john-dev - cryptmd5.cu fails to build with -arch sm

Follow @Openwall on Twitter for new release announcements and other news

[<prev] [next>] [thread-next>] [day] [month] [year] [list]

Message-ID: <20120411034940.GA16046@openwall.com>
Date: Wed, 11 Apr 2012 07:49:40 +0400
From: Solar Designer <solar@...nwall.com>
To: john-dev@...ts.openwall.com
Subject: cryptmd5.cu fails to build with -arch sm_13

Lukas, magnum -

The current magnum-jumbo fails to build with the CUDA target on
cryptmd5.cu as follows:

cd cuda; nvcc -c -Xptxas -v -arch sm_13 cryptmd5.cu
ptxas info    : Compiling entry function '_Z14kernel_crypt_rP18crypt_md5_passwordPj' for 'sm_13'
ptxas info    : Used 38 registers, 64+0 bytes lmem, 30736+16 bytes smem, 21 bytes cmem[0], 48 bytes cmem[1]
ptxas error   : Entry function '_Z14kernel_crypt_rP18crypt_md5_passwordPj' uses too much shared data (0x7810 bytes + 0x10 bytes system, 0x4000 max)
make[1]: *** [cuda_cryptmd5.o] Error 255

Changing sm_13 to sm_20 makes it build.

I am not sure what fix is best (reduce the required data size, use sm_20
by default, something else), but we must have the entire tree build fine
with its default settings.  Also, even if we use sm_20 by default, it
should be fairly easy for someone with an older card to use sm_13 or
lower instead.  And I recall that sm_13 provided better speed for some
of the formats on some cards, although I guess this may/will change.

So please fix this such that the tree builds with no tweaks necessary.

Thanks,

Alexander

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.