/**************************************************************************
 *                                                                        *
 *               Copyright (C) 1994, 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 *  unpublished  proprietary  information of Silicon Graphics, Inc., and  *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
 * verification. The idea is that each frame possibly uses a different
 * dataset (see static.c) to render a particular test.
 *
 * Since this 'application' is not an animation, it is a sequence
 * of single-frame tests, it is best to run it without double buffering (-s):
 *
 *		emulate -s -a "-o 2" rdpverif rom
 *
 * Use the -a flag to emulate 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.
 *
 *  FILES THAT ARE SHARED WITH OTHER TESTS: (don't modify)
 *
 *	main.c
 *	init.c
 *	dynamic.c
 *	zbuffer.c
 *	spec
 *
 *  FILES THAT ARE TEST-SPECIFIC: (may modify)
 *
 *	static.c		(holds static display list data for tests)
 *	rdpverif.h		(exports info to main.c, etc.)
 * 	rdpverif.c		(holds test frame procedures, etc.)
 *	<any related include files>	(textures, etc.)
 *	Makefile			(make test rules)
 *
 *
 * TO ADD A TEST CASE:
 * New test cases require that you add:
 *
 *	static.c	- add static display list for the test.
 *	rdpverif.h	- extern the static display list
 *	rdpverif.c	- add the procedure to be called to
 *			  perform the test case. (also put it
 *			  in the testCaseProc table)
 *			- put the static display list pointer
 *			  in the testList table.
 *
 * plus 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 Oct 14 11:18:33 PDT 1994
 *
 */
#include <ultra64.h>
#include <verify.h>
#include <ramrom.h>

#include "rdpverif.h"

#define PRINTF  osSyncPrintf

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

/* 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[];

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

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)];

/* 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:
 */
u64 dram_stack[SP_DRAM_STACK_SIZE64] __attribute__((aligned (16)));
u64 rdp_output[4096*16] __attribute__((aligned (16)));

/*
 * 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) */
    NULL,		        /* 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) */
};

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

/*
 * global variables for arguments, to control test cases
 */
static int	frame_count = 0;
static int	rdp_DRAM_io = 0;
static int	rdp_regressionFlag = 0;
static int	rdp_mspanFlushFlag = 0;
static int	debugflag = 0;
static int      draw_buffer = 0;
static int      uselines  = 0;
static int      realClear = 0;
static int	cfb_size = G_IM_SIZ_16b;

static void     *cfb_ptrs[2];
static u64	ramrombuf[RAMROM_MSG_SIZE/8];

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

#define PI_BASE	((u32)(0xb0700000))
u32		dummy_val;

/*
 * 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;
    char	*arglist[32], **argv = arglist;	/* max 32 args */
    char	*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]) {
	  case 'c':
	    realClear = 1;
	    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 = 0;
		PRINTF("bad frame_count [-f]!\n");
	    }
	    argc--;
	    argv++;
	    break;

	  case 'g':
	    debugflag = 1;
	    break;

	  case 't':		/* "true color" ie 24 bit RGB */
	    cfb_size = G_IM_SIZ_32b;
	    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;
	  default:
	    PRINTF("option [%s] not recognized.\n", argv[1]);
	    break;
	}
	argc--;
	argv++;
    }
}

boot(void *arg)
{
    int i, *pr;
    u32 *argp;
    u32 argbuf[16];
    
    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 */
    }

    /* Parse the  ptions */
    parse_args((char *)argbuf);

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

    /* never reached */
}

static void
idle(void *arg)
{
    /* Initialize video */
    osCreateViManager(OS_PRIORITY_VIMGR);
    osViSetMode(&osViModeTable[OS_VI_NTSC_LAN1]);
    
    /*
     * 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)
	osStartThread(&mainThread);

    /*
     * Become the idle thread
     */
    osSetThreadPri( 0, 0 );
    for(;;);
}

static void
mainproc(void *arg)
{
    int		keep_going = 1;
    OSTask	*tlistp;
    Dynamic	*dynamicp;
    char	*staticSegment;
    int 	testListCount;

    /* this structure must be aligned to 64-bit boundary: */
    VerifyInfo	*info = (VerifyInfo *)osPhysicalToVirtual(VERIFY_INFO_PHYSADDR);

    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);
    
    /*
     * Stick the static segment right after the code/data segment
     */
    staticSegment = _codeSegmentEnd;
    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);
    
    /* 
     * count the test cases. Can't use sizeof(), do it
     * here.
     */
    testListCount = 0;
    while (testList[testListCount] != (Gfx *) NULL)
      testListCount++;
    testListCount--;

    /*
     * Main game loop
     */
    while (keep_going) {

	/*
	 * pointers to build the display list.
	 */
	tlistp = &tlist;
	dynamicp = &dynamic;
	glistp = &(dynamicp->glist[0]);
	
	/*
	 * Tell RCP where each segment is
	 */
	gSPSegment(glistp++, STATIC_SEGMENT,
		   osVirtualToPhysical(staticSegment));
	
	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++, setup_rspstate);
        gSPDisplayList(glistp++, setup_rdpstate);
	
	/*
	 * 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);
	} 

	if (realClear) {
	  gSPDisplayList(glistp++, clear_fb);
	  if (cfb_size == G_IM_SIZ_32b) {
	    gSPDisplayList(glistp++, clear_32fb);
	  }
	} else {
	  gSPDisplayList(glistp++, fake_clear_fb);
	  if (cfb_size == G_IM_SIZ_32b) {
	    gSPDisplayList(glistp++, fake_clear_32fb);
	  }
	}

	/* end of standard display list part. */
	/*
	 * Setup our dynamic segment, depending on which test:
	 * For frame_count greater than the testlist count we just
	 * do the init part - useful to get it to clear.
	 */
	if (frame_count <= testListCount) {
	    (*testCaseProcs[frame_count])(dynamicp, frame_count);
	}

	/*
	 * Force top-level display list to have an END.
	 */
	gDPFullSync(glistp++);
	gSPEndDisplayList(glistp++);
	
	/* 
	 * 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 *) &(info->rdpListSize);
	} 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;

	/*
	 * Fill in verify info structure: (for verification ONLY)
	 */
	info->rspListAddr    = (unsigned int) tlistp->t.data_ptr;
	info->rspListSize    = tlistp->t.data_size;
	info->rspListType    = ((4096-128) << 12) | tlistp->t.type;
	info->ucodeTextAddr  = (unsigned int) tlistp->t.ucode;
	info->ucodeDataAddr  = (unsigned int) tlistp->t.ucode_data;
	info->ucodeDataSize  = tlistp->t.ucode_data_size;
	info->dramStackAddr  = (unsigned int) tlistp->t.dram_stack;
	info->dramStackSize  = tlistp->t.dram_stack_size;
	info->rdpListSize    = 0;
	info->rdpListAddr    = (unsigned int) tlistp->t.output_buff;
	info->bootUcodeAddr  = (unsigned int) tlistp->t.ucode_boot;
	info->width          = SCREEN_WD;
	info->height         = SCREEN_HT;
	info->frameBuffer0Addr = osVirtualToPhysical(cfb_16_a);
	info->frameBuffer1Addr = osVirtualToPhysical(cfb_16_b);
	info->zBufferAddr    = osVirtualToPhysical(zbuffer);
	info->magicNumber    = MAGICNUMBER; /* Kaelyn's birthday: 0x102594 */
	info->version        = VERSION;
	info->frameBufferFormat = G_IM_FMT_RGBA;
	info->frameBufferSize =  cfb_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);
	else
	    /* ignore SP completion */
	    (void)osRecvMesg(&rdpMessageQ, &dummyMessage, OS_MESG_BLOCK);

	osInvalDCache(&info->rdpListSize, 16);
	if (rdp_DRAM_io) {
	    PRINTF("Dp display list addr: 0x%x, len %d\n", &(rdp_output[0]),
		   info->rdpListSize);
	    dummy_val = osDpSetNextBuffer(&(rdp_output[0]),
					  info->rdpListSize);
	    (void)osRecvMesg(&rdpMessageQ, &dummyMessage, OS_MESG_BLOCK);
	}

	/* setup to swap buffers */
	osViSwapBuffer(cfb_ptrs[draw_buffer]);

	/*
	 * Only do 1 frame, then break out and wait forever
	 */
	keep_going = 0;

	/* Make sure there isn't an old retrace in queue 
	 * (assumes queue has a depth of 1) 
	 */
	if (MQ_IS_FULL(&retraceMessageQ))
	    (void)osRecvMesg(&retraceMessageQ, &dummyMessage, OS_MESG_BLOCK);
	
	/* Wait for Vertical retrace to finish swap buffers */
	(void)osRecvMesg(&retraceMessageQ, &dummyMessage, OS_MESG_BLOCK);

    }

    for(;;);

}