TMS320F280038C-Q1: CLA LUT Array access time very slow

Jason Galtieri1

Part Number: TMS320F280038C-Q1
Other Parts Discussed in Thread: C2000WARE

Hello,

Still looking for help on this topic. Previous thread was locked due to no response.

Copied question from previous thread

I am having an issue where I would like to read values from a lookup table placed in LSRAM which is handed over to the CLA. However, I am finding that access times are very slow with it taking hundreds of clock cycles to do a single read from the table.

I have modified the "cla_ex4_pwm_control" example and am seeing the same behavior as in my project. The program reads from a constant float array placed in LS5-7 memory and accesses it from the CLA. I am timing the access with PWM3 which is operating at 120Mhz. On my setup I am registering ~600 clock counts (minus whatever PWM3 access overhead) to read one float value from the array. This is observed in "timer2" in the expression windows. This seems pretty high, so I am guessing I made a mistake in my setup. Any help is appreciated.
28003xcla_ex4__pwm_LUT.zip
https://e2e.ti.com/support/microcontrollers/c2000-microcontrollers-group/c2000/f/c2000-microcontrollers-forum/1578736/tms320f280038c-q1-cla-lut-array-access-time-very-slow

6 months ago

0 Lori Heustess 6 months ago

TI__Guru* 93800 points

Hi Jason - it doesn't look like the compiler's optimization is enabled. Can you try enabling and see how it changes the performance? I would suggest -o2 but you can also see what impact -o0 and -o1 have.

Regards

Lori

0 Jason Galtieri1 6 months ago in reply to Lori Heustess

Prodigy 175 points

Seeing following computation times with various optimizations levels (as measured with timer2 variable)

Off - 617 counts

o0 - 169 counts

o1 - 172

o2 - 165

optimization definitely helped but those counts still seem way too high?

Previous linked project wasn't importing correctly for this, but this one did.

cla_ex4_pwm_control_lut_ex.zip

0 Lori Heustess 6 months ago in reply to Jason Galtieri1

TI__Guru* 93800 points

I'll take a closer look. I suspect the offset/address calculation is taking up the majority of the cycles and not the access itself. Is it possible to use a 1D array instead? I think it will be faster.

+1 Lori Heustess 6 months ago in reply to Lori Heustess

TI__Guru* 93800 points

After reviewing the code and assembly - it is the address generation that is taking the cycles up. The code has to figure out where [x][y] is in memory before it can be accessed. I spoke to our compiler team about this and it is difficult on the CLA to get this to loop flatten (loops are involved which has a bit of overhead for the CLA). I think you could try to play some tricks where the address is manually calculated in the application and a pointer is used to grab the value. It wouldn't be a very generic solution, though.

A 1D array should be much faster if that is a possibility for your application.

0 Jason Galtieri1 5 months ago in reply to Lori Heustess

Prodigy 175 points

Indexing seems like it was the issue, but is it just the integer multiply (that the CLA lacks) causing the issue here?

I am getting similar results (ie ~160 counts) for a 2d array and a 1d flattened array with index calculation. However, if I change the index calculation to floats and then cast the final result as an integer I get <10 counts

// ~160 counts
duty = LUT_D1[x][y];

// ~160 counts
duty = LUT_D1_1d[100*x + y];

// <10 counts
z = (uint32_t)(100.0f *x_float + y_float);

duty = LUT_D1_1d[z];

0 Lori Heustess 5 months ago in reply to Jason Galtieri1

TI__Guru* 93800 points

I think it might be a combo of both but from your experiment it does look like the float calculations solved the issue. Thank you for the feedback.

0 Lori Heustess 5 months ago in reply to Jason Galtieri1

TI__Guru* 93800 points

Jason,

Could you share a .pp file for each test case? We would like to take a closer look at what is going on. The instructions are here:

https://software-dl.ti.com/ccs/esd/documents/sdto_cgt_How-to-Submit-a-Compiler-Test-Case.html

0 Jason Galtieri1 5 months ago in reply to Lori Heustess

Prodigy 175 points

Here they are

// Test Case 1
// duty = LUT_D1[x][y];

// Test Case 2
// z = y + 100*x;

//Test Case 3
z = (uint32_t)(100.0f*x_f+y_f);

// Duty for Test Case 2 and 3
duty = array_start[z];

timer2 = EPWM_getTimeBaseCounterValue(EPWM3_BASE)-timer1;

Compiler Version

TMS320C2000 Linker PC v22.6.1

Compiler Options (Replaced my filepath with MyPath)

-v28 -ml -mt --cla_support=cla2 --float_support=fpu32 --tmu_support=tmu0 --vcu_support=vcrc -O2 --include_path="MyPath/cla_ex4_pwm_control_lut_ex/cla_ex4_pwm_control" --include_path="C:/ti/c2000/C2000Ware_5_01_00_00" --include_path="MyPath/cla_ex4_pwm_control_lut_ex/cla_ex4_pwm_control/device" --include_path="C:/ti/c2000/C2000Ware_5_01_00_00/driverlib/f28003x/driverlib" --include_path="C:/ti/ccs1260/ccs/tools/compiler/ti-cgt-c2000_22.6.1.LTS/include" --define=_FLASH --preproc_with_comment --preproc_with_compile --diag_suppress=10063 --diag_warning=225 --diag_wrap=off --display_error_number --gen_func_subsections=on --abi=eabi --cla_background_task=off --cla_signed_compare_workaround=on --include_path="MyPath/cla_ex4_pwm_control_lut_ex/cla_ex4_pwm_control/CPU1_FLASH/syscfg"

TestCase3_cla_ex4_pwm_control.pp.txt TestCase3_cla_ex4_pwm_control_cla.pp.txt TestCase1_cla_ex4_pwm_control.pp.txt TestCase1_cla_ex4_pwm_control_cla.pp.txt TestCase2_cla_ex4_pwm_control.pp.txt TestCase2_cla_ex4_pwm_control_cla.pp.txt

0 Lori Heustess 5 months ago in reply to Jason Galtieri1

TI__Guru* 93800 points

Thank you.

C2000™︎ microcontrollers

C2000 microcontrollers forum

TMS320F280038C-Q1: CLA LUT Array access time very slow