Refactor libpsxgpu in C, add OT drawing queue

32 files changed, 722 insertions, 1661 deletions

diff --git a/examples/demos/n00bdemo/disp.c b/examples/demos/n00bdemo/disp.c
index c134163..d8d2bbf 100644
--- a/examples/demos/n00bdemo/disp.c
+++ b/examples/demos/n00bdemo/disp.c
@@ -19,7 +19,7 @@ MATRIX mtx;
 
 void initDisplay() {
 	
-	ResetGraph( 3 );
+	ResetGraph( 0 );
 	
 	if( GetVideoMode() == MODE_NTSC ) {
 		SetDefDispEnv( &disp, 0, 0, 640, 480 );
diff --git a/examples/sound/vagsample/main.c b/examples/sound/vagsample/main.c
index 2b04c34..5764541 100644
--- a/examples/sound/vagsample/main.c
+++ b/examples/sound/vagsample/main.c
@@ -127,7 +127,7 @@ void init(void)
 	// Begin pad polling
 	InitPAD( pad_buff[0], 34, pad_buff[1], 34 );
 	StartPAD();
-	
+	ChangeClearPAD(0);
 } /* init */
 
 // Display function
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/include/psxgpu.h b/libpsn00b/include/psxgpu.h
index f061219..53a24c9 100644
--- a/libpsn00b/include/psxgpu.h
+++ b/libpsn00b/include/psxgpu.h
@@ -1,6 +1,7 @@
 #ifndef __PSXGPU_H
 #define __PSXGPU_H
 
+#include <stddef.h>
 #include <sys/types.h>
 
 // Low-level display parameters for DISPENV_RAW. A leftover from prototyping
@@ -17,10 +18,11 @@
 #define DISP_MODE_NTSC		0
 #define DISP_MODE_PAL		8
 
-
-#define MODE_NTSC			0
-#define MODE_PAL			1
-
+typedef enum _VIDEO_MODE
+{
+	MODE_NTSC	= 0,
+	MODE_PAL	= 1
+} VIDEO_MODE;
 
 // Vector macros
 
@@ -89,6 +91,12 @@
 #define setWH( p, _w, _h ) \
 	(p)->w = _w, (p)->h = _h
 
+#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)
+
 /*
  *	Set texture coordinates
  */
@@ -569,50 +577,61 @@ typedef struct _TIM_IMAGE
 	u_long	*paddr;
 } TIM_IMAGE;
 
+typedef struct _GsIMAGE
+{
+	u_long	pmode;
+	short	px, py, pw, ph;
+	u_long	*pixel;
+	short	cx, cy, cw, ch;
+	u_long	*clut;
+} GsIMAGE;
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-// Function definitions (asm)
+// Function definitions
 
 void ResetGraph(int mode);
 
-int GetVideoMode(void);
-void SetVideoMode(int mode);
+VIDEO_MODE GetVideoMode(void);
+void SetVideoMode(VIDEO_MODE mode);
 
 int GetODE(void);
 
-void PutDispEnvRaw(DISPENV_RAW *disp);	/* obsolete */
-void PutDispEnv(DISPENV *disp);
-void PutDrawEnv(DRAWENV *draw);
+void PutDispEnvRaw(const DISPENV_RAW *env);	/* obsolete */
+void PutDispEnv(const DISPENV *env);
+void PutDrawEnv(DRAWENV *env);
+void PutDrawEnvFast(DRAWENV *env);
 
 void SetDispMask(int mask);
 
-int VSync(int m);
-int DrawSync(int m);
-void WaitGPUcmd(void);
-void WaitGPUdma(void);
+int VSync(int mode);
+int DrawSync(int mode);
+//void WaitGPUcmd(void);
+//void WaitGPUdma(void);
 
 // Callback hook functions
 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 LoadImage(const RECT *rect, const u_long *data);
+void StoreImage(const RECT *rect, u_long *data);
 
-void AddPrim(u_long* ot, void* pri);
+void ClearOTagR(u_long *ot, size_t length);
+void ClearOTag(u_long *ot, size_t length);
+void DrawOTag(const u_long *ot);
+void DrawOTag2(const u_long *ot);
+void DrawOTagEnv(const u_long *ot, DRAWENV *env);
+void DrawPrim(const u_long *pri);
 
-// Function definitions (C)
+void AddPrim(u_long *ot, const void *pri);
 
-int GetTimInfo(const u_long *tim, TIM_IMAGE *timimg);	/* ORIGINAL */
+int GsGetTimInfo(const u_long *tim, GsIMAGE *info);
+int GetTimInfo(const u_long *tim, TIM_IMAGE *info);	/* deprecated */
 
-DISPENV *SetDefDispEnv(DISPENV *disp, int x, int y, int w, int h);
-DRAWENV *SetDefDrawEnv(DRAWENV *draw, int x, int y, int w, int h);
+DISPENV *SetDefDispEnv(DISPENV *env, int x, int y, int w, int h);
+DRAWENV *SetDefDrawEnv(DRAWENV *env, int x, int y, int w, int h);
 
 // Debug font functions
 
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..ee90225
--- /dev/null
+++ b/libpsn00b/psxgpu/common.c
@@ -0,0 +1,293 @@
+/*
+ * PSn00bSDK GPU library (common functions)
+ * (C) 2022 spicyjpeg - MPL licensed
+ */
+
+#include <stdint.h>
+#include <sys/types.h>
+#include <stdio.h>
+#include <psxetc.h>
+#include <psxapi.h>
+#include <psxgpu.h>
+#include <hwregs_c.h>
+
+#define QUEUE_LENGTH		8
+#define DMA_CHUNK_LENGTH	8
+#define VSYNC_TIMEOUT		0x100000
+
+/* Internal globals */
+
+VIDEO_MODE _gpu_video_mode;
+
+static void (*_vsync_callback)(void);
+static void (*_drawsync_callback)(void);
+
+static const u_long *volatile _draw_queue[QUEUE_LENGTH];
+static volatile uint8_t  _queue_head, _queue_tail, _queue_length;
+static volatile uint32_t _vblank_counter, _last_hblank;
+
+/* Interrupt handlers */
+
+static void _vblank_handler(void) {
+	_vblank_counter++;
+
+	if (_vsync_callback)
+		_vsync_callback();
+}
+
+static void _gpu_dma_handler(void) {
+	//while (DMA_CHCR(2) & (1 << 24))
+		//__asm__ volatile("");
+	while (!(GPU_GP1 & (1 << 28)))
+		__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.
+	static int setup_done = 0;
+	if (!setup_done) {
+		EnterCriticalSection();
+
+		DMA_DPCR = 0x03333333;
+		DMA_DICR = 0;
+		IRQ_MASK = 0;
+
+		InterruptCallback(0, &_vblank_handler);
+		DMACallback(2, &_gpu_dma_handler);
+		RestartCallback();
+		_96_remove();
+
+		_gpu_video_mode = (GPU_GP1 >> 20) & 1;
+		setup_done      = 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 _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);
+	return;
+}
+
+int VSync(int mode) {
+	if (mode < 0)
+		return _vblank_counter;
+
+	uint32_t status = GPU_GP1;
+
+	// If mode = 0, wait for one vertical blank event to occur. If mode = 1, do
+	// not wait.
+	for (int i = ((mode < 2) ? (mode ^ 1) : mode); i; i--) {
+		_vsync_halt();
+
+		// 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("");
+		}
+	}
+
+	// Update the horizontal blank counter and return the time elapsed since
+	// the last time it was updated.
+	uint16_t counter = TIMER_VALUE(1);
+	uint16_t delta   = counter - _last_hblank;
+
+	_last_hblank = counter;
+	return delta;
+}
+
+int DrawSync(int mode) {
+	if (mode)
+		return (DMA_BCR(2) >> 16);
+
+	// Wait for the queue to become empty, to make sure no .
+	// 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 (DMA_CHCR(2) & (1 << 24))
+			__asm__ volatile("");
+		while (!(GPU_GP1 & (1 << 28)))
+			__asm__ volatile("");
+	}
+
+	while (!(GPU_GP1 & (1 << 26)))
+		__asm__ volatile("");
+
+	return 0;
+}
+
+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(u_long *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(u_long *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] = (u_long) &ot[i + 1] & 0x00ffffff;
+
+	ot[length - 1] = 0x00ffffff;
+}
+
+void DrawOTag(const u_long *ot) {
+	// If GPU DMA is currently busy, append the OT to the queue instead of
+	// drawing it immediately.
+	if (DMA_CHCR(2) & (1 << 24)) {
+		if (_queue_length >= QUEUE_LENGTH) {
+			printf("psxgpu: DrawOTag() failed, draw queue full\n");
+			return;
+		}
+
+		_draw_queue[_queue_tail++] = ot;
+		_queue_length++;
+		_queue_tail %= QUEUE_LENGTH;
+		return;
+	}
+
+	DrawOTag2(ot);
+}
+
+void DrawOTag2(const u_long *ot) {
+	GPU_GP1 = 0x04000002;
+
+	while (!(GPU_GP1 & (1 << 26)))
+		__asm__ volatile("");
+
+	DMA_MADR(2) = (uint32_t) ot;
+	DMA_BCR(2)  = 0;
+	DMA_CHCR(2) = 0x01000401;
+}
+
+void DrawPrim(const u_long *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(u_long *ot, const void *pri) {
+	addPrim(ot, pri);
+}
+
+/* Misc. functions */
+
+VIDEO_MODE GetVideoMode(void) {
+	return _gpu_video_mode;
+}
+
+void SetVideoMode(VIDEO_MODE 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..8a74b7e
--- /dev/null
+++ b/libpsn00b/psxgpu/env.c
@@ -0,0 +1,213 @@
+/*
+ * PSn00bSDK GPU library (DRAWENV/DISPENV functions)
+ * (C) 2022 spicyjpeg - MPL licensed
+ */
+
+#include <stdint.h>
+#include <sys/types.h>
+#include <psxgpu.h>
+#include <hwregs_c.h>
+
+#define _min(x, y) (((x) < (y)) ? (x) : (y))
+
+extern VIDEO_MODE _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 u_long *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(prim);
+}
+
+void PutDrawEnv(DRAWENV *env) {
+	DrawOTagEnv((const u_long *) 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 u_long *) 0x00ffffff, env);
+		return;
+	}
+
+	DrawOTag(&(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..7a8137c 100644
--- a/libpsn00b/psxgpu/font.c
+++ b/libpsn00b/psxgpu/font.c
@@ -21,14 +21,14 @@ static int _nstreams = 0;
 u_short _font_tpage;
 u_short _font_clut;
 
-extern u_char dbugfont[];
+extern u_char _gpu_debug_font[];
 
 void FntLoad(int x, int y) {
 
 	RECT pos;
 	TIM_IMAGE tim;
 	
-	GetTimInfo( (u_long*)dbugfont, &tim );
+	GetTimInfo( (const u_long *) _gpu_debug_font, &tim );
 	
 	// Load font image
 	pos = *tim.prect;
@@ -223,4 +223,45 @@ char *FntFlush(int id) {
 	
 	return _stream[id].pribuff;
 	
-}
\ No newline at end of file
+}
+
+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;
+	
+}
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..3633c7c
--- /dev/null
+++ b/libpsn00b/psxgpu/image.c
@@ -0,0 +1,116 @@
+/*
+ * PSn00bSDK GPU library (image and VRAM transfer functions)
+ * (C) 2022 spicyjpeg - MPL licensed
+ */
+
+#include <stdint.h>
+#include <sys/types.h>
+#include <stdio.h>
+#include <psxgpu.h>
+#include <hwregs_c.h>
+
+#define DMA_CHUNK_LENGTH 8
+
+/* Common internal load/store function */
+
+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\n", length, DMA_CHUNK_LENGTH);
+
+	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);
+}
+
+/* Public VRAM API */
+
+void LoadImage(const RECT *rect, const u_long *data) {
+	_load_store_image(0xa0000000, 2, rect, (uint32_t *) data);
+}
+
+void StoreImage(const RECT *rect, u_long *data) {
+	_load_store_image(0xc0000000, 3, rect, (uint32_t *) 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 u_long *tim, GsIMAGE *info) {
+	if ((*(tim++) & 0xffff) != 0x0010)
+		return 1;
+
+	info->pmode = *(tim++);
+	if (info->pmode & 8) {
+		const u_long *palette_end = tim;
+		palette_end += *(tim++) / 4;
+
+		*((u_long *) &(info->cx)) = *(tim++);
+		*((u_long *) &(info->cw)) = *(tim++);
+		info->clut = (u_long *) tim;
+
+		tim = palette_end;
+	} else {
+		info->clut = 0;
+	}
+
+	tim++;
+	*((u_long *) &(info->px)) = *(tim++);
+	*((u_long *) &(info->pw)) = *(tim++);
+	info->pixel = (u_long *) tim;
+
+	return 0;
+}
+
+int GetTimInfo(const u_long *tim, TIM_IMAGE *info) {
+	if ((*(tim++) & 0xffff) != 0x0010)
+		return 1;
+
+	info->mode = *(tim++);
+	if (info->mode & 8) {
+		const u_long *palette_end = tim;
+		palette_end += *(tim++) / 4;
+
+		info->crect = (RECT *)   tim;
+		info->caddr = (u_long *) &tim[2];
+
+		tim = palette_end;
+	} else {
+		info->caddr = 0;
+	}
+
+	tim++;
+	info->prect = (RECT *)   tim;
+	info->paddr = (u_long *) &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..2963b6a 100644
--- a/libpsn00b/psxgpu/readme.txt
+++ b/libpsn00b/psxgpu/readme.txt
@@ -3,25 +3,19 @@ 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
-