Tutorial

  • 2022.3
  • 09/08/2022
  • Public Content

Debugging Options

Tutorial: Debugging with Intel® Distribution for GDB*

Auto-Attach

The auto-attach feature enables listening to debug events from the GPU.
This feature allows the debugger to launch a
gdbserver-gt
process to listen to GPU debug events and add the
gdbserver-gt
to the debugger as an inferior. This feature is designed to improve your debugging experience and ensure that you can debug the kernel offloaded to GPU.
The auto-attach feature is enabled by default.
The feature does not affect the debugging capability on the CPU device. However, to eliminate the extra output the feature creates, you can turn it off with the
INTELGT_AUTO_ATTACH_DISABLE
environment variable. To do it, execute the following command on the shell before starting
gbd-oneapi
:
export INTELGT_AUTO_ATTACH_DISABLE=1
To enable the feature again:
unset INTELGT_AUTO_ATTACH_DISABLE

Pretty-Printing

The pretty-printing feature simplifies the display of complex objects. If a pretty-printer is registered for the type of value you are going to print, the debugger simplifies the output. Otherwise, the debugger prints the value normally.
Intel® Distribution for GDB* supports pretty-printing for SYCL* types
id
,
buffer
, and
range
from the
cl::sycl
namespace.
To display the list of pretty-printers available, run the following command:
info pretty-print
Example output:
global pretty-printers: SYCL cl::sycl::buffer cl::sycl::id cl::sycl::range builtin mpx_bound128
Pretty-printing is enabled by default. For example, when you print a value of the
index
variable:
print index
The output is the following:
$10 = cl::sycl::id<1> = {32}
To disable pretty-printing and display raw content instead, use the
/r
flag:
print /r index
Example output:
$11 = {<cl::sycl::detail::array<1>> = {common_array = {32, <No data fields>}}}
To disable all pretty-printers, use the following command:
disable pretty-printer
To enable pretty-printers:
enable pretty-printer

Product and Performance Information

1

Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex.