topic Re: Has anybody implemented CMSIS FFT library Successfully? in Kinetis Microcontrollers

Has anybody implemented CMSIS FFT library Successfully?

Amit_Kumar1 — Thu, 26 Jun 2014 10:51:37 GMT

I want to use arm_cfft_radix4_init_q15() and arm_cfft_radix4_q15() function from CMSIS 3.2 library . I am unable to find any good resources or any examples. I have looked in the documents provided by ARM i.e CMSIS-DSP: Complex FFT Functions but I didn't find it useful. What are the inputs given to these functions I am still not clear with it. I tried writing the codes but it didn't worked

this is how I am calling the functions,

arm_cfft_radix4_init_q15(&arm_cfft_sR_q15_len256, 256, 0, 1); // For init

arm_cfft_radix4_q15(&arm_cfft_sR_q15_len256, az); // For transformation where,

arm_cfft_radix4_instance_q15 arm_cfft_sR_q15_len256 = {256, 0, 1, twiddleCoef_256, armBitRevIndexTable256, 4U, 4U};

and az[512] is an array where the data is stored. Just for testing

float32_t az[512] = {1.2 ,0.0

1.3, 0.0

-------

--------

1.2, 0.0};

my programs hangs and I am not able to print the o/p. I am using Codewrrior 10.5 and want to implement this on FRDM-K20D50M. I posted the query in ARM's community also but I didn't get any response which makes me think if anybody is actually using this library. Please look into the matter.

Kind Regards

Amit Kumar

Re: Has anybody implemented CMSIS FFT library Successfully?

martynhunt — Fri, 27 Jun 2014 19:57:46 GMT

Hi,

Here is the FFT bin example from an older version of CMSIS-DSP. Hopefully this helps you. It is for f32, but should work if you change the names to q15 versions.

Best regards,

Martyn

Re: Re: Has anybody implemented CMSIS FFT library Successfully?

Amit_Kumar1 — Tue, 01 Jul 2014 04:59:21 GMT

Hi martynhunt

I Tried implementing the CFFT code into my project. It gave an error that m_text overflowed. Then I configured the flex mem and increased the size but still the error was the same and the amount by which it was being overflowed was the same, then I reduced the input buffer size from 2048 to 32 but still the same message. I tested the library by implementing sine, cosine function example code and that is working fine then I tried to implement CFFT example code which still didn't worked. Currently I am using CMSIS 3.2, with codewarrior10.6 and I am implementing this on FRDM-K20D50M board. Attached is my project with built library please have a look into it.

The error snapshot when the data array was 256.

Kind Regards

Amit Kumar

Re: Re: Has anybody implemented CMSIS FFT library Successfully?

BlackNight — Tue, 01 Jul 2014 05:30:15 GMT

m_text overflowed means that your application is too big, and according to the screenshot it is over by more than 50 KByte.

I have not checked your project, but it looks like your are linking too much?

Re: Has anybody implemented CMSIS FFT library Successfully?

Amit_Kumar1 — Tue, 01 Jul 2014 05:50:39 GMT

HI Erich

Thanks for the reply. How do I overcome this error? any suggestions? I checked with reduced array size from 2048 to 32. but the size by which it is overflowing is the same. there is a warning message also which says invalid project path. Is that message of any concern?

Kind Regards

Amit Kumar

Re: Re: Has anybody implemented CMSIS FFT library Successfully?

martynhunt — Tue, 01 Jul 2014 17:03:39 GMT

Hi,

It looks like the GCC linker is pulling in all of the data tables in arm_common_tables.h. See .map file excerpt:

.rodata 0x00005e04 0xc ./Sources/main.o

.rodata 0x00005e10 0x26d58 C:/Users/B45635/Downloads/CMSIS-SP-00300-r3p2-00rel1/CMSIS/Lib/GCC\libarm_cortexM4l_math.a(arm_common_tables.o)

0x00005e10 twiddleCoef_rfft_4096

0x00009e10 twiddleCoef_rfft_2048

0x0000be10 twiddleCoef_rfft_1024

0x0000ce10 twiddleCoef_rfft_512

0x0000d610 twiddleCoef_rfft_256

0x0000da10 twiddleCoef_rfft_128

0x0000dc10 twiddleCoef_rfft_64

0x0000dd10 twiddleCoef_rfft_32

0x0000dd90 armBitRevIndexTable4096

0x0000fd10 armBitRevIndexTable2048

0x00011ad0 armBitRevIndexTable1024

0x000128e0 armBitRevIndexTable512

0x00012c60 armBitRevIndexTable256

0x00012fd0 armBitRevIndexTable128

0x00013170 armBitRevIndexTable64

0x000131e0 armBitRevIndexTable32

0x00013240 armBitRevIndexTable16

0x00013268 armRecipTableQ31

0x00013368 armRecipTableQ15

0x000133e8 twiddleCoefQ15

0x000163e8 twiddleCoefQ31

0x0001c3e8 twiddleCoef_4096

0x000243e8 twiddleCoef_2048

0x000283e8 twiddleCoef_1024

0x0002a3e8 twiddleCoef_512

0x0002b3e8 twiddleCoef_256

0x0002bbe8 twiddleCoef_128

0x0002bfe8 twiddleCoef_64

0x0002c1e8 twiddleCoef_32

0x0002c2e8 twiddleCoef_16

0x0002c368 armBitRevTable

.rodata 0x0002cb68 0x20 C:/Freescale/CW MCU v10.6/MCU/ARM_GCC_Support/ewl/lib/armv7e-m\libc.a(__strtold.o)

This appears to be an issue with how the ARM library behaves when being built with GCC compiler.

Best regards,

Martyn

Re: Re: Has anybody implemented CMSIS FFT library Successfully?

martynhunt — Wed, 02 Jul 2014 16:06:34 GMT

Hi,

I have some good news. Your build of the library has the correct compiler options set, and your applications does as well. However, it appears you are linking the application to the pre-built library provided by CMSIS. This pre-built version does not have -ffunction-sections -fdata-sections set; therefore, the unused data will not be removed.

To link your rebuild of the library correctly into your project you need to make sure you have the following properties set:

and:

Please let me know if this works for you.

Best regards,

Martyn

P.S. For reference the correct compiler settings for library and application are:

Library Compiler:

Application compiler and linker options:

Re: Re: Has anybody implemented CMSIS FFT library Successfully?

Amit_Kumar1 — Thu, 03 Jul 2014 03:29:23 GMT

Hi Martyn

Thanks for your support. It worked

Kind Regards

Amit Kumar

Re: Has anybody implemented CMSIS FFT library Successfully?

ibrahimerturk — Sun, 09 Nov 2014 02:07:58 GMT

Hi, I'm in trouble with the CMSIS DSP library bitreversal function and solved it the workaround described here. Hope this helps.

Regards

Ibrahim

Re: Has anybody implemented CMSIS FFT library Successfully?

egoodii — Tue, 11 Nov 2014 00:51:14 GMT

Are you saying that the Cortex-M 'bit reverse' single-instruction isn't working for this applications?

The IAR toolchain gives the programmer access to these 'special capabilities' via 'intrinsic functions' that let the compiler handle normal register allocation while inserting the particular instruction in the sequence like an 'inline function':

Syntax

unsigned long __RBIT(unsigned long);

Description

Inserts an RBIT instruction, which reverses the bit order in a 32-bit register.

Re: Has anybody implemented CMSIS FFT library Successfully?

ibrahimerturk — Tue, 11 Nov 2014 23:22:27 GMT

It is not single-instruction. It is a function that reverses bit order of all elements of FFT output array after calling FFT over an array. The function is member of ARM CMSIS Math (DSP) Library.

Re: Has anybody implemented CMSIS FFT library Successfully?

mjbcswitzerland — Tue, 10 Jan 2017 15:01:38 GMT

Hi All

I am hoping to find some people who have used this with floating points and could maybe explain why I get these results?

Test code:

float fft_input_buffer[512];
float fft_magnitude_buffer[256];

fnInjectFloat(fft_input_buffer, 512);

arm_cfft_f32(&arm_cfft_sR_f32_len256, fft_input_buffer, 0, 1); // perform complex fast-fourier transform (the result is in the input buffer)
arm_cmplx_mag_f32(fft_input_buffer, fft_magnitude_buffer, 256); // calculate the magnitude of each frequency component (256 bins are now available) - 48kHz sampling rate gives 0..24kHz spectrum with 93Hz resolution per bin

The function fnInjectFloat() is putting a high resolution test signal, in this specific reference case a 7'000Hz sine wave, into the input buffer.

Before calling the FFT the input buffer can be displayed to have the expected input signal (75 cycles in 512 input buffer):

After performing the FFT and calculating the magnitude of the output bins one expects to find the energy at 7kHz. This is however the result in the magnitude buffer:

There is a nice peak at 7kHz but there is a second peak at 17kHz (24kHz - 7kHz) with three times the energy.

I have tested with various FFT lengths with the same result.
I have tested on various chips (m0+, m4 - with and without FPU) and in each case with the same result.

If the test frequency is increase, eg. to 8kHz the second component is at 16kHz (24kHz - 8 kHz). If a frequency greater than 12kHz is injected the unexpected bin energy is less that 12kHz.

Finally, if real sampled input is used - rather than preparing a signal in the input buffer - the same results are also obtained in each case.

I can't believe that the CMSIS FFT doesn't work (I tried version 3.2 and also the very latest V5.1) but the results are presently not usable.... The only code used is the CMSIS library code so I presently see no influences that can cause errors to be introduced.

Can anyone verify or explain what is happening?

Regards

Mark

Re: Has anybody implemented CMSIS FFT library Successfully?

mjbcswitzerland — Wed, 11 Jan 2017 11:23:49 GMT

I would like to add that when the ARM test input "arm_fft_bin_data.c" is used instead of other inputs or real signals the output does pass the ARM test case (it checks for maximum energy in bin 213 as criteria) but it seems that passing the test case doesn't prove that it actually works.... (???????)

Regards

Mark

Re: Has anybody implemented CMSIS FFT library Successfully?

mjbcswitzerland — Sat, 14 Jan 2017 00:03:24 GMT

Hi All

Here are a further simple comparison result.

Another example where we have taken a simple 16 point FFT with a windowed signal (32 samples of a simple test single).

The same 32 input samples into the same FFT in Matlab gives:

There are certain similarities and their is a feeling that something is inverted but no attempts to work out what needs to be moved around has been successful yet!

The present situation is than no one here (several engineers with DSP/signal processing experience) have any idea what is going on to give such results....

Regards

Mark

Re: Has anybody implemented CMSIS FFT library Successfully?

egoodii — Mon, 16 Jan 2017 23:00:20 GMT

Without actually delving into any details myself(!), I will agree that this test makes it look like some array addressing is swapped 'end for end', shifting some results to the wrong end --- what was the complete resolution to the 11/11/14 'address reversal' question/answer?

Re: Has anybody implemented CMSIS FFT library Successfully?

mjbcswitzerland — Mon, 16 Jan 2017 23:33:25 GMT

Earl

I have realised that the CMSIS CFFT requires that the inputs samples are complex. If I insert a 0 between each sample (representing a zero complex valid) I then get correct results. The ARM reference also has 0.0 at every second array location in its input, which is why it works.

There is also a RFFT version (real) which I was hoping would save having to put in the zeros but initial tests didn't confirm this. For the moment I am continuing with the CFFT and might retry the RFFT later again (it may be faster since it performs a CFFT of half the size and then extracts real value from it [somehow]).

Since I wasn't getting an responses here or at mbed I tried "DSPrelated" and quickly got some ideas that led to the resolution.

Regards

Mark

Re: Has anybody implemented CMSIS FFT library Successfully?

egoodii — Tue, 17 Jan 2017 01:31:15 GMT

Cool! Glad to hear you're making forward progress again! They don't call it 'complex' for nothin'....

Re: Has anybody implemented CMSIS FFT library Successfully?

carstengroen — Fri, 29 May 2020 08:01:49 GMT

I'm sorry to re-activate a 3 year old thread, but I wonder if any of you got the rfft to work ? I have a project where I get data from a sensor, initially I used the rfft function from CMSIS-DSP (vers 1.8.0), but I also see the spurious at the top of the bins. I switched over to cfft (and interleaved the real data with zeros) and everything works as expected.

As I'm a little tight on execution speed, I would love to get the rfft to work properly.

Any ideas, or is there something fishy with the rfft ?

Re: Has anybody implemented CMSIS FFT library Successfully?

mjbcswitzerland — Fri, 29 May 2020 12:40:30 GMT

Carsten

Personally I didn't revisit the RFFT and also moved to a different library due to the requirement to be able to do arbitrary length FFTs (the CMSIS ones can only do power of two lengths and are also limit in maximum length) which means that I can't give any feedback about the RFFT operation or accuracy.

Sorry,

Mark
[uTasker project developer for Kinetis and i.MX RT]

Re: Has anybody implemented CMSIS FFT library Successfully?

carstengroen — Mon, 02 Nov 2020 13:10:04 GMT

Thanks mjbcswitzerland‌

I spent a couple days messing around with the rfft before I found your post above which saved my day

Anyway, I will continue for now with the cfft, and maybe later, look once more at the rfft (I have seen a couple of demos that uses this function) when I have the time.

I still need to understand the scaling of the output from the cfft as there seems not to be a 1:1 relationship between input and output (any idea maybe ?)

My code (and comment) so far:

//++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
// Using cfft (and mag) functions, a 1.0 in input level yields 256.0 in the mag output level.
// Looking at the output of tempBuffer after the cfft (feeding a sine at 1.0 peakvalue, 
// the real part of the peak bin is -0.0012 and the img. part is -256)
// Process the data through the CFFT/CIFFT module
arm_cfft_f32(&varInstCfftF32, tempBuffer, 0, 1); 
// Convert the complex output from the cfft function to real numbers, we only do half the points as the second half is negative/mirrored data)
arm_cmplx_mag_f32(tempBuffer, freqBuffer, FFT_SIZE / 2);
//++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
‍‍‍‍‍‍‍‍‍

PS: Thanks for your many articles & videos on the RT series, very nice indeed !