sponk_psm4_mtba.cpp

// Tests MTBA with MXL 0 and MMIN 0
// (MTBA still works with these set to 0)
#include "defs.h"

#include <kernel.h>
#include <tamtypes.h>
#include <stdio.h>

#include <gif_tags.h>

#include <gs_gp.h>
#include <gs_psm.h>

#include <dma.h>
#include <dma_tags.h>

#include <draw.h>
#include <graph.h>

#include <packet.h>
#include <stdlib.h>
#include <fcntl.h>

#include "clut.c"
#include "tex4_1.c"
#include "tex4_2.c"
#include "tex4_3.c"
#include "tex4_4.c"

const u32 WIDTH = 640;
const u32 HEIGHT = 480;
const s32 FRAME_PSM = GS_PSM_32;
#undef GS_SET_MIPTBP1
#define GS_SET_MIPTBP1(TBA1, TBW1, TBA2, TBW2, TBA3, TBW3)                \
	(u64)((TBA1)&0x00003FFF) << 0 | (u64)((TBW1)&0x0000003F) << 14 |      \
		(u64)((TBA2)&0x00003FFF) << 20 | (u64)((TBW2)&0x0000003F) << 34 | \
		(u64)((TBA3)&0x00003FFF) << 40 | (u64)((TBW3)&0x0000003F) << 54
#define GS_SETREG_TEST(ate, atst, aref, afail, date, datm, zte, ztst) \
	((u64)(ate) | ((u64)(atst) << 1) |                                \
	 ((u64)(aref) << 4) | ((u64)(afail) << 12) |                      \
	 ((u64)(date) << 14) | ((u64)(datm) << 15) |                      \
	 ((u64)(zte) << 16) | ((u64)(ztst) << 17))
qword_t *draw_texture_transfer(qword_t *q, void *src, int width, int height, int psm, int dest, int dest_width, u32 dx, u32 dy);
int download_from_gs(u32 address, u32 format, const char *fname, int psm_size, int width, int height)
{
	u32 destBuffer[width * height];

	static union
	{
		u32 value_u32[4];
		u128 value;
	} enable_path3 ALIGNED(16) = {
		{VIF1_MSKPATH3(0), VIF1_NOP, VIF1_NOP, VIF1_NOP}};

	u32 prev_imr;
	u32 prev_chcr;

	u32 dmaChain[20 * 2] ALIGNED(16);
	u32 i = 0; // Used to index through the chain

	u32 *pdma32 = (u32 *)&dmaChain;
	u64 *pdma64 = (u64 *)(pdma32 + 4);

	// Set up VIF packet
	pdma32[i++] = VIF1_NOP;
	pdma32[i++] = VIF1_MSKPATH3(0x8000);
	pdma32[i++] = VIF1_FLUSHA;
	pdma32[i++] = VIF1_DIRECT(6);

	// Set up our GS packet
	i = 0;

	pdma64[i++] = GIFTAG(5, 1, 0, 0, 0, 1);
	pdma64[i++] = GIF_AD;
	// pdma64[i++] = GSBITBLTBUF_SET(address / 64, width / 64, format, 0, 0, 0);
	pdma64[i++] = GSBITBLTBUF_SET(address / 64, width < 64 ? 1 : width / 64, format, 0, 0, 0);
	pdma64[i++] = GSBITBLTBUF;
	pdma64[i++] = GSTRXPOS_SET(0, 0, 0, 0, 0); // SSAX, SSAY, DSAX, DSAY, DIR
	pdma64[i++] = GSTRXPOS;
	pdma64[i++] = GSTRXREG_SET(width, height); // RRW, RRh
	pdma64[i++] = GSTRXREG;
	pdma64[i++] = 0;
	pdma64[i++] = GSFINISH;
	pdma64[i++] = GSTRXDIR_SET(1); // XDIR
	pdma64[i++] = GSTRXDIR;

	prev_imr = GsPutIMR(GsGetIMR() | 0x0200);
	prev_chcr = *D1_CHCR;

	if ((*D1_CHCR & 0x0100) != 0)
		return 0;

	// Finish event
	*GS_CSR = CSR_FINISH;

	// DMA crap

	FlushCache(0);
	printf("DMA stuff\n");
	*D1_QWC = 0x7;
	*D1_MADR = (u32)pdma32;
	*D1_CHCR = 0x101;

	asm __volatile__("sync.l\n");

	// check if DMA is complete (STR=0)

	printf("Waiting for DMA channel completion\n");
	while (*D1_CHCR & 0x0100)
		;
	printf("Waiting for GS_CSR\n");
	while ((*GS_CSR & CSR_FINISH) == 0)
		;
	printf("Waiting for the vif fifo to empty\n");
	// Wait for viffifo to become empty
	while ((*VIF1_STAT & (0x1f000000)))
		;

	// Reverse busdir and transfer image to host
	*VIF1_STAT = VIF1_STAT_FDR;
	*GS_BUSDIR = (u64)0x00000001;

	FlushCache(0);

	u32 trans_size = ((width * height * psm_size) / 0x100);

	printf("Download the frame in 2 transfers of qwc 0x%x\n", trans_size);

	*D1_QWC = trans_size;
	*D1_MADR = (u32)destBuffer;
	*D1_CHCR = 0x100;

	asm __volatile__(" sync.l\n");

	// check if DMA is complete (STR=0)
	while (*D1_CHCR & 0x0100)
		;

	*D1_QWC = trans_size;
	*D1_CHCR = 0x100;

	asm __volatile__(" sync.l\n");

	// Wait for viffifo to become empty
	while ((*VIF1_STAT & (0x1f000000)))
		;

	// check if DMA is complete (STR=0)
	while (*D1_CHCR & 0x0100)
		;

	*D1_CHCR = prev_chcr;

	asm __volatile__(" sync.l\n");

	*VIF1_STAT = 0;
	*GS_BUSDIR = (u64)0;
	// Put back prew imr and set finish event

	GsPutIMR(prev_imr);
	*GS_CSR = CSR_FINISH;
	// Enable path3 again
	*VIF1_FIFO = enable_path3.value;

	/*
		Write out to a file
	*/

	FILE *f = fopen(fname, "w+");
	if ((u32)f < 0)
	{
		printf("??? couldn't open file, trying usb\n");
	}
	printf("File should be size %d\n", (psm_size / 8) * width * height);
	fwrite(destBuffer, psm_size / 8, width * height, f);
	fclose(f);
	return 0;
}

int main(void)
{

	(*(volatile u_int *)0x10003000) = 1;

	/*
		Initialize graph
	*/
	framebuffer_t fb;
	fb.width = WIDTH;
	fb.height = HEIGHT;
	fb.psm = FRAME_PSM;
	fb.mask = 0x00000000;

	fb.address = graph_vram_allocate(fb.width, fb.height, fb.psm, GRAPH_ALIGN_PAGE);

	zbuffer_t zb;
	zb.enable = 1;
	zb.zsm = GS_ZBUF_32;
	zb.address = graph_vram_allocate(fb.width, fb.height, zb.zsm, GRAPH_ALIGN_PAGE);
	zb.mask = 0;
	zb.method = ZTEST_METHOD_ALLPASS;

	graph_initialize(fb.address, fb.width, fb.height, fb.psm, 0, 0);

	dma_channel_initialize(DMA_CHANNEL_GIF, NULL, 0);
	dma_channel_fast_waits(DMA_CHANNEL_GIF);

	/*
		Init drawing stuff
	*/

	packet_t *packet = packet_init(100, PACKET_NORMAL);

	qword_t *q = packet->data;

	q = draw_setup_environment(q, 0, &fb, &zb);
	q = draw_primitive_xyoffset(q, 0, 0, 0);
	q = draw_enable_tests(q, 0, &zb);
	q = draw_finish(q);

	dma_channel_send_normal(DMA_CHANNEL_GIF, packet->data, q - packet->data, 0, 0);
	draw_wait_finish();

	/*
		Upload texture
	*/

	q = packet->data;

	FlushCache(0);

	int clutAddress = graph_vram_allocate(16, 16, GS_PSM_32, GRAPH_ALIGN_BLOCK);

	q = draw_texture_transfer(q, &clut[0], 16, 16, GS_PSM_32, clutAddress, 64);

	q = draw_texture_flush(q);

	dma_channel_send_chain(DMA_CHANNEL_GIF, packet->data, q - packet->data, 0, 0);
	dma_wait_fast();

	q = packet->data;

	/*int texAddress32 = graph_vram_allocate(32, 32, GS_PSM_4, GRAPH_ALIGN_BLOCK);
	int texAddress16 = graph_vram_allocate(16, 16, GS_PSM_4, GRAPH_ALIGN_BLOCK);
	int texAddress8 = graph_vram_allocate(8, 8, GS_PSM_4, GRAPH_ALIGN_BLOCK);
	int texAddress4 = graph_vram_allocate(4, 4, GS_PSM_4, GRAPH_ALIGN_BLOCK);*/

	 int texAddress32 = graph_vram_allocate(32, 32, GS_PSM_4, GRAPH_ALIGN_BLOCK);
    int texAddress16 = texAddress32  + 0x80;
    int texAddress8 = texAddress16  + 0x40;
    int texAddress4 =texAddress8 + 0x40;

	printf("%x,%x,%x,%x\n", texAddress32, texAddress16, texAddress8, texAddress4);
	q = draw_texture_transfer(q, (void *)tex4_1, 32, 32, GS_PSM_4, texAddress32, 64);
	q = draw_texture_transfer(q, (void *)tex4_2, 16, 16, GS_PSM_4, texAddress16, 64);
	q = draw_texture_transfer(q, (void *)tex4_3, 8, 8, GS_PSM_4, texAddress8, 64);
	q = draw_texture_transfer(q, (void *)tex4_4, 4, 8, GS_PSM_4, texAddress4, 64);
	//q = draw_texture_transfer(q, (void *)tex4_4, 4, 4, GS_PSM_4, texAddress32, 64);
	q = draw_texture_flush(q);

	dma_channel_send_chain(DMA_CHANNEL_GIF, packet->data, q - packet->data, 0, 0);
	dma_wait_fast();

	/*
		Now draw something
	*/

	int frame = 0;
	 while (1)
	{
	q = packet->data;

	// Draw a full screen sprite

	PACK_GIFTAG(q, GIF_SET_TAG(1, 1, GIF_PRE_ENABLE, GIF_PRIM_SPRITE, GIF_FLG_PACKED, 7),
				GIF_REG_AD | (GIF_REG_AD << 4) | (GIF_REG_AD << 8) | (GIF_REG_AD << 12) | (GIF_REG_RGBAQ << 16) | (GIF_REG_XYZ2 << 20) | (GIF_REG_XYZ2 << 24));

	q++;
	q->dw[0] = GS_SET_TEX1(1, 3, 0, 2, 1, 0, 0); // MXL 0
	q->dw[1] = GS_REG_TEX1;
	q++;
	q->dw[0] = GS_SET_MIPTBP1(texAddress16 / 64, 1, texAddress8 / 64, 1, texAddress4 / 64, 1); // MXL 3
	q->dw[1] = GS_REG_NOP;
	q++;
	q->dw[0] = GS_SET_TEX0(texAddress32 / 64, 1, GS_PSM_4, 5, 5, 0, 1, clutAddress / 64, GS_PSM_32, 0, 0, 1);
	q->dw[1] = GS_REG_TEX0;
	q++;
	q->dw[0] = GS_SET_TEST(0, 0, 0, 0, 0, 0, 1, 1);
	q->dw[1] = GS_REG_TEST;
	q++;
	// RGBAQ
	q->dw[0] = (u64)((0x33) | ((u64)0x33 << 32));
	q->dw[1] = (u64)((0x33) | ((u64)0 << 32));
	q++;
	// XYZ2
	q->dw[0] = (u64)((((0 << 4)) | (((u64)(0 << 4)) << 32)));
	q->dw[1] = (u64)(0);
	q++;
	// XYZ2
	q->dw[0] = (u64)((((640 << 4)) | (((u64)(480 << 4)) << 32)));
	q->dw[1] = (u64)(0);
	q++;

	q = draw_finish(q);

	dma_channel_send_normal(DMA_CHANNEL_GIF, packet->data, q - packet->data, 0, 0);

	draw_wait_finish();

		
			frame++;
			if (!(frame % 1000))
			{
				printf("1000Frames\n");
		// allows me to reset ps2link
			}

		q = packet->data;

		PACK_GIFTAG(q, GIF_SET_TAG(4, 1, GIF_PRE_ENABLE, GIF_SET_PRIM(GIF_PRIM_SPRITE, 0, 1, 0, 0, 0, 0, 0, 0), GIF_FLG_PACKED, 6),
					GIF_REG_AD | (GIF_REG_ST << 4) | (GIF_REG_RGBAQ << 8) | (GIF_REG_XYZ2 << 12) | (GIF_REG_ST << 16) | (GIF_REG_XYZ2 << 20));
		q++;

		{ // 128X128
			// TEST1
			q->dw[0] = GS_SET_TEX1(1, 3, 0, 2, 1, 0, 0);
			q->dw[1] = GS_REG_TEX1;
			q++;
			// UV
			q->dw[0] = GIF_SET_ST(0 << 4, 0 << 4);
			q->dw[1] = 0x3f800000;
			q++;
			// RGBAQ (Q will default to 1.0f on new tag)
			q->dw[0] = (u64)((0) | ((u64)0xcc << 32));
			q->dw[1] = (u64)((0xcc) | ((u64)0xFF << 32));
			q++;
			// XYZ2
			q->dw[0] = (u64)((((0 << 4)) | (((u64)(0 << 4)) << 32)));
			q->dw[1] = (u64)(0x1000);
			q++;
			// UV
			q->dw[0] = GIF_SET_ST(0x3f800000, 0x3f800000);
			q->dw[1] = 0x3f800000;
			q++;
			// XYZ2
			q->dw[0] = (u64)(((128 << 4)) | (((u64)(128 << 4)) << 32));
			q->dw[1] = (u64)(0x1000);
			q++;
		}
		{ // 64X64
			// TEST1
			q->dw[0] = GS_SET_TEX1(1, 3, 0, 2, 1, 0, 1 << 4);
			q->dw[1] = GS_REG_TEX1;
			q++;
			// UV
			q->dw[0] = GIF_SET_ST(0 << 4, 0 << 4);
			q->dw[1] = 0x3f800000;
			q++;
			// RGBAQ (Q will default to 1.0f on new tag)
			q->dw[0] = (u64)((0) | ((u64)0xcc << 32));
			q->dw[1] = (u64)((0xcc) | ((u64)0xFF << 32));
			q++;
			// XYZ2
			q->dw[0] = (u64)((((128 << 4)) | (((u64)(0 << 4)) << 32)));
			q->dw[1] = (u64)(0x1000);
			q++;
			// UV
			q->dw[0] = GIF_SET_ST(0x3f800000, 0x3f800000);
			q->dw[1] = 0x3f800000;
			q++;
			// XYZ2
			q->dw[0] = (u64)(((256 << 4)) | (((u64)(128 << 4)) << 32));
			q->dw[1] = (u64)(0x1000);
			q++;
		}
		{ // 32X32
			// TEST1
			q->dw[0] = GS_SET_TEX1(1, 3, 0, 2, 1, 0, 2 << 4);
			q->dw[1] = GS_REG_TEX1;
			q++;
			// UV
			q->dw[0] = GIF_SET_ST(0 << 4, 0 << 4);
			q->dw[1] = 0x3f800000;
			q++;
			// RGBAQ (Q will default to 1.0f on new tag)
			q->dw[0] = (u64)((0) | ((u64)0xcc << 32));
			q->dw[1] = (u64)((0xcc) | ((u64)0xFF << 32));
			q++;
			// XYZ2
			q->dw[0] = (u64)((((256 << 4)) | (((u64)(0 << 4)) << 32)));
			q->dw[1] = (u64)(0x1000);
			q++;
			// UV
			q->dw[0] = GIF_SET_ST(0x3f800000, 0x3f800000);
			q->dw[1] = 0x3f800000;
			q++;
			// XYZ2
			q->dw[0] = (u64)(((386 << 4)) | (((u64)(128 << 4)) << 32));
			q->dw[1] = (u64)(0x1000);
			q++;
		}
		{ // 16X16
			// TEST1
			q->dw[0] = GS_SET_TEX1(1, 3, 0, 2, 1, 0, 3 << 4);
			q->dw[1] = GS_REG_TEX1;
			q++;
			// UV
			q->dw[0] = GIF_SET_ST(0 << 4, 0 << 4);
			q->dw[1] = 0x3f800000;
			q++;
			// RGBAQ (Q will default to 1.0f on new tag)
			q->dw[0] = (u64)((0) | ((u64)0xcc << 32));
			q->dw[1] = (u64)((0xcc) | ((u64)0xFF << 32));
			q++;
			// XYZ2
			q->dw[0] = (u64)((((386 << 4)) | (((u64)(0 << 4)) << 32)));
			q->dw[1] = (u64)(0x1000);
			q++;
			// UV
			q->dw[0] = GIF_SET_ST(0x3f800000, 0x3f800000);
			q->dw[1] = 0x3f800000;
			q++;
			// XYZ2
			q->dw[0] = (u64)(((512 << 4)) | (((u64)(128 << 4)) << 32));
			q->dw[1] = (u64)(0x1000);
			q++;
		}
		q = draw_finish(q);

		dma_channel_send_normal(DMA_CHANNEL_GIF, packet->data, q - packet->data, 0, 0);

		// Wait until the drawing is finished.
		draw_wait_finish();
		{
			/*	q = packet->data;

				PACK_GIFTAG(q, GIF_SET_TAG(1, 1, GIF_PRE_DISABLE, GIF_PRIM_SPRITE, GIF_FLG_PACKED, 3),
						GIF_REG_AD | (GIF_REG_AD << 4) | (GIF_REG_AD << 8));
				q++;
				q->dw[0] = GS_SET_TEX1(1, 3, 0, 2, 1, 0, 0);
				q->dw[1] = GS_REG_TEX1;
				q++;
				q->dw[0] = GS_SET_MIPTBP1(0,1,0,1,0,1);
				q->dw[1] = GS_REG_MIPTBP1;
				q++;
				q->dw[0] = GS_SET_TEX0(texAddress128 / 64, 2, GS_PSM_24, 7, 7, 0, 1, 0, 0, 0, 0, 0);
				q->dw[1] = GS_REG_TEX0;
				q++;
				q = draw_finish(q);
				dma_channel_send_normal(DMA_CHANNEL_GIF, packet->data, q - packet->data, 0, 0);
				draw_wait_finish();*/
		}
		q = packet->data;
		PACK_GIFTAG(q, GIF_SET_TAG(4, 1, GIF_PRE_ENABLE, GIF_SET_PRIM(GIF_PRIM_SPRITE, 0, 1, 0, 0, 0, 0, 0, 0), GIF_FLG_PACKED, 6),
					GIF_REG_AD | (GIF_REG_ST << 4) | (GIF_REG_RGBAQ << 8) | (GIF_REG_XYZ2 << 12) | (GIF_REG_ST << 16) | (GIF_REG_XYZ2 << 20));
		q++;

		{ // 128X128
			// TEST1
			q->dw[0] = GS_SET_TEX1(1, 3, 0, 2, 1, 0, 0);
			q->dw[1] = GS_REG_TEX1;
			q++;
			// UV
			q->dw[0] = GIF_SET_ST(0 << 4, 0 << 4);
			q->dw[1] = 0x3f800000;
			q++;
			// RGBAQ (Q will default to 1.0f on new tag)
			q->dw[0] = (u64)((0) | ((u64)0xcc << 32));
			q->dw[1] = (u64)((0xcc) | ((u64)0xFF << 32));
			q++;
			// XYZ2
			q->dw[0] = (u64)((((0 << 4)) | (((u64)(128 << 4)) << 32)));
			q->dw[1] = (u64)(0x1000);
			q++;
			// UV
			q->dw[0] = GIF_SET_ST(0x3f800000, 0x3f800000);
			q->dw[1] = 0x3f800000;
			q++;
			// XYZ2
			q->dw[0] = (u64)(((128 << 4)) | (((u64)(256 << 4)) << 32));
			q->dw[1] = (u64)(0x1000);
			q++;
		}
		{ // 64X64
			// TEST1
			q->dw[0] = GS_SET_TEX1(1, 3, 0, 2, 1, 0, 1 << 4);
			q->dw[1] = GS_REG_TEX1;
			q++;
			// UV
			q->dw[0] = GIF_SET_ST(0 << 4, 0 << 4);
			q->dw[1] = 0x3f800000;
			q++;
			// RGBAQ (Q will default to 1.0f on new tag)
			q->dw[0] = (u64)((0) | ((u64)0xcc << 32));
			q->dw[1] = (u64)((0xcc) | ((u64)0xFF << 32));
			q++;
			// XYZ2
			q->dw[0] = (u64)((((128 << 4)) | (((u64)(128 << 4)) << 32)));
			q->dw[1] = (u64)(0x1000);
			q++;
			// UV
			q->dw[0] = GIF_SET_ST(0x3f800000, 0x3f800000);
			q->dw[1] = 0x3f800000;
			q++;
			// XYZ2
			q->dw[0] = (u64)(((256 << 4)) | (((u64)(256 << 4)) << 32));
			q->dw[1] = (u64)(0x1000);
			q++;
		}
		{ // 32X32
			// TEST1
			q->dw[0] = GS_SET_TEX1(1, 3, 0, 2, 1, 0, 2 << 4);
			q->dw[1] = GS_REG_TEX1;
			q++;
			// UV
			q->dw[0] = GIF_SET_ST(0 << 4, 0 << 4);
			q->dw[1] = 0x3f800000;
			q++;
			// RGBAQ (Q will default to 1.0f on new tag)
			q->dw[0] = (u64)((0) | ((u64)0xcc << 32));
			q->dw[1] = (u64)((0xcc) | ((u64)0xFF << 32));
			q++;
			// XYZ2
			q->dw[0] = (u64)((((256 << 4)) | (((u64)(128 << 4)) << 32)));
			q->dw[1] = (u64)(0x1000);
			q++;
			// UV
			q->dw[0] = GIF_SET_ST(0x3f800000, 0x3f800000);
			q->dw[1] = 0x3f800000;
			q++;
			// XYZ2
			q->dw[0] = (u64)(((386 << 4)) | (((u64)(256 << 4)) << 32));
			q->dw[1] = (u64)(0x1000);
			q++;
		}
		{ // 16X16
			// TEST1
			q->dw[0] = GS_SET_TEX1(1, 3, 0, 2, 1, 0, 3 << 4);
			q->dw[1] = GS_REG_TEX1;
			q++;
			// UV
			q->dw[0] = GIF_SET_ST(0 << 4, 0 << 4);
			q->dw[1] = 0x3f800000;
			q++;
			// RGBAQ (Q will default to 1.0f on new tag)
			q->dw[0] = (u64)((0) | ((u64)0xcc << 32));
			q->dw[1] = (u64)((0xcc) | ((u64)0xFF << 32));
			q++;
			// XYZ2
			q->dw[0] = (u64)((((386 << 4)) | (((u64)(128 << 4)) << 32)));
			q->dw[1] = (u64)(0x1000);
			q++;
			// UV
			q->dw[0] = GIF_SET_ST(0x3f800000, 0x3f800000);
			q->dw[1] = 0x3f800000;
			q++;
			// XYZ2
			q->dw[0] = (u64)(((512 << 4)) | (((u64)(256 << 4)) << 32));
			q->dw[1] = (u64)(0x1000);
			q++;
		}
		q = draw_finish(q);

		dma_channel_send_normal(DMA_CHANNEL_GIF, packet->data, q - packet->data, 0, 0);

		// Wait until the drawing is finished.
		draw_wait_finish();

		/* Flat shaded triangle, for testing
				q = packet->data;

				PACK_GIFTAG(q, GIF_SET_TAG(1, 1, GIF_PRE_ENABLE, GIF_PRIM_TRIANGLE, GIF_FLG_PACKED, 5),
							GIF_REG_AD | (GIF_REG_RGBAQ << 4) | (GIF_REG_XYZ2 << 8) | (GIF_REG_XYZ2 << 12) | (GIF_REG_XYZ2 << 16));
				q++;

				// TEST1
				q->dw[0] = GS_SET_TEST(0, 0, 0, 0, 0, 0, 1, 1);
				q->dw[1] = GS_REG_TEST;
				q++;
				// RGBAQ (Q will default to 1.0f on new tag)
				q->dw[0] = (u64)((0) | ((u64)0xcc << 32));
				q->dw[1] = (u64)((0xcc) | ((u64)0xFF << 32));
				q++;

				// XYZ2
				q->dw[0] = (u64)((((200 << 4)) | (((u64)(100 << 4)) << 32)));
				q->dw[1] = (u64)(0x1000);
				q++;
				// XYZ2
				q->dw[0] = (u64)(((100 << 4)) | (((u64)(200 << 4)) << 32));
				q->dw[1] = (u64)(0x1000);
				q++;
				// XYZ2
				q->dw[0] = (u64)(((200 << 4)) | (((u64)(200 << 4)) << 32));
				q->dw[1] = (u64)(0x1000);
				q++;

				q = draw_finish(q);

				dma_channel_send_normal(DMA_CHANNEL_GIF, packet->data, q - packet->data, 0, 0);

				// Wait until the drawing is finished.
				draw_wait_finish();
		*/
		// Now initiate vsync.
		graph_wait_vsync();
	}

	download_from_gs(fb.address, GS_PSM_32, "framebuffer.raw", 32, WIDTH, HEIGHT);
	SleepThread();
}

qword_t *draw_texture_transfer(qword_t *q, void *src, int width, int height, int psm, int dest, int dest_width, u32 dx, u32 dy)
{

	int i;
	int remaining;
	int qwords = 0;

	switch (psm)
	{
	case GS_PSM_8:
	{
		qwords = (width * height) >> 4;
		break;
	}

	case GS_PSM_32:
	case GS_PSM_24:
	{
		qwords = (width * height) >> 2;
		break;
	}

	case GS_PSM_4:
	{
		qwords = (width * height) >> 5;
		break;
	}

	case GS_PSM_16:
	case GS_PSM_16S:
	{
		qwords = (width * height) >> 3;
		break;
	}

	default:
	{
		switch (psm)
		{
		case GS_PSM_8H:
		{
			qwords = (width * height) >> 4;
			break;
		}

		case GS_PSMZ_32:
		case GS_PSMZ_24:
		{
			qwords = (width * height) >> 2;
			break;
		}

		case GS_PSMZ_16:
		case GS_PSMZ_16S:
		{
			qwords = (width * height) >> 3;
			break;
		}

		case GS_PSM_4HL:
		case GS_PSM_4HH:
		{
			qwords = (width * height) >> 5;
			break;
		}
		}
		break;
	}
	}

	// Determine number of iterations based on the number of qwords
	// that can be handled per dmatag
	i = qwords / GIF_BLOCK_SIZE;

	// Now calculate the remaining image data left over
	remaining = qwords % GIF_BLOCK_SIZE;

	// Setup the transfer
	DMATAG_CNT(q, 5, 0, 0, 0);
	q++;
	PACK_GIFTAG(q, GIF_SET_TAG(4, 0, 0, 0, GIF_FLG_PACKED, 1), GIF_REG_AD);
	q++;
	PACK_GIFTAG(q, GS_SET_BITBLTBUF(0, 0, 0, dest >> 6, dest_width >> 6, psm), GS_REG_BITBLTBUF);
	q++;
	PACK_GIFTAG(q, GS_SET_TRXPOS(0, 0, dx, dy, 0), GS_REG_TRXPOS);
	q++;
	PACK_GIFTAG(q, GS_SET_TRXREG(width, height), GS_REG_TRXREG);
	q++;
	PACK_GIFTAG(q, GS_SET_TRXDIR(0), GS_REG_TRXDIR);
	q++;

	while (i-- > 0)
	{

		// Setup image data dma chain
		DMATAG_CNT(q, 1, 0, 0, 0);
		q++;
		PACK_GIFTAG(q, GIF_SET_TAG(GIF_BLOCK_SIZE, 0, 0, 0, 2, 0), 0);
		q++;
		DMATAG_REF(q, GIF_BLOCK_SIZE, (unsigned int)src, 0, 0, 0);
		q++;

		// Now increment the address by the number of qwords in bytes
		src = (void *)src + (GIF_BLOCK_SIZE * 16);
	}

	if (remaining)
	{

		// Setup remaining image data dma chain
		DMATAG_CNT(q, 1, 0, 0, 0);
		q++;
		PACK_GIFTAG(q, GIF_SET_TAG(remaining, 0, 0, 0, 2, 0), 0);
		q++;
		DMATAG_REF(q, remaining, (unsigned int)src, 0, 0, 0);
		q++;
	}

	return q;
}