profiling

[NaN]

The profiling counters are inconspicuous. GetCollapsedDrawList has a CPI of two. It is the amount of retired instructions that bothers me. It looks like it is doing a lot of work!?

Code:

CS:EIP       Symbol + Offset                                     64-bit         CPU clocks     Ret inst     
0x18b6e0     SCENENODE::GetCollapsedDrawList                                10000          5542         
0x122080     GRAPHICS_SDLGL::RENDER_INPUT_SCENE::FrustumCull                3016           1799         

2 functions, 215 instructions, Total: 20357 samples, 51.96% of samples in the module, 13.95% of total session samples

The bad code segment:

Code:

    CPU clocks     Ret inst     
for (map <DRAWABLE_FILTER *, vector <SCENEDRAW> >::iterator mi = drawlist_output_map.begin(); mi != drawlist_output_map.end(); ++mi) 241            168          
{                                                                                                                                    0              0            
    for (list <DRAWABLE>::const_iterator i = drawlist.begin(); i != drawlist.end(); ++i)                                             2343           846          
    {                                                                                                                                0              0            
        if (mi->first->Matches(*i))                                                                                                  4608           3814         
        {                                                                                                                            0              0            
            if (i->GetDrawEnable())                                                                                                  1622           433          
            {                                                                                                                        0              0            
                mi->second.push_back(SCENEDRAW(*i,this_transform));                                                                  30             8            
            }                                                                                                                        0              0            
        }                                                                                                                            0              0            
    }                                                                                                                                0              0            
}                                                                                                                                    0              0

The Matches() function:

Code:

bool DRAWABLE_FILTER::Matches(const DRAWABLE & drawable) const
{
    return (drawable.filterspeedup.filtervalue & filtermask) == (filtervalue & filtermask);
}

I added a counter to see how often the sort in GetCollapsedDrawList is executed. It is run if drawlist or childlist of the scenenode are not empty:
245 times/frame in GUI
251 times/frame in car selection
253 times/frame in game

Edit:
It won't solve the problem with the sort loop, but maybe it's worth to move GUI into an own scenenode like the track. I don't understand why it is faster on linux. Will install a distro asap.

Edit2:
After rearranging the loop body:

Code:

CPU clocks     Ret inst
for (map <DRAWABLE_FILTER *, vector <SCENEDRAW> >::iterator mi = drawlist_output_map.begin(); mi != drawlist_output_map.end(); ++mi) 252            240
{                                                                                                                                    0              0
    for (list <DRAWABLE>::const_iterator i = drawlist.begin(); i != drawlist.end(); ++i)                                             2879           2586
    {                                                                                                                                0              0
        if (i->GetDrawEnable())                                                                                                      5389           1661
        {                                                                                                                            0              0
            if (mi->first->Matches(*i))                                                                                              68             75
            {                                                                                                                        0              0
                mi->second.push_back(SCENEDRAW(*i, this_transform));                                                                 25             15
            }                                                                                                                        0              0
        }                                                                                                                            0              0
    }                                                                                                                                0              0
}

It looks like there is a serious penalty accessing the drawlist elemets.