Merge pull request #55 from spicyjpeg/psxmdec

Full MDEC support, C library refactors, cleanups and bugfixes (v0.20)
author: John "Lameguy" Wilbert Villamor <lameguy64@gmail.com> 2022-09-26 16:49:56 +0800
committer: GitHub <noreply@github.com> 2022-09-26 16:49:56 +0800
commit: c4a2533d21dfd05cde841ea48c67b05e0e6a853f (patch)
tree: c7ef61653b157b69fb0956709366996ddbc4ecfa /libpsn00b
parent: a8b404b3400c3ebd8e0b923dcaefcc49ea563e36 (diff)
parent: 86f0064afb8200e60dd80827535cac30d0eab028 (diff)
download: psn00bsdk-c4a2533d21dfd05cde841ea48c67b05e0e6a853f.tar.gz
104 files changed, 5059 insertions, 4058 deletions
diff --git a/libpsn00b/CMakeLists.txt b/libpsn00b/CMakeLists.txt
index f9c4f9d..a662448 100644
--- a/libpsn00b/CMakeLists.txt
+++ b/libpsn00b/CMakeLists.txt
@@ -40,6 +40,8 @@ foreach(_library IN LISTS PSN00BSDK_LIBRARIES)
 	psn00bsdk_add_library(${_library} STATIC ${_sources})
 endforeach()
 
+psn00bsdk_target_incbin(psxgpu PRIVATE _gpu_debug_font psxgpu/dbugfont.tim)
+
 # Extract libgcc's contents and merge them into libc after building.
 # Unfortunately glob expressions won't work on Windows, so we have to manually
 # enumerate the contents of libgcc and save the list to a temporary file (as it
diff --git a/libpsn00b/cmake/flags.cmake b/libpsn00b/cmake/flags.cmake
index e31773f..5d9c751 100644
--- a/libpsn00b/cmake/flags.cmake
+++ b/libpsn00b/cmake/flags.cmake
@@ -42,6 +42,8 @@ target_compile_options(
 		-march=r3000
 		-mtune=r3000
 		-mabi=32
+		-mno-mt
+		-mno-llsc
 		-mdivide-breaks
 		-O2
 		# Standard library options
@@ -49,6 +51,7 @@ target_compile_options(
 		-fno-builtin
 		-nostdlib
 		# Other options
+		-g
 		-fdata-sections
 		-ffunction-sections
 		-fsigned-char
diff --git a/libpsn00b/cmake/sdk.cmake b/libpsn00b/cmake/sdk.cmake
index d6d9bcd..8965e79 100644
--- a/libpsn00b/cmake/sdk.cmake
+++ b/libpsn00b/cmake/sdk.cmake
@@ -38,6 +38,8 @@ find_program(
 	HINTS
 		${PSN00BSDK_TC}/bin
 		${PSN00BSDK_TC}/../bin
+		# Same as ${CMAKE_INSTALL_PREFIX}/bin
+		${CMAKE_CURRENT_LIST_DIR}/../../../bin
 		# Same as ${CMAKE_INSTALL_PREFIX}/${PSN00BSDK_TARGET}/bin
 		${CMAKE_CURRENT_LIST_DIR}/../../../${PSN00BSDK_TARGET}/bin
 	PATHS
diff --git a/libpsn00b/include/ctype.h b/libpsn00b/include/ctype.h
index b79498a..24ee9d9 100644
--- a/libpsn00b/include/ctype.h
+++ b/libpsn00b/include/ctype.h
@@ -1,7 +1,20 @@
-#ifndef _CTYPE_H
-#define _CTYPE_H
+/*
+ * PSn00bSDK standard library
+ * (C) 2019-2022 PSXSDK authors, Lameguy64, spicyjpeg - MPL licensed
+ */
 
-extern int tolower(int chr);
-extern int toupper(int chr);
+#ifndef __CTYPE_H
+#define __CTYPE_H
 
-#endif
-\ No newline at end of file
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int tolower(int chr);
+int toupper(int chr);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/libpsn00b/include/gtereg.h b/libpsn00b/include/gtereg.inc
index 0d051fc..0d051fc 100644
--- a/libpsn00b/include/gtereg.h
+++ b/libpsn00b/include/gtereg.inc
diff --git a/libpsn00b/include/hwregs_a.h b/libpsn00b/include/hwregs_a.inc
index 8a504f5..c78b41a 100644
--- a/libpsn00b/include/hwregs_a.h
+++ b/libpsn00b/include/hwregs_a.inc
@@ -1,10 +1,11 @@
-# Hardware register definitions for GNU assembler (as)
-#
-# Part of the PSn00bSDK Project by Lameguy64
-# 2019 Meido-Tek Productions
+# PSn00bSDK hardware registers definitions
+# (C) 2019-2022 Lameguy64, spicyjpeg - MPL licensed
 
+## Constants
 
-.set IOBASE,			0x1f80		# IO segment base
+.set IOBASE,			0x1f80
+.set F_CPU,				33868800
+.set F_GPU,				53222400
 
 ## GPU
 
diff --git a/libpsn00b/include/hwregs_c.h b/libpsn00b/include/hwregs_c.h
index e533c56..b205b87 100644
--- a/libpsn00b/include/hwregs_c.h
+++ b/libpsn00b/include/hwregs_c.h
@@ -14,116 +14,117 @@
 
 /* Constants */
 
+#define IOBASE				0xbf800000
 #define F_CPU				33868800UL
 #define F_GPU				53222400UL
 
 /* GPU */
 
-#define GPU_GP0				_MMIO32(0x1f801810)
-#define GPU_GP1				_MMIO32(0x1f801814)
+#define GPU_GP0				_MMIO32(IOBASE | 0x1810)
+#define GPU_GP1				_MMIO32(IOBASE | 0x1814)
 
 /* CD drive */
 
-#define CD_STAT				_MMIO8(0x1f801800)
-#define CD_CMD				_MMIO8(0x1f801801)
-#define CD_DATA				_MMIO8(0x1f801802)
-#define CD_IRQ				_MMIO8(0x1f801803)
+#define CD_STAT				_MMIO8(IOBASE | 0x1800)
+#define CD_CMD				_MMIO8(IOBASE | 0x1801)
+#define CD_DATA				_MMIO8(IOBASE | 0x1802)
+#define CD_IRQ				_MMIO8(IOBASE | 0x1803)
 
-#define CD_REG(N)			_MMIO8(0x1f801800 + (N))
+#define CD_REG(N)			_MMIO8(IOBASE | 0x1800 + (N))
 
 /* SPU */
 
-#define SPU_MASTER_VOL_L	_MMIO16(0x1f801d80)
-#define SPU_MASTER_VOL_R	_MMIO16(0x1f801d82)
-#define SPU_REVERB_VOL_L	_MMIO16(0x1f801d84)
-#define SPU_REVERB_VOL_R	_MMIO16(0x1f801d86)
-#define SPU_KEY_ON			_MMIO32(0x1f801d88)
-#define SPU_KEY_OFF			_MMIO32(0x1f801d8c)
-#define SPU_FM_MODE			_MMIO32(0x1f801d90)
-#define SPU_NOISE_MODE		_MMIO32(0x1f801d94)
-#define SPU_REVERB_ON		_MMIO32(0x1f801d98)
-#define SPU_CHAN_STATUS		_MMIO32(0x1f801d9c)
-
-#define SPU_REVERB_ADDR		_MMIO16(0x1f801da2)
-#define SPU_IRQ_ADDR		_MMIO16(0x1f801da4)
-#define SPU_ADDR			_MMIO16(0x1f801da6)
-#define SPU_DATA			_MMIO16(0x1f801da8)
-
-#define SPU_CTRL			_MMIO16(0x1f801daa)
-#define SPU_DMA_CTRL		_MMIO16(0x1f801dac)
-#define SPU_STAT			_MMIO16(0x1f801dae)
-
-#define SPU_CD_VOL_L		_MMIO16(0x1f801db0)
-#define SPU_CD_VOL_R		_MMIO16(0x1f801db2)
-#define SPU_EXT_VOL_L		_MMIO16(0x1f801db4)
-#define SPU_EXT_VOL_R		_MMIO16(0x1f801db6)
-#define SPU_CURRENT_VOL_L	_MMIO16(0x1f801db8)
-#define SPU_CURRENT_VOL_R	_MMIO16(0x1f801dba)
+#define SPU_MASTER_VOL_L	_MMIO16(IOBASE | 0x1d80)
+#define SPU_MASTER_VOL_R	_MMIO16(IOBASE | 0x1d82)
+#define SPU_REVERB_VOL_L	_MMIO16(IOBASE | 0x1d84)
+#define SPU_REVERB_VOL_R	_MMIO16(IOBASE | 0x1d86)
+#define SPU_KEY_ON			_MMIO32(IOBASE | 0x1d88)
+#define SPU_KEY_OFF			_MMIO32(IOBASE | 0x1d8c)
+#define SPU_FM_MODE			_MMIO32(IOBASE | 0x1d90)
+#define SPU_NOISE_MODE		_MMIO32(IOBASE | 0x1d94)
+#define SPU_REVERB_ON		_MMIO32(IOBASE | 0x1d98)
+#define SPU_CHAN_STATUS		_MMIO32(IOBASE | 0x1d9c)
+
+#define SPU_REVERB_ADDR		_MMIO16(IOBASE | 0x1da2)
+#define SPU_IRQ_ADDR		_MMIO16(IOBASE | 0x1da4)
+#define SPU_ADDR			_MMIO16(IOBASE | 0x1da6)
+#define SPU_DATA			_MMIO16(IOBASE | 0x1da8)
+
+#define SPU_CTRL			_MMIO16(IOBASE | 0x1daa)
+#define SPU_DMA_CTRL		_MMIO16(IOBASE | 0x1dac)
+#define SPU_STAT			_MMIO16(IOBASE | 0x1dae)
+
+#define SPU_CD_VOL_L		_MMIO16(IOBASE | 0x1db0)
+#define SPU_CD_VOL_R		_MMIO16(IOBASE | 0x1db2)
+#define SPU_EXT_VOL_L		_MMIO16(IOBASE | 0x1db4)
+#define SPU_EXT_VOL_R		_MMIO16(IOBASE | 0x1db6)
+#define SPU_CURRENT_VOL_L	_MMIO16(IOBASE | 0x1db8)
+#define SPU_CURRENT_VOL_R	_MMIO16(IOBASE | 0x1dba)
 
 // These are not named SPU_VOICE_* to avoid name clashes with SPU attribute
 // flags defined in psxspu.h.
-#define SPU_CH_VOL_L(N)		_MMIO16(0x1f801c00 + 16 * (N))
-#define SPU_CH_VOL_R(N)		_MMIO16(0x1f801c02 + 16 * (N))
-#define SPU_CH_FREQ(N)		_MMIO16(0x1f801c04 + 16 * (N))
-#define SPU_CH_ADDR(N)		_MMIO16(0x1f801c06 + 16 * (N))
-#define SPU_CH_ADSR(N)		_MMIO32(0x1f801c08 + 16 * (N))
-#define SPU_CH_LOOP_ADDR(N)	_MMIO16(0x1f801c0e + 16 * (N))
+#define SPU_CH_VOL_L(N)		_MMIO16(IOBASE | 0x1c00 + 16 * (N))
+#define SPU_CH_VOL_R(N)		_MMIO16(IOBASE | 0x1c02 + 16 * (N))
+#define SPU_CH_FREQ(N)		_MMIO16(IOBASE | 0x1c04 + 16 * (N))
+#define SPU_CH_ADDR(N)		_MMIO16(IOBASE | 0x1c06 + 16 * (N))
+#define SPU_CH_ADSR(N)		_MMIO32(IOBASE | 0x1c08 + 16 * (N))
+#define SPU_CH_LOOP_ADDR(N)	_MMIO16(IOBASE | 0x1c0e + 16 * (N))
 
 /* MDEC */
 
-#define MDEC0				_MMIO32(0x1f801820)
-#define MDEC1				_MMIO32(0x1f801824)
+#define MDEC0				_MMIO32(IOBASE | 0x1820)
+#define MDEC1				_MMIO32(IOBASE | 0x1824)
 
 /* SPI controller port */
 
 // IMPORTANT: even though JOY_TXRX is a 32-bit register, it should only be
 // accessed as 8-bit. Reading it as 16 or 32-bit works fine on real hardware,
 // but leads to problems in some emulators.
-#define JOY_TXRX			_MMIO8(0x1f801040)
-#define JOY_STAT			_MMIO16(0x1f801044)
-#define JOY_MODE			_MMIO16(0x1f801048)
-#define JOY_CTRL			_MMIO16(0x1f80104a)
-#define JOY_BAUD			_MMIO16(0x1f80104e)
+#define JOY_TXRX			_MMIO8 (IOBASE | 0x1040)
+#define JOY_STAT			_MMIO16(IOBASE | 0x1044)
+#define JOY_MODE			_MMIO16(IOBASE | 0x1048)
+#define JOY_CTRL			_MMIO16(IOBASE | 0x104a)
+#define JOY_BAUD			_MMIO16(IOBASE | 0x104e)
 
 /* Serial port */
 
-#define SIO_TXRX			_MMIO8(0x1f801050)
-#define SIO_STAT			_MMIO16(0x1f801054)
-#define SIO_MODE			_MMIO16(0x1f801058)
-#define SIO_CTRL			_MMIO16(0x1f80105a)
-#define SIO_BAUD			_MMIO16(0x1f80105e)
+#define SIO_TXRX			_MMIO8 (IOBASE | 0x1050)
+#define SIO_STAT			_MMIO16(IOBASE | 0x1054)
+#define SIO_MODE			_MMIO16(IOBASE | 0x1058)
+#define SIO_CTRL			_MMIO16(IOBASE | 0x105a)
+#define SIO_BAUD			_MMIO16(IOBASE | 0x105e)
 
 /* IRQ controller */
 
-#define IRQ_STAT			_MMIO32(0x1f801070)
-#define IRQ_MASK			_MMIO32(0x1f801074)
+#define IRQ_STAT			_MMIO16(IOBASE | 0x1070)
+#define IRQ_MASK			_MMIO16(IOBASE | 0x1074)
 
 /* DMA */
 
-#define DMA_DPCR			_MMIO32(0x1f8010f0)
-#define DMA_DICR			_MMIO32(0x1f8010f4)
+#define DMA_DPCR			_MMIO32(IOBASE | 0x10f0)
+#define DMA_DICR			_MMIO32(IOBASE | 0x10f4)
 
-#define DMA_MADR(N)			_MMIO32(0x1f801080 + 16 * (N))
-#define DMA_BCR(N)			_MMIO32(0x1f801084 + 16 * (N))
-#define DMA_CHCR(N)			_MMIO32(0x1f801088 + 16 * (N))
+#define DMA_MADR(N)			_MMIO32(IOBASE | 0x1080 + 16 * (N))
+#define DMA_BCR(N)			_MMIO32(IOBASE | 0x1084 + 16 * (N))
+#define DMA_CHCR(N)			_MMIO32(IOBASE | 0x1088 + 16 * (N))
 
 /* Timers */
 
-#define TIMER_VALUE(N)		_MMIO32(0x1f801100 + 16 * (N))
-#define TIMER_CTRL(N)		_MMIO32(0x1f801104 + 16 * (N))
-#define TIMER_RELOAD(N)		_MMIO32(0x1f801108 + 16 * (N))
+#define TIMER_VALUE(N)		_MMIO32(IOBASE | 0x1100 + 16 * (N))
+#define TIMER_CTRL(N)		_MMIO32(IOBASE | 0x1104 + 16 * (N))
+#define TIMER_RELOAD(N)		_MMIO32(IOBASE | 0x1108 + 16 * (N))
 
 /* Memory control */
 
-#define EXP1_ADDR			_MMIO32(0x1f801000)
-#define EXP2_ADDR			_MMIO32(0x1f801004)
-#define EXP1_DELAY_SIZE		_MMIO32(0x1f801008)
-#define EXP3_DELAY_SIZE		_MMIO32(0x1f80100c)
-#define BIOS_DELAY_SIZE		_MMIO32(0x1f801010)
-#define SPU_DELAY_SIZE		_MMIO32(0x1f801014)
-#define CD_DELAY_SIZE		_MMIO32(0x1f801018)
-#define EXP2_DELAY_SIZE		_MMIO32(0x1f80101c)
-#define COM_DELAY_CFG		_MMIO32(0x1f801020)
-#define RAM_SIZE_CFG		_MMIO32(0x1f801060)
+#define EXP1_ADDR			_MMIO32(IOBASE | 0x1000)
+#define EXP2_ADDR			_MMIO32(IOBASE | 0x1004)
+#define EXP1_DELAY_SIZE		_MMIO32(IOBASE | 0x1008)
+#define EXP3_DELAY_SIZE		_MMIO32(IOBASE | 0x100c)
+#define BIOS_DELAY_SIZE		_MMIO32(IOBASE | 0x1010)
+#define SPU_DELAY_SIZE		_MMIO32(IOBASE | 0x1014)
+#define CD_DELAY_SIZE		_MMIO32(IOBASE | 0x1018)
+#define EXP2_DELAY_SIZE		_MMIO32(IOBASE | 0x101c)
+#define COM_DELAY_CFG		_MMIO32(IOBASE | 0x1020)
+#define RAM_SIZE_CFG		_MMIO32(IOBASE | 0x1060)
 
 #endif
diff --git a/libpsn00b/include/inline_c.h b/libpsn00b/include/inline_c.h
index 177faf1..c5eaa59 100644
--- a/libpsn00b/include/inline_c.h
+++ b/libpsn00b/include/inline_c.h
@@ -1,32 +1,23 @@
-/* Inline GTE macros for the GNU C compiler.
- *
- * Part of the PSn00bSDK Project by Lameguy64.
- * 2019 Meido-Tek Production
- *
- * All GTE commands can be used without having to pass your object file
- * through some stupid tool such as DMPSX. Perhaps it was Sony's attempt
- * to prevent people from quickly discovering the GTE commands from the 
- * official SDK easily? Though people could just extract the cop2 opcodes
- * of an object file after it has been passed through DMPSX.
- *
- * Todo: A couple of GTE operation macros are still missing such as
- *  gte_rtv*() though they appear to be just variants of gte_mvmva more or
- *  less (gte_rtv0() is actually gte_mvmva(1, 0, 0, 3, 0) for example).
+/*
+ * PSn00bSDK GTE macros
+ * (C) 2019 Lameguy64
+ * (C) 2021-2022 Soapy (tweaked by spicyjpeg)
  *
+ * This header is basically identical to Nugget's inline_n.h. All GTE commands
+ * can be used right away without having to run DMPSX or any other tool on
+ * object files.
  */
 
 #ifndef _INLINE_C_H
 #define _INLINE_C_H
 
-/*
- *	GTE load macros
- */
+/* GTE load macros */
 
 /* Load a SVECTOR (passed as a pointer) to GTE V0
  */
 #define gte_ldv0( r0 ) __asm__ volatile ( \
-	"lwc2	$0 , 0( %0 );"	\
-	"lwc2	$1 , 4( %0 );"	\
+	"lwc2	$0, 0( %0 );"	\
+	"lwc2	$1, 4( %0 );"	\
 	:						\
 	: "r"( r0 )				\
 	: "$t0" )
@@ -34,8 +25,8 @@
 /* Load a SVECTOR (passed as a pointer) to GTE V1
  */
 #define gte_ldv1( r0 ) __asm__ volatile ( \
-	"lwc2	$2 , 0( %0 );"	\
-	"lwc2	$3 , 4( %0 );"	\
+	"lwc2	$2, 0( %0 );"	\
+	"lwc2	$3, 4( %0 );"	\
 	:						\
 	: "r"( r0 )				\
 	: "$t0" )
@@ -43,8 +34,8 @@
 /* Load a SVECTOR (passed as a pointer) to GTE V2
  */
 #define gte_ldv2( r0 ) __asm__ volatile ( \
-	"lwc2	$4 , 0( %0 );"	\
-	"lwc2	$5 , 4( %0 );"	\
+	"lwc2	$4, 0( %0 );"	\
+	"lwc2	$5, 4( %0 );"	\
 	:						\
 	: "r"( r0 )				\
 	: "$t0" )
@@ -52,20 +43,187 @@
 /* Load three SVECTORs (passed as a pointer) to the GTE at once
  */
 #define gte_ldv3( r0, r1, r2 ) __asm__ volatile ( \
-	"lwc2	$0 , 0( %0 );"	\
-	"lwc2	$1 , 4( %0 );"	\
-	"lwc2	$2 , 0( %1 );"	\
-	"lwc2	$3 , 4( %1 );"	\
-	"lwc2	$4 , 0( %2 );"	\
-	"lwc2	$5 , 4( %2 );"	\
+	"lwc2	$0, 0( %0 );"	\
+	"lwc2	$1, 4( %0 );"	\
+	"lwc2	$2, 0( %1 );"	\
+	"lwc2	$3, 4( %1 );"	\
+	"lwc2	$4, 0( %2 );"	\
+	"lwc2	$5, 4( %2 );"	\
 	:						\
 	: "r"( r0 ), "r"( r1 ), "r"( r2 ) )
-	
+
+#define gte_ldv3c( r0 ) __asm__ volatile ( \
+	"lwc2	$0, 0( %0 );"	\
+	"lwc2	$1, 4( %0 );"	\
+	"lwc2	$2, 8( %0 );"	\
+	"lwc2	$3, 12( %0 );"	\
+	"lwc2	$4, 16( %0 );"	\
+	"lwc2	$5, 20( %0 );"	\
+	:						\
+	: "r"( r0 ) )
+
+#define gte_ldv3c_vertc( r0 ) __asm__ volatile ( \
+	"lwc2	$0, 0( %0 );"	\
+	"lwc2	$1, 4( %0 );"	\
+	"lwc2	$2, 12( %0 );"	\
+	"lwc2	$3, 16( %0 );"	\
+	"lwc2	$4, 24( %0 );"	\
+	"lwc2	$5, 28( %0 );"	\
+	:						\
+	: "r"( r0 ) )
+
+#define gte_ldv01( r0, r1 ) __asm__ volatile ( \
+	"lwc2	$0, 0( %0 );"	\
+	"lwc2	$1, 4( %0 );"	\
+	"lwc2	$2, 0( %1 );"	\
+	"lwc2	$3, 4( %1 );"	\
+	:						\
+	: "r"( r0 ), "r"( r1 ) )
+
+#define gte_ldv01c( r0 ) __asm__ volatile ( \
+	"lwc2	$0, 0( %0 );"	\
+	"lwc2	$1, 4( %0 );"	\
+	"lwc2	$2, 8( %0 );"	\
+	"lwc2	$3, 12( %0 );"	\
+	:						\
+	: "r"( r0 ) )
+
 #define gte_ldrgb( r0 ) __asm__ volatile ( \
 	"lwc2	$6 , 0( %0 );"	\
 	:						\
 	: "r"( r0 ) )
-	
+
+#define gte_ldrgb3( r0, r1, r2 ) __asm__ volatile ( \
+	"lwc2	$20, 0( %0 );"	\
+	"lwc2	$21, 0( %1 );"	\
+	"lwc2	$22, 0( %2 );"	\
+	"lwc2	$6, 0( %2 );"	\
+	:						\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 ) )
+
+#define gte_ldrgb3c( r0 ) __asm__ volatile ( \
+	"lwc2	$20, 0( %0 );"	\
+	"lwc2	$21, 4( %0 );"	\
+	"lwc2	$22, 8( %0 );"	\
+	"lwc2	$6, 8( %0 );"	\
+	:						\
+	: "r"( r0 ) )
+
+#define gte_ldlv0( r0 ) __asm__ volatile ( \
+	"lhu	$13, 4( %0 );"	\
+	"lhu	$12, 0( %0 );"	\
+	"sll	$13, $13, 16;"	\
+	"or	$12, $12, $13;"		\
+	"mtc2	$12, $0;"		\
+	"lwc2	$1, 8( %0 );"	\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13" )
+
+#define gte_ldlvl( r0 ) __asm__ volatile ( \
+	"lwc2	$9, 0( %0 );"	\
+	"lwc2	$10, 4( %0 );"	\
+	"lwc2	$11, 8( %0 );"	\
+	:						\
+	: "r"( r0 ) )
+
+#define gte_ldsv( r0 ) __asm__ volatile ( \
+	"lhu	$12, 0( %0 );"	\
+	"lhu	$13, 2( %0 );"	\
+	"lhu	$14, 4( %0 );"	\
+	"mtc2	$12, $9;"		\
+	"mtc2	$13, $10;"		\
+	"mtc2	$14, $11;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14" )
+
+#define gte_ldbv( r0 ) __asm__ volatile ( \
+	"lbu	$12, 0( %0 );"	\
+	"lbu	$13, 1( %0 );"	\
+	"mtc2	$12, $9;"		\
+	"mtc2	$13, $10;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13" )
+
+#define gte_ldcv( r0 ) __asm__ volatile ( \
+	"lbu	$12, 0( %0 );"	\
+	"lbu	$13, 1( %0 );"	\
+	"lbu	$14, 2( %0 );"	\
+	"mtc2	$12, $9;"		\
+	"mtc2	$13, $10;"		\
+	"mtc2	$14, $11;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14" )
+
+#define gte_ldclmv( r0 ) __asm__ volatile ( \
+	"lhu	$12, 0( %0 );"	\
+	"lhu	$13, 6( %0 );"	\
+	"lhu	$14, 12( %0 );"	\
+	"mtc2	$12, $9;"		\
+	"mtc2	$13, $10;"		\
+	"mtc2	$14, $11;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14" )
+
+#define gte_ldsxy0( r0 ) __asm__ volatile ( \
+	"mtc2	%0, $12;"		\
+	:						\
+	: "r"( r0 ) )
+
+#define gte_ldsxy1( r0 ) __asm__ volatile ( \
+	"mtc2	%0, $13;"		\
+	:						\
+	: "r"( r0 ) )
+
+#define gte_ldsxy2( r0 ) __asm__ volatile ( \
+	"mtc2	%0, $14;"		\
+	:						\
+	: "r"( r0 ) )
+
+#define gte_ldsxy3( r0, r1, r2 ) __asm__ volatile ( \
+	"mtc2	%0, $12;"		\
+	"mtc2	%2, $14;"		\
+	"mtc2	%1, $13;"		\
+	:						\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 ) )
+
+#define gte_ldsxy3c( r0 ) __asm__ volatile ( \
+	"lwc2	$12, 0( %0 );"	\
+	"lwc2	$13, 4( %0 );"	\
+	"lwc2	$14, 8( %0 );"	\
+	:						\
+	: "r"( r0 ) )
+
+#define gte_ldsz3( r0, r1, r2 ) __asm__ volatile ( \
+	"mtc2	%0, $17;"		\
+	"mtc2	%1, $18;"		\
+	"mtc2	%2, $19;"		\
+	:						\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 ) )
+
+#define gte_ldsz4( r0, r1, r2, r3 ) __asm__ volatile ( \
+	"mtc2	%0, $16;"		\
+	"mtc2	%1, $17;"		\
+	"mtc2	%2, $18;"		\
+	"mtc2	%3, $19;"		\
+	:						\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 ), "r"( r3 ) )
+
+#define gte_ldopv1( r0 ) __asm__ volatile ( \
+	"lw	$12, 0( %0 );"		\
+	"lw	$13, 4( %0 );"		\
+	"ctc2	$12, $0;"		\
+	"lw	$14, 8( %0 );"		\
+	"ctc2	$13, $2;"		\
+	"ctc2	$14, $4;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14" )
+
 #define gte_ldopv2( r0 ) __asm__ volatile ( \
 	"lwc2	$11, 8( %0 );"	\
 	"lwc2	$9 , 0( %0 );"	\
@@ -77,9 +235,53 @@
 	"mtc2	%0, $8;"		\
 	:						\
 	: "r"( r0 ) )
-	
-/* Sets the GTE offset
- */
+
+#define gte_ldlzc( r0 ) __asm__ volatile ( \
+	"mtc2	%0, $30;"		\
+	:						\
+	: "r"( r0 ) )
+
+#define gte_SetRGBcd( r0 ) __asm__ volatile ( \
+	"lwc2	$6, 0( %0 );"	\
+	:						\
+	: "r"( r0 ) )
+
+#define gte_ldbkdir( r0, r1, r2 ) __asm__ volatile ( \
+	"ctc2	%0, $13;"		\
+	"ctc2	%1, $14;"		\
+	"ctc2	%2, $15;"		\
+	:						\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 ) )
+
+#define gte_SetBackColor( r0, r1, r2 ) __asm__ volatile ( \
+	"sll	$t0, %0, 4;"	\
+	"sll	$t1, %1, 4;"	\
+	"sll	$t2, %2, 4;"	\
+	"ctc2	$t0, $13;"		\
+	"ctc2	$t1, $14;"		\
+	"ctc2	$t2, $15;"		\
+	:						\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 )	\
+	: "$t0", "$t1", "$t2" )
+
+#define gte_ldfcdir( r0, r1, r2 ) __asm__ volatile ( \
+	"ctc2	%0, $21;"		\
+	"ctc2	%1, $22;"		\
+	"ctc2	%2, $23;"		\
+	:						\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 ) )
+
+#define gte_SetFarColor( r0, r1, r2 ) __asm__ volatile ( \
+	"sll	$12, %0, 4;"	\
+	"sll	$13, %1, 4;"	\
+	"sll	$14, %2, 4;"	\
+	"ctc2	$12, $21;"		\
+	"ctc2	$13, $22;"		\
+	"ctc2	$14, $23;"		\
+	:						\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 )	\
+	: "$12", "$13", "$14" )
+
 #define gte_SetGeomOffset( r0, r1 ) __asm__ volatile ( \
 	"sll	$t0, %0, 16;"	\
 	"sll	$t1, %1, 16;"	\
@@ -88,23 +290,21 @@
 	:						\
 	: "r"( r0 ), "r"( r1 )	\
 	: "$t0", "$t1" )
-	
+
 #define gte_SetGeomScreen( r0 ) __asm__ volatile ( \
 	"ctc2	%0, $26;"		\
 	:						\
 	: "r"( r0 ) )
 
-#define gte_SetTransMatrix( r0 ) __asm__ volatile ( \
-	"lw		$t0, 20( %0 );"	\
-	"lw		$t1, 24( %0 );"	\
-	"ctc2	$t0, $5;"		\
-	"lw		$t2, 28( %0 );"	\
-	"ctc2	$t1, $6;"		\
-	"ctc2	$t2, $7;"		\
+#define gte_ldsvrtrow0( r0 ) __asm__ volatile ( \
+	"lw	$12, 0( %0 );"		\
+	"lw	$13, 4( %0 );"		\
+	"ctc2	$12, $0;"		\
+	"ctc2	$13, $1;"		\
 	:						\
 	: "r"( r0 )				\
-	: "$t2" )
-	
+	: "$12", "$13" )
+
 #define gte_SetRotMatrix( r0 ) __asm__ volatile ( \
 	"lw		$t0, 0( %0 );"	\
 	"lw		$t1, 4( %0 );"	\
@@ -120,6 +320,15 @@
 	: "r"( r0 )				\
 	: "$t2" )
 
+#define gte_ldsvllrow0( r0 ) __asm__ volatile ( \
+	"lw	$12, 0( %0 );"		\
+	"lw	$13, 4( %0 );"		\
+	"ctc2	$12, $8;"		\
+	"ctc2	$13, $9;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13" )
+
 #define gte_SetLightMatrix( r0 ) __asm__ volatile ( \
 	"lw		$t0, 0( %0 );"	\
 	"lw		$t1, 4( %0 );"	\
@@ -134,7 +343,16 @@
 	:						\
 	: "r"( r0 )				\
 	: "$t2" )
-	
+
+#define gte_ldsvlcrow0( r0 ) __asm__ volatile ( \
+	"lw	$12, 0( %0 );"		\
+	"lw	$13, 4( %0 );"		\
+	"ctc2	$12, $16;"		\
+	"ctc2	$13, $17;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13" )
+
 #define gte_SetColorMatrix( r0 ) __asm__ volatile ( \
 	"lw		$t0, 0( %0 );"	\
 	"lw		$t1, 4( %0 );"	\
@@ -149,44 +367,153 @@
 	:						\
 	: "r"( r0 )				\
 	: "$t2" )
-	
-#define gte_SetBackColor( r0, r1, r2 ) __asm__ volatile ( \
-	"sll	$t0, %0, 4;"	\
-	"sll	$t1, %1, 4;"	\
-	"sll	$t2, %2, 4;"	\
-	"ctc2	$t0, $13;"		\
-	"ctc2	$t1, $14;"		\
-	"ctc2	$t2, $15;"		\
+
+#define gte_SetTransMatrix( r0 ) __asm__ volatile ( \
+	"lw		$t0, 20( %0 );"	\
+	"lw		$t1, 24( %0 );"	\
+	"ctc2	$t0, $5;"		\
+	"lw		$t2, 28( %0 );"	\
+	"ctc2	$t1, $6;"		\
+	"ctc2	$t2, $7;"		\
 	:						\
-	: "r"( r0 ), "r"( r1 ), "r"( r2 )	\
-	: "$t0", "$t1", "$t2" )
-	
-/*
- *	GTE store macros
- */
-	
-#define gte_otz( r0 ) __asm__ volatile ( \
-	"swc2	$7, 0( %0 );"	\
+	: "r"( r0 )				\
+	: "$t2" )
+
+#define gte_ldtr( r0, r1, r2 ) __asm__ volatile ( \
+	"ctc2	%0, $5;"		\
+	"ctc2	%1, $6;"		\
+	"ctc2	%2, $7;"		\
+	:						\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 ) )
+
+#define gte_SetTransVector( r0 ) __asm__ volatile ( \
+	"lw	$12, 0( %0 );"		\
+	"lw	$13, 4( %0 );"		\
+	"lw	$14, 8( %0 );"		\
+	"ctc2	$12, $5;"		\
+	"ctc2	$13, $6;"		\
+	"ctc2	$14, $7;"		\
 	:						\
 	: "r"( r0 )				\
-	: "memory" )
-	
-#define gte_stflg( r0 ) __asm__ volatile ( \
-	"cfc2	$t0, $31;"		\
-	"nop;"					\
-	"sw		$t0, 0( %0 );"	\
+	: "$12", "$13", "$14" )
+
+#define gte_ld_intpol_uv0( r0 ) __asm__ volatile ( \
+	"lbu	$12, 0( %0 );"	\
+	"lbu	$13, 1( %0 );"	\
+	"ctc2	$12, $21;"		\
+	"ctc2	$13, $22;"		\
 	:						\
 	: "r"( r0 )				\
-	: "memory" )
-	
+	: "$12", "$13" )
+
+#define gte_ld_intpol_uv1( r0 ) __asm__ volatile ( \
+	"lbu	$12, 0( %0 );"	\
+	"lbu	$13, 1( %0 );"	\
+	"mtc2	$12, $9;"		\
+	"mtc2	$13, $10;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13" )
+
+#define gte_ld_intpol_bv0( r0 ) __asm__ volatile ( \
+	"lbu	$12, 0( %0 );"	\
+	"lbu	$13, 1( %0 );"	\
+	"ctc2	$12, $21;"		\
+	"ctc2	$13, $22;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13" )
+
+#define gte_ld_intpol_bv1( r0 ) __asm__ volatile ( \
+	"lbu	$12, 0( %0 );"	\
+	"lbu	$13, 1( %0 );"	\
+	"mtc2	$12, $9;"		\
+	"mtc2	$13, $10;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13" )
+
+#define gte_ld_intpol_sv0( r0 ) __asm__ volatile ( \
+	"lh	$12, 0( %0 );"		\
+	"lh	$13, 2( %0 );"		\
+	"lh	$14, 4( %0 );"		\
+	"ctc2	$12, $21;"		\
+	"ctc2	$13, $22;"		\
+	"ctc2	$14, $23;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14" )
+
+#define gte_ld_intpol_sv1( r0 ) __asm__ volatile ( \
+	"lh	$12, 0( %0 );"		\
+	"lh	$13, 2( %0 );"		\
+	"lh	$14, 4( %0 );"		\
+	"mtc2	$12, $9;"		\
+	"mtc2	$13, $10;"		\
+	"mtc2	$14, $11;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14" )
+
+#define gte_ldfc( r0 ) __asm__ volatile ( \
+	"lw	$12, 0( %0 );"		\
+	"lw	$13, 4( %0 );"		\
+	"lw	$14, 8( %0 );"		\
+	"ctc2	$12, $21;"		\
+	"ctc2	$13, $22;"		\
+	"ctc2	$14, $23;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14" )
+
+#define gte_ldopv2SV( r0 ) __asm__ volatile ( \
+	"lh	$12, 0( %0 );"		\
+	"lh	$13, 2( %0 );"		\
+	"lh	$14, 4( %0 );"		\
+	"mtc2	$12, $9;"		\
+	"mtc2	$13, $10;"		\
+	"mtc2	$14, $11;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14" )
+
+#define gte_ldopv1SV( r0 ) __asm__ volatile ( \
+	"lh	$12, 0( %0 );"		\
+	"lh	$13, 2( %0 );"		\
+	"ctc2	$12, $0;"		\
+	"lh	$14, 4( %0 );"		\
+	"ctc2	$13, $2;"		\
+	"ctc2	$14, $4;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14" )
+
+/* GTE store macros */
+
 #define gte_stsxy( r0 ) __asm__ volatile ( \
 	"swc2	$14, 0( %0 );"	\
 	:						\
 	: "r"( r0 )				\
 	: "memory" )
-	
-#define gte_stsxy0( r0 ) __asm__ volatile ( \
+
+#define gte_stsxy3( r0, r1, r2 ) __asm__ volatile ( \
 	"swc2	$12, 0( %0 );"	\
+	"swc2	$13, 0( %1 );"	\
+	"swc2	$14, 0( %2 );"	\
+	:						\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 )	\
+	: "memory" )
+
+#define gte_stsxy3c( r0 ) __asm__ volatile ( \
+	"swc2	$12, 0( %0 );"	\
+	"swc2	$13, 4( %0 );"	\
+	"swc2	$14, 8( %0 );"	\
+	:						\
+	: "r"( r0 )				\
+	: "memory" )
+
+#define gte_stsxy2( r0 ) __asm__ volatile ( \
+	"swc2	$14, 0( %0 );"	\
 	:						\
 	: "r"( r0 )				\
 	: "memory" )
@@ -197,140 +524,526 @@
 	: "r"( r0 )				\
 	: "memory" )
 
-#define gte_stsxy2( r0 ) __asm__ volatile ( \
-	"swc2	$14, 0( %0 );"	\
+#define gte_stsxy0( r0 ) __asm__ volatile ( \
+	"swc2	$12, 0( %0 );"	\
 	:						\
 	: "r"( r0 )				\
 	: "memory" )
 
-#define gte_stsxy3( r0, r1, r2 ) __asm__ volatile ( \
+#define gte_stsxy01( r0, r1 ) __asm__ volatile ( \
 	"swc2	$12, 0( %0 );"	\
 	"swc2	$13, 0( %1 );"	\
-	"swc2	$14, 0( %2 );"	\
 	:						\
-	: "r"( r0 ), "r"( r1 ), "r"( r2 ) \
+	: "r"( r0 ), "r"( r1 )	\
+	: "memory" )
+
+#define gte_stsxy01c( r0 ) __asm__ volatile ( \
+	"swc2	$12, 0( %0 );"	\
+	"swc2	$13, 4( %0 );"	\
+	:						\
+	: "r"( r0 )				\
 	: "memory" )
 
-#define gte_stsxy3_f3( r0 ) __asm__ volatile (	\
+#define gte_stsxy3_f3( r0 ) __asm__ volatile ( \
 	"swc2	$12, 8( %0 );"	\
 	"swc2	$13, 12( %0 );"	\
-	"swc2	$14, 16( %0 )"	\
+	"swc2	$14, 16( %0 );"	\
 	:						\
-	: "r"( r0 )	\
+	: "r"( r0 )				\
 	: "memory" )
 
-#define gte_stsxy3_g3( r0 ) __asm__ volatile (	\
+#define gte_stsxy3_g3( r0 ) __asm__ volatile ( \
 	"swc2	$12, 8( %0 );"	\
 	"swc2	$13, 16( %0 );"	\
-	"swc2	$14, 24( %0 )"	\
+	"swc2	$14, 24( %0 );"	\
 	:						\
-	: "r"( r0 )	\
+	: "r"( r0 )				\
 	: "memory" )
 
-#define gte_stsxy3_ft3( r0 ) __asm__ volatile (	\
+#define gte_stsxy3_ft3( r0 ) __asm__ volatile ( \
 	"swc2	$12, 8( %0 );"	\
 	"swc2	$13, 16( %0 );"	\
-	"swc2	$14, 24( %0 )"	\
+	"swc2	$14, 24( %0 );"	\
 	:						\
-	: "r"( r0 )	\
+	: "r"( r0 )				\
 	: "memory" )
 
-#define gte_stsxy3_gt3( r0 ) __asm__ volatile (	\
+#define gte_stsxy3_gt3( r0 ) __asm__ volatile ( \
 	"swc2	$12, 8( %0 );"	\
 	"swc2	$13, 20( %0 );"	\
-	"swc2	$14, 32( %0 )"	\
+	"swc2	$14, 32( %0 );"	\
 	:						\
-	: "r"( r0 )	\
+	: "r"( r0 )				\
 	: "memory" )
 
-#define gte_stsxy3_f4( r0 ) __asm__ volatile (	\
+#define gte_stsxy3_f4( r0 ) __asm__ volatile ( \
 	"swc2	$12, 8( %0 );"	\
 	"swc2	$13, 12( %0 );"	\
-	"swc2	$14, 16( %0 )"	\
+	"swc2	$14, 16( %0 );"	\
 	:						\
-	: "r"( r0 )	\
+	: "r"( r0 )				\
 	: "memory" )
 
-#define gte_stsxy3_g4( r0 ) __asm__ volatile (	\
+#define gte_stsxy3_g4( r0 ) __asm__ volatile ( \
 	"swc2	$12, 8( %0 );"	\
 	"swc2	$13, 16( %0 );"	\
-	"swc2	$14, 24( %0 )"	\
+	"swc2	$14, 24( %0 );"	\
 	:						\
-	: "r"( r0 )	\
+	: "r"( r0 )				\
 	: "memory" )
 
-#define gte_stsxy3_ft4( r0 ) __asm__ volatile (	\
+#define gte_stsxy3_ft4( r0 ) __asm__ volatile ( \
 	"swc2	$12, 8( %0 );"	\
 	"swc2	$13, 16( %0 );"	\
-	"swc2	$14, 24( %0 )"	\
+	"swc2	$14, 24( %0 );"	\
 	:						\
-	: "r"( r0 )	\
+	: "r"( r0 )				\
 	: "memory" )
 
-#define gte_stsxy3_gt4( r0 ) __asm__ volatile (	\
+#define gte_stsxy3_gt4( r0 ) __asm__ volatile ( \
 	"swc2	$12, 8( %0 );"	\
 	"swc2	$13, 20( %0 );"	\
-	"swc2	$14, 32( %0 )"	\
+	"swc2	$14, 32( %0 );"	\
 	:						\
-	: "r"( r0 )	\
+	: "r"( r0 )				\
 	: "memory" )
 
+#define gte_stdp( r0 ) __asm__ volatile ( \
+	"swc2	$8, 0( %0 );"	\
+	:						\
+	: "r"( r0 )				\
+	: "memory" )
+
+#define gte_stflg( r0 ) __asm__ volatile ( \
+	"cfc2	$12, $31;"		\
+	"nop;"					\
+	"sw	$12, 0( %0 );"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "memory" )
+
+#define gte_stflg_4( r0 ) __asm__ volatile ( \
+	"cfc2	$12, $31;"		\
+	"addi	$13, $0, 4;"	\
+	"sll	$13, $13, 16;"	\
+	"and	$12, $12, $13;"	\
+	"sw	$12, 0( %0 );"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "memory" )
+
 #define gte_stsz( r0 ) __asm__ volatile ( \
 	"swc2	$19, 0( %0 );"	\
 	:						\
 	: "r"( r0 )				\
 	: "memory" )
-	
+
+#define gte_stsz3( r0, r1, r2 ) __asm__ volatile ( \
+	"swc2	$17, 0( %0 );"	\
+	"swc2	$18, 0( %1 );"	\
+	"swc2	$19, 0( %2 );"	\
+	:						\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 )	\
+	: "memory" )
+
+#define gte_stsz4( r0, r1, r2, r3 ) __asm__ volatile ( \
+	"swc2	$16, 0( %0 );"	\
+	"swc2	$17, 0( %1 );"	\
+	"swc2	$18, 0( %2 );"	\
+	"swc2	$19, 0( %3 );"	\
+	:						\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 ), "r"( r3 )	\
+	: "memory" )
+
+#define gte_stsz3c( r0 ) __asm__ volatile ( \
+	"swc2	$17, 0( %0 );"	\
+	"swc2	$18, 4( %0 );"	\
+	"swc2	$19, 8( %0 );"	\
+	:						\
+	: "r"( r0 )				\
+	: "memory" )
+
+#define gte_stsz4c( r0 ) __asm__ volatile ( \
+	"swc2	$16, 0( %0 );"	\
+	"swc2	$17, 4( %0 );"	\
+	"swc2	$18, 8( %0 );"	\
+	"swc2	$19, 12( %0 );"	\
+	:						\
+	: "r"( r0 )				\
+	: "memory" )
+
+#define gte_stszotz( r0 ) __asm__ volatile ( \
+	"mfc2	$12, $19;"		\
+	"nop;"					\
+	"sra	$12, $12, 2;"	\
+	"sw	$12, 0( %0 );"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "memory" )
+
 #define gte_stotz( r0 ) __asm__ volatile ( \
 	"swc2	$7, 0( %0 );"	\
 	:						\
-	: "r"( r0 ) 			\
+	: "r"( r0 )				\
 	: "memory" )
-	
+
 #define gte_stopz( r0 ) __asm__ volatile ( \
 	"swc2	$24, 0( %0 );"	\
 	:						\
 	: "r"( r0 )				\
 	: "memory" )
-	
+
+#define gte_stlvl( r0 ) __asm__ volatile ( \
+	"swc2	$9, 0( %0 );"	\
+	"swc2	$10, 4( %0 );"	\
+	"swc2	$11, 8( %0 );"	\
+	:						\
+	: "r"( r0 )				\
+	: "memory" )
+
+#define gte_stlvnl( r0 ) __asm__ volatile ( \
+	"swc2	$25, 0( %0 );"	\
+	"swc2	$26, 4( %0 );"	\
+	"swc2	$27, 8( %0 );"	\
+	:						\
+	: "r"( r0 )				\
+	: "memory" )
+
+#define gte_stlvnl0( r0 ) __asm__ volatile ( \
+	"swc2	$25, 0( %0 );"	\
+	:						\
+	: "r"( r0 )				\
+	: "memory" )
+
+#define gte_stlvnl1( r0 ) __asm__ volatile ( \
+	"swc2	$26, 0( %0 );"	\
+	:						\
+	: "r"( r0 )				\
+	: "memory" )
+
+#define gte_stlvnl2( r0 ) __asm__ volatile ( \
+	"swc2	$27, 0( %0 );"	\
+	:						\
+	: "r"( r0 )				\
+	: "memory" )
+
+#define gte_stsv( r0 ) __asm__ volatile ( \
+	"mfc2	$12, $9;"		\
+	"mfc2	$13, $10;"		\
+	"mfc2	$14, $11;"		\
+	"sh	$12, 0( %0 );"		\
+	"sh	$13, 2( %0 );"		\
+	"sh	$14, 4( %0 );"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14", "memory" )
+
+#define gte_stclmv( r0 ) __asm__ volatile ( \
+	"mfc2	$12, $9;"		\
+	"mfc2	$13, $10;"		\
+	"mfc2	$14, $11;"		\
+	"sh	$12, 0( %0 );"		\
+	"sh	$13, 6( %0 );"		\
+	"sh	$14, 12( %0 );"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14", "memory" )
+
+#define gte_stbv( r0 ) __asm__ volatile ( \
+	"mfc2	$12, $9;"		\
+	"mfc2	$13, $10;"		\
+	"sb	$12, 0( %0 );"		\
+	"sb	$13, 1( %0 );"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "memory" )
+
+#define gte_stcv( r0 ) __asm__ volatile ( \
+	"mfc2	$12, $9;"		\
+	"mfc2	$13, $10;"		\
+	"mfc2	$14, $11;"		\
+	"sb	$12, 0( %0 );"		\
+	"sb	$13, 1( %0 );"		\
+	"sb	$14, 2( %0 );"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14", "memory" )
+
 #define gte_strgb( r0 ) __asm__ volatile ( \
 	"swc2	$22, 0( %0 );"	\
 	:						\
 	: "r"( r0 )				\
 	: "memory" )
-	
+
 #define gte_strgb3( r0, r1, r2 ) __asm__ volatile ( \
 	"swc2	$20, 0( %0 );"	\
 	"swc2	$21, 0( %1 );"	\
 	"swc2	$22, 0( %2 );"	\
 	:						\
-	: "r"( r0 ), "r"( r1 ), "r" ( r2 )	\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 )	\
 	: "memory" )
 
-#define gte_stsv( r0 ) __asm__ volatile ( \
-	"mfc2	$t0, $9;"		\
-	"mfc2	$t1, $10;"		\
-	"mfc2	$t2, $11;"		\
-	"sh		$t0, 0( %0 );"	\
-	"sh		$t1, 2( %0 );"	\
-	"sh		$t2, 4( %0 );"	\
+#define gte_strgb3_g3( r0 ) __asm__ volatile ( \
+	"swc2	$20, 4( %0 );"	\
+	"swc2	$21, 12( %0 );"	\
+	"swc2	$22, 20( %0 );"	\
 	:						\
 	: "r"( r0 )				\
 	: "memory" )
-	
-#define gte_stlvnl( r0 ) __asm__ volatile ( \
-	"swc2	$25, 0( %0 );"	\
-	"swc2	$26, 4( %0 );"	\
-	"swc2	$27, 8( %0 );"	\
+
+#define gte_strgb3_gt3( r0 ) __asm__ volatile ( \
+	"swc2	$20, 4( %0 );"	\
+	"swc2	$21, 16( %0 );"	\
+	"swc2	$22, 28( %0 );"	\
 	:						\
 	: "r"( r0 )				\
 	: "memory" )
-	
-	
-/*
- *	GTE operation macros (does not need a stupid tool such as dmpsx!)
- */
- 
+
+#define gte_strgb3_g4( r0 ) __asm__ volatile ( \
+	"swc2	$20, 4( %0 );"	\
+	"swc2	$21, 12( %0 );"	\
+	"swc2	$22, 20( %0 );"	\
+	:						\
+	: "r"( r0 )				\
+	: "memory" )
+
+#define gte_strgb3_gt4( r0 ) __asm__ volatile ( \
+	"swc2	$20, 4( %0 );"	\
+	"swc2	$21, 16( %0 );"	\
+	"swc2	$22, 28( %0 );"	\
+	:						\
+	: "r"( r0 )				\
+	: "memory" )
+
+#define gte_ReadGeomOffset( r0, r1 ) __asm__ volatile ( \
+	"cfc2	$12, $24;"		\
+	"cfc2	$13, $25;"		\
+	"sra	$12, $12, 16;"	\
+	"sra	$13, $13, 16;"	\
+	"sw	$12, 0( %0 );"		\
+	"sw	$13, 0( %1 );"		\
+	:						\
+	: "r"( r0 ), "r"( r1 )	\
+	: "$12", "$13", "memory" )
+
+#define gte_ReadGeomScreen( r0 ) __asm__ volatile ( \
+	"cfc2	$12, $26;"		\
+	"nop;"					\
+	"sw	$12, 0( %0 );"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "memory" )
+
+#define gte_ReadRotMatrix( r0 ) __asm__ volatile ( \
+	"cfc2	$12, $0;"		\
+	"cfc2	$13, $1;"		\
+	"sw	$12, 0( %0 );"		\
+	"sw	$13, 4( %0 );"		\
+	"cfc2	$12, $2;"		\
+	"cfc2	$13, $3;"		\
+	"cfc2	$14, $4;"		\
+	"sw	$12, 8( %0 );"		\
+	"sw	$13, 12( %0 );"		\
+	"sw	$14, 16( %0 );"		\
+	"cfc2	$12, $5;"		\
+	"cfc2	$13, $6;"		\
+	"cfc2	$14, $7;"		\
+	"sw	$12, 20( %0 );"		\
+	"sw	$13, 24( %0 );"		\
+	"sw	$14, 28( %0 );"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14", "memory" )
+
+#define gte_sttr( r0 ) __asm__ volatile ( \
+	"cfc2	$12, $5;"		\
+	"cfc2	$13, $6;"		\
+	"cfc2	$14, $7;"		\
+	"sw	$12, 0( %0 );"		\
+	"sw	$13, 4( %0 );"		\
+	"sw	$14, 8( %0 );"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14", "memory" )
+
+#define gte_ReadLightMatrix( r0 ) __asm__ volatile ( \
+	"cfc2	$12, $8;"		\
+	"cfc2	$13, $9;"		\
+	"sw	$12, 0( %0 );"		\
+	"sw	$13, 4( %0 );"		\
+	"cfc2	$12, $10;"		\
+	"cfc2	$13, $11;"		\
+	"cfc2	$14, $12;"		\
+	"sw	$12, 8( %0 );"		\
+	"sw	$13, 12( %0 );"		\
+	"sw	$14, 16( %0 );"		\
+	"cfc2	$12, $13;"		\
+	"cfc2	$13, $14;"		\
+	"cfc2	$14, $15;"		\
+	"sw	$12, 20( %0 );"		\
+	"sw	$13, 24( %0 );"		\
+	"sw	$14, 28( %0 );"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14", "memory" )
+
+#define gte_ReadColorMatrix( r0 ) __asm__ volatile ( \
+	"cfc2	$12, $16;"			\
+	"cfc2	$13, $17;"		\
+	"sw	$12, 0( %0 );"		\
+	"sw	$13, 4( %0 );"		\
+	"cfc2	$12, $18;"		\
+	"cfc2	$13, $19;"		\
+	"cfc2	$14, $20;"		\
+	"sw	$12, 8( %0 );"		\
+	"sw	$13, 12( %0 );"		\
+	"sw	$14, 16( %0 );"		\
+	"cfc2	$12, $21;"		\
+	"cfc2	$13, $22;"		\
+	"cfc2	$14, $23;"		\
+	"sw	$12, 20( %0 );"		\
+	"sw	$13, 24( %0 );"		\
+	"sw	$14, 28( %0 );"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14", "memory" )
+
+#define gte_stlzc( r0 ) __asm__ volatile ( \
+	"swc2	$31, 0( %0 );"	\
+	:						\
+	: "r"( r0 )				\
+	: "memory" )
+
+#define gte_stfc( r0 ) __asm__ volatile ( \
+	"cfc2	$12, $21;"		\
+	"cfc2	$13, $22;"		\
+	"cfc2	$14, $23;"		\
+	"sw	$12, 0( %0 );"		\
+	"sw	$13, 4( %0 );"		\
+	"sw	$14, 8( %0 );"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14", "memory" )
+
+#define gte_mvlvtr() __asm__ volatile ( \
+	"mfc2	$12, $25;"		\
+	"mfc2	$13, $26;"		\
+	"mfc2	$14, $27;"		\
+	"ctc2	$12, $5;"		\
+	"ctc2	$13, $6;"		\
+	"ctc2	$14, $7;"		\
+	:						\
+	:						\
+	: "$12", "$13", "$14" )
+
+/*#define gte_nop() __asm__ volatile ( \
+	"nop;" )*/
+
+#define gte_subdvl( r0, r1, r2 ) __asm__ volatile ( \
+	"lw	$12, 0( %0 );"		\
+	"lw	$13, 0( %1 );"		\
+	"mtc2	$12, $9;"		\
+	"mtc2	$13, $10;"		\
+	"sra	$12, $12, 16;"	\
+	"sra	$13, $13, 16;"	\
+	"subu	$15, $12, $13;"	\
+	"mfc2	$12, $9;"		\
+	"mfc2	$13, $10;"		\
+	"sw	$15, 4( %2 );"		\
+	"subu	$12, $12, $13;"	\
+	"sw	$12, 0( %2 );"		\
+	:						\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 )	\
+	: "$12", "$13", "$14", "$15", "memory" )
+
+#define gte_subdvd( r0, r1, r2 ) __asm__ volatile ( \
+	"lw	$12, 0( %0 );"		\
+	"lw	$13, 0( %1 );"		\
+	"mtc2	$12, $9;"		\
+	"mtc2	$13, $10;"		\
+	"sra	$12, $12, 16;"	\
+	"sra	$13, $13, 16;"	\
+	"subu	$15, $12, $13;"	\
+	"mfc2	$12, $9;"		\
+	"mfc2	$13, $10;"		\
+	"sh	$15, 2( %2 );"		\
+	"subu	$12, $12, $13;"	\
+	"sh	$12, 0( %2 );"		\
+	:						\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 )	\
+	: "$12", "$13", "$14", "$15", "memory" )
+
+#define gte_adddvl( r0, r1, r2 ) __asm__ volatile ( \
+	"lw	$12, 0( %0 );"		\
+	"lw	$13, 0( %1 );"		\
+	"mtc2	$12, $9;"		\
+	"mtc2	$13, $10;"		\
+	"sra	$12, $12, 16;"	\
+	"sra	$13, $13, 16;"	\
+	"addu	$15, $12, $13;"	\
+	"mfc2	$12, $9;"		\
+	"mfc2	$13, $10;"		\
+	"sw	$15, 4( %2 );"		\
+	"addu	$12, $12, $13;"	\
+	"sw	$12, 0( %2 );"		\
+	:						\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 )	\
+	: "$12", "$13", "$14", "$15", "memory" )
+
+#define gte_adddvd( r0, r1, r2 ) __asm__ volatile ( \
+	"lw	$12, 0( %0 );"		\
+	"lw	$13, 0( %1 );"		\
+	"mtc2	$12, $9;"		\
+	"mtc2	$13, $10;"		\
+	"sra	$12, $12, 16;"	\
+	"sra	$13, $13, 16;"	\
+	"addu	$15, $12, $13;"	\
+	"mfc2	$12, $9;"		\
+	"mfc2	$13, $10;"		\
+	"sh	$15, 2( %2 );"		\
+	"addu	$12, $12, $13;"	\
+	"sh	$12, 0( %2 );"		\
+	:						\
+	: "r"( r0 ), "r"( r1 ), "r"( r2 )	\
+	: "$12", "$13", "$14", "$15", "memory" )
+
+#define gte_FlipRotMatrixX() __asm__ volatile ( \
+	"cfc2	$12, $0;"		\
+	"cfc2	$13, $1;"		\
+	"sll	$14, $12, 16;"	\
+	"sra	$14, $14, 16;"	\
+	"subu	$14, $0, $14;"	\
+	"sra	$15, $12, 16;"	\
+	"subu	$15, $0, $15;"	\
+	"sll	$15, $15, 16;"	\
+	"sll	$14, $14, 16;"	\
+	"srl	$14, $14, 16;"	\
+	"or	$14, $14, $15;"		\
+	"ctc2	$14, $0;"		\
+	"sll	$14, $13, 16;"	\
+	"sra	$14, $14, 16;"	\
+	"subu	$14, $0, $14;"	\
+	"sra	$15, $13, 16;"	\
+	"sll	$15, $15, 16;"	\
+	"sll	$14, $14, 16;"	\
+	"srl	$14, $14, 16;"	\
+	"or	$14, $14, $15;"		\
+	"ctc2	$14, $1;"		\
+	:						\
+	:						\
+	: "$12", "$13", "$14", "$15" )
+
+#define gte_FlipTRX() __asm__ volatile ( \
+	"cfc2	$12, $5;"		\
+	"nop;"					\
+	"subu	$12, $0, $12;"	\
+	"ctc2	$12, $5;"		\
+	:						\
+	:						\
+	: "$12" )
+
+/* GTE operation macros */
+
 #define gte_rtps() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
@@ -340,169 +1053,405 @@
 	"nop;"					\
 	"nop;"					\
 	"cop2 0x0280030;" )
-	
-#define gte_nclip() __asm__ volatile ( \
+
+#define gte_rt() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2 0x1400006;" )
-	
-#define gte_avsz3() __asm__ volatile ( \
+	"cop2 0x0480012;" )
+
+#define gte_rtv0() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2 0x158002D;" )
-	
-#define gte_avsz4() __asm__ volatile ( \
+	"cop2 0x0486012;" )
+
+#define gte_rtv1() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2 0x168002E;" )
-	
-#define gte_sqr0() __asm__ volatile ( \
+	"cop2 0x048E012;" )
+
+#define gte_rtv2() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x0A00428;" )
-	
-#define gte_sqr12() __asm__ volatile ( \
+	"cop2 0x0496012;" )
+
+#define gte_rtir() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x0A80428;" )
-	
-#define gte_op0() __asm__ volatile ( \
+	"cop2 0x049E012;" )
+
+#define gte_rtir_sf0() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x170000C;"	)
+	"cop2 0x041E012;" )
 
-#define gte_op12() __asm__ volatile ( \
+#define gte_rtv0tr() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x178000C;"	)
-	
-#define gte_ncs() __asm__ volatile ( \
+	"cop2 0x0480012;" )
+
+#define gte_rtv1tr() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x0C8041E;" )
-	
-#define gte_nct() __asm__ volatile ( \
+	"cop2 0x0488012;" )
+
+#define gte_rtv2tr() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x0D80420;" )
-	
-#define gte_nccs() __asm__ volatile ( \
+	"cop2 0x0490012;" )
+
+#define gte_rtirtr() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x108041B;" )	\
-	
-#define gte_ncct() __asm__ volatile ( \
+	"cop2 0x0498012;" )
+
+#define gte_rtv0bk() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x118043F;"	)
-	
-#define gte_ncds() __asm__ volatile ( \
+	"cop2 0x0482012;" )
+
+#define gte_rtv1bk() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x0E80413;"	)
-	
-#define gte_ncdt() __asm__ volatile ( \
+	"cop2 0x048A012;" )
+
+#define gte_rtv2bk() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x0F80416;" )
-	
-#define gte_cc() __asm__ volatile ( \
+	"cop2 0x0492012;" )
+
+#define gte_rtirbk() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x138041C;" )
-	
-#define gte_cdp() __asm__ volatile ( \
+	"cop2 0x049A012;" )
+
+#define gte_ll() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04A6412;" )
+
+#define gte_llv0() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04A6012;" )
+
+#define gte_llv1() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04AE012;" )
+
+#define gte_llv2() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04B6012;" )
+
+#define gte_llir() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04BE012;" )
+
+#define gte_llv0tr() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04A0012;" )
+
+#define gte_llv1tr() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04A8012;" )
+
+#define gte_llv2tr() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04B0012;" )
+
+#define gte_llirtr() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04B8012;" )
+
+#define gte_llv0bk() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04A2012;" )
+
+#define gte_llv1bk() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04AA012;" )
+
+#define gte_llv2bk() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04B2012;" )
+
+#define gte_llirbk() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04BA012;" )
+
+
+#define gte_lc() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04DA412;" )
+
+#define gte_lcv0() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04C6012;" )
+
+#define gte_lcv1() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04CE012;" )
+
+#define gte_lcv2() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04D6012;" )
+
+#define gte_lcir() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x1280414;" )
-	
-#define gte_dcpl() __asm__ volatile ( \
+	"cop2 0x04DE012;" )
+
+#define gte_lcv0tr() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04C0012;" )
+
+#define gte_lcv1tr() __asm__ volatile ( \
+	"nop;"					\
 	"nop;"					\
+	"cop2 0x04C8012;" )
+
+#define gte_lcv2tr() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04D0012;" )
+
+#define gte_lcirtr() __asm__ volatile ( \
 	"nop;"					\
-	"cop2	0x0680029;"	)
-	
+	"nop;"					\
+	"cop2 0x04D8012;" )
+
+#define gte_lcv0bk() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04C2012;" )
+
+#define gte_lcv1bk() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04CA012;" )
+
+#define gte_lcv2bk() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04D2012;" )
+
+#define gte_lcirbk() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x04DA012;" )
+
+#define gte_dpcl() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x0680029;" )
+
 #define gte_dpcs() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x0780010;" )
-	
+	"cop2 0x0780010;" )
+
 #define gte_dpct() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x0180001;" )
+	"cop2 0x0F8002A;" )
 
 #define gte_intpl() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x0980011;" )
+	"cop2 0x0980011;" )
 
-#define gte_gpf0() __asm__ volatile ( \
+#define gte_sqr12() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x0A80428;" )
+
+#define gte_sqr0() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x190003D;"	)
+	"cop2 0x0A00428;" )
+
+#define gte_ncs() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x0C8041E;" )
+
+#define gte_nct() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x0D80420;" )
+
+#define gte_ncds() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x0E80413;" )
+
+#define gte_ncdt() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x0F80416;" )
+
+#define gte_nccs() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x0108041B;" )
+
+#define gte_ncct() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x0118043F;" )
+
+#define gte_cdp() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x01280414;" )
+
+#define gte_cc() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x0138041C;" )
+
+#define gte_nclip() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x01400006;" )
+
+#define gte_avsz3() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x0158002D;" )
+
+#define gte_avsz4() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x0168002E;" )
+
+#define gte_op12() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x0178000C;" )
+
+#define gte_op0() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x0170000C;" )
 
 #define gte_gpf12() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x198003D;"	)
-	
-#define gte_gpl0() __asm__ volatile ( \
+	"cop2 0x0198003D;" )
+
+#define gte_gpf0() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x1A0003E;" )
+	"cop2 0x0190003D;" )
 
 #define gte_gpl12() __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2	0x1A8003E;" )
+	"cop2 0x01A8003E;" )
+
+#define gte_gpl0() __asm__ volatile ( \
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x01A0003E;" )
 
 #define gte_mvmva_core( r0 ) __asm__ volatile ( \
 	"nop;"					\
 	"nop;"					\
-	"cop2 %0"				\
+	"cop2 %0;"				\
 	:						\
 	: "g"( r0 ) )
-	
+
 #define gte_mvmva(sf, mx, v, cv, lm) gte_mvmva_core( 0x0400012 | \
 	((sf)<<19) | ((mx)<<17) | ((v)<<15) | ((cv)<<13) | ((lm)<<10) )
-	
-	
-/*
- *	GTE operation macros without leading nops
- *
- *	Checking assembler output when using these is advised.
- */
+
+/* GTE operation macros without leading nops */
  
-#define gte_rtps_b()	__asm__ volatile ( "cop2 0x0180001;" )
-#define gte_rtpt_b()	__asm__ volatile ( "cop2 0x0280030;" )
-#define gte_nclip_b()	__asm__ volatile ( "cop2 0x1400006;" )
-#define gte_avsz3_b()	__asm__ volatile ( "cop2 0x158002D;" )
-#define gte_avsz4_b()	__asm__ volatile ( "cop2 0x168002E;" )
-#define gte_sqr0_b()	__asm__ volatile ( "cop2 0x0A00428;" )
-#define gte_sqr12_b()	__asm__ volatile ( "cop2 0x0A80428;" )
-#define gte_op0_b()		__asm__ volatile ( "cop2 0x170000C;" )
-#define gte_op12_b()	__asm__ volatile ( "cop2 0x178000C;" )
-#define gte_ncs_b()		__asm__ volatile ( "cop2 0x0C8041E;" )
-#define gte_nct_b()		__asm__ volatile ( "cop2 0x0D80420;" )
-#define gte_nccs_b()	__asm__ volatile ( "cop2 0x108041B;" )
-#define gte_ncct_b()	__asm__ volatile ( "cop2 0x118043F;" )
-#define gte_ncds_b()	__asm__ volatile ( "cop2 0x0E80413;" )
-#define gte_ncdt_b()	__asm__ volatile ( "cop2 0x0F80416;" )
-#define gte_cc_b()		__asm__ volatile ( "cop2 0x138041C;" )
-#define gte_cdp_b()		__asm__ volatile ( "cop2 0x1280414;" )
-#define gte_dcpl_b()	__asm__ volatile ( "cop2 0x0680029;" )
-#define gte_dpcs_b()	__asm__ volatile ( "cop2 0x0780010;" )
-#define gte_dpct_b()	__asm__ volatile ( "cop2 0x0180001;" )
-#define gte_intpl_b()	__asm__ volatile ( "cop2 0x0980011;" )
-#define gte_gpf0_b()	__asm__ volatile ( "cop2 0x190003D;" )
-#define gte_gpf12_b()	__asm__ volatile ( "cop2 0x198003D;" )
-#define gte_gpl0_b()	__asm__ volatile ( "cop2 0x1A0003E;" )
-#define gte_gpl12_b()	__asm__ volatile ( "cop2 0x1A8003E;" )
+#define gte_rtps_b()		__asm__ volatile ( "cop2 0x0180001;" )
+#define gte_rtpt_b()		__asm__ volatile ( "cop2 0x0280030;" )
+#define gte_rt_b()			__asm__ volatile ( "cop2 0x0480012;" )
+#define gte_rtv0_b()		__asm__ volatile ( "cop2 0x0486012;" )
+#define gte_rtv1_b()		__asm__ volatile ( "cop2 0x048E012;" )
+#define gte_rtv2_b()		__asm__ volatile ( "cop2 0x0496012;" )
+#define gte_rtir_b()		__asm__ volatile ( "cop2 0x049E012;" )
+#define gte_rtir_sf0_b()	__asm__ volatile ( "cop2 0x041E012;" )
+#define gte_rtv0tr_b()		__asm__ volatile ( "cop2 0x0480012;" )
+#define gte_rtv1tr_b()		__asm__ volatile ( "cop2 0x0488012;" )
+#define gte_rtv2tr_b()		__asm__ volatile ( "cop2 0x0490012;" )
+#define gte_rtirtr_b()		__asm__ volatile ( "cop2 0x0498012;" )
+#define gte_rtv0bk_b()		__asm__ volatile ( "cop2 0x0482012;" )
+#define gte_rtv1bk_b()		__asm__ volatile ( "cop2 0x048A012;" )
+#define gte_rtv2bk_b()		__asm__ volatile ( "cop2 0x0492012;" )
+#define gte_rtirbk_b()		__asm__ volatile ( "cop2 0x049A012;" )
+#define gte_ll_b()			__asm__ volatile ( "cop2 0x04A6412;" )
+#define gte_llv0_b()		__asm__ volatile ( "cop2 0x04A6012;" )
+#define gte_llv1_b()		__asm__ volatile ( "cop2 0x04AE012;" )
+#define gte_llv2_b()		__asm__ volatile ( "cop2 0x04B6012;" )
+#define gte_llir_b()		__asm__ volatile ( "cop2 0x04BE012;" )
+#define gte_llv0tr_b()		__asm__ volatile ( "cop2 0x04A0012;" )
+#define gte_llv1tr_b()		__asm__ volatile ( "cop2 0x04A8012;" )
+#define gte_llv2tr_b()		__asm__ volatile ( "cop2 0x04B0012;" )
+#define gte_llirtr_b()		__asm__ volatile ( "cop2 0x04B8012;" )
+#define gte_llv0bk_b()		__asm__ volatile ( "cop2 0x04A2012;" )
+#define gte_llv1bk_b()		__asm__ volatile ( "cop2 0x04AA012;" )
+#define gte_llv2bk_b()		__asm__ volatile ( "cop2 0x04B2012;" )
+#define gte_llirbk_b()		__asm__ volatile ( "cop2 0x04BA012;" )
+#define gte_lc_b()			__asm__ volatile ( "cop2 0x04DA412;" )
+#define gte_lcv0_b()		__asm__ volatile ( "cop2 0x04C6012;" )
+#define gte_lcv1_b()		__asm__ volatile ( "cop2 0x04CE012;" )
+#define gte_lcv2_b()		__asm__ volatile ( "cop2 0x04D6012;" )
+#define gte_lcir_b()		__asm__ volatile ( "cop2 0x04DE012;" )
+#define gte_lcv0tr_b()		__asm__ volatile ( "cop2 0x04C0012;" )
+#define gte_lcv1tr_b()		__asm__ volatile ( "cop2 0x04C8012;" )
+#define gte_lcv2tr_b()		__asm__ volatile ( "cop2 0x04D0012;" )
+#define gte_lcirtr_b()		__asm__ volatile ( "cop2 0x04D8012;" )
+#define gte_lcv0bk_b()		__asm__ volatile ( "cop2 0x04C2012;" )
+#define gte_lcv1bk_b()		__asm__ volatile ( "cop2 0x04CA012;" )
+#define gte_lcv2bk_b()		__asm__ volatile ( "cop2 0x04D2012;" )
+#define gte_lcirbk_b()		__asm__ volatile ( "cop2 0x04DA012;" )
+#define gte_dpcl_b()		__asm__ volatile ( "cop2 0x0680029;" )
+#define gte_dpcs_b()		__asm__ volatile ( "cop2 0x0780010;" )
+#define gte_dpct_b()		__asm__ volatile ( "cop2 0x0F8002A;" )
+#define gte_intpl_b()		__asm__ volatile ( "cop2 0x0980011;" )
+#define gte_sqr12_b()		__asm__ volatile ( "cop2 0x0A80428;" )
+#define gte_sqr0_b()		__asm__ volatile ( "cop2 0x0A00428;" )
+#define gte_ncs_b()			__asm__ volatile ( "cop2 0x0C8041E;" )
+#define gte_nct_b()			__asm__ volatile ( "cop2 0x0D80420;" )
+#define gte_ncds_b()		__asm__ volatile ( "cop2 0x0E80413;" )
+#define gte_ncdt_b()		__asm__ volatile ( "cop2 0x0F80416;" )
+#define gte_nccs_b()		__asm__ volatile ( "cop2 0x0108041B;" )
+#define gte_ncct_b()		__asm__ volatile ( "cop2 0x0118043F;" )
+#define gte_cdp_b()			__asm__ volatile ( "cop2 0x01280414;" )
+#define gte_cc_b()			__asm__ volatile ( "cop2 0x0138041C;" )
+#define gte_nclip_b()		__asm__ volatile ( "cop2 0x01400006;" )
+#define gte_avsz3_b()		__asm__ volatile ( "cop2 0x0158002D;" )
+#define gte_avsz4_b()		__asm__ volatile ( "cop2 0x0168002E;" )
+#define gte_op12_b()		__asm__ volatile ( "cop2 0x0178000C;" )
+#define gte_op0_b()			__asm__ volatile ( "cop2 0x0170000C;" )
+#define gte_gpf12_b()		__asm__ volatile ( "cop2 0x0198003D;" )
+#define gte_gpf0_b()		__asm__ volatile ( "cop2 0x0190003D;" )
+#define gte_gpl12_b()		__asm__ volatile ( "cop2 0x01A8003E;" )
+#define gte_gpl0_b()		__asm__ volatile ( "cop2 0x01A0003E;" )
 #define gte_mvmva_core_b( r0 ) __asm__ volatile ( \
-	"cop2 %0"				\
+	"cop2 %0;"				\
 	:						\
-	: "g"( r0 ) )	
+	: "g"( r0 ) )
 #define gte_mvmva_b(sf, mx, v, cv, lm) gte_mvmva_core_b( 0x0400012 | \
 	((sf)<<19) | ((mx)<<17) | ((v)<<15) | ((cv)<<13) | ((lm)<<10) )
-	
-#endif // _INLINE_C_H
-\ No newline at end of file
+
+#endif // _INLINE_C_H
diff --git a/libpsn00b/include/inline_s.h b/libpsn00b/include/inline_s.inc
index 68e0d07..68e0d07 100644
--- a/libpsn00b/include/inline_s.h
+++ b/libpsn00b/include/inline_s.inc
diff --git a/libpsn00b/include/lzp/lzp.h b/libpsn00b/include/lzp/lzp.h
index cfeeb72..456de02 100644
--- a/libpsn00b/include/lzp/lzp.h
+++ b/libpsn00b/include/lzp/lzp.h
@@ -16,7 +16,7 @@
 #ifndef _LZPACK_H
 #define _LZPACK_H
 
-#include <sys/types.h>
+#include <stdint.h>
 #ifdef _WIN32
 #include <windows.h>
 #endif
@@ -64,9 +64,9 @@
 typedef struct {
 
 	//! File ID (must always be 'LZP')
-	char	id[3];
+	char	    id[3];
 	//! File count
-	u_char	numFiles;
+	uint8_t 	numFiles;
 
 } LZP_HEAD;
 
@@ -74,15 +74,15 @@ typedef struct {
 typedef struct {
 
 	//! File name
-    char	fileName[16];
+    char	    fileName[16];
     //! CRC32 checksum of file
-    u_int	crc;
+    uint32_t	crc;
     //! Original size of file in bytes
-    u_int	fileSize;
+    uint32_t	fileSize;
     //! Compressed size of file
-    u_int	packedSize;
+    uint32_t	packedSize;
     //! File data offset
-    u_int	offset;
+    uint32_t	offset;
 
 } LZP_FILE;
 
@@ -162,7 +162,7 @@ void lzResetHashSizes();
  *
  *	\returns CRC16 hash of specified buffer.
  */
-unsigned short lzCRC16(const void* buff, int bytes, unsigned short crc);
+uint16_t lzCRC16(const void* buff, int bytes, uint16_t crc);
 
 /*!	Calculates a CRC32 hash of the specified buffer.
  *
@@ -172,7 +172,7 @@ unsigned short lzCRC16(const void* buff, int bytes, unsigned short crc);
  *
  *	\returns CRC32 hash of specified buffer.
  */
-unsigned int lzCRC32(const void* buff, int bytes, unsigned int crc);
+uint32_t lzCRC32(const void* buff, int bytes, uint32_t crc);
 
 /*!	@}	*/
 
diff --git a/libpsn00b/include/lzp/lzqlp.h b/libpsn00b/include/lzp/lzqlp.h
index 5b70b40..32ce0d7 100644
--- a/libpsn00b/include/lzp/lzqlp.h
+++ b/libpsn00b/include/lzp/lzqlp.h
@@ -1,7 +1,7 @@
 #ifndef _QLP_H
 #define _QLP_H
 
-#include <sys/types.h>
+#include <stdint.h>
 #ifdef _WIN32
 #include <windows.h>
 #endif
@@ -13,14 +13,14 @@
 #define PACK_ERR_READ_FAULT		-4
 
 typedef struct {
-	char	id[3];
-	u_char	numfiles;
+	char		id[3];
+	uint8_t		numfiles;
 } QLP_HEAD;
 
 typedef struct {
-	char	name[16];
-	u_int	size;
-	u_int	offs;
+	char		name[16];
+	uint32_t	size;
+	uint32_t	offs;
 } QLP_FILE;
 
 int qlpFileCount(const QLP_HEAD* qlpfile);
diff --git a/libpsn00b/include/psxapi.h b/libpsn00b/include/psxapi.h
index 1298d29..5d1097e 100644
--- a/libpsn00b/include/psxapi.h
+++ b/libpsn00b/include/psxapi.h
@@ -1,5 +1,15 @@
-#ifndef __PSXAPI__
-#define __PSXAPI__
+/*
+ * PSn00bSDK kernel API library
+ * (C) 2019-2022 Lameguy64, spicyjpeg - MPL licensed
+ */
+
+#ifndef __PSXAPI_H
+#define __PSXAPI_H
+
+#include <stdint.h>
+#include <stddef.h>
+
+/* Definitions */
 
 #define DescHW			0xf0000000
 #define DescSW			0xf4000000
@@ -32,29 +42,31 @@
 #define RCntMdFR		0x0000
 #define RCntMdGATE		0x0010
 
+/* Structure definitions */
+
 typedef struct {	// Thread control block
-	int						status;
-	int						mode;
+	int					status;
+	int					mode;
 	union {
-		unsigned int		reg[37];
+		uint32_t		reg[37];
 		struct {
-			unsigned int	zero, at;
-			unsigned int	v0, v1;
-			unsigned int	a0, a1, a2, a3;
-			unsigned int	t0, t1, t2, t3, t4, t5, t6, t7;
-			unsigned int	s0, s1, s2, s3, s4, s5, s6, s7;
-			unsigned int	t8, t9;
-			unsigned int	k0, k1;
-			unsigned int	gp, sp, fp, ra;
-
-			unsigned int	cop0r14;
-			unsigned int	hi;
-			unsigned int	lo;
-			unsigned int	cop0r12;
-			unsigned int	cop0r13;
+			uint32_t	zero, at;
+			uint32_t	v0, v1;
+			uint32_t	a0, a1, a2, a3;
+			uint32_t	t0, t1, t2, t3, t4, t5, t6, t7;
+			uint32_t	s0, s1, s2, s3, s4, s5, s6, s7;
+			uint32_t	t8, t9;
+			uint32_t	k0, k1;
+			uint32_t	gp, sp, fp, ra;
+
+			uint32_t	cop0r14;
+			uint32_t	hi;
+			uint32_t	lo;
+			uint32_t	cop0r12;
+			uint32_t	cop0r13;
 		};
 	};
-	int						_reserved[9];
+	int					_reserved[9];
 } TCB;
 
 typedef struct {	// Process control block
@@ -85,17 +97,17 @@ typedef struct {					// Device control block
 } DCB;
 
 typedef struct {			// File control block
-	int				status;
-	unsigned int	diskid;
-	void			*trns_addr;
-	unsigned int	trns_len;
-	unsigned int	filepos;
-	unsigned int	flags;
-	unsigned int	lasterr;
-	DCB				*dcb;
-	unsigned int	filesize;
-	unsigned int	lba;
-	unsigned int	fcbnum;
+	int			status;
+	uint32_t	diskid;
+	void		*trns_addr;
+	uint32_t	trns_len;
+	uint32_t	filepos;
+	uint32_t	flags;
+	uint32_t	lasterr;
+	DCB			*dcb;
+	uint32_t	filesize;
+	uint32_t	lba;
+	uint32_t	fcbnum;
 } FCB;
 
 struct DIRENTRY {			// Directory entry
@@ -108,34 +120,31 @@ struct DIRENTRY {			// Directory entry
 };
 
 struct EXEC {
-	unsigned int pc0;
-	unsigned int gp0;
-	unsigned int t_addr;
-	unsigned int t_size;
-	unsigned int d_addr;
-	unsigned int d_size;
-	unsigned int b_addr;
-	unsigned int b_size;
-	unsigned int s_addr;
-	unsigned int s_size;
-	unsigned int sp,fp,rp,ret,base;
+	uint32_t pc0, gp0;
+	uint32_t t_addr, t_size;
+	uint32_t d_addr, d_size;
+	uint32_t b_addr, b_size;
+	uint32_t s_addr, s_size;
+	uint32_t sp, fp, rp, ret, base;
 };
 
 struct JMP_BUF {
-	unsigned int	ra, sp, fp;
-	unsigned int	s0, s1, s2, s3, s4, s5, s6, s7;
-	unsigned int	gp;
+	uint32_t ra, sp, fp;
+	uint32_t s0, s1, s2, s3, s4, s5, s6, s7;
+	uint32_t gp;
 };
 
 // Not recommended to use these functions to install IRQ handlers
 
 typedef struct {
-	unsigned int* next;
-	unsigned int* func2;
-	unsigned int* func1;
-	unsigned int pad;
+	uint32_t	*next;
+	uint32_t	*func2;
+	uint32_t	*func1;
+	int			_reserved;
 } INT_RP;
 
+/* API */
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -143,30 +152,27 @@ extern "C" {
 void SysEnqIntRP(int pri, INT_RP *rp);
 void SysDeqIntRP(int pri, INT_RP *rp);
 
-// Event handler stuff
-
-int OpenEvent(unsigned int cl, int spec, int mode, void (*func)());
-int CloseEvent(int ev_desc);
-int EnableEvent(int ev_desc);
-int DisableEvent(int ev_desc);
-
-// BIOS file functions
+int OpenEvent(uint32_t cl, uint32_t spec, int mode, void (*func)());
+int CloseEvent(int event);
+int WaitEvent(int event);
+int TestEvent(int event);
+int EnableEvent(int event);
+int DisableEvent(int event);
+void DeliverEvent(uint32_t cl, uint32_t spec);
+void UnDeliverEvent(uint32_t cl, uint32_t spec);
 
 int open(const char *name, int mode);
 int close(int fd);
-int seek(int fd, unsigned int offset, int mode);
-int read(int fd, char *buff, unsigned int len);
-int write(int fd, const char *buff, unsigned int len);
+int seek(int fd, uint32_t offset, int mode);
+int read(int fd, uint8_t *buff, size_t len);
+int write(int fd, const uint8_t *buff, size_t len);
 int ioctl(int fd, int cmd, int arg);
 struct DIRENTRY *firstfile(const char *wildcard, struct DIRENTRY *entry);
 struct DIRENTRY *nextfile(struct DIRENTRY *entry);
 int erase(const char *name);
 int chdir(const char *path);
 
-//#define delete( p )	erase( p )	// May conflict with delete operator in C++
-#define cd( p )		chdir( p )		// For compatibility
-
-// BIOS device functions
+//#define cd(p) chdir(p)
 
 int AddDev(DCB *dcb);
 int DelDev(const char *name);
@@ -175,18 +181,17 @@ void AddDummyTty(void);
 
 void EnterCriticalSection(void);
 void ExitCriticalSection(void);
+void SwEnterCriticalSection(void);
+void SwExitCriticalSection(void);
 
-// BIOS CD functions
 void _InitCd(void);
 void _96_init(void);
 void _96_remove(void);
 
-// BIOS pad functions
-void InitPAD(char *buff1, int len1, char *buff2, int len2);
+void InitPAD(uint8_t *buff1, int len1, uint8_t *buff2, int len2);
 void StartPAD(void);
 void StopPAD(void);
 
-// BIOS memory card functions
 void InitCARD(int pad_enable);
 void StartCARD(void);
 void StopCARD(void);
@@ -198,32 +203,32 @@ int _card_status(int chan);
 int _card_wait(int chan);
 int _card_clear(int chan);
 int _card_chan(void);
-int _card_read(int chan, int sector, unsigned char *buf);
-int _card_write(int chan, int sector, unsigned char *buf);
+int _card_read(int chan, int sector, uint8_t *buf);
+int _card_write(int chan, int sector, uint8_t *buf);
 void _new_card(void);
 
-// Timers
-int SetRCnt(int spec, unsigned short target, int mode);
+int SetRCnt(int spec, uint16_t target, int mode);
 int GetRCnt(int spec);
 int StartRCnt(int spec);
 int StopRCnt(int spec);
 int ResetRCnt(int spec);
 
-// BIOS IRQ acknowledge control
 void ChangeClearPAD(int mode);
 void ChangeClearRCnt(int t, int m);
 
-// Executable functions
+uint32_t OpenTh(uint32_t (*func)(), uint32_t sp, uint32_t gp);
+int CloseTh(uint32_t thread);
+int ChangeTh(uint32_t thread);
+
 int Exec(struct EXEC *exec, int argc, char **argv);
 void FlushCache(void);
 
-// BIOS setjmp functions
 void b_setjmp(struct JMP_BUF *buf);
-void b_longjmp(struct JMP_BUF *buf, int param);
+void b_longjmp(const struct JMP_BUF *buf, int param);
 void SetDefaultExitFromException(void);
-void SetCustomExitFromException(struct JMP_BUF *buf);
+void SetCustomExitFromException(const struct JMP_BUF *buf);
+void ReturnFromException(void);
 
-// Misc functions
 int GetSystemInfo(int index);
 void *GetB0Table(void);
 void *GetC0Table(void);
diff --git a/libpsn00b/include/psxcd.h b/libpsn00b/include/psxcd.h
index 3336963..03ee792 100644
--- a/libpsn00b/include/psxcd.h
+++ b/libpsn00b/include/psxcd.h
@@ -1,7 +1,12 @@
-#ifndef _LIBPSXCD_H
-#define _LIBPSXCD_H
+/*
+ * PSn00bSDK CD-ROM drive library
+ * (C) 2019-2022 Lameguy64, spicyjpeg - MPL licensed
+ */
 
-#include <sys/types.h>
+#ifndef __PSXCD_H
+#define __PSXCD_H
+
+#include <stdint.h>
 
 /*
  * CD-ROM control commands
@@ -74,18 +79,18 @@
 #define CdlIsoInvalidFs		0x03
 #define CdlIsoLidOpen		0x04
 
-#define btoi(b)		((b)/16*10+(b)%16)	/* Convert BCD value to integer */
-#define itob(i)		((i)/10*16+(i)%10)	/* Convert integer to BCD value */
+#define btoi(b) ((b)/16*10+(b)%16)	/* Convert BCD value to integer */
+#define itob(i) ((i)/10*16+(i)%10)	/* Convert integer to BCD value */
 
 /*
  * CD-ROM disc location struct
  */
 typedef struct _CdlLOC
 {
-	u_char	minute;
-	u_char	second;
-	u_char	sector;
-	u_char	track;
+	uint8_t minute;
+	uint8_t second;
+	uint8_t sector;
+	uint8_t track;
 } CdlLOC;
 
 /*
@@ -93,10 +98,10 @@ typedef struct _CdlLOC
  */
 typedef struct _CdlATV
 {
-	u_char	val0;	/* L -> SPU L */
-	u_char	val1;	/* L -> SPU R */
-	u_char	val2;	/* R -> SPU R */
-	u_char	val3;	/* R -> SPU L */
+	uint8_t val0;	/* L -> SPU L */
+	uint8_t val1;	/* L -> SPU R */
+	uint8_t val2;	/* R -> SPU R */
+	uint8_t val3;	/* R -> SPU L */
 } CdlATV;
 
 /*
@@ -104,68 +109,68 @@ typedef struct _CdlATV
  */
 typedef struct _CdlFILE
 {
-	CdlLOC	pos;
-	u_long	size;
-	char	name[16];
+	CdlLOC		pos;
+	uint32_t	size;
+	char		name[16];
 } CdlFILE;
 
 typedef struct _CdlFILTER
 {
-	u_char	file;
-	u_char	chan;
-	u_short	pad;
+	uint8_t		file;
+	uint8_t		chan;
+	uint16_t	pad;
 } CdlFILTER;
 
 /* Directory query context */
 typedef void* CdlDIR;
 
 /* Data callback */
-typedef void (*CdlCB)(int, u_char *);
+typedef void (*CdlCB)(int, uint8_t *);
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-int		CdInit(void);
+int CdInit(void);
 
-CdlLOC*	CdIntToPos(int i, CdlLOC *p);
-int		CdPosToInt(CdlLOC *p);
-int		CdGetToc(CdlLOC *toc);
+CdlLOC* CdIntToPos(int i, CdlLOC *p);
+int CdPosToInt(CdlLOC *p);
+int CdGetToc(CdlLOC *toc);
 
-int		CdControl(u_char com, const void *param, u_char *result);
-int		CdControlB(u_char com, const void *param, u_char *result);
-int		CdControlF(u_char com, const void *param);
-int		CdSync(int mode, u_char *result);
-u_long	CdSyncCallback(CdlCB func);
+int CdControl(uint8_t com, const void *param, uint8_t *result);
+int CdControlB(uint8_t com, const void *param, uint8_t *result);
+int CdControlF(uint8_t com, const void *param);
+int CdSync(int mode, uint8_t *result);
+uint32_t CdSyncCallback(CdlCB func);
 
-long	CdReadyCallback(CdlCB func);
-int		CdGetSector(void *madr, int size);
+long CdReadyCallback(CdlCB func);
+int CdGetSector(void *madr, int size);
 
 CdlFILE* CdSearchFile(CdlFILE *loc, const char *filename);
 
-int		CdRead(int sectors, u_long *buf, int mode);
-int		CdReadSync(int mode, u_char *result);
-u_long	CdReadCallback(CdlCB func);
+int CdRead(int sectors, uint32_t *buf, int mode);
+int CdReadSync(int mode, uint8_t *result);
+uint32_t CdReadCallback(CdlCB func);
 
-int		CdStatus(void);
-int		CdMode(void);
+int CdStatus(void);
+int CdMode(void);
 
-int		CdMix(CdlATV *vol);
+int CdMix(CdlATV *vol);
 
 /* ORIGINAL CODE */
-CdlDIR*	CdOpenDir(const char* path);
-int		CdReadDir(CdlDIR* dir, CdlFILE* file);
-void	CdCloseDir(CdlDIR* dir);
+CdlDIR* CdOpenDir(const char* path);
+int CdReadDir(CdlDIR* dir, CdlFILE* file);
+void CdCloseDir(CdlDIR* dir);
 
-int		CdGetVolumeLabel(char* label);
+int CdGetVolumeLabel(char* label);
 
-long*	CdAutoPauseCallback(void(*func)());
-int		CdIsoError();
+long* CdAutoPauseCallback(void(*func)());
+int CdIsoError();
 
-int		CdLoadSession(int session);
+int CdLoadSession(int session);
 
 #ifdef __cplusplus
 }
 #endif
 
-#endif /* _LIBPSXCD_H */
+#endif
diff --git a/libpsn00b/include/psxetc.h b/libpsn00b/include/psxetc.h
index a55e593..24485d9 100644
--- a/libpsn00b/include/psxetc.h
+++ b/libpsn00b/include/psxetc.h
@@ -1,15 +1,25 @@
-#ifndef _PSXETC_H
-#define _PSXETC_H
+/*
+ * PSn00bSDK interrupt management library
+ * (C) 2019-2022 Lameguy64, spicyjpeg - MPL licensed
+ */
+
+#ifndef __PSXETC_H
+#define __PSXETC_H
+
+/* Public API */
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-// Interrupt callback functions
-void *DMACallback(int dma, void (*func)(void));
 void *InterruptCallback(int irq, void (*func)(void));
-void *GetInterruptCallback(int irq);				// Original
-void RestartCallback();
+void *GetInterruptCallback(int irq);
+void *DMACallback(int dma, void (*func)(void));
+void *GetDMACallback(int dma);
+
+int ResetCallback(void);
+void RestartCallback(void);
+void StopCallback(void);
 
 #ifdef __cplusplus
 }
diff --git a/libpsn00b/include/psxgpu.h b/libpsn00b/include/psxgpu.h
index f061219..0e7ec00 100644
--- a/libpsn00b/include/psxgpu.h
+++ b/libpsn00b/include/psxgpu.h
@@ -1,623 +1,519 @@
+/*
+ * PSn00bSDK GPU library
+ * (C) 2019-2022 Lameguy64, spicyjpeg - MPL licensed
+ */
+
 #ifndef __PSXGPU_H
 #define __PSXGPU_H
 
-#include <sys/types.h>
-
-// Low-level display parameters for DISPENV_RAW. A leftover from prototyping
-#define DISP_WIDTH_256		0
-#define DISP_WIDTH_320		1
-#define DISP_WIDTH_384		64
-#define DISP_WIDTH_512		2
-#define DISP_WIDTH_640		3
-
-#define DISP_HEIGHT_LOW		0	// Could be 240 for NTSC, 256 for PAL
-#define DISP_HEIGHT_HIGH	4	// Could be 480 for NTSC, 512 for PAL
-#define DISP_INTERLACE		32
-#define DISP_24BIT_COLOR	16
-#define DISP_MODE_NTSC		0
-#define DISP_MODE_PAL		8
+#include <stdint.h>
+#include <stddef.h>
 
+/* Definitions */
 
-#define MODE_NTSC			0
-#define MODE_PAL			1
+typedef enum _GPU_DispFlags {
+	DISP_WIDTH_256		= 0,
+	DISP_WIDTH_320		= 1,
+	DISP_WIDTH_512		= 2,
+	DISP_WIDTH_640		= 3,
+	DISP_HEIGHT_HIGH	= 1 << 2,
+	DISP_MODE_PAL		= 1 << 3,
+	DISP_24BIT_COLOR	= 1 << 4,
+	DISP_INTERLACE		= 1 << 5,
+	DISP_WIDTH_384		= 1 << 6
+} GPU_DispFlags;
 
+typedef enum _GPU_VideoMode {
+	MODE_NTSC	= 0,
+	MODE_PAL	= 1
+} GPU_VideoMode;
 
-// Vector macros
+/* Structure macros */
 
-#define setVector( v, _x, _y, _z ) \
+#define setVector(v, _x, _y, _z) \
 	(v)->vx = _x, (v)->vy = _y, (v)->vz = _z
 
-#define setRECT( v, _x, _y, _w, _h ) \
+#define setRECT(v, _x, _y, _w, _h) \
 	(v)->x = _x, (v)->y = _y, (v)->w = _w, (v)->h = _h
 
+#define setTPage(p, tp, abr, x, y)	((p)->tpage = getTPage(tp, abr, x, y))
+#define setClut(p, x, y)			((p)->clut = getClut(x, y))
 
-// Primitive macros	
-
-#define setDrawTPage( p, dfe, dtd, tpage ) \
-	( (p)->code[0] = tpage|(dfe<<10)|(dtd<<9), \
-	setlen( p, 1 ), setcode( p, 0xe1 ) )
-	
-/*#define setVram2Vram( p ) ( setlen( p, 8 ), setcode( p, 0x80 ), \
-	(p)->nop[0] = 0, (p)->nop[1] = 0, (p)->nop[2] = 0, (p)->nop[3] = 0 )*/
-	
-/*
-	
-#define setTPagePri2( p, dth, tp, abr, x, y ) \
-	( (p)->code[0] = getTPage( tp, abr, x, y )|(dth<<9), \
-	setlen( p, 1 ), setcode( p, 0xe1 ) )*/
-
-/*
- *	Set primitive attributes
- */
-#define setTPage( p, tp, abr, x, y ) \
-	( (p)->tpage = getTPage( tp, abr, x, y ) )
-
-#define setClut( p, x, y ) \
-	( (p)->clut = getClut(x, y) )
-	
-
-/*
- *	Set primitive colors
- */
-#define setRGB0( p, r, g, b ) ( (p)->r0 = r, (p)->g0 = g, (p)->b0 = b )
-#define setRGB1( p, r, g, b ) ( (p)->r1 = r, (p)->g1 = g, (p)->b1 = b )
-#define setRGB2( p, r, g, b ) ( (p)->r2 = r, (p)->g2 = g, (p)->b2 = b )
-#define setRGB3( p, r, g, b ) ( (p)->r3 = r, (p)->g3 = g, (p)->b3 = b )
-
+#define setRGB0(p, r, g, b) ((p)->r0 = r, (p)->g0 = g, (p)->b0 = b)
+#define setRGB1(p, r, g, b) ((p)->r1 = r, (p)->g1 = g, (p)->b1 = b)
+#define setRGB2(p, r, g, b) ((p)->r2 = r, (p)->g2 = g, (p)->b2 = b)
+#define setRGB3(p, r, g, b) ((p)->r3 = r, (p)->g3 = g, (p)->b3 = b)
 
-/*
- *	Set primitive screen coordinates
- */
-#define setXY0( p, _x0, _y0 ) \
+#define setXY0(p, _x0, _y0) \
 	(p)->x0 = _x0, (p)->y0 = _y0
 
-#define setXY2( p, _x0, _y0, _x1, _y1 ) \
+#define setXY2(p, _x0, _y0, _x1, _y1) \
 	(p)->x0 = _x0, (p)->y0 = _y0, \
 	(p)->x1 = _x1, (p)->y1 = _y1
-	
-#define setXY3( p, _x0, _y0, _x1, _y1, _x2, _y2 ) \
+
+#define setXY3(p, _x0, _y0, _x1, _y1, _x2, _y2) \
 	(p)->x0 = _x0, (p)->y0 = _y0, \
 	(p)->x1 = _x1, (p)->y1 = _y1, \
 	(p)->x2 = _x2, (p)->y2 = _y2
 
-#define setXY4( p, _x0, _y0, _x1, _y1, _x2, _y2, _x3, _y3 ) \
+#define setXY4(p, _x0, _y0, _x1, _y1, _x2, _y2, _x3, _y3) \
 	(p)->x0 = _x0, (p)->y0 = _y0, \
 	(p)->x1 = _x1, (p)->y1 = _y1, \
 	(p)->x2 = _x2, (p)->y2 = _y2, \
 	(p)->x3 = _x3, (p)->y3 = _y3
 
-#define setWH( p, _w, _h ) \
+#define setWH(p, _w, _h) \
 	(p)->w = _w, (p)->h = _h
 
-/*
- *	Set texture coordinates
- */
-#define setUV0( p, _u0, _v0 ) \
+#define setXYWH(p, _x0, _y0, _w, _h) \
+	(p)->x0 = _x0,			(p)->y0 = _y0, \
+	(p)->x1 = (_x0 + (_w)),	(p)->y1 = _y0, \
+	(p)->x2 = _x0,			(p)->y2 = (_y0 + (_h)), \
+	(p)->x3 = (_x0 + (_w)),	(p)->y3 = (_y0 + (_h))
+
+#define setUV0(p, _u0, _v0) \
 	(p)->u0 = _u0, (p)->v0 = _v0
 
-#define setUV3( p, _u0, _v0, _u1, _v1, _u2, _v2 ) \
-	(p)->u0 = _u0, (p)->v0 = _v0,	\
-	(p)->u1 = _u1, (p)->v1 = _v1,	\
+#define setUV3(p, _u0, _v0, _u1, _v1, _u2, _v2) \
+	(p)->u0 = _u0, (p)->v0 = _v0, \
+	(p)->u1 = _u1, (p)->v1 = _v1, \
 	(p)->u2 = _u2, (p)->v2 = _v2
 	
-#define setUV4( p, _u0, _v0, _u1, _v1, _u2, _v2, _u3, _v3 ) \
-	(p)->u0 = _u0, (p)->v0 = _v0,	\
-	(p)->u1 = _u1, (p)->v1 = _v1,	\
-	(p)->u2 = _u2, (p)->v2 = _v2,	\
+#define setUV4(p, _u0, _v0, _u1, _v1, _u2, _v2, _u3, _v3) \
+	(p)->u0 = _u0, (p)->v0 = _v0, \
+	(p)->u1 = _u1, (p)->v1 = _v1, \
+	(p)->u2 = _u2, (p)->v2 = _v2, \
 	(p)->u3 = _u3, (p)->v3 = _v3
 
-#define setUVWH( p, _u0, _v0, _w, _h ) 		\
-	(p)->u0 = _u0, 		(p)->v0 = _v0,		\
-	(p)->u1 = _u0+(_w),	(p)->v1 = _v0,		\
-	(p)->u2 = _u0, 		(p)->v2 = _v0+(_h),	\
-	(p)->u3 = _u0+(_w), (p)->v3 = _v0+(_h)
-
-
-/*
- *	Primitive handling macros
- */
-#define setlen( p, _len )		( ((P_TAG*)(p))->len = (unsigned char)(_len) )
-#define setaddr( p, _addr )		( ((P_TAG*)(p))->addr = (unsigned int)(_addr) )
-#define setcode( p, _code )		( ((P_TAG*)(p))->code = (unsigned char)(_code) )
-
-#define getlen( p )				( ((P_TAG*)(p))->len )
-#define getaddr( p )			( ((P_TAG*)(p))->addr )
-#define getcode( p )			( ((P_TAG*)(p))->code )
-
-#define nextPrim( p )			(void*)((((P_TAG*)(p))->addr)|0x80000000)
-#define isendprim( p )			((((P_TAG*)(p))->addr)==0xffffff)
-
-#define addPrim( ot, p )		setaddr( p, getaddr( ot ) ), setaddr( ot, p )
-#define addPrims( ot, p0, p1 )	setaddr( p1, getaddr( ot ) ), setaddr( ot, p0 )
-
-#define catPrim( p0, p1 )		setaddr( p0, p1 )
-#define termPrim( p )			setaddr( p, 0xffffff )
-
-#define setSemiTrans( p, abe ) \
-	( (abe)?setcode( p, getcode( p )|0x2 ):setcode( p, getcode( p )&~0x2 ) )
-
-#define setShadeTex( p, tge ) \
-	( (tge)?setcode( p, getcode( p )|0x1 ):setcode( p, getcode( p )&~0x1 ) )
-
-	
-/* ORIGINAl CODE */
-#define setDrawMask( p, sb, mt ) \
-	setlen( p, 1 ), (p)->code[0] = sb|(mt<<1), \
-	setcode( p, 0xe6 )
-	
-	
-#define getTPage( tp, abr, x, y ) \
-	( (((x)&0x3ff)>>6) | (((y)>>8)<<4) | (((abr)&0x3)<<5) | (((tp)&0x3)<<7) )
-
-#define getClut( x, y ) \
-	( ((y)<<6)|(((x)>>4)&0x3f) )
-
-
-/*
- *	Primitive initializers
- */ 
-#define setPolyF3( p )	setlen( p, 4 ), 	setcode( p, 0x20 )
-#define setPolyFT3( p )	setlen( p, 7 ), 	setcode( p, 0x24 )
-#define setPolyG3( p )	setlen( p, 6 ), 	setcode( p, 0x30 )
-#define setPolyGT3( p )	setlen( p, 9 ), 	setcode( p, 0x34 )
-
-#define setPolyF4( p )	setlen( p, 5 ), 	setcode( p, 0x28 )
-#define setPolyFT4( p )	setlen( p, 9 ), 	setcode( p, 0x2c )
-#define setPolyG4( p )	setlen( p, 8 ), 	setcode( p, 0x38 )
-#define setPolyGT4( p )	setlen( p, 12 ),	setcode( p, 0x3c )
-
-#define setSprt8( p )	setlen( p, 3 ), 	setcode( p, 0x74 )
-#define setSprt16( p )	setlen( p, 3 ), 	setcode( p, 0x7c )
-#define setSprt( p )	setlen( p, 4 ), 	setcode( p, 0x64 )
-
-#define setTile1( p )	setlen( p, 2 ), 	setcode( p, 0x68 )
-#define setTile8( p )	setlen( p, 2 ), 	setcode( p, 0x70 )
-#define setTile16( p )	setlen( p, 2 ), 	setcode( p, 0x78 )
-#define setTile( p )	setlen( p, 3 ), 	setcode( p, 0x60 )
-
-#define setLineF2( p )	setlen( p, 3 ),	setcode( p, 0x40 )
-#define setLineG2( p )	setlen( p, 4 ),	setcode( p, 0x50 )
-
-#define setLineF3( p )	setlen( p, 5 ), setcode( p, 0x48 ), (p)->pad = 0x55555555
-#define setLineG3( p )	setlen( p, 7 ), setcode( p, 0x58 ), (p)->pad = 0x55555555, \
-	(p)->p1 = 0, (p)->p2 = 0
-
-#define setLineF4( p )	setlen( p, 6 ), setcode( p, 0x4c ), (p)->pad = 0x55555555
-#define setLineG4( p )	setlen( p, 9 ), setcode( p, 0x5c ), (p)->pad = 0x55555555, \
-	(p)->p1 = 0, (p)->p2 = 0, (p)->p3 = 0
-			
-#define setFill( p ) 	setlen( p, 3 ), 	setcode( p, 0x02 )
-
-#define setDrawOffset( p, _x, _y ) \
-	setlen( p, 1 ), \
-	(p)->code[0] = (_x&0x3FF)|((_y&0x3FF)<<11), \
-	((char*)(p)->code)[3] = 0xE5
-	
-#define setDrawArea( p, r ) \
-	setlen( p, 2 ), \
-	(p)->code[0] = ((r)->x&0x3FF)|(((r)->y&0x1FF)<<10),	\
-	(p)->code[1] = (((r)->x+(r)->w-1)&0x3FF)|((((r)->y+(r)->h-1)&0x1FF)<<10), \
-	((char*)&(p)->code[0])[3] = 0xE3, \
-	((char*)&(p)->code[1])[3] = 0xE4
-
-#define setTexWindow( p, r ) \
-	setlen( p, 1 ), \
-	(p)->code[0] = ((r)->w&0x1F)|(((r)->h&0x1F)<<5)|(((r)->x&0x1F)<<10)|(((r)->y&0x1F)<<15), \
-	((char*)&(p)->code[0])[3] = 0xE2
-	
-/*
- *	Primitive definitions
- */
-typedef struct _P_TAG
-{
-	unsigned	addr:24;
-	unsigned	len:8;
-	u_char		r,g,b,code;
+#define setUVWH(p, _u0, _v0, _w, _h) \
+	(p)->u0 = _u0,			(p)->v0 = _v0, \
+	(p)->u1 = (_u0 + (_w)),	(p)->v1 = _v0, \
+	(p)->u2 = _u0,			(p)->v2 = (_v0 + (_h)), \
+	(p)->u3 = (_u0 + (_w)),	(p)->v3 = (_v0 + (_h))
+
+/* Primitive handling macros */
+
+#define setlen(p, _len)		(((P_TAG *) (p))->len = (uint8_t) (_len))
+#define setaddr(p, _addr)	(((P_TAG *) (p))->addr = (uint32_t) (_addr))
+#define setcode(p, _code)	(((P_TAG *) (p))->code = (uint8_t) (_code))
+#define getlen(p)			(((P_TAG *) (p))->len)
+#define getaddr(p)			(((P_TAG *) (p))->addr)
+#define getcode(p)			(((P_TAG *) (p))->code)
+
+#define nextPrim(p)			(void *) (0x80000000 | (((P_TAG *) (p))->addr))
+#define isendprim(p)		((((P_TAG *) (p))->addr) == 0xffffff)
+#define addPrim(ot, p)		setaddr(p, getaddr(ot)), setaddr(ot, p)
+#define addPrims(ot, a, b)	setaddr(b, getaddr(ot)), setaddr(ot, a)
+#define catPrim(a, b)		setaddr(a, b)
+#define termPrim(p)			setaddr(p, 0xffffff)
+
+#define setSemiTrans(p, abe) \
+	((abe) ? (getcode(p) |= 2) : (getcode(p) &= ~2))
+
+#define setShadeTex(p, tge) \
+	((tge) ? (getcode(p) |= 1) : (getcode(p) &= ~1))
+
+#define getTPage(tp, abr, x, y) \
+	((((x) & 0x3ff) >> 6) | (((y) >> 8) << 4) | (((abr) & 3) << 5) | (((tp) & 3) << 7))
+
+#define getClut(x, y) (((y) << 6) | (((x) >> 4) & 0x3f))
+
+/* Primitive initializer macros */
+
+#define setPolyF3(p)	setlen(p,  4), setcode(p, 0x20)
+#define setPolyFT3(p)	setlen(p,  7), setcode(p, 0x24)
+#define setPolyG3(p)	setlen(p,  6), setcode(p, 0x30)
+#define setPolyGT3(p)	setlen(p,  9), setcode(p, 0x34)
+#define setPolyF4(p)	setlen(p,  5), setcode(p, 0x28)
+#define setPolyFT4(p)	setlen(p,  9), setcode(p, 0x2c)
+#define setPolyG4(p)	setlen(p,  8), setcode(p, 0x38)
+#define setPolyGT4(p)	setlen(p, 12), setcode(p, 0x3c)
+#define setSprt8(p)		setlen(p,  3), setcode(p, 0x74)
+#define setSprt16(p)	setlen(p,  3), setcode(p, 0x7c)
+#define setSprt(p)		setlen(p,  4), setcode(p, 0x64)
+#define setTile1(p)		setlen(p,  2), setcode(p, 0x68)
+#define setTile8(p)		setlen(p,  2), setcode(p, 0x70)
+#define setTile16(p)	setlen(p,  2), setcode(p, 0x78)
+#define setTile(p)		setlen(p,  3), setcode(p, 0x60)
+#define setLineF2(p)	setlen(p,  3), setcode(p, 0x40)
+#define setLineG2(p)	setlen(p,  4), setcode(p, 0x50)
+#define setLineF3(p)	setlen(p,  5), setcode(p, 0x48), \
+	(p)->pad = 0x55555555
+#define setLineG3(p)	setlen(p,  7), setcode(p, 0x58), \
+	(p)->pad = 0x55555555, (p)->p1 = 0, (p)->p2 = 0
+#define setLineF4(p)	setlen(p,  6), setcode(p, 0x4c), \
+	(p)->pad = 0x55555555
+#define setLineG4(p)	setlen(p,  9), setcode(p, 0x5c), \
+	(p)->pad = 0x55555555, (p)->p1 = 0, (p)->p2 = 0, (p)->p3 = 0
+#define setFill(p) 		setlen(p,  3), setcode(p, 0x02)
+#define setVram2Vram(p)	setlen(p,  8), setcode(p, 0x80), \
+	(p)->pad[0] = 0, (p)->pad[1] = 0, (p)->pad[2] = 0, (p)->pad[3] = 0
+
+#define setDrawTPage(p, dfe, dtd, tpage) \
+	setlen(p, 1), \
+	(p)->code[0] = 0xe1000000 | tpage | (dfe << 10) | (dtd << 9)
+
+#define setDrawOffset(p, _x, _y) \
+	setlen(p, 1), \
+	(p)->code[0] = 0xe5000000 | (_x & 0x3ff) | ((_y & 0x3ff) << 11)
+
+#define setDrawMask(p, sb, mt) \
+	setlen(p, 1), \
+	(p)->code[0] = 0xe6000000 | sb | (mt << 1)
+
+#define setDrawArea(p, r) \
+	setlen(p, 2), \
+	(p)->code[0] = 0xe3000000 | ((r)->x & 0x3ff) | (((r)->y & 0x1ff) << 10), \
+	(p)->code[1] = 0xe4000000 | (((r)->x + (r)->w - 1) & 0x3ff) | ((((r)->y + (r)->h - 1) & 0x1ff) << 10)
+
+#define setTexWindow(p, r) \
+	setlen(p, 1), \
+	(p)->code[0] = 0xe2000000 | ((r)->w & 0x1f) | (((r)->h & 0x1f) << 5) | (((r)->x & 0x1f) << 10) | (((r)->y & 0x1f) << 15)
+
+/* Primitive structure definitions */
+
+typedef struct _P_TAG {
+	uint32_t	addr:24;
+	uint32_t	len:8;
+	uint8_t		r, g, b, code;
 } P_TAG;
 
-/*
- *	Polygon primitive definitions
- */
-typedef struct _POLY_F3
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	short		x1,y1;
-	short		x2,y2;
+typedef struct _POLY_F3 {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+	int16_t		x1, y1;
+	int16_t		x2, y2;
 } POLY_F3;
 
-typedef struct _POLY_F4
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	short		x1,y1;
-	short		x2,y2;
-	short		x3,y3;
+typedef struct _POLY_F4 {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+	int16_t		x1, y1;
+	int16_t		x2, y2;
+	int16_t		x3, y3;
 } POLY_F4;
 
-typedef struct _POLY_FT3
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	u_char		u0,v0;
-	u_short		clut;
-	short		x1,y1;
-	u_char		u1,v1;
-	u_short		tpage;
-	short		x2,y2;
-	u_char		u2,v2;
-	u_short		pad;
+typedef struct _POLY_FT3 {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+	uint8_t		u0, v0;
+	uint16_t	clut;
+	int16_t		x1, y1;
+	uint8_t		u1, v1;
+	uint16_t	tpage;
+	int16_t		x2, y2;
+	uint8_t		u2, v2;
+	uint16_t	pad;
 } POLY_FT3;
 
-typedef struct _POLY_FT4
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	u_short		x0,y0;
-	u_char		u0,v0;
-	u_short		clut;
-	short		x1,y1;
-	u_char		u1,v1;
-	u_short		tpage;
-	short		x2,y2;
-	u_char		u2,v2;
-	u_short		pad0;
-	short		x3,y3;
-	u_char		u3,v3;
-	u_short		pad1;
+typedef struct _POLY_FT4 {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	uint16_t	x0, y0;
+	uint8_t		u0, v0;
+	uint16_t	clut;
+	int16_t		x1, y1;
+	uint8_t		u1, v1;
+	uint16_t	tpage;
+	int16_t		x2, y2;
+	uint8_t		u2, v2;
+	uint16_t	pad0;
+	int16_t		x3, y3;
+	uint8_t		u3, v3;
+	uint16_t	pad1;
 } POLY_FT4;
 
-typedef struct _POLY_G3
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	u_char		r1,g1,b1,pad0;
-	short		x1,y1;
-	u_char		r2,g2,b2,pad1;
-	short		x2,y2;
+typedef struct _POLY_G3 {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+	uint8_t		r1, g1, b1, pad0;
+	int16_t		x1, y1;
+	uint8_t		r2, g2, b2, pad1;
+	int16_t		x2, y2;
 } POLY_G3;
 
-typedef struct _POLY_G4
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	u_char		r1,g1,b1,pad0;
-	short		x1,y1;
-	u_char		r2,g2,b2,pad1;
-	short		x2,y2;
-	u_char		r3,g3,b3,pad2;
-	short		x3,y3;
+typedef struct _POLY_G4 {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+	uint8_t		r1, g1, b1, pad0;
+	int16_t		x1, y1;
+	uint8_t		r2, g2, b2, pad1;
+	int16_t		x2, y2;
+	uint8_t		r3, g3, b3, pad2;
+	int16_t		x3, y3;
 } POLY_G4;
 
-typedef struct _POLY_GT3
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	u_char		u0,v0;
-	u_short		clut;
-	u_char		r1,g1,b1,pad0;
-	short		x1,y1;
-	u_char		u1,v1;
-	u_short		tpage;
-	u_char		r2,g2,b2,pad1;
-	short		x2,y2;
-	u_char		u2,v2;
-	u_short		pad2;
+typedef struct _POLY_GT3 {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+	uint8_t		u0, v0;
+	uint16_t	clut;
+	uint8_t		r1, g1, b1, pad0;
+	int16_t		x1, y1;
+	uint8_t		u1, v1;
+	uint16_t	tpage;
+	uint8_t		r2, g2, b2, pad1;
+	int16_t		x2, y2;
+	uint8_t		u2, v2;
+	uint16_t	pad2;
 } POLY_GT3;
 
-typedef struct _POLY_GT4
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	u_char		u0,v0;
-	u_short		clut;
-	u_char		r1,g1,b1,pad0;
-	short		x1,y1;
-	u_char		u1,v1;
-	u_short		tpage;
-	u_char		r2,g2,b2,pad1;
-	short		x2,y2;
-	u_char		u2,v2;
-	u_short		pad2;
-	u_char		r3,g3,b3,pad3;
-	short		x3,y3;
-	u_char		u3,v3;
-	u_short		pad4;
+typedef struct _POLY_GT4 {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+	uint8_t		u0, v0;
+	uint16_t	clut;
+	uint8_t		r1, g1, b1, pad0;
+	int16_t		x1, y1;
+	uint8_t		u1, v1;
+	uint16_t	tpage;
+	uint8_t		r2, g2, b2, pad1;
+	int16_t		x2, y2;
+	uint8_t		u2, v2;
+	uint16_t	pad2;
+	uint8_t		r3, g3, b3, pad3;
+	int16_t		x3, y3;
+	uint8_t		u3, v3;
+	uint16_t	pad4;
 } POLY_GT4;
 
-/*
- *	Line primitive definitions
- */
-typedef struct _LINE_F2
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	short		x1,y1;
+typedef struct _LINE_F2 {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+	int16_t		x1, y1;
 } LINE_F2;
 
-typedef struct _LINE_G2
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	u_char		r1,g1,b1,p1;
-	short		x1,y1;
+typedef struct _LINE_G2 {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+	uint8_t		r1, g1, b1, p1;
+	int16_t		x1, y1;
 } LINE_G2;
 
-typedef struct _LINE_F3
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	short		x1,y1;
-	short		x2,y2;
-	u_long		pad;			/* actually a terminator for line loops */
+typedef struct _LINE_F3 {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+	int16_t		x1, y1;
+	int16_t		x2, y2;
+	uint32_t	pad;
 } LINE_F3;
 
-typedef struct _LINE_G3
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	u_char		r1,g1,b1,p1;
-	short		x1,y1;
-	u_char		r2,g2,b2,p2;
-	short		x2,y2;
-	u_long		pad;			/* actually a terminator for line loops */
+typedef struct _LINE_G3 {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+	uint8_t		r1, g1, b1, p1;
+	int16_t		x1, y1;
+	uint8_t		r2, g2, b2, p2;
+	int16_t		x2, y2;
+	uint32_t	pad;
 } LINE_G3;
 
-typedef struct _LINE_F4
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	short		x1,y1;
-	short		x2,y2;
-	short		x3,y3;
-	u_long		pad;
+typedef struct _LINE_F4 {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+	int16_t		x1, y1;
+	int16_t		x2, y2;
+	int16_t		x3, y3;
+	uint32_t	pad;
 } LINE_F4;
 
-typedef struct _LINE_G4
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	u_char		r1,g1,b1,p1;
-	short		x1,y1;
-	u_char		r2,g2,b2,p2;
-	short		x2,y2;
-	u_char		r3,g3,b3,p3;
-	short		x3,y3;
-	u_long		pad;
+typedef struct _LINE_G4 {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+	uint8_t		r1, g1, b1, p1;
+	int16_t		x1, y1;
+	uint8_t		r2, g2, b2, p2;
+	int16_t		x2, y2;
+	uint8_t		r3, g3, b3, p3;
+	int16_t		x3, y3;
+	uint32_t	pad;
 } LINE_G4;
 
-/*
- *	Tile and sprite primitive definitions
- */
-typedef struct _TILE
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	short		w,h;
+typedef struct _TILE {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+	int16_t		w, h;
 } TILE;
 
-typedef struct _TILE_1
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-} TILE_1;
-
-typedef struct _TILE_8
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-} TILE_8;
-
-typedef struct _TILE_16
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-} TILE_16;
-
-/*
- *	Sprite primitive definitions
- */
-typedef struct _SPRT
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	u_char		u0,v0;
-	u_short		clut;
-	u_short		w,h;
+struct _TILE_FIXED {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+};
+typedef struct _TILE_FIXED TILE_1;
+typedef struct _TILE_FIXED TILE_8;
+typedef struct _TILE_FIXED TILE_16;
+
+typedef struct _SPRT {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+	uint8_t		u0, v0;
+	uint16_t	clut;
+	uint16_t	w, h;
 } SPRT;
 
-typedef struct _SPRT_8
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	u_char		u0,v0;
-	u_short		clut;
-} SPRT_8;
-
-typedef struct _SPRT_16
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	short		x0,y0;
-	u_char		u0,v0;
-	u_short		clut;
-} SPRT_16;
-
-/*
- *	VRAM fill and transfer primitive definitions
- */
-
-typedef struct _DR_ENV
-{
-	u_long		tag;
-	u_long		code[15];
+struct _SPRT_FIXED {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	int16_t		x0, y0;
+	uint8_t		u0, v0;
+	uint16_t	clut;
+};
+typedef struct _SPRT_FIXED SPRT_8;
+typedef struct _SPRT_FIXED SPRT_16;
+
+typedef struct _DR_ENV {
+	uint32_t tag;
+	uint32_t code[15];
 } DR_ENV;
 
-typedef struct _DR_AREA
-{
-	u_long		tag;
-	u_long		code[2];
+typedef struct _DR_AREA {
+	uint32_t tag;
+	uint32_t code[2];
 } DR_AREA;
 
-typedef struct _DR_OFFSET
-{
-	u_long		tag;
-	u_long		code[1];
+typedef struct _DR_OFFSET {
+	uint32_t tag;
+	uint32_t code[1];
 } DR_OFFSET;
 
-typedef struct _DR_TWIN
-{
-	u_long		tag;
-	u_long		code[2];
+typedef struct _DR_TWIN {
+	uint32_t tag;
+	uint32_t code[2];
 } DR_TWIN;
 
-typedef struct _DR_TPAGE
-{
-	u_long		tag;
-	u_long		code[1];
+typedef struct _DR_TPAGE {
+	uint32_t tag;
+	uint32_t code[1];
 } DR_TPAGE;
 
-typedef struct _DR_MASK		/* ORIGINAL */
-{
-	u_long		tag;
-	u_long		code[1];
+typedef struct _DR_MASK {
+	uint32_t tag;
+	uint32_t code[1];
 } DR_MASK;
 
-typedef struct _FILL		/* ORIGINAL */
-{
-	u_long		tag;
-	u_char		r0,g0,b0,code;
-	u_short		x0,y0;		// Note: coordinates must be in 16 pixel steps
-	u_short		w,h;
+typedef struct _FILL {
+	uint32_t	tag;
+	uint8_t		r0, g0, b0, code;
+	uint16_t	x0, y0; // Note: coordinates must be in 16 pixel steps
+	uint16_t	w, h;
 } FILL;
 
-typedef struct _VRAM2VRAM	/* ORIGINAL */
-{
-	u_long		tag;
-	u_char		p0,p1,p2,code;
-	u_short		x0,y0;
-	u_short		x1,y1;
-	u_short		w,h;
-	u_long		nop[4];
+typedef struct _VRAM2VRAM {
+	uint32_t	tag;
+	uint8_t		p0, p1, p2, code;
+	uint16_t	x0, y0;
+	uint16_t	x1, y1;
+	uint16_t	w, h;
+	uint32_t	pad[4];
 } VRAM2VRAM;
 
-/*
- *	General structs
- */
+/* Structure definitions */
 
-typedef struct _RECT
-{
-	short x,y;
-	short w,h;
+typedef struct _RECT {
+	int16_t x, y, w, h;
 } RECT;
 
-typedef struct _DISPENV_RAW	/* obsolete */
-{
-	unsigned int vid_mode;		// Video mode
-	short vid_xpos,vid_ypos;	// Video position (not framebuffer)
-	short fb_x,fb_y;			// Framebuffer display position	
+typedef struct _DISPENV_RAW {
+	uint32_t	vid_mode;
+	int16_t		vid_xpos, vid_ypos;
+	int16_t		fb_x, fb_y;
 } DISPENV_RAW;
 
-typedef struct _DISPENV
-{
-	RECT	disp;
-	RECT	screen;
-	char	isinter;
-	char	isrgb24;
-	char	reverse;
-	char	pad;
+typedef struct _DISPENV {
+	RECT		disp, screen;
+	uint8_t		isinter, isrgb24, reverse;
+	uint8_t		_reserved;
 } DISPENV;
 
-typedef struct _DRAWENV
-{
-	RECT	clip;			// Drawing area
-	short	ofs[2];			// GPU draw offset (relative to draw area)
-	RECT	tw;				// Texture window (doesn't do anything atm)
-	u_short	tpage;			// Initial tpage value
-	u_char	dtd;			// Dither processing flag (simply OR'ed to tpage)
-	u_char	dfe;			// Drawing to display area blocked/allowed (simply OR'ed to tpage)
-	u_char	isbg;			// Clear draw area if non-zero
-	u_char	r0,g0,b0;		// Draw area clear color (if isbg iz nonzero)
-	DR_ENV	dr_env;			// Draw mode packet area (used by PutDrawEnv)
+typedef struct _DRAWENV {
+	RECT		clip;		// Drawing area
+	int16_t		ofs[2];		// GPU draw offset (relative to draw area)
+	RECT		tw;			// Texture window (doesn't do anything atm)
+	uint16_t	tpage;		// Initial tpage value
+	uint8_t		dtd;		// Dither processing flag (simply OR'ed to tpage)
+	uint8_t		dfe;		// Drawing to display area blocked/allowed (simply OR'ed to tpage)
+	uint8_t		isbg;		// Clear draw area if non-zero
+	uint8_t		r0, g0, b0;	// Draw area clear color (if isbg iz nonzero)
+	DR_ENV		dr_env;		// Draw mode packet area (used by PutDrawEnv)
 } DRAWENV;
 
-typedef struct _TIM_IMAGE
-{
-	u_long	mode;
-	RECT	*crect;
-	u_long	*caddr;
-	RECT	*prect;
-	u_long	*paddr;
+typedef struct _TIM_IMAGE {
+	uint32_t	mode;
+	RECT		*crect;
+	uint32_t	*caddr;
+	RECT		*prect;
+	uint32_t	*paddr;
 } TIM_IMAGE;
 
+typedef struct _GsIMAGE {
+	uint32_t	pmode;
+	int16_t		px, py, pw, ph;
+	uint32_t	*pixel;
+	int16_t		cx, cy, cw, ch;
+	uint32_t	*clut;
+} GsIMAGE;
+
+/* Public API */
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-// Function definitions (asm)
-
 void ResetGraph(int mode);
 
-int GetVideoMode(void);
-void SetVideoMode(int mode);
-
-int GetODE(void);
-
-void PutDispEnvRaw(DISPENV_RAW *disp);	/* obsolete */
-void PutDispEnv(DISPENV *disp);
-void PutDrawEnv(DRAWENV *draw);
-
+GPU_VideoMode GetVideoMode(void);
+void SetVideoMode(GPU_VideoMode mode);
 void SetDispMask(int mask);
 
-int VSync(int m);
-int DrawSync(int m);
-void WaitGPUcmd(void);
-void WaitGPUdma(void);
+//void PutDispEnvRaw(const DISPENV_RAW *env);
+void PutDispEnv(const DISPENV *env);
+void PutDrawEnv(DRAWENV *env);
+void PutDrawEnvFast(DRAWENV *env);
 
-// Callback hook functions
+int GetODE(void);
+int VSync(int mode);
+int DrawSync(int mode);
+
+void *VSyncHaltFunction(void (*func)(void));
 void *VSyncCallback(void (*func)(void));
 void *DrawSyncCallback(void (*func)(void));
 
-void LoadImage(RECT *rect, u_long *data);
-void StoreImage(RECT *rect, u_long *data);
-
-void ClearOTagR(u_long* ot, int n);
-void DrawOTag(u_long* ot);
-void DrawPrim(void *pri);
-
-void AddPrim(u_long* ot, void* pri);
+void LoadImage(const RECT *rect, const uint32_t *data);
+void StoreImage(const RECT *rect, uint32_t *data);
 
-// Function definitions (C)
+void ClearOTagR(uint32_t *ot, size_t length);
+void ClearOTag(uint32_t *ot, size_t length);
+void DrawOTag(const uint32_t *ot);
+void DrawOTag2(const uint32_t *ot);
+void DrawOTagEnv(const uint32_t *ot, DRAWENV *env);
+void DrawPrim(const uint32_t *pri);
 
-int GetTimInfo(const u_long *tim, TIM_IMAGE *timimg);	/* ORIGINAL */
+void AddPrim(uint32_t *ot, const void *pri);
 
-DISPENV *SetDefDispEnv(DISPENV *disp, int x, int y, int w, int h);
-DRAWENV *SetDefDrawEnv(DRAWENV *draw, int x, int y, int w, int h);
+int GsGetTimInfo(const uint32_t *tim, GsIMAGE *info);
+int GetTimInfo(const uint32_t *tim, TIM_IMAGE *info);
 
-// Debug font functions
+DISPENV *SetDefDispEnv(DISPENV *env, int x, int y, int w, int h);
+DRAWENV *SetDefDrawEnv(DRAWENV *env, int x, int y, int w, int h);
 
 void FntLoad(int x, int y);
-char *FntSort(u_long *ot, char *pri, int x, int y, const char *text);
+char *FntSort(uint32_t *ot, char *pri, int x, int y, const char *text);
 int FntOpen(int x, int y, int w, int h, int isbg, int n);
 int FntPrint(int id, const char *fmt, ...);
 char *FntFlush(int id);
diff --git a/libpsn00b/include/psxgte.h b/libpsn00b/include/psxgte.h
index 7c10533..ddc988d 100644
--- a/libpsn00b/include/psxgte.h
+++ b/libpsn00b/include/psxgte.h
@@ -1,44 +1,45 @@
-#ifndef _PSXGTE_H
-#define _PSXGTE_H
+/*
+ * PSn00bSDK GTE library
+ * (C) 2019-2022 Lameguy64 - MPL licensed
+ */
 
+#ifndef __PSXGTE_H
+#define __PSXGTE_H
 
-#define ONE		4096
+#include <stdint.h>
 
+#define ONE (1 << 12)
 
-// For compatibility with official library syntax
-#define csin(a) isin(a)
-#define ccos(a) icos(a)
-#define rsin(a) isin(a)
-#define rcos(a) icos(a)
-
+/* Structure definitions */
 
-typedef struct MATRIX {
-	short	m[3][3];
-	int		t[3];
+typedef struct _MATRIX {
+	int16_t m[3][3];
+	int32_t t[3];
 } MATRIX;
 
-typedef struct VECTOR {
-	int		vx, vy, vz;
+typedef struct _VECTOR {
+	int32_t vx, vy, vz;
 } VECTOR;
 
-typedef struct SVECTOR {
-	short	vx, vy, vz, pad;
+typedef struct _SVECTOR {
+	int16_t vx, vy, vz, pad;
 } SVECTOR;
 
-typedef struct CVECTOR {
-	unsigned char r, g, b, cd;
+typedef struct _CVECTOR {
+	uint8_t r, g, b, cd;
 } CVECTOR;
 
-typedef struct DVECTOR {
-	short vx, vy;
+typedef struct _DVECTOR {
+	int16_t vx, vy;
 } DVECTOR;
 
+/* Public API */
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-void InitGeom();
+void InitGeom(void);
 
 // Integer SIN/COS functions (4096 = 360 degrees)
 // Does not use tables!
@@ -70,11 +71,16 @@ void VectorNormalS(VECTOR *v0, SVECTOR *v1);
 
 void Square0(VECTOR *v0, VECTOR *v1);
 
-int SquareRoot12( int v );
-int SquareRoot0( int v );
+int SquareRoot12(int v);
+int SquareRoot0(int v);
+
+#define csin(a) isin(a)
+#define ccos(a) icos(a)
+#define rsin(a) isin(a)
+#define rcos(a) icos(a)
 
 #ifdef __cplusplus
 }
 #endif
 
-#endif // _PSXGTE_H
+#endif
diff --git a/libpsn00b/include/psxpress.h b/libpsn00b/include/psxpress.h
index b060170..2106a53 100644
--- a/libpsn00b/include/psxpress.h
+++ b/libpsn00b/include/psxpress.h
@@ -17,6 +17,28 @@ typedef struct _DECDCTENV {
 	int16_t dct[64];	// Inverse DCT matrix (2.14 fixed-point)
 } DECDCTENV;
 
+// This is the "small" lookup table used by DecDCTvlc(). It can be copied to
+// the scratchpad.
+typedef struct _DECDCTTAB {
+	uint16_t	lut0[2];
+	uint32_t	lut2[8];
+	uint32_t	lut3[64];
+	uint16_t	lut4[8];
+	uint16_t	lut5[8];
+	uint16_t	lut7[16];
+	uint16_t	lut8[32];
+	uint16_t	lut9[32];
+	uint16_t	lut10[32];
+	uint16_t	lut11[32];
+	uint16_t	lut12[32];
+} DECDCTTAB;
+
+// This is the "large" table used by DecDCTvlc2().
+typedef struct _DECDCTTAB2 {
+	uint32_t	lut[8192];
+	uint32_t	lut00[512];
+} DECDCTTAB2;
+
 typedef enum _DECDCTMODE {
 	DECDCT_MODE_24BPP		= 1,
 	DECDCT_MODE_16BPP		= 0,
@@ -24,6 +46,23 @@ typedef enum _DECDCTMODE {
 	DECDCT_MODE_RAW			= -1
 } DECDCTMODE;
 
+typedef struct _VLC_Context {
+	const uint32_t	*input;
+	uint32_t		window, next_window, remaining;
+	uint16_t		quant_scale;
+	int8_t			is_v3, bit_offset, block_index, coeff_index;
+} VLC_Context;
+
+// Despite what some docs claim, the "number of 32-byte blocks" and "always
+// 0x3800" fields are actually a single 32-bit field which is copied over to
+// the output buffer, then parsed by DecDCTin() and written to the MDEC0
+// register.
+typedef struct {
+	uint32_t mdec0_header;
+	uint16_t quant_scale;
+	uint16_t version;
+} BS_Header;
+
 /* Public API */
 
 #ifdef __cplusplus
@@ -40,7 +79,7 @@ extern "C" {
  *
  * @param mode
  */
-void DecDCTReset(int32_t mode);
+void DecDCTReset(int mode);
 
 /**
  * @brief Uploads the specified decoding environment's quantization tables and
@@ -61,15 +100,15 @@ void DecDCTReset(int32_t mode);
  * @param env Pointer to DECDCTENV or 0 for default tables
  * @param mono 0 for color (normal), 1 for monochrome
  */
-void DecDCTPutEnv(const DECDCTENV *env, int32_t mono);
+void DecDCTPutEnv(const DECDCTENV *env, int mono);
 
 /**
- * @brief Sets up the MDEC to start fetching and decoding a stream from the
- * given address in main RAM. The first 32-bit word is initially copied to the
- * MDEC0 register, then all subsequent data is read in 128-byte (32-word)
- * chunks. The length of the stream (in 32-bit units, minus the first word)
- * must be encoded in the lower 16 bits of the first word, as expected by the
- * MDEC.
+ * @brief Sets up the MDEC to start fetching and decoding the given buffer.
+ * This function is meant to be used with buffers generated by DecDCTvlc(): the
+ * first 32-bit word of the buffer is initially copied to the MDEC0 register,
+ * then all subsequent data is read in 128-byte (32-word) chunks. The length of
+ * the stream (in 32-bit units, minus the first word) is encoded by DecDCTvlc()
+ * in the lower 16 bits of the first word.
  *
  * The mode argument optionally specifies the output color depth (0 for 16bpp,
  * 1 for 24bpp) if not already set in the first word. Passing -1 will result in
@@ -79,7 +118,7 @@ void DecDCTPutEnv(const DECDCTENV *env, int32_t mono);
  * @param data
  * @param mode DECDCT_MODE_* or -1
  */
-void DecDCTin(const uint32_t *data, int32_t mode);
+void DecDCTin(const uint32_t *data, int mode);
 
 /**
  * @brief Configures the MDEC to automatically fetch data (the input stream,
@@ -111,9 +150,9 @@ void DecDCTinRaw(const uint32_t *data, size_t length);
  * stream (usually a whole frame) is being written to main RAM.
  *
  * @param mode
- * @return 0 or -1 in case of a timeout (mode = 0) / MDEC busy flag (mode = 1)
+ * @return 0 or -1 in case of a timeout (mode = 0), MDEC busy flag (mode = 1)
  */
-int32_t DecDCTinSync(int32_t mode);
+int DecDCTinSync(int mode);
 
 /**
  * @brief Configures the MDEC to automatically transfer decoded image data in
@@ -142,9 +181,204 @@ void DecDCTout(uint32_t *data, size_t length);
  * to register a callback that calls DecDCTin() to feed the MDEC.
  *
  * @param mode
- * @return 0 or -1 in case of a timeout (mode = 0) / DMA busy flag (mode = 1)
+ * @return 0 or -1 in case of a timeout (mode = 0), DMA busy flag (mode = 1)
+ */
+int DecDCToutSync(int mode);
+
+/**
+ * @brief Begins decompressing the contents of a .BS file (or of a single .STR
+ * frame) into a buffer that can be passed to DecDCTin(). This function uses a
+ * small (<1 KB) lookup table combined with the GTE to accelerate the process;
+ * performance is roughly on par with DecDCTvlcStart2() if the lookup table
+ * is copied to the scratchpad beforehand by calling DecDCTvlcCopyTable(). The
+ * contents of the GTE's LZCR register, if any, will be destroyed.
+ *
+ * A VLC_Context object must be created and passed to this function, which will
+ * then proceed to initialize its fields. The max_size argument sets the
+ * maximum number of words that will be written to the output buffer; if more
+ * data needs to be written, this function will return 1. To continue decoding
+ * call DecDCTvlcContinue() with the same VLC_Context object (the output buffer
+ * can be different). If max_size = 0, the entire frame will always be decoded
+ * in one shot.
+ *
+ * Only bitstream version 2 is currently supported.
+ *
+ * WARNING: InitGeom() must be called prior to using DecDCTvlcStart() for the
+ * first time. Attempting to call this function with the GTE disabled will
+ * result in a crash.
+ *
+ * @param ctx Pointer to VLC_Context structure (which will be initialized)
+ * @param buf
+ * @param max_size Maximum number of 32-bit words to output
+ * @param bs
+ * @return 0, 1 if more data needs to be output or -1 in case of failure
+ */
+int DecDCTvlcStart(VLC_Context *ctx, uint32_t *buf, size_t max_size, const uint32_t *bs);
+
+/**
+ * @brief Resumes the decompression process started by DecDCTvlcStart(). The
+ * state of the decompressor is contained entirely in the VLC_Context structure
+ * so an arbitrary number of bitstreams can be decoded concurrently (although
+ * the limited CPU power makes it impractical to do so) by keeping a separate
+ * context for each bitstream.
+ *
+ * This function behaves like DecDCTvlcStart(), returning 1 if more data has to
+ * be written or 0 otherwise. DecDCTvlcContinue() shall not be called after a
+ * previous call to DecDCTvlcStart() or DecDCTvlcContinue() with the same
+ * context returned 0; in that case the context shall be discarded or reused to
+ * decode another bitstream.
+ *
+ * The contents of the GTE's LZCR register, if any, will be destroyed.
+ *
+ * See DecDCTvlcStart() for more details.
+ *
+ * @param ctx Pointer to already initialized VLC_Context structure
+ * @param buf
+ * @param max_size Maximum number of 32-bit words to output
+ * @return 0, 1 if more data needs to be output or -1 in case of failure
+ */
+int DecDCTvlcContinue(VLC_Context *ctx, uint32_t *buf, size_t max_size);
+
+/**
+ * A wrapper around DecDCTvlcStart() and DecDCTvlcContinue() for compatibility
+ * with the official SDK. This function uses an internal context; additionally,
+ * the maximum output buffer size is not passed as an argument but is instead
+ * set by calling DecDCTvlcSize().
+ *
+ * This function behaves identically to DecDCTvlcContinue() if bs = 0 and
+ * DecDCTvlcStart() otherwise.
+ *
+ * See DecDCTvlcStart() for more details.
+ *
+ * WARNING: InitGeom() must be called prior to using DecDCTvlc() for the first
+ * time. Attempting to call this function with the GTE disabled will result in
+ * a crash.
+ *
+ * @param bs Pointer to bitstream data or 0 to resume decoding
+ * @param buf
+ * @return 0, 1 if more data needs to be output or -1 in case of failure
+ */
+int DecDCTvlc(const uint32_t *bs, uint32_t *buf);
+
+/**
+ * @brief Sets the maximum number of 32-bit words that a single call to
+ * DecDCTvlc() will output. If size = 0, the entire frame will always be
+ * decoded in one shot.
+ *
+ * @param size  Maximum number of 32-bit words to output
+ * @return Previously set value 
+ */
+size_t DecDCTvlcSize(size_t size);
+
+/**
+ * @brief Copies the small (<1 KB) lookup table used by DecDCTvlcContinue(),
+ * DecDCTvlcStart() and DecDCTvlc() (a DECDCTTAB structure) to the specified
+ * address. A copy of this table is always present in main RAM, however this
+ * function can be used to copy it to the scratchpad region to boost
+ * decompression performance.
+ *
+ * The address passed to this function is saved. Calls to DecDCTvlcStart(),
+ * DecDCTvlcContinue() and DecDCTvlc() will automatically use the last table
+ * copied. Call DecDCTvlcCopyTable(0) to revert to using the library's internal
+ * table in main RAM.
+ *
+ * @param addr Pointer to free area in scratchpad region or 0 to reset
+ */
+void DecDCTvlcCopyTable(DECDCTTAB *addr);
+
+/**
+ * @brief Begins decompressing the contents of a .BS file (or of a single .STR
+ * frame) into a buffer that can be passed to DecDCTin(). This function uses a
+ * large (34 KB) lookup table that must be loaded into main RAM beforehand by
+ * calling DecDCTvlcBuild(), but does not use the GTE nor the scratchpad.
+ * Depending on the specific bitstream being decoded DecDCTvlcStart2() might be
+ * slightly faster or slower than DecDCTvlcStart() with its lookup table copied
+ * to the scratchpad (see DecDCTvlcCopyTable()). DecDCTvlcStart() with the
+ * table in main RAM tends to be much slower.
+ *
+ * A VLC_Context object must be created and passed to this function, which will
+ * then proceed to initialize its fields. The max_size argument sets the
+ * maximum number of words that will be written to the output buffer; if more
+ * data needs to be written, this function will return 1. To continue decoding
+ * call DecDCTvlcContinue2() with the same VLC_Context object (the output
+ * buffer can be different). If max_size = 0, the entire frame will always be
+ * decoded in one shot.
+ *
+ * Only bitstream version 2 is currently supported.
+ *
+ * @param ctx Pointer to VLC_Context structure (which will be initialized)
+ * @param buf
+ * @param max_size Maximum number of 32-bit words to output
+ * @param bs
+ * @return 0, 1 if more data needs to be output or -1 in case of failure
+ */
+int DecDCTvlcStart2(VLC_Context *ctx, uint32_t *buf, size_t max_size, const uint32_t *bs);
+
+/**
+ * @brief Resumes the decompression process started by DecDCTvlcStart2(). The
+ * state of the decompressor is contained entirely in the VLC_Context structure
+ * so an arbitrary number of bitstreams can be decoded concurrently (although
+ * the limited CPU power makes it impractical to do so) by keeping a separate
+ * context for each bitstream.
+ *
+ * This function behaves like DecDCTvlcStart2(), returning 1 if more data has
+ * to be written or 0 otherwise. DecDCTvlcContinue2() shall not be called after
+ * a previous call to DecDCTvlcStart2() or DecDCTvlcContinue2() with the same
+ * context returned 0; in that case the context shall be discarded or reused to
+ * decode another bitstream.
+ *
+ * See DecDCTvlcStart2() for more details.
+ *
+ * @param ctx Pointer to already initialized VLC_Context structure
+ * @param buf
+ * @param max_size Maximum number of 32-bit words to output
+ * @return 0, 1 if more data needs to be output or -1 in case of failure
+ */
+int DecDCTvlcContinue2(VLC_Context *ctx, uint32_t *buf, size_t max_size);
+
+/**
+ * A wrapper around DecDCTvlcStart2() and DecDCTvlcContinue2() for
+ * compatibility with the official SDK. This function uses an internal context;
+ * additionally, the maximum output buffer size is not passed as an argument
+ * but is instead set by calling DecDCTvlcSize2().
+ *
+ * This function behaves identically to DecDCTvlcContinue() if bs = 0 and
+ * DecDCTvlcStart() otherwise. The table argument can optionally be passed to
+ * use a custom lookup table. If zero, the last pointer passed to
+ * DecDCTvlcBuild() will be used.
+ *
+ * See DecDCTvlcStart2() for more details.
+ *
+ * @param bs Pointer to bitstream data or 0 to resume decoding
+ * @param buf
+ * @param table Pointer to decompressed table or 0 to use last table used
+ * @return 0, 1 if more data needs to be output or -1 in case of failure
+ */
+int DecDCTvlc2(const uint32_t *bs, uint32_t *buf, DECDCTTAB2 *table);
+
+/**
+ * @brief Sets the maximum number of 32-bit words that a single call to
+ * DecDCTvlc2() will output. If size = 0, the entire frame will always be
+ * decoded in one shot.
+ *
+ * @param size Maximum number of 32-bit words to output
+ * @return Previously set value 
+ */
+size_t DecDCTvlcSize2(size_t size);
+
+/**
+ * @brief Generates the lookup table required by DecDCTvlcStart2(),
+ * DecDCTvlcContinue2() and DecDCTvlc2() (a DECDCTTAB2 structure) into the
+ * specified buffer. Since the table is relatively large (34 KB), it is
+ * recommended to only generate it in a dynamically-allocated buffer when
+ * needed and deallocate the buffer afterwards.
+ *
+ * The address passed to this function is saved. Calls to DecDCTvlcStart2() and
+ * DecDCTvlcContinue2() will automatically use the last table decompressed.
+ *
+ * @param table
  */
-int32_t DecDCToutSync(int32_t mode);
+void DecDCTvlcBuild(DECDCTTAB2 *table);
 
 #ifdef __cplusplus
 }
diff --git a/libpsn00b/include/psxspu.h b/libpsn00b/include/psxspu.h
index da000e3..cf78e3d 100644
--- a/libpsn00b/include/psxspu.h
+++ b/libpsn00b/include/psxspu.h
@@ -1,138 +1,124 @@
+/*
+ * PSn00bSDK SPU library
+ * (C) 2019-2022 Lameguy64, spicyjpeg - MPL licensed
+ */
+
 #ifndef __PSXSPU_H
 #define __PSXSPU_H
 
-#include <sys/types.h>
-
-// Mask settings bits for specifying voice channels
-
-#define SPU_00CH	(1<<0)
-#define SPU_01CH	(1<<1)
-#define SPU_02CH	(1<<2)
-#define SPU_03CH	(1<<3)
-#define SPU_04CH	(1<<4)
-#define	SPU_05CH	(1<<5)
-#define SPU_06CH	(1<<6)
-#define SPU_07CH	(1<<7)
-#define SPU_08CH	(1<<8)
-#define SPU_09CH	(1<<9)
-#define SPU_10CH	(1<<10)
-#define SPU_11CH	(1<<11)
-#define SPU_12CH	(1<<12)
-#define SPU_13CH	(1<<13)
-#define SPU_14CH	(1<<14)
-#define SPU_15CH	(1<<15)
-#define SPU_16CH	(1<<16)
-#define SPU_17CH	(1<<17)
-#define SPU_18CH	(1<<18)
-#define SPU_19CH	(1<<19)
-#define SPU_20CH	(1<<20)
-#define SPU_21CH	(1<<21)
-#define SPU_22CH	(1<<22)
-#define SPU_23CH	(1<<23)
-
-#define SPU_0CH		SPU_00CH
-#define SPU_1CH		SPU_01CH
-#define SPU_2CH		SPU_02CH
-#define SPU_3CH		SPU_03CH
-#define SPU_4CH		SPU_04CH
-#define SPU_5CH		SPU_05CH
-#define SPU_6CH		SPU_06CH
-#define SPU_7CH		SPU_07CH
-#define SPU_8CH		SPU_08CH
-#define SPU_9CH		SPU_09CH
-
-#define SPU_KEYCH(x)	(1<<(x))
-#define SPU_VOICECH(x)	SPU_KEYCH(x)
-
-
-// Mask setting bits for SpuVoiceAttr.mask
-
-#define SPU_VOICE_VOLL			(1<<0)	// Left volume
-#define SPU_VOICE_VOLR			(1<<1)	// Right volume
-#define SPU_VOICE_VOLMODEL		(1<<2)	// Left volume mode
-#define SPU_VOICE_VOLMODER		(1<<3)	// Right volume mode
-#define SPU_VOICE_PITCH			(1<<4)	// Pitch tone
-#define SPU_VOICE_NOTE			(1<<5)	// Pitch note
-#define SPU_VOICE_SAMPLE_NOTE	(1<<6)	// Sample base frequency?
-#define SPU_VOICE_WDSA			(1<<7)	// Sample start address (in SPU RAM)
-#define SPU_VOICE_ADSR_AMODE	(1<<8)	// ADSR attack mode
-#define SPU_VOICE_ADSR_SMODE	(1<<9)	// ADSR sustain mode
-#define SPU_VOICE_ADSR_RMODE	(1<<10)	// ADSR release mode
-#define SPU_VOICE_ADSR_AR		(1<<11)	// ADSR attack rate
-#define SPU_VOICE_ADSR_DR		(1<<12)	// ADSR decay rate
-#define SPU_VOICE_ADSR_SR		(1<<13)	// ADSR sustain rate
-#define SPU_VOICE_ADSR_RR		(1<<14)	// ADSR release rate
-#define SPU_VOICE_ADSR_SL		(1<<15) // ADSR sustain level
-#define SPU_VOICE_LSAX			(1<<16)	// Loop start address (in SPU RAM)
-#define SPU_VOICE_ADSR_ADSR1	(1<<17)	// adsr1 for VagAtr (?)
-#define SPU_VOICE_ADSR_ADSR2	(1<<18) // adsr2 for VagAtr (?)
-
-
-#define	SPU_TRANSFER_BY_DMA		0
-
-
-typedef struct {
-	short left;
-	short right;
+#include <stdint.h>
+#include <stddef.h>
+#include <hwregs_c.h>
+
+/* Definitions */
+
+typedef enum _SPU_AttrMask {
+	SPU_VOICE_VOLL			= 1 << 0,	// Left volume
+	SPU_VOICE_VOLR			= 1 << 1,	// Right volume
+	SPU_VOICE_VOLMODEL		= 1 << 2,	// Left volume mode
+	SPU_VOICE_VOLMODER		= 1 << 3,	// Right volume mode
+	SPU_VOICE_PITCH			= 1 << 4,	// Pitch tone
+	SPU_VOICE_NOTE			= 1 << 5,	// Pitch note
+	SPU_VOICE_SAMPLE_NOTE	= 1 << 6,	// Sample base frequency?
+	SPU_VOICE_WDSA			= 1 << 7,	// Sample start address (in SPU RAM)
+	SPU_VOICE_ADSR_AMODE	= 1 << 8,	// ADSR attack mode
+	SPU_VOICE_ADSR_SMODE	= 1 << 9,	// ADSR sustain mode
+	SPU_VOICE_ADSR_RMODE	= 1 << 10,	// ADSR release mode
+	SPU_VOICE_ADSR_AR		= 1 << 11,	// ADSR attack rate
+	SPU_VOICE_ADSR_DR		= 1 << 12,	// ADSR decay rate
+	SPU_VOICE_ADSR_SR		= 1 << 13,	// ADSR sustain rate
+	SPU_VOICE_ADSR_RR		= 1 << 14,	// ADSR release rate
+	SPU_VOICE_ADSR_SL		= 1 << 15,	// ADSR sustain level
+	SPU_VOICE_LSAX			= 1 << 16,	// Loop start address (in SPU RAM)
+	SPU_VOICE_ADSR_ADSR1	= 1 << 17,
+	SPU_VOICE_ADSR_ADSR2	= 1 << 18
+} SPU_AttrMask;
+
+typedef enum _SPU_TransferMode {
+	SPU_TRANSFER_BY_DMA	= 0,
+	SPU_TRANSFER_BY_IO	= 1
+} SPU_TransferMode;
+
+typedef enum _SPU_WaitMode {
+	SPU_TRANSFER_PEEK	= 0,
+	SPU_TRANSFER_WAIT	= 1
+} SPU_WaitMode;
+
+/* Structure definitions */
+
+typedef struct _SpuVolume {
+	int16_t left, right;
 } SpuVolume;
 
-typedef struct {
-	SpuVolume		vol;			// 0
-	unsigned short	freq;			// 4
-	unsigned short	addr;			// 6
-	unsigned short	loop_addr;		// 8
-	unsigned short	res;			// 10
-	unsigned int	adsr_param;		// 12
-} SpuVoiceRaw;
-
-typedef struct {
-	u_int voice;
-	u_int mask;
-	SpuVolume volume;
-	SpuVolume volmode;
-	SpuVolume volumex;
-	u_short	pitch;
-	u_short	note;
-	u_short	sample_note;
-	short envx;
-	u_int addr;
-	u_int loop_addr;
-	int	a_mode;
-	int	s_mode;
-	int	r_mode;
-	u_short	ar;
-	u_short	dr;
-	u_short	sr;
-	u_short	rr;
-	u_short	sl;
-	u_short	adsr1;
-	u_short	adsr2;
+typedef struct _SpuVoiceAttr {
+	uint32_t	voice;
+	uint32_t	mask;
+	SpuVolume	volume, volmode, volumex;
+	uint16_t	pitch, note, sample_note;
+	int16_t		envx;
+	uint32_t	addr, loop_addr;
+	int			a_mode, s_mode, r_mode;
+	uint16_t	ar, dr, sr, rr, sl;
+	uint16_t	adsr1, adsr2;
 } SpuVoiceAttr;
 
+typedef struct _SpuExtAttr {
+	SpuVolume	volume;
+	int			reverb, mix;
+} SpuExtAttr;
+
+typedef struct _SpuCommonAttr {
+	uint32_t	mask;
+	SpuVolume	mvol, mvolmode, mvolx;
+	SpuExtAttr	cd, ext;
+} SpuCommonAttr;
+
+/* "Useless" macros for official SDK compatibility */
+
+#define SpuSetCommonMasterVolume(left, right) \
+	(SPU_MASTER_VOL_L = (left), SPU_MASTER_VOL_R = (right))
+#define SpuSetCommonCDVolume(left, right) \
+	(SPU_CD_VOL_L = (left), SPU_CD_VOL_R = (right))
+#define SpuSetCommonCDReverb(enable) \
+	((enable) ? (SPU_CTRL |= 0x0004) : (SPU_CTRL &= 0xfffb))
+#define SpuSetCommonExtVolume(left, right) \
+	(SPU_EXT_VOL_L = (left), SPU_EXT_VOL_R = (right))
+#define SpuSetCommonExtReverb(enable) \
+	((enable) ? (SPU_CTRL |= 0x0002) : (SPU_CTRL &= 0xfffd))
+
+#define SpuSetReverbAddr(addr) \
+	(SPU_REVERB_ADDR = ((addr) + 7) / 8)
+#define SpuSetIRQAddr(addr) \
+	(SPU_IRQ_ADDR = ((addr) + 7) / 8)
+
+#define SpuSetVoiceVolume(ch, left, right) \
+	(SPU_CH_VOL_L(ch) = (left), SPU_CH_VOL_R(ch) = (right))
+#define SpuSetVoicePitch(ch, pitch) \
+	(SPU_CH_FREQ(ch) = (pitch))
+#define SpuSetVoiceStartAddr(ch, addr) \
+	(SPU_CH_ADDR(ch) = ((addr) + 7) / 8)
+#define SpuSetVoiceADSR(ch, ar, dr, sr, rr, sl) \
+	(SPU_CH_ADSR(ch) = ((sl)) | ((dr) << 4) | ((ar) << 8) | ((rr) << 16) | ((sr) << 22) | (1 << 30))
+
+#define SpuSetKey(enable, voice_bit) \
+	((enable) ? (SPU_KEY_ON = (voice_bit)) : (SPU_KEY_OFF = (voice_bit)))
+
+/* Public API */
+
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-void SpuInit();
-
-void SpuSetVoiceRaw( int voice, const SpuVoiceRaw* param );
-void SpuReverbOn( int voice );
-void SpuSetReverb();
-
-void SpuSetReverbAddr( int addr );
-void SpuSetReverbVolume( int left, int right );
+void SpuInit(void);
 
-
-void SpuSetKey(int on_off, u_int voice_bit);
-
-// SPU transfer functions
-int SpuSetTransferMode(int mode);
-int SpuSetTransferStartAddr(int addr);
-int SpuWrite(const unsigned char* addr, int size);
-void SpuWait();
+void SpuRead(uint32_t *data, size_t size);
+void SpuWrite(const uint32_t *data, size_t size);
+SPU_TransferMode SpuSetTransferMode(SPU_TransferMode mode);
+uint32_t SpuSetTransferStartAddr(uint32_t addr);
+int SpuIsTransferCompleted(int mode);
 
 #ifdef __cplusplus
 }
 #endif
 
-#endif
-\ No newline at end of file
+#endif
diff --git a/libpsn00b/include/stdlib.h b/libpsn00b/include/stdlib.h
index 4c4fcd3..1888c69 100644
--- a/libpsn00b/include/stdlib.h
+++ b/libpsn00b/include/stdlib.h
@@ -1,36 +1,27 @@
 /*
- * stdlib.h
- *
- * Standard library functions
- *
- * Inherited from PSXSDK
+ * PSn00bSDK standard library
+ * (C) 2019-2022 PSXSDK authors, Lameguy64, spicyjpeg - MPL licensed
  */
 
-#ifndef _STDLIB_H
-#define _STDLIB_H
+#ifndef __STDLIB_H
+#define __STDLIB_H
 
-#define RAND_MAX	0x7fff
+#include <stddef.h>
 
-/* Conversion functions (not yet implemented) */
+/* Definitions */
 
-/*
-extern int atoi(char *s);
-extern long atol(char *s);
-extern char atob(char *s); // Is this right?
-*/
-
-// Quick sort (not yet implemented)
+#define RAND_MAX 0x7fff
 
-//void qsort(void *base , int nel , int width , int (*cmp)(const void *,const void *));
+/* API */
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-extern int __argc;
-extern const char **__argv;
+extern int			__argc;
+extern const char	**__argv;
 
-int rand();
+int rand(void);
 void srand(unsigned long seed);
 
 int abs(int j);
@@ -39,16 +30,17 @@ long long strtoll(const char *nptr, char **endptr, int base);
 long strtol(const char *nptr, char **endptr, int base);
 long double strtold(const char *nptr, char **endptr);
 
-// Note: these use floats internally!
 double strtod(const char *nptr, char **endptr);
 float strtof(const char *nptr, char **endptr);
 
-// Memory allocation functions
-void _mem_init(int ram_size, int stack_max_size);
-void InitHeap(unsigned int *addr, int size);
-int SetHeapSize(int size);
-void *malloc(int size);
-void *calloc(int number, int size);
+void _mem_init(size_t ram_size, size_t stack_max_size);
+void InitHeap(void *addr, size_t size);
+//int SetHeapSize(size_t size);
+void *sbrk(ptrdiff_t incr);
+
+void *malloc(size_t size);
+void *calloc(size_t num, size_t size);
+void *realloc(void *ptr, size_t size);
 void free(void *ptr);
 
 #ifdef __cplusplus
@@ -56,4 +48,3 @@ void free(void *ptr);
 #endif
 
 #endif
-
diff --git a/libpsn00b/include/string.h b/libpsn00b/include/string.h
index 365d238..ceee066 100644
--- a/libpsn00b/include/string.h
+++ b/libpsn00b/include/string.h
@@ -1,13 +1,10 @@
 /*
- * string.h
- *
- * Prototypes for string functions of the C library
- *
- * PSXSDK
+ * PSn00bSDK standard library
+ * (C) 2019-2022 PSXSDK authors, Lameguy64, spicyjpeg - MPL licensed
  */
 
-#ifndef _STRING_H
-#define _STRING_H
+#ifndef __STRING_H
+#define __STRING_H
 
 #ifdef __cplusplus
 extern "C" {
@@ -19,11 +16,6 @@ char *strpbrk(const char *dst , const char *src);
 char *strtok(char *s , char *set);
 char *strstr(const char *big , const char *little);
 
-//int strspn(char *s , char *set);
-//int strcspn(char *s , char *set);
-//int index(char *s , int c);
-//int rindex(char *s , int c);
-
 char *strcat(char *s , const char *append);
 char *strncat(char *s , const char *append, int n);
 char *strcpy(char *dst , const char *src);
@@ -34,9 +26,6 @@ char *strrchr(const char *s , int c);
 
 void *memmove(void *dst , const void *src , int n);
 void *memchr(void *s , int c , int n);
-
-// Functions converted to ASM
-
 void *memcpy(void *dst , const void *src , int n);
 void *memset(void *dst , char c , int n);
 int memcmp(const void *b1 , const void *b2 , int n);
@@ -46,4 +35,3 @@ int memcmp(const void *b1 , const void *b2 , int n);
 #endif
 
 #endif
-
diff --git a/libpsn00b/include/strings.h b/libpsn00b/include/strings.h
index e5e88d4..7223ab9 100644
--- a/libpsn00b/include/strings.h
+++ b/libpsn00b/include/strings.h
@@ -1,18 +1,19 @@
-/* 
- * strings.h
- *
- * PSXSDK
+/*
+ * PSn00bSDK standard library
+ * (C) 2019-2022 PSXSDK authors, Lameguy64, spicyjpeg - MPL licensed
  */
 
-#ifndef _STRINGS_H
-#define _STRINGS_H
+#ifndef __STRINGS_H
+#define __STRINGS_H
 
 #include <string.h>
 
-#define bcopy(src,dst,len)				memmove(dst,src,len)
-#define bzero(ptr, len)					memset(ptr, 0, len)
-#define bcmp(b1,b2,len)					memcmp(b1,b2,len)
-#define index(s, c)						strchr(s, c)
-#define rindex(s, c)						strrchr(s, c)
+/* Compatibility macros (this header is useless) */
+
+#define bcopy(src, dst, len)	memmove(dst, src, len)
+#define bzero(ptr, len)			memset(ptr, 0, len)
+#define bcmp(b1, b2, len)		memcmp(b1, b2, len)
+#define index(s, c)				strchr(s, c)
+#define rindex(s, c)			strrchr(s, c)
 
 #endif
diff --git a/libpsn00b/ldscripts/exe.ld b/libpsn00b/ldscripts/exe.ld
index 583d76a..a8dfccf 100644
--- a/libpsn00b/ldscripts/exe.ld
+++ b/libpsn00b/ldscripts/exe.ld
@@ -6,6 +6,10 @@
  * using $gp to reference them) is fully supported; the block is made up of
  * sections .sdata and .sbss. Note that GP-relative addressing is not
  * compatible with dynamic linking, as DLLs require GP to be unused.
+ *
+ * This linker script assumes main RAM is 8 MB to allow executables meant for
+ * devkits or arcade systems to be built, however the executable conversion
+ * tool (elf2x) will throw a warning if the size exceeds 2 MB.
  */
 
 /*OUTPUT_FORMAT(elf32-littlemips)*/
@@ -15,7 +19,7 @@ ENTRY(_start)
 MEMORY {
 	/* Mapped into KSEG0 */
 	KERNEL_RAM (rwx) : ORIGIN = 0x80000000, LENGTH = 0x010000
-	APP_RAM    (rwx) : ORIGIN = 0x80010000, LENGTH = 0x1f0000
+	APP_RAM    (rwx) : ORIGIN = 0x80010000, LENGTH = 0x7f0000
 }
 
 SECTIONS {
diff --git a/libpsn00b/libc/_start.s b/libpsn00b/libc/_start.s
new file mode 100644
index 0000000..56075c8
--- /dev/null
+++ b/libpsn00b/libc/_start.s
@@ -0,0 +1,18 @@
+# PSn00bSDK _start() trampoline
+# (C) 2022 spicyjpeg - MPL licensed
+#
+# This file provides a weak function that can be easily overridden to e.g. set
+# $sp or perform additional initialization before the "real" _start()
+# (_start_inner()) is called.
+
+.set noreorder
+.section .text
+
+.global _start
+.type _start, @function
+.weak _start
+_start:
+	la $gp, _gp
+
+	j  _start_inner
+	nop
diff --git a/libpsn00b/libc/c++-support.cxx b/libpsn00b/libc/cpp_support.cpp
index 38354dd..f451044 100644
--- a/libpsn00b/libc/c++-support.cxx
+++ b/libpsn00b/libc/cpp_support.cpp
@@ -1,8 +1,22 @@
+/*
+ * PSn00bSDK C++ support library
+ * (C) 2019-2022 Lameguy64, spicyjpeg - MPL licensed
+ */
 
 #include <stdint.h>
 #include <stdlib.h>
 #include <stdio.h>
 
+/* GCC builtins */
+
+extern "C" void *__builtin_new(size_t size) {
+	return malloc(size);
+}
+
+extern "C" void __builtin_delete(void *ptr) {
+	free(ptr);
+}
+
 /* Default new/delete operators */
 
 void *operator new(size_t size) noexcept {
@@ -21,20 +35,15 @@ void operator delete[](void *ptr) noexcept {
 	free(ptr);
 }
 
-/*
- * https://en.cppreference.com/w/cpp/memory/new/operator_delete
- *
- * Called if a user-defined replacement is provided, except that it's
- * unspecified whether other overloads or this overload is called when deleting
- * objects of incomplete type and arrays of non-class and trivially
- * destructible class types.
- *
- * A memory allocator can use the given size to be more efficient.
- */
+// https://en.cppreference.com/w/cpp/memory/new/operator_delete
 void operator delete(void *ptr, size_t size) noexcept {
 	free(ptr);
 }
 
+void operator delete[](void *ptr, size_t size) noexcept {
+	free(ptr);
+}
+
 /* Placement new operators */
 
 void *operator new(size_t size, void *ptr) noexcept {
diff --git a/libpsn00b/libc/malloc.c b/libpsn00b/libc/malloc.c
new file mode 100644
index 0000000..9d538cd
--- /dev/null
+++ b/libpsn00b/libc/malloc.c
@@ -0,0 +1,235 @@
+/*
+ * PSn00bSDK default memory allocator
+ * (C) 2022 Nicolas Noble, spicyjpeg
+ *
+ * This code is based on psyqo's malloc implementation, available here:
+ * https://github.com/grumpycoders/pcsx-redux/blob/main/src/mips/psyqo/src/alloc.c
+ *
+ * Heap management and memory allocation are completely separate, with the
+ * latter being built on top of the former. This makes it possible to override
+ * only InitHeap() and sbrk() while still using the default allocator, or
+ * override malloc()/realloc()/free() while using the default heap manager.
+ */
+
+#include <stddef.h>
+#include <stdint.h>
+#include <stdlib.h>
+
+#define _align(x, n) (((x) + ((n) - 1)) & ~((n) - 1))
+
+/* Private types */
+
+typedef struct _BlockHeader {
+	struct _BlockHeader	*prev, *next;
+	void				*ptr;
+	size_t				size;
+} BlockHeader;
+
+/* Data */
+
+static void			*_heap_start, *_heap_end, *_heap_limit;
+static void			*_alloc_start = 0;
+static BlockHeader	*_alloc_head = 0, *_alloc_tail = 0;
+
+/* Heap management API */
+
+__attribute__((weak)) void InitHeap(void *addr, size_t size) {
+	_heap_start = addr;
+	_heap_end   = addr;
+	_heap_limit = (void *) ((uintptr_t) addr + size);
+}
+
+__attribute__((weak)) void *sbrk(ptrdiff_t incr) {
+	void *old_end = _heap_end;
+	void *new_end = (void *) _align((uintptr_t) old_end + incr, 8);
+
+	if (new_end > _heap_limit)
+		return 0;
+
+	_heap_end = new_end;
+	return old_end;
+}
+
+/* Memory allocator */
+
+static BlockHeader *_find_fit(BlockHeader *head, size_t size) {
+	BlockHeader *prev = head;
+
+	for (; prev; prev = prev->next) {
+		if (prev->next) {
+			uintptr_t next_bot = (uintptr_t) prev->next;
+			next_bot          -= (uintptr_t) prev->ptr + prev->size;
+
+			if (next_bot >= size)
+				return prev;
+		}
+	}
+
+	return prev;
+}
+
+__attribute__((weak)) void *malloc(size_t size) {
+	size_t _size = _align(size + sizeof(BlockHeader), 8);
+
+	// Nothing's initialized yet? Let's just initialize the bottom of our heap,
+	// flag it as allocated.
+	if (!_alloc_head) {
+		if (!_alloc_start)
+			_alloc_start = sbrk(0);
+
+		BlockHeader *new = (BlockHeader *) sbrk(_size);
+		if (!new)
+			return 0;
+
+		void *ptr = (void *) &new[1];
+		new->ptr  = ptr;
+		new->size = _size - sizeof(BlockHeader);
+		new->prev = 0;
+		new->next = 0;
+
+		_alloc_head = new;
+		_alloc_tail = new;
+		return ptr;
+	}
+
+	// We *may* have the bottom of our heap that has shifted, because of a free.
+	// So let's check first if we have free space there, because I'm nervous
+	// about having an incomplete data structure.
+	if (((uintptr_t) _alloc_start + _size) < ((uintptr_t) _alloc_head)) {
+		BlockHeader *new = (BlockHeader *) _alloc_start;
+
+		void *ptr = (void *) &new[1];
+		new->ptr  = ptr;
+		new->size = _size - sizeof(BlockHeader);
+		new->prev = 0;
+		new->next = _alloc_head;
+
+		_alloc_head->prev = new;
+		_alloc_head       = new;
+		return ptr;
+	}
+
+	// No luck at the beginning of the heap, let's walk the heap to find a fit.
+	BlockHeader *prev = _find_fit(_alloc_head, _size);
+	if (prev) {
+		BlockHeader *new = (BlockHeader *) ((uintptr_t) prev->ptr + prev->size);
+
+		void *ptr = (void *)((uintptr_t) new + sizeof(BlockHeader));
+		new->ptr  = ptr;
+		new->size = _size - sizeof(BlockHeader);
+		new->prev = prev;
+		new->next = prev->next;
+
+		(new->next)->prev = new;
+		prev->next        = new;
+		return ptr;
+	}
+
+	// Time to extend the size of the heap.
+	BlockHeader *new = (BlockHeader *) sbrk(_size);
+	if (!new)
+		return 0;
+
+	void *ptr = (void *) &new[1];
+	new->ptr  = ptr;
+	new->size = _size - sizeof(BlockHeader);
+	new->prev = _alloc_tail;
+	new->next = 0;
+
+	_alloc_tail->next = new;
+	_alloc_tail       = new;
+	return ptr;
+}
+
+__attribute__((weak)) void *calloc(size_t num, size_t size) {
+	return malloc(num * size);
+}
+
+__attribute__((weak)) void *realloc(void *ptr, size_t size) {
+	if (!size) {
+		free(ptr);
+		return 0;
+	}
+	if (!ptr)
+		return malloc(size);
+
+	size_t _size = _align(size + sizeof(BlockHeader), 8);
+
+	BlockHeader *prev = (BlockHeader *) ((uintptr_t) ptr - sizeof(BlockHeader));
+
+	// New memory block shorter?
+	if (prev->size >= _size) {
+		prev->size = _size;
+		if (!prev->next)
+			sbrk((ptr - sbrk(0)) + _size);
+
+		return ptr;
+	}
+
+	// New memory block larger; is it the last one?
+	if (!prev->next) {
+		void *new = sbrk(_size - prev->size);
+		if (!new)
+			return 0;
+
+		prev->size = _size;
+		return ptr;
+	}
+
+	// Do we have free memory after it?
+	if (((prev->next)->ptr - ptr) > _size) {
+		prev->size = _size;
+		return ptr;
+	}
+
+	// No luck.
+	void *new = malloc(_size);
+	if (!new)
+		return 0;
+
+	__builtin_memcpy(new, ptr, prev->size);
+	free(ptr);
+	return new;
+}
+
+__attribute__((weak)) void free(void *ptr) {
+	if (!ptr || !_alloc_head)
+		return;
+
+	// First block; bumping head ahead.
+	if (ptr == _alloc_head->ptr) {
+		size_t size = _alloc_head->size;
+		size       += (uintptr_t) _alloc_head->ptr - (uintptr_t) _alloc_head;
+		_alloc_head = _alloc_head->next;
+
+		if (_alloc_head) {
+			_alloc_head->prev = 0;
+		} else {
+			_alloc_tail = 0;
+			sbrk(-size);
+		}
+
+		return;
+	}
+
+	// Finding the proper block
+	BlockHeader *cur = _alloc_head;
+	for (cur = _alloc_head; ptr != cur->ptr; cur = cur->next) {
+		if (!cur->next)
+			return;
+	}
+
+	if (cur->next) {
+		// In the middle, just unlink it
+		cur->next->prev = cur->prev;
+	} else {
+		// At the end, shrink heap
+		_alloc_tail = cur->prev;
+
+		void  *top  = sbrk(0);
+		size_t size = (top - (cur->prev)->ptr) - (cur->prev)->size;
+		sbrk(-size);
+	}
+
+	(cur->prev)->next = cur->next;
+}
diff --git a/libpsn00b/libc/malloc.s b/libpsn00b/libc/malloc.s
deleted file mode 100644
index e441bbe..0000000
--- a/libpsn00b/libc/malloc.s
+++ /dev/null
@@ -1,242 +0,0 @@
-# Custom first-fit malloc routines by Lameguy64
-# Part of the PSn00bSDK Project
-#
-# NOTE: there reportedly is a GCC bug which messes up .weak functions written
-# in assembly if LTO is enabled. I haven't tested but, according to the
-# internet, this bug has never been fixed.
-# https://gcc.gnu.org/legacy-ml/gcc-help/2019-10/msg00092.html
-
-.set noreorder
-
-.set ND_PREV,	0	# Address to previous block (NULL if starting block)
-.set ND_NEXT,	4	# Address to next block (NULL if end block)
-.set ND_SIZE,	8	# Size of block
-.set ND_HSIZ,	12
-
-.section .text
-
-# Stupid small function just to get bss end
-# due to GCC insisting externs to be gp relative
-.global GetBSSend
-.type GetBSSend, @function
-GetBSSend:
-	la		$v0, _end
-	jr		$ra
-	nop
-	
-
-# Initializes the heap for malloc
-#   a0 - Starting address of heap
-#   a1 - Size of memory heap
-#
-.global InitHeap
-.type InitHeap, @function
-.weak InitHeap
-InitHeap:
-	la		$v0, _malloc_addr
-	sw		$a0, 0($v0)
-	la		$v0, _malloc_size
-	sw		$a1, 0($v0)
-	
-	sw		$0 , ND_PREV($a0)		# Set heap header
-	sw		$0 , ND_NEXT($a0)
-	jr		$ra
-	sw		$0 , ND_SIZE($a0)
-	
-	
-# Changes the heap size without clearing or relocating the heap
-#   a0 - Size of memory heap in bytes
-.global SetHeapSize
-.type SetHeapSize, @function
-.weak SetHeapSize
-SetHeapSize:
-	la		$v1, _malloc_size
-	lw		$v0, 0($v1)
-	jr		$ra
-	sw		$a1, 0($v1)
-
-
-# Allocates a block of memory in the heap
-#   a0 - Size of memory block to allocate.
-#
-.global malloc
-.type malloc, @function
-.weak malloc
-malloc:
-	addiu	$a0, 3					# Round size to a multiple of 4
-	srl		$a0, 2
-	
-	la		$a2, _malloc_addr
-	lw		$a2, 0($a2)
-	sll		$a0, 2
-	
-.Lfind_next:
-	
-	move	$a1, $a2
-	
-	lw		$a2, ND_NEXT($a1)		# Get block header
-	lw		$v1, ND_SIZE($a1)
-	
-	subu	$v0, $a2, $a1			# Compute space between current and next
-	
-	beqz	$v1, .Lempty_block		# Occupy empty block (if size = 0)
-	nop
-	
-	beqz	$a2, .Lnew_block		# Allocate a new block (if no next)
-	nop
-	
-	addiu	$v0, -(ND_HSIZ*2)		# Compute remaining space of block
-	subu	$v0, $v1
-	
-	blt		$v0, $a0, .Lfind_next	# Search for the next block if space is not big enough
-	nop
-	
-	# Perform a block split using remaining space of current block
-
-	addiu	$v0, $a1, ND_HSIZ		# Compute address for new header
-	addu	$v0, $v1
-	
-	sw		$a1, ND_PREV($v0)		# Set the new block header
-	sw		$a2, ND_NEXT($v0)
-	sw		$a0, ND_SIZE($v0)
-	
-	sw		$v0, ND_NEXT($a1)		# Update previous and next blocks
-	sw		$v0, ND_PREV($a2)
-	
-	jr		$ra
-	addiu	$v0, ND_HSIZ
-	
-.Lempty_block:						# Occupy an empty block
-
-	beqz	$a2, .Lno_next			# Skip size calculation if there's no next
-	nop
-	
-	addiu	$v0, -ND_HSIZ
-	blt		$v0, $a0, .Lfind_next
-	nop
-	
-	b		.Lskip_space_check
-	nop
-	
-.Lno_next:
-
-	la		$v1, _malloc_addr		# Check if there's enough space for a block
-	lw		$v1, 0($v1)
-	la		$v0, _malloc_size
-	lw		$v0, 0($v0)
-	
-	subu	$v1, $a1, $v1
-	addu	$v1, $a0
-	addiu	$v1, ND_HSIZ
-	
-	bgt		$v1, $v0, .Lno_space
-	nop
-	
-.Lskip_space_check:
-
-	sw		$a0, ND_SIZE($a1)
-	jr		$ra						# Return address
-	addiu	$v0, $a1, ND_HSIZ
-	
-.Lnew_block:						# Create a new block
-
-	addu	$a2, $a1, $v1			# Compute address for new block
-	addiu	$a2, ND_HSIZ
-	
-	la		$v1, _malloc_addr
-	lw		$v1, 0($v1)
-	la		$v0, _malloc_size
-	lw		$v0, 0($v0)
-	
-	subu	$v1, $a2, $v1
-	addu	$v1, $a0
-	addiu	$v1, ND_HSIZ
-	
-	bgt		$v1, $v0, .Lno_space	# Reject if it exceeds specified size
-	nop
-	
-	sw		$a1, ND_PREV($a2)
-	sw		$0 , ND_NEXT($a2)
-	sw		$a0, ND_SIZE($a2)
-	
-	sw		$a2, ND_NEXT($a1)
-	
-	jr		$ra						# Return address
-	addiu	$v0, $a2, ND_HSIZ
-	
-.Lno_space:							# Return a null if no space can be found
-
-	jr		$ra
-	move	$v0, $0
-
-	
-# Allocates a block of memory in block units and zero fills the
-# allocated block.
-#   a0 - Block size.
-#	a1 - Number of blocks to allocate
-#
-.global calloc
-.type calloc, @function
-.weak calloc
-calloc:
-	mult	$a0, $a1
-	addiu	$sp, -4
-	sw		$ra, 0($sp)
-	
-	jal		malloc
-	mflo	$a0
-	
-	move	$a0, $v0
-	mflo	$a1
-
-.Lclear_loop:
-
-	sw		$0 , 0($a0)
-	addi	$a1, 4
-	bgtz	$a1, .Lclear_loop
-	addiu	$a0, 4
-	
-	lw		$ra, 0($sp)
-	addiu	$sp, 4
-	jr		$ra
-	nop
-
-	
-# Deallocates an allocated block
-#	a0 - An address returned by malloc to deallocate
-#
-.global free
-.type free, @function
-.weak free
-free:
-
-	addiu	$a0, -ND_HSIZ
-	lw		$a1, ND_PREV($a0)
-	lw		$a2, ND_NEXT($a0)
-	
-	beqz	$a1, .Lis_start			# Check if block is a starting block
-	nop
-	
-	beqz	$a2, .Lis_end
-	nop
-
-	# Unlink
-	
-	sw		$a2, ND_NEXT($a1)
-	jr		$ra
-	sw		$a1, ND_PREV($a2)
-	
-.Lis_end:							# Unlinks the ending block
-
-	jr		$ra
-	sw		$0 , ND_NEXT($a1)
-	
-.Lis_start:							# Simply set size to 0 if starting block
-
-	jr		$ra
-	sw		$0 , ND_SIZE($a0)
-
-
-# Internal variables
-.comm _malloc_addr, 4, 4
-.comm _malloc_size, 4, 4
diff --git a/libpsn00b/libc/start.c b/libpsn00b/libc/start.c
index fd6fe33..87ac951 100644
--- a/libpsn00b/libc/start.c
+++ b/libpsn00b/libc/start.c
@@ -62,10 +62,10 @@ extern uint8_t _end[];
 // useful though to change the stack size and/or reinitialize the heap on
 // systems that have more than 2 MB of RAM (e.g. emulators, devkits, PS1-based
 // arcade boards).
-void _mem_init(int ram_size, int stack_max_size) {
-	void *exe_end = _end + 4;
-	int  exe_size = (int) exe_end - (int) __text_start;
-	int  ram_used = (0x10000 + exe_size + stack_max_size) & 0xfffffffc;
+void _mem_init(size_t ram_size, size_t stack_max_size) {
+	void   *exe_end = _end + 4;
+	size_t exe_size = (size_t) exe_end - (size_t) __text_start;
+	size_t ram_used = (0x10000 + exe_size + stack_max_size) & 0xfffffffc;
 
 	InitHeap(exe_end, ram_size - ram_used);
 }
@@ -80,8 +80,8 @@ extern int main(int argc, const char* argv[]);
 // Even though _start() usually takes no arguments, this implementation allows
 // parent executables to pass args directly to child executables without having
 // to overwrite the arg strings in kernel RAM.
-void _start(int32_t override_argc, const char **override_argv) {
-	__asm__ volatile("la $gp, _gp;");
+void _start_inner(int32_t override_argc, const char **override_argv) {
+	//__asm__ volatile("la $gp, _gp;");
 
 	// Clear BSS 4 bytes at a time. BSS is always aligned to 4 bytes by the
 	// linker script.
diff --git a/libpsn00b/lzp/crc.c b/libpsn00b/lzp/crc.c
index 3c1ae57..6b00c8e 100644
--- a/libpsn00b/lzp/crc.c
+++ b/libpsn00b/lzp/crc.c
@@ -1,14 +1,15 @@
+#include <stdint.h>
 #include "lzp.h"
 
-void initTable16(unsigned short* table) {
+void initTable16(uint16_t *table) {
 
 	int i, j;
-    unsigned short crc, c;
+    uint16_t crc, c;
 
     for (i=0; i<256; i++) {
 
         crc = 0;
-        c   = (unsigned short) i;
+        c   = (uint16_t) i;
 
         for (j=0; j<8; j++) {
 
@@ -25,10 +26,10 @@ void initTable16(unsigned short* table) {
 
 }
 
-void initTable32(unsigned int* table) {
+void initTable32(uint32_t *table) {
 
 	int i,j;
-	unsigned int crcVal;
+	uint32_t crcVal;
 
 	for(i=0; i<256; i++) {
 
@@ -49,17 +50,17 @@ void initTable32(unsigned int* table) {
 
 }
 
-unsigned short lzCRC16(const void* buff, int bytes, unsigned short crc) {
+uint16_t lzCRC16(const void* buff, int bytes, uint16_t crc) {
 
 	int i;
-	unsigned short tmp, short_c;
-	unsigned short crcTable[256];
+	uint16_t tmp, short_c;
+	uint16_t crcTable[256];
 
 	initTable16(crcTable);
 
 	for(i=0; i<bytes; i++) {
 
-		short_c = 0x00ff & (unsigned short)((const unsigned char*)buff)[i];
+		short_c = 0x00ff & (uint16_t)((const uint8_t *)buff)[i];
 
 		tmp =  crc       ^ short_c;
 		crc = (crc >> 8) ^ crcTable[tmp&0xff];
@@ -70,18 +71,18 @@ unsigned short lzCRC16(const void* buff, int bytes, unsigned short crc) {
 
 }
 
-unsigned int lzCRC32(const void* buff, int bytes, unsigned int crc) {
+uint32_t lzCRC32(const void* buff, int bytes, uint32_t crc) {
 
 	int	i;
-	const unsigned char*	byteBuff = (const unsigned char*)buff;
-	unsigned int			byte;
-	unsigned int			crcTable[256];
+	const uint8_t	*byteBuff = (const uint8_t *)buff;
+	uint32_t		byte;
+	uint32_t		crcTable[256];
 
     initTable32(crcTable);
 
 	for(i=0; i<bytes; i++) {
 
-		byte = 0x000000ffL&(unsigned int)byteBuff[i];
+		byte = 0x000000ffL&(uint32_t)byteBuff[i];
 		crc = (crc>>8)^crcTable[(crc^byte)&0xff];
 
 	}
diff --git a/libpsn00b/lzp/lzp.h b/libpsn00b/lzp/lzp.h
index cfeeb72..456de02 100644
--- a/libpsn00b/lzp/lzp.h
+++ b/libpsn00b/lzp/lzp.h
@@ -16,7 +16,7 @@
 #ifndef _LZPACK_H
 #define _LZPACK_H
 
-#include <sys/types.h>
+#include <stdint.h>
 #ifdef _WIN32
 #include <windows.h>
 #endif
@@ -64,9 +64,9 @@
 typedef struct {
 
 	//! File ID (must always be 'LZP')
-	char	id[3];
+	char	    id[3];
 	//! File count
-	u_char	numFiles;
+	uint8_t 	numFiles;
 
 } LZP_HEAD;
 
@@ -74,15 +74,15 @@ typedef struct {
 typedef struct {
 
 	//! File name
-    char	fileName[16];
+    char	    fileName[16];
     //! CRC32 checksum of file
-    u_int	crc;
+    uint32_t	crc;
     //! Original size of file in bytes
-    u_int	fileSize;
+    uint32_t	fileSize;
     //! Compressed size of file
-    u_int	packedSize;
+    uint32_t	packedSize;
     //! File data offset
-    u_int	offset;
+    uint32_t	offset;
 
 } LZP_FILE;
 
@@ -162,7 +162,7 @@ void lzResetHashSizes();
  *
  *	\returns CRC16 hash of specified buffer.
  */
-unsigned short lzCRC16(const void* buff, int bytes, unsigned short crc);
+uint16_t lzCRC16(const void* buff, int bytes, uint16_t crc);
 
 /*!	Calculates a CRC32 hash of the specified buffer.
  *
@@ -172,7 +172,7 @@ unsigned short lzCRC16(const void* buff, int bytes, unsigned short crc);
  *
  *	\returns CRC32 hash of specified buffer.
  */
-unsigned int lzCRC32(const void* buff, int bytes, unsigned int crc);
+uint32_t lzCRC32(const void* buff, int bytes, uint32_t crc);
 
 /*!	@}	*/
 
diff --git a/libpsn00b/lzp/lzqlp.h b/libpsn00b/lzp/lzqlp.h
index 5b70b40..32ce0d7 100644
--- a/libpsn00b/lzp/lzqlp.h
+++ b/libpsn00b/lzp/lzqlp.h
@@ -1,7 +1,7 @@
 #ifndef _QLP_H
 #define _QLP_H
 
-#include <sys/types.h>
+#include <stdint.h>
 #ifdef _WIN32
 #include <windows.h>
 #endif
@@ -13,14 +13,14 @@
 #define PACK_ERR_READ_FAULT		-4
 
 typedef struct {
-	char	id[3];
-	u_char	numfiles;
+	char		id[3];
+	uint8_t		numfiles;
 } QLP_HEAD;
 
 typedef struct {
-	char	name[16];
-	u_int	size;
-	u_int	offs;
+	char		name[16];
+	uint32_t	size;
+	uint32_t	offs;
 } QLP_FILE;
 
 int qlpFileCount(const QLP_HEAD* qlpfile);
diff --git a/libpsn00b/psxapi/_initcd.s b/libpsn00b/psxapi/_initcd.s
index c3a2861..600d686 100644
--- a/libpsn00b/psxapi/_initcd.s
+++ b/libpsn00b/psxapi/_initcd.s
@@ -1,7 +1,7 @@
 .section .text
 .set noreorder
 
-.include "hwregs_a.h"
+.include "hwregs_a.inc"
 
 .global _InitCd
 .type _InitCd, @function
diff --git a/libpsn00b/psxapi/_syscalls.s b/libpsn00b/psxapi/_syscalls.s
new file mode 100644
index 0000000..24864f3
--- /dev/null
+++ b/libpsn00b/psxapi/_syscalls.s
@@ -0,0 +1,50 @@
+# PSn00bSDK syscall wrappers
+# (C) 2022 spicyjpeg - MPL licensed
+
+.set noreorder
+
+.section .text.EnterCriticalSection
+.global EnterCriticalSection
+.type EnterCriticalSection, @function
+EnterCriticalSection:
+	li $a0, 0x01
+	syscall 0
+
+	jr $ra
+	nop
+
+.section .text.ExitCriticalSection
+.global ExitCriticalSection
+.type ExitCriticalSection, @function
+ExitCriticalSection:
+	li $a0, 0x02
+	syscall 0
+
+	jr $ra
+	nop
+
+.section .text.SwEnterCriticalSection
+.global SwEnterCriticalSection
+.type SwEnterCriticalSection, @function
+SwEnterCriticalSection:
+	mfc0 $a0, $12 # cop0r12 &= ~0x00000401
+	li   $a1, -1026
+	and  $a0, $a1
+	mtc0 $a0, $12
+	nop
+
+	jr   $ra
+	nop
+
+.section .text.SwExitCriticalSection
+.global SwExitCriticalSection
+.type SwExitCriticalSection, @function
+SwExitCriticalSection:
+	mfc0 $a0, $12 # cop0r12 |= 0x00000401
+	nop
+	ori  $a0, 0x0401
+	mtc0 $a0, $12
+	nop
+
+	jr   $ra
+	nop
diff --git a/libpsn00b/psxapi/stubs.json b/libpsn00b/psxapi/stubs.json
index 9198b06..200cbc7 100644
--- a/libpsn00b/psxapi/stubs.json
+++ b/libpsn00b/psxapi/stubs.json
@@ -133,6 +133,12 @@
 	},
 	{
 		"type": "a",
+		"id":   156,
+		"name": "SetConf",
+		"file": "sys.s"
+	},
+	{
+		"type": "a",
 		"id":   160,
 		"name": "_boot",
 		"file": "sys.s"
@@ -205,12 +211,36 @@
 	},
 	{
 		"type": "b",
+		"id":   7,
+		"name": "DeliverEvent",
+		"file": "sys.s"
+	},
+	{
+		"type": "b",
 		"id":   8,
 		"name": "OpenEvent",
 		"file": "sys.s"
 	},
 	{
 		"type": "b",
+		"id":   9,
+		"name": "CloseEvent",
+		"file": "sys.s"
+	},
+	{
+		"type": "b",
+		"id":   10,
+		"name": "WaitEvent",
+		"file": "sys.s"
+	},
+	{
+		"type": "b",
+		"id":   11,
+		"name": "TestEvent",
+		"file": "sys.s"
+	},
+	{
+		"type": "b",
 		"id":   12,
 		"name": "EnableEvent",
 		"file": "sys.s"
@@ -223,6 +253,24 @@
 	},
 	{
 		"type": "b",
+		"id":   14,
+		"name": "OpenTh",
+		"file": "sys.s"
+	},
+	{
+		"type": "b",
+		"id":   15,
+		"name": "CloseTh",
+		"file": "sys.s"
+	},
+	{
+		"type": "b",
+		"id":   16,
+		"name": "ChangeTh",
+		"file": "sys.s"
+	},
+	{
+		"type": "b",
 		"id":   18,
 		"name": "InitPAD",
 		"file": "sys.s"
@@ -259,6 +307,12 @@
 	},
 	{
 		"type": "b",
+		"id":   32,
+		"name": "UnDeliverEvent",
+		"file": "sys.s"
+	},
+	{
+		"type": "b",
 		"id":   64,
 		"name": "chdir",
 		"file": "fs.s"
@@ -391,20 +445,20 @@
 	},
 	{
 		"type": "c",
-		"id":   10,
-		"name": "ChangeClearRCnt",
+		"id":   7,
+		"name": "InstallExceptionHandlers",
 		"file": "sys.s"
 	},
 	{
-		"type": "syscall",
-		"id":   1,
-		"name": "EnterCriticalSection",
+		"type": "c",
+		"id":   8,
+		"name": "SysInitMemory",
 		"file": "sys.s"
 	},
 	{
-		"type": "syscall",
-		"id":   2,
-		"name": "ExitCriticalSection",
+		"type": "c",
+		"id":   10,
+		"name": "ChangeClearRCnt",
 		"file": "sys.s"
 	}
 ]
diff --git a/libpsn00b/psxapi/sys.s b/libpsn00b/psxapi/sys.s
index e54bd98..e2505e1 100644
--- a/libpsn00b/psxapi/sys.s
+++ b/libpsn00b/psxapi/sys.s
@@ -6,7 +6,7 @@
 
 .set noreorder
 
-## A0 table functions (7)
+## A0 table functions (8)
 
 .section .text.b_setjmp
 .global b_setjmp
@@ -48,6 +48,14 @@ FlushCache:
 	jr $t2
 	li $t1, 0x44
 
+.section .text.SetConf
+.global SetConf
+.type SetConf, @function
+SetConf:
+	li $t2, 0xa0
+	jr $t2
+	li $t1, 0x9c
+
 .section .text._boot
 .global _boot
 .type _boot, @function
@@ -64,7 +72,7 @@ GetSystemInfo:
 	jr $t2
 	li $t1, 0xb4
 
-## B0 table functions (19)
+## B0 table functions (27)
 
 .section .text._kernel_malloc
 .global _kernel_malloc
@@ -122,6 +130,14 @@ ResetRCnt:
 	jr $t2
 	li $t1, 0x06
 
+.section .text.DeliverEvent
+.global DeliverEvent
+.type DeliverEvent, @function
+DeliverEvent:
+	li $t2, 0xb0
+	jr $t2
+	li $t1, 0x07
+
 .section .text.OpenEvent
 .global OpenEvent
 .type OpenEvent, @function
@@ -130,6 +146,30 @@ OpenEvent:
 	jr $t2
 	li $t1, 0x08
 
+.section .text.CloseEvent
+.global CloseEvent
+.type CloseEvent, @function
+CloseEvent:
+	li $t2, 0xb0
+	jr $t2
+	li $t1, 0x09
+
+.section .text.WaitEvent
+.global WaitEvent
+.type WaitEvent, @function
+WaitEvent:
+	li $t2, 0xb0
+	jr $t2
+	li $t1, 0x0a
+
+.section .text.TestEvent
+.global TestEvent
+.type TestEvent, @function
+TestEvent:
+	li $t2, 0xb0
+	jr $t2
+	li $t1, 0x0b
+
 .section .text.EnableEvent
 .global EnableEvent
 .type EnableEvent, @function
@@ -146,6 +186,30 @@ DisableEvent:
 	jr $t2
 	li $t1, 0x0d
 
+.section .text.OpenTh
+.global OpenTh
+.type OpenTh, @function
+OpenTh:
+	li $t2, 0xb0
+	jr $t2
+	li $t1, 0x0e
+
+.section .text.CloseTh
+.global CloseTh
+.type CloseTh, @function
+CloseTh:
+	li $t2, 0xb0
+	jr $t2
+	li $t1, 0x0f
+
+.section .text.ChangeTh
+.global ChangeTh
+.type ChangeTh, @function
+ChangeTh:
+	li $t2, 0xb0
+	jr $t2
+	li $t1, 0x10
+
 .section .text.InitPAD
 .global InitPAD
 .type InitPAD, @function
@@ -194,6 +258,14 @@ SetCustomExitFromException:
 	jr $t2
 	li $t1, 0x19
 
+.section .text.UnDeliverEvent
+.global UnDeliverEvent
+.type UnDeliverEvent, @function
+UnDeliverEvent:
+	li $t2, 0xb0
+	jr $t2
+	li $t1, 0x20
+
 .section .text.GetC0Table
 .global GetC0Table
 .type GetC0Table, @function
@@ -218,7 +290,7 @@ ChangeClearPAD:
 	jr $t2
 	li $t1, 0x5b
 
-## C0 table functions (3)
+## C0 table functions (5)
 
 .section .text.SysEnqIntRP
 .global SysEnqIntRP
@@ -236,6 +308,22 @@ SysDeqIntRP:
 	jr $t2
 	li $t1, 0x03
 
+.section .text.InstallExceptionHandlers
+.global InstallExceptionHandlers
+.type InstallExceptionHandlers, @function
+InstallExceptionHandlers:
+	li $t2, 0xc0
+	jr $t2
+	li $t1, 0x07
+
+.section .text.SysInitMemory
+.global SysInitMemory
+.type SysInitMemory, @function
+SysInitMemory:
+	li $t2, 0xc0
+	jr $t2
+	li $t1, 0x08
+
 .section .text.ChangeClearRCnt
 .global ChangeClearRCnt
 .type ChangeClearRCnt, @function
@@ -244,23 +332,3 @@ ChangeClearRCnt:
 	jr $t2
 	li $t1, 0x0a
 
-## Syscalls (2)
-
-.section .text.EnterCriticalSection
-.global EnterCriticalSection
-.type EnterCriticalSection, @function
-EnterCriticalSection:
-	li      $a0, 0x01
-	syscall 0
-	jr      $ra
-	nop
-
-.section .text.ExitCriticalSection
-.global ExitCriticalSection
-.type ExitCriticalSection, @function
-ExitCriticalSection:
-	li      $a0, 0x02
-	syscall 0
-	jr      $ra
-	nop
-
diff --git a/libpsn00b/psxcd/_cd_control.s b/libpsn00b/psxcd/_cd_control.s
index c4153ff..5fa336a 100644
--- a/libpsn00b/psxcd/_cd_control.s
+++ b/libpsn00b/psxcd/_cd_control.s
@@ -1,6 +1,6 @@
 .set noreorder
 
-.include "hwregs_a.h"
+.include "hwregs_a.inc"
 
 .section .text
 
diff --git a/libpsn00b/psxcd/cdgetsector.s b/libpsn00b/psxcd/cdgetsector.s
index 9f38e7a..6a29069 100644
--- a/libpsn00b/psxcd/cdgetsector.s
+++ b/libpsn00b/psxcd/cdgetsector.s
@@ -1,6 +1,6 @@
 .set noreorder
 
-.include "hwregs_a.h"
+.include "hwregs_a.inc"
 
 .section .text
 
diff --git a/libpsn00b/psxcd/cdmix.s b/libpsn00b/psxcd/cdmix.s
index 745fb65..40cd181 100644
--- a/libpsn00b/psxcd/cdmix.s
+++ b/libpsn00b/psxcd/cdmix.s
@@ -1,6 +1,6 @@
 .set noreorder
 
-.include "hwregs_a.h"
+.include "hwregs_a.inc"
 
 .section .text
 
diff --git a/libpsn00b/psxcd/isofs.c b/libpsn00b/psxcd/isofs.c
index d1c1b18..582b8d9 100644
--- a/libpsn00b/psxcd/isofs.c
+++ b/libpsn00b/psxcd/isofs.c
@@ -1,4 +1,4 @@
-#include <sys/types.h>
+#include <stdint.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
@@ -15,18 +15,18 @@
 
 typedef struct _CdlDIR_INT
 {
-	u_long	_pos;
-	u_long	_len;
-	u_char*	_dir;
+	uint32_t	_pos;
+	uint32_t	_len;
+	uint8_t		*_dir;
 } CdlDIR_INT;
 
 extern int _cd_media_changed;
 
 
 static int		_cd_iso_last_dir_lba;
-static u_char	_cd_iso_descriptor_buff[2048];
-static u_char*	_cd_iso_pathtable_buff=NULL;
-static u_char*	_cd_iso_directory_buff=NULL;
+static uint8_t	_cd_iso_descriptor_buff[2048];
+static uint8_t	*_cd_iso_pathtable_buff=NULL;
+static uint8_t	*_cd_iso_directory_buff=NULL;
 static int		_cd_iso_directory_len;
 static int		_cd_iso_error=0;
 
@@ -68,7 +68,7 @@ static int _CdReadIsoDescriptor(int session_offs)
 
 	// Seek to volume descriptor
 	CdIntToPos(16+session_offs, &loc);
-	if( !CdControl(CdlSetloc, (u_char*)&loc, 0) )
+	if( !CdControl(CdlSetloc, (uint8_t*)&loc, 0) )
 	{
 #ifdef DEBUG
 		printf("psxcd: Could not set seek destination.\n");
@@ -85,7 +85,7 @@ static int _CdReadIsoDescriptor(int session_offs)
 	printf("psxcd: Read sectors.\n");
 #endif
 	// Read volume descriptor
-	CdRead(1, (u_long*)_cd_iso_descriptor_buff, CdlModeSpeed);
+	CdRead(1, (uint32_t*)_cd_iso_descriptor_buff, CdlModeSpeed);
 	
 	if( CdReadSync(0, 0) )
 	{
@@ -122,7 +122,7 @@ static int _CdReadIsoDescriptor(int session_offs)
 	{
 		free(_cd_iso_pathtable_buff);
 	}
-	_cd_iso_pathtable_buff = (u_char*)malloc(i);
+	_cd_iso_pathtable_buff = (uint8_t*)malloc(i);
 	
 #ifdef DEBUG
 	printf("psxcd_dbg: Allocated %d bytes for path table.\n", i);
@@ -130,8 +130,8 @@ static int _CdReadIsoDescriptor(int session_offs)
 	
 	// Read path table
 	CdIntToPos(descriptor->pathTable1Offs, &loc);
-	CdControl(CdlSetloc, (u_char*)&loc, 0);
-	CdRead(i>>11, (u_long*)_cd_iso_pathtable_buff, CdlModeSpeed);
+	CdControl(CdlSetloc, (uint8_t*)&loc, 0);
+	CdRead(i>>11, (uint32_t*)_cd_iso_pathtable_buff, CdlModeSpeed);
 	if( CdReadSync(0, 0) )
 	{
 #ifdef DEBUG
@@ -165,7 +165,7 @@ static int _CdReadIsoDirectory(int lba)
 #ifdef DEBUG
 	printf("psxcd_dbg: Seek to sector %d\n", i);
 #endif
-	if( !CdControl(CdlSetloc, (u_char*)&loc, 0) )
+	if( !CdControl(CdlSetloc, (uint8_t*)&loc, 0) )
 	{
 #ifdef DEBUG
 		printf("psxcd: Could not set seek destination.\n");
@@ -180,8 +180,8 @@ static int _CdReadIsoDirectory(int lba)
 	}
 	
 	// Read first sector of directory record
-	_cd_iso_directory_buff = (u_char*)malloc(2048);
-	CdRead(1, (u_long*)_cd_iso_directory_buff, CdlModeSpeed);
+	_cd_iso_directory_buff = (uint8_t*)malloc(2048);
+	CdRead(1, (uint32_t*)_cd_iso_directory_buff, CdlModeSpeed);
 	if( CdReadSync(0, 0) )
 	{
 #ifdef DEBUG
@@ -201,7 +201,7 @@ static int _CdReadIsoDirectory(int lba)
 
 	if( _cd_iso_directory_len > 2048 )
 	{
-		if( !CdControl(CdlSetloc, (u_char*)&loc, 0) )
+		if( !CdControl(CdlSetloc, (uint8_t*)&loc, 0) )
 		{
 #ifdef DEBUG
 			printf("psxcd: Could not set seek destination.\n");
@@ -212,12 +212,12 @@ static int _CdReadIsoDirectory(int lba)
 	
 		free(_cd_iso_directory_buff);
 		i = ((2047+_cd_iso_directory_len)>>11)<<11;
-		_cd_iso_directory_buff = (u_char*)malloc(i);
+		_cd_iso_directory_buff = (uint8_t*)malloc(i);
 #ifdef DEBUG
 		printf("psxcd_dbg: Allocated %d bytes for directory record.\n", i);
 #endif
 
-		CdRead(i>>11, (u_long*)_cd_iso_directory_buff, CdlModeSpeed);
+		CdRead(i>>11, (uint32_t*)_cd_iso_directory_buff, CdlModeSpeed);
 		if( CdReadSync(0, 0) )
 		{
 #ifdef DEBUG
@@ -279,7 +279,7 @@ static void dump_directory(void)
 
 static void dump_pathtable(void)
 {
-	u_char *tbl_pos;
+	uint8_t *tbl_pos;
 	ISO_PATHTABLE_ENTRY *tbl_entry;
 	ISO_DESCRIPTOR *descriptor;
 	char namebuff[16];
@@ -314,7 +314,7 @@ static void dump_pathtable(void)
 static int get_pathtable_entry(int entry, ISO_PATHTABLE_ENTRY *tbl, char *namebuff)
 {
 	int i;
-	u_char *tbl_pos;
+	uint8_t *tbl_pos;
 	ISO_PATHTABLE_ENTRY *tbl_entry;
 	ISO_DESCRIPTOR *descriptor;
 	
diff --git a/libpsn00b/psxcd/psxcd.c b/libpsn00b/psxcd/psxcd.c
index 8f19c8d..6340638 100644
--- a/libpsn00b/psxcd/psxcd.c
+++ b/libpsn00b/psxcd/psxcd.c
@@ -1,4 +1,4 @@
-#include <sys/types.h>
+#include <stdint.h>
 #include <stdio.h>
 #include <psxgpu.h>
 #include "psxcd.h"
@@ -6,16 +6,16 @@
 #define READ_TIMEOUT	600		// 10 seconds for NTSC
 
 extern volatile char _cd_ack_wait;
-extern volatile u_char _cd_last_int;
-extern volatile u_char _cd_last_mode;
-extern volatile u_char _cd_status;
+extern volatile uint8_t _cd_last_int;
+extern volatile uint8_t _cd_last_mode;
+extern volatile uint8_t _cd_status;
 extern volatile CdlCB _cd_callback_int1_data;
 
-volatile u_char *_cd_result_ptr;
+volatile uint8_t *_cd_result_ptr;
 
 // For read retry
 volatile CdlLOC _cd_last_setloc;
-volatile u_long *_cd_last_read_addr;
+volatile uint32_t *_cd_last_read_addr;
 volatile int _cd_last_sector_count;
 
 int _cd_media_changed;
@@ -169,7 +169,7 @@ int CdSync(int mode, unsigned char *result)
 
 int CdGetToc(CdlLOC *toc)
 {
-	u_char track_info[8];
+	uint8_t track_info[8];
 	int i,tracks;
 	
 	// Get number of tracks
@@ -189,7 +189,7 @@ int CdGetToc(CdlLOC *toc)
 	for(i=0; i<tracks; i++)
 	{
 		int t = itob(1+i);
-		if( !CdControl(CdlGetTD, (u_char*)&t, (u_char*)&toc[i]) )
+		if( !CdControl(CdlGetTD, (uint8_t*)&t, (uint8_t*)&toc[i]) )
 		{
 			return 0;
 		}
@@ -234,11 +234,11 @@ int CdMode(void)
 
 // CD data read routines
 volatile int _cd_sector_count = 0;
-volatile u_long *_cd_read_addr;
-volatile u_char _cd_read_result[8];
-volatile u_long _cd_read_oldcb;
-volatile u_long _cd_read_sector_sz;
-volatile u_long _cd_read_counter;
+volatile uint32_t *_cd_read_addr;
+volatile uint8_t _cd_read_result[8];
+volatile uint32_t _cd_read_oldcb;
+volatile uint32_t _cd_read_sector_sz;
+volatile uint32_t _cd_read_counter;
 
 
 
@@ -279,7 +279,7 @@ static void _CdReadReadyCallback(int status, unsigned char *result)
 	}
 }
 
-int CdRead(int sectors, u_long *buf, int mode)
+int CdRead(int sectors, uint32_t *buf, int mode)
 {
 	// Set sectors to read count
 	_cd_sector_count = sectors;
@@ -307,10 +307,10 @@ int CdRead(int sectors, u_long *buf, int mode)
 	_cd_read_oldcb = CdReadyCallback(_CdReadReadyCallback);
 	
 	// Set specified mode
-	CdControl(CdlSetmode, (u_char*)&mode, 0);
+	CdControl(CdlSetmode, (uint8_t*)&mode, 0);
 	
 	// Begin reading sectors
-	CdControl(CdlReadN, 0, (u_char*)_cd_read_result);
+	CdControl(CdlReadN, 0, (uint8_t*)_cd_read_result);
 	
 	return 0;
 }
@@ -336,10 +336,10 @@ static void CdDoRetry()
 	
 	// Retry read
 	CdControl(CdlSetloc, (void*)&_cd_last_setloc, 0);
-	CdControl(CdlReadN, 0, (u_char*)_cd_read_result);
+	CdControl(CdlReadN, 0, (uint8_t*)_cd_read_result);
 }
 
-int CdReadSync(int mode, u_char *result)
+int CdReadSync(int mode, uint8_t *result)
 {
 	if( (VSync(-1)-_cd_read_counter) > READ_TIMEOUT )
 	{
@@ -371,7 +371,7 @@ int CdReadSync(int mode, u_char *result)
 	return 0;
 }
 
-u_long CdReadCallback(CdlCB func)
+uint32_t CdReadCallback(CdlCB func)
 {
 	unsigned int old_func;
 	
diff --git a/libpsn00b/psxcd/psxcd_asm.s b/libpsn00b/psxcd/psxcd_asm.s
index 129bc89..906ab32 100644
--- a/libpsn00b/psxcd/psxcd_asm.s
+++ b/libpsn00b/psxcd/psxcd_asm.s
@@ -1,6 +1,6 @@
 .set noreorder
 
-.include "hwregs_a.h"
+.include "hwregs_a.inc"
 
 .section .text
 
diff --git a/libpsn00b/psxetc/_dl_resolve_wrapper.s b/libpsn00b/psxetc/_dl_resolve_wrapper.s
index 01ebf3a..eedfa10 100644
--- a/libpsn00b/psxetc/_dl_resolve_wrapper.s
+++ b/libpsn00b/psxetc/_dl_resolve_wrapper.s
@@ -10,11 +10,11 @@
 # - $t8 = index of the function in the .dynsym symbol table
 # - $t9 = _dl_resolve_wrapper itself's address
 
-.set		noreorder
-.section	.text
+.set noreorder
 
-.global	_dl_resolve_wrapper
-.type	_dl_resolve_wrapper, @function
+.section .text._dl_resolve_wrapper
+.global _dl_resolve_wrapper
+.type _dl_resolve_wrapper, @function
 _dl_resolve_wrapper:
 	# Push the registers we're going to use onto the stack.
 	addiu $sp, -16
@@ -47,10 +47,3 @@ _dl_resolve_wrapper:
 
 	jr    $t0
 	nop
-
-.section	.data
-
-.global	_dl_credits
-.type	_dl_credits, @object
-_dl_credits:
-	.asciiz "psxetc runtime dynamic linker by spicyjpeg\n"
diff --git a/libpsn00b/psxetc/dmacallback.s b/libpsn00b/psxetc/dmacallback.s
deleted file mode 100644
index 8ea8ec0..0000000
--- a/libpsn00b/psxetc/dmacallback.s
+++ /dev/null
@@ -1,191 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.section .text
-
-.global DMACallback
-.type DMACallback, @function
-DMACallback:
-
-	# a0 - DMA channel
-	# a1 - Callback function
-	
-	addiu	$sp, -8
-	sw		$ra, 0($sp)
-
-	beqz	$a1, .Lremove_cb		# Remove callback if function is NULL
-	nop
-	
-	addiu	$sp, -8					# Install IRQ handler for DMA handler
-	sw		$a0, 0($sp)				# if not set installed yet
-	sw		$a1, 4($sp)
-	
-	jal		GetInterruptCallback
-	li		$a0, 3
-	
-	bnez	$v0, .Lskip_install
-	nop
-	
-	la		$a1, _dma_handler
-	jal		InterruptCallback
-	li		$a0, 3
-
-.Lskip_install:
-	
-	lw		$a0, 0($sp)
-	lw		$a1, 4($sp)
-	addiu	$sp, 8
-	
-	la		$v0, _dma_func_table
-	sll		$v1, $a0, 2
-	addu	$v0, $v1
-	lw		$v1, 0($v0)
-	sw		$a1, 0($v0)
-	sw		$v1, 4($sp)
-	
-	lui		$a2, IOBASE
-	
-	lw		$v0, DMA_DICR($a2)		# Enable DMA interrupt
-	lui		$v1, 0x1
-	sll		$v1, $a0
-	or		$v0, $v1
-	lui		$v1, 0x80
-	or		$v0, $v1
-	sw		$v0, DMA_DICR($a2)
-	
-	b		.Lskip_remove
-	nop
-	
-.Lremove_cb:
-
-	la		$v0, _dma_func_table	# Set callback address
-	sll		$v1, $a0, 2
-	addu	$v0, $v1
-	lw		$v1, 0($v0)
-	sw		$a1, 0($v0)
-	sw		$v1, 4($sp)
-	
-	lui		$a2, IOBASE				# Disable DMA interrupt
-	lw		$v0, DMA_DICR($a2)
-	lui		$v1, 0x1
-	sll		$v1, $a0
-	.set noat
-	addiu	$at, $0, -1
-	xor		$v1, $at
-	and		$v0, $v1
-	lui		$v1, 0x7f00
-	xor		$v1, $at
-	and		$v0, $v1
-	.set at
-	sw		$v0, DMA_DICR($a2)
-	
-	jal		_dma_has_cb				# Check if callbacks are present
-	nop
-	bnez	$v0, .Lskip_remove
-	nop
-	sw		$0 , DMA_DICR($a2)
-	
-	jal		GetInterruptCallback	# Check if callback is the DMA handler
-	li		$a0, 3
-	la		$v1, _dma_handler
-	bne		$v0, $v1, .Lskip_remove
-	nop
-	
-	li		$a0, 3
-	jal		InterruptCallback
-	move	$a1, $0
-	
-.Lskip_remove:
-	
-	lw		$ra, 0($sp)
-	lw		$v0, 4($sp)
-	jr		$ra
-	addiu	$sp, 8
-
-	
-.type _dma_has_cb, @function
-_dma_has_cb:
-
-	la		$v1, _dma_func_table
-	li		$t0, 6
-	
-.Lscan_loop:
-
-	lw		$v0, 0($v1)
-	addiu	$v1, 4
-	bnez	$v0, .Lhas_cb
-	nop
-	
-	bgtz	$t0, .Lscan_loop
-	addiu	$t0, -1
-	
-	jr		$ra
-	move	$v0, $0
-	
-.Lhas_cb:
-
-	jr		$ra
-	li		$v0, 1
-	
-	
-.type _dma_handler, @function
-_dma_handler:
-
-	addiu	$sp, -12
-	sw		$ra, 0($sp)
-	sw		$s0, 4($sp)
-	sw		$s1, 8($sp)
-	
-	move	$s0, $0
-	la		$s1, _dma_func_table
-	
-.Lhandler_loop:
-
-	lui		$a0, IOBASE
-	lw		$v0, DMA_DICR($a0)
-	li		$v1, 24
-	addu	$v1, $s0
-	srl		$v0, $v1
-	andi	$v0, 0x1
-	
-	lw		$v1, 0($s1)
-	
-	beqz	$v0, .Lno_irq
-	addiu	$s1, 4
-	
-	beqz	$v1, .Lno_irq
-	nop
-	
-	jalr	$v1
-	nop
-	
-.Lno_irq:
-	
-	blt		$s0, 6, .Lhandler_loop
-	addi	$s0, 1
-	
-	lui		$a0, IOBASE
-	lw		$v0, DMA_DICR($a0)
-	nop
-	sw		$v0, DMA_DICR($a0)
-	
-	lw		$ra, 0($sp)
-	lw		$s0, 4($sp)
-	lw		$s1, 8($sp)
-	
-	jr		$ra
-	addiu	$sp, -12
-	
-	
-.section .data
-
-_dma_func_table:
-	.word	0
-	.word	0
-	.word	0
-	.word	0
-	.word	0
-	.word	0
-	.word	0
-	
-\ No newline at end of file
diff --git a/libpsn00b/psxetc/getinterruptcallback.s b/libpsn00b/psxetc/getinterruptcallback.s
deleted file mode 100644
index 510447f..0000000
--- a/libpsn00b/psxetc/getinterruptcallback.s
+++ /dev/null
@@ -1,19 +0,0 @@
-.set noreorder
-
-.section .text
-
-.global GetInterruptCallback
-.type GetInterruptCallback, @function
-GetInterruptCallback:
-	
-	# a0 - Interrupt number
-	
-	la		$a1, _irq_func_table
-	sll		$a0, 2
-	addu	$a1, $a0
-	lw		$v0, 0($a1)
-	nop
-	jr		$ra
-	nop
-	
-	
-\ No newline at end of file
diff --git a/libpsn00b/psxetc/interruptcallback.s b/libpsn00b/psxetc/interruptcallback.s
deleted file mode 100644
index 78e5e6e..0000000
--- a/libpsn00b/psxetc/interruptcallback.s
+++ /dev/null
@@ -1,48 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.section .text
-
-.global InterruptCallback
-.type InterruptCallback, @function
-InterruptCallback:
-
-	# a0 - Interrupt number
-	# a1 - Callback function
-
-	lui		$a2, IOBASE
-	
-	beqz	$a1, .Ldisable_irq
-	nop
-
-	lw		$v0, IRQ_MASK($a2)			# Enable interrupt mask
-	li		$v1, 1
-	sll		$v1, $a0
-	or		$v0, $v1
-
-	b		.Lcont
-	sw		$v0, IRQ_MASK($a2)
-	
-.Ldisable_irq:
-
-.set noat
-	lw		$v0, IRQ_MASK($a2)			# Disable interrupt mask
-	li		$v1, 1
-	sll		$v1, $a0
-	addiu	$at, $0 , -1
-	xor		$v1, $at
-.set at
-	and		$v0, $v1
-	sw		$v0, IRQ_MASK($a2)
-	
-.Lcont:
-
-	la		$a2, _irq_func_table		# Get address to IRQ function table
-	
-	sll		$v1, $a0, 2					# Compute the slot
-	addu	$v1, $a2, $v1
-	lw		$v0, 0($v1)					# Get old handler address
-	
-	jr		$ra							# Return and set new IRQ handler
-	sw		$a1, 0($v1)
diff --git a/libpsn00b/psxetc/interrupts.c b/libpsn00b/psxetc/interrupts.c
new file mode 100644
index 0000000..859209a
--- /dev/null
+++ b/libpsn00b/psxetc/interrupts.c
@@ -0,0 +1,225 @@
+/*
+ * PSn00bSDK interrupt management library
+ * (C) 2022 spicyjpeg - MPL licensed
+ */
+
+#include <stdint.h>
+#include <psxapi.h>
+#include <psxetc.h>
+#include <hwregs_c.h>
+
+#define NUM_IRQ_CHANNELS	11
+#define NUM_DMA_CHANNELS	7
+#define ISR_STACK_SIZE		0x1000
+
+/* Internal globals */
+
+static void (*_irq_handlers[NUM_IRQ_CHANNELS])(void);
+static void (*_dma_handlers[NUM_DMA_CHANNELS])(void);
+static int  _num_dma_handlers = 0;
+
+static uint16_t _saved_irq_mask;
+static uint32_t _saved_dma_dpcr, _saved_dma_dicr;
+static int _isr_installed = 0;
+
+/* Custom ISR jmp_buf */
+
+// The ISR and all functions called by it (thus, all callbacks registered using
+// InterruptCallback() and DMACallback()) use an independent stack, isolated
+// from the main thread's stack. As the size of this stack is limited, custom
+// callbacks shall keep the number of nested subroutine calls to a minimum and
+// avoid allocating large buffers (e.g. for receiving a sector from the CD
+// drive) on the stack.
+static uint8_t _isr_stack[ISR_STACK_SIZE];
+
+extern uint8_t _gp[];
+static void _global_isr(void);
+
+static const struct JMP_BUF _isr_jmp_buf = {
+	.ra = (uint32_t) &_global_isr,
+	.sp = (uint32_t) &_isr_stack[ISR_STACK_SIZE],
+	.fp = 0,
+	.s0 = 0,
+	.s1 = 0,
+	.s2 = 0,
+	.s3 = 0,
+	.s4 = 0,
+	.s5 = 0,
+	.s6 = 0,
+	.s7 = 0,
+	.gp = (uint32_t) _gp
+};
+
+/* Internal IRQ and DMA handlers */
+
+static void _global_isr(void) {
+	uint16_t stat = IRQ_STAT, mask = IRQ_MASK;
+
+	// Clear all IRQ flags in one shot. This is not the "proper" way to do it
+	// but it's much faster than clearing one flag at a time.
+	IRQ_STAT = ~mask;
+
+	//for (int i = 0; i < NUM_IRQ_CHANNELS; i++) {
+	for (int i = 0; stat; i++, stat >>= 1) {
+		if (!(stat & 1))
+			continue;
+
+		if (_irq_handlers[i])
+			_irq_handlers[i]();
+	}
+
+	ReturnFromException();
+}
+
+static void _global_dma_handler(void) {
+	uint32_t stat = DMA_DICR;
+
+	// Clear all DMA IRQ flags in one shot (note that flags are cleared by
+	// writing 1 to them rather than 0).
+	stat    &= 0x7fff0000;
+	DMA_DICR = stat;
+	stat   >>= 24;
+
+	//for (int i = 0; i < NUM_DMA_CHANNELS; i++) {
+		for (int i = 0; stat; i++, stat >>= 1) {
+		if (!(stat & 1))
+			continue;
+
+		if (_dma_handlers[i])
+			_dma_handlers[i]();
+	}
+}
+
+/* Callback registration API */
+
+void *InterruptCallback(int irq, void (*func)(void)) {
+	if ((irq < 0) || (irq >= NUM_IRQ_CHANNELS))
+		return 0;
+
+	void *old_callback = _irq_handlers[irq];
+	_irq_handlers[irq] = func;
+
+	// Enable or disable the IRQ in the IRQ_MASK register depending on whether
+	// the callback is being registered or removed.
+	if (func)
+		IRQ_MASK |= 1 << irq;
+	else
+		IRQ_MASK &= ~(1 << irq);
+
+	return old_callback;
+}
+
+void *GetInterruptCallback(int irq) {
+	if ((irq < 0) || (irq >= NUM_IRQ_CHANNELS))
+		return 0;
+
+	return _irq_handlers[irq];
+}
+
+void *DMACallback(int dma, void (*func)(void)) {
+	if ((dma < 0) || (dma >= NUM_DMA_CHANNELS))
+		return 0;
+
+	void *old_callback = _dma_handlers[dma];
+	_dma_handlers[dma] = func;
+
+	// Enable or disable the IRQ in the DMA_DICR register depending on whether
+	// the callback is being registered or removed. The main DMA IRQ dispatcher
+	// is also registered if this is the first DMA callback being configured,
+	// or disabled if it's the last one being removed.
+	if (func) {
+		DMA_DICR |= (0x10000 << dma) | (1 << 23);
+
+		if (!(_num_dma_handlers++))
+			InterruptCallback(3, &_global_dma_handler);
+	} else {
+		if (--_num_dma_handlers) {
+			DMA_DICR &= ~(0x10000 << dma);
+		} else {
+			DMA_DICR = 0;
+			InterruptCallback(3, 0);
+		}
+	}
+
+	return old_callback;
+}
+
+void *GetDMACallback(int dma) {
+	if ((dma < 0) || (dma >= NUM_DMA_CHANNELS))
+		return 0;
+
+	return _dma_handlers[dma];
+}
+
+/* Hook installation/removal API */
+
+int ResetCallback(void) {
+	if (_isr_installed)
+		return -1;
+
+	EnterCriticalSection();
+	_saved_irq_mask = 1 << 3; // Enable DMA IRQ by default
+	_saved_dma_dpcr = 0x03333333;
+	_saved_dma_dicr = 0;
+
+	for (int i = 0; i < NUM_IRQ_CHANNELS; i++)
+		_irq_handlers[i] = (void *) 0;
+	for (int i = 0; i < NUM_DMA_CHANNELS; i++)
+		_dma_handlers[i] = (void *) 0;
+
+	// Set up the DMA IRQ handler. This handler shall *not* be overridden using
+	// InterruptCallback().
+	_irq_handlers[3] = &_global_dma_handler;
+
+	_96_remove();
+	RestartCallback();
+	return 0;
+}
+
+void RestartCallback(void) {
+	if (_isr_installed)
+		return;
+
+	IRQ_STAT = 0;
+	IRQ_MASK = _saved_irq_mask;
+	DMA_DPCR = _saved_dma_dpcr;
+	DMA_DICR = _saved_dma_dicr;
+
+	// Install the ISR hook and prevent the kernel's internal handlers from
+	// automatically acknowledging SPI and timer IRQs.
+	SetCustomExitFromException(&_isr_jmp_buf);
+	ChangeClearPAD(0);
+	ChangeClearRCnt(0, 0);
+	ChangeClearRCnt(1, 0);
+	ChangeClearRCnt(2, 0);
+	ChangeClearRCnt(3, 0);
+
+	_isr_installed = 1;
+	ExitCriticalSection();
+}
+
+void StopCallback(void) {
+	if (!_isr_installed)
+		return;
+
+	// Save the state of IRQ and DMA registers, then reset them and undo the
+	// changes that were made to the kernel's state.
+	EnterCriticalSection();
+	_saved_irq_mask = IRQ_MASK;
+	_saved_dma_dpcr = DMA_DPCR;
+	_saved_dma_dicr = DMA_DICR;
+
+	IRQ_STAT = 0;
+	IRQ_MASK = 0;
+	DMA_DPCR = _saved_dma_dpcr & 0x07777777;
+	DMA_DICR = 0;
+
+	SetDefaultExitFromException();
+	ChangeClearPAD(1);
+	ChangeClearRCnt(0, 1);
+	ChangeClearRCnt(1, 1);
+	ChangeClearRCnt(2, 1);
+	ChangeClearRCnt(3, 1);
+
+	_isr_installed = 0;
+}
diff --git a/libpsn00b/psxetc/isr.s b/libpsn00b/psxetc/isr.s
deleted file mode 100644
index 440be50..0000000
--- a/libpsn00b/psxetc/isr.s
+++ /dev/null
@@ -1,107 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.set ISR_STACK_SIZE, 4096
-
-
-.section .text
-
-# Global ISR handler of PSn00bSDK
-
-.set at
-
-.type _global_isr, @function
-_global_isr:
-	
-.Lisr_loop:
-
-	#la		$gp, _gp					# Keep restoring GP since it gets
-										# changed elsewhere sometimes
-	
-	lui		$a0, IOBASE					# Get IRQ status
-	lw		$v0, IRQ_MASK($a0)
-	nop
-	
-	srl		$v0, $s1					# Check IRQ mask bit if set
-	andi	$v0, 0x1
-	
-	beqz	$v0, .Lno_irq				# Don't execute callback if IRQ not enabled
-	nop
-	
-	lw		$v0, IRQ_STAT($a0)
-	nop
-	srl		$v0, $s1					# Check IRQ status bit if set
-	andi	$v0, 0x1
-	beqz	$v0, .Lno_irq				# Don't execute callback if no IRQ
-	nop
-	
-	lw		$v1, 0($s0)					# Load IRQ callback function
-	nop
-	
-	lw		$v0, IRQ_STAT($a0)			# Acknowledge the IRQ (by writing a 0 bit)
-	li		$a1, 1
-	sll		$a1, $s1
-	addiu	$a2, $0 , -1
-	xor		$a1, $a2
-	sw		$a1, IRQ_STAT($a0)
-	
-	beqz	$v1, .Lno_irq				# Don't execute if callback is not set
-	nop
-
-	jalr	$v1							# Call interrupt handler
-	nop
-	
-.Lno_irq:
-
-	addiu	$s0, 4
-	
-	blt		$s1, 11, .Lisr_loop
-	addiu	$s1, 1
-
-	j		ReturnFromException
-	nop
-
-
-.section .data
-
-# Global ISR callback table
-
-.global _irq_func_table
-.type _irq_func_table, @object
-_irq_func_table:
-	.word	0
-	.word	0
-	.word	0
-	.word	0
-	.word	0
-	.word	0
-	.word	0
-	.word	0
-	.word	0
-	.word	0
-	.word	0
-	.word	0
-
-# Global ISR hook structure
-.global _custom_exit
-.type _custom_exit, @object
-_custom_exit:
-	.word _global_isr			# pc
-	.word _isr_stack+ISR_STACK_SIZE	# sp
-	.word 0						# fp
-	.word _irq_func_table		# s0
-	.word 0						# s1
-	.word 0						# s2
-	.word 0						# s3
-	.word 0						# s4
-	.word 0						# s5
-	.word 0						# s6
-	.word 0						# s7
-	.word _gp					# gp
-	
-# Global ISR stack
-#	.fill 1024
-#_custom_exit_stack:
-#	.fill 4
-.comm _isr_stack, ISR_STACK_SIZE+4
diff --git a/libpsn00b/psxetc/restartcallback.s b/libpsn00b/psxetc/restartcallback.s
deleted file mode 100644
index 036a5a0..0000000
--- a/libpsn00b/psxetc/restartcallback.s
+++ /dev/null
@@ -1,54 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.section .text
-
-.global RestartCallback
-.type RestartCallback, @function
-RestartCallback:
-
-	addiu	$sp, -4
-	sw		$ra, 0($sp)
-	
-	la		$a0, _custom_exit
-	jal		SetCustomExitFromException
-	addiu	$sp, -4
-	addiu	$sp, 4
-	
-	move	$a0, $0
-	jal		ChangeClearPAD
-	addiu	$sp, -4
-	addiu	$sp, 4
-	
-	li		$a0, 3
-	move	$a1, $0
-	jal		ChangeClearRCnt
-	addiu	$sp, -8
-	addiu	$sp, 8
-	
-	la		$a0, _irq_func_table
-	move	$a1, $0
-	move	$v0, $0
-	
-.Lcheck_cbs:							# Set up the interrupt masks
-	lw		$v1, 0($a0)
-	nop
-	beqz	$v1, .Lno_cb
-	addiu	$a0, 4
-	li		$v1, 1
-	sll		$v1, $a1
-	or		$v0, $v1
-.Lno_cb:
-	blt		$a1, 10, .Lcheck_cbs
-	addiu	$a1, 1
-	
-	lui		$a0, IOBASE
-	sw		$0 , IRQ_STAT($a0)
-	sw		$v0, IRQ_MASK($a0)
-	
-	lw		$ra, 0($sp)
-	addiu	$sp, 4
-	jr		$ra
-	nop
-	
-\ No newline at end of file
diff --git a/libpsn00b/psxgpu/addprim.s b/libpsn00b/psxgpu/addprim.s
deleted file mode 100644
index 1b66274..0000000
--- a/libpsn00b/psxgpu/addprim.s
+++ /dev/null
@@ -1,26 +0,0 @@
-.set noreorder
-.set noat
-
-.section .text
-
-
-.global AddPrim
-.type AddPrim, @function
-AddPrim:
-
-	lw		$v0, 0($a0)				# Load OT entry
-	lw		$v1, 0($a1)				# Set packet length value (in words)
-	lui		$at, 0x00ff
-	or		$at, 0xffff
-	and		$v0, $at				# Mask off the upper 8 bits of OT entry
-	or		$v1, $v0				# OR values together
-	sw		$v1, 0($a1)				# Store new address to primitive tag
-	lw		$v0, 0($a0)				# Load OT entry
-	and		$a1, $at				# Mask off the upper 8 bits of primitive tag
-	lui		$at, 0xff00
-	and		$v0, $at				# Mask off the first 24 bits of OT entry
-	or		$v0, $a1				# OR values together
-
-	jr		$ra
-	sw		$v0, 0($a0)				# Store result to OT
-
diff --git a/libpsn00b/psxgpu/clearotagr.s b/libpsn00b/psxgpu/clearotagr.s
deleted file mode 100644
index 562cad4..0000000
--- a/libpsn00b/psxgpu/clearotagr.s
+++ /dev/null
@@ -1,21 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.section .text
-
-
-.global ClearOTagR
-.type ClearOTagR, @function
-ClearOTagR:
-	lui		$a2, IOBASE
-	addi	$v0, $a1, -1
-	sll		$v0, 2
-	addu	$a0, $v0
-	sw		$a0, DMA6_MADR($a2)
-	andi	$a1, 0xffff
-	sw		$a1, DMA6_BCR($a2)
-	lui		$v0, 0x1100
-	addiu	$v0, 2
-	jr		$ra
-	sw		$v0, DMA6_CHCR($a2)
diff --git a/libpsn00b/psxgpu/common.c b/libpsn00b/psxgpu/common.c
new file mode 100644
index 0000000..cef1508
--- /dev/null
+++ b/libpsn00b/psxgpu/common.c
@@ -0,0 +1,299 @@
+/*
+ * PSn00bSDK GPU library (common functions)
+ * (C) 2022 spicyjpeg - MPL licensed
+ */
+
+#include <stdint.h>
+#include <stdio.h>
+#include <psxetc.h>
+#include <psxapi.h>
+#include <psxgpu.h>
+#include <hwregs_c.h>
+
+#define QUEUE_LENGTH		16
+#define DMA_CHUNK_LENGTH	8
+#define VSYNC_TIMEOUT		0x100000
+
+static void _default_vsync_halt(void);
+
+/* Internal globals */
+
+GPU_VideoMode _gpu_video_mode;
+
+static void (*_vsync_halt_func)(void)   = &_default_vsync_halt;
+static void (*_vsync_callback)(void)    = (void *) 0;
+static void (*_drawsync_callback)(void) = (void *) 0;
+
+static const uint32_t *volatile _draw_queue[QUEUE_LENGTH];
+static volatile uint8_t  _queue_head, _queue_tail, _queue_length;
+static volatile uint32_t _vblank_counter;
+static volatile uint16_t _last_hblank;
+
+/* Interrupt handlers */
+
+static void _vblank_handler(void) {
+	_vblank_counter++;
+
+	if (_vsync_callback)
+		_vsync_callback();
+}
+
+static void _gpu_dma_handler(void) {
+	//while (!(GPU_GP1 & (1 << 26)) || (DMA_CHCR(2) & (1 << 24)))
+	while (!(GPU_GP1 & (1 << 26)))
+		__asm__ volatile("");
+
+	if (_queue_length) {
+		DrawOTag2(_draw_queue[_queue_head++]);
+
+		_queue_length--;
+		_queue_head %= QUEUE_LENGTH;
+	} else {
+		GPU_GP1 = 0x04000000; // Disable DMA request
+
+		if (_drawsync_callback)
+			_drawsync_callback();
+	}
+}
+
+/* GPU reset and system initialization */
+
+void ResetGraph(int mode) {
+	// Perform some basic system initialization when ResetGraph() is called for
+	// the first time.
+	if (!ResetCallback()) {
+		EnterCriticalSection();
+		InterruptCallback(0, &_vblank_handler);
+		DMACallback(2, &_gpu_dma_handler);
+
+		_gpu_video_mode = (GPU_GP1 >> 20) & 1;
+		ExitCriticalSection();
+
+		printf("psxgpu: setup done, default mode is %s\n", _gpu_video_mode ? "PAL" : "NTSC");
+	}
+
+	if (mode == 3) {
+		GPU_GP1 = 0x01000000; // Reset command buffer
+		return;
+	}
+
+	DMA_DPCR   |= 0x0b000b00; // Enable DMA2 and DMA6
+	DMA_CHCR(2) = 0x00000201; // Stop DMA2
+	DMA_CHCR(6) = 0x00000200; // Stop DMA6
+
+	if (mode == 1) {
+		GPU_GP1 = 0x01000000; // Reset command buffer
+		return;
+	}
+
+	GPU_GP1       = 0x00000000; // Reset GPU
+	TIMER_CTRL(0) = 0x0500;
+	TIMER_CTRL(1) = 0x0500;
+
+	_queue_head     = 0;
+	_queue_tail     = 0;
+	_queue_length   = 0;
+	_vblank_counter = 0;
+	_last_hblank    = 0;
+}
+
+/* Syncing API */
+
+// TODO: add support for no$psx's "halt" register
+static void _default_vsync_halt(void) {
+	int counter = _vblank_counter;
+
+	for (int i = VSYNC_TIMEOUT; i; i--) {
+		if (counter != _vblank_counter)
+			return;
+	}
+
+	printf("psxgpu: VSync() timeout\n");
+	ChangeClearPAD(0);
+	ChangeClearRCnt(3, 0);
+}
+
+int VSync(int mode) {
+	uint16_t delta = (TIMER_VALUE(1) - _last_hblank) & 0xffff;
+	if (mode == 1)
+		return delta;
+	if (mode < 0)
+		return _vblank_counter;
+
+	uint32_t status = GPU_GP1;
+
+	// Wait for at least one vertical blank event to occur.
+	do {
+		_vsync_halt_func();
+
+		// If interlaced mode is enabled, wait until the GPU starts displaying
+		// the next field.
+		if (status & (1 << 22)) {
+			while (!((GPU_GP1 ^ status) & (1 << 31)))
+				__asm__ volatile("");
+		}
+	} while ((--mode) > 0);
+
+	_last_hblank = TIMER_VALUE(1);
+	return delta;
+}
+
+int DrawSync(int mode) {
+	if (mode)
+		return (DMA_BCR(2) >> 16);
+
+	// Wait for the queue to become empty.
+	// TODO: add a timeout
+	while (_queue_length)
+		__asm__ volatile("");
+
+	// Wait for any DMA transfer to finish if DMA is enabled.
+	if (GPU_GP1 & (3 << 29)) {
+		while (!(GPU_GP1 & (1 << 28)) || (DMA_CHCR(2) & (1 << 24)))
+			__asm__ volatile("");
+	}
+
+	while (!(GPU_GP1 & (1 << 26)))
+		__asm__ volatile("");
+
+	return 0;
+}
+
+void *VSyncHaltFunction(void (*func)(void)) {
+	void *old_callback  = _vsync_halt_func;
+	_vsync_halt_func    = func;
+
+	return old_callback;
+}
+
+void *VSyncCallback(void (*func)(void)) {
+	EnterCriticalSection();
+
+	void *old_callback  = _vsync_callback;
+	_vsync_callback     = func;
+
+	ExitCriticalSection();
+	return old_callback;
+}
+
+void *DrawSyncCallback(void (*func)(void)) {
+	EnterCriticalSection();
+
+	void *old_callback = _drawsync_callback;
+	_drawsync_callback = func;
+
+	ExitCriticalSection();
+	return old_callback;
+}
+
+/* OT and primitive drawing API */
+
+void ClearOTagR(uint32_t *ot, size_t length) {
+	DMA_MADR(6) = (uint32_t) &ot[length - 1];
+	DMA_BCR(6)  = length & 0xffff;
+	DMA_CHCR(6) = 0x11000002;
+
+	//while (DMA_CHCR(6) & (1 << 24))
+		//__asm__ volatile("");
+}
+
+void ClearOTag(uint32_t *ot, size_t length) {
+	// DMA6 only supports writing to RAM in reverse order (last to first), so
+	// the OT has to be cleared in software here. This function is thus much
+	// slower than ClearOTagR().
+	// https://problemkaputt.de/psx-spx.htm#dmachannels
+	for (int i = 0; i < (length - 1); i++)
+		ot[i] = (uint32_t) &ot[i + 1] & 0x00ffffff;
+
+	ot[length - 1] = 0x00ffffff;
+}
+
+void DrawOTag(const uint32_t *ot) {
+	// If GPU DMA is currently busy, append the OT to the queue instead of
+	// drawing it immediately. Note that interrupts must be disabled *prior* to
+	// checking if DMA is busy; disabling them afterwards would create a race
+	// condition where the DMA transfer could end while interrupts are being
+	// disabled. Interrupts are disabled through the IRQ_MASK register rather
+	// than by calling EnterCriticalSection() for performance reasons.
+	uint16_t mask = IRQ_MASK;
+	IRQ_MASK      = 0;
+
+	if (DMA_CHCR(2) & (1 << 24)) {
+		if (_queue_length < QUEUE_LENGTH) {
+			_draw_queue[_queue_tail++] = ot;
+
+			_queue_length++;
+			_queue_tail %= QUEUE_LENGTH;
+
+			IRQ_MASK = mask;
+			return;
+		}
+
+		IRQ_MASK = mask;
+		printf("psxgpu: DrawOTag() failed, draw queue full\n");
+		return;
+	}
+
+	IRQ_MASK = mask;
+	DrawOTag2(ot);
+}
+
+void DrawOTag2(const uint32_t *ot) {
+	GPU_GP1 = 0x04000002;
+
+	while (!(GPU_GP1 & (1 << 26)) || (DMA_CHCR(2) & (1 << 24)))
+		__asm__ volatile("");
+
+	DMA_MADR(2) = (uint32_t) ot;
+	DMA_BCR(2)  = 0;
+	DMA_CHCR(2) = 0x01000401;
+}
+
+void DrawPrim(const uint32_t *pri) {	
+	size_t length = getlen(pri);
+
+	DrawSync(0);
+	GPU_GP1 = 0x04000002;
+
+	// NOTE: if length >= DMA_CHUNK_LENGTH then it also has to be a multiple of
+	// DMA_CHUNK_LENGTH, otherwise the DMA channel will get stuck waiting for
+	// more data indefinitely.
+	DMA_MADR(2) = (uint32_t) &pri[1];
+	if (length < DMA_CHUNK_LENGTH)
+		DMA_BCR(2) = 0x00010000 | length;
+	else
+		DMA_BCR(2) = DMA_CHUNK_LENGTH | ((length / DMA_CHUNK_LENGTH) << 16);
+
+	DMA_CHCR(2) = 0x01000201;
+}
+
+void AddPrim(uint32_t *ot, const void *pri) {
+	addPrim(ot, pri);
+}
+
+/* Misc. functions */
+
+GPU_VideoMode GetVideoMode(void) {
+	return _gpu_video_mode;
+}
+
+void SetVideoMode(GPU_VideoMode mode) {
+	uint32_t _mode, stat = GPU_GP1;
+
+	_gpu_video_mode = mode & 1;
+
+	_mode  = (mode & 1) << 3;
+	_mode |= (stat >> 17) & 0x37; // GPUSTAT 17-22  -> cmd bits 0-5
+	_mode |= (stat >> 10) & 0x40; // GPUSTAT bit 16 -> cmd bit 6
+	_mode |= (stat >>  7) & 0x80; // GPUSTAT bit 14 -> cmd bit 7
+
+	GPU_GP1 = 0x08000000 | mode;
+}
+
+int GetODE(void) {
+	return (GPU_GP1 >> 31);
+}
+
+void SetDispMask(int mask) {
+	GPU_GP1 = 0x03000000 | (mask ? 0 : 1);
+}
diff --git a/libpsn00b/psxgpu/dbugfont.c b/libpsn00b/psxgpu/dbugfont.c
deleted file mode 100644
index ff21d84..0000000
--- a/libpsn00b/psxgpu/dbugfont.c
+++ /dev/null
@@ -1,144 +0,0 @@
-unsigned int dbugfont_size=2112;
-unsigned char dbugfont[] = {
-0x10,0x00,0x00,0x00,0x08,0x00,0x00,0x00,0x2c,0x00,0x00,0x00,0x00,0x00,0xe0,
-0x01,0x10,0x00,0x01,0x00,0x00,0x00,0xff,0xff,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x0c,0x08,0x00,0x00,0x80,0x02,0x00,0x00,
-0x20,0x00,0x20,0x00,0x00,0x01,0x00,0x00,0x00,0x01,0x01,0x00,0x00,0x01,0x01,
-0x00,0x10,0x11,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x11,0x00,0x00,0x00,0x11,
-0x00,0x00,0x00,0x10,0x00,0x00,0x10,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x11,0x01,0x00,0x00,0x01,0x00,0x00,0x00,0x01,0x01,
-0x00,0x00,0x01,0x01,0x00,0x01,0x01,0x01,0x00,0x10,0x00,0x10,0x00,0x10,0x00,
-0x01,0x00,0x00,0x11,0x00,0x00,0x00,0x01,0x00,0x00,0x00,0x01,0x00,0x00,0x00,
-0x01,0x00,0x00,0x00,0x01,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x00,0x01,0x00,
-0x00,0x00,0x00,0x00,0x00,0x10,0x11,0x11,0x00,0x01,0x01,0x00,0x00,0x11,0x01,
-0x01,0x00,0x10,0x00,0x00,0x00,0x00,0x01,0x00,0x00,0x10,0x00,0x00,0x00,0x00,
-0x10,0x00,0x00,0x01,0x01,0x01,0x00,0x00,0x01,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x01,0x00,0x10,0x00,0x11,
-0x00,0x00,0x01,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x01,0x01,0x00,0x10,0x11,
-0x00,0x00,0x10,0x10,0x00,0x00,0x01,0x01,0x01,0x00,0x00,0x00,0x00,0x00,0x10,
-0x00,0x00,0x00,0x00,0x10,0x00,0x00,0x10,0x11,0x00,0x00,0x11,0x11,0x01,0x00,
-0x00,0x00,0x00,0x00,0x11,0x11,0x01,0x00,0x00,0x00,0x00,0x00,0x00,0x10,0x00,
-0x00,0x10,0x10,0x10,0x00,0x00,0x01,0x00,0x00,0x00,0x00,0x00,0x00,0x10,0x11,
-0x11,0x00,0x00,0x01,0x01,0x00,0x00,0x01,0x01,0x00,0x01,0x10,0x01,0x00,0x00,
-0x00,0x00,0x00,0x10,0x00,0x00,0x00,0x00,0x10,0x00,0x00,0x01,0x01,0x01,0x00,
-0x00,0x01,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x01,0x00,0x00,0x10,0x01,0x10,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x01,0x01,0x00,0x01,0x01,0x01,0x00,0x10,0x10,0x11,0x00,0x01,
-0x10,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x01,0x00,0x00,0x00,0x01,0x00,0x00,
-0x00,0x01,0x00,0x00,0x00,0x01,0x00,0x00,0x00,0x11,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x11,0x00,0x00,0x10,0x00,0x00,0x00,0x10,0x00,0x10,0x00,0x00,0x01,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x01,0x01,0x00,0x10,0x11,0x00,0x00,0x01,
-0x00,0x01,0x00,0x10,0x01,0x11,0x00,0x00,0x00,0x00,0x00,0x00,0x10,0x00,0x00,
-0x10,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x11,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x11,0x00,0x00,0x01,0x00,0x00,0x00,0x00,0x11,
-0x01,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x10,0x01,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x10,0x00,0x00,0x00,0x11,0x01,0x00,0x00,
-0x11,0x01,0x00,0x00,0x10,0x01,0x00,0x10,0x11,0x11,0x00,0x00,0x10,0x01,0x00,
-0x10,0x11,0x11,0x00,0x00,0x11,0x01,0x00,0x00,0x11,0x01,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x01,0x00,0x00,0x00,0x00,0x00,0x10,0x00,
-0x00,0x00,0x00,0x11,0x01,0x00,0x00,0x11,0x11,0x00,0x00,0x11,0x00,0x00,0x10,
-0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x00,0x01,0x01,0x00,0x10,0x00,0x00,0x00,
-0x00,0x01,0x00,0x00,0x00,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,
-0x00,0x00,0x11,0x00,0x00,0x00,0x11,0x00,0x00,0x00,0x10,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x01,0x00,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x00,0x01,0x10,
-0x10,0x00,0x00,0x00,0x00,0x10,0x00,0x00,0x00,0x10,0x00,0x10,0x00,0x01,0x00,
-0x10,0x00,0x00,0x00,0x10,0x00,0x00,0x00,0x00,0x00,0x10,0x00,0x10,0x00,0x10,
-0x00,0x10,0x00,0x10,0x00,0x00,0x11,0x00,0x00,0x00,0x11,0x00,0x00,0x00,0x01,
-0x00,0x00,0x10,0x11,0x11,0x00,0x00,0x10,0x00,0x00,0x00,0x00,0x10,0x00,0x10,
-0x10,0x01,0x01,0x00,0x10,0x00,0x00,0x00,0x10,0x01,0x00,0x00,0x10,0x01,0x00,
-0x10,0x11,0x11,0x00,0x10,0x11,0x01,0x00,0x10,0x11,0x01,0x00,0x00,0x00,0x01,
-0x00,0x00,0x11,0x01,0x00,0x00,0x11,0x11,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x10,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x01,0x00,0x00,
-0x00,0x01,0x00,0x10,0x10,0x01,0x01,0x00,0x10,0x00,0x00,0x00,0x01,0x00,0x00,
-0x00,0x00,0x10,0x00,0x00,0x00,0x01,0x00,0x00,0x00,0x10,0x00,0x10,0x00,0x10,
-0x00,0x00,0x10,0x00,0x00,0x10,0x00,0x10,0x00,0x00,0x00,0x10,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x01,0x00,0x00,0x10,0x11,0x11,0x00,0x00,
-0x10,0x00,0x00,0x00,0x10,0x00,0x00,0x10,0x10,0x11,0x00,0x00,0x10,0x00,0x00,
-0x10,0x00,0x00,0x00,0x10,0x00,0x10,0x00,0x00,0x00,0x01,0x00,0x00,0x00,0x10,
-0x00,0x10,0x00,0x10,0x00,0x00,0x10,0x00,0x00,0x10,0x00,0x10,0x00,0x00,0x00,
-0x01,0x00,0x00,0x11,0x00,0x00,0x00,0x11,0x00,0x00,0x00,0x10,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x01,0x00,0x00,0x00,0x00,0x00,0x00,0x10,0x00,0x00,0x00,
-0x10,0x11,0x11,0x00,0x10,0x11,0x11,0x00,0x00,0x11,0x01,0x00,0x00,0x00,0x01,
-0x00,0x10,0x11,0x01,0x00,0x00,0x11,0x01,0x00,0x00,0x10,0x00,0x00,0x00,0x11,
-0x01,0x00,0x00,0x11,0x00,0x00,0x00,0x11,0x00,0x00,0x00,0x11,0x00,0x00,0x00,
-0x00,0x01,0x00,0x00,0x00,0x00,0x00,0x10,0x00,0x00,0x00,0x00,0x10,0x00,0x00,
-0x00,0x11,0x11,0x01,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x10,
-0x01,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x10,0x00,0x00,0x10,0x11,0x01,
-0x00,0x00,0x11,0x01,0x00,0x10,0x11,0x01,0x00,0x10,0x11,0x11,0x00,0x10,0x11,
-0x11,0x00,0x00,0x11,0x11,0x00,0x10,0x00,0x10,0x00,0x10,0x11,0x11,0x00,0x00,
-0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x00,0x00,0x10,0x00,0x10,0x00,
-0x10,0x00,0x10,0x00,0x00,0x11,0x01,0x00,0x10,0x11,0x01,0x00,0x00,0x01,0x01,
-0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,
-0x00,0x00,0x10,0x00,0x00,0x00,0x10,0x00,0x00,0x00,0x10,0x00,0x10,0x00,0x00,
-0x10,0x00,0x00,0x00,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x00,0x00,
-0x10,0x01,0x11,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,
-0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x00,0x00,0x10,0x00,
-0x10,0x00,0x10,0x00,0x00,0x00,0x10,0x00,0x00,0x00,0x10,0x00,0x00,0x00,0x10,
-0x00,0x10,0x00,0x00,0x10,0x00,0x00,0x00,0x00,0x10,0x00,0x10,0x00,0x01,0x00,
-0x10,0x00,0x00,0x00,0x10,0x10,0x10,0x00,0x10,0x01,0x10,0x00,0x10,0x00,0x10,
-0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x11,0x01,0x00,0x10,0x00,
-0x00,0x00,0x10,0x00,0x10,0x00,0x10,0x11,0x00,0x00,0x10,0x11,0x00,0x00,0x10,
-0x10,0x11,0x00,0x10,0x11,0x11,0x00,0x00,0x10,0x00,0x00,0x00,0x00,0x10,0x00,
-0x10,0x11,0x00,0x00,0x10,0x00,0x00,0x00,0x10,0x00,0x10,0x00,0x10,0x10,0x10,
-0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x11,0x11,0x00,0x10,0x00,
-0x10,0x00,0x10,0x00,0x00,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x00,0x00,0x10,
-0x00,0x00,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x00,0x10,0x00,0x00,
-0x10,0x00,0x10,0x00,0x10,0x00,0x01,0x00,0x10,0x00,0x00,0x00,0x10,0x00,0x10,
-0x00,0x10,0x00,0x11,0x00,0x10,0x00,0x10,0x00,0x10,0x11,0x01,0x00,0x10,0x00,
-0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,
-0x00,0x00,0x00,0x10,0x00,0x00,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,
-0x00,0x10,0x00,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x00,
-0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,
-0x00,0x00,0x10,0x00,0x10,0x00,0x10,0x11,0x01,0x00,0x00,0x11,0x01,0x00,0x10,
-0x11,0x01,0x00,0x10,0x11,0x11,0x00,0x10,0x00,0x00,0x00,0x00,0x11,0x01,0x00,
-0x10,0x00,0x10,0x00,0x10,0x11,0x11,0x00,0x00,0x11,0x01,0x00,0x10,0x00,0x10,
-0x00,0x10,0x11,0x11,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x00,0x11,
-0x01,0x00,0x10,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x11,0x01,0x00,0x10,
-0x11,0x01,0x00,0x00,0x11,0x01,0x00,0x10,0x11,0x11,0x00,0x10,0x00,0x10,0x00,
-0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,
-0x00,0x10,0x11,0x11,0x00,0x00,0x11,0x01,0x00,0x01,0x00,0x00,0x00,0x00,0x11,
-0x01,0x00,0x00,0x10,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x11,0x00,0x00,0x10,
-0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x00,0x10,0x00,0x00,
-0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,
-0x00,0x10,0x00,0x10,0x00,0x00,0x00,0x10,0x00,0x00,0x01,0x00,0x00,0x10,0x00,
-0x00,0x00,0x00,0x00,0x01,0x00,0x00,0x11,0x01,0x00,0x00,0x00,0x00,0x00,0x00,
-0x11,0x00,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x00,0x00,
-0x00,0x10,0x00,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,
-0x00,0x00,0x01,0x01,0x00,0x10,0x00,0x10,0x00,0x00,0x00,0x01,0x00,0x00,0x01,
-0x00,0x00,0x00,0x01,0x00,0x00,0x00,0x00,0x01,0x00,0x10,0x01,0x11,0x00,0x00,
-0x00,0x00,0x00,0x00,0x10,0x00,0x00,0x10,0x00,0x10,0x00,0x10,0x11,0x01,0x00,
-0x00,0x11,0x01,0x00,0x00,0x10,0x00,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,
-0x00,0x10,0x00,0x10,0x00,0x00,0x10,0x00,0x00,0x00,0x01,0x01,0x00,0x00,0x10,
-0x00,0x00,0x00,0x01,0x00,0x00,0x00,0x10,0x00,0x00,0x00,0x00,0x01,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x10,0x10,0x10,0x00,
-0x10,0x00,0x10,0x00,0x00,0x00,0x10,0x00,0x00,0x10,0x00,0x00,0x10,0x00,0x10,
-0x00,0x10,0x00,0x10,0x00,0x10,0x10,0x10,0x00,0x00,0x01,0x01,0x00,0x00,0x10,
-0x00,0x00,0x00,0x01,0x00,0x00,0x00,0x01,0x00,0x00,0x00,0x00,0x01,0x00,0x00,
-0x00,0x01,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x10,0x00,0x01,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x10,0x00,0x00,0x10,0x00,
-0x00,0x10,0x00,0x10,0x00,0x00,0x01,0x01,0x00,0x10,0x01,0x11,0x00,0x10,0x00,
-0x10,0x00,0x00,0x10,0x00,0x00,0x10,0x00,0x00,0x00,0x00,0x01,0x00,0x00,0x00,
-0x00,0x10,0x00,0x00,0x00,0x01,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x11,0x10,0x00,0x10,0x00,0x10,0x00,0x00,0x11,0x01,
-0x00,0x00,0x10,0x00,0x00,0x00,0x11,0x01,0x00,0x00,0x10,0x00,0x00,0x10,0x00,
-0x10,0x00,0x10,0x00,0x10,0x00,0x00,0x10,0x00,0x00,0x10,0x11,0x11,0x00,0x00,
-0x11,0x01,0x00,0x00,0x00,0x00,0x01,0x00,0x11,0x01,0x00,0x00,0x00,0x00,0x00,
-0x11,0x11,0x11,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00
-};
diff --git a/libpsn00b/psxgpu/drawotag.s b/libpsn00b/psxgpu/drawotag.s
deleted file mode 100644
index 3cb0db0..0000000
--- a/libpsn00b/psxgpu/drawotag.s
+++ /dev/null
@@ -1,38 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.section .text
-
-
-.global DrawOTag
-.type DrawOTag, @function
-DrawOTag:
-	addiu	$sp, -4
-	sw		$ra, 0($sp)
-
-	lui		$a3, IOBASE			# I/O segment base
-
-	lui		$v0, 0x0400			# Set DMA direction to CPUtoGPU
-	ori		$v0, 0x2
-	sw		$v0, GPU_GP1($a3)
-	
-.Lgpu_wait:						# Wait for GPU to be ready for commands & DMA
-	jal		ReadGPUstat
-	nop
-	srl		$v0, 26
-	andi	$v0, 1
-	beqz	$v0, .Lgpu_wait
-	nop
-
-	sw		$a0, DMA2_MADR($a3)	# Set DMA base address to specified OT
-	sw		$0 , DMA2_BCR($a3)
-
-	lui		$v0, 0x0100			# Begin OT transfer!
-	ori		$v0, 0x0401
-	sw		$v0, DMA2_CHCR($a3)
-
-	lw		$ra, 0($sp)
-	addiu	$sp, 4
-	jr		$ra
-	nop
diff --git a/libpsn00b/psxgpu/drawprim.s b/libpsn00b/psxgpu/drawprim.s
deleted file mode 100644
index d62c202..0000000
--- a/libpsn00b/psxgpu/drawprim.s
+++ /dev/null
@@ -1,41 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.text
-
-.global DrawPrim
-.type DrawPrim, @function
-DrawPrim:
-
-	addiu	$sp, -8
-	sw		$ra, 0($sp)
-	sw		$s0, 4($sp)
-	
-	move	$s0, $a0			# Wait for GPU to complete
-	jal		DrawSync
-	move	$a0, $0
-	
-	lui		$a3, IOBASE
-	lui		$v0, 0x0400			# Set transfer direction to off
-	sw		$v0, GPU_GP1($a3)
-	
-	move	$a0, $s0
-	lbu		$a1, 3($a0)			# Get length of primitive packet
-	addiu	$a0, 4
-	addiu	$a1, -1
-	
-.Ltransfer_loop:
-	lw		$v0, 0($a0)
-	addiu	$a0, 4
-	sw		$v0, GPU_GP0($a3)
-	bgtz	$a1, .Ltransfer_loop
-	addiu	$a1, -1
-
-	jal		DrawSync
-	move	$a0, $0
-	
-	lw		$ra, 0($sp)
-	lw		$s0, 4($sp)
-	jr		$ra
-	addiu	$sp, 8
-\ No newline at end of file
diff --git a/libpsn00b/psxgpu/drawsync.s b/libpsn00b/psxgpu/drawsync.s
deleted file mode 100644
index b671b03..0000000
--- a/libpsn00b/psxgpu/drawsync.s
+++ /dev/null
@@ -1,67 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.section .text
-
-
-.global DrawSync
-.type DrawSync, @function
-DrawSync:
-
-	bnez	$a0, .Lgetwords
-	lui		$a0, IOBASE
-
-	addiu	$sp, -4
-	sw		$ra, 0($sp)
-
-	jal		ReadGPUstat			# Check if DMA enabled
-	nop
-	srl		$v0, 29
-	andi	$v0, 0x3
-	
-	beqz	$v0, .Lsimple_wait
-	nop
-	
-.Ldma_wait:
-	lw		$v0, DMA2_CHCR($a0)
-	nop
-	srl		$v0, 24
-	andi	$v0, 0x1
-	bnez	$v0, .Ldma_wait
-	nop
-
-.Lgpu_wait:
-	jal		ReadGPUstat
-	nop
-	srl		$v0, 26
-	andi	$v0, 0x5
-	bne		$v0, 5, .Lgpu_wait
-	nop
-	
-	b		.Lexit
-	nop
-	
-.Lsimple_wait:					# Wait for GPU to be ready for next DMA
-	jal		ReadGPUstat
-	nop
-	srl		$v0, 28
-	andi	$v0, 0x1
-	beqz	$v0, .Lsimple_wait
-	nop
-
-.Lexit:
-
-	lw		$ra, 0($sp)
-	addiu	$sp, 4
-	jr		$ra
-	nop
-	
-.Lgetwords:
-
-	lw		$v0, DMA2_BCR($a0)
-	nop
-
-	jr		$ra
-	srl		$v0, 16
-	
diff --git a/libpsn00b/psxgpu/drawsynccallback.s b/libpsn00b/psxgpu/drawsynccallback.s
deleted file mode 100644
index 22cfb7d..0000000
--- a/libpsn00b/psxgpu/drawsynccallback.s
+++ /dev/null
@@ -1,105 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.text
-
-.global DrawSyncCallback
-.type DrawSyncCallback, @function
-DrawSyncCallback:
-
-	addiu	$sp, -8
-	sw		$ra, 0($sp)
-	sw		$a0, 4($sp)
-	
-	jal		EnterCriticalSection
-	nop
-	
-	beqz	$a0, .Luninstall
-	nop
-	
-	la		$a1, _drawsync_handler
-	lw		$a1, 4($sp)
-	jal		DMACallback
-	li		$a0, 2
-	
-	b		.Lcontinue
-	nop
-	
-.Luninstall:
-
-	move	$a1, $0
-	jal		DMACallback
-	li		$a0, 2
-
-.Lcontinue:
-
-	lw		$a0, 4($sp)
-	la		$v1, _drawsync_func
-	lw		$v0, 0($v1)
-	sw		$a0, 0($v1)
-	sw		$v0, 4($sp)
-	
-.Lexit:
-
-	jal		ExitCriticalSection
-	nop
-	
-	lw		$ra, 0($sp)
-	lw		$v0, 4($sp)
-	jr		$ra
-	addiu	$sp, 8
-	
-	
-.type _drawsync_handler, @function
-_drawsync_handler:
-
-.Ldma_wait:
-	la		$v0, _drawsync_func
-	lw		$v0, 0($v0)
-	nop
-	beqz	$v0, .Lskip
-	nop
-
-	addiu	$sp, -4
-	sw		$ra, 0($sp)
-
-	lw		$v0, DMA2_CHCR($a0)
-	nop
-	srl		$v0, 24
-	andi	$v0, 0x1
-	
-	bnez	$v0, .Ldma_wait
-	nop
-
-.Lgpu_wait:
-	jal		ReadGPUstat
-	nop
-	srl		$v0, 28
-	andi	$v0, 0x1
-	beqz	$v0, .Lgpu_wait
-	nop
-	
-	la		$v1, _drawsync_func
-	lw		$v1, 0($v1)
-	
-	lui		$v0, 0x0400			# Set DMA direction to off
-	sw		$v0, GPU_GP1($a0)
-	
-	jalr	$v1
-	nop
-	
-	lw		$ra, 0($sp)
-	addiu	$sp, 4
-	
-.Lskip:
-
-	jr		$ra
-	nop
-	
-
-.data
-
-_drawsync_func:
-	.word 0
-	
-\ No newline at end of file
diff --git a/libpsn00b/psxgpu/env.c b/libpsn00b/psxgpu/env.c
new file mode 100644
index 0000000..5642ad4
--- /dev/null
+++ b/libpsn00b/psxgpu/env.c
@@ -0,0 +1,212 @@
+/*
+ * PSn00bSDK GPU library (DRAWENV/DISPENV functions)
+ * (C) 2022 spicyjpeg - MPL licensed
+ */
+
+#include <stdint.h>
+#include <psxgpu.h>
+#include <hwregs_c.h>
+
+#define _min(x, y) (((x) < (y)) ? (x) : (y))
+
+extern GPU_VideoMode _gpu_video_mode;
+
+/* Drawing API */
+
+DRAWENV *SetDefDrawEnv(DRAWENV *env, int x, int y, int w, int h) {
+	env->clip.x = x;
+	env->clip.y = y;
+	env->clip.w = w;
+	env->clip.h = h;
+
+	env->ofs[0] = 0;
+	env->ofs[1] = 0;
+
+	env->tw.x = 0;
+	env->tw.y = 0;
+	env->tw.w = 0;
+	env->tw.h = 0;
+
+	env->tpage = 0x0a;
+	env->dtd   = 1;
+	env->dfe   = 0;
+	env->isbg  = 0;
+	setRGB0(env, 0, 0, 0);
+
+	env->dr_env.tag = 0;
+	return env;
+}
+
+void DrawOTagEnv(const uint32_t *ot, DRAWENV *env) {
+	DR_ENV *prim = &(env->dr_env);
+
+	// All commands are grouped into a single display list packet for
+	// performance reasons (keep in mind that the GPU doesn't care about this
+	// as the display list is parsed by the DMA unit in the CPU and only the
+	// payload is sent to the GPU).
+	setaddr(prim, ot);
+	setlen(prim, 4);
+
+	// Set drawing area top left
+	prim->code[0]  = 0xe3000000;
+	prim->code[0] |=  env->clip.x & 0x3ff;
+	prim->code[0] |= (env->clip.y & 0x3ff) << 10;
+
+	// Set drawing area bottom right
+	prim->code[1]  = 0xe4000000;
+	prim->code[1] |=  (env->clip.x + (env->clip.w - 1)) & 0x3ff;
+	prim->code[1] |= ((env->clip.y + (env->clip.h - 1)) & 0x3ff) << 10;
+
+	// Set drawing offset
+	prim->code[2]  = 0xe5000000;
+	prim->code[2] |=  (env->clip.x + env->ofs[0]) & 0x7ff;
+	prim->code[2] |= ((env->clip.y + env->ofs[1]) & 0x7ff) << 11;
+
+	// Texture page (reset active page and set dither/mask bits)
+	prim->code[3]  = 0xe1000000;
+	prim->code[3] |= (env->dtd & 1) << 9;
+	prim->code[3] |= (env->dfe & 1) << 10;
+
+	if (env->isbg) {
+		setlen(prim, 7);
+
+		// Rectangle fill
+		// FIXME: reportedly this command doesn't accept height values >511...
+		prim->code[4]  = 0x02000000;
+		//prim->code[4] |= env->r0 | (env->g0 << 8) | (env->b0 << 16);
+		prim->code[4] |= *((const uint32_t *) &(env->isbg)) >> 8;
+		//prim->code[5]  = env->clip.x;
+		//prim->code[5] |= env->clip.y << 16;
+		prim->code[5]  = *((const uint32_t *) &(env->clip.x));
+		prim->code[6]  = env->clip.w;
+		prim->code[6] |= _min(env->clip.h, 0x1ff) << 16;
+	}
+
+	//while (!(GPU_GP1 & (1 << 26)))
+		//__asm__ volatile("");
+
+	DrawOTag((const uint32_t *) prim);
+}
+
+void PutDrawEnv(DRAWENV *env) {
+	DrawOTagEnv((const uint32_t *) 0x00ffffff, env);
+}
+
+// This function skips rebuilding the cached packet whenever possible and is
+// useful if the DRAWENV structure is never modified (which is the case most of
+// the time).
+void PutDrawEnvFast(DRAWENV *env) {
+	if (!(env->dr_env.tag)) {
+		DrawOTagEnv((const uint32_t *) 0x00ffffff, env);
+		return;
+	}
+
+	DrawOTag((const uint32_t *) &(env->dr_env));
+}
+
+/* Display API */
+
+DISPENV *SetDefDispEnv(DISPENV *env, int x, int y, int w, int h) {
+	env->disp.x = x;
+	env->disp.y = y;
+	env->disp.w = w;
+	env->disp.h = h;
+
+	env->screen.x = 0;
+	env->screen.y = 0;
+	env->screen.w = 0;
+	env->screen.h = 0;
+
+	env->isinter = 0;
+	env->isrgb24 = 0;
+	env->reverse = 0;
+
+	return env;
+}
+
+void PutDispEnv(const DISPENV *env) {
+	uint32_t h_range, v_range, mode, fb_pos;
+
+	mode  = _gpu_video_mode << 3;
+	mode |= (env->isrgb24 & 1) << 4;
+	mode |= (env->isinter & 1) << 5;
+	mode |= (env->reverse & 1) << 7;
+
+	if (env->disp.h >= 256)
+		mode |= 1 << 2;
+
+	// Calculate the horizontal display range values. The original code was
+	// this bad; in actual fact it was even worse due to being written in
+	// assembly and using slow multiplication even when not necessary.
+	int offset, span, default_span = 2560;
+
+	if (env->disp.w > 560) {
+		// 640 pixels
+		mode  |= 3;
+		offset = 620;
+		span   = env->screen.w * 4;
+	} else if (env->disp.w > 400) {
+		// 512 pixels
+		mode  |= 2;
+		offset = 615;
+		span   = env->screen.w * 4 + env->screen.w;
+	} else if (env->disp.w > 352) {
+		// 384 pixels (this mode is weird)
+		mode  |= 1 << 6;
+		offset = 539;
+		span   = env->screen.w * 8 - env->screen.w;
+		default_span = 2688;
+	} else if (env->disp.w > 280) {
+		// 320 pixels
+		mode  |= 1;
+		offset = 600;
+		span   = env->screen.w * 8;
+	} else {
+		// 256 pixels
+		offset = 590;
+		span   = env->screen.w * 8 + env->screen.w * 2;
+	}
+
+	offset += env->screen.x * 4;
+	if (!span)
+		span = default_span;
+
+	h_range  = offset & 0xfff;
+	h_range |= ((offset + span) & 0xfff) << 12;
+
+	// Calculate the vertical display range values.
+	offset = 16 + env->screen.y;
+	span   = env->screen.h ? env->screen.h : 240;
+
+	v_range  = offset & 0x3ff;
+	v_range |= ((offset + span) & 0x3ff) << 10;
+
+	fb_pos  =  env->disp.x & 0x3ff;
+	fb_pos |= (env->disp.y & 0x1ff) << 10;
+
+	GPU_GP1 = 0x06000000 | h_range; // Set horizontal display range
+	GPU_GP1 = 0x07000000 | v_range; // Set vertical display range
+	GPU_GP1 = 0x08000000 | mode;    // Set video mode
+	GPU_GP1 = 0x05000000 | fb_pos;  // Set VRAM location to display
+}
+
+/* Deprecated "raw" display API */
+
+void PutDispEnvRaw(const DISPENV_RAW *env) {
+	uint32_t h_range, v_range, fb_pos;
+
+	h_range  =   608 + env->vid_xpos;
+	h_range |= (3168 + env->vid_xpos) << 12;
+
+	// FIXME: these hardcoded values are for NTSC displays.
+	v_range  =  (136 - 120 + env->vid_ypos) & 0x3ff;
+	v_range |= ((136 + 120 + env->vid_ypos) & 0x3ff) << 12;
+
+	fb_pos  =  env->fb_x & 0x3ff;
+	fb_pos |= (env->fb_y & 0x1ff) << 10;
+
+	GPU_GP1 = 0x06000000 | h_range;       // Set horizontal display range
+	GPU_GP1 = 0x07000000 | v_range;       // Set vertical display range
+	GPU_GP1 = 0x08000000 | env->vid_mode; // Set video mode
+	GPU_GP1 = 0x05000000 | fb_pos;        // Set VRAM location to display
+}
diff --git a/libpsn00b/psxgpu/fntsort.c b/libpsn00b/psxgpu/fntsort.c
deleted file mode 100644
index 9358793..0000000
--- a/libpsn00b/psxgpu/fntsort.c
+++ /dev/null
@@ -1,48 +0,0 @@
-#include <sys/types.h>
-#include <stdio.h>
-#include <ctype.h>
-#include <psxgpu.h>
-
-extern unsigned short _font_tpage;
-extern unsigned short _font_clut;
-
-char *FntSort(u_long *ot, char *pri, int x, int y, const char *text) {
-	
-	DR_TPAGE *tpage;
-	SPRT_8 *sprt = (SPRT_8*)pri;
-	int	i;
-	
-	while( *text != 0 ) {
-	
-		i = toupper( *text )-32;
-	
-		if( i > 0 ) {
-			
-			i--;
-			setSprt8( sprt );
-			setRGB0( sprt, 128, 128, 128 );
-			setXY0( sprt, x, y );
-			setUV0( sprt, (i%16)<<3, (i>>4)<<3 );
-			sprt->clut = _font_clut;
-			addPrim( ot, sprt );
-			sprt++;
-			
-		}
-	
-		x += 8;
-		text++;
-		
-	}
-	
-	pri = (char*)sprt;
-	
-	tpage = (DR_TPAGE*)pri;
-	tpage->code[0] = _font_tpage;
-	setlen( tpage, 1 );
-	setcode( tpage, 0xe1 );
-	addPrim( ot, pri );
-	pri += sizeof(DR_TPAGE);
-	
-	return pri;
-	
-}
-\ No newline at end of file
diff --git a/libpsn00b/psxgpu/font.c b/libpsn00b/psxgpu/font.c
index 4c715a9..2d4105f 100644
--- a/libpsn00b/psxgpu/font.c
+++ b/libpsn00b/psxgpu/font.c
@@ -1,4 +1,4 @@
-#include <sys/types.h>
+#include <stdint.h>
 #include <stdio.h>
 #include <string.h>
 #include <stdlib.h>
@@ -6,29 +6,28 @@
 #include <psxgpu.h>
 
 typedef struct _fnt_stream {
-	char *txtbuff;
-	char *txtnext;
-	char *pribuff;
-	short x,y;
-	short w,h;
-	int bg;
-	int maxchars;
+	char	*txtbuff;
+	char	*txtnext;
+	char	*pribuff;
+	int16_t	x, y;
+	int16_t	w, h;
+	int		bg, maxchars;
 } _fnt_stream;
 
 static _fnt_stream _stream[8];
 static int _nstreams = 0;
 
-u_short _font_tpage;
-u_short _font_clut;
+uint16_t _font_tpage;
+uint16_t _font_clut;
 
-extern u_char dbugfont[];
+extern uint8_t _gpu_debug_font[];
 
 void FntLoad(int x, int y) {
 
 	RECT pos;
 	TIM_IMAGE tim;
 	
-	GetTimInfo( (u_long*)dbugfont, &tim );
+	GetTimInfo( (const uint32_t *) _gpu_debug_font, &tim );
 	
 	// Load font image
 	pos = *tim.prect;
@@ -215,7 +214,7 @@ char *FntFlush(int id) {
 	
 	// Draw the primitives
 	DrawSync(0);
-	DrawOTag((u_long*)_stream[id].pribuff);
+	DrawOTag((uint32_t*)_stream[id].pribuff);
 	DrawSync(0);
 	
 	_stream[id].txtnext = _stream[id].txtbuff;
@@ -223,4 +222,45 @@ char *FntFlush(int id) {
 	
 	return _stream[id].pribuff;
 	
-}
-\ No newline at end of file
+}
+
+char *FntSort(uint32_t *ot, char *pri, int x, int y, const char *text) {
+	
+	DR_TPAGE *tpage;
+	SPRT_8 *sprt = (SPRT_8*)pri;
+	int	i;
+	
+	while( *text != 0 ) {
+	
+		i = toupper( *text )-32;
+	
+		if( i > 0 ) {
+			
+			i--;
+			setSprt8( sprt );
+			setRGB0( sprt, 128, 128, 128 );
+			setXY0( sprt, x, y );
+			setUV0( sprt, (i%16)<<3, (i>>4)<<3 );
+			sprt->clut = _font_clut;
+			addPrim( ot, sprt );
+			sprt++;
+			
+		}
+	
+		x += 8;
+		text++;
+		
+	}
+	
+	pri = (char*)sprt;
+	
+	tpage = (DR_TPAGE*)pri;
+	tpage->code[0] = _font_tpage;
+	setlen( tpage, 1 );
+	setcode( tpage, 0xe1 );
+	addPrim( ot, pri );
+	pri += sizeof(DR_TPAGE);
+	
+	return pri;
+	
+}
diff --git a/libpsn00b/psxgpu/getode.s b/libpsn00b/psxgpu/getode.s
deleted file mode 100644
index 5dc1e70..0000000
--- a/libpsn00b/psxgpu/getode.s
+++ /dev/null
@@ -1,20 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.section .text
-
-
-.global GetODE
-.type GetODE, @function
-GetODE:
-	addiu	$sp, -4
-	sw		$ra, 0($sp)
-	jal		ReadGPUstat
-	nop
-	srl		$v0, 31
-	andi	$v0, 1
-	lw		$ra, 0($sp)
-	addiu	$sp, 4
-	jr		$ra
-	nop
diff --git a/libpsn00b/psxgpu/gettimimage.c b/libpsn00b/psxgpu/gettimimage.c
deleted file mode 100644
index 5598e07..0000000
--- a/libpsn00b/psxgpu/gettimimage.c
+++ /dev/null
@@ -1,40 +0,0 @@
-#include <sys/types.h>
-#include <psxgpu.h>
-
-int GetTimInfo(const u_long *tim, TIM_IMAGE *timimg) {
-
-	u_long *rtim;
-
-	// Check ID
-	if( ( tim[0]&0xff ) != 0x10 ) {
-		return 1;
-	}
-
-	// Check version
-	if( ( (tim[0]>>8)&0xff ) != 0x0 ) {
-		return 2;
-	}
-
-	timimg->mode = tim[1];
-	rtim = tim+2;
-
-	// Clut present?
-	if( timimg->mode & 0x8 ) {
-
-		timimg->crect = (RECT*)(rtim+1);
-		timimg->caddr = (u_long*)(rtim+3);
-
-		rtim += rtim[0]>>2;
-
-	} else {
-	
-		timimg->caddr = 0;
-		
-	}
-
-	timimg->prect = (RECT*)(rtim+1);
-	timimg->paddr = (u_long*)(rtim+3);
-
-	return 0;
-
-}
diff --git a/libpsn00b/psxgpu/getvideomode.s b/libpsn00b/psxgpu/getvideomode.s
deleted file mode 100644
index 6f1613c..0000000
--- a/libpsn00b/psxgpu/getvideomode.s
+++ /dev/null
@@ -1,14 +0,0 @@
-.set noreorder
-
-
-.section .text
-
-.global GetVideoMode
-.type GetVideoMode, @function
-GetVideoMode:
-
-	la		$v0, _gpu_standard
-	lw		$v0, 0($v0)
-	
-	jr		$ra
-	nop
diff --git a/libpsn00b/psxgpu/image.c b/libpsn00b/psxgpu/image.c
new file mode 100644
index 0000000..da51e7d
--- /dev/null
+++ b/libpsn00b/psxgpu/image.c
@@ -0,0 +1,115 @@
+/*
+ * PSn00bSDK GPU library (image and VRAM transfer functions)
+ * (C) 2022 spicyjpeg - MPL licensed
+ */
+
+#include <stdint.h>
+#include <stdio.h>
+#include <psxgpu.h>
+#include <hwregs_c.h>
+
+#define DMA_CHUNK_LENGTH 8
+
+/* VRAM transfer API */
+
+static void _load_store_image(
+	uint32_t	command,
+	int			mode,
+	const RECT	*rect,
+	uint32_t	*data
+) {
+	size_t length = rect->w * rect->h;
+	if (length % 2)
+		printf("psxgpu: can't transfer an odd number of pixels\n");
+
+	length /= 2;
+	if ((length >= DMA_CHUNK_LENGTH) && (length % DMA_CHUNK_LENGTH)) {
+		printf("psxgpu: transfer data length (%d) is not a multiple of %d, rounding\n", length, DMA_CHUNK_LENGTH);
+		length += DMA_CHUNK_LENGTH - 1;
+	}
+
+	DrawSync(0);
+	GPU_GP1 = 0x04000000; // Disable DMA request
+	GPU_GP0 = 0x01000000; // Flush cache
+
+	GPU_GP0 = command;
+	//GPU_GP0 = rect->x | (rect->y << 16);
+	GPU_GP0 = *((const uint32_t *) &(rect->x));
+	//GPU_GP0 = rect->w | (rect->h << 16);
+	GPU_GP0 = *((const uint32_t *) &(rect->w));
+
+	// Enable DMA request, route to GP0 (2) or from GPU_READ (3)
+	GPU_GP1 = 0x04000000 | mode;
+
+	DMA_MADR(2) = (uint32_t) data;
+	if (length < DMA_CHUNK_LENGTH)
+		DMA_BCR(2) = 0x00010000 | length;
+	else
+		DMA_BCR(2) = DMA_CHUNK_LENGTH | ((length / DMA_CHUNK_LENGTH) << 16);
+
+	DMA_CHCR(2) = 0x01000200 | ((mode & 1) ^ 1);
+}
+
+void LoadImage(const RECT *rect, const uint32_t *data) {
+	_load_store_image(0xa0000000, 2, rect, (uint32_t *) data);
+}
+
+void StoreImage(const RECT *rect, uint32_t *data) {
+	_load_store_image(0xc0000000, 3, rect, data);
+}
+
+/* .TIM image parsers */
+
+// This is the only libgs function PSn00bSDK is ever going to implement. The
+// difference from GetTimInfo() is that it copies RECTs rather than merely
+// returning pointers to them, which become useless once the .TIM file is
+// unloaded from main RAM.
+int GsGetTimInfo(const uint32_t *tim, GsIMAGE *info) {
+	if ((*(tim++) & 0xffff) != 0x0010)
+		return 1;
+
+	info->pmode = *(tim++);
+	if (info->pmode & 8) {
+		const uint32_t *palette_end = tim;
+		palette_end += *(tim++) / 4;
+
+		*((uint32_t *) &(info->cx)) = *(tim++);
+		*((uint32_t *) &(info->cw)) = *(tim++);
+		info->clut = (uint32_t *) tim;
+
+		tim = palette_end;
+	} else {
+		info->clut = 0;
+	}
+
+	tim++;
+	*((uint32_t *) &(info->px)) = *(tim++);
+	*((uint32_t *) &(info->pw)) = *(tim++);
+	info->pixel = (uint32_t *) tim;
+
+	return 0;
+}
+
+int GetTimInfo(const uint32_t *tim, TIM_IMAGE *info) {
+	if ((*(tim++) & 0xffff) != 0x0010)
+		return 1;
+
+	info->mode = *(tim++);
+	if (info->mode & 8) {
+		const uint32_t *palette_end = tim;
+		palette_end += *(tim++) / 4;
+
+		info->crect = (RECT *)     tim;
+		info->caddr = (uint32_t *) &tim[2];
+
+		tim = palette_end;
+	} else {
+		info->caddr = 0;
+	}
+
+	tim++;
+	info->prect = (RECT *)     tim;
+	info->paddr = (uint32_t *) &tim[2];
+
+	return 0;
+}
diff --git a/libpsn00b/psxgpu/loadimage.s b/libpsn00b/psxgpu/loadimage.s
deleted file mode 100644
index 45f152f..0000000
--- a/libpsn00b/psxgpu/loadimage.s
+++ /dev/null
@@ -1,70 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.set RECT_x,	0
-.set RECT_y,	2
-.set RECT_w,	4
-.set RECT_h,	6
-
-.section .text
-
-
-.global LoadImage
-.type LoadImage, @function
-LoadImage:
-	addiu	$sp, -8
-	sw		$ra, 0($sp)
-	sw		$s0, 4($sp)
-
-	lui		$s0, IOBASE			# Set I/O segment base address
-
-.Lgpu_wait:						# Wait for GPU to be ready for commands and DMA
-	jal		ReadGPUstat
-	nop
-	srl		$v0, 0x1a
-	andi	$v0, 0x5
-	li		$v1, 5
-	#srl		$v0, 28
-	#andi	$v0, 1
-	bne		$v0, $v1, .Lgpu_wait
-	nop
-
-	lui		$v0, 0x400			# Set DMA direction to off
-	sw		$v0, GPU_GP1($s0)
-
-	lui		$v0, 0x0100			# Clear GPU cache
-	sw		$v0, GPU_GP0($s0)
-
-	lui		$v1, 0xa000			# Load image to VRAM
-	sw		$v1, GPU_GP0($s0)
-	lw		$v0, RECT_x($a0)	# Set XY and dimensions of image
-	lw		$v1, RECT_w($a0)
-	sw		$v0, GPU_GP0($s0)
-	sw		$v1, GPU_GP0($s0)
-
-	lui		$v0, 0x400			# Set DMA direction to CPUtoVRAM
-	ori		$v0, 0x2
-	sw		$v0, GPU_GP1($s0)
-
-	lhu		$v0, RECT_w($a0)	# Get rectangle size
-	lhu		$v1, RECT_h($a0)
-	nop
-	mult	$v0, $v1			# Calculate BCR value
-	mflo	$v1
-	srl		$v1, 0x4
-	sll		$v1, 0x10
-	ori		$v1, 0x8
-
-	sw		$a1, DMA2_MADR($s0)	# Set DMA base address and transfer length
-	sw		$v1, DMA2_BCR($s0)
-
-	lui		$v0, 0x100			# Start DMA transfer
-	ori		$v0, 0x201
-	sw		$v0, DMA2_CHCR($s0)
-
-	lw		$ra, 0($sp)
-	lw		$s0, 4($sp)
-	jr		$ra
-	addiu	$sp, 8
-
diff --git a/libpsn00b/psxgpu/putdispenv.s b/libpsn00b/psxgpu/putdispenv.s
deleted file mode 100644
index fc09454..0000000
--- a/libpsn00b/psxgpu/putdispenv.s
+++ /dev/null
@@ -1,174 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.set DISP_dx,		0
-.set DISP_dy,		2
-.set DISP_dw,		4
-.set DISP_dh,		6
-.set DISP_sx,		8
-.set DISP_sy,		10
-.set DISP_sw,		12
-.set DISP_sh,		14
-.set DISP_inter,	16
-.set DISP_isrgb24,	17
-.set DISP_reverse,	18
-
-.section .text
-
-
-.global PutDispEnv
-.type PutDispEnv, @function
-PutDispEnv:
-
-	lui		$a3, IOBASE
-
-	# Horizontal resolution stuff
-
-	lh		$a2, DISP_dw($a0)			# Get X resolution
-
-	lh		$v0, DISP_sx($a0)
-	lh		$v1, DISP_sw($a0)			# Get X screen width
-
-	move	$a1, $0						# To use as mode value
-
-	bgt		$a2, 560, .Lmode_640
-	nop
-	bgt		$a2, 400, .Lmode_512
-	nop
-	bgt		$a2, 352, .Lmode_384
-	nop
-	bgt		$a2, 280, .Lmode_320
-	nop
-
-	.set noat
-
-.Lmode_256:
-	li		$at, 10
-	mult	$at, $v1
-	li		$a2, 0x24e
-	sll		$v0, 2
-	add		$a2, $v0
-	b		.Lmode_end
-	li		$v1, 0xa00
-.Lmode_320:
-	li		$at, 8
-	mult	$at, $v1
-	li		$a2, 0x258
-	ori		$a1, 0x01
-	sll		$v0, 2
-	add		$a2, $v0
-	b		.Lmode_end
-	li		$v1, 0xa00
-.Lmode_384:
-	li		$at, 7
-	mult	$at, $v1
-	li		$a2, 0x21b
-	ori		$a1, 0x64
-	sll		$v0, 2
-	add		$a2, $v0
-	b		.Lmode_end
-	li		$v1, 0xa80
-.Lmode_512:
-	li		$at, 5
-	mult	$at, $v1
-	li		$a2, 0x267
-	ori		$a1, 0x02
-	sll		$v0, 2
-	add		$a2, $v0
-	b		.Lmode_end
-	li		$v1, 0xa00
-.Lmode_640:
-	li		$at, 4
-	mult	$at, $v1
-	li		$a2, 0x26c
-	ori		$a1, 0x03
-	sll		$v0, 2
-	add		$a2, $v0
-	li		$v1, 0xa00
-.Lmode_end:
-
-	.set at
-
-	mflo	$v0
-	bnez	$v0, .Lno_default			# Check if screen with is non zero
-	nop
-	move	$v0, $v1					# Use default if screen width is 0
-.Lno_default:
-
-	addu	$v0, $a2					# Apply horizontal display coordinates
-	sll		$v0, 12
-	andi	$a2, 0xfff
-	or		$a2, $v0
-	lui		$v0, 0x0600
-	or		$v0, $a2
-	sw		$v0, GPU_GP1($a3)
-
-	# Vertical resolution
-
-	lh		$v0, DISP_dh($a0)
-	li		$a2, 0x10
-	ble		$v0, 256, .Lmode_low
-	nop
-
-.Lmode_high:
-	ori		$a1, 0x04
-.Lmode_low:
-	lh		$v0, DISP_sy($a0)
-	lh		$v1, DISP_sh($a0)
-	add		$a2, $v0
-	bnez	$v1, .Lno_default_vert
-	nop
-	li		$v1, 0xf0
-.Lno_default_vert:
-	add		$v1, $a2
-	and		$a2, 0x3ff
-	sll		$v1, 10
-	or		$v1, $a2
-	lui		$v0, 0x0700
-	or		$v1, $v0
-	sw		$v1, GPU_GP1($a3)
-
-	# Video mode
-
-	la		$v0, _gpu_standard
-	lbu		$v0, 0($v0)
-	nop
-	beqz	$v0, .Lconfig_ntsc
-	nop
-.Lconfig_pal:
-	ori		$a1, 0x08
-.Lconfig_ntsc:
-
-	lbu		$v0, DISP_inter($a0)
-	lbu		$v1, DISP_isrgb24($a0)
-	beqz	$v0, .Lno_inter
-	nop
-	or		$a1, 0x20
-.Lno_inter:
-	beqz	$v1, .Lno_rgb24
-	nop
-	or		$a1, 0x10
-.Lno_rgb24:
-	lbu		$v0, DISP_inter($a0)
-	nop
-	beqz	$v0, .Lno_reverse
-	nop
-	or		$a1, 0x80
-.Lno_reverse:
-
-	lui		$v0, 0x800					# Apply mode
-	or		$a1, $v0
-	sw		$a1, GPU_GP1($a3)
-
-	lhu		$v0, DISP_dx($a0)			# Set VRAM XY offset
-	lhu		$v1, DISP_dy($a0)
-	andi	$v0, 0x3ff
-	andi	$v1, 0x1ff
-	sll		$v1, 10
-	or		$v0, $v1
-	lui		$v1, 0x500
-	or		$v0, $v1
-
-	jr		$ra
-	sw		$v0, GPU_GP1($a3)
diff --git a/libpsn00b/psxgpu/putdispenvraw.s b/libpsn00b/psxgpu/putdispenvraw.s
deleted file mode 100644
index 747796f..0000000
--- a/libpsn00b/psxgpu/putdispenvraw.s
+++ /dev/null
@@ -1,71 +0,0 @@
-.set noreorder
-.set noat
-
-.include "hwregs_a.h"
-
-.set DISP_mode,		0
-.set DISP_vxpos,	4
-.set DISP_vypos,	6
-.set DISP_fbx,		8
-.set DISP_fby,		10
-
-.section .text
-
-
-.global PutDispEnvRaw
-.type PutDispEnvRaw, @function
-PutDispEnvRaw:
-	addiu	$sp, -8
-	sw		$ra, 0($sp)
-	sw		$s0, 4($sp)
-
-	lui		$s0, IOBASE
-
-	lh		$at, DISP_vxpos($a0)	# Set horizontal display range
-	li		$v0, 608
-	add		$v0, $at
-	li		$v1, 3168
-	add		$v1, $at
-	sll		$v1, 12
-	or		$v0, $v1
-	lui		$v1, 0x600
-	or		$v1, $v0
-	sw		$v1, GPU_GP1($s0)
-
-	lh		$at, DISP_vypos($a0)	# Set vertical display range (for NTSC)
-	li		$v1, 120				# (values differet for PAL modes)
-	sub		$v1, $at
-	li		$v0, 136
-	sub		$v0, $v1
-	andi	$v0, 0x1ff
-	li		$v1, 120
-	add		$v1, $at
-	li		$at, 136
-	add		$at, $v1
-	andi	$at, 0x1ff
-	sll		$at, 10
-	or		$v0, $at
-	lui		$at, 0x700
-	or		$v0, $at
-	sw		$v0, GPU_GP1($s0)
-
-	lw		$v0, DISP_mode($a0)		# Set video mode
-	lui		$at, 0x800
-	or		$v0, $at
-	sw		$v0, GPU_GP1($s0)
-
-	lhu		$v0, DISP_fbx($a0)		# Set VRAM XY offset
-	lhu		$v1, DISP_fby($a0)
-	andi	$v0, 0x3ff
-	andi	$v1, 0x1ff
-	sll		$v1, 10
-	or		$v0, $v1
-	lui		$v1, 0x500
-	or		$v0, $v1
-	sw		$v0, GPU_GP1($s0)
-
-	lw		$ra, 0($sp)
-	lw		$s0, 4($sp)
-	jr		$ra
-	addiu	$sp, 8
-
diff --git a/libpsn00b/psxgpu/putdrawenv.s b/libpsn00b/psxgpu/putdrawenv.s
deleted file mode 100644
index c0d5676..0000000
--- a/libpsn00b/psxgpu/putdrawenv.s
+++ /dev/null
@@ -1,142 +0,0 @@
-.set noreorder
-.set noat
-
-.include "hwregs_a.h"
-
-.set DRAW_x,		0		# Drawing area
-.set DRAW_y,		2
-.set DRAW_w,		4
-.set DRAW_h,		6
-.set DRAW_ofx,		8		# Draw offset
-.set DRAW_ofy,		10
-.set DRAW_tx,		12		# Texture window
-.set DRAW_ty,		14
-.set DRAW_tw,		16
-.set DRAW_th,		18
-.set DRAW_tpage,	20		# TPage values
-.set DRAW_dtd,		22
-.set DRAW_dfe,		23
-.set DRAW_isbg,		24		# Clear draw area
-.set DRAW_r0,		25
-.set DRAW_g0,		26
-.set DRAW_b0,		27
-.set DRAW_env,		28
-
-
-.section .text
-
-.global PutDrawEnv
-.type PutDrawEnv, @function
-PutDrawEnv:
-	addiu	$sp, -4
-	sw		$ra, 0($sp)
-
-	addiu	$a1, $a0, DRAW_env
-
-	li		$v0, 0x04ffffff				# Packet header (length+terminator)
-	sw		$v0, 0($a1)
-
-	lhu		$v0, DRAW_x($a0)			# Set draw area top-left
-	lhu		$v1, DRAW_y($a0)
-	andi	$v0, 0x3ff
-	andi	$v1, 0x1ff
-	sll		$v1, 10
-	or		$v0, $v1
-	lui		$v1, 0xe300
-	or		$v0, $v1
-	sw		$v0, 4($a1)					# 1
-
-	.set noat
-
-	lhu		$v0, DRAW_w($a0)			# Set draw area bottom-right
-	lhu		$at, DRAW_x($a0)
-	addiu	$v0, -1
-	addu	$at, $v0
-	andi	$at, 0x3ff
-	lhu		$v1, DRAW_h($a0)
-	lhu		$v0, DRAW_y($a0)
-	addiu	$v1, -1
-	addu	$v0, $v1
-	andi	$v0, 0x1ff
-	sll		$v0, 10
-	or		$at, $v0
-	lui		$v0, 0xe400
-	or		$at, $v0
-	sw		$at, 8($a1)					# 2
-
-	lhu		$v0, DRAW_x($a0)			# Set drawing offset
-	lhu		$v1, DRAW_ofx($a0)
-	nop
-	add		$v0, $v1
-	andi	$at, $v0, 0x7ff
-	lhu		$v0, DRAW_y($a0)
-	lhu		$v1, DRAW_ofy($a0)
-	nop
-	add		$v0, $v1
-	andi	$v0, 0x7ff
-	sll		$v0, 11
-	or		$at, $v0
-	lui		$v0, 0xe500
-	or		$at, $v0
-	sw		$at, 12($a1)				# 3
-
-	lhu		$at, DRAW_tpage($a0)		# Set tpage
-	lbu		$v0, DRAW_dtd($a0)
-	lbu		$v1, DRAW_dfe($a0)
-	andi	$v0, 1
-	and		$v1, 1
-	sll		$v0, 9
-	sll		$v1, 10
-	or		$at, $v0
-	or		$at, $v1
-	lui		$v0, 0xe100
-	or		$at, $v0
-	sw		$at, 16($a1)				# 4
-
-	.set at
-
-	lbu		$v0, DRAW_isbg($a0)
-	nop
-	beqz	$v0, .Lno_fillVRAM
-	nop
-
-	lw		$v0, DRAW_isbg($a0)			# FillVRAM
-	lui		$v1, 0x0200
-	srl		$v0, 8
-	or		$v0, $v1
-	sw		$v0, 20($a1)				# 5
-	lw		$v0, DRAW_x($a0)
-	lw		$v1, DRAW_w($a0)
-	sw		$v0, 24($a1)				# 6
-
-	srl		$v0, $v1, 16				# Workaround as rectangle primitives
-	blt		$v0, 511, .Lno_overflow		# don't accept a height of 512
-	nop
-
-	li		$v0, 511
-	sll		$v0, 16
-	andi	$v1, 0xffff
-	or		$v1, $v0
-
-.Lno_overflow:
-	sw		$v1, 28($a1)				# 7
-	li		$v0, 0x07ffffff				# Packet header (length+terminator)
-	sw		$v0, 0($a1)
-
-.Lno_fillVRAM:
-
-.Lgpu_wait:								# Wait for GPU to become ready for commands and DMA
-	jal		ReadGPUstat
-	nop
-	srl		$v0, 26
-	andi	$v0, 1
-	beqz	$v0, .Lgpu_wait
-	nop
-
-	jal		DrawOTag
-	move	$a0, $a1
-
-	lw		$ra, 0($sp)
-	addiu	$sp, 4
-	jr		$ra
-	nop
diff --git a/libpsn00b/psxgpu/readgpustat.s b/libpsn00b/psxgpu/readgpustat.s
deleted file mode 100644
index ffff4d7..0000000
--- a/libpsn00b/psxgpu/readgpustat.s
+++ /dev/null
@@ -1,14 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.section .text
-
-
-.global ReadGPUstat
-.type ReadGPUstat, @function
-ReadGPUstat:
-	lui     $v0, IOBASE
-	lw      $v0, GPU_GP1($v0)
-	jr      $ra
-	nop
diff --git a/libpsn00b/psxgpu/readme.txt b/libpsn00b/psxgpu/readme.txt
index b626e1d..67aa110 100644
--- a/libpsn00b/psxgpu/readme.txt
+++ b/libpsn00b/psxgpu/readme.txt
@@ -3,25 +3,18 @@ PSX GPU library, part of PSn00bSDK
 
 Licensed under Mozilla Public License
 
-	Open source implementation of the GPU library written mostly in MIPS
-assembly. Supports DMA transfers for ordering table draw and transferring
-image data to and from VRAM. The syntax is intentionally made to closely
-resemble Sony's syntax for familiarity and to make porting homebrew made
-using the official SDK to PSn00bSDK a little easier.
-
+Open source implementation of the GPU library written entirely in C. Supports
+DMA transfers for drawing OTs (with an internal queue so DrawOTag() can be
+called even when another OT is being drawn) and transferring image data to and
+from VRAM. The syntax is intentionally made to closely resemble Sony's syntax
+for familiarity and to make porting homebrew made using the official SDK to
+PSn00bSDK a little easier.
 
 Library developer(s):
 
-	Lameguy64
-
+	Lameguy64 (initial implementation in assembly, debug font API)
+	spicyjpeg
 
 Library header(s):
 
-	hwregs_a.h (GNU assembler port defs)
 	psxgpu.h
-
-
-Todo list:
-
-	* ClearOTag() function (non reverse version of ClearOTagR()) yet to be
-	  implemented (but should be trivial).
diff --git a/libpsn00b/psxgpu/resetgraph.s b/libpsn00b/psxgpu/resetgraph.s
deleted file mode 100644
index 6327f02..0000000
--- a/libpsn00b/psxgpu/resetgraph.s
+++ /dev/null
@@ -1,363 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.section .text
-
-.set ISR_STACK_SIZE, 4096
-
-.global ResetGraph						# Resets the GPU and installs a
-.type ResetGraph, @function				# VSync event handler
-ResetGraph:
-	addiu	$sp, -8						# C style stack allocation (required if
-	sw		$ra, 0($sp)					# you call BIOS functions from asm)
-	sw		$a0, 4($sp)
-
-	la		$a0, resetgraph_msg
-	move	$a1, $0
-	move	$a2, $0
-	la		$a1, _irq_func_table
-	la		$a2, _custom_exit
-	jal		printf
-	addiu	$sp, -16
-	addiu	$sp, 16
-	
-	la		$a0, sr_msg
-	mfc0	$a1, $12
-	jal		printf
-	addiu	$sp, -16
-	addiu	$sp, 16
-	
-	la		$v0, _hooks_installed		# Skip installing hooks if this function
-	lbu		$v0, 0($v0)					# has already been called before once
-	nop
-	bnez	$v0, .Lskip_hook_init
-	nop
-	
-	jal		EnterCriticalSection		# Disable interrupts as LoadExec() keeps
-	nop									# interrupts enabled when transferring 
-										# execution to the loaded program
-
-	lui		$a3, IOBASE					# Base address for I/O
-
-	lui		$v0, 0x3b33					# Enables DMA channel 6 (for ClearOTag)
-	ori		$v0, 0x3b33					# Enables DMA channel 2
-	sw		$v0, DMA_DPCR($a3)
-	sw		$0 , DMA_DICR($a3)			# Clear DICR (not needed)
-
-	sw		$0 , IRQ_MASK($a3)			# Clear IRQ settings
-	
-	la		$v0, _hooks_installed		# Set installed flag
-	li		$v1, 0x1
-	sb		$v1, 0($v0)
-
-	la		$v0, _vsync_cb_func			# Clear VSync callback function
-	sw		$0 , 0($v0)
-
-	la		$a1, _vsync_irq_callback	# Install VSync interrupt callback
-	jal		InterruptCallback
-	li		$a0, 0
-
-	jal		RestartCallback
-	nop
-
-	la		$a0, cbhooks_msg
-	jal		printf
-	addiu	$sp, -16
-	addiu	$sp, 16
-	
-	jal		_96_remove					# Remove CD handling left by the BIOS
-	nop
-	
-	la		$a0, abouttoen_msg
-	jal		printf
-	addiu	$sp, -16
-	addiu	$sp, 16
-	
-	jal		ExitCriticalSection			# Re-enable interrupts
-	nop
-	
-	la		$a0, enableint_msg
-	jal		printf
-	addiu	$sp, -16
-	addiu	$sp, 16
-	
-.Lskip_hook_init:
-	
-	lui		$a3, IOBASE
-
-	lw		$v0, GPU_GP1($a3)				# Get video standard
-	lui		$v1, 0x0010
-	and		$v0, $v1
-	la		$v1, _gpu_standard
-	beqz	$v0, .Lnot_pal
-	sw		$0 , 0($v1)
-	li		$v0, 1
-	sw		$v0, 0($v1)
-.Lnot_pal:
-
-	lw		$a0, 4($sp)					# Get argument value
-
-	lui		$a3, IOBASE					# Set base I/O again (likely destroyed
-										# by previous calls)
-
-	li		$v0, 0x1d00					# Configure timer 1 as Hblank counter
-	sw		$v0, TIMER1_CTRL($a3)		# Set timer 1 value
-	
-	beq		$a0, 1, .Lgpu_init_1
-	nop
-	beq		$a0, 3, .Lgpu_init_3
-	nop
-
-	sw		$0 , GPU_GP1($a3)				# Reset the GPU
-
-	b		.Linit_done
-	nop
-
-.Lgpu_init_1:
-
-	sw		$0 , DMA2_CHCR($a3)			# Stop any DMA
-
-.Lgpu_init_3:
-
-	li		$v0, 0x1					# Reset the command buffer
-	sw		$v0, GPU_GP1($a3)
-
-.Linit_done:
-
-	lw		$ra, 0($sp)
-	lw		$a0, 4($sp)					# Return
-	jr		$ra
-	addiu	$sp, 8
-
-
-.global VSync							# VSync function
-.type VSync, @function
-VSync:
-	
-	addiu	$sp, -12
-	sw		$ra, 0($sp)
-	sw		$s0, 4($sp)
-	
-	lui		$a3, IOBASE					# Get GPU status (for interlace sync)
-	lw		$s0, GPU_GP1($a3)
-	
-.Lhwait_loop:							# Get Hblank time
-	lw		$v0, TIMER1_VALUE($a3)
-	nop
-	lw		$v1, TIMER1_VALUE($a3)
-	nop
-	bne		$v0, $v1, .Lhwait_loop
-	nop
-	
-	la		$a3, _vsync_lasthblank		# Calculate Hblank time since last
-	lw		$v1, 0($a3)
-	nop
-	subu	$v0, $v1
-	andi	$v0, 0xffff
-	
-	beq		$a0, 1, .Lhblank_exit		# Return Hblank time only, no VSync
-	sw		$v0, 8($sp)					# Stored as return value
-	
-	bgez	$a0, .Lvsync				# Vsync if argument is 0 and up
-	nop
-	
-	la		$v0, _vsync_rcnt			# Return VSync count only
-	lw		$v0, 0($v0)
-	nop
-	b		.Lvsync_exit
-	sw		$v0, 8($sp)
-	
-.Lvsync:
-
-	bnez	$a0, .Lnot_zero
-	nop
-	li		$a0, 1
-	
-.Lnot_zero:
-	
-	la		$v0, _vsync_rcnt			# Call vsync sub function (with timeout)
-	lw		$v0, 0($v0)
-	addiu	$a1, $a0, 1
-	jal		_vsync_sub
-	addu	$a0, $v0, $a0
-	
-	lui		$v0, 0x40
-	and		$v0, $s0, $v0
-	beqz	$v0, .Lhblank_exit
-	nop
-	
-	lui		$a3, IOBASE					# Interlace wait logic
-	
-	lw		$v0, GPU_GP1($a3)
-	nop
-	xor		$v0, $s0, $v0
-	bltz	$v0, .Lhblank_exit
-	lui		$a0, 0x8000
-	
-.Linterlace_wait:
-	lw		$v0, GPU_GP1($a3)
-	nop
-	xor		$v0, $s0, $v0
-	and		$v0, $a0
-	beqz	$v0, .Linterlace_wait
-	nop
-	
-.Lhblank_exit:							# Set current Hblank as last value
-
-	la		$a2, _vsync_lasthblank
-	
-.Lhwait2_loop:
-	lw		$v0, TIMER1_VALUE($a3)
-	nop
-	lw		$v1, TIMER1_VALUE($a3)
-	sw		$v0, 0($a2)
-	bne		$v0, $v1, .Lhwait2_loop
-	nop
-
-.Lvsync_exit:
-
-	lw		$ra, 0($sp)
-	lw		$s0, 4($sp)
-	lw		$v0, 8($sp)
-	jr		$ra
-	addiu	$sp, 12
-
-	
-.type _vsync_sub, @function	
-_vsync_sub:
-
-	# a0 - VSync destination count
-	# a1 - Timeout ratio (number of vsyncs to wait relative to vsync count)
-
-	addiu	$sp, -4
-	sw		$ra, 0($sp)
-
-	sll		$a1, 15						# Timeout counter
-	
-	la		$v0, _vsync_rcnt
-	lw		$v0, 0($v0)
-	nop
-	bge		$v0, $a0, .Lvsync_sub_exit
-	nop
-	
-.Lvsync_wait:
-
-	addiu	$a1, -1
-	
-	la		$v1, 0xffffffff
-	bne		$a1, $v1, .Lnot_timeout
-	nop
-
-	la		$a0, vsynctimeout_msg
-	jal		puts
-	addiu	$sp, -8
-	
-	jal		ChangeClearPAD
-	move	$a0, $0
-	
-	li		$a0, 3
-	jal		ChangeClearRCnt
-	move	$a1, $0
-	
-	addiu	$sp, 8
-	b		.Lvsync_sub_exit
-	li		$v0, -1
-	
-.Lnot_timeout:
-
-	la		$v0, _vsync_rcnt
-	lw		$v0, 0($v0)
-	nop
-	blt		$v0, $a0, .Lvsync_wait
-	nop
-	
-.Lvsync_sub_exit:
-
-	lw		$ra, 0($sp)
-	addiu	$sp, 4
-	jr		$ra
-	move	$v0, $0
-
-
-	
-.type _vsync_irq_callback, @function
-_vsync_irq_callback:
-	
-	lui		$a0, IOBASE
-	
-	la		$v1, _vsync_rcnt			# Increment VSync root counter
-	lw		$v0, 0($v1)
-	nop
-	addiu	$v0, 1
-	sw		$v0, 0($v1)
-
-	la		$v0, _vsync_cb_func			# Check if a callback function is set
-	lw		$v0, 0($v0)
-	nop
-	beqz	$v0, .Lno_callback
-	nop
-	
-	addiu	$sp, -4						# Save return address
-	sw		$ra, 0($sp)
-	jalr	$v0							# Execute user callback function
-	nop
-	lw		$ra, 0($sp)					# Restore previous return address
-	addiu	$sp, 4
-	
-	lui		$a0, IOBASE
-	
-.Lno_callback:
-
-	jr		$ra
-	nop
-	
-
-.section .data
-
-# VSync root counter
-.type _vsync_rcnt, @object
-_vsync_rcnt:
-	.word 0
-
-.type _vsync_lasthblank, @object
-_vsync_lasthblank:
-	.word 0
-	
-.comm	_vsync_cb_func, 4, 4
-
-.comm	_gpu_standard, 4, 4
-.comm	_gpu_current_field, 4, 4
-.comm	_hooks_installed, 4, 4
-
-	
-.type vsynctimeout_msg, @object
-vsynctimeout_msg:
-	.asciiz "VSync: timeout\n"
-	
-.type resetgraph_msg, @object
-resetgraph_msg:
-	.asciiz "ResetGraph:itb=%08x,ehk=%08x\n"
-	
-.type enableint_msg, @object
-enableint_msg:
-	.asciiz "ResetGraph:Interrupts enabled!\n"
-	
-.type cbhooks_msg, @object
-cbhooks_msg:
-	.asciiz "ResetGraph:Interrupt hooks enabled.\n"
-	
-.type abouttoen_msg, @object
-abouttoen_msg:
-	.asciiz "ResetGraph:About to init interrupts.\n"
-	
-.type sr_msg, @object
-sr_msg:
-	.asciiz "ResetGraph:SR=%x\n"
-	
-.global psxgpu_credits
-.type psxgpu_credits, @object
-psxgpu_credits:
-	.ascii "psxgpu programs by Lameguy64\n"
-	.asciiz "2020 PSn00bSDK Project / Meido-Tek Productions\n"
-	
-\ No newline at end of file
diff --git a/libpsn00b/psxgpu/setdefdispenv.c b/libpsn00b/psxgpu/setdefdispenv.c
deleted file mode 100644
index 2d7b2b4..0000000
--- a/libpsn00b/psxgpu/setdefdispenv.c
+++ /dev/null
@@ -1,22 +0,0 @@
-#include <sys/types.h>
-#include <psxgpu.h>
-
-DISPENV *SetDefDispEnv(DISPENV *disp, int x, int y, int w, int h) {
-
-	disp->disp.x = x;
-	disp->disp.y = y;
-	disp->disp.w = w;
-	disp->disp.h = h;
-
-	disp->screen.x = 0;
-	disp->screen.y = 0;
-	disp->screen.w = 0;
-	disp->screen.h = 0;
-
-	disp->isinter = 0;
-	disp->isrgb24 = 0;
-	disp->reverse = 0;
-
-	return disp;
-
-}
diff --git a/libpsn00b/psxgpu/setdefdrawenv.c b/libpsn00b/psxgpu/setdefdrawenv.c
deleted file mode 100644
index 6fd6086..0000000
--- a/libpsn00b/psxgpu/setdefdrawenv.c
+++ /dev/null
@@ -1,27 +0,0 @@
-#include <sys/types.h>
-#include <psxgpu.h>
-
-DRAWENV *SetDefDrawEnv(DRAWENV *draw, int x, int y, int w, int h) {
-
-	draw->clip.x = x;
-	draw->clip.y = y;
-	draw->clip.w = w;
-	draw->clip.h = h;
-
-	draw->ofs[0] = 0;
-	draw->ofs[1] = 0;
-
-	draw->tw.x = 0;
-	draw->tw.y = 0;
-	draw->tw.w = 0;
-	draw->tw.h = 0;
-
-	draw->tpage = 0x0a;
-	draw->dtd = 1;
-	draw->dfe = 0;
-	draw->isbg = 0;
-	setRGB0( draw, 0, 0, 0 );
-
-	return draw;
-
-}
diff --git a/libpsn00b/psxgpu/setdispmask.s b/libpsn00b/psxgpu/setdispmask.s
deleted file mode 100644
index d79006c..0000000
--- a/libpsn00b/psxgpu/setdispmask.s
+++ /dev/null
@@ -1,19 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.section .text
-
-
-.global SetDispMask
-.type SetDispMask, @function
-SetDispMask:
-	lui		$v1, IOBASE
-	andi	$a0, 0x1
-	lui		$v0, 0x300
-	ori		$v0, 0x1
-	sub		$v0, $a0
-	sw		$v0, GPU_GP1($v1)
-	jr		$ra
-	nop
-
diff --git a/libpsn00b/psxgpu/setvideomode.s b/libpsn00b/psxgpu/setvideomode.s
deleted file mode 100644
index b89b285..0000000
--- a/libpsn00b/psxgpu/setvideomode.s
+++ /dev/null
@@ -1,50 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-
-.section .text
-
-.global SetVideoMode
-.type SetVideoMode, @function
-SetVideoMode:
-	addiu	$sp, -4
-	sw		$ra, 0($sp)
-
-	jal		ReadGPUstat
-	nop
-
-	srl		$a1, $v0, 17
-	andi	$a1, 0x1f
-
-	srl		$v1, $v0, 14			# Reverse flag
-	andi	$v1, 1
-	sll		$v1, 6
-	or		$a1, $v1
-
-	srl		$v1, $v0, 16			# Horizontal resolution 2
-	andi	$v1, 1
-	sll		$v1, 6
-	or		$a1, $v1
-
-	andi	$a1, 0xf7				# Mask off PAL bit
-
-	la		$v0, _gpu_standard
-	beqz	$a0, .Lset_done
-	sw		$0 , 0($v0)
-	li		$v1, 1
-	sw		$v1, 0($v0)
-	b		.Lset_done
-	or		$a1, 0x8
-.Lset_done:
-
-	lui		$v0, 0x800				# Apply new mode
-	or		$a1, $v0
-	lui		$v0, IOBASE
-	sw		$a1, GPU_GP1($v0)
-
-	lw		$ra, 0($sp)
-	addiu	$sp, 4
-	jr		$ra
-	nop
-
diff --git a/libpsn00b/psxgpu/storeimage.s b/libpsn00b/psxgpu/storeimage.s
deleted file mode 100644
index 554e83c..0000000
--- a/libpsn00b/psxgpu/storeimage.s
+++ /dev/null
@@ -1,76 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.set RECT_x,	0
-.set RECT_y,	2
-.set RECT_w,	4
-.set RECT_h,	6
-
-.section .text
-
-
-.global StoreImage
-.type StoreImage, @function
-StoreImage:
-	addiu	$sp, -8
-	sw		$ra, 0($sp)
-	sw		$s0, 4($sp)
-
-	lui		$s0, IOBASE			# Set I/O segment base address
-
-.Lgpu_wait:						# Wait for GPU to be ready for commands and DMA
-	jal		ReadGPUstat
-	nop
-	srl		$v0, 0x1a
-	andi	$v0, 0x5
-	li		$v1, 5
-	bne		$v0, $v1, .Lgpu_wait
-	nop
-
-	lui		$v0, 0x400			# Set DMA direction to off
-	sw		$v0, GPU_GP1($s0)
-
-	lui		$v0, 0x0100			# Clear GPU cache
-	sw		$v0, GPU_GP0($s0)
-
-	lui		$v1, 0xc000			# Store image from VRAM
-	sw		$v1, GPU_GP0($s0)
-	lw		$v0, RECT_x($a0)	# Set XY and dimensions of image
-	lw		$v1, RECT_w($a0)
-	sw		$v0, GPU_GP0($s0)
-	sw		$v1, GPU_GP0($s0)
-
-	lui		$v0, 0x400			# Set DMA direction to VRAMtoCPU
-	ori		$v0, 0x3
-	sw		$v0, GPU_GP1($s0)
-
-	lhu		$v0, RECT_w($a0)	# Get rectangle size
-	lhu		$v1, RECT_h($a0)
-	nop
-	mult	$v0, $v1			# Calculate BCR value
-	mflo	$v1
-	srl		$v1, 0x4
-	sll		$v1, 0x10
-	ori		$v1, 0x8
-
-	sw		$a1, DMA2_MADR($s0)	# Set DMA base address and transfer length
-	sw		$v1, DMA2_BCR($s0)
-
-.Lgpu_wait_2:					# Wait for GPU to be ready for commands and DMA
-	jal		ReadGPUstat
-	nop
-	srl		$v0, 27
-	andi	$v0, 0x1
-	beqz	$v0, .Lgpu_wait_2
-	nop
-	
-	lui		$v0, 0x100			# Start DMA transfer
-	ori		$v0, 0x200
-	sw		$v0, DMA2_CHCR($s0)
-
-	lw		$ra, 0($sp)
-	lw		$s0, 4($sp)
-	jr		$ra
-	addiu	$sp, 8
-
diff --git a/libpsn00b/psxgpu/vsynccallback.s b/libpsn00b/psxgpu/vsynccallback.s
deleted file mode 100644
index 4be29c8..0000000
--- a/libpsn00b/psxgpu/vsynccallback.s
+++ /dev/null
@@ -1,25 +0,0 @@
-.set noreorder
-
-.section .text
-
-.global VSyncCallback
-.type VSyncCallback, @function
-VSyncCallback:
-	addiu	$sp, -8
-	sw		$ra, 0($sp)
-
-	jal		EnterCriticalSection
-	sw		$a0, 4($sp)
-
-	lw		$a0, 4($sp)
-	la		$v0, _vsync_cb_func
-	sw		$a0, 0($v0)
-
-	jal		ExitCriticalSection
-	nop
-
-	lw		$ra, 0($sp)
-	addiu	$sp, 8
-	jr		$ra
-	nop
-
diff --git a/libpsn00b/psxgte/applymatrixlv.s b/libpsn00b/psxgte/applymatrixlv.s
index 332a2f8..3180d0f 100644
--- a/libpsn00b/psxgte/applymatrixlv.s
+++ b/libpsn00b/psxgte/applymatrixlv.s
@@ -1,7 +1,7 @@
 .set noreorder
 
-.include "gtereg.h"
-.include "inline_s.h"
+.include "gtereg.inc"
+.include "inline_s.inc"
 
 .section .text
 
diff --git a/libpsn00b/psxgte/compmatrixlv.s b/libpsn00b/psxgte/compmatrixlv.s
index 95da5e9..2908eb9 100644
--- a/libpsn00b/psxgte/compmatrixlv.s
+++ b/libpsn00b/psxgte/compmatrixlv.s
@@ -1,7 +1,7 @@
 .set noreorder
 
-.include "gtereg.h"
-.include "inline_s.h"
+.include "gtereg.inc"
+.include "inline_s.inc"
 
 .set MATRIX_r11r12,	0
 .set MATRIX_r13r21,	4
diff --git a/libpsn00b/psxgte/initgeom.s b/libpsn00b/psxgte/initgeom.s
index 14ca293..d004ecc 100644
--- a/libpsn00b/psxgte/initgeom.s
+++ b/libpsn00b/psxgte/initgeom.s
@@ -1,6 +1,6 @@
 .set noreorder
 
-.include "gtereg.h"
+.include "gtereg.inc"
 
 .section .text
 
diff --git a/libpsn00b/psxgte/mulmatrix.s b/libpsn00b/psxgte/mulmatrix.s
index 19dabe8..08c79c2 100644
--- a/libpsn00b/psxgte/mulmatrix.s
+++ b/libpsn00b/psxgte/mulmatrix.s
@@ -1,7 +1,7 @@
 .set noreorder
 
-.include "gtereg.h"
-.include "inline_s.h"
+.include "gtereg.inc"
+.include "inline_s.inc"
 
 .section .text
 
diff --git a/libpsn00b/psxgte/mulmatrix0.s b/libpsn00b/psxgte/mulmatrix0.s
index 874226b..c2fd859 100644
--- a/libpsn00b/psxgte/mulmatrix0.s
+++ b/libpsn00b/psxgte/mulmatrix0.s
@@ -1,7 +1,7 @@
 .set noreorder
 
-.include "gtereg.h"
-.include "inline_s.h"
+.include "gtereg.inc"
+.include "inline_s.inc"
 
 .section .text
 
diff --git a/libpsn00b/psxgte/pushpopmatrix.s b/libpsn00b/psxgte/pushpopmatrix.s
index d10687a..ca6b992 100644
--- a/libpsn00b/psxgte/pushpopmatrix.s
+++ b/libpsn00b/psxgte/pushpopmatrix.s
@@ -1,7 +1,7 @@
 .set noreorder
 
-.include "gtereg.h"
-.include "inline_s.h"
+.include "gtereg.inc"
+.include "inline_s.inc"
 
 .section .text
 
diff --git a/libpsn00b/psxgte/readme.txt b/libpsn00b/psxgte/readme.txt
index 13067ee..3242474 100644
--- a/libpsn00b/psxgte/readme.txt
+++ b/libpsn00b/psxgte/readme.txt
@@ -21,9 +21,9 @@ Library developer(s):
 	
 Library header(s):
 
-	gtereg.h
+	gtereg.inc
 	inline_c.h
-	inline_s.h
+	inline_s.inc
 	psxgte.h
 	
 	
diff --git a/libpsn00b/psxgte/scalematrix.s b/libpsn00b/psxgte/scalematrix.s
index 3e83800..1b2b6dd 100644
--- a/libpsn00b/psxgte/scalematrix.s
+++ b/libpsn00b/psxgte/scalematrix.s
@@ -1,7 +1,7 @@
 .set noreorder
 
-.include "gtereg.h"
-.include "inline_s.h"
+.include "gtereg.inc"
+.include "inline_s.inc"
 
 .section .text
 
diff --git a/libpsn00b/psxgte/scalematrixl.s b/libpsn00b/psxgte/scalematrixl.s
index 014b85a..53c2d14 100644
--- a/libpsn00b/psxgte/scalematrixl.s
+++ b/libpsn00b/psxgte/scalematrixl.s
@@ -1,7 +1,7 @@
 .set noreorder
 
-.include "gtereg.h"
-.include "inline_s.h"
+.include "gtereg.inc"
+.include "inline_s.inc"
 
 .section .text
 
diff --git a/libpsn00b/psxgte/square0.s b/libpsn00b/psxgte/square0.s
index d037b7e..a8ca107 100644
--- a/libpsn00b/psxgte/square0.s
+++ b/libpsn00b/psxgte/square0.s
@@ -1,7 +1,7 @@
 .set noreorder
 
-.include "gtereg.h"
-.include "inline_s.h"
+.include "gtereg.inc"
+.include "inline_s.inc"
 
 .section .text
 
diff --git a/libpsn00b/psxgte/squareroot.s b/libpsn00b/psxgte/squareroot.s
index a038e18..72198b4 100644
--- a/libpsn00b/psxgte/squareroot.s
+++ b/libpsn00b/psxgte/squareroot.s
@@ -1,7 +1,7 @@
 .set noreorder
 
-.include "gtereg.h"
-.include "inline_s.h"
+.include "gtereg.inc"
+.include "inline_s.inc"
 
 .section .text
 
diff --git a/libpsn00b/psxgte/vectornormals.s b/libpsn00b/psxgte/vectornormals.s
index 939c4d1..85e94e6 100644
--- a/libpsn00b/psxgte/vectornormals.s
+++ b/libpsn00b/psxgte/vectornormals.s
@@ -1,8 +1,8 @@
 .set noreorder
 .set noat
 
-.include "gtereg.h"
-.include "inline_s.h"
+.include "gtereg.inc"
+.include "inline_s.inc"
 
 .section .text
 
diff --git a/libpsn00b/psxpress/README.md b/libpsn00b/psxpress/README.md
new file mode 100644
index 0000000..a894874
--- /dev/null
+++ b/libpsn00b/psxpress/README.md
@@ -0,0 +1,42 @@
+
+# PSn00bSDK MDEC library
+
+This is a fully open source reimplementation of the official SDK's "data
+compression" library. This library is made up of two parts, the MDEC API and
+functions to decompress Huffman-encoded bitstreams (.BS files, or frames in
+.STR files) into data to be fed to the MDEC. FMV playback is not part of this
+library (nor the official one) per se, but can implemented by using these APIs
+alongside some code to stream data from the CD drive.
+
+**Currently only version 1 and 2 bitstreams are supported**.
+
+## MDEC API
+
+The MDEC data input/output API is almost identical to the official one, with
+only two minor differences:
+
+- `DecDCTPutEnv()` takes a second argument specifying whether the MDEC shall be
+  put into monochrome or color mode (the original API only supported color).
+- A `DecDCTinRaw()` function was added for easier feeding of headerless data
+  buffers to the MDEC. This function does not affect how `DecDCTin()` works.
+
+## Decompression API
+
+The following functions are currently provided:
+
+- `DecDCTvlcStart()`, `DecDCTvlcContinue()`: a decompressor implementation that
+  uses a small (<1 KB) lookup table and leverages the GTE, written in assembly.
+  `DecDCTvlcCopyTable()` can optionally be called to temporarily move the table
+  to the scratchpad region to improve decompression speed.
+- `DecDCTvlcStart2()`, `DecDCTvlcContinue2()`: a different implementation using
+  a large (34 KB) lookup table in main RAM, written in C. The table must be
+  decompressed ahead of time using `DecDCTvlcBuild()`, but can be deallocated
+  when no longer needed.
+- `DecDCTvlc()`, `DecDCTvlc2()`: wrappers around the functions listed above,
+  for compatibility with the Sony SDK. Using them is not recommended.
+
+## SPU ADPCM encoding API
+
+The Sony library has functions that can be used to convert raw 16-bit PCM audio
+data to SPU ADPCM. These are currently unimplemented due to their limited
+usefulness, but might be added at some point.
diff --git a/libpsn00b/psxpress/generate_lookup_table.py b/libpsn00b/psxpress/generate_lookup_table.py
new file mode 100644
index 0000000..b40771f
--- /dev/null
+++ b/libpsn00b/psxpress/generate_lookup_table.py
@@ -0,0 +1,297 @@
+#!/usr/bin/env python3
+# Huffman lookup table generator script for psxpress
+# (C) 2022 spicyjpeg - MPL licensed
+
+import sys, json
+from array     import array
+from itertools import repeat
+from argparse  import ArgumentParser, FileType
+
+HUFFMAN_TREE = {
+	"10":			0xfe00, # End of block
+	"11":			(  0,  1 ),
+	"01": {
+		"1":		(  1,  1 ),
+		"00":		(  0,  2 ),
+		"01":		(  2,  1 )
+	},
+	"001": {
+		"01":		(  0,  3 ),
+		"10":		(  4,  1 ),
+		"11":		(  3,  1 ),
+		"00000":	( 13,  1 ),
+		"00001":	(  0,  6 ),
+		"00010":	( 12,  1 ),
+		"00011":	( 11,  1 ),
+		"00100":	(  3,  2 ),
+		"00101":	(  1,  3 ),
+		"00110":	(  0,  5 ),
+		"00111":	( 10,  1 )
+	},
+	"0001": {
+		"00":		(  7,  1 ),
+		"01":		(  6,  1 ),
+		"10":		(  1,  2 ),
+		"11":		(  5,  1 )
+	},
+	"00001": {
+		"00":		(  2,  2 ),
+		"01":		(  9,  1 ),
+		"10":		(  0,  4 ),
+		"11":		(  8,  1 )
+	},
+	"0000001": {
+		"000":		( 16,  1 ),
+		"001":		(  5,  2 ),
+		"010":		(  0,  7 ),
+		"011":		(  2,  3 ),
+		"100":		(  1,  4 ),
+		"101":		( 15,  1 ),
+		"110":		( 14,  1 ),
+		"111":		(  4,  2 )
+	},
+	"00000001": {
+		"0000":		(  0, 11 ),
+		"0001":		(  8,  2 ),
+		"0010":		(  4,  3 ),
+		"0011":		(  0, 10 ),
+		"0100":		(  2,  4 ),
+		"0101":		(  7,  2 ),
+		"0110":		( 21,  2 ),
+		"0111":		( 20,  1 ),
+		"1000":		(  0,  9 ),
+		"1001":		( 19,  1 ),
+		"1010":		( 18,  1 ),
+		"1011":		(  1,  5 ),
+		"1100":		(  3,  3 ),
+		"1101":		(  0,  8 ),
+		"1110":		(  6,  2 ),
+		"1111":		( 17,  1 )
+	},
+	"000000001": {
+		"0000":		( 10,  2 ),
+		"0001":		(  9,  2 ),
+		"0010":		(  5,  3 ),
+		"0011":		(  3,  4 ),
+		"0100":		(  2,  5 ),
+		"0101":		(  1,  7 ),
+		"0110":		(  1,  6 ),
+		"0111":		(  0, 15 ),
+		"1000":		(  0, 14 ),
+		"1001":		(  0, 13 ),
+		"1010":		(  0, 12 ),
+		"1011":		( 26,  1 ),
+		"1100":		( 25,  1 ),
+		"1101":		( 24,  1 ),
+		"1110":		( 23,  1 ),
+		"1111":		( 22,  1 )
+	},
+	"0000000001": {
+		"0000":		(  0, 31 ),
+		"0001":		(  0, 30 ),
+		"0010":		(  0, 29 ),
+		"0011":		(  0, 28 ),
+		"0100":		(  0, 27 ),
+		"0101":		(  0, 26 ),
+		"0110":		(  0, 25 ),
+		"0111":		(  0, 24 ),
+		"1000":		(  0, 23 ),
+		"1001":		(  0, 22 ),
+		"1010":		(  0, 21 ),
+		"1011":		(  0, 20 ),
+		"1100":		(  0, 19 ),
+		"1101":		(  0, 18 ),
+		"1110":		(  0, 17 ),
+		"1111":		(  0, 16 )
+	},
+	"00000000001": {
+		"0000":		(  0, 40 ),
+		"0001":		(  0, 39 ),
+		"0010":		(  0, 38 ),
+		"0011":		(  0, 37 ),
+		"0100":		(  0, 36 ),
+		"0101":		(  0, 35 ),
+		"0110":		(  0, 34 ),
+		"0111":		(  0, 33 ),
+		"1000":		(  0, 32 ),
+		"1001":		(  1, 14 ),
+		"1010":		(  1, 13 ),
+		"1011":		(  1, 12 ),
+		"1100":		(  1, 11 ),
+		"1101":		(  1, 10 ),
+		"1110":		(  1,  9 ),
+		"1111":		(  1,  8 )
+	},
+	"000000000001": {
+		"0000":		(  1, 18 ),
+		"0001":		(  1, 17 ),
+		"0010":		(  1, 16 ),
+		"0011":		(  1, 15 ),
+		"0100":		(  6,  3 ),
+		"0101":		( 16,  2 ),
+		"0110":		( 15,  2 ),
+		"0111":		( 14,  2 ),
+		"1000":		( 13,  2 ),
+		"1001":		( 12,  2 ),
+		"1010":		( 11,  2 ),
+		"1011":		( 31,  1 ),
+		"1100":		( 30,  1 ),
+		"1101":		( 29,  1 ),
+		"1110":		( 28,  1 ),
+		"1111":		( 27,  1 )
+	}
+}
+
+## Utilities
+
+def to_int10(value):
+	clamped = min(max(int(value), -0x200), 0x1ff)
+	return clamped + (0 if clamped >= 0 else 0x400)
+
+def uint32_to_lines(data, indent = "\t", columns = 6):
+	for offset in range(0, len(data), columns):
+		line = f"{indent}0x{data[offset]:08x}"
+
+		for item in data[(offset + 1):(offset + columns)]:
+			line += f", 0x{item:08x}"
+
+		yield line
+
+## Table generation
+
+def iterate_tree(tree):
+	for code, value in tree.items():
+		if type(value) is dict:
+			# Iterate through any subtree recursively.
+			for suffix, _value in iterate_tree(value):
+				yield f"{code}{suffix}", _value
+
+		elif type(value) is tuple:
+			run_length, ac = value
+			yield f"{code}0", (run_length << 10) | to_int10(ac)
+			yield f"{code}1", (run_length << 10) | to_int10(-ac)
+
+		else:
+			yield code, value
+
+def generate_table(codes, table_bits, prefix_bits = 0):
+	table = array("I", repeat(0, 2 ** table_bits))
+
+	for code, value in codes:
+		used_bits = len(code)
+		free_bits = table_bits - (used_bits - prefix_bits)
+		index     = int(code[prefix_bits:], 2) << free_bits
+
+		# Fill out every entry in the table whose index starts with the same
+		# string of bits.
+		for combo in range(2 ** free_bits):
+			table[index | combo] = (used_bits << 16) | value
+
+	return table
+
+def compress_table(table):
+	values     = []
+	last_value = table[0]
+	run_length = 0
+
+	for value in table[1:]:
+		if value == last_value and run_length < 0x7ff:
+			run_length += 1
+			continue
+
+		# The run length is stored in the top 11 bits of each value, which are
+		# otherwise unused.
+		values.append((run_length << 21) | last_value)
+		last_value = value
+		run_length = 0
+
+	values.append((run_length << 21) | last_value)
+	return array("I", values)
+
+## Main
+
+UNCOMPRESSED_TEMPLATE = """static const DECDCTTAB {name} = {{
+	.lut = {{
+{short}
+	}},
+	.lut00 = {{
+{long}
+	}}
+}};
+"""
+COMPRESSED_TEMPLATE = """static const uint32_t {name}[{length}] = {{
+{table}
+}};
+"""
+
+def get_args():
+	parser = ArgumentParser(
+		description = "Generates a Huffman lookup table structure, to be used by DecDCTvlc2()."
+	)
+	parser.add_argument(
+		"-c", "--compress",
+		action = "store_true",
+		help   = "generate run-length compressed data instead of a DECDCTTAB struct"
+	)
+	parser.add_argument(
+		"-n", "--name",
+		type    = str,
+		default = "_default_huffman_table",
+		help    = "set the symbol name in the generated C source",
+		metavar = "file"
+	)
+	parser.add_argument(
+		"-t", "--tree",
+		type    = FileType("rt"),
+		help    = "use a custom Huffman tree from the specified JSON file",
+		metavar = "json_file"
+	)
+	parser.add_argument(
+		"-o", "--output",
+		type    = FileType("wt"),
+		default = sys.stdout,
+		help    = "where to output generated table (stdout by default)",
+		metavar = "file"
+	)
+
+	return parser.parse_args()
+
+def main():
+	args = get_args()
+	tree = json.load(args.tree) if args.tree else HUFFMAN_TREE
+
+	short_codes, short_bits = [], 0
+	long_codes, long_bits   = [], 0
+
+	for pair in iterate_tree(tree):
+		if (code := pair[0]).startswith("00000000"):
+			long_codes.append(pair)
+			long_bits = max(long_bits, len(code) - 8)
+		else:
+			short_codes.append(pair)
+			short_bits = max(short_bits, len(code))
+
+	short_table = generate_table(short_codes, short_bits, 0)
+	long_table  = generate_table(long_codes,  long_bits,  8)
+
+	if args.compress:
+		short_table.extend(long_table)
+		table = compress_table(short_table)
+
+		source = COMPRESSED_TEMPLATE.format(
+			name   = args.name,
+			length = len(table),
+			table  = ",\n".join(uint32_to_lines(table, "\t"))
+		)
+	else:
+		source = UNCOMPRESSED_TEMPLATE.format(
+			name  = args.name,
+			short = ",\n".join(uint32_to_lines(short_table, "\t\t")),
+			long  = ",\n".join(uint32_to_lines(long_table,  "\t\t"))
+		)
+
+	with args.output as _file:
+		_file.write(source)
+
+if __name__ == "__main__":
+	main()
diff --git a/libpsn00b/psxpress/mdec.c b/libpsn00b/psxpress/mdec.c
index ca4c75a..9c82d6b 100644
--- a/libpsn00b/psxpress/mdec.c
+++ b/libpsn00b/psxpress/mdec.c
@@ -9,7 +9,8 @@
 #include <psxpress.h>
 #include <hwregs_c.h>
 
-#define MDEC_SYNC_TIMEOUT 0x1000000
+#define DMA_CHUNK_LENGTH	32
+#define MDEC_SYNC_TIMEOUT	0x1000000
 
 /* Default IDCT matrix and quantization tables */
 
@@ -82,7 +83,7 @@ static const DECDCTENV _default_mdec_env = {
 
 /* Public API */
 
-void DecDCTReset(int32_t mode) {
+void DecDCTReset(int mode) {
 	EnterCriticalSection();
 
 	DMA_DPCR   |= 0x000000bb; // Enable DMA0 and DMA1
@@ -96,7 +97,7 @@ void DecDCTReset(int32_t mode) {
 		DecDCTPutEnv(0, 0);
 }
 
-void DecDCTPutEnv(const DECDCTENV *env, int32_t mono) {
+void DecDCTPutEnv(const DECDCTENV *env, int mono) {
 	const DECDCTENV *_env = env ? env : &_default_mdec_env;
 	DecDCTinSync(0);
 
@@ -109,14 +110,14 @@ void DecDCTPutEnv(const DECDCTENV *env, int32_t mono) {
 	DecDCTinSync(0);
 }
 
-void DecDCTin(const uint32_t *data, int32_t mode) {
+void DecDCTin(const uint32_t *data, int mode) {
 	uint32_t header = *data;
 	if (mode == DECDCT_MODE_RAW)
 		MDEC0 = header;
 	else if (mode & DECDCT_MODE_24BPP)
-		MDEC0 = header | 0x30000000;
+		MDEC0 = 0x30000000 | (header & 0xffff);
 	else
-		MDEC0 = header | 0x38000000 | ((mode & 2) << 24); // Bit 25 = mask
+		MDEC0 = 0x38000000 | (header & 0xffff) | ((mode & 2) << 24); // Bit 25 = mask
 
 	DecDCTinRaw((const uint32_t *) &(data[1]), header & 0xffff);
 }
@@ -125,20 +126,25 @@ void DecDCTin(const uint32_t *data, int32_t mode) {
 // data length as an argument rather than parsing it from the first 4 bytes of
 // the stream.
 void DecDCTinRaw(const uint32_t *data, size_t length) {
+	if ((length >= DMA_CHUNK_LENGTH) && (length % DMA_CHUNK_LENGTH)) {
+		printf("psxmdec: transfer data length (%d) is not a multiple of %d, rounding\n", length, DMA_CHUNK_LENGTH);
+		length += DMA_CHUNK_LENGTH - 1;
+	}
+
 	DMA_MADR(0) = (uint32_t) data;
-	if (length < 32)
+	if (length < DMA_CHUNK_LENGTH)
 		DMA_BCR(0) = 0x00010000 | length;
 	else
-		DMA_BCR(0) = 0x00000020 | ((length / 32) << 16);
+		DMA_BCR(0) = DMA_CHUNK_LENGTH | ((length / DMA_CHUNK_LENGTH) << 16);
 
 	DMA_CHCR(0) = 0x01000201;
 }
 
-int32_t DecDCTinSync(int32_t mode) {
+int DecDCTinSync(int mode) {
 	if (mode)
 		return (MDEC1 >> 29) & 1;
 
-	for (uint32_t i = MDEC_SYNC_TIMEOUT; i; i--) {
+	for (int i = MDEC_SYNC_TIMEOUT; i; i--) {
 		if (!(MDEC1 & (1 << 29)))
 			return 0;
 	}
@@ -151,19 +157,19 @@ void DecDCTout(uint32_t *data, size_t length) {
 	DecDCToutSync(0);
 
 	DMA_MADR(1) = (uint32_t) data;
-	if (length < 32)
+	if (length < DMA_CHUNK_LENGTH)
 		DMA_BCR(1) = 0x00010000 | length;
 	else
-		DMA_BCR(1) = 0x00000020 | ((length / 32) << 16);
+		DMA_BCR(1) = DMA_CHUNK_LENGTH | ((length / DMA_CHUNK_LENGTH) << 16);
 
 	DMA_CHCR(1) = 0x01000200;
 }
 
-int32_t DecDCToutSync(int32_t mode) {
+int DecDCToutSync(int mode) {
 	if (mode)
 		return (DMA_CHCR(1) >> 24) & 1;
 
-	for (uint32_t i = MDEC_SYNC_TIMEOUT; i; i--) {
+	for (int i = MDEC_SYNC_TIMEOUT; i; i--) {
 		if (!(DMA_CHCR(1) & (1 << 24)))
 			return 0;
 	}
diff --git a/libpsn00b/psxpress/vlc.c b/libpsn00b/psxpress/vlc.c
new file mode 100644
index 0000000..4e3e283
--- /dev/null
+++ b/libpsn00b/psxpress/vlc.c
@@ -0,0 +1,130 @@
+/*
+ * PSn00bSDK MDEC library (support code for the main VLC decompressor)
+ * (C) 2022 spicyjpeg - MPL licensed
+ */
+
+#include <stdint.h>
+#include <stddef.h>
+#include <string.h>
+#include <psxpress.h>
+
+/* Huffman code lookup table */
+
+#define _val1(rl, dc)		(((rl) << 10) | ((uint16_t) (dc) & 0x3ff))
+#define _val2(rl, dc, len)	(_val1(rl, dc) | (len << 16))
+
+#define _pair(rl, dc)		_val1(rl, dc), _val1(rl, -(dc))
+#define _pair2(rl, dc, len)	_val2(rl, dc, len), _val2(rl, -(dc), len)
+#define _pair3(rl, dc, len) \
+	_val2(rl, dc, len), _val2(rl, dc, len), \
+	_val2(rl, -(dc), len), _val2(rl, -(dc), len)
+#define _pair4(rl, dc, len) \
+	_val2(rl, dc, len), _val2(rl, dc, len), _val2(rl, dc, len), _val2(rl, dc, len), \
+	_val2(rl, dc, len), _val2(rl, dc, len), _val2(rl, dc, len), _val2(rl, dc, len), \
+	_val2(rl, -(dc), len), _val2(rl, -(dc), len), _val2(rl, -(dc), len), _val2(rl, -(dc), len), \
+	_val2(rl, -(dc), len), _val2(rl, -(dc), len), _val2(rl, -(dc), len), _val2(rl, -(dc), len)
+
+// This table isn't compressed since it makes no sense to compress less than a
+// kilobyte's worth of data.
+static const DECDCTTAB _default_huffman_table = {
+	.lut0 = {
+		// 11 x
+		_pair( 0,  1)
+	},
+	.lut2 = {
+		// 01 0xx
+		_pair2( 0,  2, 5), _pair2( 2,  1, 5),
+		// 01 1x-
+		_pair3( 1,  1, 4)
+	},
+	.lut3 = {
+		// 001 00xxxx
+		_pair2(13,  1, 9), _pair2( 0,  6, 9), _pair2(12,  1, 9), _pair2(11,  1, 9),
+		_pair2( 3,  2, 9), _pair2( 1,  3, 9), _pair2( 0,  5, 9), _pair2(10,  1, 9),
+		// 001 xxx---
+		_pair4( 0,  3, 6), _pair4( 4,  1, 6), _pair4( 3,  1, 6)
+	},
+	.lut4 = {
+		// 0001 xxx
+		_pair( 7,  1), _pair( 6,  1), _pair( 1,  2), _pair( 5,  1)
+	},
+	.lut5 = {
+		// 00001 xxx
+		_pair( 2,  2), _pair( 9,  1), _pair( 0,  4), _pair( 8,  1)
+	},
+	.lut7 = {
+		// 0000001 xxxx
+		_pair(16,  1), _pair( 5,  2), _pair( 0,  7), _pair( 2,  3),
+		_pair( 1,  4), _pair(15,  1), _pair(14,  1), _pair( 4,  2)
+	},
+	.lut8 = {
+		// 00000001 xxxxx
+		_pair( 0, 11), _pair( 8,  2), _pair( 4,  3), _pair( 0, 10),
+		_pair( 2,  4), _pair( 7,  2), _pair(21,  1), _pair(20,  1),
+		_pair( 0,  9), _pair(19,  1), _pair(18,  1), _pair( 1,  5),
+		_pair( 3,  3), _pair( 0,  8), _pair( 6,  2), _pair(17,  1)
+	},
+	.lut9 = {
+		// 000000001 xxxxx
+		_pair(10,  2), _pair( 9,  2), _pair( 5,  3), _pair( 3,  4),
+		_pair( 2,  5), _pair( 1,  7), _pair( 1,  6), _pair( 0, 15),
+		_pair( 0, 14), _pair( 0, 13), _pair( 0, 12), _pair(26,  1),
+		_pair(25,  1), _pair(24,  1), _pair(23,  1), _pair(22,  1)
+	},
+	.lut10 = {
+		// 0000000001 xxxxx
+		_pair( 0, 31), _pair( 0, 30), _pair( 0, 29), _pair( 0, 28),
+		_pair( 0, 27), _pair( 0, 26), _pair( 0, 25), _pair( 0, 24),
+		_pair( 0, 23), _pair( 0, 22), _pair( 0, 21), _pair( 0, 20),
+		_pair( 0, 19), _pair( 0, 18), _pair( 0, 17), _pair( 0, 16)
+	},
+	.lut11 = {
+		// 00000000001 xxxxx
+		_pair( 0, 40), _pair( 0, 39), _pair( 0, 38), _pair( 0, 37),
+		_pair( 0, 36), _pair( 0, 35), _pair( 0, 34), _pair( 0, 33),
+		_pair( 0, 32), _pair( 1, 14), _pair( 1, 13), _pair( 1, 12),
+		_pair( 1, 11), _pair( 1, 10), _pair( 1,  9), _pair( 1,  8)
+	},
+	.lut12 = {
+		// 000000000001 xxxxx
+		_pair( 1, 18), _pair( 1, 17), _pair( 1, 16), _pair( 1, 15),
+		_pair( 6,  3), _pair(16,  2), _pair(15,  2), _pair(14,  2),
+		_pair(13,  2), _pair(12,  2), _pair(11,  2), _pair(31,  1),
+		_pair(30,  1), _pair(29,  1), _pair(28,  1), _pair(27,  1)
+	}
+};
+
+/* Internal globals */
+
+// Note that DecDCTvlc() and DecDCTvlc2() do *not* share the same variables.
+static VLC_Context	_default_context;
+static size_t		_max_buffer_size = 0;
+
+const DECDCTTAB		*_vlc_huffman_table = &_default_huffman_table;
+
+/* Stateful VLC decoder API (for Sony SDK compatibility) */
+
+int DecDCTvlc(const uint32_t *bs, uint32_t *buf) {
+	if (bs)
+		return DecDCTvlcStart(&_default_context, buf, _max_buffer_size, bs);
+	else
+		return DecDCTvlcContinue(&_default_context, buf, _max_buffer_size);
+}
+
+size_t DecDCTvlcSize(size_t size) {
+	size_t old_size  = _max_buffer_size;
+	_max_buffer_size = size;
+
+	return old_size;
+}
+
+/* Lookup table relocation API */
+
+void DecDCTvlcCopyTable(DECDCTTAB *addr) {
+	if (addr) {
+		_vlc_huffman_table = addr;
+		memcpy(addr, &_default_huffman_table, sizeof(DECDCTTAB));
+	} else {
+		_vlc_huffman_table = &_default_huffman_table;
+	}
+}
diff --git a/libpsn00b/psxpress/vlc.s b/libpsn00b/psxpress/vlc.s
new file mode 100644
index 0000000..fe51642
--- /dev/null
+++ b/libpsn00b/psxpress/vlc.s
@@ -0,0 +1,404 @@
+# PSn00bSDK MDEC library (GTE-accelerated VLC decompressor)
+# (C) 2022 spicyjpeg - MPL licensed
+#
+# Register map:
+# - $a0 = ctx
+# - $a1 = output
+# - $a2 = max_size
+# - $a3 = input
+# - $t0 = window
+# - $t1 = next_window
+# - $t2 = remaining
+# - $t3 = quant_scale
+# - $t4 = is_v3
+# - $t5 = bit_offset
+# - $t6 = block_index
+# - $t7 = coeff_index
+# - $t8 = _vlc_huffman_table
+# - $t9 = &ac_jump_area
+
+.set noreorder
+
+.set VLC_Context_input,			0
+.set VLC_Context_window,		4
+.set VLC_Context_next_window,	8
+.set VLC_Context_remaining,		12
+.set VLC_Context_quant_scale,	16
+.set VLC_Context_is_v3,			18
+.set VLC_Context_bit_offset,	19
+.set VLC_Context_block_index,	20
+.set VLC_Context_coeff_index,	21
+
+.set DECDCTSMALLTAB_lut0,	0
+.set DECDCTSMALLTAB_lut2,	4
+.set DECDCTSMALLTAB_lut3,	36
+.set DECDCTSMALLTAB_lut4,	292
+.set DECDCTSMALLTAB_lut5,	308
+.set DECDCTSMALLTAB_lut7,	324
+.set DECDCTSMALLTAB_lut8,	356
+.set DECDCTSMALLTAB_lut9,	420
+.set DECDCTSMALLTAB_lut10,	484
+.set DECDCTSMALLTAB_lut11,	548
+.set DECDCTSMALLTAB_lut12,	612
+
+.section .text.DecDCTvlcStart
+.global DecDCTvlcStart
+.type DecDCTvlcStart, @function
+DecDCTvlcStart:
+	# Create a new context on-the-fly without writing it to memory then jump
+	# into DecDCTvlcContinue(), skipping context loading.
+	lw    $t0, 8($a3) # window = (bs->data[0] << 16) | (bs->data[0] >> 16)
+	nop
+	srl   $v0, $t0, 16
+	sll   $t0, 16
+
+	lw    $t1, 12($a3) # next_window = (bs->data[1] << 16) | (bs->data[1] >> 16)
+	or    $t0, $v0
+	srl   $v0, $t1, 16
+	sll   $t1, 16
+
+	lhu   $t2, 0($a3) # remaining = bs->uncomp_length * 2
+	or    $t1, $v0
+
+	lhu   $t3, 4($a3) # quant_scale = (bs->quant_scale & 63) << 10
+	sll   $t2, 1
+	andi  $t3, 63
+
+	lhu   $t4, 6($a3) # is_v3 = !(bs->version < 3)
+	sll   $t3, 10
+	sltiu $t4, $t4, 3
+	xori  $t4, 1
+
+	li    $t5, 32 # bit_offset = 32
+	li    $t6, 5 # block_index = 5
+	li    $t7, 0 # coeff_index = 0
+	j     _vlc_skip_context_load
+	addiu $a3, 16 # input = &(bs->data[2])
+
+.section .text.DecDCTvlcContinue
+.global DecDCTvlcContinue
+.type DecDCTvlcContinue, @function
+DecDCTvlcContinue:
+	lw    $a3, VLC_Context_input($a0)
+	lw    $t0, VLC_Context_window($a0)
+	lw    $t1, VLC_Context_next_window($a0)
+	lw    $t2, VLC_Context_remaining($a0)
+	lhu   $t3, VLC_Context_quant_scale($a0)
+	lb    $t4, VLC_Context_is_v3($a0)
+	lb    $t5, VLC_Context_bit_offset($a0)
+	lb    $t6, VLC_Context_block_index($a0)
+	lb    $t7, VLC_Context_coeff_index($a0)
+
+_vlc_skip_context_load:
+	# Determine how many bytes to output. This whole block of code basically
+	# does this:
+	#   max_size   = min((max_size - 1) * 2, remaining)
+	#   remaining -= max_size
+	bgtz  $a2, .Lmax_size_valid # if (max_size <= 0) max_size = 0x7ffe0000
+	addiu $a2, -1 # else max_size = (max_size - 1) * 2
+	lui   $a2, 0x3fff
+.Lmax_size_valid:
+	sll   $a2, 1
+
+	blt   $a2, $t2, .Lmax_size_ok # if (max_size > remaining) max_size = remaining
+	lui   $v1, 0x3800
+	move  $a2, $t2
+.Lmax_size_ok:
+	subu  $t2, $a2 # remaining -= max_size
+
+	# Write the length of the data that will be decoded to first 4 bytes of the
+	# output buffer, which will be then parsed by DecDCTin().
+	srl   $v0, $a2, 1 # output[0] = 0x38000000 | (max_size / 2)
+	or    $v0, $v1
+	sw    $v0, 0($a1)
+
+	# Obtain the addresses of the lookup table and jump area in advance so that
+	# they don't have to be retrieved for each coefficient decoded.
+	lw    $t8, _vlc_huffman_table
+	la    $t9, .Lac_jump_area
+
+	beqz  $a2, .Lstop_processing
+	addiu $a1, 4 # output = (uint16_t *) &output[1]
+
+.Lprocess_next_code_loop: # while (max_size)
+	# This is the "hot" part of the decoder, executed for each code in the
+	# bitstream. The first step is to determine if the next code is a DC or AC
+	# coefficient. The GTE is also given the task of counting the number of
+	# leading zeroes/ones, which takes 2 more cycles.
+	bnez  $t7, .Lprocess_ac_coefficient
+	mtc2  $t0, $30
+	bnez  $t4, .Lprocess_dc_v3_coefficient
+	#nop
+
+.Lprocess_dc_v2_coefficient: # if (!coeff_index && !is_v3)
+	# The DC coefficient in version 2 frames is not compressed.
+	srl   $v0, $t0, 22 # *output = (window >> (32 - 10)) | quant_scale
+	or    $v0, $t3
+	addiu $t7, 1 # coeff_index++
+	sll   $t0, 10 # window <<= 10
+	addiu $t5, -10 # bit_offset -= 10
+	b     .Lwrite_value
+	addiu $t7, 1 # coeff_index++
+
+.Lprocess_dc_v3_coefficient: # if (!coeff_index && is_v3)
+	# TODO: version 3 is currently not supported.
+	jr    $ra
+	li    $v0, -1
+	#b     .Lwrite_value
+
+.Lprocess_ac_coefficient: # if (coeff_index)
+	# Check whether the prefix code is one of the shorter, more common ones.
+	srl   $v0, $t0, 30
+	li    $v1, 3
+	beq   $v0, $v1, .Lac_prefix_11
+	li    $v1, 2
+	beq   $v0, $v1, .Lac_prefix_10
+	li    $v1, 1
+	beq   $v0, $v1, .Lac_prefix_01
+	#srl   $v0, $t0, 29
+	#beq   $v0, $v1, .Lac_prefix_001
+	#nop
+
+	# If the code is longer, retrieve the number of leading zeroes from the GTE
+	# and use it as an index into the jump area. Each block in the area is 8
+	# instructions long and handles decoding a specific prefix.
+	mfc2  $v0, $31
+	nop
+	andi  $v0, 15 # jump_addr = &ac_jump_area[(prefix % 16) * 8 * sizeof(u32)]
+	sll   $v0, 5
+	addu  $v0, $t9
+	jr    $v0
+	nop
+
+.Lac_prefix_11:
+	# Prefix 11 is followed by a single bit.
+	srl   $v0, $t0, 28 # index = ((window >> (32 - 2 - 1)) & 1) * sizeof(u16)
+	andi  $v0, 2
+	addu  $v0, $t8 # value = table->lut0[index]
+	lhu   $v0, DECDCTSMALLTAB_lut0($v0)
+	sll   $t0, 3 # window <<= 3
+	addiu $t5, -3 # bit_offset -= 3
+	b     .Lwrite_value
+	addiu $t7, 1 # coeff_index++
+
+.Lac_jump_area:
+.Lac_prefix_10:
+	# Prefix 10 marks the end of a block.
+	li    $v0, 0xfe00 # value = 0xfe00
+	sll   $t0, 2 # window <<= 2
+	addiu $t5, -2 # bit_offset -= 2
+	addiu $t6, -1 # block_index--
+	bgez  $t6, .Lwrite_value
+	li    $t7, 0 # coeff_index = 0
+	b     .Lwrite_value
+	li    $t6, 5 # if (block_index < 0) block_index = 5
+
+.Lac_prefix_01:
+	# Prefix 01 can be followed by a 2-bit lookup index starting with 1, or a
+	# 3-bit lookup index starting with 0. A 32-bit lookup table is used,
+	# containing both MDEC codes and lengths.
+	srl   $v0, $t0, 25 # index = ((window >> (32 - 2 - 3)) & 7) * sizeof(u32)
+	andi  $v0, 28
+	addu  $v0, $t8 # value = table->lut2[index]
+	lw    $v0, DECDCTSMALLTAB_lut2($v0)
+	addiu $t7, 1 # coeff_index++
+	b     .Lupdate_window_and_write
+	srl   $v1, $v0, 16 # length = value >> 16
+	.word 0
+
+.Lac_prefix_001:
+	# Prefix 001 can be followed by a 6-bit lookup index starting with 00, or a
+	# 3-bit lookup index starting with 01/10/11.
+	srl   $v0, $t0, 21 # index = ((window >> (32 - 3 - 6)) & 63) * sizeof(u32)
+	andi  $v0, 252
+	addu  $v0, $t8 # value = table->lut3[index]
+	lw    $v0, DECDCTSMALLTAB_lut3($v0)
+	addiu $t7, 1 # coeff_index++
+	b     .Lupdate_window_and_write
+	srl   $v1, $v0, 16 # length = value >> 16
+	.word 0
+
+.Lac_prefix_0001:
+	# Prefix 0001 is followed by a 3-bit lookup index.
+	srl   $v0, $t0, 24 # index = ((window >> (32 - 4 - 3)) & 7) * sizeof(u16)
+	andi  $v0, 14
+	addu  $v0, $t8 # value = table->lut4[index]
+	lhu   $v0, DECDCTSMALLTAB_lut4($v0)
+	sll   $t0, 7 # window <<= 4 + 3
+	addiu $t5, -7 # bit_offset -= 4 + 3
+	b     .Lwrite_value
+	addiu $t7, 1 # coeff_index++
+
+.Lac_prefix_00001:
+	# Prefix 00001 is followed by a 3-bit lookup index.
+	srl   $v0, $t0, 23 # index = ((window >> (32 - 5 - 3)) & 7) * sizeof(u16)
+	andi  $v0, 14
+	addu  $v0, $t8 # value = table->lut5[index]
+	lhu   $v0, DECDCTSMALLTAB_lut5($v0)
+	sll   $t0, 8 # window <<= 5 + 3
+	addiu $t5, -8 # bit_offset -= 5 + 3
+	b     .Lwrite_value
+	addiu $t7, 1 # coeff_index++
+
+.Lac_prefix_000001:
+	# Prefix 000001 is an escape code followed by a full 16-bit MDEC value.
+	srl   $v0, $t0, 10 # value = window >> (32 - 6 - 16)
+	sll   $t0, 22 # window <<= 6 + 16
+	addiu $t5, -22 # bit_offset -= 6 + 16
+	b     .Lwrite_value
+	addiu $t7, 1 # coeff_index++
+	.word 0, 0, 0
+
+.Lac_prefix_0000001:
+	# Prefix 0000001 is followed by a 4-bit lookup index.
+	srl   $v0, $t0, 20 # index = ((window >> (32 - 7 - 4)) & 15) * sizeof(u16)
+	andi  $v0, 30
+	addu  $v0, $t8 # value = table->lut7[index]
+	lhu   $v0, DECDCTSMALLTAB_lut7($v0)
+	sll   $t0, 11 # window <<= 7 + 4
+	addiu $t5, -11 # bit_offset -= 7 + 4
+	b     .Lwrite_value
+	addiu $t7, 1 # coeff_index++
+
+.Lac_prefix_00000001:
+	# Prefix 00000001 is followed by a 5-bit lookup index.
+	srl   $v0, $t0, 18 # index = ((window >> (32 - 8 - 5)) & 31) * sizeof(u16)
+	andi  $v0, 62
+	addu  $v0, $t8 # value = table->lut8[index]
+	lhu   $v0, DECDCTSMALLTAB_lut8($v0)
+	sll   $t0, 13 # window <<= 8 + 5
+	addiu $t5, -13 # bit_offset -= 8 + 5
+	b     .Lwrite_value
+	addiu $t7, 1 # coeff_index++
+
+.Lac_prefix_000000001:
+	# Prefix 000000001 is followed by a 5-bit lookup index.
+	srl   $v0, $t0, 17 # index = ((window >> (32 - 9 - 5)) & 31) * sizeof(u16)
+	andi  $v0, 62
+	addu  $v0, $t8 # value = table->lut9[index]
+	lhu   $v0, DECDCTSMALLTAB_lut9($v0)
+	sll   $t0, 14 # window <<= 9 + 5
+	addiu $t5, -14 # bit_offset -= 9 + 5
+	b     .Lwrite_value
+	addiu $t7, 1 # coeff_index++
+
+.Lac_prefix_0000000001:
+	# Prefix 0000000001 is followed by a 5-bit lookup index.
+	srl   $v0, $t0, 16 # index = ((window >> (32 - 10 - 5)) & 31) * sizeof(u16)
+	andi  $v0, 62
+	addu  $v0, $t8 # value = table->lut10[index]
+	lhu   $v0, DECDCTSMALLTAB_lut10($v0)
+	sll   $t0, 15 # window <<= 10 + 5
+	addiu $t5, -15 # bit_offset -= 10 + 5
+	b     .Lwrite_value
+	addiu $t7, 1 # coeff_index++
+
+.Lac_prefix_00000000001:
+	# Prefix 00000000001 is followed by a 5-bit lookup index.
+	srl   $v0, $t0, 15 # index = ((window >> (32 - 11 - 5)) & 31) * sizeof(u16)
+	andi  $v0, 62
+	addu  $v0, $t8 # value = table->lut11[index]
+	lhu   $v0, DECDCTSMALLTAB_lut11($v0)
+	sll   $t0, 16 # window <<= 11 + 5
+	addiu $t5, -16 # bit_offset -= 11 + 5
+	b     .Lwrite_value
+	addiu $t7, 1 # coeff_index++
+
+.Lac_prefix_000000000001:
+	# Prefix 000000000001 is followed by a 5-bit lookup index.
+	srl   $v0, $t0, 14 # index = ((window >> (32 - 12 - 5)) & 31) * sizeof(u16)
+	andi  $v0, 62
+	addu  $v0, $t8 # value = table->lut12[index]
+	lhu   $v0, DECDCTSMALLTAB_lut12($v0)
+	sll   $t0, 17 # window <<= 12 + 5
+	addiu $t5, -17 # bit_offset -= 12 + 5
+	b     .Lwrite_value
+	addiu $t7, 1 # coeff_index++
+
+	# Prefix 0000000000001 is not valid.
+	beqz  $t0, .Lstop_processing
+	nop
+	jr    $ra
+	li    $v0, -1
+	.word 0, 0, 0, 0
+
+	# Prefix 00000000000001 is not valid.
+	beqz  $t0, .Lstop_processing
+	nop
+	jr    $ra
+	li    $v0, -1
+	.word 0, 0, 0, 0
+
+	# Prefix 000000000000001 is not valid.
+	beqz  $t0, .Lstop_processing
+	nop
+	jr    $ra
+	li    $v0, -1
+	.word 0, 0, 0, 0
+
+	# Prefix 0000000000000001 is not valid.
+	beqz  $t0, .Lstop_processing
+	nop
+	jr    $ra
+	li    $v0, -1
+	#.word 0, 0, 0, 0
+
+.Lupdate_window_and_write:
+	sllv  $t0, $t0, $v1 # window <<= length
+	subu  $t5, $v1 # bit_offset -= length
+.Lwrite_value:
+	sh    $v0, 0($a1)
+.Lfeed_bitstream:
+	# Update the window. This makes sure the next iteration of the loop will be
+	# able to read up to 32 bits from the bitstream.
+	bgez  $t5, .Lskip_feeding # if (bit_offset < 0)
+	addiu $a2, -1 # max_size--
+
+	subu  $v0, $0, $t5 # window = next_window << (-bit_offset)
+	sllv  $t0, $t1, $v0
+	lw    $t1, 0($a3) # next_window = (*input << 16) | (*input >> 16)
+	addiu $t5, 32 # bit_offset += 32
+	srl   $v0, $t1, 16
+	sll   $t1, 16
+	or    $t1, $v0
+	addiu $a3, 4 # input++
+
+.Lskip_feeding:
+	srlv  $v0, $t1, $t5 # window |= next_window >> bit_offset
+	or    $t0, $v0
+
+	bnez  $a2, .Lprocess_next_code_loop
+	addiu $a1, 2 # output++
+
+.Lstop_processing:
+	# If remaining = 0, skip flushing the context, pad the output buffer with
+	# end-of-block codes if necessary and return 0. Otherwise flush the context
+	# and return 1.
+	beqz  $t2, .Lpad_output_buffer
+	nop
+
+	sw    $a3, VLC_Context_input($a0)
+	sw    $t0, VLC_Context_window($a0)
+	sw    $t1, VLC_Context_next_window($a0)
+	sw    $t2, VLC_Context_remaining($a0)
+	sh    $t3, VLC_Context_quant_scale($a0)
+	sb    $t4, VLC_Context_is_v3($a0)
+	sb    $t5, VLC_Context_bit_offset($a0)
+	sb    $t6, VLC_Context_block_index($a0)
+	sb    $t7, VLC_Context_coeff_index($a0)
+
+	jr    $ra
+	li    $v0, 1
+
+.Lpad_output_buffer:
+	beqz  $a2, .Lreturn_zero
+	li    $v0, 0xfe00
+.Lpad_output_buffer_loop: # while (max_size)
+	sh    $v0, 0($a1) # *output = 0xfe00
+	addiu $a2, -1 # max_size--
+	bnez  $a2, .Lpad_output_buffer_loop
+	addiu $a1, 2 # output++
+
+.Lreturn_zero:
+	jr    $ra
+	li    $v0, 0
diff --git a/libpsn00b/psxpress/vlc2.c b/libpsn00b/psxpress/vlc2.c
new file mode 100644
index 0000000..73b54b2
--- /dev/null
+++ b/libpsn00b/psxpress/vlc2.c
@@ -0,0 +1,240 @@
+/*
+ * PSn00bSDK MDEC library (alternate VLC decompressor and support code)
+ * (C) 2022 spicyjpeg - MPL licensed
+ */
+
+#include <stdint.h>
+#include <stddef.h>
+#include <psxpress.h>
+
+#define _min(x, y) (((x) < (y)) ? (x) : (y))
+
+/* Huffman code lookup table */
+
+#define TABLE_LENGTH 226
+
+// This table is run-length compressed, with the number of repetitions of each
+// value stored in the upper 11 bits which would be otherwise unused. It is
+// decompressed at runtime by DecDCTvlcBuild().
+static const uint32_t _compressed_table[TABLE_LENGTH] = {
+	0x03e00000, 0x000d000b, 0x000d03f5, 0x000d2002, 0x000d23fe, 0x000d1003,
+	0x000d13fd, 0x000d000a, 0x000d03f6, 0x000d0804, 0x000d0bfc, 0x000d1c02,
+	0x000d1ffe, 0x000d5402, 0x000d57fe, 0x000d5001, 0x000d53ff, 0x000d0009,
+	0x000d03f7, 0x000d4c01, 0x000d4fff, 0x000d4801, 0x000d4bff, 0x000d0405,
+	0x000d07fb, 0x000d0c03, 0x000d0ffd, 0x000d0008, 0x000d03f8, 0x000d1802,
+	0x000d1bfe, 0x000d4401, 0x000d47ff, 0x006b4001, 0x006b43ff, 0x006b1402,
+	0x006b17fe, 0x006b0007, 0x006b03f9, 0x006b0803, 0x006b0bfd, 0x006b0404,
+	0x006b07fc, 0x006b3c01, 0x006b3fff, 0x006b3801, 0x006b3bff, 0x006b1002,
+	0x006b13fe, 0x0fe00000, 0x03e80802, 0x03e80bfe, 0x03e82401, 0x03e827ff,
+	0x03e80004, 0x03e803fc, 0x03e82001, 0x03e823ff, 0x07e71c01, 0x07e71fff,
+	0x07e71801, 0x07e71bff, 0x07e70402, 0x07e707fe, 0x07e71401, 0x07e717ff,
+	0x01e93401, 0x01e937ff, 0x01e90006, 0x01e903fa, 0x01e93001, 0x01e933ff,
+	0x01e92c01, 0x01e92fff, 0x01e90c02, 0x01e90ffe, 0x01e90403, 0x01e907fd,
+	0x01e90005, 0x01e903fb, 0x01e92801, 0x01e92bff, 0x0fe60003, 0x0fe603fd,
+	0x0fe61001, 0x0fe613ff, 0x0fe60c01, 0x0fe60fff, 0x1fe50002, 0x1fe503fe,
+	0x1fe50801, 0x1fe50bff, 0x3fe40401, 0x3fe407ff, 0xffe2fe00, 0x7fe30001,
+	0x7fe303ff, 0x03e00000, 0x00110412, 0x001107ee, 0x00110411, 0x001107ef,
+	0x00110410, 0x001107f0, 0x0011040f, 0x001107f1, 0x00111803, 0x00111bfd,
+	0x00114002, 0x001143fe, 0x00113c02, 0x00113ffe, 0x00113802, 0x00113bfe,
+	0x00113402, 0x001137fe, 0x00113002, 0x001133fe, 0x00112c02, 0x00112ffe,
+	0x00117c01, 0x00117fff, 0x00117801, 0x00117bff, 0x00117401, 0x001177ff,
+	0x00117001, 0x001173ff, 0x00116c01, 0x00116fff, 0x00300028, 0x003003d8,
+	0x00300027, 0x003003d9, 0x00300026, 0x003003da, 0x00300025, 0x003003db,
+	0x00300024, 0x003003dc, 0x00300023, 0x003003dd, 0x00300022, 0x003003de,
+	0x00300021, 0x003003df, 0x00300020, 0x003003e0, 0x0030040e, 0x003007f2,
+	0x0030040d, 0x003007f3, 0x0030040c, 0x003007f4, 0x0030040b, 0x003007f5,
+	0x0030040a, 0x003007f6, 0x00300409, 0x003007f7, 0x00300408, 0x003007f8,
+	0x006f001f, 0x006f03e1, 0x006f001e, 0x006f03e2, 0x006f001d, 0x006f03e3,
+	0x006f001c, 0x006f03e4, 0x006f001b, 0x006f03e5, 0x006f001a, 0x006f03e6,
+	0x006f0019, 0x006f03e7, 0x006f0018, 0x006f03e8, 0x006f0017, 0x006f03e9,
+	0x006f0016, 0x006f03ea, 0x006f0015, 0x006f03eb, 0x006f0014, 0x006f03ec,
+	0x006f0013, 0x006f03ed, 0x006f0012, 0x006f03ee, 0x006f0011, 0x006f03ef,
+	0x006f0010, 0x006f03f0, 0x00ee2802, 0x00ee2bfe, 0x00ee2402, 0x00ee27fe,
+	0x00ee1403, 0x00ee17fd, 0x00ee0c04, 0x00ee0ffc, 0x00ee0805, 0x00ee0bfb,
+	0x00ee0407, 0x00ee07f9, 0x00ee0406, 0x00ee07fa, 0x00ee000f, 0x00ee03f1,
+	0x00ee000e, 0x00ee03f2, 0x00ee000d, 0x00ee03f3, 0x00ee000c, 0x00ee03f4,
+	0x00ee6801, 0x00ee6bff, 0x00ee6401, 0x00ee67ff, 0x00ee6001, 0x00ee63ff,
+	0x00ee5c01, 0x00ee5fff, 0x00ee5801, 0x00ee5bff
+};
+
+/* Internal globals */
+
+// Note that DecDCTvlc() and DecDCTvlc2() do *not* share the same variables.
+static VLC_Context	_default_context;
+static size_t		_max_buffer_size = 0;
+
+const DECDCTTAB2	*_vlc_huffman_table2 = 0;
+
+/* VLC decoder */
+
+#define _get_bits_unsigned(length)	(((uint32_t) window) >> (32 - (length)))
+#define _get_bits_signed(length)	(((int32_t)  window) >> (32 - (length)))
+#define _advance_window(num) \
+	window    <<= (num); \
+	bit_offset -= (num);
+
+int __attribute__((optimize(3))) DecDCTvlcContinue2(
+	VLC_Context *ctx, uint32_t *buf, size_t max_size
+) {
+	const uint32_t	*input		= ctx->input;
+	uint32_t		remaining	= ctx->remaining;
+	uint32_t		window		= ctx->window;
+	uint32_t		next_window	= ctx->next_window;
+	uint16_t		quant_scale	= ctx->quant_scale;
+	int				block_index	= ctx->block_index;
+	int				coeff_index	= ctx->coeff_index;
+	int				bit_offset	= ctx->bit_offset;
+	int				is_v3		= ctx->is_v3;
+
+	//if (!_vlc_huffman_table2)
+		//return -1;
+	if (!max_size)
+		max_size = 0x7fffffff;
+
+	// Write the length of the data that will be decoded to first 4 bytes of
+	// the output buffer, which will be then parsed by DecDCTin().
+	max_size   = _min((max_size - 1) * 2, remaining);
+	remaining -= max_size;
+
+	*buf = 0x38000000 | (max_size / 2);
+	uint16_t *output = (uint16_t *) &buf[1];
+
+	for (; max_size; max_size--) {
+		uint32_t value;
+
+		if (coeff_index) {
+			// Parse the next AC coefficient. Most codes are decompressed via
+			// the lookup table, however some need special handling.
+			if ((window >> 30) == 0b10) {
+				// Prefix 10 marks the end of a block.
+				*output = 0xfe00;
+				_advance_window(2);
+
+				coeff_index = -1;
+				block_index++;
+				if (block_index > 5)
+					block_index = 0;
+			} else if ((window >> 26) == 0b000001) {
+				// Prefix 000001 is an escape code followed by a full 16-bit
+				// MDEC value.
+				*output = (uint16_t) _get_bits_unsigned(22);
+				_advance_window(22);
+			} else if (window >> 24) {
+				// The first lookup table is for codes that not start with
+				// 00000000.
+				value = _vlc_huffman_table2->lut[_get_bits_unsigned(13)];
+				_advance_window(value >> 16);
+				*output = (uint16_t) value;
+			} else {
+				// If the code starts with 00000000, use the second lookup
+				// table.
+				value = _vlc_huffman_table2->lut00[_get_bits_unsigned(17)];
+				_advance_window(value >> 16);
+				*output = (uint16_t) value;
+			}
+		} else {
+			// Parse the DC (first) coefficient for this block. Version 2
+			// simply stores the signed 10-bit value as-is, while version 3
+			// uses a delta encoding combined with a compression method similar
+			// to exp-Golomb.
+			if (is_v3) {
+				// TODO: version 3 is currently not supported.
+				return -1;
+			} else {
+				value   = _get_bits_unsigned(10);
+				*output = value | quant_scale;
+				_advance_window(10);
+			}
+		}
+
+		output++;
+		coeff_index++;
+
+		// Update the bitstream window. For whatever reason Sony's DecDCTvlc()
+		// implementation inefficiently reads the input stream 16 bits at a
+		// time and processes each 16-bit word starting from the the MSB, so an
+		// endianness conversion is necessary to preserve bit order when
+		// reading 32 bits at a time. Also note that the PS1 CPU is not capable
+		// of shifting by more than 31 bits - it will shift by 0 bits instead!
+		if (bit_offset < 0) {
+			window      = next_window << (-bit_offset);
+			bit_offset += 32;
+			next_window = (*input << 16) | (*input >> 16);
+			input++;
+		};
+		window |= next_window >> bit_offset;
+	}
+
+	// Pad the buffer with end-of-block codes if necessary.
+	for (; max_size; max_size--)
+		*(output++) = 0xfe00;
+
+	if (!remaining)
+		return 0;
+
+	ctx->input			= input;
+	ctx->remaining		= remaining;
+	ctx->window			= window;
+	ctx->next_window	= next_window;
+	ctx->block_index	= block_index;
+	ctx->coeff_index	= coeff_index;
+	ctx->bit_offset		= bit_offset;
+	return 1;
+}
+
+int DecDCTvlcStart2(
+	VLC_Context *ctx, uint32_t *buf, size_t max_size, const uint32_t *bs
+) {
+	const BS_Header *header = (const BS_Header *) bs;
+	const uint32_t  *input  = (const uint32_t *) &header[1];
+
+	if (!_vlc_huffman_table2)
+		return -1;
+	if (header->version > 3)
+		return -1;
+
+	ctx->input			= &input[2];
+	ctx->remaining		= (header->mdec0_header & 0xffff) * 2;
+	ctx->window			= (input[0] << 16) | (input[0] >> 16);
+	ctx->next_window	= (input[1] << 16) | (input[1] >> 16);
+	ctx->quant_scale	= (header->quant_scale & 63) << 10;
+	ctx->block_index	= 0;
+	ctx->coeff_index	= 0;
+	ctx->bit_offset		= 32;
+	ctx->is_v3			= (header->version == 3);
+
+	return DecDCTvlcContinue2(ctx, buf, max_size);
+}
+
+/* Stateful VLC decoder API (for Sony SDK compatibility) */
+
+int DecDCTvlc2(const uint32_t *bs, uint32_t *buf, DECDCTTAB2 *table) {
+	if (table)
+		_vlc_huffman_table2 = table;
+
+	if (bs)
+		return DecDCTvlcStart2(&_default_context, buf, _max_buffer_size, bs);
+	else
+		return DecDCTvlcContinue2(&_default_context, buf, _max_buffer_size);
+}
+
+size_t DecDCTvlcSize2(size_t size) {
+	size_t old_size  = _max_buffer_size;
+	_max_buffer_size = size;
+
+	return old_size;
+}
+
+/* Lookup table decompressor */
+
+void DecDCTvlcBuild(DECDCTTAB2 *table) {
+	uint32_t *output    = (uint32_t *) table;
+	_vlc_huffman_table2 = table;
+
+	for (int i = 0; i < TABLE_LENGTH; i++) {
+		uint32_t value = _compressed_table[i] & 0x001fffff;
+
+		for (int j = (_compressed_table[i] >> 21); j >= 0; j--)
+			*(output++) = value;
+	}
+}
diff --git a/libpsn00b/psxsio/_sio_control.s b/libpsn00b/psxsio/_sio_control.s
index bf3b9b9..6378def 100644
--- a/libpsn00b/psxsio/_sio_control.s
+++ b/libpsn00b/psxsio/_sio_control.s
@@ -1,6 +1,6 @@
 .set noreorder
 
-.include "hwregs_a.h"
+.include "hwregs_a.inc"
 
 .section .text
 
diff --git a/libpsn00b/psxspu/common.c b/libpsn00b/psxspu/common.c
new file mode 100644
index 0000000..55a3dba
--- /dev/null
+++ b/libpsn00b/psxspu/common.c
@@ -0,0 +1,157 @@
+/*
+ * PSn00bSDK SPU library (common functions)
+ * (C) 2022 spicyjpeg - MPL licensed
+ */
+
+#include <stdint.h>
+#include <stdio.h>
+#include <psxspu.h>
+#include <hwregs_c.h>
+
+#define WRITABLE_AREA_ADDR	0x200
+#define DMA_CHUNK_LENGTH	16
+#define STATUS_TIMEOUT		0x100000
+
+/* Internal globals */
+
+static SPU_TransferMode	_transfer_mode = SPU_TRANSFER_BY_DMA;
+static uint16_t			_transfer_addr = WRITABLE_AREA_ADDR;
+
+/* SPU initialization */
+
+static void _wait_status(uint16_t mask, uint16_t value) {
+	for (int i = STATUS_TIMEOUT; i; i--) {
+		if ((SPU_STAT & mask) == value)
+			return;
+	}
+
+	printf("psxspu: status register timeout (0x%04x)\n", SPU_STAT);
+}
+
+void SpuInit(void) {
+	SPU_CTRL = 0x0000; // SPU disabled
+	_wait_status(0x001f, 0x0000);
+
+	SPU_MASTER_VOL_L	= 0;
+	SPU_MASTER_VOL_R	= 0;
+	SPU_REVERB_VOL_L	= 0;
+	SPU_REVERB_VOL_R	= 0;
+	SPU_KEY_OFF			= 0x00ffffff;
+	SPU_FM_MODE			= 0;
+	SPU_NOISE_MODE		= 0;
+	SPU_REVERB_ON		= 0;
+	SPU_REVERB_ADDR		= 0xfffe;
+	SPU_CD_VOL_L		= 0;
+	SPU_CD_VOL_R		= 0;
+	SPU_EXT_VOL_L		= 0;
+	SPU_EXT_VOL_R		= 0;
+
+	DMA_DPCR   |= 0x000b0000; // Enable DMA4
+	DMA_CHCR(4) = 0x00000201; // Stop DMA4
+
+	SPU_CTRL = 0xc011; // Enable SPU, DAC, CD audio, set manual transfer mode
+	_wait_status(0x001f, 0x0011);
+
+	// Upload a dummy ADPCM block to the first 16 bytes of SPU RAM. This may be
+	// freely used or overwritten.
+	SPU_ADDR = WRITABLE_AREA_ADDR;
+	_wait_status(0x0400, 0x0000);
+
+	SPU_DATA = 0x0500;
+	for (int i = 7; i; i--)
+		SPU_DATA = 0x0000;
+
+	// "Play" the dummy block on all channels. This will reset the start
+	// address and ADSR envelope status of each channel.
+	for (int i = 0; i < 24; i++) {
+		SPU_CH_VOL_L(i) = 0;
+		SPU_CH_VOL_R(i) = 0;
+		SPU_CH_FREQ(i)  = 0x1000;
+		SPU_CH_ADDR(i)  = WRITABLE_AREA_ADDR;
+	}
+
+	// Sony's implementation leaves everything muted, however it makes sense to
+	// turn up at least the master and CD audio volume by default.
+	SPU_KEY_ON			= 0x00ffffff;
+	SPU_MASTER_VOL_L	= 0x3fff;
+	SPU_MASTER_VOL_R	= 0x3fff;
+	SPU_CD_VOL_L		= 0x3fff;
+	SPU_CD_VOL_R		= 0x3fff;
+}
+
+/* SPU RAM transfer API */
+
+static void _load_store_data(uint32_t *data, size_t length, int mode) {
+	if (length % 4)
+		printf("psxspu: can't transfer a number of bytes that isn't multiple of 4\n");
+
+	length /= 4;
+	if ((length >= DMA_CHUNK_LENGTH) && (length % DMA_CHUNK_LENGTH)) {
+		printf("psxspu: transfer data length (%d) is not a multiple of %d, rounding\n", length, DMA_CHUNK_LENGTH);
+		length += DMA_CHUNK_LENGTH - 1;
+	}
+
+	SPU_CTRL &= 0xffcf; // Disable DMA request
+	_wait_status(0x0030, 0x0000);
+
+	// Enable DMA request for writing (2) or reading (3)
+	SPU_ADDR  = _transfer_addr;
+	SPU_CTRL |= mode << 4;
+	_wait_status(0x0400, 0x0000);
+
+	DMA_MADR(4) = (uint32_t) data;
+	if (length < DMA_CHUNK_LENGTH)
+		DMA_BCR(4) = 0x00010000 | length;
+	else
+		DMA_BCR(4) = DMA_CHUNK_LENGTH | ((length / DMA_CHUNK_LENGTH) << 16);
+
+	DMA_CHCR(4) = 0x01000200 | ((mode & 1) ^ 1);
+}
+
+void SpuRead(uint32_t *data, size_t size) {
+	_load_store_data(data, size, 3);
+}
+
+void SpuWrite(const uint32_t *data, size_t size) {
+	if (_transfer_addr < WRITABLE_AREA_ADDR)
+		return;
+
+	// I/O transfer mode is not that useful, but whatever.
+	if (_transfer_mode) {
+		SPU_ADDR = _transfer_addr;
+		SPU_CTRL = (SPU_CTRL & 0xffcf) | 0x0010; // Manual transfer mode
+		_wait_status(0x0400, 0x0000);
+
+		for (int i = size; i; i -= 4) {
+			uint32_t value = *(data++);
+
+			SPU_DATA = (uint16_t) value;
+			SPU_DATA = (uint16_t) (value >> 16);
+		}
+
+		return;
+	}
+
+	_load_store_data((uint32_t *) data, size, 2);
+}
+
+SPU_TransferMode SpuSetTransferMode(SPU_TransferMode mode) {
+	_transfer_mode = mode;
+	return mode;
+}
+
+uint32_t SpuSetTransferStartAddr(uint32_t addr) {
+	if (addr > 0x7ffff)
+		return 0;
+
+	_transfer_addr = (addr + 7) / 8;
+	return addr;
+}
+
+int SpuIsTransferCompleted(int mode) {
+	if (!mode)
+		return ((SPU_STAT >> 10) & 1) ^ 1;
+
+	_wait_status(0x0400, 0x0000);
+	return 1;
+}
diff --git a/libpsn00b/psxspu/readme.txt b/libpsn00b/psxspu/readme.txt
index 3ed90d0..07bbf97 100644
--- a/libpsn00b/psxspu/readme.txt
+++ b/libpsn00b/psxspu/readme.txt
@@ -3,29 +3,24 @@ PSX SPU Library, part of PSn00bSDK
 
 Licensed under Mozilla Public License
 
-	Open source implementation of the SPU library written mostly in MIPS
-assembly. Currently only supports SPU init, uploading sample data using DMA
-transfer and basic sample playback but is currently lacking a bunch of
-important functions.
-
-	Very work in progress currently.
-	
+Open source implementation of the SPU library written entirely in C. Currently
+only supports SPU initialization, reading/writing SPU RAM using DMA and basic
+sample playback. Most of the official API is not going to be implemented as the
+vast majority of it is just inefficient wrappers around accessing SPU registers
+directly, which can be done already using the macros defined in hwregs_c.h.
 
 Library developer(s):
 
-	Lameguy64
-	
-	
+	Lameguy64 (initial implementation in assembly)
+	spicyjpeg
+
 Library header(s):
 
 	psxspu.h
 
-
 Todo list:
 
 	* SPU RAM allocation routines yet to be implemented (heap must only be
 	  stored in main RAM and not SPU RAM like in the official SDK).
 
-	* SpuKeyOn() is actually not part of the official library.
-
 	* SPU reverb configuration functions yet to be implemented.
diff --git a/libpsn00b/psxspu/spuinit.s b/libpsn00b/psxspu/spuinit.s
deleted file mode 100644
index 6966213..0000000
--- a/libpsn00b/psxspu/spuinit.s
+++ /dev/null
@@ -1,124 +0,0 @@
-.set noreorder
-.set noat
-
-.include "hwregs_a.h"
-
-.section .text
-
-
-.global SpuInit
-.type SpuInit, @function
-SpuInit:
-	
-	addiu	$sp, -4
-	sw		$ra, 0($sp)
-	
-	lui		$v1, IOBASE
-	
-	# Stop and mute everything
-	
-	sh		$0 , SPU_CTRL($v1)			# Clear control settings
-	jal		SpuCtrlSync
-	move	$a0, $0
-	
-	sh		$0 , SPU_MASTER_VOL_L($v1)	# Clear master volume
-	sh		$0 , SPU_MASTER_VOL_R($v1)
-
-	sh		$0 , SPU_REVERB_VOL_L($v1)	# Clear reverb volume
-	sh		$0 , SPU_REVERB_VOL_R($v1)
-	
-	sh		$0 , SPU_CD_VOL_L($v1)		# Clear CD volume
-	sh		$0 , SPU_CD_VOL_R($v1)
-	
-	sh		$0 , SPU_EXT_VOL_L($v1)		# Clear external audio volume
-	sh		$0 , SPU_EXT_VOL_R($v1)
-	
-	sw		$0 , SPU_FM_MODE($v1)		# Turn off FM modes
-	sw		$0 , SPU_NOISE_MODE($v1)	# Turn off noise modes
-	sw		$0 , SPU_REVERB_ON($v1)		# Turn off reverb modes
-	
-	li		$v0, 0xfffe
-	sh		$v0, SPU_REVERB_ADDR($v1)
-	
-	lui		$v0, 0x0200;
-	ori		$v0, 0x3fff;
-	
-	# Clear all voices
-	
-	addiu	$a1, $sp, -20
-	sw		$0 , 0($a1)
-	sw		$0 , 4($a1)
-	sw		$0 , 8($a1)
-	sw		$0 , 12($a1)
-	
-	li		$a2, 23
-	
-.Lclear_voices:
-	jal		SpuSetVoiceRaw
-	move	$a0, $a2
-	addiu	$a2, -1
-	bgez	$a2, .Lclear_voices
-	nop
-	
-	addiu	$v0, $0, -1					# Set all keys to off
-	sw		$v0, SPU_KEY_OFF($v1)
-	
-	li		$v0, 0x4					# Set SPU data transfer control
-	sh		$v0, SPU_DMA_CTRL($v1)			# (usually always 0x4)
-	
-	lw		$v0, DMA_DPCR($v1)			# Enable DMA channel 4 (SPU DMA)
-	lui		$at, 0xb
-	or		$v0, $at
-	sw		$v0, DMA_DPCR($v1)
-	
-	li		$v0, 0xC001					# Enable SPU
-	sh		$v0, SPU_CTRL($v1)
-	jal		SpuCtrlSync
-	move	$a0, $v0
-	
-	li		$v0, 0x3fff					# Activate master volume
-	sh		$v0, SPU_MASTER_VOL_L($v1)
-	sh		$v0, SPU_MASTER_VOL_R($v1)
-	
-	sh		$v0, SPU_CD_VOL_L($v1)		# Activate CD volume
-	sh		$v0, SPU_CD_VOL_R($v1)
-	
-	lw		$ra, 0($sp)
-	addiu	$sp, 4
-	jr		$ra
-	nop
-	
-
-# Waits until bits 0-5 of SPUSTAT are equal to SPUCNT
-#
-# Destroys v0, v1, a0
-#
-.global SpuCtrlSync		
-.type SpuCtrlSync, @function
-SpuCtrlSync:
-	lui		$v1, IOBASE
-	andi	$a0, 0x3f
-.Lctrl_wait:
-	lhu		$v0, SPU_STAT($v1)		# Get SPUSTAT value
-	nop
-	andi	$v0, 0x3f
-	bne		$v0, $a0, .Lctrl_wait	# Wait until SPUCNT and SPUSTAT are equal
-	nop
-	jr		$ra
-	nop
-
-	
-# Waits until SPU has finished transfers
-#
-.global SpuWait
-.type SpuWait, @function
-SpuWait:
-	lui		$v0, IOBASE
-	lhu		$v0, SPU_STAT($v0)
-	nop
-	andi	$v0, 0x400
-	bnez	$v0, SpuWait
-	nop
-	jr		$ra
-	nop
-	
-\ No newline at end of file
diff --git a/libpsn00b/psxspu/spukeyon.s b/libpsn00b/psxspu/spukeyon.s
deleted file mode 100644
index 33592c2..0000000
--- a/libpsn00b/psxspu/spukeyon.s
+++ /dev/null
@@ -1,17 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.section .data
-
-
-.global SpuKeyOn
-.type SpuKeyOn, @function
-SpuKeyOn:
-	lui		$v1, IOBASE
-	li		$v0, 1
-	sll		$v0, $a0
-	sw		$v0, SPU_KEY_ON($v1)
-	sw		$v0, SPU_KEY_ON($v1)
-	jr		$ra
-	nop
-\ No newline at end of file
diff --git a/libpsn00b/psxspu/spureverbon.s b/libpsn00b/psxspu/spureverbon.s
deleted file mode 100644
index 635fac3..0000000
--- a/libpsn00b/psxspu/spureverbon.s
+++ /dev/null
@@ -1,16 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.section .data
-
-
-.global SpuReverbOn
-.type SpuReverbOn, @function
-SpuReverbOn:
-	lui		$v1, IOBASE
-	li		$v0, 1
-	sll		$v0, $a0
-	sw		$v0, SPU_REVERB_ON($v1)
-	jr		$ra
-	nop
-\ No newline at end of file
diff --git a/libpsn00b/psxspu/spusetkey.s b/libpsn00b/psxspu/spusetkey.s
deleted file mode 100644
index 4ad0cff..0000000
--- a/libpsn00b/psxspu/spusetkey.s
+++ /dev/null
@@ -1,26 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.section .data
-
-
-.global SpuSetKey
-.type SpuSetKey, @function
-SpuSetKey:
-	# a0 - 0: key off, 1: key on
-	# a1 - Voice bit mask
-	
-	lui		$a2, IOBASE
-	
-	beqz	$a0, .Lkey_off
-	nop
-	
-	jr		$ra
-	sw		$a1, SPU_KEY_ON($a2)
-	
-.Lkey_off:
-
-	jr		$ra
-	sw		$a1, SPU_KEY_OFF($a2)
-	
-\ No newline at end of file
diff --git a/libpsn00b/psxspu/spusetreverb.s b/libpsn00b/psxspu/spusetreverb.s
deleted file mode 100644
index 8257812..0000000
--- a/libpsn00b/psxspu/spusetreverb.s
+++ /dev/null
@@ -1,25 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-.section .data
-
-
-.global SpuSetReverb
-.type SpuSetReverb, @function
-SpuSetReverb:
-	addiu	$sp, -4
-	sw		$ra, 0($sp)
-
-	lui		$v1, IOBASE
-	lhu		$v0, SPU_CTRL($v1)
-	nop
-	ori		$v0, 0x80					# Enable reverb
-	sh		$v0, SPU_CTRL($v1)
-	jal		SpuCtrlSync
-	move	$a0, $v0
-	
-	lw		$ra, 0($sp)
-	addiu	$sp, 4
-	jr		$ra
-	nop
-\ No newline at end of file
diff --git a/libpsn00b/psxspu/spusetreverbaddr.s b/libpsn00b/psxspu/spusetreverbaddr.s
deleted file mode 100644
index 089a91a..0000000
--- a/libpsn00b/psxspu/spusetreverbaddr.s
+++ /dev/null
@@ -1,25 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-
-.section .text
-
-.global SpuSetReverbAddr
-.type SpuSetReverbAddr, @function
-SpuSetReverbAddr:
-	lui		$a3, IOBASE
-	srl		$a0, 3
-	sh		$a0, SPU_REVERB_ADDR($a3)
-	jr		$ra
-	nop
-	
-	
-.global SpuSetReverbVolume
-.type SpuSetReverbVolume, @function
-SpuSetReverbVolume:
-	lui		$a3, IOBASE
-	sh		$a0, SPU_REVERB_VOL_L($a3)
-	sh		$a1, SPU_REVERB_VOL_R($a3)
-	jr		$ra
-	nop
-\ No newline at end of file
diff --git a/libpsn00b/psxspu/spusetvoiceraw.s b/libpsn00b/psxspu/spusetvoiceraw.s
deleted file mode 100644
index 43450f6..0000000
--- a/libpsn00b/psxspu/spusetvoiceraw.s
+++ /dev/null
@@ -1,60 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-
-.set PARAM_L,		0
-.set PARAM_R,		2
-.set PARAM_FREQ,	4
-.set PARAM_ADDR,	6
-.set PARAM_LOOP,	8
-.set PARAM_RES,		10
-.set PARAM_ADSR,	12
-
-
-.section .text
-
-.global SpuSetVoiceRaw
-.type SpuSetVoiceRaw, @function
-SpuSetVoiceRaw:
-	
-	# a0 - Voice number
-	# a1 - Address to parameters
-	
-	sll		$a0, 4
-	addiu	$a0, SPU_VOICE_BASE
-	
-	lui		$v1, IOBASE
-	or		$a0, $v1
-	
-	lhu		$v0, PARAM_L($a1)
-	nop
-	sh		$v0, SPU_VOICE_VOL_L($a0)
-	
-	lhu		$v0, PARAM_R($a1)
-	nop
-	sh		$v0, SPU_VOICE_VOL_R($a0)
-	
-	lhu		$v0, PARAM_FREQ($a1)
-	nop
-	sh		$v0, SPU_VOICE_FREQ($a0)
-	
-	lhu		$v0, PARAM_ADDR($a1)
-	nop
-	sh		$v0, SPU_VOICE_ADDR($a0)
-	
-	lhu		$v0, PARAM_LOOP($a1)
-	nop
-	sh		$v0, SPU_VOICE_LOOP($a0)
-	
-	
-	lw		$v0, PARAM_ADSR($a1)
-	nop
-	sh		$v0, SPU_VOICE_ADSR_L($a0)
-	srl		$v0, 16
-	sh		$v0, SPU_VOICE_ADSR_H($a0)
-	
-	
-	jr		$ra
-	nop
-	
-\ No newline at end of file
diff --git a/libpsn00b/psxspu/transfer.s b/libpsn00b/psxspu/transfer.s
deleted file mode 100644
index adcdb33..0000000
--- a/libpsn00b/psxspu/transfer.s
+++ /dev/null
@@ -1,108 +0,0 @@
-.set noreorder
-
-.include "hwregs_a.h"
-
-
-.section .text
-
-.global SpuSetTransferMode
-.type SpuSetTransferMode, @function
-SpuSetTransferMode:
-	la		$v0, _spu_transfer_mode
-	sb		$a0, 0($v0)
-	jr		$ra
-	move	$v0, $a0
-	
-	
-.global SpuSetTransferStartAddr
-.type SpuSetTransferStartAddr, @function
-SpuSetTransferStartAddr:
-	li		$v0, 0x1000				# Check if value is valid
-	blt		$a0, $v0, .Lbad_value
-	nop
-	lui		$v0, 8					# 0x7ffff = (8<<16)-1
-	addiu	$v0, -1
-	bgt		$a0, $v0, .Lbad_value
-	nop
-	
-	la		$v1, _spu_transfer_addr
-	srl		$v0, $a0, 3					# Set transfer destination address
-	sh		$v0, 0($v1)
-	
-	jr		$ra
-	move	$v0, $a0
-	
-.Lbad_value:
-	jr		$ra
-	move	$v0, $0
-	
-	
-.global SpuWrite
-.type SpuWrite, @function
-SpuWrite:
-	addiu	$sp, -8
-	sw		$ra, 0($sp)
-	sw		$a0, 4($sp)
-	
-	lui		$a3, IOBASE	
-	
-	lhu		$v0, SPU_CTRL($a3)			# Set transfer mode to Stop
-	nop
-	andi	$v0, 0xffcf
-	sh		$v0, SPU_CTRL($a3)
-	jal		SpuCtrlSync
-	move	$a0, $v0
-	
-	la		$v1, _spu_transfer_addr		# Set SPU write address
-	lhu		$v1, 0($v1)
-	nop
-	sh		$v1, SPU_ADDR($a3)
-	
-	lhu		$v0, SPU_CTRL($a3)			# Set transfer mode to DMA write
-	nop
-	ori		$v0, 0x20
-	sh		$v0, SPU_CTRL($a3)
-	#jal		SpuCtrlSync				# Locks up on most emulators (bit 5 in
-	#move	$a0, $v0					# SPUSTAT likely not updating, seems to
-										# be okay to not wait for it on real HW)
-	
-	lw		$a0, 4($sp)
-	
-.Ldma_wait:								# Wait for SPU to be ready for DMA
-	lhu		$v0, SPU_STAT($a3)
-	nop
-	andi	$v0, 0x400					# Bit 8 in SPUSTAT never changes to 1 on
-	bnez	$v0, .Ldma_wait				# emulators so use bit 10 instead
-	nop
-	
-	sw		$a0, DMA4_MADR($a3)			# Set DMA source address
-	
-	li		$v0, 0x10					# 16 words per block (64 bytes)
-	addiu	$a1, 63						# Add by 63 to ensure all bytes get sent
-	srl		$a1, 6						# Equivalent to divide by 64
-	andi	$a1, 0xffff
-	sll		$a1, 16
-	or		$v0, $a1
-	sw		$v0, DMA4_BCR($a3)
-	
-	lui		$v0, 0x0100					# Commence transfer
-	ori		$v0, 0x0201
-	sw		$v0, DMA4_CHCR($a3)
-	
-	lw		$ra, 0($sp)
-	addiu	$sp, 8
-	jr		$ra
-	nop
-	
-	
-.section .data
-
-.global _spu_transfer_mode
-.type _spu_transfer_mode, @object
-_spu_transfer_mode:
-	.word	0x0
-	
-.global _spu_transfer_addr
-.type _spu_transfer_addr, @object
-_spu_transfer_addr:
-	.word	0x200
-\ No newline at end of file
author	John "Lameguy" Wilbert Villamor <lameguy64@gmail.com>	2022-09-26 16:49:56 +0800
committer	GitHub <noreply@github.com>	2022-09-26 16:49:56 +0800
commit	c4a2533d21dfd05cde841ea48c67b05e0e6a853f (patch)
tree	c7ef61653b157b69fb0956709366996ddbc4ecfa /libpsn00b
parent	a8b404b3400c3ebd8e0b923dcaefcc49ea563e36 (diff)
parent	86f0064afb8200e60dd80827535cac30d0eab028 (diff)
download	psn00bsdk-c4a2533d21dfd05cde841ea48c67b05e0e6a853f.tar.gz