Tutorial: Debugging with Intel® Distribution for GDB*

Basic Debugging

Consider the array-transform.cpp example again:

54        h.parallel_for(data_range, [=](id<1> index) {
55            size_t id0 = GetDim(index, 0);
56            int element = in[index]; // breakpoint-here
57            int result = element + 50;
58            if (id0 % 2 == 0) {
59                result = result + 50; // then-branch
60            } else {
61                result = -1; // else-branch
62            }
63            out[index] = result;
64        });

Start the debugger, set two breakpoints inside the kernel (one for each conditional branch) as follows:

  1. break 59

    Expected output:

    Breakpoint 1 at 0x40583c: file /path/to/array-transform.cpp, line 59.
  2. break 61

    Expected output:

    Breakpoint 2 at 0x40584a: file /path/to/array-transform.cpp, line 61.

Note

  • Do not expect the output you receive will match exactly the one provided in the tutorial. The output may vary due to the nature of parallelism and different machine properties. The ellipsis [...] denotes output omitted for brevity
  • For your convenience, common Intel® Distribution for GDB* commands are provided in the reference sheet.

To start the program, execute:

run gpu

You should see the following output:

Starting program: /path/to/array-transform gpu
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
[New Thread 0x7ffff37dc700 (LWP 9479)]
intelgt: gdbserver-gt started for process 28837.
Will listen for an attached process
[New Thread 0x7fffe21e9700 (LWP 9599)]
[SYCL] Using device: [Intel® Iris® Plus Graphics 650 [0x5927]] from [Intel® Level-Zero]
intelgt: attached to device 1 of 1; id 0x5927 (Gen9)
[New inferior 2]
[New Thread 1.1073741824]
[New Thread 1.1073741888]
[New Thread 1.1073742080]
[New Thread 1.1073742144]
[New Thread 1.1073742336]
[New Thread 1.1073745920]
[New Thread 1.1073746176]
[New Thread 1.1073746432]
[Switching to Thread 1.1073741824 lane 1]
Thread 2.1 hit Breakpoint 2, with SIMD lanes [1 3 5 7], main::$_1::operator()[...]
  at array-transform.cpp:61
61        result = -1;          // else-branch

The auto-attach mechanism is triggered and gdbserver-gt begins to listen to GPU events.

Check the presence of gdbserver-gt as follows:

info inferiors
Expected output:
 Num  Description    Connection                            Executable
  1   process 9463   1 (native)                            <path_to_array-transform>
* 2   device 1      2 (extended-remote gdbserver-gt --multi --hostpid=9463 -)

Note

The auto-attach feature sets schedule-multiple to on.

The breakpoint event is received from the gdbserver-gt process. The thread ID 2.1:1 points to the thread 1 of the inferior 2 and indicates that the first active SIMD lane is now in focus.

The breakpoint at line 61 is hit first (the order of branch execution is defined by the Intel® Graphics Compiler).

Check which SIMD lanes are currently active with the following command:

info threads

In the example, thread 2.1 has 4 active SIMD lanes: 1, 3, 5, and 7. The asterisk * marks the current SIMD lane. See the expected output below.

Note

SIMD lane enumeration starts from 0.
  Id             Target Id               Frame
  1.1            Thread [...]         [...]
  1.2            Thread [...]         [...]
  2.1:1          Thread 1.1073741824	    <frame> at array-transform.cpp:61  
  2.1:[3 5 7]    Thread 1.1073741824	    <frame> at array-transform.cpp:61  
  2.2:[1 3 5 7]  Thread 1.1073741888	    <frame> at array-transform.cpp:61  
  2.4:[1 3 5 7]  Thread 1.1073742080	    <frame> at array-transform.cpp:61  
  2.5:[1 3 5 7]  Thread 1.1073742144	    <frame> at array-transform.cpp:61  
  2.6:[1 3 5 7]  Thread 1.1073742336	    <frame> at array-transform.cpp:61  
  2.7:[1 3 5 7]  Thread 1.1073745920	    <frame> at array-transform.cpp:61  
  2.8:[1 3 5 7]  Thread 1.1073746176	    <frame> at array-transform.cpp:61 
  2.9:[1 3 5 7]  Thread 1.1073746432	    <frame> at array-transform.cpp:61

To switch focus to a different SIMD lane, use the thread <thread_ID> command. Thread ID is specified by a triple: inferior.thread:lane. See examples of working with particular lanes:

Note

In the thread ID, the inferior number can be skipped. In this case, the current inferior ID is used. The thread number can also be skipped in case of switching to a lane in the current thread. Thus, the command below can be used to switch to the desired thread:
thread :5
Expected output:
[Switching to thread 2.1:5 (Thread 1.1073741824 lane 5)]
#0  main::$_1::operator()[...] at array-transform.cpp:61
61        result = -1;       // else-branch

As you are now inside the kernel running on the GPU, you can get the GPU assembly code by executing the following command:

disassemble

See an example output below:

Dump of assembler code for function _ZTSN2cl4sycl6kernelE(...):
   0x00000000fffad000 <+0>:         mov (1|M0)    null<1>:ud    0xC72C169A:ud
   0x00000000fffad010 <+16>:  (W)   mov (8|M0)    r22.0<1>:ud   r0.0<1;1,0>:ud
   0x00000000fffad020 <+32>:  (W)   or (1|M0)     cr0.0<1>:ud   cr0.0<0;1,0>:ud   0x4C0:uw     {Switch}
   0x00000000fffad030 <+48>:  (W)   mov (8|M0)    r9.0<1>:w     0x76543210:v
   0x00000000fffad040 <+64>:  (W)   and (1|M0)    r8.1<1>:d     r22.5<0;1,0>:d    511:w
   0x00000000fffad050 <+80>:  (W)   mul (1|M0)    r8.1<1>:d     r8.1<0;1,0>:d     0xC440:uw
   0x00000000fffad060 <+96>:  (W)   add (1|M0)    r8.2<1>:d     r8.1<0;1,0>:d     0x8440:uw
   0x00000000fffad070 <+112>:       mov (8|M0)    r9.0<1>:d     r9.0<8;8,1>:uw
   0x00000000fffad080 <+128>:       mul (8|M0)    r10.0<1>:d    r9.0<8;8,1>:d     8:w

To display a list of GPU registers, run the following command:

info reg

Additionally, inspect the execution mask ($emask register), which shows active lanes. To print the result in binary format, use the /t format flag as follows:

print/t $emask

Example output:

$3 = 10101010

Recall that you have stopped at line 61: the else-branch of the condition that checks evenness of the work item index. Hence, every other SIMD lane is inactive, as indicated by the $emask bit pattern.

To move forward and stop at the then-branch, set the scheduler-locking mode to step and execute the next command:

set scheduler-locking step
next

You should see the following output:

[Switching to SIMD lane 0]
Thread 2.1 hit Breakpoint 1, with SIMD lanes [0 2 4 6], main::$_1::operator()[...]
  at array-transform.cpp:59
59        result = result + 50;  // then-branch

Due to the breakpoint event, the SIMD lane focus switches to the first active lane in the then-branch, which is SIMD lane 0. Other threads of inferior 2 stayed at the line 61:

info threads 2.*

Example output:

  Id             Target Id           Frame
 *2.1:0          Thread 1.1073741824	  <frame> at array-transform.cpp:59
  2.1:[2 4 6]    Thread 1.1073741824	  <frame> at array-transform.cpp:59
  2.2:[1 3 5 7]  Thread 1.1073741888	  <frame> at array-transform.cpp:61
  2.3:[1 3 5 7]  Thread 1.1073742080	  <frame> at array-transform.cpp:61
  2.4:[1 3 5 7]  Thread 1.1073742144	  <frame> at array-transform.cpp:61
  2.5:[1 3 5 7]  Thread 1.1073742336	  <frame> at array-transform.cpp:61
  2.6:[1 3 5 7]  Thread 1.1073745920	  <frame> at array-transform.cpp:61
  2.7:[1 3 5 7]  Thread 1.1073746176	  <frame> at array-transform.cpp:61
  2.8:[1 3 5 7]  Thread 1.1073746432	  <frame> at array-transform.cpp:61

Since the thread is vectorized, you can also inspect the vector of a local variable:

x /8dw &result

Example output:

0x7fffe3f972c0:	 150	  -1	  152	  -1
0x7fffe3f972d0:	 154	  -1	  156	  -1

SIMD Lanes

To investigate all active SIMD lanes at once, use the thread apply command:

thread apply 2.1 print element

Example output:

Thread 2.1:0 (Thread 1.1073741824 lane 0):
$4 = 100

You can specify a SIMD lane as a number:

thread apply 2.1:2 print element

Example output:

Thread 2.1:2 (Thread 1.1073741824 lane 2):
$5 = 102

You can also specify a SIMD lane as a range. In this case, only active SIMD lanes from the range are considered:

thread apply 2.1:2-5 print element

Example output:

Thread 2.1:2 (Thread 1.1073741824 lane 2):
$6 = 102
warning: SIMD lane 3 is inactive in thread 1.2

Thread 2.1:4 (Thread 1.1073741824 lane 4):
$7 = 104
warning: SIMD lane 5 is inactive in thread 1.2

To denote all active SIMD lanes, use the wildcard:

thread apply 2.1:* print element

Example output:

Thread 2.1:0 (Thread 1.1073741824 lane 0):
$8 = 100

Thread 2.1:2 (Thread 1.1073741824 lane 2):
$9 = 102

Thread 2.1:4 (Thread 1.1073741824 lane 4):
$10 = 104

Thread 2.1:6 (Thread 1.1073741824 lane 6):
$11 = 106

To apply the command to all active SIMD lanes of all threads, use all-lanes parameter:

thread apply all-lanes print element

Example output:

Thread 2.8:7 (Thread 1.1073741888 lane 7):
$12 = 155
Thread 2.8:5 (Thread 1.1073741888 lane 5):
$13 = 153

[...]
Thread 2.1:2 (Thread 1.1073741824 lane 2):
$42 = 102

Thread 2.1:0 (Thread 1.1073741824 lane 0):
$43 = 100

Thread 1.2 (Thread 0x7ffff26dc700 (LWP 30173) "array-transform"):
No symbol "element" in current context.

You can mix SIMD lane ranges with thread ranges and the thread wildcard. For example, to apply the command to all active lanes of all threads of inferior 2, you can use any of the following commands:

If the current inferior is 2, the inferior number can be skipped:

Breakpoint Actions

You can define a set of actions for a breakpoint to be executed when the breakpoint is hit. By default, the actions are executed in the context of the SIMD lane selected after the hit.

Start a new debugging session and define two temporary breakpoints with actions for if and else branches:

    1. Set a temporary breakpoint:
      tbreak 61
      Example output:
      Temporary breakpoint 1 at 0x40584a: file /path/to/array-transform.cpp, line 61.
    2. Define an action:
      commands
      When you are asked to type commands, enter the following:
      print element
      end
    1. Set another temporary breakpoint:
      tbreak 59
      Example output:
      Temporary breakpoint 2 at 0x40583c: file /path/to/array-transform.cpp, line 59.
    2. Define an action to be executed for all SIMD lines by adding /a modifier:
      commands /a
      When you are asked to type commands, enter the following:
      print element
      end

Start the program:

run gpu

Example output:

[...]
Thread 2.1 hit Temporary breakpoint 1, with SIMD lanes [1 3 5 7],
  main::$_1::operator()[...] at array-transform.cpp:61
61	          result = -1;  // else-branch
$1 = 101

Continue to hit both breakpoints:

continue

Example output:

Continuing.
[Switching to SIMD lane 0]
Thread 2.1 hit Temporary breakpoint 2, with SIMD lanes [0 2 4 6],
  main::$_1::operator()[...] at array-transform.cpp:59
59	          result = result + 50;  // then-branch
$2 = 100
$3 = 102
$4 = 104
$5 = 106

The action for the breakpoint at the else branch was executed for a single SIMD lane 1, while the action at the then branch was executed for all active SIMD lanes.

Note

For conditional breakpoints, the actions are executed only for SIMD lanes that meet the condition.

Conditional Breakpoints

Quit the debugging session and start the program from the beginning. This time set a breakpoint at line 59 with the condition element==106:

break 59 if element == 106

Example output:

Breakpoint 1 at 0x40583c: file /path/to/array-transform.cpp, line 59.

Run the program (execute the run gpu command) and check if the output looks as follows:

Starting program: <path_to_array-transform> gpu

[...]

[Switching to Thread 1.1073741824 lane 6]

Thread 2.1 hit Breakpoint 1, with SIMD lane 6, main::$_1::operator()[...]
  at array-transform.cpp:59
59        result = result + 50; // then-branch

The condition is true for the lane 6 in thread 2.1.

Note

A breakpoint condition is evaluated only for active SIMD lanes, meaning that (gdb) break 59 if element == 107 does not cause a stop, since element == 107 is true for the lane 7 in thread 2.1, and it is inactive at line 59.