Re: the most efficient way to send data to a GPU
- From: "jbwest" <jbwest@xxxxxxxxxxx>
- Date: Wed, 4 Jul 2007 09:03:52 -0700
"Philipp Klaus Krause" <pkk@xxxxxxx> wrote in message
news:468a9b20$0$4750$6e1ede2f@xxxxxxxxxxxxxxxxx
jbwest wrote:
Well, neither MESA nor the Intel are very strong. MESA is an "indirect"
renderer, usually, and is more or less software emulation. A $25 NVIDIA
AGP
card with proprietary drivers will probably blow the doors off of that
setup. You should be able to easily get 100 mbytes/sec. Maybe there's a
direct renderer for the Intel chip -- make sure you research it and
verify
that you have the DRI/DRM driver for the intel in the kernel & etc.
Mesa is not "more or less a software emulation". On Radeon cards it has
had hardware TCL for years. Today even indirect rendering over the
network will give you hardware acceleration with Mesa. It supports
hardware-accelerated shaders on Radeon and some Intel cards. Since Intel
is the only company that still releases documentation of their hardware
the Intel drivers are one of Mesa's strongpoints. AFAIK Mesa's drivers
for intel graphics chips are better (both feature- and speedwise) than
Intel's Windows drivers.
Philipp
I stand corrected (for Intel graphics). There's no comparison between MESA
and ATI proprietary drivers for new ATI/FIREGL cards, though. And last I
tried, the Intel/MESA combo was just too buggy & slow except for limited
applications.
Be that all as it may, I strongly suspect the OP would be better off
spending $25-$50 on a more highly advanced NV or ATI card -- A new FX 6200
AGP or similar PCIEx card is $50+/-. His Xfer rates are in the pits, and
the proof indeed is in the bottom line.
Dear OP: OK, you need 12 bits. So, you need to try to minimize data
conversions. converting 12 bits to 32 bits and then to 16 bits or whatever
isn't exactly cheap.
1) Fetch your original data in unsigned 16 bit shorts & present it as
GL_UNSIGNED_SHORT
2) Store it as GL_LUMINANCE_12 or GL_LUMINANCE_16, depending on how the 12
bits are scaled in (1), and to what extent these formats are really
supported. These formats may silently only store 8 bits anyway ... but
you'll find out. Any of the "generic" GL_LUMINANCE, etc formats is almost
guaranteed to be 8 bits only.
-jbw
.
- References:
- the most efficient way to send data to a GPU
- From: anon
- Re: the most efficient way to send data to a GPU
- From: jbwest
- Re: the most efficient way to send data to a GPU
- From: Philipp Klaus Krause
- the most efficient way to send data to a GPU
- Prev by Date: Re: Camera, Spheres, Headaches..
- Next by Date: Re: How to save a texture to a file
- Previous by thread: Re: the most efficient way to send data to a GPU
- Next by thread: Re: the most efficient way to send data to a GPU
- Index(es):
Relevant Pages
|
