.equ vertexTop, 0x3f0
.equ numInAttribs, 3
.equ numOutAttribs, 3
.equ numOutBuf, 2
.equ vertCount, ((vertexTop-numOutBuf)/(numInAttribs*2+numOutAttribs*numOutBuf))
.equ offset, (vertCount*numInAttribs)
.equ outBuf1, (2*offset) 
;.equ outSize, ((vertexTop-outBuf1-numOutBuf)/numOutBuf)
.equ outSize, (1 + vertCount*numOutAttribs)
.equ outBuf2, (outBuf1+outSize) 

.equ clipVertCount, ((vertexTop-outBuf1-numOutBuf)/(numInAttribs+numOutAttribs*numOutBuf))
.equ clipOutSize, (1 + clipVertCount*numOutAttribs)
.equ clipOutBuf2, (outBuf1+clipOutSize)
.equ clipBuf, (clipOutBuf2+clipOutSize)	

; up to 9 verts => 7 tris per poly in worst case, a lot for lists unfortunately
.equ clipVertLimit, (clipVertCount-21)

.include "defines.inc"

.balign 16,0

.global vu1_im3d_TLclip
vu1_im3d_TLclip:


;;; Register allocation
;;;  vf25	colorScale
;;;  vf26	xyzwOffset
;;;  vf27	xyzwScale
;;;  vf28-vf31	matrix

DMAret *
MPG 0, *
.vu
start:
	nop				iaddiu vi12,vi00,outBuf1
	nop				iaddiu vi13,vi00,clipOutBuf2

	nop				lq vf28,matrix0(vi00)
	nop				lq vf29,matrix1(vi00)
	nop				lq vf30,matrix2(vi00)
	nop				lq vf31,matrix3(vi00)
	nop				lq vf27,xyzwScale(vi00)
	nop				lq vf26,xyzwOffset(vi00)
	nop				lq vf25,colorScale(vi00)
restart:
	nop				xtop vi02				; input pointer
	nop				xitop vi04				; vertex count
	nop				nop

	;; Init input - we're just converting RGBA to float here
loop:
	nop				lq vf02,2(vi02)				; rgba
	nop				nop
	nop				nop
	nop				nop
	itof0 vf02,vf02			nop
	nop				isubiu vi04,vi04,1
	nop				nop
	nop				nop
	nop				sq vf02,2(vi02)
	nop				ibne vi04,vi00,loop
	nop				iaddiu vi02,vi02,numInAttribs

.include "TLclip.vu"

	nop[e]				nop
	nop				nop
	nop				b restart
	nop				nop


.include "cliptri.vu"
.include "im3d_clipproc.vu"

.EndMPG
.EndDmaData