/**************************************************************************
 *                                                                        *
 *               Copyright (C) 1995, Silicon Graphics, Inc.               *
 *                                                                        *
 *  These coded instructions, statements, and computer programs  contain  *
 *  unpublished  proprietary  information of Silicon Graphics, Inc., and  *
 *  are protected by Federal copyright  law.  They  may not be disclosed  *
 *  to  third  parties  or copied or duplicated in any form, in whole or  *
 *  in part, without the prior written consent of Silicon Graphics, Inc.  *
 *                                                                        *
 *************************************************************************/

/*
 * File:	main.c
 *
 *
 * This application is a test-bed of single frame tests for RDP
 * test vector generation. The idea is that each frame possibly uses a 
 * different dataset (see static.c) to render a particular test.
 *
 * Under argc, argv control, you can choose to run a single frame test, or
 * (by default) run all the "frames" of vector generation at once.  When
 * sending multiple "frames" down, we always wait for vertical retrace before
 * drawing the next frame (frames are always drawn in the first cfb, so that we
 * know how to extract the image data when running simulations/test vectors).
 *
 * At the user's option, they may specify that RDRAM (e.g. the frame buffer)
 * has been preinitialized, thus eliminating the need for a screen/zbuffer 
 * clear.  Use the -m flag when this is the case.
 *
 * A similar strategy was used in RDP verification, so we are leveraging
 * heavily from the old rdpverif program.
 *
 * The "-c <number>" flag performs checksum screen compares:
 *
 * -c 0 dumps the checksums
 *
 * -c 1 compares the generated checksums with a previously acquired set 
 *      compiled into this program.  If the checsums don't match, we PRINTF
 *	and/or print the results to the screen.
 *      
 *
 * Use gload's -a flag to pass arguments through to this app.
 *
 * NOTE:
 *
 * This test-bed is composed of some shared code, and test-specific
 * code. To add or modify a test case, you should only touch the
 * test-specific code. Modifying shared code might break other tests, should
 * we ever create other directories of test vector generators based upon this
 * program.
 *
 *  FILES THAT MAY BE SHARED WITH OTHER TESTS: (don't modify)
 *
 *	main.c
 *	init.c
 *	dynamic.c
 *	zbuffer.c
 *	spec
 *
 *  FILES THAT ARE TEST-SPECIFIC: (may modify)
 *
 * 	<module_name>_static.c	(holds static display list data for tests)
 *	<module_name>.h		(exports info to main.c, etc.)
 * 	<module_name.c		(holds test frame procedures, etc.)
 *	<any related include files>	(textures, etc.)
 *	Makefile			(make test rules)
 *
 * A separate set of files should be modified when producing display lists for
 * per module testing (e.g. cs.h, cs.c, cs_static.c).  A template display list
 * test has been provided; some real gfx dl's are provided as samples in the ms
 * test series.
 *
 * TO ADD A TEST CASE:
 *
 * decide which module your test will focus on, then tweak:
 *
 * <module_name>.h		- extern the static display list, so that it is
 *				  visible from <module_name>_static.c
 *
 * <module_name>_static.c	- add the static display list for the test.
 *
 * <module_name>.c		- add the procedure to be called to
 *				  perform the test case. (also put it
 *				  in the testCaseProc table appropriate for
 *				  this module)
 *
 *				- put the static display list pointer
 *				  in the testList table (appropriately named
 *				  for your module under test) as a Gfx pointer 
 *				  in the GfxTest_t struct; also provide a name
 *				  for your test so that the rdpascii2rdp tool
 *				  can provide a filename for the resulting
 *				  binary .rdp file.
 *
 * ... plus add any new textures, include files, etc.
 *
 * All locations are marked with comments that say "ADD TEST CASE"
 * and give some hints what you need to do there.
 *
 * ADVANCED USAGE:
 * You could re-use test case procedures or data, as long as you
 * coordinate with all the tests that share it. It is a bad idea
 * to share data or procedures with a test that someone else may
 * go in and modify.
 *
 * If your test case needs a command line argument, that isn't handled
 * too cleanly. You would have to add it to main.c, and export the
 * global variables, etc. (messy) 
 *
 * This test suite favors single-frame tests. Animation support is clumsy.
 *
 * Fri Sep 22 11:39:18 PDT 1995
 */

#include <ultra64.h>
#include <ramrom.h>
#include <assert.h>
#include "controller.h"

#include "rdpvector.h"
#include "ms.h"

#include "checksum.h"

/*
 * Symbol genererated by "makerom" to indicate the end of the code segment
 * in virtual (and physical) memory
 */
extern char _textureSegmentEnd[];

/* these are fixed sizes: */
#define SP_UCODE_SIZE		4096
#define SP_UCODE_DATA_SIZE	2048

/*
 * Symbols generated by "makerom" to tell us where the static segment is
 * in ROM.
 */
extern char _staticSegmentRomStart[], _staticSegmentRomEnd[];
extern char _textureSegmentRomStart[], _textureSegmentRomEnd[];
extern char _textureSegmentStart[];

/*
 * Stacks for the threads as well as message queues for synchronization
 */
u64	bootStack[STACKSIZE/sizeof(u64)];

static int      uselines  = 0;
static int      useblack  = 0;
static int      skipZclear  = 0;

static void	idle(void *);
static void	mainproc(void *);

static OSThread	idleThread;
static u64	idleThreadStack[STACKSIZE/sizeof(u64)];

static OSThread	mainThread;
static u64	mainThreadStack[STACKSIZE/sizeof(u64)];

static OSThread	rmonThread;
static u64      rmonStack[RMON_STACKSIZE/sizeof(u64)];

static char     dummy[1024*1024] = {0}; /* idiotic array to pad out image size for makemask */

/* this number (the depth of the message queue) needs to be equal
 * to the maximum number of possible overlapping PI requests.
 * For this app, 1 or 2 is probably plenty, other apps might
 * require a lot more.
 */
#define NUM_PI_MSGS     8
    
static OSMesg PiMessages[NUM_PI_MSGS];
static OSMesgQueue PiMessageQ;

OSMesgQueue	dmaMessageQ, rspMessageQ, rdpMessageQ, retraceMessageQ;
OSMesg		dmaMessageBuf, rspMessageBuf, rdpMessageBuf, retraceMessageBuf;
OSMesg		dummyMessage;
OSIoMesg	dmaIOMessageBuf;

/*
 * Dynamic segment in code space. Needs OS_K0_TO_PHYSICAL()...
 */
Dynamic dynamic;

/*
 * necessary for RSP tasks:
 */
#ifdef __GNUC__
u64 dram_stack[SP_DRAM_STACK_SIZE64] __attribute__((aligned (16)));
u64 dram_yield[OS_YIELD_DATA_SIZE];

u64 rdp_output_len_foo[4] __attribute__((aligned (16)));
#define rdp_output_len	rdp_output_len_foo[0]
u64 rdp_output[4096*16] __attribute__((aligned (16)));
#else
u64 dram_stack[SP_DRAM_STACK_SIZE64];
u64 dram_yield[OS_YIELD_DATA_SIZE];

u64 rdp_output_len;
u64 rdp_output[4096*16];
#endif

/*
 * must be in BSS, not on the stack for this to work:
 */
OSTask	tlist =
{
    M_GFXTASK,			/* task type */
    OS_TASK_DP_WAIT,		/* task flags */
    NULL,			/* boot ucode pointer (fill in later) */
    0,				/* boot ucode size (fill in later) */
    NULL,			/* task ucode pointer (fill in later) */
    SP_UCODE_SIZE,		/* task ucode size */
    NULL,			/* task ucode data pointer (fill in later) */
    SP_UCODE_DATA_SIZE,		/* task ucode data size */
    &(dram_stack[0]),		/* task dram stack pointer */
    SP_DRAM_STACK_SIZE8,	/* task dram stack size */
    &(rdp_output[0]),		/* task output buffer ptr (not always used) */
    &rdp_output_len,		/* task output buffer size ptr */
    NULL,			/* task data pointer (fill in later) */
    0,				/* task data size (fill in later) */
    NULL,			/* task yield buffer ptr (not used here) */
    0				/* task yield buffer size (not used here) */
};

static char *staticSegment;
static char *textureSegment;

Gfx		*glistp;	/* global for test case procs */

/*
 * global variables for arguments, to control test cases
 */
static int	frame_count = -1;
static int	module_num = -1;
       int	rdp_DRAM_io = 0;
static int	checksumRequest = -1;
static int	rdp_regressionFlag = 0;
static int	rdp_mspanFlushFlag = 0;
static int	debugflag = 0;
static int	dumpImage = 0;
static int      dumpImage2 = 0;
static int      draw_buffer = 0;
static void     *cfb_ptrs[2];
static int      cfb_size = G_IM_SIZ_16b;
static u64	ramrombuf[RAMROM_MSG_SIZE/8];

int             controlFrame = 0;
int		lastFrameNum = -2;

		/*
		 * When TRUE, toplevel.c routines will issue a full sync after
		 * the frame's DL has been rendered.  When dumping DL's, we
		 * will want to inhibit the full sync's for every frame of a
		 * particular module until the last frame, so that we can
		 * accumulate all the frame's test vectors together in one 
		 * monster DL.
		 */
int		generateFullSync = 0;

typedef union {
    u32		word[2];
    u64		force_alignment;
} scratch_t;
scratch_t scratch_buff[1024], *scrp;

#define PI_BASE	((u32)(0xb0700000))
static u32	*piAddr = (u32 *)(PI_BASE);
u32		dummy_val;

/* 
 * Declare an array of ModuleTest_t structs, so that we can cycle through all
 * the tests designed for each of the rdp's major functional blocks.
 */
ModuleTest_t moduleTestSet[MODULE_COUNT];

u32 argbuf[16];

/*
 * The boot() function MUST be the first function in this module, otherwise
 * the original boot thread (loaded at 2MB in rdram) will be unable to invoke 
 * this secondary boot procedure, which loads the program low in rdram.
 */
boot(void *arg)
{
    int i, *pr;
    char *ap;
    u32 *argp;
    
    osInitialize();

   for (pr = (int *)0xb0700000; (u32)pr < 0xb0702000; pr++)
       __osPiRawWriteIo((u32)pr, 0x55555555);

    argp = (u32 *)RAMROM_APP_WRITE_ADDR;
    for (i=0; i<sizeof(argbuf)/4; i++, argp++) {
	__osPiRawReadIo((u32)argp, &argbuf[i]); /* Assume no DMA */
    }

    osCreateThread(&idleThread, 1, idle, arg, idleThreadStack+STACKSIZE/sizeof(u64), 10);
    osStartThread(&idleThread);

    /* never reached */
}

/*
 * private routine used to parse args.
 */
static int
myatoi(char *str)
{
    int		log10[5], logn = 0, val = 0, i, pow10 = 1;

    if (str == NULL || *str == '\0')
	return(-1);
    
    while (*str != '\0') {
	if (!(*str == '0' ||
	      *str == '1' ||
	      *str == '2' ||
	      *str == '3' ||
	      *str == '4' ||
	      *str == '5' ||
	      *str == '6' ||
	      *str == '7' ||
	      *str == '8' ||
	      *str == '9')) {
	    logn = 0;
	    break;
	}

	log10[logn++] = *str - '0';
	str++;
    }
    
    if (logn == 0)
	return(-1);
    
    for (i=logn-1; i>= 0; i--) {
	val += log10[i] * pow10;
	pow10 *= 10;
    }
    return (val);
}

/*
 * private routine used to parse args.
 */
static void
parse_args(char *argstring)
{
    int		argc = 1, i;
    char	*arglist[32], **argv = arglist;	/* max 32 args */
    char	*c, *ptr;

    if (argstring == NULL || argstring[0] == '\0')
	return;

    /* re-organize argstring to be like main(argv,argc) */
    ptr = argstring;
    while (*ptr != '\0') {
	while (*ptr != '\0' && (*ptr == ' ')) {
	    *ptr = '\0';
	    ptr++;
	}
	if (*ptr != '\0')
	    arglist[argc++] = ptr;
	while (*ptr != '\0' && (*ptr != ' ')) {
	    ptr++;
	}
    }

    /* ADD TEST CASE:
     * might want to pick off special arguments here.
     */

    /* process the arguments: */
    while ((argc > 1) && (argv[1][0] == '-')) {
	switch(argv[1][1]) {

	  /*
	   * -c controls checksum behavior; legitimate values are "0" to
	   * dump checksums, "1" to compare generated checksums with 
	   * old values compiled in to this program.
	   */
	  case 'c':
	    checksumRequest = myatoi(argv[2]);
	    if ( (checksumRequest < 0) || (checksumRequest > 1) ) {
		checksumRequest = -1;
	    }
	    argc--;
	    argv++;
	    break;

	  case 'd':
	    rdp_DRAM_io = 1;	/* route RSP output for RDP back to DRAM */
	    break;
	    
	  case 'f':
	    frame_count = myatoi(argv[2]); /* start frame */
	    if (frame_count < 0) {
		frame_count = -1;
	    }
	    controlFrame = frame_count;
	    argc--;
	    argv++;
	    break;

	  case 'g':
	    debugflag = 1;
	    break;

	  case 'i':
	    dumpImage = 1;
	    break;

	  case 'I':
	    dumpImage2 = 1;
	    break;

	  case 'F':
	    rdp_mspanFlushFlag = 1; /* flush mspan cache by enabling 1PRIM */
            break;
	    
	  case 'r':
	    rdp_regressionFlag = 1; /* when set, shrink viewport for brevity */
	    break;
	    
	  case 'l':
	     uselines = 1;
	     break;

	/*
	 * -m specifies that we run a single module's set of tests.
	 */
	case 'm':
	    module_num = myatoi(argv[2]);
	    if (module_num < 0) {
		module_num = -1;
	    }
	    argc--;
	    argv++;
	    break;

	case 'b':
	    useblack = 1;
	    break;
	    
	case 'z':
	    skipZclear = 1;
	    break;

	  default:
	    /*
	    PRINTF("option [%s] not recognized.\n", argv[1]);
	    */
	    break;
	}
	argc--;
	argv++;
    }
}

static void
idle(void *arg)
{
    /* Initialize video */
    osCreateViManager(OS_PRIORITY_VIMGR);
    osViSetMode(&osViModeTable[OS_VI_NTSC_LAN1]);
    
    /*
     * Parse option arguments before we start the main thread, so that if we
     * set the debugflag, we won't start the main thread (debugger will have to
     * do that for us).  Unfortunately this means we can't PRINTF any
     * errors in argument parsing...
     */
    parse_args((char *)argbuf);

    /*
     * Start PI Mgr for access to cartridge
     */
    osCreatePiManager((OSPri)OS_PRIORITY_PIMGR, &PiMessageQ, PiMessages, 
		      NUM_PI_MSGS);
    
    /*
     * Start RMON for debugging & data xfer (make sure to start 
     * PI Mgr first)
     */
    osCreateThread(&rmonThread, 0, rmonMain, (void *)0,
		   rmonStack+RMON_STACKSIZE/sizeof(u64), OS_PRIORITY_RMON);
    osStartThread(&rmonThread);
    
    /*
     * Create main thread
     */
    osCreateThread(&mainThread, 3, mainproc, arg,
		   mainThreadStack+STACKSIZE/sizeof(u64), 10);
    
    if (debugflag != 1) {
	osStartThread(&mainThread);
    }
    /*
     * Become the idle thread
     */
    osSetThreadPri( 0, 0 );
    for(;;);
}

/* 
 * Dumps texture memory out as raw data.  We expect texture data to always be
 * 64 bit aligned, and dump a little extra if the size says there's less.
 */
static void
dump_texture_memory(u64 *textureSegment, u64 size)
{
    char String[128];

    sprintf(String, 
	    "TEXTURE DUMP 0x%x\n", osVirtualToPhysical(textureSegment));
    guParseString(String, 128);
    if (size % sizeof(u64)) {
	size += sizeof(u64);
	size &= ~(sizeof(u64));
    }
    guParseRdpDL(textureSegment, size, GU_PARSERDP_DUMPONLY);
}

/*
 * Build up an RDP task to clear the frame buffer, in whatever video mode the
 * current test requires.
 */
void 
clear_framebuffers(int cfb_pixSize)
{
    OSTask	*tlistp;
    Dynamic	*dynamicp;

    /*
     * set up pointers to build the display list.
     */
    tlistp = &tlist;
    dynamicp = &dynamic;
    glistp = &(dynamicp->glist[0]);
    
    /*
     * Tell RCP where each segment is
     */
    gSPSegment(glistp++, 0, 0x0); /* K0 (physical) address segment */
    gSPSegment(glistp++, STATIC_SEGMENT,
	       osVirtualToPhysical(staticSegment));
    gSPSegment(glistp++, TEXTURE_SEGMENT,
	       osVirtualToPhysical(textureSegment));
    
    /*
     * XXX Probably not necessary to do this each time, but too hard to figure
     *     out what's needed and what's not, so we do it every time & wonder...
     */
    gSPDisplayList(glistp++, rdpinit_dl);
    gSPDisplayList(glistp++, rspinit_dl);

    if (cfb_pixSize == G_IM_SIZ_16b) {
	if (skipZclear == 0) {
	    gSPDisplayList(glistp++, clear_16fb_z);
	}
	gSPDisplayList(glistp++, clear_16fb);
    } else {
	if (skipZclear == 0) {
	    gSPDisplayList(glistp++, clear_32fb_z);
	}
	gSPDisplayList(glistp++, clear_32fb);
    }

    gDPFullSync(glistp++);
    gSPEndDisplayList(glistp++);

    assert((glistp - dynamicp->glist) < MAX_GRAPHICS_SIZE);

    /* 
     * Build a graphics task to execute the frame buffer clears, fire it off,
     * then wait for RDP completion.
     */
    tlistp->t.type = M_GFXTASK;
    tlistp->t.flags = OS_TASK_DP_WAIT;
    tlistp->t.ucode_boot = (u64 *) rspbootTextStart;
    tlistp->t.ucode_boot_size = ((int)rspbootTextEnd - 
				 (int)rspbootTextStart);

    /* 
     * RSP output over XBUS to RDP.
     */
    tlistp->t.ucode = (u64 *) gspFast3DTextStart;
    tlistp->t.ucode_data = (u64 *) gspFast3DDataStart;
    
    tlistp->t.ucode_size = SP_UCODE_SIZE;
    tlistp->t.ucode_data_size = SP_UCODE_DATA_SIZE;
    
    tlistp->t.dram_stack = (u64 *) &(dram_stack[0]);
    tlistp->t.dram_stack_size = SP_DRAM_STACK_SIZE8;

    tlistp->t.output_buff = (u64 *) 0x0;
    tlistp->t.output_buff_size = (u64 *) 0x0;

    /* initial display list: */
    tlistp->t.data_ptr = (u64 *) dynamicp->glist;
    tlistp->t.data_size = ((int)(glistp - dynamicp->glist) *
			   sizeof (Gfx));
    tlistp->t.yield_data_ptr = (u64 *) NULL;
    tlistp->t.yield_data_size = 0xDA0;

    /*
     * Can just flush 16KB and forget about each individual pieces
     * of data to flush.
     */
    osWritebackDCacheAll();
    
    /* unwrap this macro for debugging:
     *
     *	osSpTaskStart(tlistp);
     *
     */
    osSpTaskLoad(tlistp);
    osSpTaskStartGo(tlistp);
    
    /* 
     * ignore SP completion, wait for DP completion.
     */
    (void)osRecvMesg(&rdpMessageQ, &dummyMessage, OS_MESG_BLOCK);

    /* 
     * Now wait for a vertical retrace to go by.
     *
     * Make sure there isn't an old retrace in queue 
     * (assumes queue has a depth of 1) by checking to see if the MQ is full.
     */

    if (MQ_IS_FULL(&retraceMessageQ))
	(void)osRecvMesg(&retraceMessageQ, &dummyMessage, OS_MESG_BLOCK);

    /* 
     * Now really wait for vertical retrace to finish.
     */
    (void)osRecvMesg(&retraceMessageQ, &dummyMessage, OS_MESG_BLOCK);
}

#define DUMP_REQUEST	0xdeadbeef
#define DUMP_REPLY	0xbaaddeed

void    
dumpRGB( int pixSize, int width, int height, unsigned long *fb)
                                     
{       
    unsigned long       *longPixels; 
    unsigned long       pixelValue;
    int                 row, column;
    int                 channel;
    char                str[32];
    char String[128];
            
    /*  
     * checksum calculates 32 bits, so if 16 bit fb, than 2 pixels at a time.
     * if 8 bit fb, than 4 pixels at a time
     */     

    PRINTF("START IMAGE\n");
        
    if ( pixSize == G_IM_SIZ_8b ) { 
        PRINTF("Dimensions %d %d 8\n",width,height);
        width >>= 2;
    } else
    if ( pixSize == G_IM_SIZ_16b ) {
        PRINTF("Dimensions %d %d 16\n",width,height);
        width >>= 1;
    } else  {   
        PRINTF("Dimensions %d %d 32\n",width,height);
    }           
        
    longPixels = (unsigned long *) (fb);
            
    for (column = 0; column < height; column++){
      for (row = 0; row < width; row += 8) {
          PRINTF("0x%x 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x ",
                      longPixels[0],longPixels[1],longPixels[2],longPixels[3],
                      longPixels[4],longPixels[5],longPixels[6],longPixels[7]);
        longPixels += 8;
        }
        PRINTF("\n");
      }
    
       PRINTF("END IMAGE\n");
}   

static void
computeChecksum(int pixSize, int moduleNum, int frameNum)
{
    unsigned short *shortPixels;
    unsigned long  *longPixels, *longPixels1;
    unsigned long long checksumValue;
    int row, column, i, checksumIndex;

    unsigned int hostBuffer[2];

    static stored_cfb = 0;

    /*
     * Only spin cycles to compute/compare the checksum upon user request.
     */
    if ( (checksumRequest == 0) || (checksumRequest == 1) ) {
#define HUSH 1
#ifdef HUSH
    PRINTF("moduleNum = %d, frameNum = %d, pixSize = %d\n", 
	moduleNum, frameNum, pixSize);
#endif
	if ( pixSize == G_IM_SIZ_16b ) {

	    shortPixels = (unsigned short *) (cfb_16_a);
	    checksumValue = 0;

	    for (row = 0; row < SCREEN_WD; row++) {
		for (column = 0; column < SCREEN_HT; column++) {
		    checksumValue += *shortPixels++;
		}
	    }
	} else if ( pixSize == G_IM_SIZ_32b ) {

	    longPixels = (unsigned long *) (cfb_16_a);
	    checksumValue = 0;

	    for (row = 0; row < SCREEN_WD; row++) {
		for (column = 0; column < SCREEN_HT; column++) {
		    checksumValue += *longPixels++;
		}
	    }
	} else {
	    /*
	     * 1024 x 1024 framebuffer renders in 32 bits.
	     */

	    longPixels = (unsigned long *) (cfb_16_a);
	    checksumValue = 0;

	    for (row = 0; row < 1024; row++) {
		for (column = 0; column < 1024; column++) {
		    checksumValue += *longPixels++;
		}
	    }
	}
	PRINTF("checksumValue = 0x%llx\n", checksumValue);

	if (checksumRequest == 1) {
	    /*
	     * Figure out which checksum value to use.
	     */

	    for (i = 0, checksumIndex = 0; i <  moduleNum; i++) {
	        checksumIndex += moduleTestSet[i].testListCount;
	    }
	    checksumIndex += frameNum;
#ifdef HUSH
	    if (checksumValues[checksumIndex] != checksumValue) {
		PRINTF("ERROR: (%d): exp 0x%llx, computed 0x%llx\n", 
		checksumIndex, checksumValues[checksumIndex], checksumValue);
	    } else {
		PRINTF("PASS checksum test.\n");
	    }
#endif
	}

	/* 
	 * Wait for a vertical retrace to go by.  SP sometimes hangs if we
	 * don't do this after computing a checksum (hack, I know, sigh).
	 *
	 * Make sure there isn't an old retrace in queue 
	 * (assumes queue has a depth of 1) by checking to see if the 
	 * MQ is full.
	 */

	if (MQ_IS_FULL(&retraceMessageQ))
	    /*
	     * Eat the old retrace message prior to the real wait below.
	     */
	    (void)osRecvMesg(&retraceMessageQ, &dummyMessage, 
		OS_MESG_BLOCK);

	/* 
	 * Now really wait for vertical retrace to finish.
	 */
	(void)osRecvMesg(&retraceMessageQ, &dummyMessage, 
		OS_MESG_BLOCK);
    }

    /* Alternate route for dumping frame buffer since below code looks like
       debugging crap left over by someone */

    if (dumpImage)  {
      dumpRGB(pixSize, 320, 240, cfb_ptrs[draw_buffer]);
      PRINTF("End frame dump \n");

      /* Now fill up output buffers so we get a flush of real data */
      for (i = 0; i < 1024; i++) PRINTF("================\n");

      dumpImage = 0; /* Don't dump any more frame */
    }
    

    /*
     * Hack to store framebuffer, so that we can read & compare, & see what's
     * different, since the checksums are differing inexplicably.
     */

    if ( /*(moduleNum == 5) && (frameNum == 3)*/ dumpImage2 ) {

	longPixels = (unsigned long *) (cfb_16_a);
	longPixels1 = (unsigned long *) (0x80300000);
	shortPixels = (unsigned short *) (cfb_16_a);

	if (stored_cfb == 0) {
	    PRINTF("image size %s\n", pixSize == G_IM_SIZ_16b ? "16" : "32");
	    if (pixSize == G_IM_SIZ_16b) {
		longPixels = longPixels1;
	        for (column = 0; column < SCREEN_HT; column++) {
		    for (row = 0; row < SCREEN_WD; row++) {
		        *longPixels1++ = ((*shortPixels&0xf800) << 16) |
					 ((*shortPixels&0x7c0) << (8+5)) |
					 ((*shortPixels&0x3e) << (0+10))|
					 ((*shortPixels&0x1) << (15));
			shortPixels++;
		    }
		}
	    }
	    stored_cfb = 1;

	    /*
	     * If we are dumping an image, negotiate with the host side
	     * process now & dump the image using the debug port.
	     */
	    if (dumpImage2) {
		osReadHost(hostBuffer, 4);
		if (hostBuffer[0] != DUMP_REQUEST) {
		    PRINTF("host request 0x%x received\n", hostBuffer[0]);
		    return;
		} 
		hostBuffer[0] = DUMP_REPLY;
		hostBuffer[1] = (SCREEN_WD * SCREEN_HT * sizeof(unsigned int));
		osWriteHost(hostBuffer, 8);
		osWriteHost(longPixels, 
		    (SCREEN_WD * SCREEN_HT * sizeof(unsigned int)));
		osExit();
	    }

	} else {
#ifdef NOTDEF
	    for (row = 0; row < SCREEN_WD; row++) {
		for (column = 0; column < SCREEN_HT; column++) {
		    if (*longPixels1 != *longPixels) {
PRINTF("ERROR (%d, %d): exp 0x%x, rd 0x%x, xor 0x%x\n",
row, column, *longPixels1, *longPixels, (*longPixels1 ^ *longPixels));
		    }
		    ++longPixels;
		    ++longPixels1;
		}
	    }
#endif
	}
    }
}

static void
mainproc(void *arg)
{
    OSTask	*tlistp;
    Dynamic	*dynamicp;
    int 	controllerSlot;
    int		testListCount;
    int		frameNum, endFrame;
    int		moduleNum, startModule, endModule;
    int         keep_going = 1;
    int		pixSize;

    cfb_ptrs[0] = cfb_16_a;
    cfb_ptrs[1] = cfb_16_b;
    /*
     * Setup the message queues
     */
    osCreateMesgQueue(&dmaMessageQ, &dmaMessageBuf, 1);
    
    osCreateMesgQueue(&rspMessageQ, &rspMessageBuf, 1);
    osSetEventMesg(OS_EVENT_SP, &rspMessageQ, dummyMessage);
    
    osCreateMesgQueue(&rdpMessageQ, &rdpMessageBuf, 1);
    osSetEventMesg(OS_EVENT_DP, &rdpMessageQ, dummyMessage);
    
    osCreateMesgQueue(&retraceMessageQ, &retraceMessageBuf, 1);
    osViSetEvent(&retraceMessageQ, dummyMessage, 1);

    /*
     * Load the texture segment wherever it ended up (it will be right after 
     * the codeSegment as per our spec file directives to makerom).
     *
     * We'll dump out it's address when dumping texture data, for inclusion
     * in all .rdram files to be generated by this program and rdpascii2rdram.
     */
    textureSegment = (char *)_textureSegmentStart;
    osInvalDCache( textureSegment, 
	(_textureSegmentRomEnd - _textureSegmentRomStart) );
    osPiStartDma(&dmaIOMessageBuf, OS_MESG_PRI_NORMAL, OS_READ,
		 (u32)_textureSegmentRomStart, textureSegment,
		 _textureSegmentRomEnd - _textureSegmentRomStart, &dmaMessageQ);

    /*
     * Wait for DMA to finish
     */
    (void)osRecvMesg(&dmaMessageQ, &dummyMessage, OS_MESG_BLOCK);

    /*
     * Stick the static segment right after the texture segment.
     */
    staticSegment = _textureSegmentEnd;
    osInvalDCache( staticSegment, 
	(_staticSegmentRomEnd - _staticSegmentRomStart) );
    osPiStartDma(&dmaIOMessageBuf, OS_MESG_PRI_NORMAL, OS_READ,
		 (u32)_staticSegmentRomStart, staticSegment,
		 _staticSegmentRomEnd - _staticSegmentRomStart, &dmaMessageQ);
    
    /*
     * Wait for DMA to finish
     */
    (void)osRecvMesg(&dmaMessageQ, &dummyMessage, OS_MESG_BLOCK);

    /*
     * Initialize controller
     */
    controllerSlot = initControllers();

    /*
     * Tell the VI to display the first frame buffer.
     */
    osViSwapBuffer(cfb_16_a);

    /*
     * Dump the contents of texture memory, so that rdpascii2rdram can load
     * each .rdram file with the contents of texture memory, as well the
     * pre-initialized frame buffer.
     */
    if (rdp_DRAM_io) {
	dump_texture_memory( (u64 *)textureSegment, 
	    (u64)(_textureSegmentRomEnd - _textureSegmentRomStart) );
    }

    /* 
     * Count number of tests to run for each module under test, and stuff
     * an array of pointers to the individual module test entry points, so 
     * that we can invoke them from the main game loop below.
     */

    /*
     * CS_MODULE
     */
    testListCount = 0;
    moduleTestSet[CS_MODULE].gfxTests = testList_cs;

    while ( (testList_cs[testListCount].gfx_dl != (Gfx *) NULL) &&
	    (testList_cs[testListCount].toplevel_dl != (CaseProc_t) NULL) ) {
	testListCount++;
    }
    moduleTestSet[CS_MODULE].testListCount = testListCount;

    /*
     * EW_MODULE
     */
    testListCount = 0;
    moduleTestSet[EW_MODULE].gfxTests = testList_ew;

    while ( (testList_ew[testListCount].gfx_dl != (Gfx *) NULL) &&
	    (testList_ew[testListCount].toplevel_dl != (CaseProc_t) NULL) ) {
	testListCount++;
    }
    moduleTestSet[EW_MODULE].testListCount = testListCount;

    /*
     * ST_MODULE
     */
    testListCount = 0;
    moduleTestSet[ST_MODULE].gfxTests = testList_st;

    while ( (testList_st[testListCount].gfx_dl != (Gfx *) NULL) &&
	    (testList_st[testListCount].toplevel_dl != (CaseProc_t) NULL) ) {
	testListCount++;
    }
    moduleTestSet[ST_MODULE].testListCount = testListCount;

    /*
     * TM_MODULE
     */
    testListCount = 0;
    moduleTestSet[TM_MODULE].gfxTests = testList_tm;

    while ( (testList_tm[testListCount].gfx_dl != (Gfx *) NULL) &&
	    (testList_tm[testListCount].toplevel_dl != (CaseProc_t) NULL) ) {
	testListCount++;
    }
    moduleTestSet[TM_MODULE].testListCount = testListCount;

    /*
     * TC_MODULE
     */
    testListCount = 0;
    moduleTestSet[TC_MODULE].gfxTests = testList_tc;

    while ( (testList_tc[testListCount].gfx_dl != (Gfx *) NULL) &&
	    (testList_tc[testListCount].toplevel_dl != (CaseProc_t) NULL) ) {
	testListCount++;
    }
    moduleTestSet[TC_MODULE].testListCount = testListCount;

    /*
     * TF_MODULE
     */
    testListCount = 0;
    moduleTestSet[TF_MODULE].gfxTests = testList_tf;

    while ( (testList_tf[testListCount].gfx_dl != (Gfx *) NULL) &&
	    (testList_tf[testListCount].toplevel_dl != (CaseProc_t) NULL) ) {
	testListCount++;
    }
    moduleTestSet[TF_MODULE].testListCount = testListCount;

    /*
     * CC_MODULE
     */
    testListCount = 0;
    moduleTestSet[CC_MODULE].gfxTests = testList_cc;

    while ( (testList_cc[testListCount].gfx_dl != (Gfx *) NULL) &&
	    (testList_cc[testListCount].toplevel_dl != (CaseProc_t) NULL) ) {
	testListCount++;
    }
    moduleTestSet[CC_MODULE].testListCount = testListCount;

    /*
     * BL_MODULE
     */
    testListCount = 0;
    moduleTestSet[BL_MODULE].gfxTests = testList_bl;

    while ( (testList_bl[testListCount].gfx_dl != (Gfx *) NULL) &&
	    (testList_bl[testListCount].toplevel_dl != (CaseProc_t) NULL) ) {
	testListCount++;
    }
    moduleTestSet[BL_MODULE].testListCount = testListCount;

    /*
     * MS_MODULE
     */
    testListCount = 0;
    moduleTestSet[MS_MODULE].gfxTests = testList_ms;

    while ( (testList_ms[testListCount].gfx_dl != (Gfx *) NULL) &&
	    (testList_ms[testListCount].toplevel_dl != (CaseProc_t) NULL) ) {
	testListCount++;
    }
    moduleTestSet[MS_MODULE].testListCount = testListCount;

    /*
     * Main game loop.  Send one (if specified) or all frames worth of DL to
     * be processed into test vectors, one after the other.  Optionally 
     * restrict ourselves to a single module as well; you can also run a single
     * frame from multiple module's, but that's a weird combination of -m and 
     * -f that isn't terribly interesting (but not worth ruling out).
     *
     * When dumping DL information, the commands will be sequentially numbered
     * such that we can generate separate .rdp files for each display list,
     * and thus generate separate tests to run as test vectors (providing
     * better fault isolation).
     *
     * Externally, we can write a tool to collapse redundant RDP initialization
     * from a sequence of DL's, so that if a second DL does not require as much
     * initialization as the first, we'll skip the unecessary set of mode
     * twiddling DL operations.
     */

    if (module_num < 0 || module_num > MODULE_COUNT) {
	startModule = 0;
	endModule = MODULE_COUNT;
    } else {
	startModule = module_num;
	endModule = module_num + 1;
    }

    for (moduleNum = startModule; moduleNum < endModule; moduleNum++) {

	if (controlFrame >  moduleTestSet[moduleNum].testListCount) {
	    controlFrame = 0;
	}
	if ( frame_count >= moduleTestSet[moduleNum].testListCount ) {
PRINTF("bogus -f <frame_count> %d ignored (testListCount for module %d is %d)\n",
	    frame_count, moduleNum,  moduleTestSet[moduleNum].testListCount);
PRINTF("NOTE: -f <frame_count> numbers frames from 0 .. (n-1)\n");
	    frame_count = -1;
	}

	if (frame_count < 0) {
	    frame_count = -1;
	    frameNum = -1;
	    endFrame = moduleTestSet[moduleNum].testListCount;
	}
	else {
	    frameNum = frame_count;
	}
	/*
	 * Reset this variable (which keeps us from drawing the same frame over
	 * and over again) each time we step to a new module.
	 */
	lastFrameNum = -2;

	/*
	 * IF
	 *    we are rendering the last frame for this module,
	 * THEN
	 *    generate a fullsync in toplevel.c
	 * ELSE
	 *    inhibit fullsync
	 */

	generateFullSync = 0;

	while (keep_going) {

	    /*
	     * Read controller
	     */
	    if (frame_count > -1) {
		readController(controllerSlot);
		/*
		 * If the user has requested a particular frame AND we are
		 * dumping to dram, generate the full sync because we are
		 * going to exit after the rendering of this single frame.
		 */
		if (rdp_DRAM_io) {
		    generateFullSync = 1;
		}
	    }
	    else {
		frameNum++;
		if (frameNum == endFrame) break;

		if (frameNum == (endFrame - 1) || 1 ) {
		    generateFullSync = 1;
		}
	    }
/*
PRINTF("moduleNum %d, frameNum %d, lastFrameNum %d %s\n", 
moduleNum, frameNum, lastFrameNum, moduleTestSet[moduleNum].gfxTests[frameNum].testname);
*/
	    if (frameNum != lastFrameNum) {

	    /*
	     * Always clear the frame buffer, but do it as a separate task 
	     * so that the RDP instructions don't make it into the RDP DL 
	     * we dump out as input for the test vector generator.
	     */

	    pixSize = moduleTestSet[moduleNum].gfxTests[frameNum].cfb_pixSize;
	    clear_framebuffers(pixSize);

	    /*
	     * Set up the video mode for either 16 bit or 32 bit pixels.
	     */
	    if ( pixSize == G_IM_SIZ_16b ) {
		osViSetMode(&osViModeTable[OS_VI_NTSC_LAN1]);
	    } else {
		osViSetMode(&osViModeTable[OS_VI_NTSC_LAN2]);
	    }

	    /*
	     * Provide a name for each display list segment we are to dump, so
	     * that rdpasciitordram can provide meaningful filenames.  Only the
	     * first frame's name is used for each module.  If a module needs
	     * to render in 16 bit and 32 bit, there will have to be two 
	     * sequences of frames (and two .rdram files), since the rdram
	     * file can only support 16 or 32 bit rendering (not both).
	     *
	     * Finally, a third type of rdram file could be a 1024x1024 cfb,
	     * which we use for edge walker testing.
	     */
	    {
	      char String[128];

	      if ( ( (rdp_DRAM_io) /*&& (frameNum == 0)*/ ) ||
		   ( (rdp_DRAM_io) && (frame_count > -1) ) ) {
		if ( pixSize == G_IM_SIZ_16b ) {
		  sprintf(String, "TEST %s %d bit framebuffer\n", 
			 moduleTestSet[moduleNum].gfxTests[frameNum].testname, 16);
		} else if ( pixSize == G_IM_SIZ_32b ) {
		  sprintf(String, "TEST %s %d bit framebuffer\n", 
			 moduleTestSet[moduleNum].gfxTests[frameNum].testname, 32);
		} else if ( pixSize == 1024 ) {
		  sprintf(String, "TEST %s 1024 x 1024 32 bit framebuffer\n", 
			 moduleTestSet[moduleNum].gfxTests[frameNum].testname);
		} else {
		  sprintf(String, "bogus pixel size %d for test %s\n", pixSize, 
			 moduleTestSet[moduleNum].gfxTests[frameNum].testname);
		}
		guParseString(String, 128);
	      }
	    }

	    /*
	     * set up pointers to build the display list.
	     */
	    tlistp = &tlist;
	    dynamicp = &dynamic;
	    glistp = &(dynamicp->glist[0]);
	    
	    /*
	     * Tell RCP where each segment is
	     */
	    gSPSegment(glistp++, 0, 0x0); /* K0 (physical) address segment */
	    gSPSegment(glistp++, STATIC_SEGMENT,
		       osVirtualToPhysical(staticSegment));
	    /*
	    gSPSegment(glistp++, TEXTURE_SEGMENT,
		       osVirtualToPhysical(textureSegment));
	     */
	    
	    if (rdp_mspanFlushFlag) {
		/*
		 * Force the pipeline to empty after the rendering of each
		 * display list item, to avoid incurring span buffer cache 
		 * coherency problems with DRAM r-m-w cycles.
		 */
		gDPPipelineMode(glistp++, G_PM_1PRIMITIVE);
	    }
	    
	    gSPDisplayList(glistp++, rdpinit_dl);
	    gSPDisplayList(glistp++, rspinit_dl);

	    /*
	     * If the rdp_regressionFlag is set, apply a 1/4 shrink to the 
	     * viewport so that the tests will render more quickly (because they
	     * won't be as fill limited).  Do this after we invoke the display
	     * list pointed to by setup_rspstate.
	     */
	    if (rdp_regressionFlag) {
		gSPDisplayList(glistp++, regression_viewport);
	    } 

	    /*
	     * Determine 16/32 bit rendering, then issue the appropriate
	     * gsDPSetColorImage() calls.  Width of Z buffer is determined
	     * by width of color frame buffer.
	     */

	    gDPSetDepthImage( glistp++, APP_ZBUFFER );
	    gDPPipeSync(glistp++);

	    if ( pixSize == G_IM_SIZ_16b ) {
		gDPSetColorImage( glistp++, G_IM_FMT_RGBA, G_IM_SIZ_16b, 
		    SCREEN_WD, OS_K0_TO_PHYSICAL(cfb_16_a) );
	    } else if ( pixSize == G_IM_SIZ_32b ) {
		gDPSetColorImage( glistp++, G_IM_FMT_RGBA, G_IM_SIZ_32b, 
		    SCREEN_WD, OS_K0_TO_PHYSICAL(cfb_16_a) );
	    } else {
		/*
		 * 1024 x 1024 framebuffer; we'll render in 32 bits.
		 */
		gDPSetColorImage( glistp++, G_IM_FMT_RGBA, G_IM_SIZ_32b, 
		    1024, OS_K0_TO_PHYSICAL(cfb_16_a) );
	    }
	    gDPPipeSync(glistp++);

	    /* end of standard display list part. */
	    
	    /* 
	     * Build graphics task:
	     */
	    tlistp->t.type = M_GFXTASK;
	    tlistp->t.flags = OS_TASK_DP_WAIT;
	    tlistp->t.ucode_boot = (u64 *) rspbootTextStart;
	    tlistp->t.ucode_boot_size = ((int)rspbootTextEnd - 
					 (int)rspbootTextStart);

	    /*
	     * choose which ucode to run:
	     */
	    if (rdp_DRAM_io) {
		/* re-direct output to DRAM: */
		if (uselines) {
		    tlistp->t.ucode = (u64 *) gspLine3D_dramTextStart;
		    tlistp->t.ucode_data = (u64 *) gspLine3D_dramDataStart;
		} else {
		    tlistp->t.ucode = (u64 *) gspFast3D_dramTextStart;
		    tlistp->t.ucode_data = (u64 *) gspFast3D_dramDataStart; 
		}
	    } else {
		/* RSP output over XBUS to RDP: */
		if (uselines) {
		    tlistp->t.ucode = (u64 *) gspLine3DTextStart;
		    tlistp->t.ucode_data = (u64 *) gspLine3DDataStart;
		} else {
		    tlistp->t.ucode = (u64 *) gspFast3DTextStart;
		    tlistp->t.ucode_data = (u64 *) gspFast3DDataStart;
		}
	    }
	    
	    tlistp->t.ucode_size = SP_UCODE_SIZE;
	    tlistp->t.ucode_data_size = SP_UCODE_DATA_SIZE;
	    
	    tlistp->t.dram_stack = (u64 *) &(dram_stack[0]);
	    tlistp->t.dram_stack_size = SP_DRAM_STACK_SIZE8;

	    if (rdp_DRAM_io) {
		tlistp->t.output_buff = (u64 *) &(rdp_output[0]);
		tlistp->t.output_buff_size = (u64 *) &rdp_output_len;
	    } else {
		tlistp->t.output_buff = (u64 *) 0x0;
		tlistp->t.output_buff_size = (u64 *) 0x0;
	    }

	    /* initial display list: */
	    tlistp->t.data_ptr = (u64 *) dynamicp->glist;
	    tlistp->t.data_size = ((int)(glistp - dynamicp->glist) *
				   sizeof (Gfx));
	    tlistp->t.yield_data_ptr = (u64 *) NULL;
	    tlistp->t.yield_data_size = 0xDA0;

	    /*
	     * Render the display list stored in the dynamic segment.
	     *
	     * The top level DL driver routines in toplevel.c will decide 
	     * whether to issue a FullSync or not depending on the value of
	     * generateFullSync.
	     *
	     * NOTE: this is done after setting up the task header so that
	     * the testcase procedure can modify the task list if neccessary
	     */

	    (*(moduleTestSet[moduleNum].gfxTests[frameNum].toplevel_dl)) 
	     (dynamicp, (moduleTestSet[moduleNum].gfxTests[frameNum].gfx_dl));

	    /*
	     * Force top-level display list to have an END.
	     */
	    gSPEndDisplayList(glistp++);
	    
	    assert((glistp - dynamicp->glist) < MAX_GRAPHICS_SIZE);


	    /*
	     * Can just flush 16KB and forget about each individual pieces
	     * of data to flush.
	     */
	    osWritebackDCacheAll();
	    
	    /* unwrap this macro for debugging:
	     *
	     *	osSpTaskStart(tlistp);
	     *
	     */

	    osSpTaskLoad(tlistp);
	    osSpTaskStartGo(tlistp);
	    
	    if (rdp_DRAM_io) {
		/* wait for SP completion */
		(void)osRecvMesg(&rspMessageQ, &dummyMessage, OS_MESG_BLOCK);
	/* 
	 * Wait for a vertical retrace to go by (Total hack, but necessary, as
	 * rdp_output_len is not updated properly unless we wait for a vertical,
	 * even though the SP interrupt has arrived).  Maybe we need to flush
	 * a data cache?
	 *
	 * Make sure there isn't an old retrace in queue 
	 * (assumes queue has a depth of 1) by checking to see if the 
	 * MQ is full.
	 */

	if (MQ_IS_FULL(&retraceMessageQ))
	    /*
	     * Eat the old retrace message prior to the real wait below.
	     */
	    (void)osRecvMesg(&retraceMessageQ, &dummyMessage, 
		OS_MESG_BLOCK);

	/* 
	 * Now really wait for vertical retrace to finish.
	 */
	(void)osRecvMesg(&retraceMessageQ, &dummyMessage, 
		OS_MESG_BLOCK);

PRINTF("SP done: rdp_output_len = 0x%llx\n", rdp_output_len);
		guParseRdpDL(&(rdp_output[0]),rdp_output_len,
                    GU_PARSERDP_DUMPONLY);
		dummy_val = osDpSetNextBuffer((void *)( &(rdp_output[0]) ), 
		    (u32) rdp_output_len);

	    } 

	    /* 
	     * wait for DP completion, but only if this DL had a fullsync.
	     */
	    if (generateFullSync == 1) {
		(void)osRecvMesg(&rdpMessageQ, &dummyMessage, 
		    OS_MESG_BLOCK);
	    }

	    /* 
	     * Wait for a vertical retrace to go by.
	     *
	     * Make sure there isn't an old retrace in queue 
	     * (assumes queue has a depth of 1) by checking to see if the 
	     * MQ is full.
	     */

	    if (MQ_IS_FULL(&retraceMessageQ))
		/*
		 * Eat the old retrace message prior to the real wait below.
		 */
		(void)osRecvMesg(&retraceMessageQ, &dummyMessage, 
		    OS_MESG_BLOCK);

	    /* 
	     * Now really wait for vertical retrace to finish.
	     */
	    (void)osRecvMesg(&retraceMessageQ, &dummyMessage, 
		    OS_MESG_BLOCK);
	    }
	    lastFrameNum = frameNum;
#if 0
#define SEE_X	30
#define SEE_Y	40
{
u32 *fbp;
u16 *zbp;
int c;
fbp=(u32 *)cfb_16_a;
zbp=zbuffer;
fbp += (320*SEE_Y) + SEE_X;
zbp += (320*SEE_Y) + SEE_X;
for (c=0; c<270; c++){
 PRINTF("%08x : %08x    (%04x)-> %04x , %01x",(int)fbp,*fbp,*zbp,*zbp>>2,*zbp&0x3);
 if (!(c%8)) PRINTF(" === %x",c/8);
 PRINTF("\n");
 fbp++;
 zbp++;
}
}
while(1) PRINTF ("-\n");
#endif /* 0 */

	    computeChecksum(pixSize, moduleNum, frameNum);

	    if (frame_count > -1) {
		if (rdp_DRAM_io) {
		    /* 
		     * User has requested a DRAM dump of a particular frame.
		     * Clear the "keep_going" variable so that we will call
		     * osExit(), which will force gload to exit & flush all
		     * pending character I/O.
		     */
		    keep_going = 0;
		} else {
		    frameNum = controlFrame;
		    if ((frameNum < 0) || (frameNum > 
		      moduleTestSet[moduleNum].testListCount - 1)) {
			frameNum = 0;
			controlFrame = 0;
		    }
		}
	    }
	}
    }

    /*
     * This will cause gload to terminate, possibly ending our ascii dump of 
     * the rdp display list.
     */
    if (rdp_DRAM_io) {
      char String[128];
      
      sprintf(String, "END DUMP\n");
      guParseString(String, 128);    
    }

    osExit();
}