Using gprof with Nios II

Hi Steven,

Here are the basic steps to using GPROF in Nios II:

(In the IDE):

1. Add a compiler switch, "-pg", for your src code project (in the C/C++ build area of the project properties). This is standard fare for using GPROF regardless of processor.
2. Add the same "-pg" switch for the compiler in your system library project.
3. Check the "use profiling library" checkbox in syslib properties.

(From the SDK Shell):

1. Run your program: Use nios2-terminal's "--gmon" switch which will automatically capture the profiling data into a 'gmon.out' file after your program finishes execution. This has to be done from the command line, not the IDE.

Note: nios2-terminal must be started with the processor in a paused state. You can do this by opening two SDK shell windows, downloading your program using "nios2-download " (which downloads your code and leaves the processor paused), then running "nios2-terminal

--gmon" in the second window, and finally running "nios2-download -g" in the first to start execution.

More notes: Your application must return from main() because the profiling library's data dump is triggered by reaching atexit().

Also, the above notes assume that you're using the JTAG UART for STDIO; for the conventional UART we cannot support GPROF with the above flow just yet (this is scheduled for the next release)

1. Place the "gmon.out" file (generated in step 4) into the same dir as your .elf file
2. Run nios2-elf-gprof gmon.out >
3. Examine the output file for results.

The above flow was actually easier than I had expected given my GPROF experience with other processors :) However, I am sending an enhancement request to our engineering team to make the above flow all happen from within the IDE so that the SDK shell business isn't necessary.

Also, I'll see to it that this is looked at for a relevant app note when the time comes; in addition to GPROF there is a performance counter peripheral (which is documented in the Nios II kit) which is useful for profiling sections of code -- such an app note will likely describe the use of both tools.

Jesse Kempa Altera Corp. jkempa at altera dot com

Thank you for your help.

Steven

Jesse Kempa wrote:

Jesse,

The performance counter is very usefull. I used it already in Nios1, with a custom peripheral just counting clock cycles. The point in my design was that I use a 64 bit counter. So it is expected to never roll over. The counter part of this is that you have to do two read actions to it to get the time.

If you then add some classes, you can define a kind of PseudoTimer class. Each instance keeps track of when it is started and stopped. Finde code below.

Hope this can be of any use for someone

Stefaan.

Verilog code for counter :

module never_ending_timer(Clk, Reset, Out, A, nRD, CS);

input Clk, Reset; input A, nRD, CS; output [31:0] Out;

parameter bits = 64;

reg [bits-1:0] cnt; reg [bits - 32 - 1 : 0] temp;

always @(posedge Clk or posedge Reset) if (Reset) begin cnt Ticks(); delta = t - start; start = t; if (delta > 0xFFFFFFFFFFFFFFFFL) return 0xFFFFFFFF; Running = 0; return (DWORD) delta; };

DWORD PeekDWORD() { //peak doesn't affect timer operation if (Running) { QWORD t = hw->Ticks(); delta = t - start; } if (delta > 0xFFFFFFFFFFFFFFFFL) return 0xFFFFFFFF; return (DWORD) delta; };

QWORD Peek() { //peak doesn't affect timer operation if (Running) { QWORD t = hw->Ticks(); delta = t - start; } return delta; };

QWORD Stop() { //you can use stop as a brand new starting point QWORD t = hw->Ticks(); delta = t - start; start = t; Running = 0; return delta; }; };