16x tri binning setup

gistfile1.txt

    // Constants
    static const OM_U32 cEdgeFlags[5] =
    {
        RAST_SCISSORS_FLAG_EDGE0,
        RAST_SCISSORS_FLAG_EDGE1,
        RAST_SCISSORS_FLAG_EDGE2,
        RAST_SCISSORS_FLAG_EDGE3,
        RAST_SCISSORS_FLAG_TRIANGLE_EXTENDS_OUTSIDE_SCISSORS
    };

    __mmask     mRastScissor        = 0;
    __mmask     mAnyScissor         = 0;

    // 1D scissor logic for x
    __mmask     mScissorXMin        = _mm512_mask_cmplt_pi(mPrimPass, vBBoxXMin, vScissorXMin);
    __mmask     mScissorXMax        = _mm512_mask_cmplt_pi(mPrimPass, vScissorXMax, vBBoxXMax);
    _M512I XFullIntersectionMin     = vIndentedSnap4BBoxXMin;
    _M512I XFullIntersectionMax     = vIndentedSnap4BBoxXMax; 
    if (!_mm512_kortestz(mScissorXMin, mScissorXMax))
    {
        // At least some tris extend outside of scissor rect
        __mmask mScissored          = _mm512_kor(mScissorXMin, mScissorXMax);

        // Disable prim if it gets rejected by either original or indented bbox edges
        // NOTE original code only did 
        _M512I vMaxXMin             = _mm512_max_pi(vBBoxXMin, vIndentedSnap4BBoxXMin);
        _M512I vMinXMax             = _mm512_min_pi(vBBoxXMax, vIndentedSnap4BBoxXMax);
        mPrimPass                   = _mm512_mask_cmplt_pi(mPrimPass, vScissorXMin, vMinXMax);
        mPrimPass                   = _mm512_mask_cmplt_pi(mPrimPass, vMaxXMin, vScissorXMax);
        mAnyScissor                 = _mm512_kor(mAnyScissor, mScissored);

        // Intersection of scissor and triangle bbox
        XFullIntersectionMin        = _mm512_mask_max_pi(XFullIntersectionMin, mScissorXMin, vBBoxXMin, vScissorXMin);
        XFullIntersectionMax        = _mm512_mask_min_pi(XFullIntersectionMax, mScissorXMax, vBBoxXMax, vScissorXMax);

        // Do we need to rasterize scissor edges? (Can skip that if they happen to fall on tile boundaries)
        __mmask mRastScissorEdge0   = _mm512_mask_test_pi(mScissorXMin, vScissorXMin, _mm512_upconv_1to16(&pfers->Fix8TileXMask));
        __mmask mRastScissorEdge1   = _mm512_mask_test_pi(mScissorXMax, vScissorXMax, _mm512_upconv_1to16(&pfers->Fix8TileXMask));

        // Set scissor flag on any scissored tri we didn't reject yet
        mScissored                  = _mm512_kand(mScissored, mPrimPass);
        vFlags                      = _mm512_mask_or_pi(vFlags, mScissored, vFlags, _mm512_upconv_1to16(&cEdgeFlags[4]));

        // Also set edge scissor flags 
        vFlags                      = _mm512_mask_or_pi(vFlags, mRastScissorEdge0, vFlags, _mm512_upconv_1to16(&cEdgeFlags[0]));
        vFlags                      = _mm512_mask_or_pi(vFlags, mRastScissorEdge1, vFlags, _mm512_upconv_1to16(&cEdgeFlags[1]));

        // Keep track of whether any tris need to rasterize scissor edges
        mRastScissor                = _mm512_kor(mRastScissor, mRastScissorEdge0);
        mRastScissor                = _mm512_kor(mRastScissor, mRastScissorEdge1);
    }

    // 1D scissor logic for y
    __mmask     mScissorYMin        = _mm512_mask_cmplt_pi(mPrimPass, vBBoxYMin, vScissorYMin);
    __mmask     mScissorYMax        = _mm512_mask_cmplt_pi(mPrimPass, vScissorYMax, vBBoxYMax);
    _M512I YFullIntersectionMin     = vIndentedSnap4BBoxYMin;
    _M512I YFullIntersectionMax     = vIndentedSnap4BBoxYMax;
    if (!_mm512_kortestz(mScissorYMin, mScissorYMax))
    {
        // At least some tris extend outside of scissor rect
        __mmask mScissored          = _mm512_kor(mScissorYMin, mScissorYMax);

        // Disable prim if it gets rejected by either original or indented bbox edges
        _M512I vMaxYMin             = _mm512_max_pi(vBBoxYMin, vIndentedSnap4BBoxYMin);
        _M512I vMinYMax             = _mm512_min_pi(vBBoxYMax, vIndentedSnap4BBoxYMax);
        mPrimPass                   = _mm512_mask_cmplt_pi(mPrimPass, vScissorYMin, vMinYMax);
        mPrimPass                   = _mm512_mask_cmplt_pi(mPrimPass, vMaxYMin, vScissorYMax);
        mAnyScissor                 = _mm512_kor(mAnyScissor, mScissored);

        // Intersection of scissor and triangle bbox
        YFullIntersectionMin        = _mm512_mask_max_pi(YFullIntersectionMin, mScissorYMin, vBBoxYMin, vScissorYMin);
        YFullIntersectionMax        = _mm512_mask_min_pi(YFullIntersectionMax, mScissorYMax, vBBoxYMax, vScissorYMax);

        // Do we need to rasterize scissor edges? (Can skip that if they happen to fall on tile boundaries)
        __mmask mRastScissorEdge2   = _mm512_mask_test_pi(mScissorYMin, vScissorYMin, _mm512_upconv_1to16(&pfers->Fix8TileYMask));
        __mmask mRastScissorEdge3   = _mm512_mask_test_pi(mScissorYMax, vScissorYMax, _mm512_upconv_1to16(&pfers->Fix8TileYMask));

        // Set scissor flag on any scissored tri we didn't reject yet
        mScissored                  = _mm512_kand(mScissored, mPrimPass);
        vFlags                      = _mm512_mask_or_pi(vFlags, mScissored, vFlags, _mm512_upconv_1to16(&cEdgeFlags[4]));

        // Also set edge scissor flags 
        vFlags                      = _mm512_mask_or_pi(vFlags, mRastScissorEdge2, vFlags, _mm512_upconv_1to16(&cEdgeFlags[2]));
        vFlags                      = _mm512_mask_or_pi(vFlags, mRastScissorEdge3, vFlags, _mm512_upconv_1to16(&cEdgeFlags[3]));

        // Keep track of whether any tris need to rasterize scissor edges
        mRastScissor                = _mm512_kor(mRastScissor, mRastScissorEdge2);
        mRastScissor                = _mm512_kor(mRastScissor, mRastScissorEdge3);
    }

    // We don't care about rasterized scissor edges for prims that ended up being rejected
    mRastScissor = _mm512_kand(mRastScissor, mPrimPass);

    // If we actually rast scissor edges, need to use original not indented triangle bbox:
    // Rast can drop triangle edges it doesn't think necessary because of the scissors.
    // This can result in both scissor edge and triangle edge being dropped if the indented
    // bbox is used, which can result in a blowout in the general rasterizer.
    XFullIntersectionMin            = _mm512_mask_max_pi(XFullIntersectionMin, mRastScissor, vBBoxXMin, vScissorXMin);
    XFullIntersectionMax            = _mm512_mask_min_pi(XFullIntersectionMax, mRastScissor, vBBoxXMax, vScissorXMax);
    YFullIntersectionMin            = _mm512_mask_max_pi(YFullIntersectionMin, mRastScissor, vBBoxYMin, vScissorYMin);
    YFullIntersectionMax            = _mm512_mask_min_pi(YFullIntersectionMax, mRastScissor, vBBoxYMax, vScissorYMax);
    
    // Indented snap4 16x16 candidate test (this case can't deal with rasterized scissor edges)
    __mmask     mIndentedCandidate  = _mm512_kandnr(mPrimPass, mRastScissor);
    _M512I      vIndentedBBoxW      = _mm512_sub_pi(vIndentedSnap4BBoxXMax, vIndentedSnap4BBoxXMin);
    _M512I      vIndentedBBoxH      = _mm512_sub_pi(vIndentedSnap4BBoxYMax, vIndentedSnap4BBoxYMin);
    _M512I      vIndentedBBoxSize   = _mm512_max_pi(vIndentedBBoxW, vIndentedBBoxH);
    __mmask     mIndented16x16      = _mm512_mask_cmple_pi(mIndentedCandidate, vIndentedBBoxSize, _mm512_upconv_1to16(&pfers->c16Pixels));

    // Tile bounds for the tris
    // Because of top-left fill rule, min is inclusive whereas max is exclusive, which is
    // why we subtract 1 from the latter.
    _M512I      vFullXIsectMaxSub1  = _mm512_sub_pi(XFullIntersectionMax, _mm512_upconv_1to16(&pfers->cOne));
    _M512I      vFullYIsectMaxSub1  = _mm512_sub_pi(YFullIntersectionMax, _mm512_upconv_1to16(&pfers->cOne));

    _M512I      vTileX              = _mm512_srl_pi(XFullIntersectionMin, _mm512_upconv_1to16(&pfers->TileXShiftPlusFix8Shift));
    _M512I      vTileY              = _mm512_srl_pi(YFullIntersectionMin, _mm512_upconv_1to16(&pfers->TileYShiftPlusFix8Shift));
    _M512I      vTileMaxX           = _mm512_srl_pi(vFullXIsectMaxSub1, _mm512_upconv_1to16(&pfers->TileXShiftPlusFix8Shift));
    _M512I      vTileMaxY           = _mm512_srl_pi(vFullYIsectMaxSub1, _mm512_upconv_1to16(&pfers->TileYShiftPlusFix8Shift));
    _M512I      vTileSizeXMinus1    = _mm512_sub_pi(vTileMaxX, vTileX);
    _M512I      vTileSizeYMinus1    = _mm512_sub_pi(vTileMaxY, vTileY);
    _M512I      vTileSizesORed      = _mm512_or_pi(vTileSizeXMinus1, vTileSizeYMinus1);
    __mmask     mMoreThanOneTile    = _mm512_mask_test_pi(mPrimPass, vTileSizesORed, vTileSizesORed);