From af7856c1e3c2a96cdeb53f32b7f2b42e0cf227b9 Mon Sep 17 00:00:00 2001
From: walbourn_cp <SND\walbourn_cp>
Date: Fri, 12 Jul 2013 15:12:03 -0700
Subject: [PATCH] texconv: Enable multithreaded BC6H/BC7 compression using
 OpenMP - Added '-singleproc' command-line to turn off MT usage.

---
 Texconv/Texconv_Desktop_2010.vcxproj       | 12 +++++-----
 Texconv/Texconv_Desktop_2010_SDK80.vcxproj | 12 +++++-----
 Texconv/Texconv_Desktop_2012.vcxproj       | 12 +++++-----
 Texconv/Texconv_Desktop_2013.vcxproj       | 12 +++++-----
 Texconv/texconv.cpp                        | 26 +++++++++++++++++++++-
 5 files changed, 49 insertions(+), 25 deletions(-)

diff --git a/Texconv/Texconv_Desktop_2010.vcxproj b/Texconv/Texconv_Desktop_2010.vcxproj
index 18b842e..51adb67 100644
--- a/Texconv/Texconv_Desktop_2010.vcxproj
+++ b/Texconv/Texconv_Desktop_2010.vcxproj
@@ -129,7 +129,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>Disabled</Optimization>
       <RuntimeLibrary>MultiThreadedDebugDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
       <EnableEnhancedInstructionSet>StreamingSIMDExtensions2</EnableEnhancedInstructionSet>
@@ -169,7 +169,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>Disabled</Optimization>
       <RuntimeLibrary>MultiThreadedDebugDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
       <ExceptionHandling>Sync</ExceptionHandling>
@@ -207,7 +207,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>MaxSpeed</Optimization>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
@@ -248,7 +248,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>MaxSpeed</Optimization>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
@@ -288,7 +288,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>MaxSpeed</Optimization>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
@@ -329,7 +329,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>MaxSpeed</Optimization>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
diff --git a/Texconv/Texconv_Desktop_2010_SDK80.vcxproj b/Texconv/Texconv_Desktop_2010_SDK80.vcxproj
index 8bbcd75..dc2f243 100644
--- a/Texconv/Texconv_Desktop_2010_SDK80.vcxproj
+++ b/Texconv/Texconv_Desktop_2010_SDK80.vcxproj
@@ -117,7 +117,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>Disabled</Optimization>
       <RuntimeLibrary>MultiThreadedDebugDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
       <EnableEnhancedInstructionSet>StreamingSIMDExtensions2</EnableEnhancedInstructionSet>
@@ -157,7 +157,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>Disabled</Optimization>
       <RuntimeLibrary>MultiThreadedDebugDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
       <ExceptionHandling>Sync</ExceptionHandling>
@@ -195,7 +195,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>MaxSpeed</Optimization>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
@@ -236,7 +236,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>MaxSpeed</Optimization>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
@@ -276,7 +276,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>MaxSpeed</Optimization>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
@@ -317,7 +317,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>MaxSpeed</Optimization>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
diff --git a/Texconv/Texconv_Desktop_2012.vcxproj b/Texconv/Texconv_Desktop_2012.vcxproj
index 7980edd..ffb67ca 100644
--- a/Texconv/Texconv_Desktop_2012.vcxproj
+++ b/Texconv/Texconv_Desktop_2012.vcxproj
@@ -136,7 +136,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>Disabled</Optimization>
       <RuntimeLibrary>MultiThreadedDebugDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
       <EnableEnhancedInstructionSet>StreamingSIMDExtensions2</EnableEnhancedInstructionSet>
@@ -176,7 +176,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>Disabled</Optimization>
       <RuntimeLibrary>MultiThreadedDebugDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
       <ExceptionHandling>Sync</ExceptionHandling>
@@ -214,7 +214,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>MaxSpeed</Optimization>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
@@ -255,7 +255,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>MaxSpeed</Optimization>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
@@ -295,7 +295,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>MaxSpeed</Optimization>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
@@ -336,7 +336,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>MaxSpeed</Optimization>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
diff --git a/Texconv/Texconv_Desktop_2013.vcxproj b/Texconv/Texconv_Desktop_2013.vcxproj
index 5e255f8..3dfe0a5 100644
--- a/Texconv/Texconv_Desktop_2013.vcxproj
+++ b/Texconv/Texconv_Desktop_2013.vcxproj
@@ -136,7 +136,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>Disabled</Optimization>
       <RuntimeLibrary>MultiThreadedDebugDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
       <EnableEnhancedInstructionSet>StreamingSIMDExtensions2</EnableEnhancedInstructionSet>
@@ -176,7 +176,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>Disabled</Optimization>
       <RuntimeLibrary>MultiThreadedDebugDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
       <ExceptionHandling>Sync</ExceptionHandling>
@@ -214,7 +214,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>MaxSpeed</Optimization>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
@@ -255,7 +255,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>MaxSpeed</Optimization>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
@@ -295,7 +295,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>MaxSpeed</Optimization>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
@@ -336,7 +336,7 @@
       <WarningLevel>Level4</WarningLevel>
       <Optimization>MaxSpeed</Optimization>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
-      <OpenMPSupport>false</OpenMPSupport>
+      <OpenMPSupport>true</OpenMPSupport>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <FloatingPointModel>Fast</FloatingPointModel>
diff --git a/Texconv/texconv.cpp b/Texconv/texconv.cpp
index 0a5ed1a..0c42012 100644
--- a/Texconv/texconv.cpp
+++ b/Texconv/texconv.cpp
@@ -42,6 +42,7 @@ enum OPTIONS    // Note: dwOptions below assumes 32 or less options.
     OPT_EXPAND_LUMINANCE,
     OPT_TA_WRAP,
     OPT_TA_MIRROR,
+    OPT_FORCE_SINGLEPROC,
 };
 
 struct SConversion
@@ -88,6 +89,7 @@ SValue g_pOptions[] =
     { L"xlum",          OPT_EXPAND_LUMINANCE },
     { L"wrap",          OPT_TA_WRAP },
     { L"mirror",        OPT_TA_MIRROR },
+    { L"singleproc",    OPT_FORCE_SINGLEPROC },
     { nullptr,          0             }
 };
 
@@ -366,6 +368,9 @@ void PrintUsage()
     wprintf( L"\n                       (DDS output only)\n");
     wprintf( L"   -dx10               Force use of 'DX10' extended header\n");
     wprintf( L"\n   -nologo             suppress copyright message\n");
+#ifdef _OPENMP
+    wprintf( L"   -singleproc         Do not use multi-threaded compression\n");
+#endif
 
     wprintf( L"\n");
     wprintf( L"   <format>: ");
@@ -448,6 +453,7 @@ int __cdecl wmain(_In_ int argc, _In_z_count_(argc) wchar_t* argv[])
             if( (OPT_NOLOGO != dwOption) && (OPT_TYPELESS_UNORM != dwOption) && (OPT_TYPELESS_FLOAT != dwOption)
                 && (OPT_SEPALPHA != dwOption) && (OPT_PREMUL_ALPHA != dwOption) && (OPT_EXPAND_LUMINANCE != dwOption)
                 && (OPT_TA_WRAP != dwOption) && (OPT_TA_MIRROR != dwOption)
+                && (OPT_FORCE_SINGLEPROC != dwOption)
                 && (OPT_SRGB != dwOption) && (OPT_SRGBI != dwOption) && (OPT_SRGBO != dwOption)
                 && (OPT_HFLIP != dwOption) && (OPT_VFLIP != dwOption)
                 && (OPT_DDS_DWORD_ALIGN != dwOption) && (OPT_USE_DX10 != dwOption) )
@@ -1087,7 +1093,25 @@ int __cdecl wmain(_In_ int argc, _In_z_count_(argc) wchar_t* argv[])
                 goto LError;
             }
 
-            hr = Compress( img, nimg, info, tformat, TEX_COMPRESS_DEFAULT, 0.5f, *timage );
+            DWORD cflags = TEX_COMPRESS_DEFAULT;
+#ifdef _OPENMP
+            switch( tformat )
+            {
+            case DXGI_FORMAT_BC6H_TYPELESS:
+            case DXGI_FORMAT_BC6H_UF16:
+            case DXGI_FORMAT_BC6H_SF16:
+            case DXGI_FORMAT_BC7_TYPELESS:
+            case DXGI_FORMAT_BC7_UNORM:
+            case DXGI_FORMAT_BC7_UNORM_SRGB:
+                if ( !(dwOptions & (1 << OPT_FORCE_SINGLEPROC) ) )
+                {
+                    cflags |= TEX_COMPRESS_PARALLEL;
+                }
+                break;
+            }
+#endif
+
+            hr = Compress( img, nimg, info, tformat, cflags, 0.5f, *timage );
             if ( FAILED(hr) )
             {
                 wprintf( L" FAILED [compress] (%x)\n", hr);