Slower floating point instructions after converting project from CCS3.3 to CCS6

Hi Vishal,

Thank you for the tips, I'll try that and report back.

Regarding the math routines I'm running them in the RAM.

Best regards, Jorge Lopes

Do you have interrupts runnings? The way i benchmark these functions is to turn on the clock feature in CCS (i think you will find it under the Run tab in the debug perspective). The clock will show up as a little clock icon in the bottom right corner of the CCS window. Then i run code all the way to the actual calling of the functions....so if i were becnhmarking the divide function i would keep stepping code until this point

   LCR #_FS$$DIV

or something like that - the syntax may not be exactly correct. Once im at the call instruction, i reset the clock by double clicking on the clock icon, then step over the LCR instruction and check what the cycles are. Can you try this and see if you get the same cycles mentioned in the user's guide - if you have interrupts running disable them and try this again to make sure that the interrupts arent responsible for the slowing down.

Hello Lori,

I have updated the compiler to the version 6.4.6 and it didn't work yet.

I've tried with diferent versions of the compiler with similar results.

Best regards, Jorge Lopes

Hello Vishal, how are you?

No, I haven't solved my problem.

I've made some extra tests to get more information about the situation.

I made a simple routine with math routines runing in an interruption and compiled this in CCS v3.3 and in the CCS v6.1:

sqrt(sin(2.3f));
sqrt(sin(3.4f));
sqrt(sin(5.6f));
sqrt(sin(6.7f));
sqrt(sin(7.8f));
sqrt(sin(8.9f));
sqrt(sin(10.11f));
sqrt(asin(2.3f));
sqrt(atan(3.4f));
sqrt(cos(5.6f));
sqrt(6.7f/45.6f);
sqrt(acos(7.8f));
sqrt(atan(8.9f));
sqrt(exp(10.11f));

I've tested the execution time for this routine and got the following results:

Execution time in CCS v3.3 = 13 uSeconds

Execution time in CCS v6.1 = 30 uSeconds

I have the following configuration on the CCS v3.3 project:

Compiler: v5.0.1

Compiler options:

-pm -pdsw225 -pden -o3 -fr"$(Proj_dir)\DSP\2833x_FLASH" -i"$(Proj_dir)\DSP\include" -i"$(Proj_dir)\DSP\librarias\lib\PowerLIB" -i"$(Proj_dir)\SOURCE\include" -d"_DEBUG" -d"FLASH" -mf -v28 --float_support=fpu32

Linker options:

-b -c -m".\DSP\2833x_FLASH\Conversor.map" -o".\DSP\2833x_FLASH\Conversor.out" -stack0x400 -w -i"$(Proj_dir)\DSP\lib" -priority

Linker order:

rts2800_fpu32_fast_supplement.lib

F28335_FLASH_Conversor.cmd

I have the following configuration on the CCS v6.1 project:

Compiler: v6.4.6

Compiler options:

-v28 --float_support=fpu32 --vcu_support=vcu0 -O2 --opt_for_speed=2 --fp_reassoc=off --fp_mode=relaxed --include_path="C:/ti/ccsv6/tools/compiler/ti-cgt-c2000_6.4.6/include" --include_path="C:/SVN/dsp/branches/_CC6/DSP/include" --include_path="C:/SVN/dsp/branches/_CC6/SOURCE/include" -g --float_operations_allowed=all --define="_DEBUG" --define="FLASH" --diag_warning=225 --display_error_number --issue_remarks --diag_wrap=on --c_src_interlist --obj_directory="C:/SVN/dsp/branches/_CC6/_DSP/F2833x_FLASH"

Linker options:

-v28 --float_support=fpu32 --vcu_support=vcu0 -O2 --opt_for_speed=2 --fp_reassoc=off --fp_mode=relaxed -g --float_operations_allowed=all --define="_DEBUG" --define="FLASH" --diag_warning=225 --display_error_number --issue_remarks --diag_wrap=on --c_src_interlist --obj_directory="C:/SVN/dsp/branches/_CC6/_DSP/F2833x_FLASH" -z -m"C:/SVN/dsp/branches/_CC6/DSP/2833x_FLASH/Conversor.map" --stack_size=0x400 --warn_sections -i"C:/ti/ccsv6/tools/compiler/ti-cgt-c2000_6.4.6/DSP/lib" -i"C:/ti/ccsv6/tools/compiler/ti-cgt-c2000_6.4.6/lib" -i"C:/ti/ccsv6/tools/compiler/ti-cgt-c2000_6.4.6/include" -i"C:/SVN/dsp/branches/_CC6/_DSP" -i"C:/SVN/dsp/branches/_CC6" --priority --reread_libs --display_error_number --xml_link_info="_DSP_linkInfo.xml" --no_sym_merge --rom_model

Linker order:

rts2800_fpu32_fast_supplement.lib

rts2800_fpu32.lib

IQmath_fpu32.lib

Hi Jorge,

This is probably the differences in the compiler that is causing the code to bloat. To be sure, you need to benchmark the actual square root function itself and not the ISR. So in my previous post i suggested using the CCS clock to do this. You would enable the clock (Run-> Clock) and then in the disassembly window you run to the point where it calls the square root function

LCR #_sqrt

Then you would do two things:

1. make sure its actually calling the fastRTS library

2. double click the clock icon (bottom right corner of CCS) to reset it then step-over this line (F6) and see how many cycles it takes, and that it matches what the fastRTS user guide says. That way we eliminate the issues with the math routines

If the execution time for a single sqrt matches the user guide numbers then its probably the newer compiler that is putting in more code than before. 

Jorge,

Also did you check the .map file?  Can you post it for us to review?

Thank you

Lori

Yes I can:

For the compiler 6.4.6:

For the v5.0.1 compiler: