Yidhra/DirectXTex
2
0
Fork 0
mirror of https://github.com/microsoft/DirectXTex.git synced 2025-07-12 13:10:14 +02:00
Code Issues Actions Packages Projects Releases Wiki Activity

BC7 compressor optimziation for only using mode 6

Browse Source
This commit is contained in:
Chuck Walbourn 2016-09-20 16:40:49 -07:00
parent bfa53ff6c0
commit 1b27a1b819
9 changed files with 94 additions and 63 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
DirectXTex/BC.h
View File

@ -55,11 +55,12 @@ extern const int g_aWeights4[16];
enum BC_FLAGS
{
BC_FLAGS_NONE = 0x0,
BC_FLAGS_DITHER_RGB = 0x10000, // Enables dithering for RGB colors for BC1-3
BC_FLAGS_DITHER_A = 0x20000, // Enables dithering for Alpha channel for BC1-3
BC_FLAGS_UNIFORM = 0x40000, // By default, uses perceptual weighting for BC1-3; this flag makes it a uniform weighting
BC_FLAGS_USE_3SUBSETS = 0x80000,// By default, BC7 skips mode 0 & 2; this flag adds those modes back
BC_FLAGS_NONE = 0x0,
BC_FLAGS_DITHER_RGB = 0x10000, // Enables dithering for RGB colors for BC1-3
BC_FLAGS_DITHER_A = 0x20000, // Enables dithering for Alpha channel for BC1-3
BC_FLAGS_UNIFORM = 0x40000, // By default, uses perceptual weighting for BC1-3; this flag makes it a uniform weighting
BC_FLAGS_USE_3SUBSETS = 0x80000, // By default, BC7 skips mode 0 & 2; this flag adds those modes back
BC_FLAGS_FORCE_BC7_MODE6 = 0x100000, // BC7 should only use mode 6; skip other modes
};
//-------------------------------------------------------------------------------------
@ -624,7 +625,7 @@ class D3DX_BC7 : private CBits< 16 >
{
public:
void Decode(_Out_writes_(NUM_PIXELS_PER_BLOCK) HDRColorA* pOut) const;
void Encode(bool skip3subsets, _In_reads_(NUM_PIXELS_PER_BLOCK) const HDRColorA* const pIn);
void Encode(DWORD flags, _In_reads_(NUM_PIXELS_PER_BLOCK) const HDRColorA* const pIn);
private:
struct ModeInfo

16
DirectXTex/BC6HBC7.cpp
View File

@ -634,7 +634,7 @@ namespace
_In_reads_(NUM_PIXELS_PER_BLOCK) const HDRColorA* const pPoints,
_Out_ HDRColorA* pX,
_Out_ HDRColorA* pY,
size_t cSteps,
_In_range_(3, 4) size_t cSteps,
size_t cPixels,
_In_reads_(cPixels) const size_t* pIndex)
{
@ -830,7 +830,7 @@ namespace
_In_reads_(NUM_PIXELS_PER_BLOCK) const HDRColorA* const pPoints,
_Out_ HDRColorA* pX,
_Out_ HDRColorA* pY,
size_t cSteps,
_In_range_(3, 4) size_t cSteps,
size_t cPixels,
_In_reads_(cPixels) const size_t* pIndex)
{
@ -2176,7 +2176,7 @@ void D3DX_BC7::Decode(HDRColorA* pOut) const
}
_Use_decl_annotations_
void D3DX_BC7::Encode(bool skip3subsets, const HDRColorA* const pIn)
void D3DX_BC7::Encode(DWORD flags, const HDRColorA* const pIn)
{
assert(pIn);
@ -2194,12 +2194,18 @@ void D3DX_BC7::Encode(bool skip3subsets, const HDRColorA* const pIn)
for (EP.uMode = 0; EP.uMode < 8 && fMSEBest > 0; ++EP.uMode)
{
if (skip3subsets && (EP.uMode == 0 || EP.uMode == 2))
if (!(flags & BC_FLAGS_USE_3SUBSETS) && (EP.uMode == 0 || EP.uMode == 2))
{
// 3 subset modes tend to be used rarely and add significant compression time
continue;
}
if ((flags & TEX_COMPRESS_BC7_QUICK) && (EP.uMode != 6))
{
// Use only mode 6
continue;
}
const size_t uShapes = size_t(1) << ms_aInfo[EP.uMode].uPartitionBits;
assert(uShapes <= BC7_MAX_SHAPES);
_Analysis_assume_(uShapes <= BC7_MAX_SHAPES);
@ -2902,5 +2908,5 @@ void DirectX::D3DXEncodeBC7(uint8_t *pBC, const XMVECTOR *pColor, DWORD flags)
{
assert(pBC && pColor);
static_assert(sizeof(D3DX_BC7) == 16, "D3DX_BC7 should be 16 bytes");
reinterpret_cast<D3DX_BC7*>(pBC)->Encode(!(flags& BC_FLAGS_USE_3SUBSETS), reinterpret_cast<const HDRColorA*>(pColor));
reinterpret_cast<D3DX_BC7*>(pBC)->Encode(flags, reinterpret_cast<const HDRColorA*>(pColor));
}

72
DirectXTex/BCDirectCompute.cpp
View File

@ -175,8 +175,7 @@ HRESULT GPUCompressBC::Initialize(ID3D11Device* pDevice)
//-------------------------------------------------------------------------------------
_Use_decl_annotations_
HRESULT GPUCompressBC::Prepare(size_t width, size_t height, DXGI_FORMAT format, float alphaWeight, bool skip3subsets)
HRESULT GPUCompressBC::Prepare(size_t width, size_t height, DWORD flags, DXGI_FORMAT format, float alphaWeight)
{
if (!width || !height || alphaWeight < 0.f)
return E_INVALIDARG;
@ -189,7 +188,16 @@ HRESULT GPUCompressBC::Prepare(size_t width, size_t height, DXGI_FORMAT format,
m_alphaWeight = alphaWeight;
m_skip3Subsets = skip3subsets;
if (flags & TEX_COMPRESS_BC7_QUICK)
{
m_bc7_mode02 = false;
m_bc7_mode137 = false;
}
else
{
m_bc7_mode02 = (flags & TEX_COMPRESS_BC7_USE_3SUBSETS) != 0;
m_bc7_mode137 = true;
}
size_t xblocks = std::max<size_t>(1, (width + 3) >> 2);
size_t yblocks = std::max<size_t>(1, (height + 3) >> 2);
@ -332,7 +340,6 @@ HRESULT GPUCompressBC::Prepare(size_t width, size_t height, DXGI_FORMAT format,
//-------------------------------------------------------------------------------------
_Use_decl_annotations_
HRESULT GPUCompressBC::Compress(const Image& srcImage, const Image& destImage)
{
if (!srcImage.pixels || !destImage.pixels)
@ -456,40 +463,43 @@ HRESULT GPUCompressBC::Compress(const Image& srcImage, const Image& destImage)
RunComputeShader(pContext, m_BC7_tryMode456CS.Get(), pSRVs, 2, m_constBuffer.Get(),
m_err1UAV.Get(), std::max<UINT>((uThreadGroupCount + 3) / 4, 1));
for (UINT i = 0; i < 3; ++i)
if (m_bc7_mode137)
{
static const UINT modes[] = { 1, 3, 7 };
// Mode 1: err1 -> err2
// Mode 3: err2 -> err1
// Mode 7: err1 -> err2
for (UINT i = 0; i < 3; ++i)
{
D3D11_MAPPED_SUBRESOURCE mapped;
HRESULT hr = pContext->Map(m_constBuffer.Get(), 0, D3D11_MAP_WRITE_DISCARD, 0, &mapped);
if (FAILED(hr))
static const UINT modes[] = { 1, 3, 7 };
// Mode 1: err1 -> err2
// Mode 3: err2 -> err1
// Mode 7: err1 -> err2
{
ResetContext(pContext);
return hr;
D3D11_MAPPED_SUBRESOURCE mapped;
HRESULT hr = pContext->Map(m_constBuffer.Get(), 0, D3D11_MAP_WRITE_DISCARD, 0, &mapped);
if (FAILED(hr))
{
ResetContext(pContext);
return hr;
}
ConstantsBC6HBC7 param;
param.tex_width = static_cast<UINT>(srcImage.width);
param.num_block_x = static_cast<UINT>(xblocks);
param.format = m_bcformat;
param.mode_id = modes[i];
param.start_block_id = start_block_id;
param.num_total_blocks = num_total_blocks;
param.alpha_weight = m_alphaWeight;
memcpy(mapped.pData, &param, sizeof(param));
pContext->Unmap(m_constBuffer.Get(), 0);
}
ConstantsBC6HBC7 param;
param.tex_width = static_cast<UINT>(srcImage.width);
param.num_block_x = static_cast<UINT>(xblocks);
param.format = m_bcformat;
param.mode_id = modes[i];
param.start_block_id = start_block_id;
param.num_total_blocks = num_total_blocks;
param.alpha_weight = m_alphaWeight;
memcpy(mapped.pData, &param, sizeof(param));
pContext->Unmap(m_constBuffer.Get(), 0);
pSRVs[1] = (i & 1) ? m_err2SRV.Get() : m_err1SRV.Get();
RunComputeShader(pContext, m_BC7_tryMode137CS.Get(), pSRVs, 2, m_constBuffer.Get(),
(i & 1) ? m_err1UAV.Get() : m_err2UAV.Get(), uThreadGroupCount);
}
pSRVs[1] = (i & 1) ? m_err2SRV.Get() : m_err1SRV.Get();
RunComputeShader(pContext, m_BC7_tryMode137CS.Get(), pSRVs, 2, m_constBuffer.Get(),
(i & 1) ? m_err1UAV.Get() : m_err2UAV.Get(), uThreadGroupCount);
}
if (!m_skip3Subsets)
if (m_bc7_mode02)
{
// 3 subset modes tend to be used rarely and add significant compression time
for (UINT i = 0; i < 2; ++i)
@ -524,7 +534,7 @@ HRESULT GPUCompressBC::Compress(const Image& srcImage, const Image& destImage)
}
}
pSRVs[1] = m_err2SRV.Get();
pSRVs[1] = (m_bc7_mode02 || m_bc7_mode137) ? m_err2SRV.Get() : m_err1SRV.Get();
RunComputeShader(pContext, m_BC7_encodeBlockCS.Get(), pSRVs, 2, m_constBuffer.Get(),
m_outputUAV.Get(), std::max<UINT>((uThreadGroupCount + 3) / 4, 1));
}

7
DirectXTex/BCDirectCompute.h
View File

@ -23,9 +23,9 @@ public:
HRESULT Initialize( _In_ ID3D11Device* pDevice );
HRESULT Prepare( _In_ size_t width, _In_ size_t height, _In_ DXGI_FORMAT format, _In_ float alphaWeight = 1.f, _In_ bool skip3subsets = true );
HRESULT Prepare( size_t width, size_t height, DWORD flags, DXGI_FORMAT format, float alphaWeight );
HRESULT Compress( _In_ const Image& srcImage, _In_ const Image& destImage );
HRESULT Compress( const Image& srcImage, const Image& destImage );
DXGI_FORMAT GetSourceFormat() const { return m_srcformat; }
@ -33,7 +33,8 @@ private:
DXGI_FORMAT m_bcformat;
DXGI_FORMAT m_srcformat;
float m_alphaWeight;
bool m_skip3Subsets;
bool m_bc7_mode02;
bool m_bc7_mode137;
size_t m_width;
size_t m_height;

23
DirectXTex/DirectXTex.h
View File

@ -503,30 +503,33 @@ namespace DirectX
enum TEX_COMPRESS_FLAGS
{
TEX_COMPRESS_DEFAULT = 0,
TEX_COMPRESS_DEFAULT = 0,
TEX_COMPRESS_RGB_DITHER = 0x10000,
TEX_COMPRESS_RGB_DITHER = 0x10000,
// Enables dithering RGB colors for BC1-3 compression
TEX_COMPRESS_A_DITHER = 0x20000,
TEX_COMPRESS_A_DITHER = 0x20000,
// Enables dithering alpha for BC1-3 compression
TEX_COMPRESS_DITHER = 0x30000,
TEX_COMPRESS_DITHER = 0x30000,
// Enables both RGB and alpha dithering for BC1-3 compression
TEX_COMPRESS_UNIFORM = 0x40000,
TEX_COMPRESS_UNIFORM = 0x40000,
// Uniform color weighting for BC1-3 compression; by default uses perceptual weighting
TEX_COMPRESS_BC7_USE_3SUBSETS = 0x80000,
TEX_COMPRESS_BC7_USE_3SUBSETS = 0x80000,
// Enables exhaustive search for BC7 compress for mode 0 and 2; by default skips trying these modes
TEX_COMPRESS_SRGB_IN = 0x1000000,
TEX_COMPRESS_SRGB_OUT = 0x2000000,
TEX_COMPRESS_SRGB = ( TEX_COMPRESS_SRGB_IN | TEX_COMPRESS_SRGB_OUT ),
TEX_COMPRESS_BC7_QUICK = 0x100000,
// Minimal modes (usually mode 6) for BC7 compression
TEX_COMPRESS_SRGB_IN = 0x1000000,
TEX_COMPRESS_SRGB_OUT = 0x2000000,
TEX_COMPRESS_SRGB = ( TEX_COMPRESS_SRGB_IN | TEX_COMPRESS_SRGB_OUT ),
// if the input format type is IsSRGB(), then SRGB_IN is on by default
// if the output format type is IsSRGB(), then SRGB_OUT is on by default
TEX_COMPRESS_PARALLEL = 0x10000000,
TEX_COMPRESS_PARALLEL = 0x10000000,
// Compress is free to use multithreading to improve performance (by default it does not use multithreading)
};

3
DirectXTex/DirectXTexCompress.cpp
View File

@ -33,7 +33,8 @@ namespace
static_assert(TEX_COMPRESS_DITHER == (BC_FLAGS_DITHER_RGB | BC_FLAGS_DITHER_A), "TEX_COMPRESS_* flags should match BC_FLAGS_*");
static_assert(TEX_COMPRESS_UNIFORM == BC_FLAGS_UNIFORM, "TEX_COMPRESS_* flags should match BC_FLAGS_*");
static_assert(TEX_COMPRESS_BC7_USE_3SUBSETS == BC_FLAGS_USE_3SUBSETS, "TEX_COMPRESS_* flags should match BC_FLAGS_*");
return (compress & (BC_FLAGS_DITHER_RGB | BC_FLAGS_DITHER_A | BC_FLAGS_UNIFORM | BC_FLAGS_USE_3SUBSETS));
static_assert(TEX_COMPRESS_BC7_QUICK == BC_FLAGS_FORCE_BC7_MODE6, "TEX_COMPRESS_* flags should match BC_FLAGS_*");
return (compress & (BC_FLAGS_DITHER_RGB | BC_FLAGS_DITHER_A | BC_FLAGS_UNIFORM | BC_FLAGS_USE_3SUBSETS | BC_FLAGS_FORCE_BC7_MODE6));
}
inline DWORD GetSRGBFlags(_In_ DWORD compress)

6
DirectXTex/DirectXTexCompressGPU.cpp
View File

@ -234,7 +234,7 @@ HRESULT DirectX::Compress(
if (FAILED(hr))
return hr;
hr = gpubc->Prepare(srcImage.width, srcImage.height, format, alphaWeight, !(compress & TEX_COMPRESS_BC7_USE_3SUBSETS));
hr = gpubc->Prepare(srcImage.width, srcImage.height, compress, format, alphaWeight);
if (FAILED(hr))
return hr;
@ -320,7 +320,7 @@ HRESULT DirectX::Compress(
for (size_t level = 0; level < metadata.mipLevels; ++level)
{
hr = gpubc->Prepare(w, h, format, alphaWeight, !(compress & TEX_COMPRESS_BC7_USE_3SUBSETS));
hr = gpubc->Prepare(w, h, compress, format, alphaWeight);
if (FAILED(hr))
{
cImages.Release();
@ -371,7 +371,7 @@ HRESULT DirectX::Compress(
for (size_t level = 0; level < metadata.mipLevels; ++level)
{
hr = gpubc->Prepare(w, h, format, alphaWeight, !(compress & TEX_COMPRESS_BC7_USE_3SUBSETS));
hr = gpubc->Prepare(w, h, compress, format, alphaWeight);
if (FAILED(hr))
{
cImages.Release();

1
DirectXTex/Filters.h
View File

@ -295,6 +295,7 @@ namespace TriangleFilter
}
assert(pFilter != 0);
_Analysis_assume_(pFilter != 0);
// Filter setup
size_t sizeInBytes = TF_FILTER_SIZE;

16
Texconv/texconv.cpp
View File

@ -81,6 +81,7 @@ enum OPTIONS
OPT_NORMAL_MAP_AMPLITUDE,
OPT_COMPRESS_UNIFORM,
OPT_COMPRESS_MAX,
OPT_COMPRESS_QUICK,
OPT_COMPRESS_DITHER,
OPT_WIC_QUALITY,
OPT_WIC_LOSSLESS,
@ -147,6 +148,7 @@ SValue g_pOptions[] =
{ L"nmapamp", OPT_NORMAL_MAP_AMPLITUDE },
{ L"bcuniform", OPT_COMPRESS_UNIFORM },
{ L"bcmax", OPT_COMPRESS_MAX },
{ L"bcquick", OPT_COMPRESS_QUICK },
{ L"bcdither", OPT_COMPRESS_DITHER },
{ L"wicq", OPT_WIC_QUALITY },
{ L"wiclossless", OPT_WIC_LOSSLESS },
@ -656,7 +658,8 @@ namespace
wprintf(L" -nogpu Do not use DirectCompute-based codecs\n");
wprintf(L" -bcuniform Use uniform rather than perceptual weighting for BC1-3\n");
wprintf(L" -bcdither Use dithering for BC1-3\n");
wprintf(L" -bcmax Use exchaustive compression (BC7 only)\n");
wprintf(L" -bcmax Use exhaustive compression (BC7 only)\n");
wprintf(L" -bcquick USe quick compression (BC7 only)\n");
wprintf(L" -wicq <quality> When writing images with WIC use quality (0.0 to 1.0)\n");
wprintf(L" -wiclossless When writing images with WIC use lossless mode\n");
wprintf(
@ -690,9 +693,10 @@ namespace
for (UINT adapterIndex = 0; DXGI_ERROR_NOT_FOUND != dxgiFactory->EnumAdapters(adapterIndex, adapter.ReleaseAndGetAddressOf()); ++adapterIndex)
{
DXGI_ADAPTER_DESC desc;
adapter->GetDesc(&desc);
wprintf(L" %u: VID:%04X, PID:%04X - %ls\n", adapterIndex, desc.VendorId, desc.DeviceId, desc.Description);
if (SUCCEEDED(adapter->GetDesc(&desc)))
{
wprintf(L" %u: VID:%04X, PID:%04X - %ls\n", adapterIndex, desc.VendorId, desc.DeviceId, desc.Description);
}
}
}
}
@ -1203,6 +1207,10 @@ int __cdecl wmain(_In_ int argc, _In_z_count_(argc) wchar_t* argv[])
dwCompress |= TEX_COMPRESS_BC7_USE_3SUBSETS;
break;
case OPT_COMPRESS_QUICK:
dwCompress |= TEX_COMPRESS_BC7_QUICK;
break;
case OPT_COMPRESS_DITHER:
dwCompress |= TEX_COMPRESS_DITHER;
break;