Use fast_math matrix multiplication for culling and sw transform

hrydgard · Mar 22, 2014 · bc12124 · bc12124
1 parent cbc46be
commit bc12124
Show file tree

Hide file tree

Showing 3 changed files with 10 additions and 5 deletions.
diff --git a/GPU/Math3D.h b/GPU/Math3D.h
@@ -20,6 +20,8 @@
 #include <cmath>
 #include "Common/Common.h"
 
+#include "math/fast/fast_matrix.h"
+
 #if defined(_M_SSE)
 #include <emmintrin.h>
 #endif
@@ -828,10 +830,7 @@ inline void Norm3ByMatrix43(float vecOut[3], const float v[3], const float m[12]
 }
 
 inline void Matrix4ByMatrix4(float out[16], const float a[16], const float b[16]) {
-	Vec4ByMatrix44(out, a, b);
-	Vec4ByMatrix44(out + 4, a + 4, b);
-	Vec4ByMatrix44(out + 8, a + 8, b);
-	Vec4ByMatrix44(out + 12, a + 12, b);
+	fast_matrix_mul_4x4(out, b, a);
 }
 
 inline void ConvertMatrix4x3To4x4(float *m4x4, const float *m4x3) {

diff --git a/UI/NativeApp.cpp b/UI/NativeApp.cpp
@@ -48,6 +48,7 @@
 #include "gfx/texture.h"
 #include "i18n/i18n.h"
 #include "input/input_state.h"
+#include "math/fast/fast_math.h"
 #include "math/math_util.h"
 #include "math/lin/matrix4x4.h"
 #include "ui/ui.h"
@@ -56,6 +57,7 @@
 #include "ui/view.h"
 #include "util/text/utf8.h"
 
+#include "Common/CPUDetect.h"
 #include "Common/FileUtil.h"
 #include "Common/LogManager.h"
 #include "Core/Config.h"
@@ -211,6 +213,10 @@ void NativeInit(int argc, const char *argv[],
 	monstartup("ppsspp_jni.so");
 #endif
 
+#ifdef ANDROID
+	InitFastMath(cpu_info.bNEON);
+#endif
+
 	bool skipLogo = false;
 	EnableFZ();
 	setlocale( LC_ALL, "C" );

diff --git a/native b/native