#9 Tuning transform failure for a Tld build.
https://sourceforge.net/forum/message.php?msg_id=4372894
By: vkarak
Hi!
I have downloaded oski-1.0.1b and I am actually trying the get around with its
tuning capabilities. I have configured it on an Intel Pentium 4 Xeon, using
gcc-4.1 with a rather aggressive opt '-O3 -funroll-all-loops'. I have only enabled
the long-double configuration. The normal MatMult kernel runs at a pretty fast
average of 350-400MFLOPS (thanks to unrolling, which eliminates all that overhead
needed to reach the actual kernel). I have tried to aggressively apply tuning
using oski_TuneMat to a diverse set of matrices, but with no result. I tried
also to let OSKI implicitly optimize the matrix structure by calling oski_TuneMat
in every MatMult operation (i am actually doing about a hundred of them). So
I decided to try the ApplyMatTransforms out, using an explicit simple Lua program,
just a 'return BCSR(InputMat, 2, 2)' snippet, but again I get an error. Turning
on OSKI debugging info an error 'One of the conversion arguments is invalid'
appears, but I cannot figure out why. Am I missing something with the BCSR Lua
representation? I also get the same error, when I instruct OSKI to use 1x1 blocks.
The entire debug output of OSKI follows. I would much appreciate your help,
thank you in advance!
Bill
PS: Btw, trying to get long-double configuration properly working, I think I
came across a typo bug in <oski/oski_Tld.h>. `DEF_IND_TYPE' should have been
defined to `2' (long) and not to `1' (int).
Run output with OSKI_DEBUG_LEVEL set.
Logged In: YES
user_id=1291352
Originator: YES
Additional information about the bug, extracted from: https://sourceforge.net/forum/message.php?msg_id=4373306
The user cross-checked xforms.c
and the debug output and noticed that the reason of failure is that id_CSR
!= A_source->mattype_id. I don't know what exactly id_CSR represents internally,
but I hope this will help. Anyway, I am now trying an int-double configuration
to check if it 'll work for me.