yuzu-emu/FasTC
1
0
Fork 0
mirror of https://github.com/yuzu-emu/FasTC.git synced 2024-11-30 20:04:18 +01:00
Code Issues Packages Projects Releases Wiki Activity

Repurpose the rest of our scaffolding to use Compression Jobs

Browse Source
This commit is contained in:
Pavel Krajcevski 2013-03-09 13:36:39 -05:00
parent 435f935de3
commit ae2324153d
7 changed files with 44 additions and 60 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
BPTCEncoder/include/BC7Compressor.h
View File

@ -103,20 +103,12 @@ namespace BC7C
// Compress the image given as RGBA data to BC7 format. Width and Height are the dimensions of
// the image in pixels.
void CompressImageBC7(
const unsigned char *inBuf,
unsigned char *outBuf,
unsigned int width,
unsigned int height
);
void Compress(const CompressionJob &);
void CompressImageBC7Stats(
const unsigned char *inBuf,
unsigned char *outBuf,
unsigned int width,
unsigned int height,
BlockStatManager &statManager
);
// Perform a compression while recording all of the choices the compressor made into a
// list of statistics. We can use this to see whether or not certain heuristics are working, such as
// whether or not certain modes are being chosen more often than others, etc.
void CompressWithStats(const CompressionJob &, BlockStatManager &statManager);
#ifdef HAS_SSE_41
// Compress the image given as RGBA data to BC7 format using an algorithm optimized for SIMD
@ -132,5 +124,5 @@ namespace BC7C
#endif
// Decompress the image given as BC7 data to R8G8B8A8 format. Width and Height are the dimensions of the image in pixels.
void DecompressImageBC7(const unsigned char* inBuf, unsigned char* outBuf, unsigned int width, unsigned int height);
void Decompress(const DecompressionJob &);
}

36
BPTCEncoder/src/BC7Compressor.cpp
View File

@ -1484,14 +1484,13 @@ namespace BC7C
// 4-byte RGBA format. The width and height parameters specify the size of the image in pixels.
// The buffer pointed to by outBuf should be large enough to store the compressed image. This
// implementation has an 4:1 compression ratio.
void CompressImageBC7(const unsigned char *inBuf, unsigned char *outBuf, unsigned int width, unsigned int height)
void Compress(const CompressionJob &cj)
{
const int kMaxIters = BC7CompressionMode::kMaxAnnealingIterations;
BC7CompressionMode::MaxAnnealingIterations = std::min(kMaxIters, GetQualityLevel());
for(uint32 j = 0; j < height; j += 4)
const unsigned char *inBuf = cj.inBuf;
unsigned char *outBuf = cj.outBuf;
for(uint32 j = 0; j < cj.height; j += 4)
{
for(uint32 i = 0; i < width; i += 4)
for(uint32 i = 0; i < cj.width; i += 4)
{
// ExtractBlock(inBuf + i * 4, width, block);
CompressBC7Block((const uint32 *)inBuf, outBuf);
@ -1582,19 +1581,15 @@ namespace BC7C
}
#endif // HAS_ATOMICS
void CompressImageBC7Stats(
const unsigned char *inBuf,
unsigned char *outBuf,
unsigned int width,
unsigned int height,
void CompressWithStats(
const CompressionJob &cj,
BlockStatManager &statManager
) {
const int kMaxIters = BC7CompressionMode::kMaxAnnealingIterations;
BC7CompressionMode::MaxAnnealingIterations = std::min(kMaxIters, GetQualityLevel());
for(uint32 j = 0; j < height; j += 4)
const unsigned char *inBuf = cj.inBuf;
unsigned char *outBuf = cj.outBuf;
for(uint32 j = 0; j < cj.height; j += 4)
{
for(uint32 i = 0; i < width; i += 4)
for(uint32 i = 0; i < cj.width; i += 4)
{
// ExtractBlock(inBuf + i * 4, width, block);
CompressBC7Block((const uint32 *)inBuf, outBuf, statManager);
@ -2567,16 +2562,17 @@ namespace BC7C
}
// Convert the image from a BC7 buffer to a RGBA8 buffer
void DecompressImageBC7(const uint8 *inBuf, uint8* outBuf, unsigned int width, unsigned int height) {
void Decompress(const DecompressionJob &dj) {
unsigned char *outBuf = dj.outBuf;
unsigned int blockIdx = 0;
// for(unsigned int j = 0; j < height; j += 4, outBuf += width * 3 * 4)
for(unsigned int j = 0; j < height; j += 4)
for(unsigned int j = 0; j < dj.height; j += 4)
{
for(unsigned int i = 0; i < width; i += 4)
for(unsigned int i = 0; i < dj.width; i += 4)
{
uint32 pixels[16];
DecompressBC7Block(inBuf + (16*(blockIdx++)), pixels);
DecompressBC7Block(dj.inBuf + (16*(blockIdx++)), pixels);
memcpy(outBuf, pixels, 16 * sizeof(uint32));
//memcpy(outBuf + (width * 4), pixels + 4, 4 * sizeof(uint32));

15
Core/include/TexComp.h
View File

@ -103,24 +103,13 @@ extern bool CompressImageData(
// returns the compressed image data into outData. It is assumed that there is
// enough space allocated for outData to store the compressed data. Allocation
// is dependent on the compression format.
typedef void (* CompressionFunc)(
const unsigned char *inData, // Raw image data
unsigned char *outData, // Buffer to store compressed data.
unsigned int width, // Image width
unsigned int height // Image height
);
typedef void (* CompressionFunc)(const CompressionJob &);
// A compression function format. It takes the raw data and image dimensions and
// returns the compressed image data into outData. It is assumed that there is
// enough space allocated for outData to store the compressed data. Allocation
// is dependent on the compression format.
typedef void (* CompressionFuncWithStats)(
const unsigned char *inData, // Raw image data
unsigned char *outData, // Buffer to store compressed data.
unsigned int width, // Image width
unsigned int height, // Image height
BlockStatManager &statManager// Stat manager
);
typedef void (* CompressionFuncWithStats)(const CompressionJob &, BlockStatManager &statManager);
// This function computes the Peak Signal to Noise Ratio between a
// compressed image and a raw image.

5
Core/src/CompressedImage.cpp
View File

@ -127,7 +127,10 @@ bool CompressedImage::DecompressImage(unsigned char *outBuf, unsigned int outBuf
switch(m_Format) {
case eCompressionFormat_BPTC:
BC7C::DecompressImageBC7(m_Data, outBuf, m_Width, m_Height);
{
DecompressionJob dj (m_Data, outBuf, m_Width, m_Height);
BC7C::Decompress(dj);
}
break;
default:

9
Core/src/TexComp.cpp
View File

@ -82,7 +82,7 @@ static CompressionFuncWithStats ChooseFuncFromSettingsWithStats(const SCompress
switch(s.format) {
case eCompressionFormat_BPTC:
{
return BC7C::CompressImageBC7Stats;
return BC7C::CompressWithStats;
}
break;
@ -105,7 +105,7 @@ static CompressionFunc ChooseFuncFromSettings(const SCompressionSettings &s) {
return BC7C::CompressImageBC7SIMD;
}
#endif
return BC7C::CompressImageBC7;
return BC7C::Compress;
}
break;
@ -140,11 +140,12 @@ static double CompressImageInSerial(
stopWatch.Start();
// !FIXME! We're assuming that we have 4x4 blocks here...
CompressionJob cj (imgData, outBuf, imgDataSz / 16, 4);
if(fStats && settings.pStatManager) {
(*fStats)(imgData, outBuf, imgDataSz / 16, 4, *(settings.pStatManager));
(*fStats)(cj, *(settings.pStatManager));
}
else {
(*f)(imgData, outBuf, imgDataSz / 16, 4);
(*f)(cj);
}
stopWatch.Stop();

13
Core/src/ThreadGroup.cpp
View File

@ -86,10 +86,11 @@ void CmpThread::operator()() {
return;
}
CompressionJob cj (m_InBuf, m_OutBuf, m_Width, m_Height);
if(m_CmpFunc)
(*m_CmpFunc)(m_InBuf, m_OutBuf, m_Width, m_Height);
(*m_CmpFunc)(cj);
else
(*m_CmpFuncWithStats)(m_InBuf, m_OutBuf, m_Width, m_Height, *m_StatManager);
(*m_CmpFuncWithStats)(cj, *m_StatManager);
{
TCLock lock(*m_ParentCounterLock);
@ -112,7 +113,7 @@ ThreadGroup::ThreadGroup( int numThreads, const unsigned char *inBuf, unsigned i
, m_ThreadState(eThreadState_Done)
, m_ExitFlag(false)
, m_CompressedBlockSize(
(func == BC7C::CompressImageBC7
(func == BC7C::Compress
#ifdef HAS_SSE_41
|| func == BC7C::CompressImageBC7SIMD
#endif
@ -122,7 +123,7 @@ ThreadGroup::ThreadGroup( int numThreads, const unsigned char *inBuf, unsigned i
0
)
, m_UncompressedBlockSize(
(func == BC7C::CompressImageBC7
(func == BC7C::Compress
#ifdef HAS_SSE_41
|| func == BC7C::CompressImageBC7SIMD
#endif
@ -162,13 +163,13 @@ ThreadGroup::ThreadGroup(
, m_ThreadState(eThreadState_Done)
, m_ExitFlag(false)
, m_CompressedBlockSize(
(func == BC7C::CompressImageBC7Stats)?
(func == BC7C::CompressWithStats)?
16
:
0
)
, m_UncompressedBlockSize(
(func == BC7C::CompressImageBC7Stats)?
(func == BC7C::CompressWithStats)?
64
:
0

6
Core/src/WorkerQueue.cpp
View File

@ -100,10 +100,12 @@ void WorkerThread::operator()() {
{
const uint8 *src = m_Parent->GetSrcForThread(m_ThreadIdx);
uint8 *dst = m_Parent->GetDstForThread(m_ThreadIdx);
CompressionJob cj (src, dst, 4 * m_Parent->GetNumBlocksForThread(m_ThreadIdx), 4);
if(f)
(*f)(src, dst, 4 * m_Parent->GetNumBlocksForThread(m_ThreadIdx), 4);
(*f)(cj);
else
(*fStat)(src, dst, 4 * m_Parent->GetNumBlocksForThread(m_ThreadIdx), 4, *statManager);
(*fStat)(cj, *statManager);
break;
}