Execution time issue SARAM/XINTF

Hi Josip,

If the execution code is accessing some data via CPU data bus which is shared with XINTF access then it'll definitely impact the execution performance. It depends how the these access mixed in the code. If you could provide the sequence of operation, it'll be helpful.

Regards,

Vivek Singh 

Hi Vivek,

thank you for the reply. I've experimented with measurement of execution time of a piece of code that checks if two longword locations are equal. For this, I used ePWM6 timer. It counts up at SYSCLK, so a 1 means one CPU cycle. All directly addressed locations are in internal L7 SARAM. Firstly, here is the code with mixed XINTF SRAM and internal L7 SARAM accesses:

; get ePWM6 timer value an store it to XINTF SRAM
MOVL XAR0,#0x006944
MOV ACC,*
MOVL XAR0,#0x2000DA
MOVL *,ACC

; check if two locations are equal (done on L7 SARAM)
MOVL ACC,@@120
SUBL ACC,@@118
ABS ACC
SUBB ACC,#1
MOVL @@116,ACC ; -1 if equal, >= 0 if not equal

; get ePWM6 timer value an store it to XINTF SRAM
MOVL XAR0,#0x006944
MOV ACC,*
MOVL XAR0,#0x2000DC
MOVL *,ACC

; calculate execution time by subtracting two ePWM6 timer values
MOVL XAR0,#0x2000DC
MOVL ACC,*
MOVL XAR0,#0x2000DA
SUBL ACC,*

The result in ACC is 25.

If I use internal L7 SARAM exclusively, the code looks like:

; get ePWM6 timer value an store it to L7 SARAM
MOVL XAR0,#0x006944
MOV ACC,*
MOVL @@116,ACC

; check if two locations are equal (done on L7 SARAM)
MOVL ACC,@@124
SUBL ACC,@@122
ABS ACC
SUBB ACC,#1
MOVL @@118,ACC

; get ePWM6 timer value an store it to L7 SARAM
MOVL XAR0,#0x006944
MOV ACC,*
MOVL @@120,ACC

; calculate execution time by subtracting two ePWM6 timer values
MOVL ACC,@@120
SUBL ACC,@@116

The result in ACC is 10. Clearly, one cycle is spared because storing of the 1st ePWM6 timer value is done directly, but this cannot account for the huge difference in measured execution times.

Regards,
Josip

Hi Josip,

In code where you are storing the timer values on exteranl RAMs, there are two READ operations to external RAM which will stall the pipeline and will add to the execution time. If you substract the two memory READ ACCESS time from the execution, I think you'll get the same number.

Regards,

Vivek Singh

Hi Vivek,

I do not understand your answer. You say there are two READ operations that cause this, but to my understanding there is only one WRITE operation that can influence the measurement: the writing of first timer value to external RAM prior to executing the measured code (line 05). The only reads from external RAM are when subtracting the two timer values at the end of the code section (lines 22 and 24), but these cannot influence the final result. To be clear, I copied the 1st code segment from my previous post and added line numbers that i referenced in brackets.

01   ; get ePWM6 timer value an store it to XINTF SRAM
02   MOVL XAR0,#0x006944
03   MOV ACC,*
04   MOVL XAR0,#0x2000DA
05   MOVL *,ACC
06
07   ; check if two locations are equal (done on L7 SARAM)
08   MOVL ACC,@@120
09   SUBL ACC,@@118
10   ABS ACC
11   SUBB ACC,#1
12   MOVL @@116,ACC ; -1 if equal, >= 0 if not equal
13
14   ; get ePWM6 timer value an store it to XINTF SRAM
15   MOVL XAR0,#0x006944
16   MOV ACC,*
17   MOVL XAR0,#0x2000DC
18   MOVL *,ACC
19
20   ; calculate execution time by subtracting two ePWM6 timer values
21   MOVL XAR0,#0x2000DC
22   MOVL ACC,*
23   MOVL XAR0,#0x2000DA
24   SUBL ACC,

Regards,
Josip