Super-Secret Project

Another terribly-short-and-limited blog post to follow, so I don't blame you if you skip this one. Work, parents' evening, and the super-secret-project have been taking up most of my time. In no particular order, I have managed to:

• Sort the 3D printer bed out - glass cut and bored successfully (or nearly finished, need to tidy and finish the holes)
• Nearly-finish the lighting optimisations - stencils are now working, and now just need to fiddle around with how the information about lighting is stored (radius, colour, position, shadow-casting)
• Finish most of the programming on the super-secret project for the time being. Hint: it involves multiple virtual servers, a simple database, plenty of permissions, various types of buckets, and voxels. 
• Buy far too many games - RAGE, Amnesia, L4D2, FEAR series, and a couple of indie titles. I might do a mini-review when I get around to playing them.

I leave you with an incredible video - worth watching in HD.

Charlie

Bugfixes

This weekend, I got to see a legitimate Robot Wars battle at the NEC - two flipper-armed robots throwing each other (and themselves) around. I also saw and spoke to a couple of engineers on the Bloodhound SSC project and saw the vehicle itself.

On the game, instead of coding anything new, I went through some of the very worst bits of my code and carefully removed a few of the worst-offending bugs.

• The first one was that resizing the window would not resize the stencil buffer, resulting in some fairly unusual artifacts appearing. It took me a fair while to notice that I was actually allocating the buffer twice, and corrected it. Now I have stencils, which makes culling pixels for lighting very easy now.
• The second one was a silly mistake in how I handled texture path names. I'd been doing it in a silly way (hard-coded slashes, which is death for cross-platform compatibility) - I went through and replaced most of them with the appropriate platform-independent solution. 
• The final one was a complete pain to find and kill - I found that when changing camera angle with bright lights in the scene, the lights would flicker in a fairly irritating manner. I checked and double-checked my entire lighting code several times, as well as the matrix transformation, before realising that everything was (surprisingly) perfect. The actual issue was that my normal buffer was at too low a precision to make the lights adjust properly over a wide range of viewing angles. Unfortunately, a side effect of increasing the precision is that the lights now take longer to process - now, 800MP would seem to take (very approximately) half a second to run, giving me 16MP to use for lights, not 32MP (which is fair, given that I doubled the bits). It remains to be seen whether changing some of the other calculations (inverse of the matrices, reprojection, etc), lowering precision (8-bit floating not 16-bit floating buffers) and/or constraining the areas will be enough - though there's no reason it shouldn't. 

I also managed to fix George's PS3 for the second time and started work on a super-secret project, which is going to be revealed fairly soon (on or before next week).

In Motion

The printer is finally moving of its own accord - I discovered a work-around for the firmware issue. Connecting the Arduino via USB, connecting in RepStrapper, and then finally applying power to the shield allows the board to connect. The thermistor readings work fine and I'm pretty sure the heater MOSFETS do too, but the stepper boards get too warm to work effectively - I should have some thermal epoxy and old aluminium heatsinks spare to sort that out, though I dislike making any permanent changes to hardware (and that glue is just about as permanent as it gets). At the same time I should also be able to heatsink my 1W laser driver too, and get some pretty pictures of the beam at full power, without it cutting out as the thermal limiter kicks in. The only other option is to buy a new and higher-power circuit instead (torches have higher-power boards but with much less tight regulation, which should be OK for a tough laser diode, though it may shorten the life-span somewhat).

I've also played with the lighting shader code, and worked out that changing the data types (the "quick fix" solution to the lighting speed) is simply not enough to bring the time down to reasonable lengths - it saves a about 2 ms for 800 million pixels - not nearly enough. The only reasonable solution now is to apply the lighting calculation to less pixels - using an estimate of 0.75ms/4MP, that means that lighting can "afford" to light about 32MP and still leave enough processing time for pretty effects like depth of field, ambient light, SSAO and another lighting or material pass for skin, particles, and other effects. To restrict 100 lights to 32MP, the average visible light radius can be around 300 on-screen pixels, which is pretty reasonable - in a cluttered scene many of the lit pixels will be occluded anyway, so drawing the light with a depth comparison might save even more processing time. Shadows are also relatively cheap with this method but will still need careful planning. All of these estimates are based on a fast card at high resolution though, so others' mileage may vary.

Charlie

Thermistor Trouble

In the last week, I've added multi-lights, tidied up my current lighting method and planned/written the beginnings of a skeleton-loader - the Doom 3 format is particularly nice for my purposes. Lights are currently too inefficient - 100 simultaneous non-shadowcasting lights run at about 7fps at 2560x1600. Admittedly, I've taken some horrific shortcuts in the pursuit of simplicity (right now, for every pixel on the screen, for every light in view, it generates the inverse of a 4x4 matrix - not exactly a speedy process).. When I fix that particular source of inefficiency, I'll re-time everything, add shadows, and see if it still needs more work. After that is skeletons and game logic itself.

As for the 3D printer, I (badly) drilled the holes through the glass bed - it's still usable, but has some irritating cracks. The firmware is unfortunately proving to be a bit of a pain - I haven't managed to get everything working reliably. I remember having a similar issue previously, which I think ended up tracing back to the thermistor board, of all places. The board currently sends back a "start" signal, but doesn't do anything useful afterwards, and drops the connection. With a bit of luck that issue shouldn't take much more time to trace down and I'll be all but ready to print!

 Charlie