John Carmack
id Software | Lead Programmer | Apr 29, 2000, 01:30:39 (ET) |

Name: John Carmack
Description: Programmer
Project: Quake 3 Arena
We need more bits per color component in our 3D accelerators.

I have been pushing for a couple more bits of range for several years now,
but I now extend that to wanting full 16 bit floating point colors throughout
the graphics pipeline. A sign bit, ten bits of mantissa, and five bits of
exponent (possibly trading a bit or two between the mantissa and exponent).
Even that isn't all you could want, but it is the rational step.

It is turning out that I need a destination alpha channel for a lot of the
new rendering algorithms, so intermediate solutions like 10/12/10 RGB
formats aren't a good idea. Higher internal precision with dithering to 32
bit pixels would have some benefit, but dithered intermediate results can
easily start piling up the errors when passed over many times, as we have
seen with 5/6/5 rendering.

Eight bits of precision isn't enough even for full range static image
display. Images with a wide range usually come out fine, but restricted
range images can easily show banding on a 24-bit display. Digital television
specifies 10 bits of precision, and many printing operations are performed
with 12 bits of precision.

The situation becomes much worse when you consider the losses after multiple
operations. As a trivial case, consider having multiple lights on a wall,
with their contribution to a pixel determined by a texture lookup. A single
light will fall off towards 0 some distance away, and if it covers a large
area, it will have visible bands as the light adds one unit, two units, etc.
Each additional light from the same relative distance stacks its contribution
on top of the earlier ones, which magnifies the amount of the step between
bands: instead of going 0,1,2, it goes 0,2,4, etc. Pile a few lights up like
this and look towards the dimmer area of the falloff, and you can believe you
are back in 256-color land.

There are other more subtle issues, like the loss of potential result values
from repeated squarings of input values, and clamping issues when you sum up
multiple incident lights before modulating down by a material.

Range is even more clear cut. There are some values that have intrinsic
ranges of 0.0 to 1.0, like factors of reflection and filtering. Normalized
vectors have a range of -1.0 to 1.0. However, the most central quantity in
rendering, light, is completely unbounded. We want a LOT more than a 0.0 to
1.0 range. Q3 hacks the gamma tables to sacrifice a bit of precision to get
a 0.0 to 2.0 range, but I wanted more than that for even primitive rendering
techniques. To accurately model the full human sensable range of light
values, you would need more than even a five bit exponent.

This wasn't much of an issue even a year ago, when we were happy to just
cover the screen a couple times at a high framerate, but realtime graphics
is moving away from just "putting up wallpaper" to calculating complex
illumination equations at each pixel. It is not at all unreasonable to
consider having twenty textures contribute to the final value of a pixel.
Range and precision matter.

A few common responses to this pitch:

"64 bits per pixel??? Are you crazy???" Remember, it is exactly the same
relative step as we made from 16 bit to 32 bit, which didn't take all
that long.

Yes, it will be slower. That's ok. This is an important point: we can't
continue to usefully use vastly greater fill rate without an increase in
precision. You can always crank the resolution and multisample anti-alaising
up higher, but that starts to have diminishing returns well before you use of
the couple gigatexels of fill rate we are expected to have next year. The
cool and interesting things to do with all that fill rate involves many
passes composited into less pixels, making precision important.

"Can we just put it in the texture combiners and leave the framebuffer at 32
bits?" No. There are always going to be shade trees that overflow a given
number of texture units, and they are going to be the ones that need the
extra precision. Scales and biases between the framebuffer and the higher
precision internal calculations can get you some mileage (assuming you can
bring the blend color into your combiners, which current cards can't), but
its still not what you want. There are also passes which fundamentally
aren't part of a single surface, but still combine to the same pixels, as
with all forms of translucency, and many atmospheric effects.

"Do we need it in textures as well?" Not for most image textures, but it
still needs to be supported for textures that are used as function look
up tables.

"Do we need it in the front buffer?" Probably not. Going to a 64 bit front
buffer would probably play hell with all sorts of other parts of the system.
It is probably reasonable to stay with 32 bit front buffers with a blit from
the 64 bit back buffer performing a lookup or scale and bias operation before
dithering down to 32 bit. Dynamic light adaptation can also be done during
this copy. Dithering can work quite well as long as you are only performing
a single pass.

I used to be pitching this in an abstract "you probably should be doing this"
form, but two significant things have happened that have moved this up my hit
list to something that I am fairly positive about.

Mark Peercy of SGI has shown, quite surprisingly, that all Renderman surface
shaders can be decomposed into multi-pass graphics operations if two
extensions are provided over basic OpenGL: the existing pixel texture
extension, which allows dependent texture lookups (matrox already supports a
form of this, and most vendors will over the next year), and signed, floating
point colors through the graphics pipeline. It also makes heavy use of the
existing, but rarely optimized, copyTexSubImage2D functionality for

This is a truly striking result. In retrospect, it seems obvious that with
adds, multiplies, table lookups, and stencil tests that you can perform any
computation, but most people were working under the assumption that there
were fundamentally different limitations for "realtime" renderers vs offline
renderers. It may take hundreds or thousands of passes, but it clearly
defines an approach with no fundamental limits. This is very important.
I am looking forward to his Siggraph paper this year.

Once I set down and started writing new renderers targeted at GeForce level
performance, the precision issue has started to bite me personally. There
are quite a few times where I have gotten visible banding after a set of
passes, or have had to worry about ordering operations to avoid clamping.
There is nothing like actually dealing with problems that were mostly
theoretical before...

64 bit pixels. It is The Right Thing to do. Hardware vendors: don't you be
the company that is the last to make the transition.

Whenever I start a new graphics engine, I always spend a fair amount of time
flipping back through older graphics books. It is always interesting to see
how your changed perspective with new experience impacts your appreciation of
a given article.

I was skimming through Jim Blinn's "A Trip Down The Graphics Pipeline"
tonight, and I wound up laughing out loud twice.

From the book:

P73: I then empirically found that I had to scale by -1 in x instead of in z,
and also to scale the xa and xf values by -1. (Basically I just put in enough
minus signs after the fact to make it work.) Al Barr refers to this technique
as "making sure you have made an even number of sign errors."

P131: The only lines that generate w=0 after clipping are those that pass
through the z axis, the valley of the trough. These lines are lines that
pass exactly through the eyepoint. After which you are dead and don't care
about divide-by-zero errors.

If you laughed, you are a graphics geek.

My first recollection of a Jim Blinn article many years ago was my skimming
over it and thinking "My god, what ridiculously picky minutia." Over the last
couple years, I found myself haranguing people over some fairly picky issues,
like the LSB errors with cpu vs rasterizer face culling and screen edge
clipping with guard band bit tests. After one of those pitches, I quite
distinctly thought to myself "My god, I'm turning into Jim Blinn!" :-)

Two years ago, Id was contacted by the Startlight Foundation, an organization
that tries to grant wishes to seriously ill kids. (

There was a young man with Hodgkin's Lymphoma that, instead of wanting to go
to Disneyland or other traditional wishes, wanted to visit Id and talk with
me about programming.

It turned out that Seumas McNally was already an accomplished developer.
His family company, Longbow Digital Arts (, had
been doing quite respectably selling small games directly over the internet.
It bore a strong resemblance to the early shareware days of Apogee and Id.

We spent the evening talking about graphics programmer things -- the relative
merits of voxels and triangles, procedurally generated media, level of detail
management, API and platforms.

We talked at length about the balance between technology and design, and all
the pitfalls that lie in the way of shipping a modern product.

We also took a dash out in my ferrari, thinking "this is going to be the best
excuse a cop will ever hear if we get pulled over".

Longbow continued to be successful, and eventually the entire family was
working full time on "Treadmarks", their new 3D tank game.

Over email about finishing the technology in Treadmarks, Seumas once said
"I hope I can make it". Not "be a huge success" or "beat the competition".
Just "make it".

That is a yardstick to measure oneself by.

It is all too easy to lose your focus or give up with just the ordinary
distractions and disappointments that life brings. This wasn't ordinary.
Seumas had cancer. Whatever problems you may be dealing with in your life,
they pale before having problems drawing your next breath.

He made it.

Treadmarks started shipping a couple months ago, and was entered in the
Independent Games Festival at the Game Developer's Conference this last month.
It came away with the awards for technical excellence, game design, and the
grand prize.

I went out to dinner with the McNally family the next day, and had the
opportunity to introduce Anna to them. One of the projects at Anna's new
company, Fountainhead Entertainment (, is a
documentary covering gaming, and she had been looking forward to meeting
Seumas after hearing me tell his story a few times. The McNallys invited
her to bring a film crew up to Canada and talk with everyone whenever she

Seumas died the next week.

I am proud to have been considered an influence in Seumas' work, and I think
his story should be a good example for others. Through talent and
determination, he took something he loved and made a success out of it in
many dimensions. for more information.

This is something I have been preaching for a couple years, but I
finally got around to setting all the issues down in writing.

First, the statement:

Virtualized video card local memory is The Right Thing.

Now, the argument (and a whole bunch of tertiary information):

If you had all the texture density in the world, how much texture
memory would be needed on each frame?

For directly viewed textures, mip mapping keeps the amount of
referenced texels between one and one quarter of the drawn pixels.
When anisotropic viewing angles and upper level clamping are taken into
account, the number gets smaller. Take 1/3 as a conservative estimate.

Given a fairly aggressive six texture passes over the entire screen,
that equates to needing twice as many texels as pixels. At 1024x768
resolution, well under two million texels will be referenced, no matter
what the finest level of detail is. This is the worst case, assuming
completely unique texturing with no repeating. More commonly, less
than one million texels are actually needed.

As anyone who has tried to run certain Quake 3 levels in high quality
texture mode on an eight or sixteen meg card knows, it doesn’t work out
that way in practice. There is a fixable part and some more
fundamental parts to the fall-over-dead-with-too-many-textures problem.

The fixable part is that almost all drivers perform pure LRU (least
recently used) memory management. This works correctly as long as the
total amount of textures needed for a given frame fits in the card’s
memory after they have been loaded. As soon as you need a tiny bit
more memory than fits on the card, you fall off of a performance cliff.
If you need 14 megs of textures to render a frame, and your graphics
card has 12 megs available after its frame buffers, you wind up loading
14 megs of texture data over the bus every frame, instead of just the 2
megs that don’t fit. Having the cpu generate 14 megs of command
traffic can drop you way into the single digit frame rates on most

If an application makes reasonable effort to group rendering by
texture, and there is some degree of coherence in the order of texture
references between frames, much better performance can be gotten with a
swapping algorithm that changes its behavior instead of going into a
full thrash:

While ( memory allocation for new texture fails )
Find the least recently used texture.
If the LRU texture was not needed in the previous frame,
Free it
Free the most recently used texture that isn’t bound to an
active texture unit

Freeing the MRU texture seems counterintuitive, but what it does is
cause the driver to use the last bit of memory as a sort of scratchpad
that gets constantly overwritten when there isn’t enough space. Pure
LRU plows over all the other textures that are very likely going to be
needed at the beginning of the next frame, which will then plow over
all the textures that were loaded on top of them.

If an application uses textures in a completely random order, any given
replacement policy has the some effect…

Texture priority for swapping is a non-feature. There is NO benefit to
attempting to statically prioritize textures for swapping. Either a
texture is going to be referenced in the next frame, or it isn’t.
There aren’t any useful gradations in between. The only hint that
would be useful would be a notice that a given texture is not going to
be in the next frame, and that just doesn’t come up very often or cover
very many texels.

With the MRU-on-thrash texture swapping policy, things degrade
gracefully as the total amount of textures increase but due to several
issues, the total amount of textures calculated and swapped is far
larger than the actual amount of texels referenced to draw pixels.

The primary problem is that textures are loaded as a complete unit,
from the smallest mip map level all the way up to potentially a 2048 by
2048 top level image. Even if you are only seeing 16 pixels of it off
in the distance, the entire 12 meg stack might need to be loaded.

Packing can also cause some amount of wasted texture memory. When you
want to load a two meg texture, it is likely going to require a lot
more than just two megs of free texture memory, because a lot of it is
going to be scattered around in 8k to 64k blocks. At the pathological
limit, this can waste half your texture memory, but more reasonably it
is only going to be 10% or so, and cause a few extra texture swap outs.

On a frame at a time basis, there are often significant amounts of
texels even in referenced mip levels that are not seen. The back sides
of characters, and large textures on floors can often have less than
50% of their texels used during a frame. This is only an issue as they
are being swapped in, because they will very likely be needed within
the next few frames. The result is one big hitch instead of a steady

There are schemes that can help with these problems, but they have

Packing losses can be addressed with compaction, but that has rarely
proven to be worthwhile in the history of memory management. A 128-bit
graphics accelerator could compact and sort 10 megs of texture memory
in about 10 msec if desired.

The problems with large textures can be solved by just not using large
textures. Both packing losses, and non- referenced texels can be
reduced by chopping everything up into 64x64 or 128x128 textures. This
requires preprocessing, adds geometry, and requires messy overlap of
the textures to avoid seaming problems.

It is possible to estimate which mip levels will actually be needed and
only swap those in. An application can’t calculate exactly the mip
map levels that will be referenced by the hardware, because there are
slight variations between chips and the slope calculation would add
significant processing overhead. A conservative upper bound can be
taken by looking at the minimum normal distance of any vertex
referencing a given texture in a frame. This will overestimate the
required textures by 2x or so and still leave a big hit when the top
mip level loads for big textures, but it can allow giant cathedral
style scenes to render without swapping.

Clever programmers can always work harder to overcome obstacles, but in
this case, there is a clear hardware solution that gives better
performance than anything possible with software and just makes
everyone’s lives easier: virtualize the card’s view of its local

With page tables, address fragmentation isn’t an issue, and with the
graphics rasterizer only causing a page load when something from that
exact 4k block is needed, the mip level problems and hidden texture
problems just go away. Nothing sneaky has to be done by the
application or driver, you just manage page indexes.

The hardware requirements are not very heavy. You need translation
lookaside buffers (TLB) on the graphics chip, the ability to
automatically load the TLB from a page table set up in local memory,
and the ability to move a page from AGP or PCI into graphics memory and
update the page tables and reference counts. You don’t even need that
many TLB, because graphics access patterns don’t hop all over the place
like CPU access can. Even with only a single TLB for each texture
bilerp unit, reloads would only account for about 1/32 of the memory
access if the textures were 4k blocked. All you would really want at
the upper limit would be enough TLB for each texture unit to cover the
texels referenced on a typical rasterization scan line.

Some programmers will say “I don’t want the system to manage the
textures, I want full control!” There are a couple responses to that.
First, a page level management scheme has flexibility that you just
can’t get with a software only scheme, so it is a set of brand new
capabilities. Second, you can still just choose to treat it as a fixed
size texture buffer and manage everything yourself with updates.
Third, even if it WAS slower than the craftiest possible software
scheme (and I seriously doubt it), so much of development is about
willingly trading theoretical efficiency for quicker, more robust
development. We don’t code overlays in assembly language any more…

Some hardware designers will say something along the lines of “But
the graphics engine goes idle when you are pulling the page over from
AGP!” Sure, you are always better off to just have enough texture
memory and never swap, and this feature wouldn’t let you claim any more
megapixels or megatris, but every card winds up not having enough
memory at some point. Ignoring those real world cases isn’t helping
your customers. In any case, it goes idle a hell of a lot less than if
you were loading the entire texture over the command fifo.

3Dlabs is supposed to have some form of virtual memory management in
the permedia 3, but I am not familiar with the details (if anyone from
3dlabs wants to send me the latest register specs, I would appreciate

A mouse controlled first person shooter is fairly unique in how quickly
it can change the texture composition of a scene. A 180-degree snap
turn can conceivably bring in a completely different set of textures on
a subsequent frame. Almost all other graphics applications bring
textures in at a much steadier pace.

So, given that 180-degree snap turn to a completely different and
uniquely textured scene, what would be the worst case performance? An
AGP 2x bus is theoretically supposed to have over 500 mb/sec of
bandwidth. It doesn’t get that high in practice, but linear 4k block
reads would give it the best possible conditions, and even at 300
mb/sec, reloading the entire texture working set would only take 10

Rendering is not likely to be buffered sufficiently to overlap
appreciably with page loading, and the command transport for a complex
scene will take significant time by itself, so it shows that a worst
case scene will often not be able to be rendered in 1/60th of a second.

This is roughly the same lower bound that you get from a chip texturing
directly from AGP memory. A direct AGP texture gains the benefit of
fine-grained rendering overlap, but loses the benefit of subsequent
references being in faster memory (outside of small on-chip caches).
A direct AGP texture engine doesn’t have the higher upper bounds of a
cached texture engine, though. It’s best and worst case are similar
(generally a good thing), but the cached system can bring several times
more bandwidth to bear when it isn’t forced to swap anything in.

The important point is that the lower performance bound is almost an
order of magnitude faster than swapping in the textures as a unit by
the driver.

If you just positively couldn’t deal with the chance of that much worst
case delay, some form of mip level biasing could be made to kick in, or
you could try and do pre-touching, but I don’t think it would ever be
worth it. The worst imaginable case is acceptable, and you just won’t
hit that case very often.

Unless a truly large number of TLB are provided, the textures would
need to be blocked. The reason is that with a linear texture, a 4k
page maps to only a couple scan lines on very large textures. If you
are going with the grain you get great reuse, but if you go across it,
you wind up referencing a new page every couple texel accesses. What
is wanted is an addressing mechanism that converts a 4k page into a
square area in the texture, so the page access is roughly constant for
all orientations. There is also a benefit from having a 128 bit access
also map to a square block of pixels, which several existing cards
already do. The same interleaving-of-low-order-bits approach can just
be extended a few more bits.

Dealing with blocked texture patterns is a hassle for a driver writer,
but most graphics chips have a host blit capability that should let the
chip deal with changing a linear blit into blocked writes. Application
developers should never know about it, in any case.

There are some other interesting things that could be done if the page
tables could trigger a cpu interrupt in addition to being automatically
backed by AGP or PCI memory. Textures could be paged in directly from
disk for truly huge settings, or decompressed from jpeg blocks, or even
procedurally generated. Even the size limits of the AGP aperture could
usefully be avoided if the driver wanted to manage each page’s

Aside from all the basic swapping issue, there are a couple of other
hardware trends that push things this way.

Embedded dram should be a driving force. It is possible to put several
megs of extremely high bandwidth dram on a chip or die with a video
controller, but won’t be possible (for a while) to cram a 64 meg
geforce in. With virtualized texturing, the major pressure on memory
is drastically reduced. Even an 8mb card would be sufficient for 16
bit 1024x768 or 32 bit 800x600 gaming, no matter what the texture load.

The only thing that prevents a geometry processor based card from
turning almost any set of commands in a display list into a single
static dma buffer is the fact that textures may be swapped in and out,
causing the register programming in the buffer to be wrong. With
virtual texture addressing, a texture’s address never changes, and an
arbitrarily complex model can be described in a static dma buffer.
id Software...
Timothee Besset 2007/02/02
Christian Antkow 2005/09/08
Kenneth Scott 2005/07/18
John Carmack 2005/01/02
Fred Nilsson 2004/11/08
Todd Hollenshead 2004/11/04
Robert Duffy 2004/10/15
Tim Willits 2003/09/10
Graeme Devine 2003/07/03
Jim Dose 2002/08/19
Kevin Cloud 2002/04/16
Andy Chang 2002/03/14
Matt Hooper 2002/03/14
Paul Jaquays 2002/03/08
Seneca Menard 2002/02/15
Brandon James 2002/02/15
Eric Webb 2001/08/01
Mal Blackwell 2001/02/01
John Cash 2000/03/31
Dave Kirsch 2000/03/24
Brian Hook 1999/06/01
Katherine Anna Kang 1999/03/04
Paul Steed 1998/12/29

Also Today...


Full list

Visit Webdog today!


Square Eight - Taking over the world and you don't even know it yet! Copyright © Square Eight 1998-2024. All Rights Reserved.
The BlueTracker is provided by Webdog.
We are not responsible for the content of the .plans displayed here.