How to Test the speed of floating point process on OMAPL138 DSP?

PengGuowen

Other Parts Discussed in Thread: OMAP-L138

As title.

I want to test the dsp floating point , Does anyone have program to do this work?

I have modified the sample program of dsplink (message), And add a lot of floating-point processing on dsp side program (message.out), But i found a very slow processing speed of DSP, what 's worng?

over 15 years ago

0 tlee over 15 years ago

TI__Guru 62975 points

Here was a similar discussion on floating point performance: http://e2e.ti.com/support/dsp/tms320c6000_high_performance_dsps/f/115/p/12402/48504.aspx#48504

0 Mariana over 15 years ago

TI__Mastermind 24340 points

Another thing that you can notice is that the DSPLink examples might come with the "wrong" mv option:

http://processors.wiki.ti.com/index.php/-mv_option_to_use_with_the_C674x

Please edit the file:

.../user/OMAP-L138_arm_1_00_00_11/dsplink_linux_1_65_00_02/dsplink/make/DspBios/c64xxp_5.xx_linux.mk

aroung line 68, replace:

STD_CC_FLAGS := -q -pdr -pdv -pden -ml3 -mv6400+ --disable:sploop

with:

STD_CC_FLAGS := -q -pdr -pdv -pden -ml3 -mv6740

And rebuild DSPLink and the samples:

http://processors.wiki.ti.com/index.php/Building_The_OMAP-L1_SDK

0 PengGuowen over 15 years ago in reply to Mariana

Prodigy 230 points

According to your reply, i modified the file, and rebuild , but the results did not change significantly, what happend ?

0 PengGuowen over 15 years ago in reply to PengGuowen

Prodigy 230 points

I just add the following code at the beginning of "TSKMESSAGE_execute" function ( message sample, dsp side, tskMessage.c )

double *a;

double *b;

double *c;

Int j;

double tmp;

a = (double *) calloc(1000, sizeof(double));

b = (double *) calloc(1000, sizeof(double));

c = (double *) calloc(1000, sizeof(double));

for (j = 0; j < 1000; j++) {

tmp = rand() % 1000;

a[j] = tmp + (tmp/(10000));

b[j] = a[j] + 1.0000;

}

for (i = 0; i < 100000; i++) {

for (j = 0; j < 1000; j++) {

c[j] = a[j] + b[j];

}

for (j = 0; j < 1000; j++) {

c[j] = a[j] - b[j];

}

for (j = 0; j< 1000; j++) {

c[j] = a[j] * b[j];

}

for (j = 0; j< 1000; j++) {

c[j] = a[j] / b[j];

}

free(a);

free(b);

free(c);

and use the following command:

./messagegpp ./message.out 1 0

But it took 300 seconds to execute. that is too long

0 PengGuowen over 15 years ago in reply to PengGuowen

Prodigy 230 points

Why?

0 Deepali Uppal over 15 years ago in reply to PengGuowen

TI__Expert 4930 points

Peng,

300 seconds does seem like a very long time. We have profiled the round trip time for the message sample with just MSGQ API's (no floating point ops) and it comes around 300 usec.

Deepali

Processors

Processors forum

How to Test the speed of floating point process on OMAPL138 DSP?