topic Re:big difference of kurt between float and double in Intel® oneAPI Math Kernel Library

big difference of kurt between float and double

rolandsun — Fri, 04 Mar 2022 07:50:31 GMT

I test computing kurtosis by MKL with float and double , and get strange result as following , why kurt of float and double have such big difference ?

double skew -3.2811 ,kurt 9.3626 ,mean 32.29 ,r2m 1042.57,r3m 33668.46 ,r4m 1087393.04
float skew -3.3029 ,kurt 5.7297 ,mean 32.29 ,r2m 1042.57,r3m 33668.46 ,r4m 1087393.00

#define DIM 1 #define N 16 void test_mkl_skew_double() { VSLSSTaskPtr task; double x[N]= {32.368824, 32.3732567, 32.3706436, 32.3741646, 32.3736801, 32.3755112, 32.3742905, 32.3735313, 32.3738403, 30.9899216, 32.3761978, 32.3737259, 32.3755112, 32.3710899, 32.3742638, 32.3742638}; double mean=0,skew=0,kurt={0},vari=0; MKL_INT p,n,xst; int status; p=DIM; n=N; xst=VSL_SS_MATRIX_STORAGE_ROWS; status = vsldSSNewTask( &task, &p, &n, &xst, x, 0, 0 ); double rm[8]={-1}; status=vsldSSEditMoments(task,&rm[0],&rm[1],&rm[2],&rm[3],&rm[4],&rm[6],&rm[5]); status = vsldSSEditTask( task, VSL_SS_ED_SKEWNESS, &skew ); status = vsldSSEditTask( task, VSL_SS_ED_KURTOSIS, &kurt ); MKL_UINT64 mask = VSL_SS_MEAN | VSL_SS_KURTOSIS | VSL_SS_SKEWNESS | VSL_SS_2R_MOM | VSL_SS_3R_MOM | VSL_SS_4R_MOM | VSL_SS_2C_MOM | VSL_SS_3C_MOM | VSL_SS_4C_MOM ; status = vsldSSCompute(task, mask, VSL_SS_METHOD_FAST ); status = vslSSDeleteTask( &task ); printf ("double skew %.4f ,kurt %.4f ,mean %.2f ,r2m %.2f,r3m %.2f ,r4m %.2f\n",skew,kurt,rm[0],rm[1],rm[2],rm[3]); } void test_mkl_skew_float() { VSLSSTaskPtr task; float x[N]= {32.368824, 32.3732567, 32.3706436, 32.3741646, 32.3736801, 32.3755112, 32.3742905, 32.3735313, 32.3738403, 30.9899216, 32.3761978, 32.3737259, 32.3755112, 32.3710899, 32.3742638, 32.3742638}; float mean=0,skew=0,kurt={0},vari=0; MKL_INT p,n,xst; int status; p=DIM; n=N; xst=VSL_SS_MATRIX_STORAGE_ROWS; status = vslsSSNewTask( &task, &p, &n, &xst, x, 0, 0 ); float rm[8]={-1}; status=vslsSSEditMoments(task,&rm[0],&rm[1],&rm[2],&rm[3],&rm[4],&rm[6],&rm[5]); status = vslsSSEditTask( task, VSL_SS_ED_SKEWNESS, &skew ); status = vslsSSEditTask( task, VSL_SS_ED_KURTOSIS, &kurt ); MKL_UINT64 mask = VSL_SS_MEAN | VSL_SS_KURTOSIS | VSL_SS_SKEWNESS | VSL_SS_2R_MOM | VSL_SS_3R_MOM | VSL_SS_4R_MOM | VSL_SS_2C_MOM | VSL_SS_3C_MOM | VSL_SS_4C_MOM ; status = vslsSSCompute(task, mask, VSL_SS_METHOD_FAST ); status = vslSSDeleteTask( &task ); printf ("float skew %.4f ,kurt %.4f ,mean %.2f ,r2m %.2f,r3m %.2f ,r4m %.2f\n",skew,kurt,rm[0],rm[1],rm[2],rm[3]); }

Re: big difference of kurt between float and double

VidyalathaB_Intel — Fri, 04 Mar 2022 11:37:56 GMT

Hi Roland,

The issue raised by you is reproducible with VSL_SS_METHOD_FAST method parameter.

Meanwhile could you please try changing the method parameter to the function vslSSCompute as shown below?

//kurtosis of double datatype status = vsldSSCompute(task, mask, VSL_SS_METHOD_1PASS); //kurtosis of float datatype status = vslsSSCompute(task, mask, VSL_SS_METHOD_1PASS );

We checked it from our end and could see that the results are same for both float and double kurtosis.

Reference Link:https://www.intel.com/content/www/us/en/develop/documentation/onemkl-developer-reference-c/top/statistical-functions/summary-statistics/summary-statistics-task-computation-routines.html#summary-statistics-task-computation-routines_TBL10-29

Please refer to the below screenshot and do let us know if it helps in resolving the issue.

Regards,

Vidya.

Re: big difference of kurt between float and double

rolandsun — Mon, 07 Mar 2022 00:59:26 GMT

It works ,thankyou very much.

Re:big difference of kurt between float and double

Ruqiu_C_Intel — Mon, 25 Apr 2022 08:54:06 GMT

Thank you for reaching us. This issue is closing and we will no longer respond to this thread. If you require additional assistance from Intel, please start a new thread. Any further interaction in this thread will be considered community only.