Holger, thank you for your excellent (as usual) comments.
Indeed, to convert vector:
// d0.re, d0.im, d1.re, d1.im, d2.re, d2.im, d3.re, d3.im
into vector:
// d0.im, -d0.re, d1.im, -d1.re, d2.im, -d2.re, d3.im, -d3.re
i need only to instructions, one vec_sub and one vec_perm.
Plus this instructions run on separate units!
// permute mask and the result of permute
// 0203 1011 0607 1415 0A0B 1819 0E0F 1C1D
// |----+----+----+----+----+----+----+----|
// 0im -0re 1im -1re 2im -2re 3im -3re
Thank you again for your nice ideas.
--
Sincerely,
Rustam Muginov
On Nov 25, 2005, at 6:19 PM, Holger Bettag wrote:
On Fri, 25 Nov 2005, Rustam Muginov wrote:
[...]
1. Use vec_msum onto Filter and Data vector (calculation one). This
allows me to to make 8 multiplications and 4 additions with one
instruction, and i am getting 32bit result for free.
2. Swap real and imaginary parts in one vector with vec_rl.
3. Invert signs on the odd element of the swapped vector (now to
instuctions, vec_sub and vec_sel).
4. Use vec_msum again (calculation two). Again, 8 muls and 4 adds with
one instruction.
Consider replacing step 2 and the vec_sel of step 3 with equivalent
permutes. If the loop is unrolled to exhibit enough instruction level
parallelism, the permutes can execute in parallel with the
computational
instructions. This could increase throughput notably. However,
especially
on G5, latencies will be increased by this. So unrolling is a necessary
requirement for this to be beneficial.
Holger
P.S.: The general rule here is to try and balance the work between
permute
unit and the computational units. Ideally, the machine will
approach
peak throughput of one permute plus one other vector instruction
per
cycle.
_______________________________________________
Do not post admin requests to the list. They will be ignored.
PerfOptimization-dev mailing list
(email@hidden)
Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/perfoptimization-dev/
email@hidden
