CMSIS-DSP  Version 1.10.0
CMSIS DSP Software Library
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
Revision History of CMSIS-DSP
Version Description

Float promotion issues:

  • Removed remaining issues of implicit float promotion (mainly in f16 code)

Scalar versions of arm_vlog_q31 and arm_vlog_q15

Synchronous Data Flow (scripts in github repository)

Python wrapper available on official Python repository

MFCC F32, F16, Q31 and Q15

Typecasting functions for converting between f16 and s16 in scalar.

More f64 functions

arm_cmplx_mag_fast_q15: The original arm_cmplx_mag_q15 is now more accurate but slower.

Added arm_mat_mult_opt_q31 for a faster version of arm_mat_mult_q31 when using Helium (but requiring more storage for intermediate results).




Trigonometric functions:

  • atan2 in f32, f16, q31 and q15.

Re-organization of arm_math.h

Re-organization of interpolation functions:

  • New Interpolation folder

New q7 functions:

Matrix x vector multiply

Complex Matrix Transpose

Cholesky (LL^t and LDL^t decompositions)

More support for f16 datatype:

  • arm_cfft_f16.c
  • arm_rfft_fast_f16.c
  • f16 basic Maths
  • f16 complex Maths
  • f16 filtering (FIR and BiQuad and correlate)
  • f16 matrix
  • f16 interpolation
  • f16 statistics
  • f16 support functions
  • f16 svm
  • f16 distances
  • f16 bayes

Support for Quaternions

Levinson Durbin algorithm

Q15 scalar division

New absmax and absmin functions

GCC issues: Some functions have build issues when building Helium code with gcc. So they are replaced with a less efficient scalar version when built with gcc.


  • Initialization of FIRs has changed for the Helium version (state and coefs). Read the corresponding documentations.
  • Initialization of float biquad DF1 has changed for the Helium version (two init functions are required). Read the corresponding documentations.
  • The states of the vectorized filters are different from their scalar versions (which may break some algorithms which are working on the state of a filter like Goertzel)
  • For performance reasons, some vectorized algorithms may read one vector after the end of buffer (but the data read is not used). If the buffer is close to the end of a memory zone, it is advised to leave some free memory area between the end of the buffer and the end of the memory zone.


Helium implementation of lots of functions in the CMSIS-DSP. More will be added in future releases.


New folder SVMFunctions : Support vector machines.

  • Training must be done with scikit-learn.
  • DSP/Examples/ARM/arm_svm_example is showing how to train with scikit-learn and dump the result for use with CMSIS-DSP

New folder BayesFunctions : Functions related to Bayesian probability. In this version only a naive gaussian classifier.

  • Training must be done with scikit-learn.
  • DSP/Examples/ARM/arm_bayes_example is showing how to train with scikit-learn and dump the result for use with CMSIS-DSP

New folder DistanceFunctions:

  • Distance functions for clustering algorithms.

For the new functions in SVMFunctions, BayesFunctions and DistanceFunctions, a tutorial is available.

New BasicMathFunctions:

  • Elementwise bit operations.

New support functions:

  • arm_barycenter_f32 and arm_weighted_sum_f32.
  • Sorting functions.
  • Spline interpolation.

New statistics functions:

  • arm_entropy_f32, arm_kullback_leibler_f32, arm_logsumexp_f32 and arm_logsumexp_dot_prod_f32

New testing framework:

  • Developed for our internal needs. We have released it but won't give support for the first few releases since it is a work in progress.

Improvements to the cmake build:

  • It is more easy to switch toolchains. CMAKE_PREFIX_PATH can be used.


New Neon implementations of f32 functions

  • Can be enabled with macros ARM_MATH_NEON and ARM_MATH_NEON_EXPERIMENTAL

Python wrapper compatible with Numpy

  • DSP/PythonWrapper/ is explaining how to build the wrapper

Preliminary cmake build

  • DSP/ is explaining how to use it. This first version is restricted to AC6 toolchain and assume (hard coded path) that ArmDS is used. Future versions will make it easier to switch toolchains.

Modified arm_math.h

  • Removed dependance on cores. It is no more testing for some specific cores but is now using features like ARM_MATH_LOOPUNROLL (see comments in arm_math.h)
  • Deprecated __SIMD32 and introduced new functions for SIMD accesses.

Added compilation flags for FFT:

  • It is now possible to include only the tables required for FFTs. It is explained in DSP/ since it is simpler to use it from the cmake (which is defining the right macros). But the macros can also be defined without cmake.


Reworked DSP library source files

  • added macro ARM_MATH_LOOPUNROLL
  • added const-correctness
  • replaced SIMD pointer construct with memcopy solution
  • replaced macro combination "CMSIS_INLINE __STATIC_INLINE with "__STATIC_FORCEINLINE"

Reworked DSP library documentation

Changed DSP folder structure

  • moved DSP libraries to ./DSP/Lib

ARM DSP Libraries are built with ARMCLANG

Added DSP Libraries source variant


Updated IAR projects

Note: binaries NOT rebuild!


Note: binaries NOT rebuild!

Merged PullRequests

  • #439, #430, #369, #265, #166

Modified arm_math.h

  • reworked macros ARM_MATH_CMx
  • removed macros __FPU_USED, __DSP_PRESENT
  • used different include files

Modified source files (because of new arm_math.h)

  • FilteringFunctions\arm_biquad_cascade_df2T_f32.c
  • FilteringFunctions\arm_biquad_cascade_df2T_f64.c
  • FilteringFunctions\arm_biquad_cascade_stereo_df2T_f32.c
  • FilteringFunctions\arm_conv_partial_q15.c
  • FilteringFunctions\arm_correlate_q15.c
  • FilteringFunctions\arm_fir_f32.c
  • MatrixFunctions\arm_mat_scale_q31.c
  • StatisticsFunctions\arm_rms_q31.c
  • TransformFunctions\arm_bitreversal2.S
  • TransformFunctions\arm_cfft_radix4_q31.c Reworked DSP_Lib_TestSuite
V1.5.2 Modified arm_math.h
  • fixed GNU Compiler specific diagnostics.
  • moved __SSAT and __USAT to CMSIS-Core

Changed copyright note.

Added ARMv8M DSP libraries.


Modified arm_math.h

  • removed macro TABLE_SIZE.
  • changed preprocessor conditions to select functions instead of SIMD instruction for Cortex-M3.

Modified arm_const_structs.c

  • Added structure instances for the RFFT functions (equivalent to CFFT structures).
V1.4.10 (no source code change [still labeled 1.4.5])

Modified arm_sin_cos_f32.c

Modified arm_var_f32.c

V1.4.9 (no source code change [still labeled 1.4.5]) 2016/09/26

Modified arm_mat_mult_fast_q15.c, arm_mat_mult_fast_q31.c

Modified arm_conv_partial_q15.c

Modified arm_conv_q15.c

Modified arm_correlate_q15.c

V1.4.8 (no source code change [still labeled 1.4.5]) 2015/12/18

Modified arm_math.h, arm_bitreversal2.S

Modified arm_sin_f32.c

  • Updated Texas Instruments Arm Compiler support.

Updated documentation

V1.4.7 (no source code change [still labeled 1.4.5]) 2015/10/20

Modified arm_math.h

  • Added explicit type casts to remove compiler warnings.
  • Supressed irrelevant warnings for toolchain GCC.

Updated documentation

  • functions arm_cos_f32, arm_sin_f32 use table lookup combined with linear interpolation (since V1.4.2). This is now documented.
V1.4.6 (no source code change [still labeled 1.4.5]) 2015/08/26 Modified arm_math.h
V1.4.5 2015/03/19

Added support for the Cortex-M7 processor

Fixed bug in arm_mat_inverse_f32.c and arm_mat_inverse_f64.c. They weren't properly handling diagonal matrices.

arm_cfft_f32.c - help documentation updated

Updated documentation to show deprecated functions

V1.4.4 2014/07/31

Added the following new files:

Optimizations to the following files:

V1.4.3 2014/03/12

Undid changes to arm_biquad_cascade_df1_q31.c

Added support for COSMIC

Changed 'short' to 'q15_t' where appropriate

Fixed arm_conv_partial_fast_q15.c for UNALIGNED_SUPPORT_DISABLE

Fixed arm_mat_cmplx_mult_q15.c for UNALIGNED_SUPPORT_DISABLE

Fixed arm_conv_partial_opt_q7.c for UNALIGNED_SUPPORT_DISABLE

Restored the internal fftlen of 16 to arm_rfft_fast_init_f32.c

Updated core_xxx.h files to newer versions from ARM

V1.4.2 2013/10/16

Moved const structures from arm_const_structs.h to arm_const_structs.c

Rfft_fast_f32 no longer allows fft length of 16 as it wouldn't have worked anyways

Partial convolution was producing the wrong results in some cases

arm_lms_q31 and q15 now saturate the results in the M0 code to match the M3 & M4 code

Rfft_q15 and q31 had potential overflow issues resolved

arm_biquad_cascade_df1_q31.c had a typo which resulted in incorrect outputs

fast math sine and cosine now use linear interpolation

controller sin/cos now uses a more accurate interpolation algorithm

arm_mat_inverse was reading outside its input array

arm_cmplx_dot_prod was incorrect

switched some incorrect usages of __ssat to clip_q63_to_q31

changed var & std q31 to downshift input data by 8

var q31 & q15 no longer output larger data types

arm_mat_cmplx_mult_q15.c was done incorrectly for big vs little endian

arm_mat_mult_q31.c was inconsistent with the other multiplies, so added saturation

arm_conv_partial_q15 had an incorrect comparison between signed & unsigned values

V1.4.1 2013/02/20

Updated licenses in headers to 2013

Fixed ALIGN4 macro in arm_math.h

Added files to Cortex-M0 projects so that all projects have same file list

Fixed bugs in

V1.4.0 2013/01/09

Updated with more optimizations, bug fixes and new license information in headers

Optimized functions:

  • arm_biquad_cascade_df2T_f32
  • arm_biquad_cascade_df1_q31
  • arm_fir_f32
  • arm_fir_fast_q31
  • arm_cfft_f32
  • arm_cfft_radix2_q31
  • arm_rfft_fast_f32 (new function)

Fixed compiler warnings in arm_math.h for comparing signed and unsigned ints

Fixed a saturation bug in arm_rms_q15

Simplified the code in arm_sin_cos_q31

Added a preprocessor directive to treat the Cortex M0+ just like the Cortex M0

The following functions were deprecated and will be removed in a future version

  • arm_cfft_radix2_f32
  • arm_cfft_radix2_init_f32
  • arm_cfft_radix4_f32
  • arm_cfft_radix4_init_f32

Added CMSIS DSP Software Library

The CMSIS DSP Software Library is a suite of common signal processing functions targeted to Cortex-M processor based microcontrollers. Even though the code has been specifically optimized towards using the extended DSP instruction set of the Cortex-M4 processor, the library can be compiled for any Cortex-M processor.

For more information please see CMSIS DSP Library documentation. Added Cortex-M4 Core Support

Additional folder CM4, containing the Cortex-M4 core support files, has been added. CM0 CM3 CM4 CoreSupport DeviceSupport

New naming for Core Support Files

The new Core Support Files are:

  • core_cm#.h (# = 0, 3, 4)
  • core_cmFunc.h (Cortex-M Core Register access functions)
  • core_cmInstr.h (Cortex-M Core instructions)
  • core_cm4_simd.h (Cortex-M4 SIMD instructions)

Removed CMSIS Middelware packages

CMSIS Middleware is on hold from ARM side until a agreement between all CMSIS partners is found. SystemFrequency renamed to SystemCoreClock

The variable name SystemCoreClock is more precise than SystemFrequency because the variable holds the clock value at which the core is running. Changed startup concept

The old startup concept (calling SystemInit_ExtMemCtl from startup file and calling SystemInit from main) has the weakness that it does not work for controllers which need a already configuerd clock system to configure the external memory controller.

Changed startup concept

  • SystemInit() is called from startup file before premain.
  • SystemInit() configures the clock system and also configures an existing external memory controller.
  • SystemInit() must not use global variables.
  • SystemCoreClock is initialized with a correct predefined value.
  • Additional function void SystemCoreClockUpdate (void) is provided.
  • SystemCoreClockUpdate() updates the variable SystemCoreClock and must be called whenever the core clock is changed.
  • SystemCoreClockUpdate() evaluates the clock register settings and calculates the current core clock.

Advanced Debug Functions

ITM communication channel is only capable for OUT direction. To allow also communication for IN direction a simple concept is provided.

  • Global variable volatile int ITM_RxBuffer used for IN data.
  • Function int ITM_CheckChar (void) checks if a new character is available.
  • Function int ITM_ReceiveChar (void) retrieves the new character.

For detailed explanation see file CMSIS debug support.htm.

Core Register Bit Definitions

Files core_cm3.h and core_cm0.h contain now bit definitions for Core Registers. The name for the defines correspond with the Cortex-M Technical Reference Manual.

e.g. SysTick structure with bit definitions

typedef struct
__IO uint32_t CTRL;
__IO uint32_t LOAD;
__IO uint32_t VAL;
__I uint32_t CALIB;
} SysTick_Type;
/* SysTick Control / Status Register Definitions */
#define SysTick_CTRL_COUNTFLAG_Pos 16
#define SysTick_CTRL_COUNTFLAG_Msk (1ul << SysTick_CTRL_COUNTFLAG_Pos)
#define SysTick_CTRL_CLKSOURCE_Pos 2
#define SysTick_CTRL_CLKSOURCE_Msk (1ul << SysTick_CTRL_CLKSOURCE_Pos)
#define SysTick_CTRL_TICKINT_Pos 1
#define SysTick_CTRL_TICKINT_Msk (1ul << SysTick_CTRL_TICKINT_Pos)
#define SysTick_CTRL_ENABLE_Pos 0
#define SysTick_CTRL_ENABLE_Msk (1ul << SysTick_CTRL_ENABLE_Pos)
/* SysTick Reload Register Definitions */
#define SysTick_LOAD_RELOAD_Pos 0
#define SysTick_LOAD_RELOAD_Msk (0xFFFFFFul << SysTick_LOAD_RELOAD_Pos)
/* SysTick Current Register Definitions */
#define SysTick_VAL_CURRENT_Pos 0
#define SysTick_VAL_CURRENT_Msk (0xFFFFFFul << SysTick_VAL_CURRENT_Pos)
/* SysTick Calibration Register Definitions */
#define SysTick_CALIB_NOREF_Pos 31
#define SysTick_CALIB_NOREF_Msk (1ul << SysTick_CALIB_NOREF_Pos)
#define SysTick_CALIB_SKEW_Pos 30
#define SysTick_CALIB_SKEW_Msk (1ul << SysTick_CALIB_SKEW_Pos)
#define SysTick_CALIB_TENMS_Pos 0
#define SysTick_CALIB_TENMS_Msk (0xFFFFFFul << SysTick_VAL_CURRENT_Pos)
/* end of group CMSIS_CM3_SysTick */

DoxyGen Tags

DoxyGen tags in files core_cm3.[c,h] and core_cm0.[c,h] are reworked to create proper documentation using DoxyGen. Folder Structure

The folder structure is changed to differentiate the single support packages.

    Example (optional)
V1.1.0 2012/02/15 Updated with more optimizations, bug fixes and minor API changes.
V1.0.11 2011/10/18 Bug Fix in conv, correlation, partial convolution.
V1.0.10 2011/7/15 Big Endian support added and Merged M0 and M3/M4 Source code.
V1.0.3 2010/11/29 Re-organized the CMSIS folders and updated documentation.
V1.0.2 2010/11/11 Documentation updated.
V1.0.1 2010/10/05 Production release and review comments incorporated.
V1.0.0 2010/09/20 Production release and review comments incorporated.
V0.0.9 2010/08/27 Added files:
V0.0.7 2010/06/10 Misra-C changes done
V0.0.5 2010/04/26 incorporated review comments and updated with latest CMSIS layer
V0.0.3 2010/03/10 DP Initial version