AM5718: OpenCL profiling support/tools for C66x DSP

Hi Ran,

Thanks for your response. I am using OpenCL events to benchmark the overall execution time for the program, similar to the Monte-Carlo simulation example. So I have coarse timing for the kernel execution itself, as well as fine grained timing for the IPC read/write buffers. But I am also executing a large amount of C library code (using the TI extension: Calling Standard C Code From OpenCL C Code: downloads.ti.com/mctools/esd/docs/opencl/extensions/standard-c-code.html) outside of the OpenCL kernel. It is this library code that I would like to profile.

Thanks,

Weston

Hi Ran,

The library I am calling into is a custom one. The task that the DSP is performing is an image processing operation. I enqueue an image in to a CL write buffer and a number of parameters to control the image processing into another write buffer, then perform the image processing task on the DSP (using library code), then return the processing results via read buffer. I'd like to carry along execution time-stamps for the various steps in the image processing sequence in the return read buffer as well.

Weston

Ran,

Thank you very much for your help.  I should be close to a solution with the above pseudo code. Using the document: TMS320C6000 Optimizing Compiler v8.1.x User's Guide, section: 7.5.3 The __cregister Keyword, and document: TMS320C66x DSP CPU and Instruction Set Reference Guide, section: 2.9.13 Time Stamp Counter Registers (TSCL and TSCH),  I have implemented the following:

#include "c6x.h"

// Local t0 for start time of processing.
unsigned long long int _startTime = 0;

// Control register access provided by __cregister keyword, see: c6x.h.
extern __cregister volatile unsigned int TSCL;
extern __cregister volatile unsigned int TSCH;

void enable_cpu_clock_counter(void) {
    TSCL = 0;
}

unsigned long long int get_timestamp(void)
{
    return ((unsigned long long int)TSCH << 32) + (unsigned long long int)TSCL;
}

void set_timestamp(unsigned long long int timestamp)
{
    _startTime = timestamp;
}

unsigned int uptime(void)
{
    return (unsigned int)(get_timestamp() - _startTime);
}

Which generates the following assembly code for enable_cpu_clock_counter:

;******************************************************************************
;* FUNCTION NAME: enable_cpu_clock_counter()                                  *
;*                                                                            *
;*   Regs Modified     : B4                                                   *
;*   Regs Used         : B3,B4                                                *
;*   Local Frame Size  : 0 Args + 0 Auto + 0 Save = 0 byte                    *
;******************************************************************************
_Z24enable_cpu_clock_counterv:
;** --------------------------------------------------------------------------*
;          EXCLUSIVE CPU CYCLES: 8
           ZERO    .L2     B4                ; [B_L66] |25| 
           MVC     .S2     B4,TSCL           ; [B_Sb66] |25| 
           RETNOP          B3,5              ; [] |26| 
           ; BRANCH OCCURS {B3}              ; [] |26| 
	.sect	".text"
	.clink
	.global	_Z13set_timestampy

This looks like a correct way to start the cpu clock counter per the documentation. Instead of converting clock ticks to an actual time value, I am simply returning the delta clock ticks. 

This appears to work.

Thanks,

Weston