From 0055b8f7c7a07177868a6b94bc79f5b45b070997 Mon Sep 17 00:00:00 2001
From: Floogle <18466542+skyfloogle@users.noreply.github.com>
Date: Wed, 27 Mar 2024 22:54:00 +0100
Subject: [PATCH] full-range 24-bit colour

---
 source/3ds/video.c      | 13 +++++++++----
 source/3ds/video_hard.c |  6 ++++--
 2 files changed, 13 insertions(+), 6 deletions(-)

diff --git a/source/3ds/video.c b/source/3ds/video.c
index 4703071..02cd83b 100644
--- a/source/3ds/video.c
+++ b/source/3ds/video.c
@@ -162,10 +162,6 @@ void video_init() {
         GX_TRANSFER_IN_FORMAT(GX_TRANSFER_FMT_RGB8) | GX_TRANSFER_OUT_FORMAT(GX_TRANSFER_FMT_RGB8) | \
         GX_TRANSFER_SCALING(GX_TRANSFER_SCALE_NO))
 	C3D_Init(C3D_DEFAULT_CMDBUF_SIZE * 4);
-	finalScreen[0] = C3D_RenderTargetCreate(240, 400, GPU_RB_RGB8, -1);
-	C3D_RenderTargetSetOutput(finalScreen[0], GFX_TOP, GFX_LEFT, DISPLAY_TRANSFER_FLAGS);
-	finalScreen[1] = C3D_RenderTargetCreate(240, 400, GPU_RB_RGB8, -1);
-	C3D_RenderTargetSetOutput(finalScreen[1], GFX_TOP, GFX_RIGHT, DISPLAY_TRANSFER_FLAGS);
 
 	sFinal_dvlb = DVLB_ParseFile((u32 *)final_shbin, final_shbin_size);
 	shaderProgramInit(&sFinal);
@@ -190,6 +186,15 @@ void video_init() {
 
     video_hard_init();
 	video_soft_init();
+
+	// The hardware renderer creates 1 * 0.75MB + 4 * 1MB framebuffers.
+	// The 3DS has two 3MB VRAM banks, so for this to work, 3 framebuffers must go into one bank.
+	// However, the allocator alternates between banks, so we need to allocate those first,
+	// before even the final render targets.
+	finalScreen[0] = C3D_RenderTargetCreate(240, 400, GPU_RB_RGB8, -1);
+	C3D_RenderTargetSetOutput(finalScreen[0], GFX_TOP, GFX_LEFT, DISPLAY_TRANSFER_FLAGS);
+	finalScreen[1] = C3D_RenderTargetCreate(240, 400, GPU_RB_RGB8, -1);
+	C3D_RenderTargetSetOutput(finalScreen[1], GFX_TOP, GFX_RIGHT, DISPLAY_TRANSFER_FLAGS);
 }
 
 void video_render(int alt_buf) {
diff --git a/source/3ds/video_hard.c b/source/3ds/video_hard.c
index 864d8d9..39cc182 100644
--- a/source/3ds/video_hard.c
+++ b/source/3ds/video_hard.c
@@ -105,14 +105,16 @@ void video_hard_init() {
 
 	params.width = 512;
 	params.height = 512;
-	params.format = GPU_RGBA4;
+	params.format = GPU_RGB8;
 	params.onVram = true;
 	C3D_TexInitWithParams(&screenTexHard, NULL, params);
-	screenTarget = C3D_RenderTargetCreateFromTex(&screenTexHard, GPU_TEX_2D, 0, GPU_RB_DEPTH16);
+	// Drawing backwards with a depth buffer isn't faster, so omit the depth buffer.
+	screenTarget = C3D_RenderTargetCreateFromTex(&screenTexHard, GPU_TEX_2D, 0, -1);
 	C3D_RenderTargetClear(screenTarget, C3D_CLEAR_ALL, 0, 0);
 
 	params.width = 512;
 	params.height = 512;
+	params.format = GPU_RGBA8;
 	for (int i = 0; i < AFFINE_CACHE_SIZE; i++) {
 		C3D_TexInitWithParams(&tileMapCache[i].tex, NULL, params);
 		tileMapCache[i].target = C3D_RenderTargetCreateFromTex(&tileMapCache[i].tex, GPU_TEX_2D, 0, -1);