Ejemplo n.º 1
0
/* loads the scene using DMA - blocks until done */
void load_scene(unsigned long long ea, scene_t *scene)
{
	unsigned int i = 0;
	object3d_t *objects = 0;
	pointlight_t *lights = 0;
	point_t *v = 0;

#if defined(_DEBUG) && _DEBUG > 2
	printf("Transferring %d bytes to LSaddr(%8lX) from EAadd(%8lX:%8lX) for SCENE\n",
		sizeof(scene_t),
		&scene,
		mfc_ea2h(ea),
		mfc_ea2l(ea));
#endif
	/* DMA request for scene */
	spu_mfcdma64(scene,
		mfc_ea2h(ea),
		mfc_ea2l(ea),
		sizeof(scene_t),
		SPUDMA_SCENE,
		MFC_GET_CMD);
	
	/* wait for request to complete */
	spu_writech(MFC_WrTagMask, 1 << SPUDMA_SCENE);
	mfc_read_tag_status_all();
	
	
	/* copy over objects */
	objects = _malloc_align(sizeof(object3d_t) * scene->nObjects, 4);
#if defined(_DEBUG) && _DEBUG > 2
	printf("Transferring %d bytes to LSaddr(%8lX) from EAadd(%8lX:%8lX) for OBJECTS\n",
		sizeof(object3d_t) * scene->nObjects,
		objects,
		mfc_ea2h(scene->objects_ea),
		mfc_ea2l(scene->objects_ea));
#endif
	/* initiate DMA */
	spu_mfcdma64(objects,
		mfc_ea2h(scene->objects_ea),
		mfc_ea2l(scene->objects_ea),
		sizeof(object3d_t) * scene->nObjects,
		SPUDMA_OBJECTS,
		MFC_GET_CMD);
	
	/* copy over lights */
	lights = _malloc_align(sizeof(pointlight_t) * scene->nLights, 4);	
#if defined(_DEBUG) && _DEBUG > 2
	printf("Transferring %d bytes to LSaddr(%8X) from EAadd(%8lX:%8lX) for LIGHTS\n",
		sizeof(pointlight_t) * scene->nLights,
		lights,
		mfc_ea2h(scene->lights_ea),
		mfc_ea2l(scene->lights_ea));
#endif
	/* initiate DMA for lights */
	spu_mfcdma64(lights,
		mfc_ea2h(scene->lights_ea),
		mfc_ea2l(scene->lights_ea),
		sizeof(pointlight_t) * scene->nLights,
		SPUDMA_LIGHTS,
		MFC_GET_CMD);
	
	/* wait for objects to complete */
	spu_writech(MFC_WrTagMask, 1 << SPUDMA_OBJECTS);
	mfc_read_tag_status_all();
	/* assign local store pointer to objects */
	scene->objects = objects;

	/* iterate each object locally */
	for(; i < scene->nObjects; ++i)
	{
		if(objects[i].geometryType == GEOMETRY_POLYGON)
		{
			/* allocate memory for vertex */
			v = _malloc_align(sizeof(point_t) 
				* objects[i].poly_obj.nVerticies, 4);
			/* initiate DMA to get verticies */
			spu_mfcdma64(v,
				mfc_ea2h(objects[i].poly_obj.vertex_ea),
				mfc_ea2l(objects[i].poly_obj.vertex_ea),
				sizeof(point_t)
				* objects[i].poly_obj.nVerticies,
				SPUDMA_VERTEXES,
				MFC_GET_CMD);
			/* assign local store pointer - WARNING - safe? */
			objects[i].poly_obj.vertex = v;				
		}
	}
	
	/* wait for all DMA to finish (vertexes, lights) */
	spu_writech(MFC_WrTagMask, 1 << SPUDMA_LIGHTS |
				1 << SPUDMA_VERTEXES );
	mfc_read_tag_status_all();
	/* assign local store lights pointer */
	scene->lights = lights;
}
Ejemplo n.º 2
0
void process_render_tasks(unsigned long eah_render_tasks, unsigned long eal_render_tasks)
{
    const vec_uchar16 SHUFFLE_MERGE_BYTES = (vec_uchar16) {	// merge lo bytes from unsigned shorts (array)
        1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31
    };

    const vec_uchar16 SHUFFLE_GET_BUSY_WITH_ONES = (vec_uchar16) {	// get busy flag with ones in unused bytes
        0xc0, 0xc0, 2, 3, 0xc0,0xc0,0xc0,0xc0, 0xc0,0xc0,0xc0,0xc0
    };

    const vec_uchar16 ZERO_BYTES = (vec_uchar16) spu_splats(0);

    char trianglebuffer[ 256 + TRIANGLE_MAX_SIZE ];

    char	sync_buffer[128+127];
    void*	aligned_sync_buffer = (void*) ( ((unsigned long)sync_buffer+127) & ~127 );

    RenderableCacheLine* cache = (RenderableCacheLine*) aligned_sync_buffer;
    unsigned long long cache_ea;

    spu_mfcdma64(&cache_ea, eah_render_tasks, eal_render_tasks, sizeof(cache_ea), 0, MFC_GET_CMD);
    mfc_write_tag_mask(1<<0);
    mfc_read_tag_status_all();

    while (cache_ea) {
        // terminate immediately if possible
        if (spu_stat_in_mbox())
            return;

        // read the cache line
        spu_mfcdma64(cache, mfc_ea2h(cache_ea), mfc_ea2l(cache_ea), 128, 0, MFC_GETLLAR_CMD);
        spu_readch(MFC_RdAtomicStat);

        unsigned int endTriangle = cache->endTriangle;
        vec_ushort8 testTriangle = spu_splats((unsigned short) endTriangle);

        // first look for short chunks
        vec_uchar16 next = cache->chunkNext;
        vec_uchar16 nextmask = spu_and(next, spu_splats((unsigned char)CHUNKNEXT_MASK));

        // change next to word offset, note we don't care what the low bit shifted in is
        vec_uchar16 firstshuf = (vec_uchar16) spu_sl( (vec_ushort8)nextmask, 1 );
        vec_uchar16 trishufhi = spu_or ( firstshuf, spu_splats((unsigned char) 1));
        vec_uchar16 trishuflo = spu_and( firstshuf, spu_splats((unsigned char) 254));

        vec_ushort8 start0 = cache->chunkStart[0];
        vec_ushort8 start1 = cache->chunkStart[1];

        vec_ushort8 nstart0 = spu_shuffle( start0, start1, spu_shuffle( trishuflo, trishufhi, SHUF0 ) );
        vec_ushort8 nstart1 = spu_shuffle( start0, start1, spu_shuffle( trishuflo, trishufhi, SHUF1 ) );

        vec_ushort8 starteq0 = spu_cmpeq( nstart0, spu_splats((unsigned short)0) );
        vec_ushort8 starteq1 = spu_cmpeq( nstart1, spu_splats((unsigned short)0) );

        vec_ushort8 end0 = spu_sel( nstart0, spu_splats((unsigned short)4096), starteq0);
        vec_ushort8 end1 = spu_sel( nstart1, spu_splats((unsigned short)4096), starteq1);

        vec_ushort8 len0 = spu_sub( end0, start0);
        vec_ushort8 len1 = spu_sub( end1, start1);

        vec_ushort8 small0 = spu_cmpgt( spu_splats((unsigned short)17), len0);
        vec_ushort8 small1 = spu_cmpgt( spu_splats((unsigned short)17), len1);
        vec_uchar16 small = (vec_uchar16) spu_shuffle( small0, small1, MERGE );
        vec_uint4 smallChunkGather = spu_gather(small);

        // check to see if chunk is already at the last triangle
        vec_uint4 doneChunkGather = spu_gather( (vec_uchar16) spu_shuffle(
                (vec_uchar16) spu_cmpeq(testTriangle, cache->chunkTriangle[0]),
                (vec_uchar16) spu_cmpeq(testTriangle, cache->chunkTriangle[1]),
                SHUFFLE_MERGE_BYTES) );

        // check if the chunk is free
        vec_uint4 freeChunkGather = spu_gather(
                                        spu_cmpeq( spu_splats( (unsigned char) CHUNKNEXT_FREE_BLOCK ), cache->chunkNext ) );

        // check to see if the chunk is being processed
        vec_uint4 busyChunkGather = spu_gather(
                                        spu_cmpgt( cache->chunkNext, //spu_and(cache->chunkNext, CHUNKNEXT_MASK),
                                                spu_splats( (unsigned char) (CHUNKNEXT_BUSY_BIT-1) ) ) );

        // doneChunkGather, freeChunkGather, busyChunkGather - rightmost 16 bits of word 0
        // note that if freeChunkGather is true then busyChunkGather must also be true

        // done=false, free=false, busy=false -> can process
        // free=false, busy=false -> can be merged

        // decide which chunk to process
        vec_uint4 mayProcessGather = spu_nor( doneChunkGather, busyChunkGather );
        vec_uint4 mayProcessShortGather = spu_and( mayProcessGather, smallChunkGather );

        vec_uint4 shortSelMask = spu_cmpeq( mayProcessShortGather, spu_splats(0U) );
        vec_uint4 mayProcessSelection = spu_sel( mayProcessShortGather, mayProcessGather, shortSelMask );

        /*
        		if (!spu_extract(shortSelMask, 0))
        			printf("taken short: may=%04x short=%04x mayshort=%04x mask=%04x sel=%04x\n",
        				spu_extract(mayProcessGather, 0) & 0xffff,
        				spu_extract(smallChunkGather, 0),
        				spu_extract(mayProcessShortGather, 0),
        				spu_extract(shortSelMask, 0) & 0xffff,
        				spu_extract(mayProcessSelection, 0) & 0xffff );
        */

        vec_uint4 mayProcessBits = spu_sl( mayProcessSelection, 16);
        unsigned int chunkToProcess = spu_extract( spu_cntlz( mayProcessBits ), 0);
        unsigned int freeChunk = spu_extract( spu_cntlz( spu_sl( freeChunkGather, 16 ) ), 0);

        // if there's nothing to process, try the next cache line in the rendering tasks list
        if (!spu_extract(mayProcessBits, 0)) {
trynextcacheline:
            cache_ea = cache->next;
            // sleep();
            continue;
        }

        unsigned int chunkStart    	= cache->chunkStartArray   [chunkToProcess];
        unsigned int chunkTriangle	= cache->chunkTriangleArray[chunkToProcess];
        unsigned int chunkNext		= cache->chunkNextArray	   [chunkToProcess] & CHUNKNEXT_MASK;
        unsigned int chunkEnd		= (cache->chunkStartArray  [chunkNext]-1) & (NUMBER_OF_TILES-1);
        unsigned int chunkLength	= 1 + chunkEnd-chunkStart;

        // only need an extra block if the block is especially long
        if (chunkLength <= NUMBER_OF_TILES_PER_CHUNK) {
            freeChunk = 32;
        }

        // mark this block as busy
        cache->chunkNextArray[chunkToProcess] |= CHUNKNEXT_BUSY_BIT;

        // if there's at least one free chunk, claim it
        if (freeChunk != 32) {
            cache->chunkNextArray[freeChunk] = CHUNKNEXT_RESERVED;
            cache->chunkTriangleArray[freeChunk] = chunkTriangle;
        }

        // write the cache line back
        spu_mfcdma64(cache, mfc_ea2h(cache_ea), mfc_ea2l(cache_ea), 128, 0, MFC_PUTLLC_CMD);
        if (spu_readch(MFC_RdAtomicStat) & MFC_PUTLLC_STATUS)
            continue;

#ifdef INFO
        printf("[%d] Claimed chunk %d (%d-%d len %d) at tri %x end %x with free chunk %d\n", _SPUID,
               chunkToProcess, chunkStart, chunkEnd, chunkLength, chunkTriangle, endTriangle,
               freeChunk!=32 ? freeChunk : -1 );
//		debug_render_tasks(cache);
#endif

        Triangle* triangle;
        int firstTile;
        do {
            // read the triangle data for the current triangle
            unsigned int extra = chunkTriangle & 127;
            unsigned long long trianglebuffer_ea = cache_ea + TRIANGLE_OFFSET_FROM_CACHE_LINE + (chunkTriangle & ~127);
            triangle = (Triangle*) (trianglebuffer+extra);
            unsigned int length = (extra + TRIANGLE_MAX_SIZE + 127) & ~127;

            // ensure DMA slot available
            do {} while (!spu_readchcnt(MFC_Cmd));

            spu_mfcdma64(trianglebuffer, mfc_ea2h(trianglebuffer_ea), mfc_ea2l(trianglebuffer_ea),
                         length, 0, MFC_GET_CMD);
            mfc_write_tag_mask(1<<0);
            mfc_read_tag_status_all();

            // get the triangle deltas
            firstTile = findFirstTriangleTile(triangle, chunkStart, chunkEnd);

            if (firstTile>=0)
                break;

            // no match, try next triangle
            chunkTriangle = triangle->next_triangle;
        } while (chunkTriangle != endTriangle);

        // if we actually have something to process...
        if (firstTile>=0) {
            // the "normal" splitting will now become:
            // chunkStart .. (firstTile-1)	-> triangle->next_triangle
            // firstTile .. (firstTile+NUM-1) -> chunkTriangle (BUSY)
            // (firstTile+NUM) .. chunkEnd -> chunkTriangle (FREE)

            int tailChunk;
            int thisChunk;
            int nextBlockStart;
            int thisBlockStart;
            int realBlockStart;
            do {
retry:
                // read the cache line
                spu_mfcdma64(cache, mfc_ea2h(cache_ea), mfc_ea2l(cache_ea), 128, 0, MFC_GETLLAR_CMD);
                spu_readch(MFC_RdAtomicStat);

                // calculate start of next block
                nextBlockStart = firstTile + NUMBER_OF_TILES_PER_CHUNK;
                if (nextBlockStart > chunkEnd)
                    nextBlockStart = chunkEnd+1;

                // calculate start of block to mark as busy
                thisBlockStart = nextBlockStart - NUMBER_OF_TILES_PER_CHUNK;
                if (thisBlockStart < chunkStart)
                    thisBlockStart = chunkStart;
                realBlockStart = thisBlockStart;

#ifdef INFO
                printf("[%d] nextBlockStart=%d, realBlockStart=%d, thisBlockStart=%d, chunkStart=%d\n", _SPUID,
                       nextBlockStart, realBlockStart, thisBlockStart, chunkStart);
#endif


                // allocate some more free chunks
                vec_uint4 freeChunkGather2 = spu_sl(spu_gather(spu_cmpeq(
                                                        spu_splats((unsigned char)CHUNKNEXT_FREE_BLOCK), cache->chunkNext)), 16);
                unsigned int freeChunk2 = spu_extract(spu_cntlz(freeChunkGather2), 0);

                if (freeChunk == 32) {
                    // if we didn't have one before, try again
                    freeChunk = freeChunk2;

                    // and try to get the second one
                    freeChunkGather2 = spu_andc( freeChunkGather2, spu_promote(0x80000000>>freeChunk2, 0) );
                    freeChunk2 = spu_extract(spu_cntlz(freeChunkGather2), 0);
                } else {
                    // speculatively clear the free chunk just in case we don't need it
                    cache->chunkNextArray[freeChunk] = CHUNKNEXT_FREE_BLOCK;
                }

#ifdef INFO
                printf("[%d] Free chunks %d and %d, cN=%d, nBS=%d, cE=%d, tBS=%d, cS=%d\n",
                       _SPUID, freeChunk, freeChunk2, chunkNext, nextBlockStart, chunkEnd, thisBlockStart, chunkStart );
#endif

                // mark region after as available for processing if required
                if (nextBlockStart < chunkEnd) {
                    if (freeChunk==32) {
                        // if no free chunk, relinquish entire block and write back
                        cache->chunkNextArray[chunkToProcess] = chunkNext;
                        spu_mfcdma64(cache, mfc_ea2h(cache_ea), mfc_ea2l(cache_ea), 128, 0, MFC_PUTLLC_CMD);
                        // if writeback failed, we *might* have a free block, retry
                        if (spu_readch(MFC_RdAtomicStat) & MFC_PUTLLC_STATUS)
                            goto retry;

                        // otherwise give up and try the next cache line
                        goto trynextcacheline;
                    }
                    cache->chunkStartArray[freeChunk] = nextBlockStart;
                    cache->chunkNextArray[freeChunk] = chunkNext;
                    cache->chunkTriangleArray[freeChunk] = chunkTriangle;
                    cache->chunkNextArray[chunkToProcess] = freeChunk | CHUNKNEXT_BUSY_BIT;
                    tailChunk = freeChunk;
#ifdef INFO
                    printf("[%d] Insert tail, tailChunk=%d, chunkNext=%d, chunkToProcess=%d\n", _SPUID, tailChunk, chunkNext, chunkToProcess);
                    debug_render_tasks(cache);
#endif
                } else {
                    // we're gonna use freeChunk2 for the "in front" block, as we've not
                    // used freeChunk, let's use it as it's more likely to have a free chunk
                    freeChunk2 = freeChunk;
                    tailChunk = chunkNext;
                }

                // mark region before as available if required and possible
                thisChunk = chunkToProcess;
                if (thisBlockStart > chunkStart) {
                    if (freeChunk2 != 32) {
                        // mark this region as busy
                        cache->chunkStartArray[freeChunk2]=thisBlockStart;
                        cache->chunkNextArray[freeChunk2]=tailChunk | CHUNKNEXT_BUSY_BIT;
                        cache->chunkTriangleArray[freeChunk2]=chunkTriangle;

                        // mark region before as available for processing
                        cache->chunkNextArray[chunkToProcess]=freeChunk2;
                        cache->chunkTriangleArray[chunkToProcess]=triangle->next_triangle;
                        thisChunk = freeChunk2;
#ifdef INFO
                        printf("[%d] Insert new head, tailChunk=%d, chunkNext=%d, thisChunk=%d\n", _SPUID, tailChunk, chunkNext, thisChunk);
                        debug_render_tasks(cache);
#endif
                    } else {
                        // need to keep whole block, update info and mark bust
                        cache->chunkTriangleArray[chunkToProcess]=chunkTriangle;
                        cache->chunkNextArray[chunkToProcess]=tailChunk | CHUNKNEXT_BUSY_BIT;
                        realBlockStart = chunkStart;
                        printf("[%d] Keep whole block, tailChunk=%d, chunkNext=%d, thisChunk=%d\n", _SPUID, tailChunk, chunkNext, thisChunk);
                        debug_render_tasks(cache);
#ifdef INFO
#endif
                        sleep();
                    }
                }

                // merge chunks
                merge_cache_blocks(cache);

                // write the cache line back
                spu_mfcdma64(cache, mfc_ea2h(cache_ea), mfc_ea2l(cache_ea), 128, 0, MFC_PUTLLC_CMD);
            } while (spu_readch(MFC_RdAtomicStat) & MFC_PUTLLC_STATUS);

            // finally after the write succeeded, update the variables
            chunkNext = tailChunk;
            chunkToProcess = thisChunk;
            chunkStart = firstTile; //thisBlockStart;
            chunkLength = nextBlockStart - firstTile;
            chunkEnd = chunkStart + chunkLength - 1;
            freeChunk = 32;

            // now we can process the block up to endTriangle
            initTileBuffers(thisBlockStart, chunkEnd);

            int ok=0;
            while (chunkTriangle != endTriangle) {
#ifdef INFO
                printf("[%d] Processing chunk %d at %4d len %4d, triangle %04x first=%d tbs=%d\n",
                       _SPUID, chunkToProcess, chunkStart, chunkLength,
                       chunkTriangle, firstTile, thisBlockStart);
#endif
                // and actually process that triangle on these chunks
                processTriangleChunks(triangle, cache, thisBlockStart, chunkEnd, chunkTriangle, ok);
                ok=1;
#ifdef PAUSE
                sleep();
#endif
                // and advance to the next-triangle
                chunkTriangle = triangle->next_triangle;

                // this should only ever happen if we're running really low on cache line slots
                // basically, if we pick up a block with more than NUMBER_OF_TILES_PER_CHUNK and
                // there's no slot to store the pre-NUMBER_OF_TILES_PER_CHUNK tiles.
                // in this case, we process from thisBlockStart only (because we know that from
                // chunkStart to there has no result) and then we only process one triangle
                if (chunkStart != realBlockStart) {
                    /*
                    printf("[%d] chunkStart=%d != realBlockStart %d, chunkEnd=%d, "
                    	"firstTile=%d chunk=%d\n",
                    	_SPUID, chunkStart, realBlockStart, chunkEnd,
                    	firstTile, chunkToProcess);
                    debug_render_tasks(cache);
                    */

                    // abort the while loop
                    break;
                }

                // read the next triangle
                unsigned int extra = chunkTriangle & 127;
                unsigned long long trianglebuffer_ea = cache_ea + TRIANGLE_OFFSET_FROM_CACHE_LINE + (chunkTriangle & ~127);
                triangle = (Triangle*) (trianglebuffer+extra);
                unsigned int length = (extra + TRIANGLE_MAX_SIZE + 127) & ~127;

                // ensure DMA slot available
                do {} while (!spu_readchcnt(MFC_Cmd));

                spu_mfcdma64(trianglebuffer, mfc_ea2h(trianglebuffer_ea),
                             mfc_ea2l(trianglebuffer_ea), length, 0, MFC_GET_CMD);
                mfc_write_tag_mask(1<<0);
                mfc_read_tag_status_all();
            } // until chunkTriangle == endTriangle

            // flush any output buffers
            flushTileBuffers(thisBlockStart, chunkEnd);

        } // firstTile>=0
Ejemplo n.º 3
0
Triangle* getTriangleBuffer(Context* context)
{
	// if we've already allocated a triangle buffer (and we're in the same context)
	if (context == _currentTriangleContext && _currentTriangle)
		return _currentTriangle;

	// trash the default values
	_currentTriangleContext	= context;
	_currentTriangle	= NULL;

	// read the current renderable cache line to ensure there is room for the triangle data
	// in the cache line buffer; we do this by comparing against all 16 cache line blocks
	// to make sure that extending the write pointer wouldn't clobber the data

	unsigned long long cache_ea = context->renderableCacheLine;
	if (cache_ea == 0)
		return NULL;
	char cachebuffer[128+127];
	RenderableCacheLine* cache = (RenderableCacheLine*) ( ((unsigned int)cachebuffer+127) & ~127 );

	// printf("GTB: reading to %x from %x:%x\n", cache, mfc_ea2h(cache_ea), mfc_ea2l(cache_ea));

	spu_mfcdma64(cache, mfc_ea2h(cache_ea), mfc_ea2l(cache_ea), 128, 0, MFC_GETLLAR_CMD);
	spu_readch(MFC_RdAtomicStat);

	// extendvalid = ( read<=write && test<end ) || ( read>write && test<read )
	// extendvalid = ( read>write && read>test ) || ( read<=write && end>test )
	// simplifies to	extendvalid = selb(end, read, read>write) > test
	// or			extendvalid = selb(end>test, read>test, read>write)
	// rewind = next >= end
	// rewindvalid = read != 0
	// valid = extendvalid && (!rewind || rewindvalid)
	// 	 = extendvalid && (!rewind || !rewindinvalid)
	// 	 = extendvalid && !(rewind && rewindinvalid)
	// invalid = ! (extendvalid && !(rewind && rewindinvalid))
	//         = (!extendvalid || (rewind && rewindinvalid))

	vec_ushort8 v_writeptr		= spu_splats( cache->endTriangle );
	vec_ushort8 v_readptr0		= cache->chunkTriangle[0];
	vec_ushort8 v_readptr1		= cache->chunkTriangle[1];
	vec_ushort8 v_testptr		= spu_add(v_writeptr,   TRIANGLE_MAX_SIZE);
	vec_ushort8 v_nextptr		= spu_add(v_writeptr, 2*TRIANGLE_MAX_SIZE);
	vec_ushort8 v_endptr		= spu_splats( (unsigned short)TRIANGLE_BUFFER_SIZE);

	vec_ushort8 v_zero		= spu_splats( (unsigned short) 0 );
	vec_uchar16 v_merger		= (vec_uchar16) { 1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31 };

	vec_ushort8 v_max0_test		= spu_sel( v_endptr, v_readptr0, spu_cmpgt( v_readptr0, v_writeptr ) );
	vec_ushort8 v_max1_test		= spu_sel( v_endptr, v_readptr1, spu_cmpgt( v_readptr1, v_writeptr ) );
	vec_ushort8 v_extend0_valid	= spu_cmpgt( v_max0_test, v_testptr );
	vec_ushort8 v_extend1_valid	= spu_cmpgt( v_max1_test, v_testptr );
	vec_ushort8 v_rewind0_invalid	= spu_cmpeq( v_readptr0, v_zero );
	vec_ushort8 v_rewind1_invalid	= spu_cmpeq( v_readptr1, v_zero );
	vec_ushort8 v_rewind8		= spu_cmpgt( v_nextptr, v_endptr );

	vec_uchar16 v_extend_valid	= (vec_uchar16) spu_shuffle( v_extend0_valid, v_extend1_valid, v_merger );
	vec_uchar16 v_rewind_invalid	= (vec_uchar16) spu_shuffle( v_rewind0_invalid, v_rewind1_invalid, v_merger );
	vec_uchar16 v_rewind		= (vec_uchar16) v_rewind8;

	vec_uchar16 v_valid_rhs		= spu_and( v_rewind_invalid, v_rewind );
	vec_uchar16 v_invalid		= spu_orc( v_valid_rhs, v_extend_valid );

	// check to see if the chunk is being processed
	vec_uint4 v_free = spu_gather(
		spu_cmpeq( spu_splats( (unsigned char) CHUNKNEXT_FREE_BLOCK ), cache->chunkNext ) );
	vec_uint4   v_invalid_bits	= spu_andc( spu_gather( v_invalid ), (vec_uint4) v_free );

	// if any of the bits are invalid, then no can do
	if ( spu_extract(v_invalid_bits, 0) ) {
		return NULL;
	}

	// fetch in the data before this triangle in the cache buffer
	unsigned int offset = cache->endTriangle;
	_currentTriangleBufferExtra = offset & 127;
	unsigned long long trianglebuffer_ea = cache_ea + TRIANGLE_OFFSET_FROM_CACHE_LINE + (offset & ~127);
	if (_currentTriangleBufferExtra) {
		spu_mfcdma64(_currentTriangleBuffer, mfc_ea2h(trianglebuffer_ea), mfc_ea2l(trianglebuffer_ea), 128, 0, MFC_GET_CMD);

		// ensure DMA did actually complete
		mfc_write_tag_mask(1<<0);
		mfc_read_tag_status_all();
	}

	// final bit of initialisation
	_currentTriangle = (Triangle*) (_currentTriangleBuffer+_currentTriangleBufferExtra);
	_currentTriangleOffset = offset;
	_currentTriangleRewind = v_rewind8;
	_currentTriangleCacheEndTriangleEAL = mfc_ea2l(cache_ea) + (((char*)&cache->endTriangle) - ((char*)cache));
	_currentTriangleCacheEndTriangleEAH = mfc_ea2h(cache_ea); 
	_currentTriangleBufferEA = trianglebuffer_ea; 

	// printf("Allocated new triangle buffer: %x\n", offset);

	// and return the buffer ready to go
	return _currentTriangle;
}