Hi all, I searched the web and didn't find anything useful, but is there some trick to clip a signed int to, say, +-15 or +-16, maybe using some bit pattern processing? Best regards, Andre
simple clipping to +-(2^n) or +-(2^n-1)?
Started by ●January 14, 2015
Reply by ●January 14, 20152015-01-14
On Wed, 14 Jan 2015 15:33:21 +0100, Andre <lodwig@pathme.de> wrote:>Hi all, > >I searched the web and didn't find anything useful, >but is there some trick to clip a signed int to, say, >+-15 or +-16, maybe using some bit pattern processing? > >Best regards, > >Andreif (x(n) > UpperLimit) x(n) = UpperLimit; if (x(n) < LowerLimit) x(n) = LowerLimit; Eric Jacobsen Anchor Hill Communications http://www.anchorhill.com
Reply by ●January 14, 20152015-01-14
On 14.01.15 15.33, Andre wrote:> I searched the web and didn't find anything useful, > but is there some trick to clip a signed int to, say, > +-15 or +-16, maybe using some bit pattern processing?+-15 or +-16 is an odd range. You won't get this working by bit operations. And even other saturation logic is not that trivial. That's why GPUs have hard coded hardware for this purpose. The fastest way without dedicated hardware is compare followed by conditional memory access if your CPU supports it. There you might succeed with 4 machine instructions. Without conditional memory access (e.g. x86) you won't come around branches with probably not that good branch prediction. Of course, you can use bit logic, but I am in doubt that there is any benefit, besides obfuscation of course. short x, y; x = y = i; x |= x << 6; y &= y << 6; x |= x << 3; y &= y << 3; x |= x << 1; y &= y << 1; x |= x << 1; y &= y << 1; x &= ~y; y = i; x >>= 15; y >>= 15; printf(" - %x, %x, %x; %x - ", x, y, x & ~y & 15); i |= x & ((y & -16) | (~y & 15)); i &= ~x | ((y | 15) & (~y | -16)); Now i is restricted to -16..15. This is for 16 bit input. But it works similar for 32 bit input. You only need another shift operation above. Marcel
Reply by ●January 14, 20152015-01-14
On Wed, 14 Jan 2015 15:33:21 +0100, Andre wrote:> Hi all, > > I searched the web and didn't find anything useful, but is there some > trick to clip a signed int to, say, +-15 or +-16, maybe using some bit > pattern processing?In what hardware? What are you really trying to do? There are some handy tricks you can do if n is equal to one less than your word width, the number is a result of an add or subtract (so there's an overflow bit available) and you're willing to program in assembly language. There are some handy tricks you can do if n is equal to one less than your word width and you're using an integer DSP with an overflow register (they come with different names, but yea -- that). There are some handy tricks you can do if you're doing FPGA or logic design and the number is the result of an add (or a series thereof). All these sets of handy tricks are different. -- Tim Wescott Wescott Design Services http://www.wescottdesign.com
Reply by ●January 14, 20152015-01-14
For FPGAs you can either check the whole value or better just the few MSB
bits to be ddiscarded as follows:
-- checking one MSB bit and clipping, assuming data1 is 17 bits signed
if data1(16:15) = “00” or data1(16:15) = “11” then
data <= data1(15 downto 0);
elsif data1(16) = '1' then
data <= -32767;
else
data <= +32767;
end if;
Kaz
_____________________________
Posted through www.DSPRelated.com
Reply by ●January 14, 20152015-01-14
Andre <lodwig@pathme.de> wrote:> I searched the web and didn't find anything useful, > but is there some trick to clip a signed int to, say, > +-15 or +-16, maybe using some bit pattern processing?If you want saturation arithmetic, so clip and the largest and smallest values, sense overflow/underflow (assuming add or subtract) and process accordingly. Some years ago, I was trying to do this in XC4000 series FPGAs, which have a neat trick. I don't think this works in later ones. The XC4000 carry logic is separate from the logic generating the sum (or difference). Also, there is one extra entry in the LUTs that are needed for adding. I could test for overflow or underflow (not both) and feed the resulting signal into the otherwise unused LUT input to generate the largest or smallest value. (In most cases I knew which way it would go.) In the case where it could either overflow or underflow, I would feed one signal to the following pipeline stage, which also had an extra input. That trick went away with the XC4000, and I don't know any quite as good for later series. For software, again detect the overflow and use conditional load (faster than unpredicted branch) to load the right value. Again, you have to know which way it goes, or use two such tests. For other cases, you can use bit masks to detect values. That is sometimes faster than compare instructions. -- glen
Reply by ●January 15, 20152015-01-15
Dear all, thanks for all the responses, I was just a bit upset that comparing to upper and lower limit is my best option. Hardware is a blackfin DSP, I noticed that a sharc has a native clip instruction... best regards, Andre On 15.01.2015 00:43, kaz wrote:> For FPGAs you can either check the whole value or better just the few MSB > bits to be ddiscarded as follows: > > -- checking one MSB bit and clipping, assuming data1 is 17 bits signed > if data1(16:15) = “00” or data1(16:15) = “11” then > data <= data1(15 downto 0); > elsif data1(16) = '1' then > data <= -32767; > else > data <= +32767; > end if; > > > Kaz > > _____________________________ > Posted through www.DSPRelated.com >
Reply by ●January 15, 20152015-01-15
On 15.01.2015 00:43, kaz wrote:> For FPGAs you can either check the whole value or better just the few MSB > bits to be ddiscarded as follows: > > -- checking one MSB bit and clipping, assuming data1 is 17 bits signed > if data1(16:15) = “00” or data1(16:15) = “11” then > data <= data1(15 downto 0); > elsif data1(16) = '1' then > data <= -32767; > else > data <= +32767; > end if; > > > Kaz > > _____________________________ > Posted through www.DSPRelated.com >Dear all, thanks for all the responses, I was just a bit upset that comparing to upper and lower limit is my best option. Hardware is a blackfin DSP, I noticed that a sharc has a native clip instruction... best regards, Andre
Reply by ●January 15, 20152015-01-15
On 01/15/2015 04:52 PM, Andre wrote:> On 15.01.2015 00:43, kaz wrote: >> For FPGAs you can either check the whole value or better just the few MSB >> bits to be ddiscarded as follows: >> >> -- checking one MSB bit and clipping, assuming data1 is 17 bits signed >> if data1(16:15) = “00” or data1(16:15) = “11” then >> data <= data1(15 downto 0); >> elsif data1(16) = '1' then >> data <= -32767; >> else >> data <= +32767; >> end if; >> >> >> Kaz >> >> _____________________________ >> Posted through www.DSPRelated.com >> > > Dear all, > > thanks for all the responses, I was just a bit upset that comparing to > upper and lower limit is my best option. > > Hardware is a blackfin DSP, I noticed that a sharc has > a native clip instruction... > > best regards, > > Andre >Are you sure there isn't an option to saturate to any selected bit position in a Blackfin? Its years since I touched one, but I thought it had that. Regards, Steve
Reply by ●January 15, 20152015-01-15
On Thu, 15 Jan 2015 09:52:11 +0100, Andre <lodwig@pathme.de> wrote:>On 15.01.2015 00:43, kaz wrote: >> For FPGAs you can either check the whole value or better just the few MSB >> bits to be ddiscarded as follows: >> >> -- checking one MSB bit and clipping, assuming data1 is 17 bits signed >> if data1(16:15) = “00” or data1(16:15) = “11” then >> data <= data1(15 downto 0); >> elsif data1(16) = '1' then >> data <= -32767; >> else >> data <= +32767; >> end if; >> >> >> Kaz >> >> _____________________________ >> Posted through www.DSPRelated.com >> > >Dear all, > >thanks for all the responses, I was just a bit upset that comparing to >upper and lower limit is my best option. > >Hardware is a blackfin DSP, I noticed that a sharc has >a native clip instruction...It is probably an asymetric 2's complement clip, which most are. If you want an symmetric clip like you previously described, you pretty much have to compare, although the "compare" can take different tricky forms depending on what you're doing..>best regards, > >Andre >Eric Jacobsen Anchor Hill Communications http://www.anchorhill.com
