VideoCommon: Implement proper RGBA8 texture loading from tmem.

Source/Core/VideoCommon/Src/BPStructs.cpp

@@ -497,18 +497,39 @@ 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 (int i = 0; i < tmem_cfg.preload_tile_info.count; ++i)
+				{
+					if (tmem_addr_even + TMEM_LINE_SIZE > TMEM_SIZE ||
+						tmem_addr_even + 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;
 

Source/Core/VideoCommon/Src/TextureCacheBase.cpp

@@ -333,6 +333,7 @@ TextureCache::TCacheEntryBase* TextureCache::Load(unsigned int stage,
 	if (from_tmem) src_data = &texMem[bpmem.tex[stage/4].texImage1[stage%4].tmem_even * TMEM_LINE_SIZE];
 	else src_data = Memory::GetPointer(address);
 
+	// TODO: This doesn't hash GB tiles for preloaded RGBA8 textures (instead, it's hashing more data from the low tmem bank than it should)
 	tex_hash = GetHash64(src_data, texture_size, g_ActiveConfig.iSafeTextureCache_ColorSamples);
 	if (isPaletteTexture)
 	{
@@ -413,10 +414,19 @@ TextureCache::TCacheEntryBase* TextureCache::Load(unsigned int stage,
 		}
 	}
 
-	// TODO: RGBA8 textures are stored non-continuously in tmem, that might cause problems here when preloading is enabled
 	if (!using_custom_texture)
-		pcfmt = TexDecoder_Decode(temp, src_data, expandedWidth,
-					expandedHeight, texformat, tlutaddr, tlutfmt, g_ActiveConfig.backend_info.bUseRGBATextures);
+	{
+		if (!(texformat == GX_TF_RGBA8 && from_tmem))
+		{
+			pcfmt = TexDecoder_Decode(temp, src_data, expandedWidth,
+						expandedHeight, texformat, tlutaddr, tlutfmt, g_ActiveConfig.backend_info.bUseRGBATextures);
+		}
+		else
+		{
+			u8* src_data_gb = &texMem[bpmem.tex[stage/4].texImage2[stage%4].tmem_odd * TMEM_LINE_SIZE];
+			pcfmt = TexDecoder_DecodeRGBA8FromTmem(temp, src_data, src_data_gb, expandedWidth, expandedHeight);
+		}
+	}
 
 	// TODO: Cleanup. Plus, we still autogenerate mipmaps in certain cases (we shouldn't do that)
 	bool isPow2;

Source/Core/VideoCommon/Src/TextureDecoder.cpp

@@ -2519,7 +2519,7 @@ void TexDecoder_DecodeTexelRGBA8FromTmem(u8 *dst, const u8 *src_ar, const u8* sr
 {
 	u16 sBlk = s >> 2;
 	u16 tBlk = t >> 2;
-	u16 widthBlks = (imageWidth >> 2) + 1;
+	u16 widthBlks = (imageWidth >> 2) + 1; // TODO: Looks wrong. Shouldn't this be ((imageWidth-1)>>2)+1 ?
 	u32 base_ar = (tBlk * widthBlks + sBlk) << 4;
 	u32 base_gb = (tBlk * widthBlks + sBlk) << 4;
 	u16 blkS = s & 3;
@@ -2537,6 +2537,18 @@ void TexDecoder_DecodeTexelRGBA8FromTmem(u8 *dst, const u8 *src_ar, const u8* sr
 	dst[2] = val_addr_gb[1]; // B
 }
 
+PC_TexFormat TexDecoder_DecodeRGBA8FromTmem(u8* dst, const u8 *src_ar, const u8 *src_gb, int width, int height)
+{
+	// TODO for someone who cares: Make this less slow!
+	for (int y = 0; y < height; ++y)
+		for (int x = 0; x < width; ++x)
+		{
+			TexDecoder_DecodeTexelRGBA8FromTmem(dst, src_ar, src_gb, x, y, width-1);
+			dst += 4;
+		}
+
+	return PC_TEX_FMT_RGBA32;
+}
 
 const char* texfmt[] = {
 	// pixel

Source/Core/VideoCommon/Src/TextureDecoder.h

@@ -87,6 +87,7 @@ PC_TexFormat TexDecoder_Decode(u8 *dst, const u8 *src, int width, int height, in
 PC_TexFormat GetPC_TexFormat(int texformat, int tlutfmt);
 void TexDecoder_DecodeTexel(u8 *dst, const u8 *src, int s, int t, int imageWidth, int texformat, int tlutaddr, int tlutfmt);
 void TexDecoder_DecodeTexelRGBA8FromTmem(u8 *dst, const u8 *src_ar, const u8* src_gb, int s, int t, int imageWidth);
+PC_TexFormat TexDecoder_DecodeRGBA8FromTmem(u8* dst, const u8 *src_ar, const u8 *src_gb, int width, int height);
 void TexDecoder_SetTexFmtOverlayOptions(bool enable, bool center);
 
 #endif