TMS320F28384D: CLA FP32 precision issues

Ahn JaeKyun

Part Number: TMS320F28384D

Hello TI experts! I am using TMS320F28384D.

I conducted a simple test to use CPU 1 and 2 and each CLA at the same time. The results are as follows:

Single core time=74ms
Single core res =1967.704102

multi core time=18ms
core cpu1 cpu res =502.574646
core cpu1 cla res =493.185547
core cpu2 cpu res =502.574646
core cpu2 cla res =493.185547
multi core res =1991.520386

The code performs 2000000 FP32 additions. The addition unit is 0.001.

The results of processing by CPU1 alone and the results of shared processing by four processing units are expressed as single and multi, respectively.
There appears to be a difference between the CPU and CLA, so it seems necessary to check. Please feel free to ask for any information you need prior to giving advice. thank you

10 months ago

0 Delaney Woodward 10 months ago

TI__Genius 13265 points

Hello,

What exactly is the issue or which result do you have questions about? Are you seeing a precision issue in your application? If so, please briefly explain your application and what the expected vs. actual results are and I will try to help.

Best Regards,

Delaney

0 Ahn JaeKyun 10 months ago in reply to Delaney Woodward

Prodigy 10 points

Hello Delaney Woodward.

The source code is as follows.

...
for( i = 0 ; i < 500000 ; i++) {
fp32_data = fp32_data + 0.001f;
}
...

The calculation results of the two cores are as follows:

CPU : 502.574646
CLA : 493.185547

The CPU's results output the same value as a regular PC, but a problem occurred with the CLA's results, which were different. I wonder if this result is normal or if it is a problem caused by a setting error.

+1 Delaney Woodward 10 months ago in reply to Ahn JaeKyun

TI__Genius 13265 points

Hi Ahn,

I see, I believe this is a result of the CLA including only a subset of the FPU32 instruction set. Can you view the disassembly window in CCS to compare the assembly instructions being used on each core? I suspect one of more of the operations being done on the CLA are causing the loss of precision.

Best Regards,

Delaney

0 Ahn JaeKyun 10 months ago in reply to Delaney Woodward

Prodigy 10 points

Hello Delaney Woodward.

Thank you for your attention.

First of all, another problem occurred, and I think we need to resolve this and proceed. The CLA only works with the debugger connected, and when run without the debugger, there is no response, so I am inferring that it is a configuration issue. We will share information later when it is resolved.

thank you

+1 Delaney Woodward 9 months ago in reply to Ahn JaeKyun

TI__Genius 13265 points

Hi Ahn,

Ahn JaeKyun said:
when run without the debugger, there is no response

Do you mean if you load the program directly without connecting the debugger? How are you verifying that the CLA is not connected (toggling GPIOs?)?

Best Regards,

Delaney

0 Ahn JaeKyun 9 months ago in reply to Delaney Woodward

Prodigy 10 points

This problem was solved by modifying cmd. When the MCU boots from flash, the CLA must also specify flash, but it did not work because the program was read from RAM.

0 Delaney Woodward 9 months ago in reply to Ahn JaeKyun

TI__Genius 13265 points

Hi Ahn,

Glad to hear you were able to solve the issue.

Best Regards,

Delaney

C2000™︎ microcontrollers

C2000 microcontrollers forum

TMS320F28384D: CLA FP32 precision issues