24 Jun
2004
24 Jun
'04
12:52 p.m.
I have a denormal fix without a branch but you probably don't want to see it ;-)
It's pretty simple, just OR the bits of the exponent together which gives either
0 (denormal) or 1, typecast that to float, and then multiply the original float
by that (0.0 or 1.0). Voila, no branch, but it is messy looking ;-)
Jan
On Thu, 24 Jun 2004 11:03 , Tim Blechmann <TimBlechmann(a)gmx.net> sent:
Since the
problem is denormalised numbers, has anybody thought of
adding a small DC offset (1e-15) or alternating the
addition/subraction of a small value?
of course it is possible to add a small dc
offset ... but what if we are
working in a feedback loop? or even worse with a high pass filter?
if you alter addition/substraction and send it throuth a low pass filer?
the denormal problem is not a trivial one, but afaik a macro like this
is the only one working for sure...
cheers .... tim
--
TimBlechmann@gmx.de','','','')">TimBlechmann@gmx.de
ICQ: 96771783
After one look at this planet any visitor from outer space
would say "I want to see the manager." 
25 Jun
25 Jun
4:18 p.m.
New subject: [linux-audio-dev] swh plugins and fixing undenormalize
I have a denormal fix without a branch but you
probably don't want to
see it ;-)
It's pretty simple, just OR the bits of the exponent together which
gives either
0 (denormal) or 1, typecast that to float, and then multiply the
original float by that (0.0 or 1.0). Voila, no branch, but it is
messy looking ;-)
there is one problem ... if you multiply the denormal number with the 1
or 0, you will probably have one denormal operation on the cpu instead
of a branch ... i don't know, that's worse ...
but definitely, some benchmarks would be useful for any denormal
flushing algorithm...
cheers...
--
mailto:TimBlechmann@gmx.de ICQ: 96771783
After one look at this planet any visitor from outer space
would say "I want to see the manager."
William S. Burroughs
8:42 p.m.
New subject: [linux-audio-dev] swh plugins and fixing undenormalize
On Fri, 2004-06-25 at 13:49, Tim Blechmann wrote:
The definition of denormal is that the exponent is 0 so you will never
multiply a denormal by 1, only by 0. I'm not sure whether that would be
a denormal operation or not. It depends on the compiler.
Jan
I have a
denormal fix without a branch but you probably don't want to
see it ;-)
It's pretty simple, just OR the bits of the exponent together which
gives either
0 (denormal) or 1, typecast that to float, and then multiply the
original float by that (0.0 or 1.0). Voila, no branch, but it is
messy looking ;-)
there is one problem ... if you multiply the denormal number with the 1
or 0, you will probably have one denormal operation on the cpu instead
of a branch ... i don't know, that's worse ...
but definitely, some benchmarks would be useful for any denormal
flushing algorithm...
26 Jun
26 Jun
11:50 a.m.
New subject: [linux-audio-dev] swh plugins and fixing undenormalize
On Fri, 25 Jun 2004 17:38:24 -0500
Jan Depner <eviltwin69(a)cableone.net> wrote:
On Fri, 2004-06-25 at 13:49, Tim Blechmann wrote:
The definition of denormal is that the exponent is 0 so you will
never
multiply a denormal by 1, only by 0. I'm not sure whether that would
be a denormal operation or not. It depends on the compiler.
hm ... someone should write a test for all these algorithms ... i'm
curious, how different compilers / different algorithms actually affect
the speed of the code ...
cheers.... tim
--
mailto:TimBlechmann@gmx.de ICQ: 96771783
After one look at this planet any visitor from outer space
would say "I want to see the manager."
William S. Burroughs
I have a
denormal fix without a branch but you probably don't want
to see it ;-)
It's pretty simple, just OR the bits of the exponent together
which gives either
0 (denormal) or 1, typecast that to float, and then multiply the
original float by that (0.0 or 1.0). Voila, no branch, but it is
messy looking ;-)
there is one problem ... if you multiply the denormal number with
the 1 or 0, you will probably have one denormal operation on the cpu
instead of a branch ... i don't know, that's worse ...
but definitely, some benchmarks would be useful for any denormal
flushing algorithm...
27 Jun
27 Jun
10:58 a.m.
New subject: [linux-audio-dev] swh plugins and fixing undenormalize
[Tim Blechmann]
On Fri, 25 Jun 2004 17:38:24 -0500
Jan Depner <eviltwin69(a)cableone.net> wrote:
> On Fri, 2004-06-25 at 13:49, Tim Blechmann wrote:
> > > I have a denormal fix without a branch but you probably don't want
> > > to see it ;-)
> > > It's pretty simple, just OR the bits of the exponent together
> > > which gives either
> > > 0 (denormal) or 1, typecast that to float, and then multiply the
> > > original float by that (0.0 or 1.0). Voila, no branch, but it is
> > > messy looking ;-)
indeed sounds more like a fun proposal; nonetheless i'm wondering how
many cycles 'ORing the exponent bits together' would take. with an
8-bit exponent and no assumption about its value made, 8 binary
'shift', 7 'or' and 1 'and' statement if i'm not badly
mistaken. and
if i'm not, a branch will probably hurt less.
(on a related note, is it true that the P4 units have a weakness at
binary shift operations?)
> The definition of denormal is that the exponent
is 0 so you will
> never
> multiply a denormal by 1, only by 0. I'm not sure whether that would
> be a denormal operation or not. It depends on the compiler.
i'd expect the FPU to shortcut a multiplication by zero no matter if
the other operand is denormal or not, though i'm not too sure.
hm ... someone should write a test for all these
algorithms ... i'm
curious, how different compilers / different algorithms actually affect
the speed of the code ...
i think that the best denormal-avoiding approach depends on the
DSP algorithm in question and the extents of your demand for
computational accuracy and S/N ratio. lots of tradeoffs are possible.
tim
11:22 a.m.
New subject: [linux-audio-dev] swh plugins and fixing undenormalize
Tim Goetze wrote:
[Tim Blechmann]
Three shifts, three copys, three 'or's and an 'and':
copy = value;
value |= copy >> 4;
copy = value;
value |= copy >> 2;
copy = value;
value |= copy >> 1;
value &= 0x01;
On Fri, 25 Jun 2004 17:38:24 -0500
Jan Depner <eviltwin69(a)cableone.net> wrote:
>On Fri, 2004-06-25 at 13:49, Tim Blechmann wrote:
>
>
>>>I have a denormal fix without a branch but you probably don't want
>>>to see it ;-)
>>> It's pretty simple, just OR the bits of the exponent together
>>> which gives either
>>>0 (denormal) or 1, typecast that to float, and then multiply the
>>>original float by that (0.0 or 1.0). Voila, no branch, but it is
>>>messy looking ;-)
>>>
>>>
indeed sounds more like a fun proposal; nonetheless i'm wondering how
many cycles 'ORing the exponent bits together' would take. with an
8-bit exponent and no assumption about its value made, 8 binary
'shift', 7 'or' and 1 'and' statement if i'm not badly
mistaken. and
if i'm not, a branch will probably hurt less.
11:44 a.m.
New subject: [linux-audio-dev] swh plugins and fixing undenormalize
[Simon Jenkins]
Tim Goetze wrote:
elegant indeed (though the exponent is not in the lowest 8 bits so it
needs another shift, no?)
tim
8-bit exponent and no assumption about its value
made, 8 binary
'shift', 7 'or' and 1 'and' statement if i'm not badly
mistaken. and
if i'm not, a branch will probably hurt less.
Three shifts, three copys, three 'or's and an 'and':
copy = value;
value |= copy >> 4;
copy = value;
value |= copy >> 2;
copy = value;
value |= copy >> 1;
value &= 0x01;
12:18 p.m.
New subject: [linux-audio-dev] swh plugins and fixing undenormalize
Tim Goetze wrote:
[Simon Jenkins]
Yes it does. And another copy of value because the one you start out with
gets destroyed. And the result, when its 1, is the integer 1 not the
float 1.0
so you either need to (...etc etc).
Tim Goetze wrote:
elegant indeed (though the exponent is not in the lowest 8 bits so it
needs another shift, no?)
8-bit exponent and no assumption about its value
made, 8 binary
'shift', 7 'or' and 1 'and' statement if i'm not badly
mistaken. and
if i'm not, a branch will probably hurt less.
Three shifts, three copys, three 'or's and an 'and':
copy = value;
value |= copy >> 4;
copy = value;
value |= copy >> 2;
copy = value;
value |= copy >> 1;
value &= 0x01;
1:52 p.m.
New subject: [linux-audio-dev] swh plugins and fixing undenormalize
On Sun, 2004-06-27 at 08:30, Tim Goetze wrote:
[Simon Jenkins]
You'd probably need to AND and shift by 23 prior to starting this.
This is much nicer than my first cut. I was using brute force on the
ORs. Also, I believe you're right about the FPU short-cutting a
multiply by zero.
Jan
Tim Goetze wrote:
elegant indeed (though the exponent is not in the lowest 8 bits so it
needs another shift, no?)
8-bit exponent and no assumption about its value
made, 8 binary
'shift', 7 'or' and 1 'and' statement if i'm not badly
mistaken. and
if i'm not, a branch will probably hurt less.
Three shifts, three copys, three 'or's and an 'and':
copy = value;
value |= copy >> 4;
copy = value;
value |= copy >> 2;
copy = value;
value |= copy >> 1;
value &= 0x01;
1:48 p.m.
New subject: [linux-audio-dev] swh plugins and fixing undenormalize
On Sun, 2004-06-27 at 07:43, Tim Goetze wrote:
[Tim Blechmann]
That's why I said you probably don't want to see it. I did it mostly
for fun. When I finished coding it I was laughing too hard to test it.
Jan
On Fri, 25 Jun 2004 17:38:24 -0500
Jan Depner <eviltwin69(a)cableone.net> wrote:
> On Fri, 2004-06-25 at 13:49, Tim Blechmann wrote:
> > > I have a denormal fix without a branch but you probably don't want
> > > to see it ;-)
> > > It's pretty simple, just OR the bits of the exponent together
> > > which gives either
> > > 0 (denormal) or 1, typecast that to float, and then multiply the
> > > original float by that (0.0 or 1.0). Voila, no branch, but it is
> > > messy looking ;-)
indeed sounds more like a fun proposal; nonetheless i'm wondering how
many cycles 'ORing the exponent bits together' would take. with an
8-bit exponent and no assumption about its value made, 8 binary
'shift', 7 'or' and 1 'and' statement if i'm not badly
mistaken. and
if i'm not, a branch will probably hurt less.
7545
days inactive
7548
days old
linux-audio-dev@lists.linuxaudio.org
9 comments
5 participants
participants (5)
-
eviltwin69＠cableone.net -
Jan Depner
-
Simon Jenkins -
Tim Blechmann -
Tim Goetze