CMSIS DSP is not being included

JosuGaztelu · February 23, 2023, 3:21pm

Question/Issue:
I tried to deploy a model on a new device, with a EFR32MG22C224F512IM40-C which has a Cortex-M33, and I’ve realised that CMSIS-DSP and CMSIS-NN are not being included and I’d like to know how could I use them.
Thanks
Project ID:

Context/Use case:

mateusz · March 1, 2023, 7:19am

Hi @JosuGaztelu,
Could you write what kind of deployment from the Studio you use (C++ Library, SLCC Component)?
Also, what toolchain/IDE do you use for compilation?
As a temporary workaround, try to set these three global defines:
EIDSP_USE_CMSIS_DSP=1
EIDSP_LOAD_CMSIS_DSP_SOURCES=1
EI_CLASSIFIER_TFLITE_ENABLE_CMSIS_NN=1

Best regards,
Mateusz

JosuGaztelu · March 1, 2023, 8:47am

I used the C++ library, and I did define those three global defines, but what I dont understand is why using those defines or not the inference time is always the same. The toolchain is GNU ARM v10.3.1 and I’m using the Simplicity Studio v5 as IDE.

Best Regards,
Josu

mateusz · March 1, 2023, 10:14am

OK, if you are using Simplicity Studio and GCC, then those defines should be added implicitly here:

github.com

edgeimpulse/inferencing-sdk-cpp/blob/master/dsp/config.hpp#L45

      
        
                        #define EIDSP_USE_CMSIS_DSP      1
                    #endif // Mbed OS 5.7 version check
            
            
        // Arduino on Mbed targets prior to Mbed OS 6.0.0 ship their own CMSIS-DSP sources
                    #if defined(ARDUINO) && (MBED_VERSION < MBED_ENCODE_VERSION((6), (0), (0)))
                        #define EIDSP_LOAD_CMSIS_DSP_SOURCES      0
                    #else
                        #define EIDSP_LOAD_CMSIS_DSP_SOURCES      1
                    #endif // Mbed OS 6.0 version check
                #else
                    #define EIDSP_USE_CMSIS_DSP		        1
                    #define EIDSP_LOAD_CMSIS_DSP_SOURCES    1
                #endif
            #else
                #define EIDSP_USE_CMSIS_DSP     0
            #endif // Mbed / ARM Core check
            #endif // ifndef EIDSP_USE_CMSIS_DSP
            
            
#if EIDSP_USE_CMSIS_DSP == 1
            #define EIDSP_i32                int32_t
            #define EIDSP_i16                int16_t

github.com

edgeimpulse/inferencing-sdk-cpp/blob/master/classifier/ei_classifier_config.h#L42

      
        
            #if defined(__MBED__)
                #include "mbed.h"
                #if (MBED_VERSION < MBED_ENCODE_VERSION(5, 7, 0))
                    #define EI_CLASSIFIER_TFLITE_ENABLE_CMSIS_NN      0
                #else
                    #define EI_CLASSIFIER_TFLITE_ENABLE_CMSIS_NN      1
                #endif // Mbed OS 5.7 version check
            
            
// __ARM_ARCH_PROFILE is a predefine of arm-gcc.  __TARGET_* is armcc
            #elif __ARM_ARCH_PROFILE == 'M' || defined(__TARGET_CPU_CORTEX_M0) || defined(__TARGET_CPU_CORTEX_M0PLUS) || defined(__TARGET_CPU_CORTEX_M3) || defined(__TARGET_CPU_CORTEX_M4) || defined(__TARGET_CPU_CORTEX_M7) || defined(ARDUINO_NRF52_ADAFRUIT)
                #define EI_CLASSIFIER_TFLITE_ENABLE_CMSIS_NN      1
            #else
                #define EI_CLASSIFIER_TFLITE_ENABLE_CMSIS_NN      0
            #endif
            #endif // EI_CLASSIFIER_TFLITE_ENABLE_CMSIS_NN
            
            
// CMSIS-NN falls back to reference kernels when __ARM_FEATURE_DSP and __ARM_FEATURE_MVE are not defined
            // we should never use those... So disable CMSIS-NN in that case and throw a warning
            #if EI_CLASSIFIER_TFLITE_ENABLE_CMSIS_NN == 1
                #if !defined(__ARM_FEATURE_DSP) && !defined(__ARM_FEATURE_MVE)
                    #pragma message( \

We are enabling CMSIS-DSP and CMSIS-NN for all ARM-based targets to improve their performance.

Can eventually disable CMSIS-NN and CMSIS-DSP by explicitly defining:
EIDSP_USE_CMSIS_DSP=0
EI_CLASSIFIER_TFLITE_ENABLE_CMSIS_NN=0

Best regards,
Mateusz

JosuGaztelu · March 21, 2023, 8:35am

I found out that it is due to the option of take log of spectrum, it causes to not have an impact between using or not CMSIS-DSP. I’m not sure if it is due to my processors specification so if anyone knows why could it be or could give me a hint I would be very grateful.

Thanks,
Josu

mateusz · March 21, 2023, 9:06am

Hi @JosuGaztelu
What do you mean by:

due to the option of take log of spectrum

Is it possible to share your project source code? Or if you can build a minimal example where you are observing the same inference time regardless of the use of CMSIS-DSP and CMSIS-NN.

Best regards,
Mateusz

JosuGaztelu · March 21, 2023, 9:07am

When selecting the spectral features as preprocessing block it has an otion to use logaritmic scale:

if this is selected CMSIS-DSP will not have any impact on DSP timing.
Should it have something to do maybe with the __ARM_FEATURE_DSP or any other define?

AlexE · March 21, 2023, 9:29pm

@JosuGaztelu can you post a few things to help us repro this?

For the deployment that runs faster, can you paste the serial output with the reported timings, along with model_variables.h?

And for the deployment the runs slower (with overlap fft frames enabled), can you post the same output and file?

JosuGaztelu · March 22, 2023, 7:41am

Hi I dont know what i did, but now it seems to be working correctly with new project and CMSIS-DSP now works correctly, maybe it could be that in this case i defines the __ARM_FEATURE_DSP=1.
Thanks for the answers