On 20 August 2011 18:46, Norbert Lataille nlataill@gmail.com wrote:
Performance gains (for parts 1/2/3) on WoW: http://norbert.lataille.free.fr/constants.html Results for 3dmark06, X3: http://norbert.lataille.free.fr/perf.html (line: Constants).
Do you have those in a reasonable format as well?
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
Am 20.08.2011 um 22:28 schrieb Henri Verbeet:
On 20 August 2011 18:46, Norbert Lataille nlataill@gmail.com wrote:
Performance gains (for parts 1/2/3) on WoW: http://norbert.lataille.free.fr/constants.html Results for 3dmark06, X3: http://norbert.lataille.free.fr/perf.html (line: Constants).
Do you have those in a reasonable format as well?
The graphs are nice to click through them, but the plain numbers would be nice too. The difference looks roughly like the GLSL-ARB shader performance difference.
My impression was that the drivers already set unused constants to -1. Sure, we declare all 256 constants, but unless indirect addressing is used the driver eliminates those that aren't read. Unfortunately I don't know which driver I used to test that, maybe it was OSX.
Hi,
I should add that my test system is a Linux Nvidia driver, and I don't have any ATI available. On this configuration, I do receive non negative location values for all input uniforms (even if they are not used). This may indeed vary depending on the target system.
Thanks,
2011/8/20 Stefan Dösinger stefandoesinger@gmx.at
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
Am 20.08.2011 um 22:28 schrieb Henri Verbeet:
On 20 August 2011 18:46, Norbert Lataille nlataill@gmail.com wrote:
Performance gains (for parts 1/2/3) on WoW: http://norbert.lataille.free.fr/constants.html Results for 3dmark06, X3: http://norbert.lataille.free.fr/perf.html (line: Constants).
Do you have those in a reasonable format as well?
The graphs are nice to click through them, but the plain numbers would be nice too. The difference looks roughly like the GLSL-ARB shader performance difference.
My impression was that the drivers already set unused constants to -1. Sure, we declare all 256 constants, but unless indirect addressing is used the driver eliminates those that aren't read. Unfortunately I don't know which driver I used to test that, maybe it was OSX. -----BEGIN PGP SIGNATURE----- Version: GnuPG/MacGPG2 v2.0.17 (Darwin)
iQIcBAEBAgAGBQJOUCwDAAoJEN0/YqbEcdMwB88P/jVrtqUjShgXz2m0QthJCV7/ OG6YxQZZMUgYj0fU2jYQttwk+CZt7guIkJsIE6WnIwnewOu0DWZRXoEF1io2G3gi Rcsu2d/gsmpxEtnVcHxwlE/W76NMYNAveM96xNj7hIqZZGumRR3J5cOFWijgey5G CEGpBv+ADQi/29FurTYRBNMmmyabwT3LlSD4tIcXFEMgcDriumNLUmahq/KHnam7 stwLnzdIIxE/gcChuZ7+U3u2j6UDaDpvPLXghxJcPuMTcutP4Zk7MySrCVQ/Rdwi ofQYijARCwQRedP0W9dAfGw2u8Ex9lCsyPkO6vUO+U1AsyZf05lJFeua+Lvwv7Y0 RXV6KcHqZFTc9G7yHC/8xNkRGUBTILOaPpJ7PiEFl/DFhaSASLP4P6czc/Ov1jS0 7EgMzflyw/gIkXztXVWq5/T+90beD29xeVDsB1fg/A2Y1EEVYsd6zMK8Xat4TuNG 6G9uz8UeQ+9slN0d29xc4IpMMCUZ962o6/1JSflAOmz7mG0yO/gqMj5Yh7gHk4vA Yxio/uGHoUJ06iDiMbgGlsTqTZLMokzZyCbPx4ml7ybJ97fpkXPYWc7U75ITXGgR TX3WXx9nhTf/hIGDa6vQt5VA8CtMNoYrTgIhV68shnhnheHcREFc++LfCBVb3bWn pBPnk768S2ctzIlLbn1h =tBJK -----END PGP SIGNATURE-----
On 20 August 2011 23:49, Stefan Dösinger stefandoesinger@gmx.at wrote:
The graphs are nice to click through them,
Well, it's Flash. Note that the patch has some issues, style being perhaps the most obvious, but there's no point even talking about that without seeing numbers first. Also, since people tend to get this wrong, it's perhaps worth pointing out that frame time is a more interesting metric than fps in most cases.
My impression was that the drivers already set unused constants to -1.
Yes, that's somewhat suspicious, although it's possible that the constants all being in a single array prevents that. I'd have to look up if the spec has anything to say about this.
On Sunday 21 August 2011 00:18:56 Henri Verbeet wrote:
Yes, that's somewhat suspicious, although it's possible that the constants all being in a single array prevents that. I'd have to look up if the spec has anything to say about this.
I think that was about the attributes, not the uniforms. In my regression hunt I ran across a game that declared a vshader attribute and didn't actually use it, so the compiler optimized it away and glBindAttribLocationARB records a warning in the shader's info log.
Sure.
Internally I am producing CSV files, which are later generated as XML files. That's not a problem to upload them in parallel to the HTML page, will look at it tomorrow.
Thanks,
2011/8/20 Henri Verbeet hverbeet@gmail.com
On 20 August 2011 18:46, Norbert Lataille nlataill@gmail.com wrote:
Performance gains (for parts 1/2/3) on WoW: http://norbert.lataille.free.fr/constants.html Results for 3dmark06, X3: http://norbert.lataille.free.fr/perf.html (line: Constants).
Do you have those in a reasonable format as well?
-
Henri Verbeet -
Norbert Lataille -
Stefan Dösinger