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

Fix TextureDecoder.cl to work on both NVidia and ATI video cards. 

To do so I had to re-add the casting bloat removed in revision 6102.  Also, for some odd reason the NVidia OpenCL drivers don't like 8 bit rotations, but are okay with 2, 4 bit rotations.  These are apparently bugs in the NVidia drivers that are hopefully fixed in future versions.
Also, on linux make sure the TextureDecoder.cl file is copied from the shared data directory to the users directory.


git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6611 8ced0084-cf51-0410-be5f-012b33b47a6e
This commit is contained in:
Glenn Rice 2010-12-19 14:59:57 +00:00
parent d1631bec3f
commit 05ad8d01d3
6 changed files with 58 additions and 49 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

94
Data/User/OpenCL/TextureDecoder.cl
View File

@ -24,9 +24,9 @@ kernel void DecodeI4(global uchar *dst,
{ {
uchar4 val = vload4(srcOffset, src); uchar4 val = vload4(srcOffset, src);
uchar8 res; uchar8 res;
res.even = (val >> 4) & 0x0F; res.even = (val >> (uchar4)4) & (uchar4)0x0F;
res.odd = val & 0x0F; res.odd = val & (uchar4)0x0F;
res |= res << 4; res |= res << (uchar8)4;
vstore8(res, 0, dst + ((y + iy)*width + x)); vstore8(res, 0, dst + ((y + iy)*width + x));
srcOffset++; srcOffset++;
} }
@ -41,9 +41,9 @@ kernel void DecodeI4_RGBA(global uint *dst,
{ {
uchar4 val = vload4(srcOffset, src); uchar4 val = vload4(srcOffset, src);
uchar8 res; uchar8 res;
res.even = (val >> 4) & 0x0F; res.even = (val >> (uchar4)4) & (uchar4)0x0F;
res.odd = val & 0x0F; res.odd = val & (uchar4)0x0F;
res |= res << 4; res |= res << (uchar8)4;
vstore8(upsample(upsample(res,res),upsample(res,res)), 0, dst + ((y + iy)*width + x)); vstore8(upsample(upsample(res,res),upsample(res,res)), 0, dst + ((y + iy)*width + x));
srcOffset++; srcOffset++;
} }
@ -108,8 +108,8 @@ kernel void DecodeIA4(global ushort *dst,
for (int iy = 0; iy < 4; iy++) for (int iy = 0; iy < 4; iy++)
{ {
val = vload8(srcOffset++, src); val = vload8(srcOffset++, src);
res = upsample(val >> 4, val & 0xF); res = upsample(val >> (uchar8)4, val & (uchar8)0xF);
res |= res << 4; res |= res << (ushort8)4;
vstore8(res, 0, dst + y*width + x); vstore8(res, 0, dst + y*width + x);
dst+=width; dst+=width;
} }
@ -125,10 +125,10 @@ kernel void DecodeIA4_RGBA(global uint *dst,
for (int iy = 0; iy < 4; iy++) for (int iy = 0; iy < 4; iy++)
{ {
val = vload8(srcOffset++, src); val = vload8(srcOffset++, src);
uchar8 a = val >> 4; uchar8 a = val >> (uchar8)4;
uchar8 l = val & 0xF; uchar8 l = val & (uchar8)0xF;
res = upsample(upsample(a, l), upsample(l,l)); res = upsample(upsample(a, l), upsample(l,l));
res |= res << 4; res |= res << (uint8)4;
vstore8(res, 0, dst + y*width + x); vstore8(res, 0, dst + y*width + x);
dst+=width; dst+=width;
} }
@ -142,7 +142,8 @@ kernel void DecodeRGBA8(global ushort *dst,
for (int iy = 0; iy < 4; iy++) for (int iy = 0; iy < 4; iy++)
{ {
ushort8 val = (ushort8)(vload4(srcOffset, src), vload4(srcOffset + 4, src)); ushort8 val = (ushort8)(vload4(srcOffset, src), vload4(srcOffset + 4, src));
ushort8 bgra = rotate(val,8).s40516273; ushort8 temp = rotate(val, (ushort8)4);
ushort8 bgra = rotate(temp, (ushort8)4).s40516273;
vstore8(bgra, 0, dst + ((y + iy)*width + x) * 2); vstore8(bgra, 0, dst + ((y + iy)*width + x) * 2);
srcOffset++; srcOffset++;
} }
@ -175,7 +176,8 @@ kernel void DecodeRGB565(global ushort *dst,
dst += width*y + x; dst += width*y + x;
for (int iy = 0; iy < 4; iy++) for (int iy = 0; iy < 4; iy++)
{ {
vstore4(rotate(vload4(srcOffset++, src),8), 0, dst + iy*width); ushort4 val = rotate(vload4(srcOffset++, src),(ushort4)4);
vstore4(rotate(val,(ushort4)4), 0, dst + iy*width);
} }
} }
@ -189,10 +191,10 @@ kernel void DecodeRGB565_RGBA(global uchar *dst,
uchar8 val = vload8(srcOffset++, src); uchar8 val = vload8(srcOffset++, src);
uchar16 res; uchar16 res;
res.even.even = bitselect(val.even, val.even >> 5, 7); res.even.even = bitselect(val.even, val.even >> (uchar4)5, (uchar4)7);
res.odd.even = bitselect((val.odd >> 3) | (val.even << 5), val.even >> 1, 3); res.odd.even = bitselect((val.odd >> (uchar4)3) | (val.even << (uchar4)5), val.even >> (uchar4)1, (uchar4)3);
res.even.odd = bitselect(val.odd << 3, val.odd >> 2, 7); res.even.odd = bitselect(val.odd << (uchar4)3, val.odd >> (uchar4)2, (uchar4)7);
res.odd.odd = 0xFF; res.odd.odd = (uchar4)0xFF;
vstore16(res, 0, dst + ((y + iy)*width + x) * 4); vstore16(res, 0, dst + ((y + iy)*width + x) * 4);
} }
@ -207,16 +209,16 @@ kernel void DecodeRGB5A3(global uchar *dst,
uchar16 resNoAlpha, resAlpha, choice; uchar16 resNoAlpha, resAlpha, choice;
#define iterateRGB5A3() \ #define iterateRGB5A3() \
val = vload8(srcOffset++, src); \ val = vload8(srcOffset++, src); \
resNoAlpha.s26AE = val.even << 1; \ resNoAlpha.s26AE = val.even << (uchar4)1; \
resNoAlpha.s159D = val.even << 6 | val.odd >> 2; \ resNoAlpha.s159D = val.even << (uchar4)6 | val.odd >> (uchar4)2; \
resNoAlpha.s048C = val.odd << 3; \ resNoAlpha.s048C = val.odd << (uchar4)3; \
resNoAlpha = bitselect(resNoAlpha, resNoAlpha >> 5, 0x3); \ resNoAlpha = bitselect(resNoAlpha, resNoAlpha >> (uchar16)5, (uchar16)0x3); \
resNoAlpha.s37BF = 0xFF; \ resNoAlpha.s37BF = (uchar4)(0xFF); \
resAlpha.s26AE = bitselect(val.even << 4, val.even, 0xF); \ resAlpha.s26AE = bitselect(val.even << (uchar4)4, val.even, (uchar4)0xF); \
resAlpha.s159D = bitselect(val.odd, val.odd >> 4, 0xF); \ resAlpha.s159D = bitselect(val.odd, val.odd >> (uchar4)4, (uchar4)0xF); \
resAlpha.s048C = bitselect(val.odd << 4, val.odd, 0xF); \ resAlpha.s048C = bitselect(val.odd << (uchar4)4, val.odd, (uchar4)0xF); \
resAlpha.s37BF = bitselect(val.even << 1, val.even >> 2, 0x1C); \ resAlpha.s37BF = bitselect(val.even << (uchar4)1, val.even >> (uchar4)2, (uchar4)0x1C); \
resAlpha.s37BF = bitselect(resAlpha.s37BF, val.even >> 5, 0x3); \ resAlpha.s37BF = bitselect(resAlpha.s37BF, val.even >> (uchar4)5, (uchar4)0x3); \
choice = (uchar16)((uchar4)(val.even.s0), \ choice = (uchar16)((uchar4)(val.even.s0), \
(uchar4)(val.even.s1), \ (uchar4)(val.even.s1), \
(uchar4)(val.even.s2), \ (uchar4)(val.even.s2), \
@ -237,16 +239,16 @@ kernel void DecodeRGB5A3_RGBA(global uchar *dst,
uchar16 resNoAlpha, resAlpha, choice; uchar16 resNoAlpha, resAlpha, choice;
#define iterateRGB5A3_RGBA() \ #define iterateRGB5A3_RGBA() \
val = vload8(srcOffset++, src); \ val = vload8(srcOffset++, src); \
resNoAlpha.s048C = val.even << 1; \ resNoAlpha.s048C = val.even << (uchar4)1; \
resNoAlpha.s159D = val.even << 6 | val.odd >> 2; \ resNoAlpha.s159D = val.even << (uchar4)6 | val.odd >> (uchar4)2; \
resNoAlpha.s26AE = val.odd << 3; \ resNoAlpha.s26AE = val.odd << (uchar4)3; \
resNoAlpha = bitselect(resNoAlpha, resNoAlpha >> 5, 0x3); \ resNoAlpha = bitselect(resNoAlpha, resNoAlpha >> (uchar16)5, (uchar16)0x3); \
resNoAlpha.s37BF = 0xFF; \ resNoAlpha.s37BF = (uchar4)(0xFF); \
resAlpha.s048C = bitselect(val.even << 4, val.even, 0xF); \ resAlpha.s048C = bitselect(val.even << (uchar4)4, val.even, (uchar4)0xF); \
resAlpha.s159D = bitselect(val.odd, val.odd >> 4, 0xF); \ resAlpha.s159D = bitselect(val.odd, val.odd >> (uchar4)4, (uchar4)0xF); \
resAlpha.s26AE = bitselect(val.odd << 4, val.odd, 0xF); \ resAlpha.s26AE = bitselect(val.odd << (uchar4)4, val.odd, (uchar4)0xF); \
resAlpha.s37BF = bitselect(val.even << 1, val.even >> 2, 0x1C); \ resAlpha.s37BF = bitselect(val.even << (uchar4)1, val.even >> (uchar4)2, (uchar4)0x1C); \
resAlpha.s37BF = bitselect(resAlpha.s37BF, val.even >> 5, 0x3); \ resAlpha.s37BF = bitselect(resAlpha.s37BF, val.even >> (uchar4)5, (uchar4)0x3); \
choice = (uchar16)((uchar4)(val.even.s0), \ choice = (uchar16)((uchar4)(val.even.s0), \
(uchar4)(val.even.s1), \ (uchar4)(val.even.s1), \
(uchar4)(val.even.s2), \ (uchar4)(val.even.s2), \
@ -274,13 +276,13 @@ kernel void decodeCMPRBlock(global uchar *dst,
uchar2 colora565 = (uchar2)(val.s1, val.s3); uchar2 colora565 = (uchar2)(val.s1, val.s3);
uchar2 colorb565 = (uchar2)(val.s0, val.s2); uchar2 colorb565 = (uchar2)(val.s0, val.s2);
uchar8 color32 = (uchar8)(bitselect(colora565 << 3, colora565 >> 2, 7), uchar8 color32 = (uchar8)(bitselect(colora565 << (uchar2)3, colora565 >> (uchar2)2, (uchar2)7),
bitselect((colora565 >> 3) | (colorb565 << 5), colorb565 >> 1, 3), bitselect((colora565 >> (uchar2)3) | (colorb565 << (uchar2)5), colorb565 >> (uchar2)1, (uchar2)3),
bitselect(colorb565, colorb565 >> 5, 7), bitselect(colorb565, colorb565 >> (uchar2)5, (uchar2)7),
(uchar2)0xFF); (uchar2)0xFF);
ushort4 frac2 = convert_ushort4(color32.even) - convert_ushort4(color32.odd); ushort4 frac2 = convert_ushort4(color32.even) - convert_ushort4(color32.odd);
uchar4 frac = convert_uchar4((frac2 * 3) / 8); uchar4 frac = convert_uchar4((frac2 * (ushort4)3) / (ushort4)8);
ushort4 colorAlpha = upsample((uchar4)(color32.even.s0,color32.even.s1,color32.even.s2,0), ushort4 colorAlpha = upsample((uchar4)(color32.even.s0,color32.even.s1,color32.even.s2,0),
rhadd(color32.odd, color32.even)); rhadd(color32.odd, color32.even));
@ -320,13 +322,13 @@ kernel void decodeCMPRBlock_RGBA(global uchar *dst,
uchar2 colora565 = (uchar2)(val.s1, val.s3); uchar2 colora565 = (uchar2)(val.s1, val.s3);
uchar2 colorb565 = (uchar2)(val.s0, val.s2); uchar2 colorb565 = (uchar2)(val.s0, val.s2);
uchar8 color32 = (uchar8)(bitselect(colorb565, colorb565 >> 5, 7), uchar8 color32 = (uchar8)(bitselect(colorb565, colorb565 >> (uchar2)5, (uchar2)7),
bitselect((colora565 >> 3) | (colorb565 << 5), colorb565 >> 1, 3), bitselect((colora565 >> (uchar2)3) | (colorb565 << (uchar2)5), colorb565 >> (uchar2)1, (uchar2)3),
bitselect(colora565 << 3, colora565 >> 2, 7), bitselect(colora565 << (uchar2)3, colora565 >> (uchar2)2, (uchar2)7),
(uchar2)0xFF); (uchar2)0xFF);
ushort4 frac2 = convert_ushort4(color32.even) - convert_ushort4(color32.odd); ushort4 frac2 = convert_ushort4(color32.even) - convert_ushort4(color32.odd);
uchar4 frac = convert_uchar4((frac2 * 3) / 8); uchar4 frac = convert_uchar4((frac2 * (ushort4)3) / (ushort4)8);
ushort4 colorAlpha = upsample((uchar4)(color32.even.s0,color32.even.s1,color32.even.s2,0), ushort4 colorAlpha = upsample((uchar4)(color32.even.s0,color32.even.s1,color32.even.s2,0),
rhadd(color32.odd, color32.even)); rhadd(color32.odd, color32.even));
@ -356,4 +358,4 @@ kernel void DecodeCMPR_RGBA(global uchar *dst,
decodeCMPRBlock_RGBA(dst + 16, src, width); src += 8; decodeCMPRBlock_RGBA(dst + 16, src, width); src += 8;
decodeCMPRBlock_RGBA(dst + 16 * width, src, width); src += 8; decodeCMPRBlock_RGBA(dst + 16 * width, src, width); src += 8;
decodeCMPRBlock_RGBA(dst + 16 * (width + 1), src, width); decodeCMPRBlock_RGBA(dst + 16 * (width + 1), src, width);
} }

1
Source/Core/Common/Src/CommonPaths.h
View File

@ -100,6 +100,7 @@
#define SHADERCACHE_DIR "ShaderCache" #define SHADERCACHE_DIR "ShaderCache"
#define STATESAVES_DIR "StateSaves" #define STATESAVES_DIR "StateSaves"
#define SCREENSHOTS_DIR "ScreenShots" #define SCREENSHOTS_DIR "ScreenShots"
#define OPENCL_DIR "OpenCL"
#define LOAD_DIR "Load" #define LOAD_DIR "Load"
#define HIRES_TEXTURES_DIR LOAD_DIR DIR_SEP "Textures" #define HIRES_TEXTURES_DIR LOAD_DIR DIR_SEP "Textures"
#define DUMP_DIR "Dump" #define DUMP_DIR "Dump"

4
Source/Core/Common/Src/FileUtil.cpp
View File

@ -647,6 +647,7 @@ const char *GetUserPath(int DirIDX)
static char ShadersDir[MAX_PATH] = {0}; static char ShadersDir[MAX_PATH] = {0};
static char StateSavesDir[MAX_PATH] = {0}; static char StateSavesDir[MAX_PATH] = {0};
static char ScreenShotsDir[MAX_PATH] = {0}; static char ScreenShotsDir[MAX_PATH] = {0};
static char OpenCLDir[MAX_PATH] = {0};
static char HiresTexturesDir[MAX_PATH] = {0}; static char HiresTexturesDir[MAX_PATH] = {0};
static char DumpDir[MAX_PATH] = {0}; static char DumpDir[MAX_PATH] = {0};
static char DumpFramesDir[MAX_PATH] = {0}; static char DumpFramesDir[MAX_PATH] = {0};
@ -689,6 +690,7 @@ const char *GetUserPath(int DirIDX)
snprintf(ShadersDir, sizeof(ShadersDir), "%s" SHADERS_DIR DIR_SEP, UserDir); snprintf(ShadersDir, sizeof(ShadersDir), "%s" SHADERS_DIR DIR_SEP, UserDir);
snprintf(StateSavesDir, sizeof(StateSavesDir), "%s" STATESAVES_DIR DIR_SEP, UserDir); snprintf(StateSavesDir, sizeof(StateSavesDir), "%s" STATESAVES_DIR DIR_SEP, UserDir);
snprintf(ScreenShotsDir, sizeof(ScreenShotsDir), "%s" SCREENSHOTS_DIR DIR_SEP, UserDir); snprintf(ScreenShotsDir, sizeof(ScreenShotsDir), "%s" SCREENSHOTS_DIR DIR_SEP, UserDir);
snprintf(OpenCLDir, sizeof(OpenCLDir), "%s" OPENCL_DIR DIR_SEP, UserDir);
snprintf(HiresTexturesDir, sizeof(HiresTexturesDir), "%s" HIRES_TEXTURES_DIR DIR_SEP, UserDir); snprintf(HiresTexturesDir, sizeof(HiresTexturesDir), "%s" HIRES_TEXTURES_DIR DIR_SEP, UserDir);
snprintf(DumpDir, sizeof(DumpDir), "%s" DUMP_DIR DIR_SEP, UserDir); snprintf(DumpDir, sizeof(DumpDir), "%s" DUMP_DIR DIR_SEP, UserDir);
snprintf(DumpFramesDir, sizeof(DumpFramesDir), "%s" DUMP_FRAMES_DIR DIR_SEP, UserDir); snprintf(DumpFramesDir, sizeof(DumpFramesDir), "%s" DUMP_FRAMES_DIR DIR_SEP, UserDir);
@ -732,6 +734,8 @@ const char *GetUserPath(int DirIDX)
return StateSavesDir; return StateSavesDir;
case D_SCREENSHOTS_IDX: case D_SCREENSHOTS_IDX:
return ScreenShotsDir; return ScreenShotsDir;
case D_OPENCL_IDX:
return OpenCLDir;
case D_HIRESTEXTURES_IDX: case D_HIRESTEXTURES_IDX:
return HiresTexturesDir; return HiresTexturesDir;
case D_DUMP_IDX: case D_DUMP_IDX:

1
Source/Core/Common/Src/FileUtil.h
View File

@ -38,6 +38,7 @@ enum {
D_SHADERS_IDX, D_SHADERS_IDX,
D_STATESAVES_IDX, D_STATESAVES_IDX,
D_SCREENSHOTS_IDX, D_SCREENSHOTS_IDX,
D_OPENCL_IDX,
D_HIRESTEXTURES_IDX, D_HIRESTEXTURES_IDX,
D_DUMP_IDX, D_DUMP_IDX,
D_DUMPFRAMES_IDX, D_DUMPFRAMES_IDX,

1
Source/Core/DolphinWX/Src/Main.cpp
View File

@ -300,6 +300,7 @@ bool DolphinApp::OnInit()
File::CopyDir(SHARED_USER_DIR MAPS_DIR DIR_SEP, File::GetUserPath(D_MAPS_IDX)); File::CopyDir(SHARED_USER_DIR MAPS_DIR DIR_SEP, File::GetUserPath(D_MAPS_IDX));
File::CopyDir(SHARED_USER_DIR SHADERS_DIR DIR_SEP, File::GetUserPath(D_SHADERS_IDX)); File::CopyDir(SHARED_USER_DIR SHADERS_DIR DIR_SEP, File::GetUserPath(D_SHADERS_IDX));
File::CopyDir(SHARED_USER_DIR WII_USER_DIR DIR_SEP, File::GetUserPath(D_WIIUSER_IDX)); File::CopyDir(SHARED_USER_DIR WII_USER_DIR DIR_SEP, File::GetUserPath(D_WIIUSER_IDX));
File::CopyDir(SHARED_USER_DIR OPENCL_DIR DIR_SEP, File::GetUserPath(D_OPENCL_IDX));
if (!File::Exists(File::GetUserPath(D_GCUSER_IDX))) if (!File::Exists(File::GetUserPath(D_GCUSER_IDX)))
File::CreateFullPath(File::GetUserPath(D_GCUSER_IDX)); File::CreateFullPath(File::GetUserPath(D_GCUSER_IDX));

6
Source/Core/VideoCommon/Src/OpenCL/OCLTextureDecoder.cpp
View File

@ -103,7 +103,7 @@ void TexDecoder_OpenCL_Initialize()
char **binaries = NULL; char **binaries = NULL;
char filename[1024]; char filename[1024];
sprintf(filename, "%sOpenCL/kernel.bin", File::GetUserPath(D_USER_IDX)); sprintf(filename, "%skernel.bin", File::GetUserPath(D_OPENCL_IDX));
FILE *input = NULL; FILE *input = NULL;
@ -155,7 +155,7 @@ void TexDecoder_OpenCL_Initialize()
if (err) if (err)
{ {
std::string code; std::string code;
sprintf(filename, "%sOpenCL/TextureDecoder.cl", File::GetUserPath(D_USER_IDX)); sprintf(filename, "%sTextureDecoder.cl", File::GetUserPath(D_OPENCL_IDX));
if (!File::ReadFileToString(true, filename, code)) if (!File::ReadFileToString(true, filename, code))
{ {
ERROR_LOG(VIDEO, "Failed to load OpenCL code %s - file is missing?", filename); ERROR_LOG(VIDEO, "Failed to load OpenCL code %s - file is missing?", filename);
@ -204,7 +204,7 @@ void TexDecoder_OpenCL_Initialize()
if (!err) if (!err)
{ {
sprintf(filename, "%sOpenCL/kernel.bin", File::GetUserPath(D_USER_IDX)); sprintf(filename, "%skernel.bin", File::GetUserPath(D_OPENCL_IDX));
const char *current_rev = SVN_REV_STR + '\0'; const char *current_rev = SVN_REV_STR + '\0';
FILE *output = NULL; FILE *output = NULL;