|
I'm using a 21065L and am trying to find the quickest way to split an interleaved integer array into two float arrays. Details : My input array consists of 128 samples of codec data as i1,i0,i1,i0,i1,i0,i1,i0... I need to convert this into two separate float arrays as : f1,f1,f1,f1... f0,f0,f0,f0... Currently I'm doing something like this : r2 = dm(_l_NumPairs) ; LCNTR=r2 , DO xfer UNTIL LCE ; // Loop through the input array by twos. r0 = dm(i4,m4) ; // Get sample for zero. f1 = float r0 ; // Convert it to float. dm(i2,m1) = r1 ; // Store it in output zero. r0 = dm(i4,m4) ; // Get sample for one. f1 = float r0 ; // Convert it to float. xfer: dm(i1,m1) = r1 ; // Store it in output one. Is there any way to eliminate the intermediate r0 assignment? Should either the input or output array be in code memory space? Would it make a difference? Thanks. Steve Holle Link Communications, Inc. 1035 Cerise Rd. Billings, MT 59101 |
|
|
Fast int array to float array convert.
Started by ●October 19, 2004
Reply by ●October 20, 20042004-10-20
|
Steve Holle wrote: ... > Currently I'm doing something like this : > r2 = dm(_l_NumPairs) ; > LCNTR=r2 , DO xfer UNTIL LCE ; > r0 = dm(i4,m4) ; > f1 = float r0 ; > dm(i2,m1) = r1 ; > r0 = dm(i4,m4) ; > f1 = float r0 ; > xfer: dm(i1,m1) = r1 ; > > Is there any way to eliminate the intermediate r0 assignment? No. Steve, you write assembler code like a c compiler (no offense :-). > Should either the input or output array be in code memory space? > Would it make a difference? Yes. Try this (assume that the float array is in pm, i10 is the corresponding pointer and m9 = m1): r2 = dm(_l_NumPairs) ; r2=r2-1, r0 = dm(i4,m4) ; f1 = float r0 , r0 = dm(i4,m4); LCNTR=r2 , DO xfer UNTIL LCE ; f1 = float r0, r0 = dm(i4,m4), pm(i10,m9) = f1 ; xfer: f1 = float r0, r0 = dm(i4,m4), pm(i10,m9) = f1 ; f1 = float r0, pm(i10,m9) = f1 ; pm(i10,m9) = f1 ; This code should run three times faster than your original. Regards, Andor |
|
|
Reply by ●October 20, 20042004-10-20
|
hai all, i am devoloping digital effect and recorder in one module. in that for 16 bit to 8 bit and vice versa conversion i had applied mu_law companding. when i am recording plain sound is good and delay,flange and chorous recording also very fine...but when i am recording reverb more noise adding (hissing noise).for this i had applied fir filter.hissing noise solved...but for low volume like end of guitar play more noise adding.. is there any idea to solve this problem.kindly let me know. thanks in advance... with kind regards ayyams ________________________________________________________________________ Yahoo! India Matrimony: Find your life partner online Go to: http://yahoo.shaadi.com/india-matrimony |
|
|
Reply by ●October 20, 20042004-10-20
|
On Tue, 19 Oct 2004, Steve Holle wrote: > > > I'm using a 21065L and am trying to find the quickest way to split an > interleaved integer array into two float arrays. > > Details : > My input array consists of 128 samples of codec data as > i1,i0,i1,i0,i1,i0,i1,i0... > I need to convert this into two separate float arrays as : > f1,f1,f1,f1... > f0,f0,f0,f0... > > Currently I'm doing something like this : > r2 = dm(_l_NumPairs) ; > LCNTR=r2 , DO xfer UNTIL LCE ; // > Loop through the input array by twos. > r0 = dm(i4,m4) ; // > Get sample for zero. > f1 = float r0 ; // > Convert it to float. > dm(i2,m1) = r1 ; // I'd have done it r1 = dm(i4,m4); f1 = float r1; dm(i2,m1) = f1; Which is really the same thing, but you only use one register. > > Is there any way to eliminate the intermediate r0 assignment? Nope. > Should either the input or output array be in code memory space? Would it > make a difference? It could, depending on what you do with the data later. If you take more than 32 instructions in cache it won't make any difference at all, but if you do this in a long loop of less than 32 instructions then you can go a lot faster. If you do dual ram access in one instruction then putting the data in pm rather than dm can be helpful. If you're not needing that kind of memory access, then staying in dm space is easier to think about. Patience, persistence, truth, Dr. mike |
Reply by ●October 20, 20042004-10-20
|
>No. Steve, you write assembler code like a c compiler (no offense :-). Yah, I just cannot get my brain around parallel processing. Everything looks linear to me until someone explains it. I need a brain transplant :D Your example was just what I was looking for. Thanks. > > Should either the input or output array be in code memory space? > > Would it make a difference? >Yes. Try this (assume that the float array is in pm, i10 is the >corresponding pointer and m9 = m1): > >r2 = dm(_l_NumPairs) ; >r2=r2-1, r0 = dm(i4,m4) ; >f1 = float r0 , r0 = dm(i4,m4); > >LCNTR=r2 , DO xfer UNTIL LCE ; > f1 = float r0, r0 = dm(i4,m4), pm(i10,m9) = f1 ; >xfer: f1 = float r0, r0 = dm(i4,m4), pm(i10,m9) = f1 ; > >f1 = float r0, pm(i10,m9) = f1 ; >pm(i10,m9) = f1 ; > >This code should run three times faster than your original. > >Regards, >Andor >_____________________________________ >Note: If you do a simple "reply" with your email client, only the author >of this message will receive your answer. You need to do a "reply all" if >you want your answer to be distributed to the entire group. > >_____________________________________ >About this discussion group: > >To Join: Send an email to > >To Post: Send an email to > >To Leave: Send an email to > >Archives: http://groups.yahoo.com/group/adsp > >Other Groups: http://www.dsprelated.com/groups.php3 > >Yahoo! Groups Links > > Steve Holle Link Communications, Inc. 1035 Cerise Rd. Billings, MT 59101 |
Reply by ●October 20, 20042004-10-20
|
Hi, Does your noise come from quantization, maybe? What precision are you using? 16 bit fixed point, 32 bit fixed point, 32 bit floating point? Would a noise gate help? JaaC --- ayyam perumal <> wrote: > > > hai all, > > i am devoloping digital effect and recorder > in one module. > in that for 16 bit to 8 bit and vice versa > conversion > i had applied > mu_law companding. > > when i am recording plain sound is good and > delay,flange and > chorous recording also very fine...but when i am > recording reverb more > noise adding (hissing noise).for this i had applied > fir filter.hissing > noise solved...but for low volume like end of guitar > play more noise > adding.. > > is there any idea to solve this problem.kindly > let > me know. > > thanks in advance... > > with kind regards > ayyams ________________________________________________________________________ > Yahoo! India Matrimony: Find your life partner > online > Go to: http://yahoo.shaadi.com/india-matrimony > _____________________________________ > Note: If you do a simple "reply" with your email > client, only the author of this message will receive > your answer. You need to do a "reply all" if you > want your answer to be distributed to the entire > group. > > _____________________________________ > About this discussion group: > > To Join: Send an email to > To Post: Send an email to > > To Leave: Send an email to > Archives: http://groups.yahoo.com/group/adsp > > Other Groups: http://www.dsprelated.com/groups.php3 > > Yahoo! Groups Links > ===== Jaime Andr Aranguren Cardona __________________________________________________ |
|
|
Reply by ●October 20, 20042004-10-20
|
>Yes. Try this (assume that the float array is in pm, i10 is the >corresponding pointer and m9 = m1): > >r2 = dm(_l_NumPairs) ; >r2=r2-1, r0 = dm(i4,m4) ; >f1 = float r0 , r0 = dm(i4,m4); > >LCNTR=r2 , DO xfer UNTIL LCE ; > f1 = float r0, r0 = dm(i4,m4), pm(i10,m9) = f1 ; >xfer: f1 = float r0, r0 = dm(i4,m4), pm(i10,m9) = f1 ; > >f1 = float r0, pm(i10,m9) = f1 ; >pm(i10,m9) = f1 ; > >This code should run three times faster than your original. > >Regards, >Andor Is there a way to turn this around so that my float output arrays are in dm and my integer input arrays are in pm? >_____________________________________ >Note: If you do a simple "reply" with your email client, only the author >of this message will receive your answer. You need to do a "reply all" if >you want your answer to be distributed to the entire group. > >_____________________________________ >About this discussion group: > >To Join: Send an email to > >To Post: Send an email to > >To Leave: Send an email to > >Archives: http://groups.yahoo.com/group/adsp > >Other Groups: http://www.dsprelated.com/groups.php3 > >Yahoo! Groups Links > > Steve Holle Link Communications, Inc. 1035 Cerise Rd. Billings, MT 59101 |
|
|
Reply by ●October 21, 20042004-10-21
|
I think I figured it out. Thanks for your help. Here's what I wound up with : r12 = r12 -1, r0 = pm(i12,m12) ; f1 = float r0, r0 = pm(i12,m12) ; r2 = pm(i13,m11); f4 = float r2, r2 = pm(i13,m11) ; LCNTR=r12, DO xfer UNTIL LCE ; f1 = float r0, dm(i4,m4) = f1, r0 = pm(i12,m12) ; xfer: f4 = float r2, dm(i5,m4) = f4, r2 = pm(i13,m11) ; dm(i4,m4) = f1 ; dm(i5,m4) = f4 ; At 02:50 PM 10/20/2004, Steve Holle wrote: > >Yes. Try this (assume that the float array is in pm, i10 is the > >corresponding pointer and m9 = m1): > > > >r2 = dm(_l_NumPairs) ; > >r2=r2-1, r0 = dm(i4,m4) ; > >f1 = float r0 , r0 = dm(i4,m4); > > > >LCNTR=r2 , DO xfer UNTIL LCE ; > > f1 = float r0, r0 = dm(i4,m4), pm(i10,m9) = f1 ; > >xfer: f1 = float r0, r0 = dm(i4,m4), pm(i10,m9) = f1 ; > > > >f1 = float r0, pm(i10,m9) = f1 ; > >pm(i10,m9) = f1 ; > > > >This code should run three times faster than your original. > > > >Regards, > >Andor > >Is there a way to turn this around so that my float output arrays are in dm >and my integer input arrays are in pm? > > >_____________________________________ > >Note: If you do a simple "reply" with your email client, only the author > >of this message will receive your answer. You need to do a "reply all" if > >you want your answer to be distributed to the entire group. > > > >_____________________________________ > >About this discussion group: > > > >To Join: Send an email to > > > >To Post: Send an email to > > > >To Leave: Send an email to > > > >Archives: http://groups.yahoo.com/group/adsp > > > >Other Groups: http://www.dsprelated.com/groups.php3 > > > >Yahoo! Groups Links > > > > > > > > > >Steve Holle >Link Communications, Inc. >1035 Cerise Rd. >Billings, MT 59101 > > >_____________________________________ >Note: If you do a simple "reply" with your email client, only the author >of this message will receive your answer. You need to do a "reply all" if >you want your answer to be distributed to the entire group. > >_____________________________________ >About this discussion group: > >To Join: Send an email to > >To Post: Send an email to > >To Leave: Send an email to > >Archives: http://groups.yahoo.com/group/adsp > >Other Groups: http://www.dsprelated.com/groups.php3 > >Yahoo! Groups Links > > Steve Holle Link Communications, Inc. 1035 Cerise Rd. Billings, MT 59101 |
Reply by ●October 22, 20042004-10-22
|
Hai, noise is comming from quantization,i am using 16-bit fixed point (adsp2186).. with good regards ayyams --- Jaime Andres Aranguren Cardona <> wrote: > Hi, > > Does your noise come from quantization, maybe? What > precision are you using? 16 bit fixed point, 32 bit > fixed point, 32 bit floating point? > > Would a noise gate help? > > JaaC > ________________________________________________________________________ Yahoo! India Matrimony: Find your life partner online Go to: http://yahoo.shaadi.com/india-matrimony |
|
|
Reply by ●October 22, 20042004-10-22
|
If the noise is coming from quantization at the analog-to-digital stage, the way to reduce it is to make sure your signal is as large as possible (without clipping). If the quantization is from processing, then using double-precision math can help considerably. When processing with 16-bit single precision, you may also need to pay attention to internal scaling. For example doing something that reduces the signal level considerably will add quantization noise. In either case I'm assuming that increasing the number of bits is not possible due to hardware limitations. --- ayyam perumal <> wrote: > Hai, > noise is comming from quantization,i am using > 16-bit fixed point (adsp2186).. > > with good regards > ayyams > > --- Jaime Andres Aranguren Cardona > <> wrote: > > Hi, > > > > Does your noise come from quantization, maybe? What > > precision are you using? 16 bit fixed point, 32 bit > > fixed point, 32 bit floating point? > > > > Would a noise gate help? > > > > JaaC _______________________________ |
