) < srcWidth &&
(ySrc + cyIdx) < srcHeight;
if (validAccess) {
TexOffset = CalculateTexOffset(xSrc + cxIdx, ySrc + cyIdx, srcWidth);
//pulSrcTemp++;
*pulDstTemp = TexReadRgb(pSrcFbStart, TexOffset);
}
pulDstTemp++;
cxIdx++;
}
} else {
// Read normally from source, write swapped to dest.
while (cxTemp--) {
validAccess = pulSrcTemp >= pstartSrc && pulDstTemp >= pstartDst &&
pulSrcTemp < pendSrc && pulDstTemp < pendDst &&
(xDst + cxIdx) < destWidth &&
(yDst + cyIdx) < destHeight;
if (validAccess) {
ULONG sourceVal = LoadToRegister32(*pulSrcTemp);
pulSrcTemp++;
//EfbWrite32(pulDstTemp, sourceVal);
//pulDstTemp++;
TexOffset = CalculateTexOffset(xDst + cxIdx, yDstStart + cyIdx, destWidth);
ULONG TexOffset2 = CalculateTexOffset(xDst + cxIdx + 1, yDstStart + cyIdx, destWidth);
if (((TexOffset & 3) == 0) && (TexOffset2 == (TexOffset + 2)) && (cxTemp > 1)) {
// Will be writing the next pixel too, we can optimise this to two writes from four reads and four writes.
ULONG sourceVal2 = LoadToRegister32(*pulSrcTemp);
pulSrcTemp++;
TexWriteRgb2Aligned(pDestFbStart, TexOffset, sourceVal, sourceVal2);
cxIdx++;
cxTemp--;
} else {
TexWriteRgb(pDestFbStart, TexOffset, sourceVal);
}
} else {
pulSrcTemp++;
}
cxIdx++;
}
}
}