Previous: , Up: Architectures   [Contents][Index]


21.4.9 AMD GPU

GDB supports debugging programs offloaded to AMD GPU devices using the AMD ROCm platform. GDB presents host threads alongside GPU wavefronts, allowing debugging both the host and device parts of the program simultaneously.

21.4.9.1 AMD GPU Architectures

The list of AMD GPU architectures supported by GDB depends on the version of the AMD Debugger API library used. See its documentation for more details.

21.4.9.2 AMD GPU Device Driver and AMD ROCm Runtime

GDB requires a compatible AMD GPU device driver to be installed. A warning message is displayed if either the device driver version or the version of the debug support it implements is unsupported. GDB will continue to function except no AMD GPU debugging will be possible.

GDB requires each agent to have compatible firmware installed by the device driver. A warning message is displayed if unsupported firmware is detected. GDB will continue to function except no AMD GPU debugging will be possible on the agent.

GDB requires a compatible AMD ROCm runtime to be loaded in order to detect AMD GPU code objects and wavefronts. A warning message is displayed if an unsupported AMD ROCm runtime is detected, or there is an error or restriction that prevents debugging. GDB will continue to function except no AMD GPU debugging will be possible.

21.4.9.3 AMD GPU Wavefronts

An AMD GPU wavefront is represented in GDB as a thread.

Note that some AMD GPU architectures may have restrictions on providing information about AMD GPU wavefronts created when GDB is not attached (see AMD GPU Attaching Restrictions).

When scheduler-locking is in effect (see set scheduler-locking), new wavefronts created by the resumed thread (either CPU thread or GPU wavefront) are held in the halt state.

21.4.9.4 AMD GPU Code Objects

The ‘info sharedlibrary’ command will show the AMD GPU code objects as file or memory URIs, together with the host’s shared libraries. For example:

(gdb) info sharedlibrary
From    To      Syms Read   Shared Object Library
0x1111  0x2222  Yes (*)     /lib64/ld-linux-x86-64.so.2
...
0x3333  0x4444  Yes (*)     /opt/rocm-4.5.0/.../libamd_comgr.so
0x5555  0x6666  Yes (*)     /lib/x86_64-linux-gnu/libtinfo.so.5
0x7777  0x8888  Yes         file:///tmp/a.out#offset=6477&size=10832
0x9999  0xaaaa  Yes (*)     memory://95557/mem#offset=0x1234&size=100
(*): Shared library is missing debugging information.
(gdb)

For a ‘file’ URI, the path portion is the file on disk containing the code object. The offset parameter is a 0-based offset in this file, to the start of the code object. If omitted, it defaults to 0. The size parameter is the size of the code object in bytes. If omitted, it defaults to the size of the file.

For a ‘memory’ URI, the path portion is the process id of the process owning the memory containing the code object. The offset parameter is the memory address where the code object is found, and the size parameter is its size in bytes.

AMD GPU code objects are loaded into each AMD GPU device separately. The ‘info sharedlibrary’ command may therefore show the same code object loaded multiple times. As a consequence, setting a breakpoint in AMD GPU code will result in multiple breakpoint locations if there are multiple AMD GPU devices.

21.4.9.5 AMD GPU Entity Target Identifiers and Convenience Variables

The AMD GPU entities have the following target identifier formats:

Thread Target ID

The AMD GPU thread target identifier (systag) string has the following format:

AMDGPU Wave agent-id:queue-id:dispatch-id:wave-id (work-group-x,work-group-y,work-group-z)/work-group-thread-index

21.4.9.6 AMD GPU Signals

For AMD GPU wavefronts, GDB maps target conditions to stop signals in the following way:

SIGILL

Execution of an illegal instruction.

SIGTRAP

Execution of a S_TRAP instruction other than:

SIGABRT

Execution of a S_TRAP 2 instruction.

SIGFPE

Execution of a floating point or integer instruction detects a condition that is enabled to raise a signal. The conditions include:

By default, these conditions are not enabled to raise signals. The ‘set $mode’ command can be used to change the AMD GPU wavefront’s register that has bits controlling which conditions are enabled to raise signals. The ‘print $trapsts’ command can be used to inspect which conditions have been detected even if they are not enabled to raise a signal.

SIGBUS

Execution of an instruction that accessed global memory using an address that is outside the virtual address range.

SIGSEGV

Execution of an instruction that accessed a global memory page that is either not mapped or accessed with incompatible permissions.

If a single instruction raises more than one signal, they will be reported one at a time each time the wavefront is continued.

21.4.9.7 AMD GPU Memory Violation Reporting

A wavefront can report memory violation events. However, the program location at which they are reported may be after the machine instruction that caused them. This can result in the reported source statement being incorrect. The following commands can be used to control this behavior:

set amdgpu precise-memory mode

Controls how AMD GPU devices detect memory violations, where mode can be:

off

The program location may not be immediately after the instruction that caused the memory violation. This is the default.

on

Requests that the program location will be immediately after the instruction that caused a memory violation. Enabling this mode may make the AMD GPU device execution significantly slower as it has to wait for each memory operation to complete before executing the next instruction.

The amdgpu precise-memory parameter is per-inferior. When an inferior forks or execs, or the user uses the clone-inferior command, and an inferior is created as a result, the newly created inferior inherits the parameter value of the original inferior.

show amdgpu precise-memory

Displays the currently requested AMD GPU precise memory setting.

21.4.9.8 AMD GPU Logging

The ‘set debug amd-dbgapi’ command can be used to enable diagnostic messages in the ‘amd-dbgapi’ target. The ‘show debug amd-dbgapi’ command displays the current setting. See set debug amd-dbgapi.

The ‘set debug amd-dbgapi-lib log-level level’ command can be used to enable diagnostic messages from the ‘amd-dbgapi’ library (which GDB uses under the hood). The ‘show debug amd-dbgapi-lib log-level’ command displays the current ‘amd-dbgapi’ library log level. See set debug amd-dbgapi-lib.

21.4.9.9 AMD GPU Restrictions

  1. When in non-stop mode, wavefronts may not hit breakpoints inserted while not stopped, nor see memory updates made while not stopped, until the wavefront is next stopped. Memory updated by non-stopped wavefronts may not be visible until the wavefront is next stopped.
  2. The HIP runtime performs deferred code object loading by default. AMD GPU code objects are not loaded until the first kernel is launched. Before then, all breakpoints have to be set as pending breakpoints.

    If source line positions are used that only correspond to source lines in unloaded code objects, then GDB may not set pending breakpoints, and instead set breakpoints on the next following source line that maps to host code. This can result in unexpected breakpoint hits being reported. When the code object containing the source lines is loaded, the incorrect breakpoints will be removed and replaced by the correct ones. This problem can be avoided by only setting breakpoints in unloaded code objects using symbol or function names.

    Setting the HIP_ENABLE_DEFERRED_LOADING environment variable to 0 can be used to disable deferred code object loading by the HIP runtime. This ensures all code objects will be loaded when the inferior reaches the beginning of the main function.

  3. If no CPU thread is running, then ‘Ctrl-C’ is not able to stop AMD GPU threads. This can happen for example if you enable scheduler-locking after the whole program stopped, and then resume an AMD GPU thread. The only way to unblock the situation is to kill the GDB process.
  4. By default, for some architectures, the AMD GPU device driver causes all AMD GPU wavefronts created when GDB is not attached to be unable to report the dispatch associated with the wavefront, or the wavefront’s work-group position. The ‘info threads’ command will display this missing information with a ‘?’.

    This does not affect wavefronts created while GDB is attached which are always capable of reporting this information.

    If the HSA_ENABLE_DEBUG environment variable is set to ‘1’ when the AMD ROCm runtime is initialized, then this information will be available for all architectures even for wavefronts created when GDB was not attached.


Previous: , Up: Architectures   [Contents][Index]