DirectXTex: Fix for size rounding bug in GPU BC compressor

walbourn_cp · walbourn_cp · commit 840f0d4238e6 · 2014-01-22T11:06:35.000-08:00
diff --git a/DirectXTex/BCDirectCompute.cpp b/DirectXTex/BCDirectCompute.cpp
@@ -461,7 +461,7 @@ HRESULT GPUCompressBC::Compress( const Image& srcImage, const Image& destImage )
             //--- BC7 -----------------------------------------------------------------
             ID3D11ShaderResourceView* pSRVs[] = { sourceSRV.Get(), nullptr };
             RunComputeShader( pContext, m_BC7_tryMode456CS.Get(), pSRVs, 2, m_constBuffer.Get(),
-                              m_err1UAV.Get(), std::max<UINT>(uThreadGroupCount / 4, 1) );
+                              m_err1UAV.Get(), std::max<UINT>( (uThreadGroupCount + 3) / 4, 1) );
 
             for ( UINT i = 0; i < 3; ++i )
             {
@@ -523,14 +523,14 @@ HRESULT GPUCompressBC::Compress( const Image& srcImage, const Image& destImage )
 
             pSRVs[1] = m_err2SRV.Get();
             RunComputeShader( pContext, m_BC7_encodeBlockCS.Get(), pSRVs, 2, m_constBuffer.Get(),
-                              m_outputUAV.Get(), std::max<UINT>(uThreadGroupCount / 4, 1) );
+                              m_outputUAV.Get(), std::max<UINT>( (uThreadGroupCount + 3) / 4, 1) );
         }
         else
         {
             //--- BC6H ----------------------------------------------------------------
             ID3D11ShaderResourceView* pSRVs[] = { sourceSRV.Get(), nullptr };
             RunComputeShader( pContext, m_BC6H_tryModeG10CS.Get(), pSRVs, 2, m_constBuffer.Get(),
-                              m_err1UAV.Get(), std::max<UINT>(uThreadGroupCount / 4, 1) );
+                              m_err1UAV.Get(), std::max<UINT>( (uThreadGroupCount + 3) / 4, 1) );
 
             for ( UINT i = 0; i < 10; ++i )
             {
@@ -556,12 +556,12 @@ HRESULT GPUCompressBC::Compress( const Image& srcImage, const Image& destImage )
 
                 pSRVs[1] = (i & 1) ? m_err2SRV.Get() : m_err1SRV.Get();
                 RunComputeShader( pContext, m_BC6H_tryModeLE10CS.Get(), pSRVs, 2, m_constBuffer.Get(),
-                                  (i & 1) ? m_err1UAV.Get() : m_err2UAV.Get(), std::max<UINT>(uThreadGroupCount / 2, 1) );
+                                  (i & 1) ? m_err1UAV.Get() : m_err2UAV.Get(), std::max<UINT>( (uThreadGroupCount + 1) / 2, 1) );
             }               
 
             pSRVs[1] = m_err1SRV.Get();
             RunComputeShader( pContext, m_BC6H_encodeBlockCS.Get(), pSRVs, 2, m_constBuffer.Get(),
-                              m_outputUAV.Get(), std::max<UINT>(uThreadGroupCount / 2, 1) );
+                              m_outputUAV.Get(), std::max<UINT>( (uThreadGroupCount + 1) / 2, 1) );
         }
 
         start_block_id += n;

Original file line number	Diff line number	Diff line change
`@@ -461,7 +461,7 @@ HRESULT GPUCompressBC::Compress( const Image& srcImage, const Image& destImage )`
`461`	`461`	`//--- BC7 -----------------------------------------------------------------`
`462`	`462`	`ID3D11ShaderResourceView* pSRVs[] = { sourceSRV.Get(), nullptr };`
`463`	`463`	`RunComputeShader( pContext, m_BC7_tryMode456CS.Get(), pSRVs, 2, m_constBuffer.Get(),`
`464`		`- m_err1UAV.Get(), std::max<UINT>(uThreadGroupCount / 4, 1) );`
	`464`	`+ m_err1UAV.Get(), std::max<UINT>( (uThreadGroupCount + 3) / 4, 1) );`
`465`	`465`
`466`	`466`	`for ( UINT i = 0; i < 3; ++i )`
`467`	`467`	`{`
`@@ -523,14 +523,14 @@ HRESULT GPUCompressBC::Compress( const Image& srcImage, const Image& destImage )`
`523`	`523`
`524`	`524`	`pSRVs[1] = m_err2SRV.Get();`
`525`	`525`	`RunComputeShader( pContext, m_BC7_encodeBlockCS.Get(), pSRVs, 2, m_constBuffer.Get(),`
`526`		`- m_outputUAV.Get(), std::max<UINT>(uThreadGroupCount / 4, 1) );`
	`526`	`+ m_outputUAV.Get(), std::max<UINT>( (uThreadGroupCount + 3) / 4, 1) );`
`527`	`527`	`}`
`528`	`528`	`else`
`529`	`529`	`{`
`530`	`530`	`//--- BC6H ----------------------------------------------------------------`
`531`	`531`	`ID3D11ShaderResourceView* pSRVs[] = { sourceSRV.Get(), nullptr };`
`532`	`532`	`RunComputeShader( pContext, m_BC6H_tryModeG10CS.Get(), pSRVs, 2, m_constBuffer.Get(),`
`533`		`- m_err1UAV.Get(), std::max<UINT>(uThreadGroupCount / 4, 1) );`
	`533`	`+ m_err1UAV.Get(), std::max<UINT>( (uThreadGroupCount + 3) / 4, 1) );`
`534`	`534`
`535`	`535`	`for ( UINT i = 0; i < 10; ++i )`
`536`	`536`	`{`
`@@ -556,12 +556,12 @@ HRESULT GPUCompressBC::Compress( const Image& srcImage, const Image& destImage )`
`556`	`556`
`557`	`557`	`pSRVs[1] = (i & 1) ? m_err2SRV.Get() : m_err1SRV.Get();`
`558`	`558`	`RunComputeShader( pContext, m_BC6H_tryModeLE10CS.Get(), pSRVs, 2, m_constBuffer.Get(),`
`559`		`- (i & 1) ? m_err1UAV.Get() : m_err2UAV.Get(), std::max<UINT>(uThreadGroupCount / 2, 1) );`
	`559`	`+ (i & 1) ? m_err1UAV.Get() : m_err2UAV.Get(), std::max<UINT>( (uThreadGroupCount + 1) / 2, 1) );`
`560`	`560`	`}`
`561`	`561`
`562`	`562`	`pSRVs[1] = m_err1SRV.Get();`
`563`	`563`	`RunComputeShader( pContext, m_BC6H_encodeBlockCS.Get(), pSRVs, 2, m_constBuffer.Get(),`
`564`		`- m_outputUAV.Get(), std::max<UINT>(uThreadGroupCount / 2, 1) );`
	`564`	`+ m_outputUAV.Get(), std::max<UINT>( (uThreadGroupCount + 1) / 2, 1) );`
`565`	`565`	`}`
`566`	`566`
`567`	`567`	`start_block_id += n;`