From 56289b36fb0afcc385b32400ea8dc7886a25df0b Mon Sep 17 00:00:00 2001
From: profi200 <fd3194@gmx.de>
Date: Thu, 14 Jan 2021 22:10:36 +0100
Subject: [PATCH] Make scaling easier to configure at compile time.

---
 arm11/Makefile                                |  2 +-
 include/arm11/gpu_cmd_lists.h                 | 16 +++----
 source/arm11/hardware/lgyfb.c                 | 45 +++++++++++++------
 source/arm11/hardware/lgyfb_240x160.dma330    | 35 +++++++++++++++
 .../{lgyfb.dma330 => lgyfb_360x240.dma330}    | 11 -----
 5 files changed, 72 insertions(+), 37 deletions(-)
 create mode 100644 source/arm11/hardware/lgyfb_240x160.dma330
 rename source/arm11/hardware/{lgyfb.dma330 => lgyfb_360x240.dma330} (74%)

diff --git a/arm11/Makefile b/arm11/Makefile
index 7428767..c9562c7 100644
--- a/arm11/Makefile
+++ b/arm11/Makefile
@@ -24,7 +24,7 @@ SOURCES		:=	../kernel/source ../source ../source/hardware ../source/arm11 \
 DATA		:=
 INCLUDES	:=	../kernel/include ../include ../thirdparty
 DEFINES		:=	-DARM11 -D_3DS -DVERS_STRING=\"$(VERS_STRING)\" \
-				-DVERS_MAJOR=$(VERS_MAJOR) -DVERS_MINOR=$(VERS_MINOR)
+				-DVERS_MAJOR=$(VERS_MAJOR) -DVERS_MINOR=$(VERS_MINOR) -DOAF_SCALE_SELECT=2
 ASSETS		:=
 
 ifneq ($(strip $(NO_DEBUG)),)
diff --git a/include/arm11/gpu_cmd_lists.h b/include/arm11/gpu_cmd_lists.h
index 3e19405..86fcbe7 100644
--- a/include/arm11/gpu_cmd_lists.h
+++ b/include/arm11/gpu_cmd_lists.h
@@ -21,10 +21,7 @@
 #include "types.h"
 
 
-#define SCALE_SELECT 2
-
-
-#if SCALE_SELECT == 0
+#if OAF_SCALE_SELECT == 0
 // 240x160 no scaling, no filter.
 alignas(16) static const u8 gbaGpuInitList[1136] =
 {
@@ -75,7 +72,7 @@ alignas(16) static const u8 gbaGpuInitList[1136] =
 	0x00, 0x00, 0x00, 0x00, 0x62, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00,
 	0x18, 0x01, 0x0F, 0x00, 0x00, 0x00, 0x00, 0x00, 0x61, 0x00, 0x01, 0x00,
 	0x00, 0x00, 0x00, 0x00, 0x6A, 0x00, 0x07, 0x00, 0x00, 0x00, 0x00, 0x00,
-	0x81, 0x00, 0x4F, 0x80, 0x00, 0x01, 0x00, 0x01, 0x02, 0x00, 0x00, 0x00,
+	0x81, 0x00, 0x4F, 0x80, 0x00, 0x01, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00,
 	0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x03, 0x01, 0x00, 0x00, 0x00,
 	0x8E, 0x00, 0x0F, 0x00, 0x01, 0x10, 0x01, 0x00, 0x80, 0x00, 0x0B, 0x00,
 	0x00, 0x00, 0x01, 0x00, 0x80, 0x00, 0x04, 0x00, 0x01, 0x00, 0x00, 0x00,
@@ -167,7 +164,7 @@ alignas(16) static const u8 gbaGpuList2[448] =
 	0x78, 0x56, 0x34, 0x12, 0x10, 0x00, 0x0F, 0x00, 0x78, 0x56, 0x34, 0x12,
 	0x10, 0x00, 0x0F, 0x00
 };
-#elif SCALE_SELECT == 1
+#elif OAF_SCALE_SELECT == 1
 // 240x160 with bilinear scaling to 360x240, no filter.
 alignas(16) static const u8 gbaGpuInitList[1136] =
 {
@@ -310,7 +307,7 @@ alignas(16) static const u8 gbaGpuList2[448] =
 	0x78, 0x56, 0x34, 0x12, 0x10, 0x00, 0x0F, 0x00, 0x78, 0x56, 0x34, 0x12,
 	0x10, 0x00, 0x0F, 0x00
 };
-#elif SCALE_SELECT == 2
+#elif OAF_SCALE_SELECT == 2
 // 360x240, no filter.
 alignas(16) static const u8 gbaGpuInitList[1136] =
 {
@@ -361,7 +358,7 @@ alignas(16) static const u8 gbaGpuInitList[1136] =
 	0x00, 0x00, 0x00, 0x00, 0x62, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00,
 	0x18, 0x01, 0x0F, 0x00, 0x00, 0x00, 0x00, 0x00, 0x61, 0x00, 0x01, 0x00,
 	0x00, 0x00, 0x00, 0x00, 0x6A, 0x00, 0x07, 0x00, 0x00, 0x00, 0x00, 0x00,
-	0x81, 0x00, 0x4F, 0x80, 0x00, 0x02, 0x00, 0x02, 0x02, 0x00, 0x00, 0x00,
+	0x81, 0x00, 0x4F, 0x80, 0x00, 0x02, 0x00, 0x02, 0x00, 0x00, 0x00, 0x00,
 	0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x03, 0x01, 0x00, 0x00, 0x00, // Last 4 bytes: Texture format.
 	0x8E, 0x00, 0x0F, 0x00, 0x01, 0x10, 0x01, 0x00, 0x80, 0x00, 0x0B, 0x00,
 	0x00, 0x00, 0x01, 0x00, 0x80, 0x00, 0x04, 0x00, 0x01, 0x00, 0x00, 0x00,
@@ -454,6 +451,3 @@ alignas(16) static const u8 gbaGpuList2[448] =
 	0x10, 0x00, 0x0F, 0x00
 };
 #endif
-
-
-#undef SCALE_SELECT
diff --git a/source/arm11/hardware/lgyfb.c b/source/arm11/hardware/lgyfb.c
index c4e3ec1..999d4ad 100644
--- a/source/arm11/hardware/lgyfb.c
+++ b/source/arm11/hardware/lgyfb.c
@@ -5,7 +5,11 @@
 #include "arm11/hardware/interrupt.h"
 #include "hardware/corelink_dma-330.h"
 #include "arm11/hardware/lcd.h"
-#include "lgyfb_dma330.h"
+#if OAF_SCALE_SELECT < 2
+#include "lgyfb_240x160_dma330.h"
+#else
+#include "lgyfb_360x240_dma330.h"
+#endif
 #include "kevent.h"
 
 
@@ -121,14 +125,17 @@ void LGYFB_init(KEvent *frameReadyEvent)
 
 	g_frameReadyEvent = frameReadyEvent;
 
-	//REG_LGYFB_TOP_SIZE  = LGYFB_SIZE(240u, 160u);
+#if OAF_SCALE_SELECT < 2
+	REG_LGYFB_TOP_SIZE  = LGYFB_SIZE(240u, 160u);
+#else
 	REG_LGYFB_TOP_SIZE  = LGYFB_SIZE(360u, 240u);
+#endif
 	REG_LGYFB_TOP_STAT  = LGYFB_IRQ_MASK;
 	REG_LGYFB_TOP_IRQ   = 0;
 	REG_LGYFB_TOP_ALPHA = 0xFF;
 
 	/*
-	 * Limitations:
+	 * Scale matrix limitations:
 	 * First pattern bit must be 1 and last 0 (for V-scale) or it loses sync with the DS/GBA input.
 	 * Vertical scaling is fucked with identity matrix.
 	 *
@@ -142,6 +149,21 @@ void LGYFB_init(KEvent *frameReadyEvent)
 	 *
 	 * Note: At scanline start the in FIFO is all filled with the first pixel.
 	 */
+#if OAF_SCALE_SELECT < 2
+	// Identity.
+	static const s16 scaleMatrix[6 * 6] =
+	{
+		// Identity (no scaling). Don't use for vertical scaling!
+		     0,      0,      0,      0,      0,      0,
+		     0,      0,      0,      0,      0,      0,
+		     0,      0,      0,      0,      0,      0,
+		0x4000, 0x4000, 0x4000, 0x4000, 0x4000, 0x4000,
+		     0,      0,      0,      0,      0,      0,
+		     0,      0,      0,      0,      0,      0
+	};
+	setScaleMatrix(&REG_LGYFB_TOP_H_LEN, 0xF8, 6, 0b00111111, scaleMatrix);
+#else
+	// Other
 	static const s16 scaleMatrix[6 * 6] =
 	{
 		// Original from AGB_FIRM.
@@ -168,20 +190,10 @@ void LGYFB_init(KEvent *frameReadyEvent)
 		0x4000, 0x4000, 0x2000, 0x4000, 0x4000, 0x2000,
 		     0,      0,      0,      0,      0,      0,
 		     0,      0,      0,      0,      0,      0
-
-		// Identity (no scaling). Don't use for vertical scaling!
-		/*     0,      0,      0,      0,      0,      0,
-		     0,      0,      0,      0,      0,      0,
-		     0,      0,      0,      0,      0,      0,
-		0x4000, 0x4000, 0x4000, 0x4000, 0x4000, 0x4000,
-		     0,      0,      0,      0,      0,      0,
-		     0,      0,      0,      0,      0,      0*/
 	};
-	// Identity.
-	//setScaleMatrix(&REG_LGYFB_TOP_H_LEN, 0xF8, 6, 0b00111111, scaleMatrix);
-	// Other
 	setScaleMatrix(&REG_LGYFB_TOP_V_LEN, 0xF8, 6, 0b00011011, scaleMatrix);
 	setScaleMatrix(&REG_LGYFB_TOP_H_LEN, 0x00, 6, 0b00011011, scaleMatrix);
+#endif
 
 	// With RGB8 output solid red and blue are converted to 0xF8 and green to 0xFA.
 	// The green bias exists on the whole range of green colors.
@@ -190,8 +202,13 @@ void LGYFB_init(KEvent *frameReadyEvent)
 	// RGB8:    Observed best format. Invisible dithering and best color accuracy.
 	// RGB565:  A little visible dithering. Good color accuracy.
 	// RGB5551: Lots of visible dithering. Good color accuracy (a little worse than 565).
+#if OAF_SCALE_SELECT < 2
+	REG_LGYFB_TOP_CNT = LGYFB_DMA_E | LGYFB_OUT_SWIZZLE | LGYFB_OUT_FMT_8880 |
+	                    LGYFB_HSCALE_E | LGYFB_ENABLE;
+#else
 	REG_LGYFB_TOP_CNT = LGYFB_DMA_E | LGYFB_OUT_SWIZZLE | LGYFB_OUT_FMT_8880 |
 	                    LGYFB_HSCALE_E | LGYFB_VSCALE_E | LGYFB_ENABLE;
+#endif
 
 	IRQ_registerIsr(IRQ_CDMA_EVENT0, 13, 0, lgyFbDmaIrqHandler);
 }
diff --git a/source/arm11/hardware/lgyfb_240x160.dma330 b/source/arm11/hardware/lgyfb_240x160.dma330
new file mode 100644
index 0000000..9368075
--- /dev/null
+++ b/source/arm11/hardware/lgyfb_240x160.dma330
@@ -0,0 +1,35 @@
+# 8 bytes burst with 16 transfers. Total 128 bytes per burst.
+# Source fixed address and destination incrementing.
+# Source and destination unprivileged, non-secure data access.
+MOV CCR, SB16 SS64 SAF SP2 DB16 DS64 DAI DP2
+MOV SAR, 0x10311000
+MOV DAR, 0x18200000
+
+FLUSHP 14
+
+
+# Loop until the entire frame has been transferred.
+LP 20 # For height 160.
+	# Wait for a burst request.
+	WFP 14, burst
+
+	# Transfer 8 scanlines.
+	# For width 240:
+	#LP 29 # RGB5551 & RGB565, 128 bytes burst
+	LP 44 # RGB8, 128 bytes burst
+	#LP 59 # RGBA8, 128 bytes burst
+		LD
+		ST
+	LPEND
+	LDPB 14 # This must always be conditional on bursts. No idea why. Assembler bug?
+	ST
+
+	# Skip gaps swizzle edition™.
+	# For width 240 with 256x256 texture:
+	#ADDH DAR, 0x100 # RGB5551 & RGB565
+	ADDH DAR, 0x180 # RGB8
+	#ADDH DAR, 0x200 # RGBA8
+LPEND
+WMB
+SEV 0
+END
diff --git a/source/arm11/hardware/lgyfb.dma330 b/source/arm11/hardware/lgyfb_360x240.dma330
similarity index 74%
rename from source/arm11/hardware/lgyfb.dma330
rename to source/arm11/hardware/lgyfb_360x240.dma330
index 33bdba2..002f0a9 100644
--- a/source/arm11/hardware/lgyfb.dma330
+++ b/source/arm11/hardware/lgyfb_360x240.dma330
@@ -9,17 +9,11 @@ FLUSHP 14
 
 
 # Loop until the entire frame has been transferred.
-#LP 20 # For height 160.
 LP 30 # For height 240.
 	# Wait for a burst request.
 	WFP 14, burst
 
 	# Transfer 8 scanlines.
-	# For width 240:
-	#LP 29 # RGB5551 & RGB565, 128 bytes burst
-	#LP 44 # RGB8, 128 bytes burst
-	#LP 59 # RGBA8, 128 bytes burst
-
 	# For width 360:
 	#LP 44 # RGB5551 & RGB565, 128 bytes burst
 	LP 71 # RGB8, 120 bytes burst
@@ -31,11 +25,6 @@ LP 30 # For height 240.
 	ST
 
 	# Skip gaps swizzle edition™.
-	# For width 240 with 256x256 texture:
-	#ADDH DAR, 0x100 # RGB5551 & RGB565
-	#ADDH DAR, 0x180 # RGB8
-	#ADDH DAR, 0x200 # RGBA8
-
 	#For width 360 with 512x512 texture:
 	#ADDH DAR, 0x980 # RGB5551 & RGB565
 	ADDH DAR, 0xE40 # RGB8