From 6ec42ae8a1e5a82daa098951b5ad8ccc9c506b0f Mon Sep 17 00:00:00 2001
From: fincs <fincs.alt1@gmail.com>
Date: Sat, 11 Feb 2017 18:29:41 +0100
Subject: [PATCH] Calculate total mipmap texture size with geometric series
 formula

---
 include/c3d/texture.h | 34 +++++++++++++++++++++++++++++++++-
 source/texture.c      | 35 +++--------------------------------
 2 files changed, 36 insertions(+), 33 deletions(-)

diff --git a/include/c3d/texture.h b/include/c3d/texture.h
index 33200ad..d44e02b 100644
--- a/include/c3d/texture.h
+++ b/include/c3d/texture.h
@@ -34,7 +34,6 @@ typedef struct ALIGN(8)
 } C3D_TexInitParams;
 
 bool C3D_TexInitWithParams(C3D_Tex* tex, C3D_TexInitParams p);
-void* C3D_TexGetLevel(C3D_Tex* tex, int level, u32* size);
 void C3D_TexUploadLevel(C3D_Tex* tex, const void* data, int level);
 void C3D_TexGenerateMipmap(C3D_Tex* tex);
 void C3D_TexBind(int unitId, C3D_Tex* tex);
@@ -46,6 +45,32 @@ static inline int C3D_TexCalcMaxLevel(u32 width, u32 height)
 	return (31-__builtin_clz(width < height ? width : height)) - 3; // avoid sizes smaller than 8
 }
 
+static inline u32 C3D_TexCalcLevelSize(u32 size, int level)
+{
+	return size >> (2*level);
+}
+
+static inline u32 C3D_TexCalcTotalSize(u32 size, int maxLevel)
+{
+	/*
+	S  = s + sr + sr^2 + sr^3 + ... + sr^n
+	Sr = sr + sr^2 + sr^3 + ... + sr^(n+1)
+	S-Sr = s - sr^(n+1)
+	S(1-r) = s(1 - r^(n+1))
+	S = s (1 - r^(n+1)) / (1-r)
+
+	r = 1/4
+	1-r = 3/4
+
+	S = 4s (1 - (1/4)^(n+1)) / 3
+	S = 4s (1 - 1/4^(n+1)) / 3
+	S = (4/3) (s - s/4^(n+1))
+	S = (4/3) (s - s/(1<<(2n+2)))
+	S = (4/3) (s - s>>(2n+2))
+	*/
+	return (size - C3D_TexCalcLevelSize(size,maxLevel+1)) * 4 / 3;
+}
+
 static inline bool C3D_TexInit(C3D_Tex* tex, u16 width, u16 height, GPU_TEXCOLOR format)
 {
 	return C3D_TexInitWithParams(tex,
@@ -64,6 +89,13 @@ static inline bool C3D_TexInitVRAM(C3D_Tex* tex, u16 width, u16 height, GPU_TEXC
 		(C3D_TexInitParams){ width, height, 0, format, GPU_TEX_2D, true });
 }
 
+static inline void* C3D_TexGetLevel(C3D_Tex* tex, int level, u32* size)
+{
+	if (size) *size = C3D_TexCalcLevelSize(tex->size, level);
+	if (!level) return tex->data;
+	return (u8*)tex->data + C3D_TexCalcTotalSize(tex->size, level-1);
+}
+
 static inline void C3D_TexUpload(C3D_Tex* tex, const void* data)
 {
 	C3D_TexUploadLevel(tex, data, 0);
diff --git a/source/texture.c b/source/texture.c
index 94e5534..f6078d0 100644
--- a/source/texture.c
+++ b/source/texture.c
@@ -43,19 +43,9 @@ bool C3D_TexInitWithParams(C3D_Tex* tex, C3D_TexInitParams p)
 	u32 size = fmtSize(p.format);
 	if (!size) return false;
 	size *= (u32)p.width * p.height / 8;
+	u32 total_size = C3D_TexCalcLevelSize(size, p.maxLevel);
 
-	u32 alloc_size = size;
-	{
-		int i;
-		u32 level_size = size;
-		for (i = 0; i < p.maxLevel; i ++)
-		{
-			level_size >>= 2;
-			alloc_size += level_size;
-		}
-	}
-
-	tex->data = p.onVram ? vramAlloc(alloc_size) : linearAlloc(alloc_size);
+	tex->data = p.onVram ? vramAlloc(total_size) : linearAlloc(total_size);
 	if (!tex->data) return false;
 
 	tex->width = p.width;
@@ -74,24 +64,6 @@ bool C3D_TexInitWithParams(C3D_Tex* tex, C3D_TexInitParams p)
 	return true;
 }
 
-void* C3D_TexGetLevel(C3D_Tex* tex, int level, u32* size)
-{
-	u8* data = (u8*)tex->data;
-	u32 level_size = tex->size;
-	int i;
-	for (i = 0; i <= tex->maxLevel; i ++)
-	{
-		if (i == level)
-		{
-			if (size) *size = level_size;
-			return data;
-		}
-		data += level_size;
-		level_size >>= 2;
-	}
-	return NULL;
-}
-
 void C3D_TexUploadLevel(C3D_Tex* tex, const void* data, int level)
 {
 	if (!tex->data || addrIsVRAM(tex->data))
@@ -99,8 +71,7 @@ void C3D_TexUploadLevel(C3D_Tex* tex, const void* data, int level)
 
 	u32 size = 0;
 	void* out = C3D_TexGetLevel(tex, level, &size);
-	if (out)
-		memcpy(out, data, size);
+	memcpy(out, data, size);
 }
 
 static void C3Di_DownscaleRGBA8(u32* dst, const u32* src[4])