TMS320F28P650DH: Questions on EABI VS COFF

Hello,

I brought this thread to the attention of the compiler experts for further assistance.

Thanks

ki

Hi George,

Thanks for the timely response! 

George Mock said:

This answer presumes you execute on a C28x CPU that has an instruction that can perform the floating point operation specified in the C code.

Do you mean C28x CPU that has FPU64? They are using F28P65, so I think the CPU would have the instruction for basic math （they may refer to instruction set for specific instructions) correct?

If they use float as parameter of a function that has a double parameter type. For example:

double add(double A, double B){return A+B;}
flaot x = 1.0;
x = add(x,x);

There would be type conversion (by means of calling FS$$TOFD and FD$$TOFS )when calling the function and returning, which takes extra time, correct? If x is a double in the above code, it would be faster?

Additionally, is there a document on #3?

Hang Yang said:

3. Is there any document on FPU32 vs FPU64?

Thanks,

Hang.

Hang Yang said:

Do you mean C28x CPU that has FPU64?

Yes, for the case where the C code specifies 64-bit floating point operations.  One example would be operations of type double and EABI is used.

Hang Yang said:

They are using F28P65, so I think the CPU would have the instruction for basic math （they may refer to instruction set for specific instructions) correct?

Compiler experts like me do not keep track of what hardware features are available on a specific devices.  I took a quick look at the datasheet, and it appears you are correct.

Hang Yang said:

There would be type conversion

Yes.

Hang Yang said:

(by means of calling FS$$TOFD and FD$$TOFS )

No.  There are instructions for converting between 64-bit and 32-bit floating point representations.

Hang Yang said:

If x is a double in the above code, it would be faster?

Yes.  But only by a few cycles.

I'm sorry I overlooked this question ...

Hang Yang said:

Is there any document on FPU32 vs FPU64?

The compiler team never published anything.  And I've never seen anything myself.  

Thanks and regards,

-George

Hi George,

Although you said

George Mock said:

No.  There are instructions for converting between 64-bit and 32-bit floating point representations

George Mock said:

Yes.  But only by a few cycles.

I had example code from customer that the type conversion is done by means of calling functions and take lots cycles and result in a large performance difference. Could you help explain the difference here? 

The code is:

#define MAINS_SAMPLING_FREQ_HZ    19200.0L

#define DCBUS_MAINS_SAMPLING_STEP_US  (F32)(1000000.0L/MAINS_SAMPLING_FREQ_HZ);
// vs
//#define DCBUS_MAINS_SAMPLING_STEP_US  (1000000.0L/MAINS_SAMPLING_FREQ_HZ); // without (F32)

float x;
x += DCBUS_MAINS_SAMPLING_STEP_US

They are comparing the macro with or without the (F32) conversion statement.

When with the F32 conversion, they got the assembly:

Without the F32 conversion statement, they got the assembly:

Not only it's longer, it also contains LCR that takes even more cycles.

Please prepare a test case that uses both variants of the macro DCBUS_MAINS_SAMPLING_STEP_US in a manner similar to how it appears in the actual application.  One variant uses the cast to F32, the other does not.  This test case only needs to be a single C file that compiles into one object file.  It does not need to run.  Build it with the compiler option --src_interlist.  This causes the compiler to keep the auto-generated assembly file.  This file has the same name as the source file, with the extension changed to .asm.  Inspect this assembly file to be sure it generates the code you show in the screenshots from the last post.  For this test case, please submit all the details described in the article How to Submit a Compiler Test Case.

Thanks and regards,

-George