From d68a119bc95f866b903c0fccdf0d3f976cdd572d Mon Sep 17 00:00:00 2001
From: Pavel Krajcevski <pavel@cs.unc.edu>
Date: Sat, 25 Aug 2012 12:58:20 -0400
Subject: [PATCH] Fix SIMD recognition at least with GCC targets. Still need to
 test with MSVC

---
 BPTCEncoder/CMakeLists.txt            | 47 +++++++++++++++++++++++----
 BPTCEncoder/config/BC7Config.h.in     |  2 +-
 BPTCEncoder/config/testsse4.1.cpp     | 10 ------
 BPTCEncoder/src/RGBAEndpointsSIMD.cpp |  8 +++--
 4 files changed, 48 insertions(+), 19 deletions(-)
 delete mode 100644 BPTCEncoder/config/testsse4.1.cpp

diff --git a/BPTCEncoder/CMakeLists.txt b/BPTCEncoder/CMakeLists.txt
index c053b96..de660e7 100644
--- a/BPTCEncoder/CMakeLists.txt
+++ b/BPTCEncoder/CMakeLists.txt
@@ -1,24 +1,44 @@
 INCLUDE_DIRECTORIES(${TexC_SOURCE_DIR}/BPTCEncoder/include)
+INCLUDE_DIRECTORIES(${TexC_BINARY_DIR}/BPTCEncoder/include)
 
-INCLUDE(CheckCXXSourceCompiles)
+INCLUDE(CheckCXXSourceRuns)
 
 SET(OLD_CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS})
 IF(CMAKE_COMPILER_IS_GNUCC)
-	SET(CMAKE_REQUIRED_FLAGS -msse4.1)
-	CHECK_CXX_SOURCE_COMPILES("config/testsse4.1.cpp" HAS_SSE_41)
+	SET(CMAKE_REQUIRED_FLAGS -msse4.1 -E)
+	CHECK_CXX_SOURCE_RUNS("
+		#include <smmintrin.h>
+		int main() {
+		  const __m128 fv = _mm_set1_ps(1.0f);
+		  const __m128 fv2 = _mm_set1_ps(2.0f);
+		
+		  const __m128 ans = _mm_blend_ps(fv, fv2, 2);
+
+		  return ((int *)(&ans))[0];
+		}"
+		HAS_SSE_41
+	)
 
 	IF(HAS_SSE_41)
 		SET(CMAKE_REQUIRED_FLAGS -msse4.2)
-		CHECK_CXX_SOURCE_COMPILES("config/testsse4.2.cpp"	HAS_SSE_POPCNT)
+		CHECK_CXX_SOURCE_RUNS("
+			#include <smmintrin.h>
+			int main() {
+				const unsigned int testMe = 5;
+				return _mm_popcnt_u32(testMe);
+			}"
+			HAS_SSE_POPCNT
+		)
 	ENDIF(HAS_SSE_41)
 
 ELSEIF(MSVC)
+#!FIXME!
 ENDIF()
 SET(CMAKE_REQUIRED_FLAGS ${OLD_CMAKE_REQUIRED_FLAGS})
 
 CONFIGURE_FILE(
 	"config/BC7Config.h.in"
-	"src/BC7Config.h"
+	"include/BC7Config.h"
 )
 
 IF(CMAKE_COMPILER_IS_GNUCC)
@@ -38,6 +58,21 @@ SET( SOURCES
 )
 
 IF( HAS_SSE_41 ) 
+
+	IF ( HAS_SSE_POPCNT )
+		IF( MSVC )
+			ADD_DEFINITIONS( /arch:SSE4.2 )
+		ELSE() #Assume GCC
+			ADD_DEFINITIONS( -msse4.2 )
+		ENDIF()
+	ELSE()
+		IF( MSVC )
+			ADD_DEFINITIONS( /arch:SSE4.1 )
+		ELSE() #Assume GCC
+			ADD_DEFINITIONS( -msse4.1 )
+		ENDIF()
+	ENDIF()
+
 	SET( HEADERS
 		${HEADERS}
 		src/RGBAEndpointsSIMD.h
@@ -52,6 +87,6 @@ IF( HAS_SSE_41 )
 ENDIF( HAS_SSE_41 )
 
 ADD_LIBRARY( BPTCEncoder
+	${HEADERS}
 	${SOURCES}
-	${SIMD_SOURCES}
 )
diff --git a/BPTCEncoder/config/BC7Config.h.in b/BPTCEncoder/config/BC7Config.h.in
index 158812d..599835a 100644
--- a/BPTCEncoder/config/BC7Config.h.in
+++ b/BPTCEncoder/config/BC7Config.h.in
@@ -5,4 +5,4 @@
 // explicitly by the CMake build process.
 
 // Do we have the proper popcnt instruction defined?
-#define HAS_SSE_POPCNT @HAS_SSE_POPCNT@
+#cmakedefine HAS_SSE_POPCNT
diff --git a/BPTCEncoder/config/testsse4.1.cpp b/BPTCEncoder/config/testsse4.1.cpp
deleted file mode 100644
index 2792820..0000000
--- a/BPTCEncoder/config/testsse4.1.cpp
+++ /dev/null
@@ -1,10 +0,0 @@
-#include <smmintrin.h>
-
-int main() {
-  const __m128 fv = _mm_set1_ps(1.0f);
-  const __m128 fv2 = _mm_set1_ps(2.0f);
-
-  const __m128 ans = _mm_blend_ps(fv, fv2, 2);
-
-  return ((int *)(&ans))[0];
-}
diff --git a/BPTCEncoder/src/RGBAEndpointsSIMD.cpp b/BPTCEncoder/src/RGBAEndpointsSIMD.cpp
index 7625bee..efe16ff 100755
--- a/BPTCEncoder/src/RGBAEndpointsSIMD.cpp
+++ b/BPTCEncoder/src/RGBAEndpointsSIMD.cpp
@@ -119,7 +119,7 @@ __m128i RGBAVectorSIMD::ToPixel(const __m128i &qmask) const {
 
 	// !SPEED! We should figure out a way to get rid of these scalar operations.
 #ifdef HAS_SSE_POPCNT
-	const uint32 prec = _mm_popcnt32(((uint32 *)(&qmask))[0]);
+	const uint32 prec = _mm_popcnt_u32(((uint32 *)(&qmask))[0]);
 #else
 	const uint32 prec = popcnt32(((uint32 *)(&qmask))[0]);
 #endif
@@ -160,7 +160,7 @@ __m128i RGBAVectorSIMD::ToPixel(const __m128i &qmask, const int pBit) const {
 	
 	// !SPEED! We should figure out a way to get rid of these scalar operations.
 #ifdef HAS_SSE_POPCNT
-	const uint32 prec = _mm_popcnt32(((uint32 *)(&qmask))[0]);
+	const uint32 prec = _mm_popcnt_u32(((uint32 *)(&qmask))[0]);
 #else
 	const uint32 prec = popcnt32(((uint32 *)(&qmask))[0]);
 #endif
@@ -283,7 +283,11 @@ float RGBAClusterSIMD::QuantizedError(const RGBAVectorSIMD &p1, const RGBAVector
 	// nBuckets should be a power of two.
 	assert(!(nBuckets & (nBuckets - 1)));
 
+#ifdef HAS_SSE_POPCNT
 	const uint8 indexPrec = 8-_mm_popcnt_u32(~(nBuckets - 1) & 0xFF);
+#else
+	const uint8 indexPrec = 8-popcnt32(~(nBuckets - 1) & 0xFF);
+#endif
 	assert(indexPrec >= 2 && indexPrec <= 4);
 
 	typedef __m128i tInterpPair[2];