Add IsIdleGPU(), tweak psxgpu.h, fix mdec/strvideo

6 files changed, 78 insertions, 49 deletions

diff --git a/examples/demos/n00bdemo/main.c b/examples/demos/n00bdemo/main.c
index c9ca3ce..55dbbc4 100644
--- a/examples/demos/n00bdemo/main.c
+++ b/examples/demos/n00bdemo/main.c
@@ -358,10 +358,9 @@ void stencilstuff() {
 	/* 
 	The stencil demo is achieved by utilizing the mask bit setting
 	primitive GP0(E6h). The structure of this primitive is defined as
-	DR_MASK initialized and set by setDrawMask(). These are not available
-	in Sony's SDK by default.
+	DR_STP initialized and set by setDrawStp().
 	
-	The DR_MASK primitive controls mask bit operations for drawing
+	The DR_STP primitive controls mask bit operations for drawing
 	primitives such as setting mask bits on every pixel drawn or mask
 	bit test where pixels won't be drawn on pixels with the mask bit set.
 	It applies to most graphics drawing primitives except VRAM fill.
@@ -373,10 +372,10 @@ void stencilstuff() {
 	bit operation disabled.
 	
 	The stencil effect featured in this demo is achieved by enabling set
-	mask bit with DR_MASK, drawing semi-transparent primitives using
+	mask bit with DR_STP, drawing semi-transparent primitives using
 	additive blending but color is all zero to make it completely invisible
 	but is enough to update the mask bits, disable mask set bit but enable
-	mask test with DR_MASK and then drawing a rectangle that fills the
+	mask test with DR_STP and then drawing a rectangle that fills the
 	entire screen. Semi-transparency mask in textures must not be used when
 	drawing the scene that will be 'below' the mask layer.
 	*/
@@ -384,7 +383,7 @@ void stencilstuff() {
 	
 	int spin=0;
 	
-	DR_MASK *mask;
+	DR_STP *mask;
 	TILE *rect;
 	
 	SC_OT s_ot;
@@ -430,10 +429,10 @@ void stencilstuff() {
 		
 		
 		// Sort mask primitive that enables setting mask bits
-		mask = (DR_MASK*)nextpri;
-		setDrawMask( mask, 1, 0 );
+		mask = (DR_STP*)nextpri;
+		setDrawStp( mask, 1, 0 );
 		addPrim( ot[db]+20, mask );
-		nextpri += sizeof(DR_MASK);
+		nextpri += sizeof(DR_STP);
 		
 		
 		// Sort the stars
@@ -465,10 +464,10 @@ void stencilstuff() {
 		
 		
 		// Sort mask primitive that enables mask bit test
-		mask = (DR_MASK*)nextpri;
-		setDrawMask( mask, 0, 1 );
+		mask = (DR_STP*)nextpri;
+		setDrawStp( mask, 0, 1 );
 		addPrim( ot[db]+18, mask );
-		nextpri += sizeof(DR_MASK);
+		nextpri += sizeof(DR_STP);
 
 		
 		// Sort rectangle that fills the screen
@@ -482,10 +481,10 @@ void stencilstuff() {
 		
 		
 		// Clear all mask settings
-		mask = (DR_MASK*)nextpri;
-		setDrawMask( mask, 0, 0 );
+		mask = (DR_STP*)nextpri;
+		setDrawStp( mask, 0, 0 );
 		addPrim( ot[db]+15, mask );
-		nextpri += sizeof(DR_MASK);
+		nextpri += sizeof(DR_STP);
 		
 
 		// Sort overlay then display
diff --git a/examples/mdec/strvideo/main.c b/examples/mdec/strvideo/main.c
index 57cb6ef..853e0c2 100644
--- a/examples/mdec/strvideo/main.c
+++ b/examples/mdec/strvideo/main.c
@@ -189,8 +189,15 @@ void cd_sector_handler(void) {
 		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 (sector_header.frame_id != str_ctx.frame_id) {
+	// that have been read so far and flip the bitstream data buffers. If the
+	// frame number is actually lower than the current one, assume the drive
+	// has started reading another .STR file and stop playback.
+	if ((int) sector_header.frame_id < str_ctx.frame_id) {
+		str_ctx.frame_ready = -1;
+		return;
+	}
+
+	if ((int) sector_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++;
@@ -263,11 +270,9 @@ void init_stream(void) {
 	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);
+	// Copy the lookup table used for frame decompression to the scratchpad
+	// area. This is optional but makes the decompressor slightly faster. See
+	// the libpsxpress documentation for more details.
 	DecDCTvlcCopyTableV3((VLC_TableV3 *) 0x1f800000);
 
 	str_ctx.cur_frame = 0;
diff --git a/libpsn00b/include/psxgpu.h b/libpsn00b/include/psxgpu.h
index 2329908..78d8342 100644
--- a/libpsn00b/include/psxgpu.h
+++ b/libpsn00b/include/psxgpu.h
@@ -83,7 +83,7 @@ typedef enum _GPU_VideoMode {
 	(p)->u0 = (_u0), (p)->v0 = (_v0), \
 	(p)->u1 = (_u1), (p)->v1 = (_v1), \
 	(p)->u2 = (_u2), (p)->v2 = (_v2)
-	
+
 #define setUV4(p, _u0, _v0, _u1, _v1, _u2, _v2, _u3, _v3) \
 	(p)->u0 = (_u0), (p)->v0 = (_v0), \
 	(p)->u1 = (_u1), (p)->v1 = (_v1), \
@@ -202,38 +202,48 @@ typedef enum _GPU_VideoMode {
 
 #define setTexWindow_T(p, r) \
 	(p)->code[0] = (0xe2000000 | \
-		((r)->w  % 32) | \
-		(((r)->h % 32) <<  5) | \
-		(((r)->x % 32) << 10) | \
-		(((r)->y % 32) << 15) \
+		((r)->w  & 0x1f) | \
+		(((r)->h & 0x1f) <<  5) | \
+		(((r)->x & 0x1f) << 10) | \
+		(((r)->y & 0x1f) << 15) \
 	)
 #define setTexWindow(p, r) \
 	setlen(p, 1), setTexWindow_T(p, r)
 
-#define setDrawArea_T(p, r) \
+#define setDrawAreaXY_T(p, _x0, _y0, _x1, _y1) \
 	(p)->code[0] = (0xe3000000 | \
-		((r)->x  % 1024) | \
-		(((r)->y % 1024) << 10) \
+		((_x0)  & 0x3ff) | \
+		(((_y0) & 0x3ff) << 10) \
 	), \
 	(p)->code[1] = (0xe4000000 | \
-		(((r)->x  + (r)->w - 1) % 1024) | \
-		((((r)->y + (r)->h - 1) % 1024) << 10) \
+		((_x1)  & 0x3ff) | \
+		(((_y1) & 0x3ff) << 10) \
+	)
+#define setDrawAreaXY(p, _x0, _y0, _x1, _y1) \
+	setlen(p, 2), setDrawAreaXY_T(p, _x0, _y0, _x1, _y1)
+
+#define setDrawArea_T(p, r) \
+	setDrawAreaXY_T(p, \
+		(r)->x, \
+		(r)->y, \
+		(r)->x + (r)->w - 1, \
+		(r)->y + (r)->h - 1 \
 	)
 #define setDrawArea(p, r) \
 	setlen(p, 2), setDrawArea_T(p, r)
 
 #define setDrawOffset_T(p, _x, _y) \
 	(p)->code[0] = (0xe5000000 | \
-		((_x)  % 1024) | \
-		(((_y) % 1024) << 11) \
+		((_x)  & 0x7ff) | \
+		(((_y) & 0x7ff) << 11) \
 	)
 #define setDrawOffset(p, _x, _y) \
 	setlen(p, 1), setDrawOffset_T(p, _x, _y)
 
-#define setDrawMask_T(p, sb, mt) \
-	(p)->code[0] = (0xe6000000 | (sb) | ((mt) << 1))
-#define setDrawMask(p, sb, mt) \
-	setlen(p, 1), setDrawMask_T(p, sb, mt)
+#define setDrawStp_T(p, pbw, mt) \
+	(p)->code[0] = (0xe6000000 | (pbw) | ((mt) << 1))
+#define setDrawStp(p, pbw, mt) \
+	setlen(p, 1), setDrawStp_T(p, pbw, mt)
 
 /* Primitive structure definitions */
 
@@ -469,7 +479,7 @@ _DEF_PRIM(DR_TWIN,
 _DEF_PRIM(DR_TPAGE,
 	uint32_t code[1];
 )
-_DEF_PRIM(DR_MASK,
+_DEF_PRIM(DR_STP,
 	uint32_t code[1];
 )
 
@@ -481,6 +491,9 @@ _DEF_PRIM(DR_ENV,
 	FILL_T		fill;
 )
 
+#undef _DEF_PRIM
+#undef _DEF_ALIAS
+
 /* Structure definitions */
 
 typedef struct _RECT {
@@ -545,6 +558,7 @@ void PutDrawEnv(DRAWENV *env);
 void PutDrawEnvFast(DRAWENV *env);
 
 int GetODE(void);
+int IsIdleGPU(int timeout);
 int VSync(int mode);
 void *VSyncHaltFunction(void (*func)(void));
 void *VSyncCallback(void (*func)(void));
diff --git a/libpsn00b/include/psxpress.h b/libpsn00b/include/psxpress.h
index c3b13f4..ea0c2ec 100644
--- a/libpsn00b/include/psxpress.h
+++ b/libpsn00b/include/psxpress.h
@@ -246,8 +246,6 @@ int DecDCToutSync(int mode);
  * can be different). If max_size = 0, the entire frame will always be decoded
  * in one shot.
  *
- * Only bitstream version 2 is currently supported.
- *
  * WARNING: InitGeom() must be called prior to using DecDCTvlcStart() for the
  * first time. Attempting to call this function with the GTE disabled will
  * result in a crash.
@@ -411,7 +409,8 @@ void DecDCTvlcCopyTableV3(VLC_TableV3 *addr);
  * buffer can be different). If max_size = 0, the entire frame will always be
  * decoded in one shot.
  *
- * Only bitstream version 2 is currently supported.
+ * This function only supports decoding version 1 and 2 bitstreams. Use
+ * DecDCTvlcStart() to decode a version 3 bitstream.
  *
  * @param ctx Pointer to VLC_Context structure (which will be initialized)
  * @param buf
diff --git a/libpsn00b/psxgpu/common.c b/libpsn00b/psxgpu/common.c
index 537f672..e354261 100644
--- a/libpsn00b/psxgpu/common.c
+++ b/libpsn00b/psxgpu/common.c
@@ -321,3 +321,18 @@ void DrawOTag2(const uint32_t *ot) {
 	DMA_BCR(DMA_GPU)  = 0;
 	DMA_CHCR(DMA_GPU) = 0x01000401;
 }
+
+/* Queue pause/resume API */
+
+int IsIdleGPU(int timeout) {
+	if (timeout <= 0)
+		timeout = 1;
+
+	for (; timeout; timeout--) {
+		if (GPU_GP1 & (1 << 26))
+			return 0;
+	}
+
+	//_sdk_log("IsIdleGPU() timeout\n");
+	return -1;
+}
diff --git a/libpsn00b/psxpress/vlc2.c b/libpsn00b/psxpress/vlc2.c
index 24c54ce..7d9d9f3 100644
--- a/libpsn00b/psxpress/vlc2.c
+++ b/libpsn00b/psxpress/vlc2.c
@@ -123,7 +123,7 @@ int __attribute__((optimize(3))) DecDCTvlcContinue2(
 				*output = (uint16_t) _get_bits_unsigned(22);
 				_advance_window(22);
 			} else if (window >> 24) {
-				// The first lookup table is for codes that not start with
+				// The first lookup table is for codes that do not start with
 				// 00000000.
 				value = _vlc_huffman_table2->ac[_get_bits_unsigned(13)];
 				_advance_window(value >> 16);
@@ -136,12 +136,9 @@ int __attribute__((optimize(3))) DecDCTvlcContinue2(
 				*output = (uint16_t) value;
 			}
 		} else {
-			// Parse the DC (first) coefficient for this block. Version 2
-			// simply stores the signed 10-bit value as-is, while version 3
-			// uses a delta encoding combined with a compression method similar
-			// to exp-Golomb.
+			// Parse the DC (first) coefficient for this block.
 			if (is_v3) {
-				// TODO: version 3 is currently not supported.
+				// This implementation does not support version 3.
 				return -1;
 			} else {
 				value = _get_bits_unsigned(10);
@@ -161,7 +158,7 @@ int __attribute__((optimize(3))) DecDCTvlcContinue2(
 		// time and processes each 16-bit word starting from the the MSB, so an
 		// endianness conversion is necessary to preserve bit order when
 		// reading 32 bits at a time. Also note that the PS1 CPU is not capable
-		// of shifting by more than 31 bits - it will shift by 0 bits instead!
+		// of shifting by >=31 bits - it will shift by (N % 32) bits instead!
 		if (bit_offset < 0) {
 			window      = next_window << (-bit_offset);
 			bit_offset += 32;