From 6ee55c23b042a1559e8cabfccf3b9d3320c4c5cc Mon Sep 17 00:00:00 2001
From: spicyjpeg <thatspicyjpeg@gmail.com>
Date: Thu, 6 Oct 2022 12:15:24 +0200
Subject: Replace psxgpu debug font, add CdGetSector2()

---
 libpsn00b/include/psxcd.h | 1 +
 1 file changed, 1 insertion(+)

(limited to 'libpsn00b/include')

diff --git a/libpsn00b/include/psxcd.h b/libpsn00b/include/psxcd.h
index 03ee792..cf9ecad 100644
--- a/libpsn00b/include/psxcd.h
+++ b/libpsn00b/include/psxcd.h
@@ -145,6 +145,7 @@ uint32_t CdSyncCallback(CdlCB func);
 
 long CdReadyCallback(CdlCB func);
 int CdGetSector(void *madr, int size);
+int CdGetSector2(void *madr, int size);
 
 CdlFILE* CdSearchFile(CdlFILE *loc, const char *filename);
 
-- 
cgit v1.2.3


From 82a259240d9c63d4656b9dae0b46a3689840473b Mon Sep 17 00:00:00 2001
From: spicyjpeg <thatspicyjpeg@gmail.com>
Date: Sat, 8 Oct 2022 12:44:34 +0200
Subject: Optimize memset(), add heap usage API, remove _mem_init()

---
 libpsn00b/include/stdlib.h |  15 +++++-
 libpsn00b/libc/malloc.c    |  62 ++++++++++++++++++---
 libpsn00b/libc/memset.s    | 132 ++++++++++++++++++++++++++++++++++++++-------
 libpsn00b/libc/start.c     |  25 ++-------
 4 files changed, 184 insertions(+), 50 deletions(-)

(limited to 'libpsn00b/include')

diff --git a/libpsn00b/include/stdlib.h b/libpsn00b/include/stdlib.h
index 1888c69..f0753c1 100644
--- a/libpsn00b/include/stdlib.h
+++ b/libpsn00b/include/stdlib.h
@@ -12,6 +12,16 @@
 
 #define RAND_MAX 0x7fff
 
+/* Structure definitions */
+
+typedef struct _HeapUsage {
+	size_t total;		// Total size of heap + stack
+	size_t heap;		// Amount of memory currently reserved for heap
+	size_t stack;		// Amount of memory currently reserved for stack
+	size_t alloc;		// Amount of memory currently allocated
+	size_t alloc_max;	// Maximum amount of memory ever allocated
+} HeapUsage;
+
 /* API */
 
 #ifdef __cplusplus
@@ -33,11 +43,12 @@ long double strtold(const char *nptr, char **endptr);
 double strtod(const char *nptr, char **endptr);
 float strtof(const char *nptr, char **endptr);
 
-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 TrackHeapUsage(ptrdiff_t alloc_incr);
+void GetHeapUsage(HeapUsage *usage);
+
 void *malloc(size_t size);
 void *calloc(size_t num, size_t size);
 void *realloc(void *ptr, size_t size);
diff --git a/libpsn00b/libc/malloc.c b/libpsn00b/libc/malloc.c
index 9d538cd..acac753 100644
--- a/libpsn00b/libc/malloc.c
+++ b/libpsn00b/libc/malloc.c
@@ -9,6 +9,8 @@
  * 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.
+ * Custom allocators should call TrackHeapUsage() to let the heap manager know
+ * how much memory is allocated at a given time.
  */
 
 #include <stddef.h>
@@ -25,11 +27,13 @@ typedef struct _BlockHeader {
 	size_t				size;
 } BlockHeader;
 
-/* Data */
+/* Internal globals */
 
 static void			*_heap_start, *_heap_end, *_heap_limit;
-static void			*_alloc_start = 0;
-static BlockHeader	*_alloc_head = 0, *_alloc_tail = 0;
+static size_t		_heap_alloc, _heap_alloc_max;
+
+static void			*_alloc_start;
+static BlockHeader	*_alloc_head, *_alloc_tail;
 
 /* Heap management API */
 
@@ -37,6 +41,13 @@ __attribute__((weak)) void InitHeap(void *addr, size_t size) {
 	_heap_start = addr;
 	_heap_end   = addr;
 	_heap_limit = (void *) ((uintptr_t) addr + size);
+
+	_heap_alloc     = 0;
+	_heap_alloc_max = 0;
+
+	_alloc_start = addr;
+	_alloc_head  = 0;
+	_alloc_tail  = 0;
 }
 
 __attribute__((weak)) void *sbrk(ptrdiff_t incr) {
@@ -50,6 +61,22 @@ __attribute__((weak)) void *sbrk(ptrdiff_t incr) {
 	return old_end;
 }
 
+__attribute__((weak)) void TrackHeapUsage(ptrdiff_t alloc_incr) {
+	_heap_alloc += alloc_incr;
+
+	if (_heap_alloc > _heap_alloc_max)
+		_heap_alloc_max = _heap_alloc;
+}
+
+__attribute__((weak)) void GetHeapUsage(HeapUsage *usage) {
+	usage->total = _heap_limit - _heap_start;
+	usage->heap  = _heap_end   - _heap_start;
+	usage->stack = _heap_limit - _heap_end;
+
+	usage->alloc     = _heap_alloc;
+	usage->alloc_max = _heap_alloc_max;
+}
+
 /* Memory allocator */
 
 static BlockHeader *_find_fit(BlockHeader *head, size_t size) {
@@ -69,13 +96,16 @@ static BlockHeader *_find_fit(BlockHeader *head, size_t size) {
 }
 
 __attribute__((weak)) void *malloc(size_t size) {
+	if (!size)
+		return 0;
+
 	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);
+		//if (!_alloc_start)
+			//_alloc_start = sbrk(0);
 
 		BlockHeader *new = (BlockHeader *) sbrk(_size);
 		if (!new)
@@ -89,6 +119,8 @@ __attribute__((weak)) void *malloc(size_t size) {
 
 		_alloc_head = new;
 		_alloc_tail = new;
+
+		TrackHeapUsage(_size);
 		return ptr;
 	}
 
@@ -106,6 +138,8 @@ __attribute__((weak)) void *malloc(size_t size) {
 
 		_alloc_head->prev = new;
 		_alloc_head       = new;
+
+		TrackHeapUsage(_size);
 		return ptr;
 	}
 
@@ -122,6 +156,8 @@ __attribute__((weak)) void *malloc(size_t size) {
 
 		(new->next)->prev = new;
 		prev->next        = new;
+
+		TrackHeapUsage(_size);
 		return ptr;
 	}
 
@@ -138,6 +174,8 @@ __attribute__((weak)) void *malloc(size_t size) {
 
 	_alloc_tail->next = new;
 	_alloc_tail       = new;
+
+	TrackHeapUsage(_size);
 	return ptr;
 }
 
@@ -153,13 +191,14 @@ __attribute__((weak)) void *realloc(void *ptr, size_t size) {
 	if (!ptr)
 		return malloc(size);
 
-	size_t _size = _align(size + sizeof(BlockHeader), 8);
-
+	size_t      _size = _align(size + sizeof(BlockHeader), 8);
 	BlockHeader *prev = (BlockHeader *) ((uintptr_t) ptr - sizeof(BlockHeader));
 
 	// New memory block shorter?
 	if (prev->size >= _size) {
+		TrackHeapUsage(_size - prev->size);
 		prev->size = _size;
+
 		if (!prev->next)
 			sbrk((ptr - sbrk(0)) + _size);
 
@@ -172,12 +211,14 @@ __attribute__((weak)) void *realloc(void *ptr, size_t size) {
 		if (!new)
 			return 0;
 
+		TrackHeapUsage(_size - prev->size);
 		prev->size = _size;
 		return ptr;
 	}
 
 	// Do we have free memory after it?
 	if (((prev->next)->ptr - ptr) > _size) {
+		TrackHeapUsage(_size - prev->size);
 		prev->size = _size;
 		return ptr;
 	}
@@ -209,11 +250,13 @@ __attribute__((weak)) void free(void *ptr) {
 			sbrk(-size);
 		}
 
+		TrackHeapUsage(-size);
 		return;
 	}
 
 	// Finding the proper block
 	BlockHeader *cur = _alloc_head;
+
 	for (cur = _alloc_head; ptr != cur->ptr; cur = cur->next) {
 		if (!cur->next)
 			return;
@@ -221,14 +264,17 @@ __attribute__((weak)) void free(void *ptr) {
 
 	if (cur->next) {
 		// In the middle, just unlink it
-		cur->next->prev = cur->prev;
+		(cur->next)->prev = cur->prev;
+		TrackHeapUsage(-(cur->size + sizeof(BlockHeader)));
 	} 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);
+		TrackHeapUsage(-size);
 	}
 
 	(cur->prev)->next = cur->next;
diff --git a/libpsn00b/libc/memset.s b/libpsn00b/libc/memset.s
index b3a3af3..5a1589d 100644
--- a/libpsn00b/libc/memset.s
+++ b/libpsn00b/libc/memset.s
@@ -1,25 +1,117 @@
-# High speed ASM memset implementation by Lameguy64
-#
-# Part of PSn00bSDK
+# PSn00bSDK optimized memset
+# (C) 2022 spicyjpeg - MPL licensed
 
 .set noreorder
 
-.section .text
-
-# Arguments:
-#	a0 - address to buffer
-#	a1 - value to set
-#	a2 - bytes to set
+.section .text.memset
 .global memset
-.type memset,@function
+.type memset, @function
 memset:
-	move	$v0, $a0
-	blez	$a2, .Lexit
-	addi	$a2, -1
-	sb		$a1, 0($a0)
-	b		memset
-	addiu	$a0, 1
-.Lexit:
-	jr		$ra
-	nop
-	
\ No newline at end of file
+	# If more than 16 bytes have to be written then take the "large" path,
+	# otherwise use the code below.
+	addiu $t0, $a2, -16
+	bgtz  $t0, .Llarge_fill
+	move  $v0, $a0 # return_value = dest
+
+	# Jump to one of the sb opcodes below. This is basically a cut-down Duff's
+	# device implementation with no looping.
+	la    $t0, .Lsmall_duff + 0x40 # jump_addr = &small_duff[(16 - count) * 4]
+	sll   $t1, $a2, 2
+	subu  $t0, $t1
+	addu  $a0, $a2 # dest -= 16 - count
+	jr    $t0
+	addiu $a0, -16
+
+.Lsmall_duff:
+	sb    $a1, 0x0($a0)
+	sb    $a1, 0x1($a0)
+	sb    $a1, 0x2($a0)
+	sb    $a1, 0x3($a0)
+	sb    $a1, 0x4($a0)
+	sb    $a1, 0x5($a0)
+	sb    $a1, 0x6($a0)
+	sb    $a1, 0x7($a0)
+	sb    $a1, 0x8($a0)
+	sb    $a1, 0x9($a0)
+	sb    $a1, 0xa($a0)
+	sb    $a1, 0xb($a0)
+	sb    $a1, 0xc($a0)
+	sb    $a1, 0xd($a0)
+	sb    $a1, 0xe($a0)
+	jr    $ra
+	sb    $a1, 0xf($a0)
+
+.Llarge_fill:
+	# Initialize fast filling by repeating the fill byte 4 times, so it can be
+	# written 32 bits at a time.
+	andi  $a1, 0xff # ch &= 0xff
+	sll   $t0, $a1, 8 # ch |= (ch << 8) | (ch << 16) | (ch << 24)
+	or    $a1, $t0
+	sll   $t0, $a1, 16
+	or    $a1, $t0
+
+	# Fill the first 1-4 bytes (here the swr instruction does all the magic)
+	# and update dest and count accordingly.
+	swr   $a1, 0($a0)
+	andi  $t0, $a0, 3 # align = 4 - (dest % 4)
+	addiu $t0, -4
+	addu  $a2, $t0 # count -= align
+	subu  $a0, $t0 # dest += align
+
+	la    $t1, .Llarge_duff
+	andi  $t2, $a2, 3 # remainder = count % 4
+	subu  $a2, $t2 # count -= remainder
+
+.Llarge_fill_loop:
+	# If 128 bytes or more still have to be written, skip calculating the jump
+	# offset and execute the whole block of sw opcodes.
+	addiu $a2, -0x80 # count -= 0x80
+	bgez  $a2, .Llarge_duff
+	#nop
+
+	# Jump to one of the sw opcodes below. This is the "full" Duff's device.
+	subu  $t0, $t1, $a2 # jump_addr = &large_duff[0x80 - (count + 0x80)]
+	jr    $t0
+	addu  $a0, $a2 # dest -= 0x80 - (count + 0x80)
+
+.Llarge_duff:
+	sw    $a1, 0x00($a0)
+	sw    $a1, 0x04($a0)
+	sw    $a1, 0x08($a0)
+	sw    $a1, 0x0c($a0)
+	sw    $a1, 0x10($a0)
+	sw    $a1, 0x14($a0)
+	sw    $a1, 0x18($a0)
+	sw    $a1, 0x1c($a0)
+	sw    $a1, 0x20($a0)
+	sw    $a1, 0x24($a0)
+	sw    $a1, 0x28($a0)
+	sw    $a1, 0x2c($a0)
+	sw    $a1, 0x30($a0)
+	sw    $a1, 0x34($a0)
+	sw    $a1, 0x38($a0)
+	sw    $a1, 0x3c($a0)
+	sw    $a1, 0x40($a0)
+	sw    $a1, 0x44($a0)
+	sw    $a1, 0x48($a0)
+	sw    $a1, 0x4c($a0)
+	sw    $a1, 0x50($a0)
+	sw    $a1, 0x54($a0)
+	sw    $a1, 0x58($a0)
+	sw    $a1, 0x5c($a0)
+	sw    $a1, 0x60($a0)
+	sw    $a1, 0x64($a0)
+	sw    $a1, 0x68($a0)
+	sw    $a1, 0x6c($a0)
+	sw    $a1, 0x70($a0)
+	sw    $a1, 0x74($a0)
+	sw    $a1, 0x78($a0)
+	sw    $a1, 0x7c($a0)
+
+	bgtz  $a2, .Llarge_fill_loop
+	addiu $a0, 0x80 # dest += 0x80
+
+	# Fill the remaining 1-4 bytes, using (again) an unaligned store.
+	addu  $a0, $t2 # last_byte = dest + remainder - 1
+	jr    $ra
+	swl   $a1, -1($a0)
diff --git a/libpsn00b/libc/start.c b/libpsn00b/libc/start.c
index 87ac951..9ff09c8 100644
--- a/libpsn00b/libc/start.c
+++ b/libpsn00b/libc/start.c
@@ -21,7 +21,7 @@ const char	**__argv;
 static const char	*_argv_buffer[ARGC_MAX];
 static char			_arg_string_buffer[132];
 
-static void _parse_kernel_args() {
+static void _parse_kernel_args(void) {
 	// Copy the argument string from kernel memory into a private buffer (which
 	// won't be cleared or deallocated) and trim it at the first newline.
 	memset(_arg_string_buffer, 0, 132);
@@ -48,7 +48,7 @@ static void _parse_kernel_args() {
 	}
 }
 
-/* Heap initialization */
+/* Main */
 
 // These are defined by the linker script. Note that these are *NOT* pointers,
 // they are virtual symbols whose location matches their value. The simplest
@@ -58,20 +58,6 @@ extern uint8_t __bss_start[];
 extern uint8_t _end[];
 //extern uint8_t _gp[];
 
-// This function should not be called manually in most cases. It might be
-// 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(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);
-}
-
-/* Main */
-
 extern void (*__CTOR_LIST__[])(void);
 extern void (*__DTOR_LIST__[])(void);
 
@@ -88,10 +74,9 @@ void _start_inner(int32_t override_argc, const char **override_argv) {
 	for (uint32_t *i = (uint32_t *) __bss_start; i < (uint32_t *) _end; i++)
 		*i = 0;
 
-	// Initialize the heap, assuming 2 MB of RAM and reserving 128 KB for the
-	// stack. Note that _mem_init() can be called again in main() to change
-	// these values.
-	_mem_init(0x200000, 0x20000);
+	// Initialize the heap and place it after the executable, assuming 2 MB of
+	// RAM. Note that InitHeap() can be called again in main().
+	InitHeap((void *) _end + 4, (void *) 0x801ffff8 - (void *) _end);
 
 	if (override_argv) {
 		__argc = override_argc;
-- 
cgit v1.2.3


From e0e608855b9b2e83a92047294e1f5a242ff51e88 Mon Sep 17 00:00:00 2001
From: spicyjpeg <thatspicyjpeg@gmail.com>
Date: Tue, 11 Oct 2022 11:46:24 +0200
Subject: Replace DEBUG macro with standard NDEBUG macro

---
 libpsn00b/cmake/flags.cmake |  2 +-
 libpsn00b/include/assert.h  |  6 +++---
 libpsn00b/libc/abort.c      |  6 +++---
 libpsn00b/psxcd/isofs.c     | 15 ++++++---------
 libpsn00b/psxcd/psxcd.c     | 32 ++++++++++++++++----------------
 libpsn00b/psxetc/dl.c       |  6 +++---
 libpsn00b/psxgpu/common.c   |  6 +++---
 libpsn00b/psxgpu/image.c    |  6 +++---
 libpsn00b/psxpress/mdec.c   |  6 +++---
 libpsn00b/psxspu/common.c   |  6 +++---
 10 files changed, 44 insertions(+), 47 deletions(-)

(limited to 'libpsn00b/include')

diff --git a/libpsn00b/cmake/flags.cmake b/libpsn00b/cmake/flags.cmake
index e56d3fc..1ca25c2 100644
--- a/libpsn00b/cmake/flags.cmake
+++ b/libpsn00b/cmake/flags.cmake
@@ -86,5 +86,5 @@ target_link_options(
 target_compile_definitions(
 	psn00bsdk INTERFACE
 		PSN00BSDK=1
-		$<$<CONFIG:DEBUG>:DEBUG=1>
+		$<$<CONFIG:Release>:NDEBUG=1>
 )
diff --git a/libpsn00b/include/assert.h b/libpsn00b/include/assert.h
index e27f2ed..32301e2 100644
--- a/libpsn00b/include/assert.h
+++ b/libpsn00b/include/assert.h
@@ -8,13 +8,13 @@
 
 void _assert_abort(const char *file, int line, const char *expr);
 
-#ifdef DEBUG
+#ifdef NDEBUG
+#define assert(x)
+#else
 #define assert(expr) { \
 	if (!(expr)) \
 		_assert_abort(__FILE__, __LINE__, #expr); \
 }
-#else
-#define assert(x)
 #endif
 
 #endif
diff --git a/libpsn00b/libc/abort.c b/libpsn00b/libc/abort.c
index 562b619..1108160 100644
--- a/libpsn00b/libc/abort.c
+++ b/libpsn00b/libc/abort.c
@@ -1,10 +1,10 @@
 
 #include <stdio.h>
 
-#ifdef DEBUG
-#define _LOG(...) printf(__VA_ARGS__)
-#else
+#ifdef NDEBUG
 #define _LOG(...)
+#else
+#define _LOG(...) printf(__VA_ARGS__)
 #endif
 
 /* Standard abort */
diff --git a/libpsn00b/psxcd/isofs.c b/libpsn00b/psxcd/isofs.c
index 6755d3f..fb1f6c9 100644
--- a/libpsn00b/psxcd/isofs.c
+++ b/libpsn00b/psxcd/isofs.c
@@ -7,13 +7,10 @@
 #include "psxcd.h"
 #include "isofs.h"
 
-// Uncommend to enable debug output
-//#define DEBUG
-
-#ifdef DEBUG
-#define _LOG(...) printf(__VA_ARGS__)
-#else
+#ifdef NDEBUG
 #define _LOG(...)
+#else
+#define _LOG(...) printf(__VA_ARGS__)
 #endif
 
 #define DEFAULT_PATH_SEP	'\\'
@@ -215,7 +212,7 @@ static int _CdReadIsoDirectory(int lba)
 	return 0;
 }
 
-#ifdef DEBUG
+#ifndef NDEBUG
 
 static void dump_directory(void)
 {
@@ -479,7 +476,7 @@ CdlFILE *CdSearchFile(CdlFILE *fp, const char *filename)
 	// Get number of directories in path table
 	num_dirs = get_pathtable_entry(0, NULL, NULL);
 	
-#ifdef DEBUG
+#ifndef NDEBUG
 	_LOG("psxcd: Directories in path table: %d\n", num_dirs);
 	
 	rbuff = resolve_pathtable_path(num_dirs-1, tpath_rbuff+127);
@@ -536,7 +533,7 @@ CdlFILE *CdSearchFile(CdlFILE *fp, const char *filename)
 		strcat(fp->name, ";1");
 	}
 	
-#ifdef DEBUG
+#ifndef NDEBUG
 	dump_directory();
 #endif
 	
diff --git a/libpsn00b/psxcd/psxcd.c b/libpsn00b/psxcd/psxcd.c
index ac93376..65dea48 100644
--- a/libpsn00b/psxcd/psxcd.c
+++ b/libpsn00b/psxcd/psxcd.c
@@ -20,10 +20,10 @@ volatile int _cd_last_sector_count;
 
 int _cd_media_changed;
 
-#ifdef DEBUG
-#define _LOG(...) printf(__VA_ARGS__)
-#else
+#ifdef NDEBUG
 #define _LOG(...)
+#else
+#define _LOG(...) printf(__VA_ARGS__)
 #endif
 
 void _cd_init(void);
@@ -88,7 +88,7 @@ int CdControlB(unsigned char com, const void *param, unsigned char *result)
 int CdControlF(unsigned char com, const void *param)
 {
 	int param_len=0;
-	
+
 	// Command specific parameters
 	switch(com)
 	{
@@ -116,21 +116,21 @@ int CdControlF(unsigned char com, const void *param)
 			break;
 		case CdlGetTD:
 			param_len = 1;
+			break;
+		case CdlReadN:
+		case CdlReadS:
+		case CdlSeekL:
+		case CdlSeekP:
+			if( param )
+			{
+				_cd_control(CdlSetloc, param, 3);
+				_cd_last_setloc = *((CdlLOC*)param);
+			}
 	}
-	
-	// Issue Setloc if parameters are specified on CdlReadN and CdlReadS
-	if( ( com == CdlReadN ) || ( com == CdlReadS ) )
-	{
-		if( param )
-		{
-			_cd_control(CdlSetloc, param, 3);
-			_cd_last_setloc = *((CdlLOC*)param);
-		}
-	}
-	
+
 	// Issue CD command
 	_cd_control(com, param, param_len);
-	
+
 	return 1;
 }
 
diff --git a/libpsn00b/psxetc/dl.c b/libpsn00b/psxetc/dl.c
index 6d37605..3d8f3ab 100644
--- a/libpsn00b/psxetc/dl.c
+++ b/libpsn00b/psxetc/dl.c
@@ -69,10 +69,10 @@ void *(*_dl_resolve_callback)(DLL *, const char *) = 0;
 
 /* Private utilities */
 
-#ifdef DEBUG
-#define _LOG(...) printf(__VA_ARGS__)
-#else
+#ifdef NDEBUG
 #define _LOG(...)
+#else
+#define _LOG(...) printf(__VA_ARGS__)
 #endif
 
 #define _ERROR(code, ret) { \
diff --git a/libpsn00b/psxgpu/common.c b/libpsn00b/psxgpu/common.c
index e06c63a..1e3d9e5 100644
--- a/libpsn00b/psxgpu/common.c
+++ b/libpsn00b/psxgpu/common.c
@@ -31,10 +31,10 @@ static volatile uint16_t _last_hblank;
 
 /* Private utilities and interrupt handlers */
 
-#ifdef DEBUG
-#define _LOG(...) printf(__VA_ARGS__)
-#else
+#ifdef NDEBUG
 #define _LOG(...)
+#else
+#define _LOG(...) printf(__VA_ARGS__)
 #endif
 
 static void _vblank_handler(void) {
diff --git a/libpsn00b/psxgpu/image.c b/libpsn00b/psxgpu/image.c
index 6190c7a..a0d7065 100644
--- a/libpsn00b/psxgpu/image.c
+++ b/libpsn00b/psxgpu/image.c
@@ -12,10 +12,10 @@
 
 /* Private utilities */
 
-#ifdef DEBUG
-#define _LOG(...) printf(__VA_ARGS__)
-#else
+#ifdef NDEBUG
 #define _LOG(...)
+#else
+#define _LOG(...) printf(__VA_ARGS__)
 #endif
 
 static void _load_store_image(
diff --git a/libpsn00b/psxpress/mdec.c b/libpsn00b/psxpress/mdec.c
index 53c596e..d308994 100644
--- a/libpsn00b/psxpress/mdec.c
+++ b/libpsn00b/psxpress/mdec.c
@@ -83,10 +83,10 @@ static const DECDCTENV _default_mdec_env = {
 
 /* Private utilities */
 
-#ifdef DEBUG
-#define _LOG(...) printf(__VA_ARGS__)
-#else
+#ifdef NDEBUG
 #define _LOG(...)
+#else
+#define _LOG(...) printf(__VA_ARGS__)
 #endif
 
 /* Public API */
diff --git a/libpsn00b/psxspu/common.c b/libpsn00b/psxspu/common.c
index 4809684..d6508c7 100644
--- a/libpsn00b/psxspu/common.c
+++ b/libpsn00b/psxspu/common.c
@@ -19,10 +19,10 @@ static uint16_t			_transfer_addr = WRITABLE_AREA_ADDR;
 
 /* Private utilities */
 
-#ifdef DEBUG
-#define _LOG(...) printf(__VA_ARGS__)
-#else
+#ifdef NDEBUG
 #define _LOG(...)
+#else
+#define _LOG(...) printf(__VA_ARGS__)
 #endif
 
 static void _wait_status(uint16_t mask, uint16_t value) {
-- 
cgit v1.2.3


From 03434a230d8c3ed2e32a3885128e05e42ee11769 Mon Sep 17 00:00:00 2001
From: spicyjpeg <thatspicyjpeg@gmail.com>
Date: Sun, 16 Oct 2022 23:52:19 +0200
Subject: Add mdec/strvideo example, fix psxpress bug

---
 examples/mdec/strvideo/CMakeLists.txt |  24 ++
 examples/mdec/strvideo/iso.xml        |  28 +++
 examples/mdec/strvideo/main.c         | 439 ++++++++++++++++++++++++++++++++++
 examples/mdec/strvideo/system.cnf     |   4 +
 libpsn00b/include/psxcd.h             |   7 +-
 libpsn00b/psxcd/getsector.c           |  23 ++
 libpsn00b/psxcd/psxcd.c               |  10 +-
 libpsn00b/psxpress/mdec.c             |   5 +
 libpsn00b/psxpress/vlc.s              |  10 +-
 libpsn00b/psxpress/vlc2.c             |   5 +-
 10 files changed, 539 insertions(+), 16 deletions(-)
 create mode 100644 examples/mdec/strvideo/CMakeLists.txt
 create mode 100644 examples/mdec/strvideo/iso.xml
 create mode 100644 examples/mdec/strvideo/main.c
 create mode 100644 examples/mdec/strvideo/system.cnf

(limited to 'libpsn00b/include')

diff --git a/examples/mdec/strvideo/CMakeLists.txt b/examples/mdec/strvideo/CMakeLists.txt
new file mode 100644
index 0000000..d41556b
--- /dev/null
+++ b/examples/mdec/strvideo/CMakeLists.txt
@@ -0,0 +1,24 @@
+# PSn00bSDK example CMake script
+# (C) 2021 spicyjpeg - MPL licensed
+
+cmake_minimum_required(VERSION 3.21)
+
+project(
+	strvideo
+	LANGUAGES    C
+	VERSION      1.0.0
+	DESCRIPTION  "PSn00bSDK .STR video playback example"
+	HOMEPAGE_URL "http://lameguy64.net/?page=psn00bsdk"
+)
+
+file(GLOB _sources *.c)
+psn00bsdk_add_executable(strvideo GPREL ${_sources})
+#psn00bsdk_add_cd_image(strvideo_iso strvideo iso.xml DEPENDS strvideo)
+
+install(
+	FILES
+		#${PROJECT_BINARY_DIR}/strvideo.bin
+		#${PROJECT_BINARY_DIR}/strvideo.cue
+		${PROJECT_BINARY_DIR}/strvideo.exe
+	TYPE BIN
+)
diff --git a/examples/mdec/strvideo/iso.xml b/examples/mdec/strvideo/iso.xml
new file mode 100644
index 0000000..65e0ff5
--- /dev/null
+++ b/examples/mdec/strvideo/iso.xml
@@ -0,0 +1,28 @@
+<?xml version="1.0" encoding="utf-8"?>
+<iso_project
+	image_name="${CD_IMAGE_NAME}.bin"
+	cue_sheet="${CD_IMAGE_NAME}.cue"
+>
+	<track type="data">
+		<identifiers
+			system			="PLAYSTATION"
+			volume			="STRVIDEO"
+			volume_set		="STRVIDEO"
+			publisher		="MEIDOTEK"
+			data_preparer	="PSN00BSDK ${PSN00BSDK_VERSION}"
+			application		="PLAYSTATION"
+			copyright		="README.TXT;1"
+		/>
+
+		<directory_tree>
+			<file name="SYSTEM.CNF"		type="data" source="${PROJECT_SOURCE_DIR}/system.cnf" />
+			<file name="STRVIDEO.EXE"	type="data" source="strvideo.exe" />
+
+			<file name="VIDEO.STR"		type="mixed" source="${PROJECT_SOURCE_DIR}/video.str"/>
+
+			<dummy sectors="1024"/>
+		</directory_tree>
+	</track>
+
+	<!--<track type="audio" source="track2.wav" />-->
+</iso_project>
diff --git a/examples/mdec/strvideo/main.c b/examples/mdec/strvideo/main.c
new file mode 100644
index 0000000..842bbb8
--- /dev/null
+++ b/examples/mdec/strvideo/main.c
@@ -0,0 +1,439 @@
+/*
+ * PSn00bSDK .STR FMV playback example
+ * (C) 2022 spicyjpeg - MPL licensed
+ *
+ * This example demonstrates playback of full-motion video in the standard .STR
+ * format, using the MDEC for frame decoding and XA for audio. Decoded frames
+ * are transferred directly to the main framebuffer in this example, but could
+ * also be output to another VRAM location and used as a background or texture
+ * for a 2D or 3D scene.
+ *
+ * Playing video files requires setting up a fairly complex pipeline, involving
+ * several buffers and components working in parallel:
+ *
+ * - .STR sectors are read continuously from the CD and each frame, usually
+ *   spanning multiple sectors, is reassembled (demuxed) into a buffer in
+ *   memory. In this example the task is performed by cd_sector_handler(). The
+ *   CD drive handles XA-ADPCM sectors automatically, so no CPU intervention is
+ *   necessary to play the audio track interleaved with the video.
+ * - Once a full frame has been demuxed, the bitstream data is parsed and
+ *   decompressed by the CPU (using DecDCTvlc()) to an array of run-length
+ *   codes to be fed to the MDEC. This is done in the main loop.
+ * - At the same time the last frame decompressed is read from RAM by the MDEC,
+ *   which decodes it and outputs one 16-pixel-wide vertical slice at a time.
+ * - When a slice is ready, it is uploaded by mdec_dma_handler() to the current
+ *   framebuffer in VRAM while the MDEC is decoding the next slice.
+ *   A text overlay is drawn on top of the framebuffer using the GPU after the
+ *   entire frame has been decoded.
+ *
+ * Since pretty much all buffers used are going to be read and written at the
+ * same time, double buffering is required for all of them. Every part of the
+ * pipeline must also run in lockstep with each other to prevent frame
+ * corruption, hence several functions and flag variables are used to stall the
+ * main loop until a frame is available for decoding and the MDEC is ready.
+ * Playback is stopped when a sector with the end-of-file flag set in the XA
+ * subheader (added at the end of the file by most .STR encoders) is
+ * encountered; in order to access the subheader, this example requests 2340
+ * bytes of data for each sector (rather than the usual 2048) from the drive.
+ *
+ * Note that PSn00bSDK's bitstream decoding API only supports version 1 and 2
+ * bitstreams currently, so make sure your .STR files are encoded as v2 and not
+ * v3.
+ */
+
+#include <stdint.h>
+#include <string.h>
+#include <psxetc.h>
+#include <psxapi.h>
+#include <psxgpu.h>
+#include <psxgte.h>
+#include <psxspu.h>
+#include <psxcd.h>
+#include <psxpress.h>
+#include <hwregs_c.h>
+
+// Uncomment to display the video in 24bpp mode. Note that the GPU does not
+// support 24bpp rendering, so the text overlay is only enabled in 16bpp mode.
+//#define DISP_24BPP
+
+/* Display/GPU context utilities */
+
+#define SCREEN_XRES 320
+#define SCREEN_YRES 240
+
+#define BGCOLOR_R 0
+#define BGCOLOR_G 0
+#define BGCOLOR_B 0
+
+typedef struct {
+	DISPENV disp;
+	DRAWENV draw;
+} Framebuffer;
+
+typedef struct {
+	Framebuffer db[2];
+	int         db_active;
+} RenderContext;
+
+void init_context(RenderContext *ctx) {
+	Framebuffer *db;
+
+	ResetGraph(0);
+	ctx->db_active = 0;
+
+	db = &(ctx->db[0]);
+	SetDefDispEnv(&(db->disp), 0,           0, SCREEN_XRES, SCREEN_YRES);
+	SetDefDrawEnv(&(db->draw), 0, SCREEN_YRES, SCREEN_XRES, SCREEN_YRES);
+	setRGB0(&(db->draw), BGCOLOR_R, BGCOLOR_G, BGCOLOR_B);
+	db->draw.isbg    = 1;
+	db->draw.dtd     = 1;
+
+	db = &(ctx->db[1]);
+	SetDefDispEnv(&(db->disp), 0, SCREEN_YRES, SCREEN_XRES, SCREEN_YRES);
+	SetDefDrawEnv(&(db->draw), 0,           0, SCREEN_XRES, SCREEN_YRES);
+	setRGB0(&(db->draw), BGCOLOR_R, BGCOLOR_G, BGCOLOR_B);
+	db->draw.isbg    = 1;
+	db->draw.dtd     = 1;
+
+	PutDrawEnv(&(db->draw));
+	//PutDispEnv(&(db->disp));
+
+	// Create a text stream at the top of the screen.
+	FntLoad(960, 0);
+	FntOpen(4, 12, 312, 16, 2, 256);
+}
+
+void display(RenderContext *ctx, int sync) {
+	Framebuffer *db;
+	ctx->db_active ^= 1;
+
+	DrawSync(0);
+	if (sync)
+		VSync(0);
+
+	db = &(ctx->db[ctx->db_active]);
+	PutDrawEnv(&(db->draw));
+	PutDispEnv(&(db->disp));
+	SetDispMask(1);
+}
+
+/* CD and MDEC interrupt handlers */
+
+#ifdef DISP_24BPP
+#define BLOCK_SIZE 24
+#else
+#define BLOCK_SIZE 16
+#define DRAW_OVERLAY
+#endif
+
+#define VRAM_X_COORD(x) ((x) * BLOCK_SIZE / 16)
+
+// All non-audio sectors in .STR files begin with this 32-byte header, which
+// contains metadata about the sector and is followed by a chunk of frame
+// bitstream data.
+// https://problemkaputt.de/psx-spx.htm#cdromfilevideostrstreamingandbspicturecompressionsony
+typedef struct {
+	uint16_t magic;			// Always 0x0160
+	uint16_t type;			// 0x8001 for MDEC
+	uint16_t sector_id;		// Chunk number (0 = first chunk of this frame)
+	uint16_t sector_count;	// Total number of chunks for this frame
+	uint32_t frame_id;		// Frame number
+	uint32_t bs_length;		// Total length of this frame in bytes
+
+	uint16_t width, height;
+	uint8_t  bs_header[8];
+	uint32_t _reserved;
+} STR_Header;
+
+// https://problemkaputt.de/psx-spx.htm#cdromxasubheaderfilechannelinterleave
+typedef struct {
+	uint8_t file, channel;
+	uint8_t submode, coding_info;
+} XA_Header;
+
+// https://problemkaputt.de/psx-spx.htm#cdromsectorencoding
+typedef struct {
+	CdlLOC     pos;
+	XA_Header  xa_header[2];
+	STR_Header str_header;
+	uint8_t    data[2016];
+	uint32_t   edc;
+	uint8_t    ecc[276];
+} STR_Sector;
+
+typedef struct {
+	uint16_t width, height;
+	uint32_t bs_data[0x2000];	// Bitstream data read from the disc
+	uint32_t mdec_data[0x8000];	// Decompressed data to be fed to the MDEC
+} StreamBuffer;
+
+typedef struct {
+	StreamBuffer frames[2];
+	uint32_t     slices[2][BLOCK_SIZE * SCREEN_YRES / 2];
+
+	int  frame_id, sector_count;
+	int  dropped_frames;
+	RECT slice_pos;
+	int  frame_width;
+
+	volatile int8_t sector_pending, frame_ready;
+	volatile int8_t cur_frame, cur_slice;
+} StreamContext;
+
+StreamContext str_ctx;
+
+// This buffer is used by cd_sector_handler() as a temporary area for sectors
+// read from the CD. Due to DMA limitations it can't be allocated on the stack
+// (especially not in the interrupt callbacks' stack, whose size is very
+// limited).
+STR_Sector sector_buffer;
+
+void cd_sector_handler(void) {
+	// Fetch the .STR header of the sector that has been read and check if the
+	// end-of-file bit is set in the XA header.
+	CdGetSector(&sector_buffer, sizeof(STR_Sector) / 4);
+
+	if (
+		(sector_buffer.xa_header[0].submode & (1 << 7)) ||
+		(sector_buffer.xa_header[1].submode & (1 << 7))
+	) {
+		CdControlF(CdlPause, 0);
+		str_ctx.frame_ready = -1;
+		return;
+	}
+
+	STR_Header   *header = &sector_buffer.str_header;
+	StreamBuffer *frame  = &str_ctx.frames[str_ctx.cur_frame];
+
+	// Ignore any non-MDEC sectors that might be present in the stream.
+	if (header->type != 0x8001)
+		return;
+
+	// If this sector is actually part of a new frame, validate the sectors
+	// that have been read so far and flip the bitstream data buffers.
+	if (header->frame_id != str_ctx.frame_id) {
+		// Do not set the ready flag if any sector has been missed.
+		if (str_ctx.sector_count)
+			str_ctx.dropped_frames++;
+		else
+			str_ctx.frame_ready = 1;
+
+		str_ctx.frame_id     = header->frame_id;
+		str_ctx.sector_count = header->sector_count;
+		str_ctx.cur_frame   ^= 1;
+
+		frame = &str_ctx.frames[str_ctx.cur_frame];
+
+		// Initialize the next frame. Dimensions must be rounded up to the
+		// nearest multiple of 16 as the MDEC operates on 16x16 pixel blocks.
+		frame->width  = (header->width  + 15) & 0xfff0;
+		frame->height = (header->height + 15) & 0xfff0;
+	}
+
+	// Append the payload contained in this sector to the current buffer.
+	memcpy(
+		&(frame->bs_data[2016 / 4 * header->sector_id]),
+		sector_buffer.data,
+		2016
+	);
+	str_ctx.sector_count--;
+}
+
+void mdec_dma_handler(void) {
+	// Handle any sectors that were not processed by cd_event_handler() (see
+	// below) while a DMA transfer from the MDEC was in progress. As the MDEC
+	// has just finished decoding a slice, they can be safely handled now.
+	if (str_ctx.sector_pending) {
+		cd_sector_handler();
+		str_ctx.sector_pending = 0;
+	}
+
+	// Upload the decoded slice to VRAM and start decoding the next slice (into
+	// another buffer) if any.
+	LoadImage(&str_ctx.slice_pos, str_ctx.slices[str_ctx.cur_slice]);
+
+	str_ctx.cur_slice   ^= 1;
+	str_ctx.slice_pos.x += BLOCK_SIZE;
+
+	if (str_ctx.slice_pos.x < str_ctx.frame_width)
+		DecDCTout(
+			str_ctx.slices[str_ctx.cur_slice],
+			BLOCK_SIZE * str_ctx.slice_pos.h / 2
+		);
+}
+
+void cd_event_handler(int event, uint8_t *payload) {
+	// Ignore all events other than a sector being ready.
+	if (event != CdlDataReady)
+		return;
+
+	// Only handle sectors immediately if the MDEC is not decoding a frame,
+	// otherwise defer handling to mdec_dma_handler(). This is a workaround for
+	// a hardware conflict between the DMA channels used for the CD drive and
+	// MDEC output, which shall not run simultaneously.
+	if (DecDCTinSync(1))
+		str_ctx.sector_pending = 1;
+	else
+		cd_sector_handler();
+}
+
+/* Stream helpers */
+
+void init_stream(void) {
+	EnterCriticalSection();
+	DMACallback(1, &mdec_dma_handler);
+	CdReadyCallback(&cd_event_handler);
+	ExitCriticalSection();
+
+	// Set the maximum amount of data DecDCTvlc() can output and copy the
+	// lookup table used for decompression to the scratchpad area. This is
+	// optional but makes the decompressor slightly faster. See the libpsxpress
+	// documentation for more details.
+	DecDCTvlcSize(0x8000);
+	DecDCTvlcCopyTable((DECDCTTAB *) 0x1f800000);
+
+	str_ctx.dropped_frames = 0;
+	str_ctx.cur_frame      = 0;
+	str_ctx.cur_slice      = 0;
+}
+
+StreamBuffer *get_next_frame(void) {
+	while (!str_ctx.frame_ready)
+		__asm__ volatile("");
+
+	if (str_ctx.frame_ready < 0)
+		return 0;
+
+	str_ctx.frame_ready = 0;
+	return &str_ctx.frames[str_ctx.cur_frame ^ 1];
+}
+
+void start_stream(CdlFILE *file) {
+	str_ctx.frame_id       = -1;
+	str_ctx.sector_pending =  0;
+	str_ctx.frame_ready    =  0;
+
+	CdSync(0, 0);
+
+	// Configure the CD drive to read 2340-byte sectors at 2x speed and to
+	// play any XA-ADPCM sectors that might be interleaved with the video data.
+	uint8_t mode = CdlModeSize | CdlModeRT | CdlModeSpeed;
+	CdControl(CdlSetmode, (const uint8_t *) &mode, 0);
+
+	// Start reading in real-time mode (i.e. without retrying in case of read
+	// errors) and wait for the first frame to be buffered.
+	CdControl(CdlReadS, &(file->pos), 0);
+
+	get_next_frame();
+}
+
+/* Main */
+
+static RenderContext ctx;
+
+#define SHOW_STATUS(...) { FntPrint(-1, __VA_ARGS__); FntFlush(-1); display(&ctx, 1); }
+#define SHOW_ERROR(...)  { SHOW_STATUS(__VA_ARGS__); while (1) __asm__("nop"); }
+
+int main(int argc, const char* argv[]) {
+	init_context(&ctx);
+
+	SHOW_STATUS("INITIALIZING\n");
+	SpuInit();
+	CdInit();
+	InitGeom(); // Required for PSn00bSDK's DecDCTvlc()
+	DecDCTReset(0);
+
+	SHOW_STATUS("OPENING VIDEO FILE\n");
+
+	CdlFILE file;
+	if (!CdSearchFile(&file, "\\VIDEO.STR"))
+		SHOW_ERROR("FAILED TO FIND VIDEO.STR\n");
+
+	init_stream();
+	start_stream(&file);
+
+	// Disable framebuffer clearing to get rid of flickering during playback.
+	display(&ctx, 1);
+	ctx.db[0].draw.isbg = 0;
+	ctx.db[1].draw.isbg = 0;
+#ifdef DISP_24BPP
+	ctx.db[0].disp.isrgb24 = 1;
+	ctx.db[1].disp.isrgb24 = 1;
+#endif
+
+	int decode_errors = 0;
+
+	while (1) {
+		// Wait for a full frame to be read from the disc and decompress the
+		// bitstream into the format expected by the MDEC. If the video has
+		// ended, restart playback from the beginning.
+		StreamBuffer *frame = get_next_frame();
+		if (!frame) {
+			start_stream(&file);
+			continue;
+		}
+
+#ifdef DRAW_OVERLAY
+		// Measure CPU usage of the decompressor using the hblank counter.
+		int total_time = TIMER_VALUE(1) + 1;
+		TIMER_VALUE(1) = 0;
+#endif
+
+		if (DecDCTvlc(frame->bs_data, frame->mdec_data)) {
+			decode_errors++;
+			continue;
+		}
+
+#ifdef DRAW_OVERLAY
+		int cpu_usage = TIMER_VALUE(1) * 100 / total_time;
+#endif
+
+		// Wait for the MDEC to finish decoding the previous frame, then flip
+		// the framebuffers to display it and prepare the buffer for the next
+		// frame.
+		// NOTE: you should *not* call VSync(0) during playback, as the refresh
+		// rate of the GPU is not synced to the video's frame rate. If you want
+		// to minimize screen tearing, consider triple buffering instead (i.e.
+		// always keep 2 fully decoded frames in VRAM and use VSyncCallback()
+		// to register a function that displays the next decoded frame whenever
+		// vblank occurs).
+		DecDCTinSync(0);
+		DecDCToutSync(0);
+
+#ifdef DRAW_OVERLAY
+		FntPrint(-1, "FRAME:%5d    READ ERRORS:  %5d\n", str_ctx.frame_id, str_ctx.dropped_frames);
+		FntPrint(-1, "CPU:  %5d%%   DECODE ERRORS:%5d\n", cpu_usage, decode_errors);
+		FntFlush(-1);
+#endif
+		display(&ctx, 0);
+
+		// Feed the newly decompressed frame to the MDEC. The MDEC will not
+		// actually start decoding it until an output buffer is also configured
+		// by calling DecDCTout() (see below).
+#ifdef DISP_24BPP
+		DecDCTin(frame->mdec_data, DECDCT_MODE_24BPP);
+#else
+		DecDCTin(frame->mdec_data, DECDCT_MODE_16BPP);
+#endif
+
+		// Place the frame at the center of the currently active framebuffer
+		// and start decoding the first slice. Decoded slices will be uploaded
+		// to VRAM in the background by mdec_dma_handler().
+		RECT *fb_clip = &(ctx.db[ctx.db_active].draw.clip);
+		int  x_offset = (fb_clip->w - frame->width)  / 2;
+		int  y_offset = (fb_clip->h - frame->height) / 2;
+
+		str_ctx.slice_pos.x = VRAM_X_COORD(fb_clip->x + x_offset);
+		str_ctx.slice_pos.y = fb_clip->y + y_offset;
+		str_ctx.slice_pos.w = BLOCK_SIZE;
+		str_ctx.slice_pos.h = frame->height;
+		str_ctx.frame_width = VRAM_X_COORD(frame->width);
+
+		DecDCTout(
+			str_ctx.slices[str_ctx.cur_slice],
+			BLOCK_SIZE * str_ctx.slice_pos.h / 2
+		);
+	}
+
+	return 0;
+}
diff --git a/examples/mdec/strvideo/system.cnf b/examples/mdec/strvideo/system.cnf
new file mode 100644
index 0000000..d199117
--- /dev/null
+++ b/examples/mdec/strvideo/system.cnf
@@ -0,0 +1,4 @@
+BOOT=cdrom:\strvideo.exe;1
+TCB=4
+EVENT=10
+STACK=801FFFF0
diff --git a/libpsn00b/include/psxcd.h b/libpsn00b/include/psxcd.h
index cf9ecad..429a439 100644
--- a/libpsn00b/include/psxcd.h
+++ b/libpsn00b/include/psxcd.h
@@ -55,8 +55,10 @@
 #define CdlModeAP			0x02
 #define CdlModeRept			0x04
 #define CdlModeSF			0x08
-#define CdlModeSize0		0x10
-#define CdlModeSize1		0x20
+//#define CdlModeSize0		0x10
+//#define CdlModeSize1		0x20
+#define CdlModeIgnore		0x10
+#define CdlModeSize			0x20
 #define CdlModeRT			0x40
 #define CdlModeSpeed		0x80
 
@@ -146,6 +148,7 @@ uint32_t CdSyncCallback(CdlCB func);
 long CdReadyCallback(CdlCB func);
 int CdGetSector(void *madr, int size);
 int CdGetSector2(void *madr, int size);
+int CdDataSync(int mode);
 
 CdlFILE* CdSearchFile(CdlFILE *loc, const char *filename);
 
diff --git a/libpsn00b/psxcd/getsector.c b/libpsn00b/psxcd/getsector.c
index bc1c8ae..ee5315a 100644
--- a/libpsn00b/psxcd/getsector.c
+++ b/libpsn00b/psxcd/getsector.c
@@ -7,6 +7,16 @@
 #include <psxcd.h>
 #include <hwregs_c.h>
 
+#define DATA_SYNC_TIMEOUT	0x100000
+
+/* Private utilities */
+
+#ifdef NDEBUG
+#define _LOG(...)
+#else
+#define _LOG(...) printf(__VA_ARGS__)
+#endif
+
 /* DMA transfer functions */
 
 int CdGetSector(void *madr, int size) {
@@ -33,3 +43,16 @@ int CdGetSector2(void *madr, int size) {
 
 	return 1;
 }
+
+int CdDataSync(int mode) {
+	if (mode)
+		return (DMA_CHCR(3) >> 24) & 1;
+
+	for (int i = DATA_SYNC_TIMEOUT; i; i--) {
+		if (!(DMA_CHCR(3) & (1 << 24)))
+			return 0;
+	}
+
+	_LOG("psxcd: CdDataSync() timeout\n");
+	return -1;
+}
diff --git a/libpsn00b/psxcd/psxcd.c b/libpsn00b/psxcd/psxcd.c
index 160a1ed..f48542d 100644
--- a/libpsn00b/psxcd/psxcd.c
+++ b/libpsn00b/psxcd/psxcd.c
@@ -285,18 +285,10 @@ int CdRead(int sectors, uint32_t *buf, int mode)
 	_cd_read_addr = buf;
 	
 	// Determine sector based on mode flags
-	if( mode & CdlModeSize0 )
-	{
-		_cd_read_sector_sz = 2328 / 4;
-	}
-	else if( mode & CdlModeSize1 )
-	{
+	if( mode & CdlModeSize )
 		_cd_read_sector_sz = 2340 / 4;
-	}
 	else
-	{
 		_cd_read_sector_sz = 2048 / 4;
-	}
 	
 	_cd_read_counter = VSync(-1);
 	
diff --git a/libpsn00b/psxpress/mdec.c b/libpsn00b/psxpress/mdec.c
index d308994..06510cb 100644
--- a/libpsn00b/psxpress/mdec.c
+++ b/libpsn00b/psxpress/mdec.c
@@ -164,6 +164,11 @@ int DecDCTinSync(int mode) {
 void DecDCTout(uint32_t *data, size_t length) {
 	DecDCToutSync(0);
 
+	if ((length >= DMA_CHUNK_LENGTH) && (length % DMA_CHUNK_LENGTH)) {
+		_LOG("psxpress: transfer data length (%d) is not a multiple of %d, rounding\n", length, DMA_CHUNK_LENGTH);
+		length += DMA_CHUNK_LENGTH - 1;
+	}
+
 	DMA_MADR(1) = (uint32_t) data;
 	if (length < DMA_CHUNK_LENGTH)
 		DMA_BCR(1) = 0x00010000 | length;
diff --git a/libpsn00b/psxpress/vlc.s b/libpsn00b/psxpress/vlc.s
index fe51642..885a3f7 100644
--- a/libpsn00b/psxpress/vlc.s
+++ b/libpsn00b/psxpress/vlc.s
@@ -131,10 +131,12 @@ _vlc_skip_context_load:
 	#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++
+	# The DC coefficient in version 2 frames is not compressed. Value 0x1ff is
+	# used to signal the end of the bitstream.
+	srl   $v0, $t0, 22 # prefix = (window >> (32 - 10))
+	li    $v1, 0x01ff
+	beq   $v0, $v1, .Lstop_processing # if (prefix == 0x1ff) break
+	or    $v0, $t3 # *output = prefix | quant_scale
 	sll   $t0, 10 # window <<= 10
 	addiu $t5, -10 # bit_offset -= 10
 	b     .Lwrite_value
diff --git a/libpsn00b/psxpress/vlc2.c b/libpsn00b/psxpress/vlc2.c
index 73b54b2..9eb99bf 100644
--- a/libpsn00b/psxpress/vlc2.c
+++ b/libpsn00b/psxpress/vlc2.c
@@ -141,7 +141,10 @@ int __attribute__((optimize(3))) DecDCTvlcContinue2(
 				// TODO: version 3 is currently not supported.
 				return -1;
 			} else {
-				value   = _get_bits_unsigned(10);
+				value = _get_bits_unsigned(10);
+				if (value == 0x1ff)
+					break;
+
 				*output = value | quant_scale;
 				_advance_window(10);
 			}
-- 
cgit v1.2.3


From 2f100c78c0f12b56bcd73c203e6216d415d9f772 Mon Sep 17 00:00:00 2001
From: spicyjpeg <thatspicyjpeg@gmail.com>
Date: Mon, 17 Oct 2022 22:42:00 +0200
Subject: Remove interrupt disabling calls in psxcd callback APIs

---
 examples/cdrom/cdxa/main.c  |  2 ++
 libpsn00b/include/psxcd.h   |  4 ++--
 libpsn00b/psxcd/isofs.c     | 19 ++++++++++++++-----
 libpsn00b/psxcd/psxcd.c     | 11 ++++++++---
 libpsn00b/psxcd/psxcd_asm.s | 30 ++++++------------------------
 5 files changed, 32 insertions(+), 34 deletions(-)

(limited to 'libpsn00b/include')

diff --git a/examples/cdrom/cdxa/main.c b/examples/cdrom/cdxa/main.c
index 4921658..93cf01a 100644
--- a/examples/cdrom/cdxa/main.c
+++ b/examples/cdrom/cdxa/main.c
@@ -349,7 +349,9 @@ int main(int argc, const char* argv[])
 	xa_loc = file.pos;
 	
 	/* Hook XA callback function to CdReadyCallback (for auto stop/loop */
+	EnterCriticalSection();
 	CdReadyCallback(xa_callback);
+	ExitCriticalSection();
 
 	/* Set CD mode for XA streaming (2x speed, send XA to SPU, enable filter */
 	i = CdlModeSpeed|CdlModeRT|CdlModeSF;
diff --git a/libpsn00b/include/psxcd.h b/libpsn00b/include/psxcd.h
index 429a439..0460f20 100644
--- a/libpsn00b/include/psxcd.h
+++ b/libpsn00b/include/psxcd.h
@@ -145,7 +145,7 @@ 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 CdReadyCallback(CdlCB func);
 int CdGetSector(void *madr, int size);
 int CdGetSector2(void *madr, int size);
 int CdDataSync(int mode);
@@ -168,7 +168,7 @@ void CdCloseDir(CdlDIR* dir);
 
 int CdGetVolumeLabel(char* label);
 
-long* CdAutoPauseCallback(void(*func)());
+int* CdAutoPauseCallback(void(*func)());
 int CdIsoError();
 
 int CdLoadSession(int session);
diff --git a/libpsn00b/psxcd/isofs.c b/libpsn00b/psxcd/isofs.c
index 16e64ef..4ec701c 100644
--- a/libpsn00b/psxcd/isofs.c
+++ b/libpsn00b/psxcd/isofs.c
@@ -3,7 +3,7 @@
 #include <stdlib.h>
 #include <string.h>
 #include <psxgpu.h>
-#include <psxsio.h>
+#include <psxapi.h>
 #include "psxcd.h"
 #include "isofs.h"
 
@@ -795,8 +795,10 @@ int CdLoadSession(int session)
 	}
 	
 	// Set search routine callback
+	EnterCriticalSection();
 	ready_oldcb = CdReadyCallback(_scan_callback);
-	
+	ExitCriticalSection();
+
 	_ses_scanfound = 0;
 	_ses_scancount = 0;
 	_ses_scancomplete = 0;
@@ -811,21 +813,28 @@ int CdLoadSession(int session)
 	
 	// Wait until scan complete
 	while(!_ses_scancomplete);
-		
+
+	EnterCriticalSection();
 	CdReadyCallback((void*)_ready_oldcb);
-	
+	ExitCriticalSection();
+
 	if( !_ses_scanfound )
 	{
 		_LOG("psxcd: CdLoadSession(): Did not find volume descriptor.\n");
 
 		_cd_iso_error = CdlIsoInvalidFs;
+		EnterCriticalSection();
 		CdReadyCallback((CdlCB)ready_oldcb);
+		ExitCriticalSection();
+
 		return -1;
 	}
 	
 	// Restore old callback if any
+	EnterCriticalSection();
 	CdReadyCallback((CdlCB)ready_oldcb);
-	
+	ExitCriticalSection();
+
 	// Wait until CD-ROM has completely stopped reading, to get a consistent
 	// fix of the CD-ROM pickup's current location
 	do
diff --git a/libpsn00b/psxcd/psxcd.c b/libpsn00b/psxcd/psxcd.c
index f48542d..6730531 100644
--- a/libpsn00b/psxcd/psxcd.c
+++ b/libpsn00b/psxcd/psxcd.c
@@ -1,6 +1,7 @@
 #include <stdint.h>
 #include <stdio.h>
 #include <psxgpu.h>
+#include <psxapi.h>
 #include "psxcd.h"
 
 #define READ_TIMEOUT	600		// 10 seconds for NTSC
@@ -293,8 +294,10 @@ int CdRead(int sectors, uint32_t *buf, int mode)
 	_cd_read_counter = VSync(-1);
 	
 	// Set read callback
+	EnterCriticalSection();
 	_cd_read_oldcb = CdReadyCallback(_CdReadReadyCallback);
-	
+	ExitCriticalSection();
+
 	// Set specified mode
 	CdControl(CdlSetmode, (uint8_t*)&mode, 0);
 	
@@ -320,9 +323,11 @@ static void CdDoRetry()
 	
 	// Reset timeout
 	_cd_read_counter = VSync(-1);
-	
+
+	EnterCriticalSection();
 	CdReadyCallback(_CdReadReadyCallback);
-	
+	ExitCriticalSection();
+
 	// Retry read
 	CdControl(CdlSetloc, (void*)&_cd_last_setloc, 0);
 	CdControl(CdlReadN, 0, (uint8_t*)_cd_read_result);
diff --git a/libpsn00b/psxcd/psxcd_asm.s b/libpsn00b/psxcd/psxcd_asm.s
index 16e17d8..c0a5312 100644
--- a/libpsn00b/psxcd/psxcd_asm.s
+++ b/libpsn00b/psxcd/psxcd_asm.s
@@ -413,17 +413,11 @@ CdAutoPauseCallback:
 	lw		$v0, 0($v1)
 	
 	la		$v1, _cd_callback_int4
-	
-	jal		EnterCriticalSection
-	nop
-	
+
 	lw		$a0, 4($sp)
 	nop
 	sw		$a0, 0($v1)
-	
-	jal		ExitCriticalSection
-	nop
-	
+
 	lw		$ra, 0($sp)
 	addiu	$sp, 8
 	jr		$ra
@@ -443,17 +437,11 @@ CdReadyCallback:
 	
 	la		$v1, _cd_callback_int1_data
 	sw		$v0, 8($sp)
-	
-	jal		EnterCriticalSection
-	nop
-	
+
 	lw		$a0, 4($sp)
 	nop
 	sw		$a0, 0($v1)
-	
-	jal		ExitCriticalSection
-	nop
-	
+
 	lw		$ra, 0($sp)
 	lw		$v0, 8($sp)
 	jr		$ra
@@ -472,17 +460,11 @@ CdSyncCallback:
 	
 	la		$v1, _cd_sync_cb
 	sw		$v0, 8($sp)
-	
-	jal		EnterCriticalSection
-	nop
-	
+
 	lw		$a0, 4($sp)
 	nop
 	sw		$a0, 0($v1)
-	
-	jal		ExitCriticalSection
-	nop
-	
+
 	lw		$ra, 0($sp)
 	lw		$v0, 8($sp)
 	jr		$ra
-- 
cgit v1.2.3


From b71a55bc489db6bc9beca5cee9cd584e82846ac8 Mon Sep 17 00:00:00 2001
From: spicyjpeg <thatspicyjpeg@gmail.com>
Date: Tue, 18 Oct 2022 15:51:52 +0200
Subject: Add MoveImage(), use draw queue for psxgpu VRAM APIs

---
 libpsn00b/include/psxgpu.h |  22 ++++--
 libpsn00b/psxgpu/common.c  | 176 +++++++++++++++++++++++++++------------------
 libpsn00b/psxgpu/env.c     |   4 +-
 libpsn00b/psxgpu/image.c   |  43 +++++++----
 libpsn00b/psxspu/common.c  |  58 +++++++--------
 5 files changed, 181 insertions(+), 122 deletions(-)

(limited to 'libpsn00b/include')

diff --git a/libpsn00b/include/psxgpu.h b/libpsn00b/include/psxgpu.h
index 0e7ec00..f2568b0 100644
--- a/libpsn00b/include/psxgpu.h
+++ b/libpsn00b/include/psxgpu.h
@@ -488,20 +488,30 @@ void PutDrawEnvFast(DRAWENV *env);
 
 int GetODE(void);
 int VSync(int mode);
-int DrawSync(int mode);
-
 void *VSyncHaltFunction(void (*func)(void));
 void *VSyncCallback(void (*func)(void));
+
+int EnqueueDrawOp(
+	void		(*func)(uint32_t, uint32_t, uint32_t),
+	uint32_t	arg1,
+	uint32_t	arg2,
+	uint32_t	arg3
+);
+int DrawSync(int mode);
 void *DrawSyncCallback(void (*func)(void));
 
-void LoadImage(const RECT *rect, const uint32_t *data);
-void StoreImage(const RECT *rect, uint32_t *data);
+int LoadImage(const RECT *rect, const uint32_t *data);
+int StoreImage(const RECT *rect, uint32_t *data);
+int MoveImage(const RECT *rect, int x, int y);
+void LoadImage2(const RECT *rect, const uint32_t *data);
+void StoreImage2(const RECT *rect, uint32_t *data);
+void MoveImage2(const RECT *rect, int x, int y);
 
 void ClearOTagR(uint32_t *ot, size_t length);
 void ClearOTag(uint32_t *ot, size_t length);
-void DrawOTag(const uint32_t *ot);
+int DrawOTag(const uint32_t *ot);
+int DrawOTagEnv(const uint32_t *ot, DRAWENV *env);
 void DrawOTag2(const uint32_t *ot);
-void DrawOTagEnv(const uint32_t *ot, DRAWENV *env);
 void DrawPrim(const uint32_t *pri);
 
 void AddPrim(uint32_t *ot, const void *pri);
diff --git a/libpsn00b/psxgpu/common.c b/libpsn00b/psxgpu/common.c
index 1e3d9e5..bf70b72 100644
--- a/libpsn00b/psxgpu/common.c
+++ b/libpsn00b/psxgpu/common.c
@@ -16,6 +16,13 @@
 
 static void _default_vsync_halt(void);
 
+/* Private types */
+
+typedef struct {
+	void     (*func)(uint32_t, uint32_t, uint32_t);
+	uint32_t arg1, arg2, arg3;
+} QueueEntry;
+
 /* Internal globals */
 
 GPU_VideoMode _gpu_video_mode;
@@ -24,10 +31,10 @@ 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;
+static volatile QueueEntry _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;
 
 /* Private utilities and interrupt handlers */
 
@@ -49,11 +56,11 @@ static void _gpu_dma_handler(void) {
 	while (!(GPU_GP1 & (1 << 26)))
 		__asm__ volatile("");
 
-	if (_queue_length) {
-		DrawOTag2(_draw_queue[_queue_head++]);
+	if (--_queue_length) {
+		QueueEntry *entry = &_draw_queue[_queue_head++];
+		_queue_head      %= QUEUE_LENGTH;
 
-		_queue_length--;
-		_queue_head %= QUEUE_LENGTH;
+		entry->func(entry->arg1, entry->arg2, entry->arg3);
 	} else {
 		GPU_GP1 = 0x04000000; // Disable DMA request
 
@@ -103,7 +110,7 @@ void ResetGraph(int mode) {
 	_last_hblank    = 0;
 }
 
-/* Syncing API */
+/* VSync() API */
 
 // TODO: add support for no$psx's "halt" register
 static void _default_vsync_halt(void) {
@@ -144,27 +151,6 @@ int VSync(int mode) {
 	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;
@@ -182,6 +168,80 @@ void *VSyncCallback(void (*func)(void)) {
 	return old_callback;
 }
 
+/* Command queue API */
+
+// This function is normally only used internally, but it is exposed for
+// advanced use cases.
+int EnqueueDrawOp(
+	void		(*func)(uint32_t, uint32_t, uint32_t),
+	uint32_t	arg1,
+	uint32_t	arg2,
+	uint32_t	arg3
+) {
+	// If GPU DMA is currently busy, append the command to the queue instead of
+	// executing 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 (_queue_length) {
+		if (_queue_length >= QUEUE_LENGTH) {
+			IRQ_MASK = mask;
+			_LOG("psxgpu: draw queue overflow, dropping commands\n");
+			return -1;
+		}
+
+		int length    = _queue_length;
+		_queue_length = length + 1;
+
+		QueueEntry *entry = &_draw_queue[_queue_tail++];
+		_queue_tail      %= QUEUE_LENGTH;
+
+		entry->func = func;
+		entry->arg1 = arg1;
+		entry->arg2 = arg2;
+		entry->arg3 = arg3;
+
+		IRQ_MASK = mask;
+		return length;
+	}
+
+	_queue_length = 1;
+
+	IRQ_MASK = mask;
+	func(arg1, arg2, arg3);
+	return 0;
+}
+
+int DrawSync(int mode) {
+	if (mode)
+		return _queue_length;
+
+	// Wait for the queue to become empty.
+	for (int i = VSYNC_TIMEOUT; i; i--) {
+		if (!_queue_length)
+			break;
+	}
+
+	if (!_queue_length) {
+		// 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("");
+	} else {
+		printf("psxgpu: DrawSync() timeout\n");
+	}
+
+	return _queue_length;
+}
+
 void *DrawSyncCallback(void (*func)(void)) {
 	EnterCriticalSection();
 
@@ -214,45 +274,8 @@ void ClearOTag(uint32_t *ot, size_t length) {
 	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;
-		_LOG("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 AddPrim(uint32_t *ot, const void *pri) {
+	addPrim(ot, pri);
 }
 
 void DrawPrim(const uint32_t *pri) {	
@@ -273,8 +296,19 @@ void DrawPrim(const uint32_t *pri) {
 	DMA_CHCR(2) = 0x01000201;
 }
 
-void AddPrim(uint32_t *ot, const void *pri) {
-	addPrim(ot, pri);
+int DrawOTag(const uint32_t *ot) {
+	return EnqueueDrawOp(&DrawOTag2, (uint32_t) ot, 0, 0);
+}
+
+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;
 }
 
 /* Misc. functions */
diff --git a/libpsn00b/psxgpu/env.c b/libpsn00b/psxgpu/env.c
index 5642ad4..1b97026 100644
--- a/libpsn00b/psxgpu/env.c
+++ b/libpsn00b/psxgpu/env.c
@@ -37,7 +37,7 @@ DRAWENV *SetDefDrawEnv(DRAWENV *env, int x, int y, int w, int h) {
 	return env;
 }
 
-void DrawOTagEnv(const uint32_t *ot, DRAWENV *env) {
+int DrawOTagEnv(const uint32_t *ot, DRAWENV *env) {
 	DR_ENV *prim = &(env->dr_env);
 
 	// All commands are grouped into a single display list packet for
@@ -85,7 +85,7 @@ void DrawOTagEnv(const uint32_t *ot, DRAWENV *env) {
 	//while (!(GPU_GP1 & (1 << 26)))
 		//__asm__ volatile("");
 
-	DrawOTag((const uint32_t *) prim);
+	return EnqueueDrawOp(&DrawOTag2, (uint32_t) prim, 0, 0);
 }
 
 void PutDrawEnv(DRAWENV *env) {
diff --git a/libpsn00b/psxgpu/image.c b/libpsn00b/psxgpu/image.c
index a0d7065..c09a59d 100644
--- a/libpsn00b/psxgpu/image.c
+++ b/libpsn00b/psxgpu/image.c
@@ -18,12 +18,7 @@
 #define _LOG(...) printf(__VA_ARGS__)
 #endif
 
-static void _load_store_image(
-	uint32_t	command,
-	int			mode,
-	const RECT	*rect,
-	uint32_t	*data
-) {
+static void _dma_transfer(const RECT *rect, uint32_t *data, int write) {
 	size_t length = rect->w * rect->h;
 	if (length % 2)
 		_LOG("psxgpu: can't transfer an odd number of pixels\n");
@@ -34,18 +29,17 @@ static void _load_store_image(
 		length += DMA_CHUNK_LENGTH - 1;
 	}
 
-	DrawSync(0);
 	GPU_GP1 = 0x04000000; // Disable DMA request
 	GPU_GP0 = 0x01000000; // Flush cache
 
-	GPU_GP0 = command;
+	GPU_GP0 = write ? 0xa0000000 : 0xc0000000;
 	//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;
+	GPU_GP1 = 0x04000002 | (write ^ 1);
 
 	DMA_MADR(2) = (uint32_t) data;
 	if (length < DMA_CHUNK_LENGTH)
@@ -53,17 +47,38 @@ static void _load_store_image(
 	else
 		DMA_BCR(2) = DMA_CHUNK_LENGTH | ((length / DMA_CHUNK_LENGTH) << 16);
 
-	DMA_CHCR(2) = 0x01000200 | ((mode & 1) ^ 1);
+	DMA_CHCR(2) = 0x01000200 | write;
 }
 
 /* VRAM transfer API */
 
-void LoadImage(const RECT *rect, const uint32_t *data) {
-	_load_store_image(0xa0000000, 2, rect, (uint32_t *) data);
+int LoadImage(const RECT *rect, const uint32_t *data) {
+	return EnqueueDrawOp(&_dma_transfer, (uint32_t) rect, (uint32_t) data, 1);
 }
 
-void StoreImage(const RECT *rect, uint32_t *data) {
-	_load_store_image(0xc0000000, 3, rect, data);
+int StoreImage(const RECT *rect, uint32_t *data) {
+	return EnqueueDrawOp(&_dma_transfer, (uint32_t) rect, (uint32_t) data, 0);
+}
+
+int MoveImage(const RECT *rect, int x, int y) {
+	return EnqueueDrawOp(&MoveImage2, (uint32_t) rect, x, y);
+}
+
+void LoadImage2(const RECT *rect, const uint32_t *data) {
+	_dma_transfer(rect, (uint32_t *) data, 1);
+}
+
+void StoreImage2(const RECT *rect, uint32_t *data) {
+	_dma_transfer(rect, data, 0);
+}
+
+void MoveImage2(const RECT *rect, int x, int y) {
+	GPU_GP0 = 0x80000000;
+	//GPU_GP0 = rect->x | (rect->y << 16);
+	GPU_GP0 = *((const uint32_t *) &(rect->x));
+	GPU_GP0 = (x & 0xffff) | (y << 16);
+	//GPU_GP0 = rect->w | (rect->h << 16);
+	GPU_GP0 = *((const uint32_t *) &(rect->w));
 }
 
 /* .TIM image parsers */
diff --git a/libpsn00b/psxspu/common.c b/libpsn00b/psxspu/common.c
index d6508c7..380bd3d 100644
--- a/libpsn00b/psxspu/common.c
+++ b/libpsn00b/psxspu/common.c
@@ -34,6 +34,33 @@ static void _wait_status(uint16_t mask, uint16_t value) {
 	_LOG("psxspu: status register timeout (0x%04x)\n", SPU_STAT);
 }
 
+static void _dma_transfer(uint32_t *data, size_t length, int write) {
+	if (length % 4)
+		_LOG("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)) {
+		_LOG("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 |= write ? 0x0020 : 0x0030;
+	_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 | write;
+}
+
 /* Public API */
 
 void SpuInit(void) {
@@ -87,35 +114,8 @@ void SpuInit(void) {
 	SPU_CD_VOL_R		= 0x7fff;
 }
 
-static void _load_store_data(uint32_t *data, size_t length, int mode) {
-	if (length % 4)
-		_LOG("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)) {
-		_LOG("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);
+	_dma_transfer(data, size, 0);
 }
 
 void SpuWrite(const uint32_t *data, size_t size) {
@@ -138,7 +138,7 @@ void SpuWrite(const uint32_t *data, size_t size) {
 		return;
 	}
 
-	_load_store_data((uint32_t *) data, size, 2);
+	_dma_transfer((uint32_t *) data, size, 1);
 }
 
 SPU_TransferMode SpuSetTransferMode(SPU_TransferMode mode) {
-- 
cgit v1.2.3


From 9b2ffc6078a850b7d354855cca7622090b41f30c Mon Sep 17 00:00:00 2001
From: spicyjpeg <thatspicyjpeg@gmail.com>
Date: Tue, 18 Oct 2022 17:29:42 +0200
Subject: Add debug log buffering, fix GetHeapUsage()

---
 libpsn00b/include/psxetc.h  |  15 ++++++
 libpsn00b/libc/abort.c      |  16 ++-----
 libpsn00b/libc/malloc.c     |  22 ++++-----
 libpsn00b/psxcd/getsector.c |  11 +----
 libpsn00b/psxcd/isofs.c     | 112 ++++++++++++++++++++------------------------
 libpsn00b/psxcd/psxcd.c     |  14 ++----
 libpsn00b/psxetc/dl.c       |  91 +++++++++++++++++------------------
 libpsn00b/psxetc/logging.c  |  50 ++++++++++++++++++++
 libpsn00b/psxgpu/common.c   |  31 +++++-------
 libpsn00b/psxgpu/env.c      |   2 +-
 libpsn00b/psxgpu/image.c    |  22 ++++-----
 libpsn00b/psxpress/mdec.c   |  20 +++-----
 libpsn00b/psxspu/common.c   |  14 ++----
 13 files changed, 216 insertions(+), 204 deletions(-)
 create mode 100644 libpsn00b/psxetc/logging.c

(limited to 'libpsn00b/include')

diff --git a/libpsn00b/include/psxetc.h b/libpsn00b/include/psxetc.h
index 24485d9..fcfec06 100644
--- a/libpsn00b/include/psxetc.h
+++ b/libpsn00b/include/psxetc.h
@@ -6,12 +6,27 @@
 #ifndef __PSXETC_H
 #define __PSXETC_H
 
+/* Macros */
+
+// This macro is used internally by PSn00bSDK to log debug messages to a buffer
+// which is then printed to stdout when calling VSync().
+#ifdef NDEBUG
+#define _sdk_log(...)
+#define _sdk_dump_log()
+#else
+#define _sdk_log(...)	_sdk_log_inner(__VA_ARGS__)
+#define _sdk_dump_log()	_sdk_dump_log_inner()
+#endif
+
 /* Public API */
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
+void _sdk_log_inner(const char *fmt, ...);
+void _sdk_dump_log_inner(void);
+
 void *InterruptCallback(int irq, void (*func)(void));
 void *GetInterruptCallback(int irq);
 void *DMACallback(int dma, void (*func)(void));
diff --git a/libpsn00b/libc/abort.c b/libpsn00b/libc/abort.c
index 1108160..2db5016 100644
--- a/libpsn00b/libc/abort.c
+++ b/libpsn00b/libc/abort.c
@@ -1,16 +1,10 @@
 
-#include <stdio.h>
-
-#ifdef NDEBUG
-#define _LOG(...)
-#else
-#define _LOG(...) printf(__VA_ARGS__)
-#endif
+#include <psxetc.h>
 
 /* Standard abort */
 
-void abort() {
-	_LOG("abort()\n");
+void abort(void) {
+	_sdk_log("abort()\n");
 
 	for (;;)
 		__asm__ volatile("");
@@ -19,7 +13,7 @@ void abort() {
 /* Internal function used by assert() macro */
 
 void _assert_abort(const char *file, int line, const char *expr) {
-	_LOG("%s:%d: assert(%s)\n", file, line, expr);
+	_sdk_log("%s:%d: assert(%s)\n", file, line, expr);
 
 	for (;;)
 		__asm__ volatile("");
@@ -28,7 +22,7 @@ void _assert_abort(const char *file, int line, const char *expr) {
 /* Pure virtual function call (C++) */
 
 void __cxa_pure_virtual(void) {
-	_LOG("__cxa_pure_virtual()\n");
+	_sdk_log("__cxa_pure_virtual()\n");
 
 	for (;;)
 		__asm__ volatile("");
diff --git a/libpsn00b/libc/malloc.c b/libpsn00b/libc/malloc.c
index acac753..e9fd6f4 100644
--- a/libpsn00b/libc/malloc.c
+++ b/libpsn00b/libc/malloc.c
@@ -120,7 +120,7 @@ __attribute__((weak)) void *malloc(size_t size) {
 		_alloc_head = new;
 		_alloc_tail = new;
 
-		TrackHeapUsage(_size);
+		TrackHeapUsage(size);
 		return ptr;
 	}
 
@@ -139,7 +139,7 @@ __attribute__((weak)) void *malloc(size_t size) {
 		_alloc_head->prev = new;
 		_alloc_head       = new;
 
-		TrackHeapUsage(_size);
+		TrackHeapUsage(size);
 		return ptr;
 	}
 
@@ -157,7 +157,7 @@ __attribute__((weak)) void *malloc(size_t size) {
 		(new->next)->prev = new;
 		prev->next        = new;
 
-		TrackHeapUsage(_size);
+		TrackHeapUsage(size);
 		return ptr;
 	}
 
@@ -175,7 +175,7 @@ __attribute__((weak)) void *malloc(size_t size) {
 	_alloc_tail->next = new;
 	_alloc_tail       = new;
 
-	TrackHeapUsage(_size);
+	TrackHeapUsage(size);
 	return ptr;
 }
 
@@ -196,7 +196,7 @@ __attribute__((weak)) void *realloc(void *ptr, size_t size) {
 
 	// New memory block shorter?
 	if (prev->size >= _size) {
-		TrackHeapUsage(_size - prev->size);
+		TrackHeapUsage(size - prev->size);
 		prev->size = _size;
 
 		if (!prev->next)
@@ -211,14 +211,14 @@ __attribute__((weak)) void *realloc(void *ptr, size_t size) {
 		if (!new)
 			return 0;
 
-		TrackHeapUsage(_size - prev->size);
+		TrackHeapUsage(size - prev->size);
 		prev->size = _size;
 		return ptr;
 	}
 
 	// Do we have free memory after it?
 	if (((prev->next)->ptr - ptr) > _size) {
-		TrackHeapUsage(_size - prev->size);
+		TrackHeapUsage(size - prev->size);
 		prev->size = _size;
 		return ptr;
 	}
@@ -250,7 +250,7 @@ __attribute__((weak)) void free(void *ptr) {
 			sbrk(-size);
 		}
 
-		TrackHeapUsage(-size);
+		TrackHeapUsage(-(_alloc_head->size));
 		return;
 	}
 
@@ -265,17 +265,15 @@ __attribute__((weak)) void free(void *ptr) {
 	if (cur->next) {
 		// In the middle, just unlink it
 		(cur->next)->prev = cur->prev;
-		TrackHeapUsage(-(cur->size + sizeof(BlockHeader)));
 	} else {
 		// At the end, shrink heap
-		_alloc_tail = cur->prev;
-
 		void  *top  = sbrk(0);
 		size_t size = (top - (cur->prev)->ptr) - (cur->prev)->size;
+		_alloc_tail = cur->prev;
 
 		sbrk(-size);
-		TrackHeapUsage(-size);
 	}
 
+	TrackHeapUsage(-(cur->size));
 	(cur->prev)->next = cur->next;
 }
diff --git a/libpsn00b/psxcd/getsector.c b/libpsn00b/psxcd/getsector.c
index ee5315a..31d0ac7 100644
--- a/libpsn00b/psxcd/getsector.c
+++ b/libpsn00b/psxcd/getsector.c
@@ -4,19 +4,12 @@
  */
 
 #include <stdint.h>
+#include <psxetc.h>
 #include <psxcd.h>
 #include <hwregs_c.h>
 
 #define DATA_SYNC_TIMEOUT	0x100000
 
-/* Private utilities */
-
-#ifdef NDEBUG
-#define _LOG(...)
-#else
-#define _LOG(...) printf(__VA_ARGS__)
-#endif
-
 /* DMA transfer functions */
 
 int CdGetSector(void *madr, int size) {
@@ -53,6 +46,6 @@ int CdDataSync(int mode) {
 			return 0;
 	}
 
-	_LOG("psxcd: CdDataSync() timeout\n");
+	_sdk_log("psxcd: CdDataSync() timeout\n");
 	return -1;
 }
diff --git a/libpsn00b/psxcd/isofs.c b/libpsn00b/psxcd/isofs.c
index 4ec701c..0425c0d 100644
--- a/libpsn00b/psxcd/isofs.c
+++ b/libpsn00b/psxcd/isofs.c
@@ -1,18 +1,12 @@
 #include <stdint.h>
-#include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
 #include <psxgpu.h>
 #include <psxapi.h>
+#include <psxetc.h>
 #include "psxcd.h"
 #include "isofs.h"
 
-#ifdef NDEBUG
-#define _LOG(...)
-#else
-#define _LOG(...) printf(__VA_ARGS__)
-#endif
-
 #define DEFAULT_PATH_SEP	'\\'
 #define IS_PATH_SEP(ch)		(((ch) == '/') || ((ch) == '\\'))
 
@@ -49,7 +43,7 @@ static int _CdReadIsoDescriptor(int session_offs)
 			CdControl(CdlNop, 0, 0);
 			if( (CdStatus()&0x10) )
 			{
-				_LOG("psxcd: Lid is still open.\n");
+				_sdk_log("psxcd: Lid is still open.\n");
 
 				_cd_iso_error = CdlIsoLidOpen;
 				return -1;
@@ -64,45 +58,45 @@ static int _CdReadIsoDescriptor(int session_offs)
 		return 0;
 	}
 
-	_LOG("psxcd: Parsing ISO file system.\n");
+	_sdk_log("psxcd: Parsing ISO file system.\n");
 
 	// Seek to volume descriptor
 	CdIntToPos(16+session_offs, &loc);
 	if( !CdControl(CdlSetloc, (uint8_t*)&loc, 0) )
 	{
-		_LOG("psxcd: Could not set seek destination.\n");
+		_sdk_log("psxcd: Could not set seek destination.\n");
 
 		_cd_iso_error = CdlIsoSeekError;
 		return -1;
 	}
 
-	_LOG("psxcd: Read sectors.\n");
+	_sdk_log("psxcd: Read sectors.\n");
 
 	// Read volume descriptor
 	CdRead(1, (uint32_t*)_cd_iso_descriptor_buff, CdlModeSpeed);
 	
 	if( CdReadSync(0, 0) )
 	{
-		_LOG("psxcd: Error reading ISO volume descriptor.\n");
+		_sdk_log("psxcd: Error reading ISO volume descriptor.\n");
 
 		_cd_iso_error = CdlIsoReadError;
 		return -1;
 	}
 
-	_LOG("psxcd: Read complete.\n");
+	_sdk_log("psxcd: Read complete.\n");
 
 	// Verify if volume descriptor is present
 	descriptor = (ISO_DESCRIPTOR*)_cd_iso_descriptor_buff;
 	if( strncmp("CD001", descriptor->header.id, 5) )
 	{
-		_LOG("psxcd: Disc does not contain a ISO9660 file system.\n");
+		_sdk_log("psxcd: Disc does not contain a ISO9660 file system.\n");
 
 		_cd_iso_error = CdlIsoInvalidFs;
 		return -1;
 	}
 
-	_LOG("psxcd: Path table LBA = %d\n", descriptor->pathTable1Offs);
-	_LOG("psxcd: Path table len = %d\n", descriptor->pathTableSize.lsb);
+	_sdk_log("psxcd: Path table LBA = %d\n", descriptor->pathTable1Offs);
+	_sdk_log("psxcd: Path table len = %d\n", descriptor->pathTableSize.lsb);
 
 	// Allocate path table buffer
 	i = ((2047+descriptor->pathTableSize.lsb)>>11)<<11;
@@ -112,7 +106,7 @@ static int _CdReadIsoDescriptor(int session_offs)
 	}
 	_cd_iso_pathtable_buff = (uint8_t*)malloc(i);
 
-	_LOG("psxcd: Allocated %d bytes for path table.\n", i);
+	_sdk_log("psxcd: Allocated %d bytes for path table.\n", i);
 
 	// Read path table
 	CdIntToPos(descriptor->pathTable1Offs, &loc);
@@ -120,7 +114,7 @@ static int _CdReadIsoDescriptor(int session_offs)
 	CdRead(i>>11, (uint32_t*)_cd_iso_pathtable_buff, CdlModeSpeed);
 	if( CdReadSync(0, 0) )
 	{
-		_LOG("psxcd: Error reading ISO path table.\n");
+		_sdk_log("psxcd: Error reading ISO path table.\n");
 
 		_cd_iso_error = CdlIsoReadError;
 		return -1;
@@ -148,11 +142,11 @@ static int _CdReadIsoDirectory(int lba)
 	CdIntToPos(lba, &loc);
 	i = CdPosToInt(&loc);
 
-	_LOG("psxcd: Seek to sector %d\n", i);
+	_sdk_log("psxcd: Seek to sector %d\n", i);
 
 	if( !CdControl(CdlSetloc, (uint8_t*)&loc, 0) )
 	{
-		_LOG("psxcd: Could not set seek destination.\n");
+		_sdk_log("psxcd: Could not set seek destination.\n");
 
 		_cd_iso_error = CdlIsoSeekError;
 		return -1;
@@ -168,7 +162,7 @@ static int _CdReadIsoDirectory(int lba)
 	CdRead(1, (uint32_t*)_cd_iso_directory_buff, CdlModeSpeed);
 	if( CdReadSync(0, 0) )
 	{
-		_LOG("psxcd: Error reading initial directory record.\n");
+		_sdk_log("psxcd: Error reading initial directory record.\n");
 
 		_cd_iso_error = CdlIsoReadError;
 		return -1;
@@ -177,14 +171,14 @@ static int _CdReadIsoDirectory(int lba)
 	direntry = (ISO_DIR_ENTRY*)_cd_iso_directory_buff;
 	_cd_iso_directory_len = direntry->entrySize.lsb;
 
-	_LOG("psxcd: Location of directory record = %d\n", direntry->entryOffs.lsb);
-	_LOG("psxcd: Size of directory record = %d\n", _cd_iso_directory_len);
+	_sdk_log("psxcd: Location of directory record = %d\n", direntry->entryOffs.lsb);
+	_sdk_log("psxcd: Size of directory record = %d\n", _cd_iso_directory_len);
 
 	if( _cd_iso_directory_len > 2048 )
 	{
 		if( !CdControl(CdlSetloc, (uint8_t*)&loc, 0) )
 		{
-			_LOG("psxcd: Could not set seek destination.\n");
+			_sdk_log("psxcd: Could not set seek destination.\n");
 
 			_cd_iso_error = CdlIsoSeekError;
 			return -1;
@@ -194,12 +188,12 @@ static int _CdReadIsoDirectory(int lba)
 		i = ((2047+_cd_iso_directory_len)>>11)<<11;
 		_cd_iso_directory_buff = (uint8_t*)malloc(i);
 
-		_LOG("psxcd: Allocated %d bytes for directory record.\n", i);
+		_sdk_log("psxcd: Allocated %d bytes for directory record.\n", i);
 
 		CdRead(i>>11, (uint32_t*)_cd_iso_directory_buff, CdlModeSpeed);
 		if( CdReadSync(0, 0) )
 		{
-			_LOG("psxcd: Error reading remaining directory record.\n");
+			_sdk_log("psxcd: Error reading remaining directory record.\n");
 
 			_cd_iso_error = CdlIsoReadError;
 			return -1;
@@ -221,7 +215,7 @@ static void dump_directory(void)
 	ISO_DIR_ENTRY *dir_entry;
 	char namebuff[16];
 	
-	_LOG("psxcd: Cached directory record contents:\n");
+	_sdk_log("psxcd: Cached directory record contents:\n");
 	
 	i = 0;
 	dir_pos = 0;
@@ -232,7 +226,7 @@ static void dump_directory(void)
 		strncpy(namebuff, 
 			_cd_iso_directory_buff+dir_pos+sizeof(ISO_DIR_ENTRY), dir_entry->identifierLen);
 			
-		_LOG("psxcd: P:%d L:%d %s\n", dir_pos, dir_entry->identifierLen, namebuff);
+		_sdk_log("psxcd: P:%d L:%d %s\n", dir_pos, dir_entry->identifierLen, namebuff);
 		
 		dir_pos += dir_entry->entryLength;
 		i++;
@@ -251,7 +245,7 @@ static void dump_directory(void)
 		}
 	}
 	
-	_LOG("psxcd: --\n");
+	_sdk_log("psxcd: --\n");
 	
 }
 
@@ -262,7 +256,7 @@ static void dump_pathtable(void)
 	ISO_DESCRIPTOR *descriptor;
 	char namebuff[16];
 	
-	_LOG("psxcd: Path table entries:\n");
+	_sdk_log("psxcd: Path table entries:\n");
 	
 	descriptor = (ISO_DESCRIPTOR*)_cd_iso_descriptor_buff;
 	
@@ -276,7 +270,7 @@ static void dump_pathtable(void)
 			tbl_pos+sizeof(ISO_PATHTABLE_ENTRY), 
 			tbl_entry->nameLength);
 		
-		_LOG("psxcd: %s\n", namebuff);
+		_sdk_log("psxcd: %s\n", namebuff);
 		
 		// Advance to next entry
 		tbl_pos += sizeof(ISO_PATHTABLE_ENTRY)
@@ -372,7 +366,7 @@ static int find_dir_entry(const char *name, ISO_DIR_ENTRY *dirent)
 	ISO_DIR_ENTRY *dir_entry;
 	char namebuff[16];
 
-	_LOG("psxcd: Locating file %s.\n", name);
+	_sdk_log("psxcd: Locating file %s.\n", name);
 
 	i = 0;
 	dir_pos = 0;
@@ -465,11 +459,11 @@ CdlFILE *CdSearchFile(CdlFILE *fp, const char *filename)
 		// Read ISO descriptor and path table
 	if( _CdReadIsoDescriptor(0) )
 	{
-		_LOG("psxcd: Could not read ISO file system.\n");
+		_sdk_log("psxcd: Could not read ISO file system.\n");
 		return NULL;
 	}
 
-	//	_LOG("psxcd: ISO file system cache updated.\n");
+	//	_sdk_log("psxcd: ISO file system cache updated.\n");
 	//	_cd_media_changed = 0;
 	//}
 	
@@ -477,23 +471,23 @@ CdlFILE *CdSearchFile(CdlFILE *fp, const char *filename)
 	num_dirs = get_pathtable_entry(0, NULL, NULL);
 	
 #ifndef NDEBUG
-	_LOG("psxcd: Directories in path table: %d\n", num_dirs);
+	_sdk_log("psxcd: Directories in path table: %d\n", num_dirs);
 	
 	rbuff = resolve_pathtable_path(num_dirs-1, tpath_rbuff+127);
 
 	if( !rbuff )
 	{
-		_LOG("psxcd: Could not resolve path.\n");
+		_sdk_log("psxcd: Could not resolve path.\n");
 	}
 	else
 	{
-		_LOG("psxcd: Longest path: %s|\n", rbuff);
+		_sdk_log("psxcd: Longest path: %s|\n", rbuff);
 	}
 #endif
 	
 	if( get_pathname(search_path, filename) )
 	{
-		_LOG("psxcd: Search path = %s|\n", search_path);
+		_sdk_log("psxcd: Search path = %s|\n", search_path);
 	}
 	
 	// Search the pathtable for a matching path
@@ -501,7 +495,7 @@ CdlFILE *CdSearchFile(CdlFILE *fp, const char *filename)
 	for(i=1; i<num_dirs; i++)
 	{
 		rbuff = resolve_pathtable_path(i, tpath_rbuff+127);
-		_LOG("psxcd: Found = %s|\n", rbuff);
+		_sdk_log("psxcd: Found = %s|\n", rbuff);
 
 		if( rbuff )
 		{
@@ -515,14 +509,14 @@ CdlFILE *CdSearchFile(CdlFILE *fp, const char *filename)
 	
 	if( !found_dir )
 	{
-		_LOG("psxcd: Directory path not found.\n");
+		_sdk_log("psxcd: Directory path not found.\n");
 		return NULL;
 	}
 
-	_LOG("psxcd: Found directory at record %d!\n", found_dir);
+	_sdk_log("psxcd: Found directory at record %d!\n", found_dir);
 
 	get_pathtable_entry(found_dir, &tbl_entry, NULL);
-	_LOG("psxcd: Directory LBA = %d\n", tbl_entry.dirOffs);
+	_sdk_log("psxcd: Directory LBA = %d\n", tbl_entry.dirOffs);
 
 	_CdReadIsoDirectory(tbl_entry.dirOffs);
 	get_filename(fp->name, filename);
@@ -539,12 +533,12 @@ CdlFILE *CdSearchFile(CdlFILE *fp, const char *filename)
 	
 	if( find_dir_entry(fp->name, &dir_entry) )
 	{
-		_LOG("psxcd: Could not find file.\n");
+		_sdk_log("psxcd: Could not find file.\n");
 
 		return NULL;
 	}
 
-	_LOG("psxcd: Located file at LBA %d.\n", dir_entry.entryOffs.lsb);
+	_sdk_log("psxcd: Located file at LBA %d.\n", dir_entry.entryOffs.lsb);
 
 	CdIntToPos(dir_entry.entryOffs.lsb, &fp->pos);
 	fp->size = dir_entry.entrySize.lsb;
@@ -568,11 +562,11 @@ CdlDIR *CdOpenDir(const char* path)
 	// Read ISO descriptor and path table
 	if( _CdReadIsoDescriptor( 0 ) )
 	{
-		_LOG( "psxcd: Could not read ISO file system.\n" );
+		_sdk_log( "psxcd: Could not read ISO file system.\n" );
 		return NULL;
 	}
 
-//		_LOG( "psxcd: ISO file system cache updated.\n" );
+//		_sdk_log( "psxcd: ISO file system cache updated.\n" );
 //		_cd_media_changed = 0;
 //	}
 	
@@ -582,7 +576,7 @@ CdlDIR *CdOpenDir(const char* path)
 	for( i=1; i<num_dirs; i++ )
 	{
 		rbuff = resolve_pathtable_path( i, tpath_rbuff+127 );
-		_LOG( "psxcd: Found = %s|\n", rbuff );
+		_sdk_log( "psxcd: Found = %s|\n", rbuff );
 
 		if( rbuff )
 		{
@@ -596,14 +590,14 @@ CdlDIR *CdOpenDir(const char* path)
 	
 	if( !found_dir )
 	{
-		_LOG( "psxcd: Directory path not found.\n" );
+		_sdk_log( "psxcd: Directory path not found.\n" );
 		return NULL;
 	}
 
-	_LOG( "psxcd: Found directory at record %d!\n", found_dir );
+	_sdk_log( "psxcd: Found directory at record %d!\n", found_dir );
 
 	get_pathtable_entry( found_dir, &tbl_entry, NULL );
-	_LOG( "psxcd: Directory LBA = %d\n", tbl_entry.dirOffs );
+	_sdk_log( "psxcd: Directory LBA = %d\n", tbl_entry.dirOffs );
 
 	_CdReadIsoDirectory( tbl_entry.dirOffs );
 
@@ -668,11 +662,11 @@ int CdReadDir(CdlDIR *dir, CdlFILE* file)
 	
 	file->size = dir_entry->entrySize.lsb;
 
-	_LOG("psxcd: dir_entry->entryLength = %d, ", dir_entry->entryLength);
+	_sdk_log("psxcd: dir_entry->entryLength = %d, ", dir_entry->entryLength);
 
 	d_dir->_pos += dir_entry->entryLength;
 
-	_LOG("psxcd: d_dir->_pos = %d\n", d_dir->_pos);
+	_sdk_log("psxcd: d_dir->_pos = %d\n", d_dir->_pos);
 
 	// Check if padding is reached (end of record sector)
 	if( d_dir->_dir[d_dir->_pos] == 0 )
@@ -776,15 +770,13 @@ int CdLoadSession(int session)
 	int i;
 
 	// Seek to specified session	
-	_LOG("psxcd: CdLoadSession(): Seeking to session %d...\n", session);
+	_sdk_log("psxcd: CdLoadSession(): Seeking to session %d...\n", session);
 	CdControl(CdlSetsession, (unsigned char*)&session, 
 		(unsigned char*)&resultbuff);
 	
 	if( CdSync(0, 0) == CdlDiskError )
 	{
-		_LOG("psxcd: CdLoadSession(): Session seek failed, "
-			"session does not exist.\n");
-		_LOG("psxcd: CdLoadSession(): Restarting CD-ROM...\n");
+		_sdk_log("psxcd: CdLoadSession(): Session seek failed, session does not exist. Restarting CD-ROM...\n");
 
 		// Restart CD-ROM on session seek failure
 		CdControl(CdlNop, 0, 0);
@@ -805,7 +797,7 @@ int CdLoadSession(int session)
 	_ses_scanbuff = scanbuff;
 	
 	// Begin scan for an ISO volume descriptor
-	_LOG("psxcd: CdLoadSession(): Scanning for ISO9660 volume descriptor.\n");
+	_sdk_log("psxcd: CdLoadSession(): Scanning for ISO9660 volume descriptor.\n");
 
 	i = CdlModeSpeed;
 	CdControl(CdlSetmode, (unsigned char*)&i, 0);
@@ -820,7 +812,7 @@ int CdLoadSession(int session)
 
 	if( !_ses_scanfound )
 	{
-		_LOG("psxcd: CdLoadSession(): Did not find volume descriptor.\n");
+		_sdk_log("psxcd: CdLoadSession(): Did not find volume descriptor.\n");
 
 		_cd_iso_error = CdlIsoInvalidFs;
 		EnterCriticalSection();
@@ -849,11 +841,11 @@ int CdLoadSession(int session)
 	
 	loc = (CdlLOC*)resultbuff;
 
-	_LOG("psxcd: CdLoadSession(): Session found in %02d:%02d:%02d (LBA=%d)\n",
+	_sdk_log("psxcd: CdLoadSession(): Session found in %02d:%02d:%02d (LBA=%d)\n",
 		btoi(loc->minute), btoi(loc->second), btoi(loc->sector), CdPosToInt(loc));
 
 	i = CdPosToInt(loc)-17;
-	_LOG("psxcd: CdLoadSession(): Session starting at LBA=%d\n", i);
+	_sdk_log("psxcd: CdLoadSession(): Session starting at LBA=%d\n", i);
 
 	_cd_media_changed = 1;
 	
diff --git a/libpsn00b/psxcd/psxcd.c b/libpsn00b/psxcd/psxcd.c
index 6730531..b914b5e 100644
--- a/libpsn00b/psxcd/psxcd.c
+++ b/libpsn00b/psxcd/psxcd.c
@@ -1,6 +1,6 @@
 #include <stdint.h>
-#include <stdio.h>
 #include <psxgpu.h>
+#include <psxetc.h>
 #include <psxapi.h>
 #include "psxcd.h"
 
@@ -21,12 +21,6 @@ volatile int _cd_last_sector_count;
 
 int _cd_media_changed;
 
-#ifdef NDEBUG
-#define _LOG(...)
-#else
-#define _LOG(...) printf(__VA_ARGS__)
-#endif
-
 void _cd_init(void);
 void _cd_control(unsigned char com, const void *param, int plen);
 void _cd_wait_ack(void);
@@ -45,9 +39,9 @@ int CdInit(void) {
 	
 	if(CdSync(0, 0) != CdlDiskError) {
 		CdControl(CdlDemute, 0, 0);
-		_LOG("psxcd: setup done\n");
+		_sdk_log("psxcd: setup done\n");
 	} else {
-		_LOG("psxcd: setup error, bad disc/drive or no disc inserted\n");
+		_sdk_log("psxcd: setup error, bad disc/drive or no disc inserted\n");
 	}
 	
 	return 1;
@@ -311,7 +305,7 @@ static void CdDoRetry()
 {
 	int cb;
 	
-	_LOG("psxcd: retrying read...\n");
+	_sdk_log("psxcd: retrying read...\n");
 	
 	// Stop reading
 	CdControl(CdlPause, 0, 0);
diff --git a/libpsn00b/psxetc/dl.c b/libpsn00b/psxetc/dl.c
index cf4e466..b85a7df 100644
--- a/libpsn00b/psxetc/dl.c
+++ b/libpsn00b/psxetc/dl.c
@@ -30,6 +30,7 @@
 #include <elf.h>
 #include <dlfcn.h>
 #include <string.h>
+#include <psxetc.h>
 #include <psxapi.h>
 
 /* Compile options */
@@ -66,12 +67,6 @@ void *(*_dl_resolve_callback)(DLL *, const char *) = 0;
 
 /* Private utilities */
 
-#ifdef NDEBUG
-#define _LOG(...)
-#else
-#define _LOG(...) printf(__VA_ARGS__)
-#endif
-
 #define _ERROR(code, ret) { \
 	_error_code = code; \
 	return ret; \
@@ -92,7 +87,7 @@ void *_dl_resolve_helper(DLL *dll, uint32_t index) {
 		address = DL_GetSymbolByName(_name);
 
 	if (!address) {
-		_LOG("psxetc: FATAL! can't resolve %s, locking up\n", _name);
+		_sdk_log("psxetc: FATAL! can't resolve %s, locking up\n", _name);
 		while (1)
 			__asm__ volatile("nop");
 	}
@@ -133,7 +128,7 @@ static uint32_t _elf_hash(const char *str) {
 static uint8_t *_dl_load_file(const char *filename, size_t *size_output) {
 	int32_t fd = open(filename, 1);
 	if (fd < 0) {
-		_LOG("psxetc: can't open %s, error = %d\n", filename, fd);
+		_sdk_log("psxetc: can't open %s, error = %d\n", filename, fd);
 		_ERROR(RTLD_E_FILE_OPEN, 0);
 	}
 
@@ -144,11 +139,11 @@ static uint8_t *_dl_load_file(const char *filename, size_t *size_output) {
 
 	uint8_t *buffer = malloc(size);
 	if (!buffer) {
-		_LOG("psxetc: unable to allocate %d bytes for %s\n", size, filename);
+		_sdk_log("psxetc: unable to allocate %d bytes for %s\n", size, filename);
 		_ERROR(RTLD_E_FILE_ALLOC, 0);
 	}
 
-	//_LOG("psxetc: loading %s (%d bytes)..", filename, size);
+	//_sdk_log("psxetc: loading %s (%d bytes)..", filename, size);
 
 	for (uint32_t offset = 0; offset < size; ) {
 		int32_t length = read(fd, &(buffer[offset]), 0x800);
@@ -157,16 +152,16 @@ static uint8_t *_dl_load_file(const char *filename, size_t *size_output) {
 			close(fd);
 			free(buffer);
 
-			_LOG("failed, error = %d\n", length);
+			_sdk_log("failed, error = %d\n", length);
 			_ERROR(RTLD_E_FILE_READ, 0);
 		}
 
-		//_LOG(".");
+		//_sdk_log(".");
 		offset += length;
 	}
 
 	close(fd);
-	_LOG(" done\n");
+	_sdk_log(" done\n");
 
 	if (size_output)
 		*size_output = size;
@@ -192,7 +187,7 @@ int32_t DL_ParseSymbolMap(const char *ptr, size_t size) {
 	// in order to minimize hash table size
 	_symbol_map.nbucket = entries;
 	_symbol_map.nchain  = entries;
-	_LOG(
+	_sdk_log(
 		"psxetc: allocating nbucket = %d, nchain = %d\n",
 		_symbol_map.nbucket,
 		entries
@@ -205,7 +200,7 @@ int32_t DL_ParseSymbolMap(const char *ptr, size_t size) {
 	_symbol_map.chain   = malloc(sizeof(uint32_t) * entries);
 
 	if (!_symbol_map.entries || !_symbol_map.bucket || !_symbol_map.chain) {
-		_LOG("psxetc: unable to allocate symbol map table\n");
+		_sdk_log("psxetc: unable to allocate symbol map table\n");
 		_ERROR(RTLD_E_MAP_ALLOC, -1);
 	}
 
@@ -248,7 +243,7 @@ int32_t DL_ParseSymbolMap(const char *ptr, size_t size) {
 				(_type == 'D') || // .data
 				(_type == 'B')    // .bss
 			)) {
-				//_LOG(
+				//_sdk_log(
 					//"psxetc: map sym: %08x,%08x [%c %s]\n",
 					//address, _size, _type, name
 				//);
@@ -274,7 +269,7 @@ int32_t DL_ParseSymbolMap(const char *ptr, size_t size) {
 			pos++;
 	}
 
-	_LOG("psxetc: parsed %d symbols\n", entries);
+	_sdk_log("psxetc: parsed %d symbols\n", entries);
 	if (!entries)
 		_ERROR(RTLD_E_NO_SYMBOLS, -1);
 
@@ -307,7 +302,7 @@ void DL_UnloadSymbolMap(void) {
 
 void *DL_GetSymbolByName(const char *name) {
 	if (!_symbol_map.entries) {
-		_LOG("psxetc: attempted lookup with no map loaded\n");
+		_sdk_log("psxetc: attempted lookup with no map loaded\n");
 		_ERROR(RTLD_E_NO_MAP, 0);
 	}
 
@@ -319,7 +314,7 @@ void *DL_GetSymbolByName(const char *name) {
 	// calculated.
 	for (uint32_t i = _symbol_map.bucket[hash_mod]; i != 0xffffffff;) {
 		if (i >= _symbol_map.nchain) {
-			_LOG(
+			_sdk_log(
 				"psxetc: GetSymbolByName() index out of bounds (%d >= %d)\n",
 				i, _symbol_map.nchain
 			);
@@ -329,14 +324,14 @@ void *DL_GetSymbolByName(const char *name) {
 		MapEntry *entry = &(_symbol_map.entries[i]);
 
 		if (hash == entry->hash) {
-			//_LOG("psxetc: map lookup [%s = %08x]\n", name, entry->ptr);
+			//_sdk_log("psxetc: map lookup [%s = %08x]\n", name, entry->ptr);
 			return entry->ptr;
 		}
 
 		i = _symbol_map.chain[i];
 	}
 
-	_LOG("psxetc: map lookup [%s not found]\n", name);
+	_sdk_log("psxetc: map lookup [%s not found]\n", name);
 	_ERROR(RTLD_E_MAP_SYMBOL, 0);
 }
 
@@ -352,14 +347,14 @@ DLL *DL_CreateDLL(void *ptr, size_t size, DL_ResolveMode mode) {
 
 	DLL *dll = malloc(sizeof(DLL));
 	if (!dll) {
-		_LOG("psxetc: unable to allocate DLL struct\n");
+		_sdk_log("psxetc: unable to allocate DLL struct\n");
 		_ERROR(RTLD_E_DLL_ALLOC, 0);
 	}
 
 	dll->ptr        = ptr;
 	dll->malloc_ptr = (mode & RTLD_FREE_ON_DESTROY) ? ptr : 0;
 	dll->size       = size;
-	_LOG("psxetc: initializing DLL at %08x\n", ptr);
+	_sdk_log("psxetc: initializing DLL at %08x\n", ptr);
 
 	// Interpret the key-value pairs in the .dynamic section to obtain info
 	// about all the other sections. The pairs are null-terminated, which makes
@@ -368,128 +363,128 @@ DLL *DL_CreateDLL(void *ptr, size_t size, DL_ResolveMode mode) {
 	uint32_t first_got_sym = 0;
 
 	for (Elf32_Dyn *dyn = (Elf32_Dyn *) ptr; dyn->d_tag; dyn++) {
-		//_LOG("psxetc: .dynamic %08x=%08x ", dyn->d_tag, dyn->d_un.d_val);
+		//_sdk_log("psxetc: .dynamic %08x=%08x ", dyn->d_tag, dyn->d_un.d_val);
 
 		switch (dyn->d_tag) {
 			// Offset of .got section
 			case DT_PLTGOT:
-				//_LOG("[PLTGOT]\n");
+				//_sdk_log("[PLTGOT]\n");
 
 				dll->got = (void *) (ptr + dyn->d_un.d_val);
 				break;
 
 			// Offset of .hash section
 			case DT_HASH:
-				//_LOG("[HASH]\n");
+				//_sdk_log("[HASH]\n");
 
 				dll->hash = (void *) (ptr + dyn->d_un.d_val);
 				break;
 
 			// Offset of .dynstr (NOT .strtab) section
 			case DT_STRTAB:
-				//_LOG("[STRTAB]\n");
+				//_sdk_log("[STRTAB]\n");
 
 				dll->strtab = (void *) (ptr + dyn->d_un.d_val);
 				break;
 
 			// Offset of .dynsym (NOT .symtab) section
 			case DT_SYMTAB:
-				//_LOG("[SYMTAB]\n");
+				//_sdk_log("[SYMTAB]\n");
 
 				dll->symtab = (void *) (ptr + dyn->d_un.d_val);
 				break;
 
 			// Length of .dynstr section
 			//case DT_STRSZ:
-				//_LOG("[STRSZ]\n");
+				//_sdk_log("[STRSZ]\n");
 				//break;
 
 			// Length of each .dynsym entry
 			case DT_SYMENT:
-				//_LOG("[SYMENT]\n");
+				//_sdk_log("[SYMENT]\n");
 
 				// Only 16-byte symbol table entries are supported.
 				if (dyn->d_un.d_val != sizeof(Elf32_Sym)) {
 					free(dll);
 
-					_LOG("psxetc: invalid DLL symtab entry size %d\n", dyn->d_un.d_val);
+					_sdk_log("psxetc: invalid DLL symtab entry size %d\n", dyn->d_un.d_val);
 					_ERROR(RTLD_E_DLL_FORMAT, 0);
 				}
 				break;
 
 			// MIPS ABI (?) version
 			case DT_MIPS_RLD_VERSION:
-				//_LOG("[MIPS_RLD_VERSION]\n");
+				//_sdk_log("[MIPS_RLD_VERSION]\n");
 
 				// Versions other than 1 are unsupported (do they even exist?).
 				if (dyn->d_un.d_val != 1) {
 					free(dll);
 
-					_LOG("psxetc: invalid DLL version %d\n", dyn->d_un.d_val);
+					_sdk_log("psxetc: invalid DLL version %d\n", dyn->d_un.d_val);
 					_ERROR(RTLD_E_DLL_FORMAT, 0);
 				}
 				break;
 
 			// DLL/ABI flags
 			case DT_MIPS_FLAGS:
-				//_LOG("[MIPS_FLAGS]\n");
+				//_sdk_log("[MIPS_FLAGS]\n");
 
 				// Shortcut pointers (whatever they are) are not supported.
 				if (dyn->d_un.d_val & RHF_QUICKSTART) {
 					free(dll);
 
-					_LOG("psxetc: invalid DLL flags\n");
+					_sdk_log("psxetc: invalid DLL flags\n");
 					_ERROR(RTLD_E_DLL_FORMAT, 0);
 				}
 				break;
 
 			// Number of local (not to resolve) GOT entries
 			case DT_MIPS_LOCAL_GOTNO:
-				//_LOG("[MIPS_LOCAL_GOTNO]\n");
+				//_sdk_log("[MIPS_LOCAL_GOTNO]\n");
 
 				local_got_len = dyn->d_un.d_val;
 				break;
 
 			// Base address DLL was compiled for
 			case DT_MIPS_BASE_ADDRESS:
-				//_LOG("[MIPS_BASE_ADDRESS]\n");
+				//_sdk_log("[MIPS_BASE_ADDRESS]\n");
 
 				// Base addresses other than zero are not supported. It would
 				// be easy enough to support them, but why?
 				if (dyn->d_un.d_val) {
 					free(dll);
 
-					_LOG("psxetc: invalid DLL base address %08x\n", dyn->d_un.d_val);
+					_sdk_log("psxetc: invalid DLL base address %08x\n", dyn->d_un.d_val);
 					_ERROR(RTLD_E_DLL_FORMAT, 0);
 				}
 				break;
 
 			// Number of symbol table entries
 			case DT_MIPS_SYMTABNO:
-				//_LOG("[MIPS_SYMTABNO]\n");
+				//_sdk_log("[MIPS_SYMTABNO]\n");
 
 				dll->symbol_count = dyn->d_un.d_val;
 				break;
 
 			// Index of first unresolved symbol table entry
 			//case DT_MIPS_UNREFEXTNO:
-				//_LOG("[MIPS_UNREFEXTNO]\n");
+				//_sdk_log("[MIPS_UNREFEXTNO]\n");
 				//break;
 
 			// Index of first symbol table entry which has a matching GOT entry
 			case DT_MIPS_GOTSYM:
-				//_LOG("[MIPS_GOTSYM]\n");
+				//_sdk_log("[MIPS_GOTSYM]\n");
 
 				first_got_sym = dyn->d_un.d_val;
 				break;
 
 			// Number of pages the GOT is split into (does not apply to PS1)
 			//case DT_MIPS_HIPAGENO:
-				//_LOG("[MIPS_HIPAGENO]\n");
+				//_sdk_log("[MIPS_HIPAGENO]\n");
 				//break;
 
 			//default:
-				//_LOG("[ignored]\n");
+				//_sdk_log("[ignored]\n");
 		}
 	}
 
@@ -501,7 +496,7 @@ DLL *DL_CreateDLL(void *ptr, size_t size, DL_ResolveMode mode) {
 		((uint32_t) ptr + size - (uint32_t) dll->got) / sizeof(uint32_t) - 2;
 
 	dll->got_length = local_got_len + (dll->symbol_count - first_got_sym) - 2;
-	_LOG(
+	_sdk_log(
 		"psxetc: %d symbols, %d GOT entries\n",
 		dll->symbol_count, dll->got_length
 	);
@@ -530,7 +525,7 @@ DLL *DL_CreateDLL(void *ptr, size_t size, DL_ResolveMode mode) {
 			continue;
 
 		sym->st_value += (uint32_t) ptr;
-		//_LOG(
+		//_sdk_log(
 			//"psxetc: DLL sym: %08x,%08x [%s]\n",
 			//sym->st_value, sym->st_size, _name
 		//);
@@ -641,7 +636,7 @@ void *DL_GetDLLSymbol(const DLL *dll, const char *name) {
 	// provided.
 	for (uint32_t i = bucket[hash_mod]; i != 0xffffffff;) {
 		if (i >= nchain) {
-			_LOG("psxetc: DL_GetDLLSymbol() index out of bounds (%d >= %d)\n", i, nchain);
+			_sdk_log("psxetc: DL_GetDLLSymbol() index out of bounds (%d >= %d)\n", i, nchain);
 			_ERROR(RTLD_E_HASH_LOOKUP, 0);
 		}
 
@@ -649,14 +644,14 @@ void *DL_GetDLLSymbol(const DLL *dll, const char *name) {
 		const char *_name = &(dll->strtab[sym->st_name]);
 
 		if (!strcmp(name, _name)) {
-			//_LOG("psxetc: DLL lookup [%s = %08x]\n", name, sym->st_value);
+			//_sdk_log("psxetc: DLL lookup [%s = %08x]\n", name, sym->st_value);
 			return sym->st_value;
 		}
 
 		i = chain[i];
 	}
 
-	_LOG("psxetc: DLL lookup [%s not found]\n", name);
+	_sdk_log("psxetc: DLL lookup [%s not found]\n", name);
 	_ERROR(RTLD_E_DLL_SYMBOL, 0);
 }
 
diff --git a/libpsn00b/psxetc/logging.c b/libpsn00b/psxetc/logging.c
new file mode 100644
index 0000000..5199190
--- /dev/null
+++ b/libpsn00b/psxetc/logging.c
@@ -0,0 +1,50 @@
+/*
+ * PSn00bSDK internal debug logger
+ * (C) 2022 spicyjpeg - MPL licensed
+ *
+ * This file provides the (admittedly minimal) logging system used by all
+ * PSn00bSDK libraries. Log messages and warnings are issued using the
+ * _sdk_log() macro and collected into a buffer, whose contents can be flushed
+ * by calling _sdk_dump_log() (by default this is done by VSync()). Logging is
+ * only enabled in debug builds of libpsn00b.
+ */
+
+#include <stddef.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <psxapi.h>
+#include <psxetc.h>
+
+#define LOG_BUFFER_SIZE		256
+
+#ifndef NDEBUG
+
+/* Internal globals */
+
+static char   _log_buffer[LOG_BUFFER_SIZE];
+static size_t _log_buffer_length = 0;
+
+/* Internal logging API */
+
+void _sdk_log_inner(const char *fmt, ...) {
+	va_list ap;
+
+	va_start(ap, fmt);
+	_log_buffer_length += vsnprintf(
+		&_log_buffer[_log_buffer_length],
+		LOG_BUFFER_SIZE - _log_buffer_length,
+		fmt,
+		ap
+	);
+	va_end(ap);
+}
+
+void _sdk_dump_log_inner(void) {
+	if (!_log_buffer_length)
+		return;
+
+	write(1, _log_buffer, _log_buffer_length);
+	_log_buffer_length = 0;
+}
+
+#endif
diff --git a/libpsn00b/psxgpu/common.c b/libpsn00b/psxgpu/common.c
index bf70b72..a262472 100644
--- a/libpsn00b/psxgpu/common.c
+++ b/libpsn00b/psxgpu/common.c
@@ -4,7 +4,6 @@
  */
 
 #include <stdint.h>
-#include <stdio.h>
 #include <psxetc.h>
 #include <psxapi.h>
 #include <psxgpu.h>
@@ -36,13 +35,7 @@ static volatile uint8_t    _queue_head, _queue_tail, _queue_length;
 static volatile uint32_t   _vblank_counter;
 static volatile uint16_t   _last_hblank;
 
-/* Private utilities and interrupt handlers */
-
-#ifdef NDEBUG
-#define _LOG(...)
-#else
-#define _LOG(...) printf(__VA_ARGS__)
-#endif
+/* Private interrupt handlers */
 
 static void _vblank_handler(void) {
 	_vblank_counter++;
@@ -57,8 +50,8 @@ static void _gpu_dma_handler(void) {
 		__asm__ volatile("");
 
 	if (--_queue_length) {
-		QueueEntry *entry = &_draw_queue[_queue_head++];
-		_queue_head      %= QUEUE_LENGTH;
+		volatile QueueEntry *entry = &_draw_queue[_queue_head++];
+		_queue_head %= QUEUE_LENGTH;
 
 		entry->func(entry->arg1, entry->arg2, entry->arg3);
 	} else {
@@ -82,7 +75,7 @@ void ResetGraph(int mode) {
 		_gpu_video_mode = (GPU_GP1 >> 20) & 1;
 		ExitCriticalSection();
 
-		_LOG("psxgpu: setup done, default mode is %s\n", _gpu_video_mode ? "PAL" : "NTSC");
+		_sdk_log("psxgpu: setup done, default mode is %s\n", _gpu_video_mode ? "PAL" : "NTSC");
 	}
 
 	if (mode == 3) {
@@ -115,13 +108,13 @@ void ResetGraph(int mode) {
 // 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;
 	}
 
-	_LOG("psxgpu: VSync() timeout\n");
+	_sdk_log("psxgpu: VSync() timeout\n");
+	_sdk_dump_log();
 	ChangeClearPAD(0);
 	ChangeClearRCnt(3, 0);
 }
@@ -137,6 +130,7 @@ int VSync(int mode) {
 
 	// Wait for at least one vertical blank event to occur.
 	do {
+		_sdk_dump_log();
 		_vsync_halt_func();
 
 		// If interlaced mode is enabled, wait until the GPU starts displaying
@@ -190,15 +184,15 @@ int EnqueueDrawOp(
 	if (_queue_length) {
 		if (_queue_length >= QUEUE_LENGTH) {
 			IRQ_MASK = mask;
-			_LOG("psxgpu: draw queue overflow, dropping commands\n");
+			_sdk_log("psxgpu: draw queue overflow, dropping commands\n");
 			return -1;
 		}
 
 		int length    = _queue_length;
 		_queue_length = length + 1;
 
-		QueueEntry *entry = &_draw_queue[_queue_tail++];
-		_queue_tail      %= QUEUE_LENGTH;
+		volatile QueueEntry *entry = &_draw_queue[_queue_tail++];
+		_queue_tail %= QUEUE_LENGTH;
 
 		entry->func = func;
 		entry->arg1 = arg1;
@@ -236,7 +230,8 @@ int DrawSync(int mode) {
 		while (!(GPU_GP1 & (1 << 26)))
 			__asm__ volatile("");
 	} else {
-		printf("psxgpu: DrawSync() timeout\n");
+		_sdk_log("psxgpu: DrawSync() timeout\n");
+		_sdk_dump_log();
 	}
 
 	return _queue_length;
@@ -297,7 +292,7 @@ void DrawPrim(const uint32_t *pri) {
 }
 
 int DrawOTag(const uint32_t *ot) {
-	return EnqueueDrawOp(&DrawOTag2, (uint32_t) ot, 0, 0);
+	return EnqueueDrawOp((void *) &DrawOTag2, (uint32_t) ot, 0, 0);
 }
 
 void DrawOTag2(const uint32_t *ot) {
diff --git a/libpsn00b/psxgpu/env.c b/libpsn00b/psxgpu/env.c
index 1b97026..f513727 100644
--- a/libpsn00b/psxgpu/env.c
+++ b/libpsn00b/psxgpu/env.c
@@ -85,7 +85,7 @@ int DrawOTagEnv(const uint32_t *ot, DRAWENV *env) {
 	//while (!(GPU_GP1 & (1 << 26)))
 		//__asm__ volatile("");
 
-	return EnqueueDrawOp(&DrawOTag2, (uint32_t) prim, 0, 0);
+	return EnqueueDrawOp((void *) &DrawOTag2, (uint32_t) prim, 0, 0);
 }
 
 void PutDrawEnv(DRAWENV *env) {
diff --git a/libpsn00b/psxgpu/image.c b/libpsn00b/psxgpu/image.c
index c09a59d..968dde5 100644
--- a/libpsn00b/psxgpu/image.c
+++ b/libpsn00b/psxgpu/image.c
@@ -4,7 +4,7 @@
  */
 
 #include <stdint.h>
-#include <stdio.h>
+#include <psxetc.h>
 #include <psxgpu.h>
 #include <hwregs_c.h>
 
@@ -12,20 +12,14 @@
 
 /* Private utilities */
 
-#ifdef NDEBUG
-#define _LOG(...)
-#else
-#define _LOG(...) printf(__VA_ARGS__)
-#endif
-
 static void _dma_transfer(const RECT *rect, uint32_t *data, int write) {
 	size_t length = rect->w * rect->h;
 	if (length % 2)
-		_LOG("psxgpu: can't transfer an odd number of pixels\n");
+		_sdk_log("psxgpu: can't transfer an odd number of pixels\n");
 
 	length /= 2;
 	if ((length >= DMA_CHUNK_LENGTH) && (length % DMA_CHUNK_LENGTH)) {
-		_LOG("psxgpu: transfer data length (%d) is not a multiple of %d, rounding\n", length, DMA_CHUNK_LENGTH);
+		_sdk_log("psxgpu: transfer data length (%d) is not a multiple of %d, rounding\n", length, DMA_CHUNK_LENGTH);
 		length += DMA_CHUNK_LENGTH - 1;
 	}
 
@@ -53,15 +47,19 @@ static void _dma_transfer(const RECT *rect, uint32_t *data, int write) {
 /* VRAM transfer API */
 
 int LoadImage(const RECT *rect, const uint32_t *data) {
-	return EnqueueDrawOp(&_dma_transfer, (uint32_t) rect, (uint32_t) data, 1);
+	return EnqueueDrawOp(
+		(void *) &_dma_transfer, (uint32_t) rect, (uint32_t) data, 1
+	);
 }
 
 int StoreImage(const RECT *rect, uint32_t *data) {
-	return EnqueueDrawOp(&_dma_transfer, (uint32_t) rect, (uint32_t) data, 0);
+	return EnqueueDrawOp(
+		(void *) &_dma_transfer, (uint32_t) rect, (uint32_t) data, 0
+	);
 }
 
 int MoveImage(const RECT *rect, int x, int y) {
-	return EnqueueDrawOp(&MoveImage2, (uint32_t) rect, x, y);
+	return EnqueueDrawOp((void *) &MoveImage2, (uint32_t) rect, x, y);
 }
 
 void LoadImage2(const RECT *rect, const uint32_t *data) {
diff --git a/libpsn00b/psxpress/mdec.c b/libpsn00b/psxpress/mdec.c
index 06510cb..d43436f 100644
--- a/libpsn00b/psxpress/mdec.c
+++ b/libpsn00b/psxpress/mdec.c
@@ -4,7 +4,7 @@
  */
 
 #include <stdint.h>
-#include <stdio.h>
+#include <psxetc.h>
 #include <psxapi.h>
 #include <psxpress.h>
 #include <hwregs_c.h>
@@ -81,14 +81,6 @@ static const DECDCTENV _default_mdec_env = {
 	}
 };
 
-/* Private utilities */
-
-#ifdef NDEBUG
-#define _LOG(...)
-#else
-#define _LOG(...) printf(__VA_ARGS__)
-#endif
-
 /* Public API */
 
 void DecDCTReset(int mode) {
@@ -135,7 +127,7 @@ void DecDCTin(const uint32_t *data, int mode) {
 // the stream.
 void DecDCTinRaw(const uint32_t *data, size_t length) {
 	if ((length >= DMA_CHUNK_LENGTH) && (length % DMA_CHUNK_LENGTH)) {
-		_LOG("psxpress: transfer data length (%d) is not a multiple of %d, rounding\n", length, DMA_CHUNK_LENGTH);
+		_sdk_log("psxpress: transfer data length (%d) is not a multiple of %d, rounding\n", length, DMA_CHUNK_LENGTH);
 		length += DMA_CHUNK_LENGTH - 1;
 	}
 
@@ -157,7 +149,8 @@ int DecDCTinSync(int mode) {
 			return 0;
 	}
 
-	_LOG("psxpress: DecDCTinSync() timeout\n");
+	_sdk_log("psxpress: DecDCTinSync() timeout\n");
+	_sdk_dump_log();
 	return -1;
 }
 
@@ -165,7 +158,7 @@ void DecDCTout(uint32_t *data, size_t length) {
 	DecDCToutSync(0);
 
 	if ((length >= DMA_CHUNK_LENGTH) && (length % DMA_CHUNK_LENGTH)) {
-		_LOG("psxpress: transfer data length (%d) is not a multiple of %d, rounding\n", length, DMA_CHUNK_LENGTH);
+		_sdk_log("psxpress: transfer data length (%d) is not a multiple of %d, rounding\n", length, DMA_CHUNK_LENGTH);
 		length += DMA_CHUNK_LENGTH - 1;
 	}
 
@@ -187,6 +180,7 @@ int DecDCToutSync(int mode) {
 			return 0;
 	}
 
-	_LOG("psxpress: DecDCToutSync() timeout\n");
+	_sdk_log("psxpress: DecDCToutSync() timeout\n");
+	_sdk_dump_log();
 	return -1;
 }
diff --git a/libpsn00b/psxspu/common.c b/libpsn00b/psxspu/common.c
index 380bd3d..7d90858 100644
--- a/libpsn00b/psxspu/common.c
+++ b/libpsn00b/psxspu/common.c
@@ -4,7 +4,7 @@
  */
 
 #include <stdint.h>
-#include <stdio.h>
+#include <psxetc.h>
 #include <psxspu.h>
 #include <hwregs_c.h>
 
@@ -19,28 +19,22 @@ static uint16_t			_transfer_addr = WRITABLE_AREA_ADDR;
 
 /* Private utilities */
 
-#ifdef NDEBUG
-#define _LOG(...)
-#else
-#define _LOG(...) printf(__VA_ARGS__)
-#endif
-
 static void _wait_status(uint16_t mask, uint16_t value) {
 	for (int i = STATUS_TIMEOUT; i; i--) {
 		if ((SPU_STAT & mask) == value)
 			return;
 	}
 
-	_LOG("psxspu: status register timeout (0x%04x)\n", SPU_STAT);
+	_sdk_log("psxspu: status register timeout (0x%04x)\n", SPU_STAT);
 }
 
 static void _dma_transfer(uint32_t *data, size_t length, int write) {
 	if (length % 4)
-		_LOG("psxspu: can't transfer a number of bytes that isn't multiple of 4\n");
+		_sdk_log("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)) {
-		_LOG("psxspu: transfer data length (%d) is not a multiple of %d, rounding\n", length, DMA_CHUNK_LENGTH);
+		_sdk_log("psxspu: transfer data length (%d) is not a multiple of %d, rounding\n", length, DMA_CHUNK_LENGTH);
 		length += DMA_CHUNK_LENGTH - 1;
 	}
 
-- 
cgit v1.2.3