Merge branch 'master' into perfqueries.

Conflicts:
	Source/Core/VideoCommon/Src/VideoConfig.h
	Source/Core/VideoCommon/VideoCommon.vcxproj.filters
	Source/Plugins/Plugin_VideoDX11/Src/VertexManager.cpp
	Source/Plugins/Plugin_VideoSoftware/Src/Rasterizer.cpp
	Source/Plugins/Plugin_VideoSoftware/Src/Tev.cpp

1 files changed, 40 insertions, 15 deletions

diff --git a/Source/Core/VideoCommon/Src/BPStructs.cpp b/Source/Core/VideoCommon/Src/BPStructs.cpp
index cda0f7c7c0..a8d081666f 100644
--- a/Source/Core/VideoCommon/Src/BPStructs.cpp
+++ b/Source/Core/VideoCommon/Src/BPStructs.cpp
@@ -35,9 +35,9 @@
 
 using namespace BPFunctions;
 
-u32 mapTexAddress;
-bool mapTexFound;
-int numWrites;
+static u32 mapTexAddress;
+static bool mapTexFound;
+static int numWrites;
 
 extern volatile bool g_bSkipCurrentFrame;
 
@@ -81,6 +81,9 @@ void BPWritten(const BPCmd& bp)
 		  just stuff geometry in them and don't put state changes there
 	----------------------------------------------------------------------------------------------------------------
 	*/
+	
+	// check for invalid state, else unneeded configuration are built
+	g_video_backend->CheckInvalidState();
 
 	// Debugging only, this lets you skip a bp update
 	//static int times = 0;
@@ -359,9 +362,9 @@ void BPWritten(const BPCmd& bp)
 			PixelShaderManager::SetFogColorChanged();
 		break;
 	case BPMEM_ALPHACOMPARE: // Compare Alpha Values
-		PRIM_LOG("alphacmp: ref0=%d, ref1=%d, comp0=%d, comp1=%d, logic=%d", bpmem.alphaFunc.ref0,
-				bpmem.alphaFunc.ref1, bpmem.alphaFunc.comp0, bpmem.alphaFunc.comp1, bpmem.alphaFunc.logic);
-		PixelShaderManager::SetAlpha(bpmem.alphaFunc);
+		PRIM_LOG("alphacmp: ref0=%d, ref1=%d, comp0=%d, comp1=%d, logic=%d", bpmem.alpha_test.ref0,
+				bpmem.alpha_test.ref1, bpmem.alpha_test.comp0, bpmem.alpha_test.comp1, bpmem.alpha_test.logic);
+		PixelShaderManager::SetAlpha(bpmem.alpha_test);
 		break;
 	case BPMEM_BIAS: // BIAS
 		PRIM_LOG("ztex bias=0x%x", bpmem.ztex1.bias);
@@ -458,6 +461,7 @@ void BPWritten(const BPCmd& bp)
 
 	case BPMEM_ZCOMPARE:      // Set the Z-Compare and EFB pixel format
 		g_renderer->SetColorMask(); // alpha writing needs to be disabled if the new pixel format doesn't have an alpha channel
+		g_renderer->SetBlendMode(true);
 		OnPixelFormatChange();
 		break;
 
@@ -500,18 +504,39 @@ void BPWritten(const BPCmd& bp)
 		// if this is different from 0, manual TMEM management is used (GX_PreloadEntireTexture).
 		if (bp.newvalue != 0)
 		{
-			// NOTE(neobrain): Apparently tmemodd doesn't affect hardware behavior at all (libogc uses it just as a buffer and switches its contents with tmemeven whenever this is called)
+			// TODO: Not quite sure if this is completely correct (likely not)
+			// NOTE: libogc's implementation of GX_PreloadEntireTexture seems flawed, so it's not necessarily a good reference for RE'ing this feature.
+
 			BPS_TmemConfig& tmem_cfg = bpmem.tmem_config;
-			u8* ram_ptr = Memory::GetPointer(tmem_cfg.preload_addr << 5);
-			u32 tmem_addr = tmem_cfg.preload_tmem_even * TMEM_LINE_SIZE;
-			u32 size = tmem_cfg.preload_tile_info.count * 32;
+			u8* src_ptr = Memory::GetPointer(tmem_cfg.preload_addr << 5); // TODO: Should we add mask here on GC?
+			u32 size = tmem_cfg.preload_tile_info.count * TMEM_LINE_SIZE;
+			u32 tmem_addr_even = tmem_cfg.preload_tmem_even * TMEM_LINE_SIZE;
 
-			// Check if the game has overflowed TMEM, and copy up to the limit.
-			// Paper Mario does this when entering the Great Boogly Tree (Chap 2)
-			if ((tmem_addr + size) > TMEM_SIZE)
-				size = TMEM_SIZE - tmem_addr;
+			if (tmem_cfg.preload_tile_info.type != 3)
+			{
+				if (tmem_addr_even + size > TMEM_SIZE)
+					size = TMEM_SIZE - tmem_addr_even;
 
-			memcpy(texMem + tmem_addr, ram_ptr, size);
+				memcpy(texMem + tmem_addr_even, src_ptr, size);
+			}
+			else // RGBA8 tiles (and CI14, but that might just be stupid libogc!)
+			{
+				// AR and GB tiles are stored in separate TMEM banks => can't use a single memcpy for everything
+				u32 tmem_addr_odd = tmem_cfg.preload_tmem_odd * TMEM_LINE_SIZE;
+
+				for (u32 i = 0; i < tmem_cfg.preload_tile_info.count; ++i)
+				{
+					if (tmem_addr_even + TMEM_LINE_SIZE > TMEM_SIZE ||
+						tmem_addr_odd  + TMEM_LINE_SIZE > TMEM_SIZE)
+						break;
+
+					memcpy(texMem + tmem_addr_even, src_ptr, TMEM_LINE_SIZE);
+					memcpy(texMem + tmem_addr_odd, src_ptr + TMEM_LINE_SIZE, TMEM_LINE_SIZE);
+					tmem_addr_even += TMEM_LINE_SIZE;
+					tmem_addr_odd += TMEM_LINE_SIZE;
+					src_ptr += TMEM_LINE_SIZE * 2;
+				}
+			}
 		}
 		break;