TLDS RGBA instruction broken

[andravin]

The TLDS RGBA instruction has been broken since March 4 checkin b744f88

This breaks maxDNN.

The problem is that the following lines were removed from MaxasGrammar.pm:

TLDS: chnls
0x0010000000000000 RGBA

[scott-gray]

Sorry about that. I re-factored your change a bit but missed that part.
No matter how small the change it's always worth testing.. guess I failed
on that one.

Should be all set now.

-Scott

On Wed, Mar 25, 2015 at 10:33 AM, Andrew Lavin notifications@github.com
wrote:

The TLDS RGBA instruction has been broken since March 4 checkin b744f88
b744f88

This breaks maxDNN.

The problem is that the following lines were removed from MaxasGrammar.pm:

TLDS: chnls
0x0010000000000000 RGBA

—
Reply to this email directly or view it on GitHub
#3.

[andravin]

Fix works for me, thanks.

I think if there was a coverage test in Maxas that checked the op codes of a bunch of instructions, it would have caught this regression error. Does that sound like a good idea? I seem to recall you used something like that to explore new instructions anyway.

[scott-gray]

I did have a bunch of unit tests working, but haven't been good about
keeping that up to date. I just have way too much on my plate to be able
to fill in all the little details for all my projects that I should be. If
that's something you want to contribute it would be a big help. There is
now a team at Alibaba working on a Kepler version and I'm sure they could
make use of the same test framework. Basically we just need a bunch of ptx
code that exercises all the instructions with all the potential flags
(all.ptx was the start of this). But ptx is kind of an abomination of a
program language to be working in and hence pretty tedious. There's also
the challenge of making sure your test code doesn't get optimized away by
ptxas.

-Scott

On Wed, Mar 25, 2015 at 10:52 AM, Andrew Lavin notifications@github.com
wrote:

Fix works for me, thanks.

I think if there was a coverage test in Maxas that checked the op codes of
a bunch of instructions, it would have caught this regression error. Does
that sound like a good idea? I seem to recall you used something like that
to explore new instructions anyway.

—
Reply to this email directly or view it on GitHub
#3 (comment).

[andravin]

Can't you just disassemble and reassemble a list of cubins? Like start with your lib and maxDNN.

Sent from my iPad

On Mar 25, 2015, at 11:31 AM, Scott Gray notifications@github.com wrote:

I did have a bunch of unit tests working, but haven't been good about
keeping that up to date. I just have way too much on my plate to be able
to fill in all the little details for all my projects that I should be. If
that's something you want to contribute it would be a big help. There is
now a team at Alibaba working on a Kepler version and I'm sure they could
make use of the same test framework. Basically we just need a bunch of ptx
code that exercises all the instructions with all the potential flags
(all.ptx was the start of this). But ptx is kind of an abomination of a
program language to be working in and hence pretty tedious. There's also
the challenge of making sure your test code doesn't get optimized away by
ptxas.

-Scott

On Wed, Mar 25, 2015 at 10:52 AM, Andrew Lavin notifications@github.com
wrote:

Fix works for me, thanks.

I think if there was a coverage test in Maxas that checked the op codes of
a bunch of instructions, it would have caught this regression error. Does
that sound like a good idea? I seem to recall you used something like that
to explore new instructions anyway.

—
Reply to this email directly or view it on GitHub
#3 (comment).

—
Reply to this email directly or view it on GitHub.