[Pixman] [PATCH 0/10] Floating point pipeline again
sandmann at cs.au.dk
Wed Sep 26 13:48:28 PDT 2012
Behdad Esfahbod <behdad at behdad.org> writes:
> On 09/26/2012 04:43 PM, Søren Sandmann wrote:
>> Here is a new version of the floating point patches. The main changes
>> since last time are:
>> (a) It is now explicit that the only guarantee wrt. floating point
>> exceptions is that divide-by-zero is not generated. So if an
>> application enables for example FP_INVALID, it may crash. It can
>> safely enablE FP_DIVBYZERO though.
>> (b) pixman_expand_to_float() is faster. The first version was written
>> in a way that caused gcc to generate divisions in the inner
>> loop. The new version avoids that.
>> Below are the results of some benchmark runs. From top to bottom:
>> - Regular, unmodified pixman
>> - With all compositing going through the current 16 bpc wide path
>> - With all compositign going through the new floating point wide path
> What are the columns?
[test number], [cairo backend], [test name], [min time], [median time], [std variance]
The 2/2 column I believe indicates that the test was run twice and two
of those runs were used in the data gathering (ie., not considered
More information about the Pixman