Re: Inline function much faster than UDF in C?

Hi Alastair, please be aware that your function is not an aggregate by itself. Rather, you aggregate (via standard SUM()) over the result of your function. For documentation and exmaples of how to implement scalar and bulk versions of UDFs, please see http://dev.monetdb.org/hg/MonetDB/file/tip/sql/backends/monet5/UDF/ Best, Stefan ----- Original Message -----

Hi Sjoerd,

Thanks for your response. I think you are almost certainly right about the slow loop in MAL. I hope that operating in bulk also means I create an implementation using all available SIMD lanes.

Are there any examples in the source of implementing a bulk version of this type of aggregate that I could use as a guide?

Best regards,

Alastair

________________________________________ From: users-list on behalf of Sjoerd Mullender Sent: 09 December 2013 13:08 To: users-list@monetdb.org Subject: Re: Inline function much faster than UDF in C?

-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1

To get the most out of a C implementation, you need both a scalar version (one value at the time), and a bulk version (complete BATs as input). If you only have the former, the code generator will generate a slow loop in MAL calling the scalar function. And then it is very well possible that the bulk versions of the inline expression in your second attempt is a lot faster.

On 2013-12-09 13:47, Alastair McKinley wrote:

...

Hi everyone,

I have only been using MonetDB for a few weeks now and I have come across a very unexpected result when using User Defined Functions in C.

I have a query which does essentially this:

TABLE small_table ( int col1; float p; );

TABLE big_table ( int col1; int col2; float p; );

SELECT col1,1+SUM(MY_FUNCTION(big.p,small.p)) FROM small_table small inner join big_table big on big.col2 == small.col1 group by big.col1;

My function is f(x,y) = 0.7213*(x*ln(x/x+y) + y*ln(y/x+y))

I wrote an optimised version of this function using sse intrinsics in C and compiled it into MonetDB.

To my surprise, I discovered that writing this function inline in the query like so:

1+SUM(0.7213475*(big.p*LOG(big.p/(big.p+small.p))+small.p*LOG(small.p/(big.p+small.p))))

is a lot faster!

The query takes approx 2 seconds with the C version and less than 0.5 seconds with the inline version in SQL!

Can anyone explain this strange result?

Best regards,

Alastair

_______________________________________________ users-list mailing list users-list@monetdb.org https://www.monetdb.org/mailman/listinfo/users-list

- -- Sjoerd Mullender -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.15 (GNU/Linux) Comment: Using GnuPG with Thunderbird - http://www.enigmail.net/

iQCVAwUBUqXA1j7g04AjvIQpAQLGnwQAjDYkXF/z+AIcGG4/2+WZtjF5S0Z1jvLJ /sARm3pjN9e6d9U0k/KR5MmGZ3DWJNPDqYrGFqZIWp2+B46YMB72DoXW9hAudcbP P4azUiyvG0BccaHjf3gco5CA8U2DA6nGNXpX//NOr0ievVRbdMT/sxRCPNGbXBVk L0dFS9GqaR4= =sPj/ -----END PGP SIGNATURE----- _______________________________________________ users-list mailing list users-list@monetdb.org https://www.monetdb.org/mailman/listinfo/users-list _______________________________________________ users-list mailing list users-list@monetdb.org https://www.monetdb.org/mailman/listinfo/users-list

-- | Stefan.Manegold@CWI.nl | DB Architectures (DA) | | www.CWI.nl/~manegold/ | Science Park 123 (L321) | | +31 (0)20 592-4212 | 1098 XG Amsterdam (NL) |