Introduction

The amd-dbgapi is a library that implements an AMD GPU debugger application programming interface (API). It provides the support necessary for a client of the library to control the execution and inspect the state of supported commercially available AMD GPU devices.

The term client is used to refer to the application that uses this API.

The term library is used to refer to the implementation of this interface being used by the client.

The term AMD GPU is used to refer to commercially available AMD GPU devices supported by the library.

The term inferior is used to refer to the process being debugged.

The library does not provide any operations to perform symbolic mappings, code object decoding, or stack unwinding. The client must use the AMD GPU code object ELF ABI defined in User Guide for AMDGPU Backend - Code Object, together with the AMD GPU debug information DWARF and call frame information CFI ABI define in User Guide for AMDGPU Backend - Code Object - DWARF to perform those tasks.

The library does not provide operations for inserting or managing breakpoints. The client must write the architecture specific breakpoint instruction provided by the AMD_DBGAPI_ARCHITECTURE_INFO_BREAKPOINT_INSTRUCTION query into the loaded code object memory to set breakpoints. For resuming from breakpoints the client must use the displaced stepping mechanism provided by amd_dbgapi_displaced_stepping_start and amd_dbgapi_displaced_stepping_complete in conjunction with the amd_dbgapi_wave_resume in single step mode. In order to determine the location of stopped waves the client must read the architecture specific program counter register available using the AMD_DBGAPI_ARCHITECTURE_INFO_PC_REGISTER query and adjust it by the amount specified by the AMD_DBGAPI_ARCHITECTURE_INFO_BREAKPOINT_INSTRUCTION_PC_ADJUST query.

The client is responsible for checking that only a single thread at a time invokes a function provided by the library. A callback (see Callbacks) invoked by the library must not itself invoke any function provided by the library.

The library implementation uses the native operating system to inspect and control the inferior. Therefore, the library must be executed on the same machine as the inferior.

A library instance is defined as the period between a call to amd_dbgapi_initialize and a matching call to amd_dbgapi_finalize.

The library uses opaque handles to refer to the entities that it manages. A handle value should not be modified directly. See the handle definitions for information on the lifetime and scope of handles of that type. Handles are invalidated outside their lifetime, scope, or single library instance. If a handle is returned by an operation in one library instance which then becomes invalidated, then any operation using that handle in the same library instance will return an invalid handle error code. However, it is undefined to use a handle created by an operation in one library instance in the operations of another library instance. A handle value is globally unique within each library instance. This is true even if the handle becomes invalidated: handle values are not reused within a library instance. Every handle with handle value of 0 is reserved to indicate the handle does not reference an entity.

When the library is first loaded it is in the uninitialized state with the logging level set to AMD_DBGAPI_LOG_LEVEL_NONE.

AMD GPU Execution Model

In this section the AMD GPU execution model is described to provide background to the reader if they are not familiar with this environment. The AMD GPU execution model is more complicated than that of a traditional CPU because of how GPU hardware is used to accelerate and schedule the very large number of threads of execution that are created on GPUs.

Chapter 2 of the [HSA Programmer's Reference Manual][hsa-prm] provides an introduction to this execution model. Note that the AMD ROCm compilers compile directly to ISA and do not use the HSAIL intermediate language. However, the ROCr low-level runtime and ROCgdb debugger use the same terminology.

In this model, a CPU process may interact with multiple AMD GPU devices, which are termed agents. A Process Address Space Identifier (PASID) is created for each process that interacts with agents. An agent can be executing code for multiple processes at once. This is achieved by mapping the PASID to one of a limited set of Virtual Memory Identifiers (VMIDs). Each VMID is associated with its own page table.

The AMD GPU device driver for Linux, termed the Kernel Mode Driver (KMD), manages the page tables used by each GPU so they correlate with the CPU page table for the corresponding process. The CPU and GPU page tables do not necessarily map all the same memory pages but pages they do have in common have the same virtual address. Therefore, the CPU and GPUs have a unified address space.

Each GPU includes one or more Microcode Engines (ME) that can execute microcode firmware. This firmware includes a Hardware Scheduler (HWS) that, in collaboration with the KMD, manages which processes, identified by a PASID, are mapped onto the GPU using one of the limited VMIDs. This mapping configures the VMID to use the GPU page table that corresponds to the PASID. In this way, the code executing on the GPU from different processes is isolated.

Multiple software submission queues may be created for each agent. The GPU hardware has a limited number of pipes, each of which has a fixed number of hardware queues. The HWS, in collaboration with the KMD, is responsible for mapping software queues onto hardware queues. This is done by multiplexing the software queues onto hardware queues using time slicing. The software queues provide a virtualized abstraction, allowing for more queues than are directly supported by the hardware. Each ME manages its own set of pipes and their associated hardware queues.

To execute code on the GPU, a packet must be created and placed in a software queue. This is achieved using regular user space atomic memory operations. No Linux kernel call is required. For this reason, the queues are termed user mode queues.

The AMD ROCm platform uses the Asynchronous Queuing Language (AQL) packet format defined in the [HSA Platform System Architecture Specification][hsa-sysarch]. Packets can request GPU management actions (for example, manage memory coherence) and the execution of kernel functions. The ME firmware includes the Command Processor (CP) which, together with fixed-function hardware support, is responsible for detecting when packets are added to software queues that are mapped to hardware queues. Once detected, CP is responsible for initiating actions requested by the packet, using the appropriate VMID when performing all memory operations.

Dispatch packets are used to request the execution of a kernel function. Each dispatch packet specifies the address of a kernel descriptor, the address of the kernel argument block holding the arguments to the kernel function, and the number of threads of execution to create to execute the kernel function. The kernel descriptor describes how the CP must configure the hardware to execute the kernel function and the starting address of the kernel function code. The compiler generates a kernel descriptor in the code object for each kernel function and determines the kernel argument block layout. The number of threads of execution is specified as a grid, such that each thread of execution can identify its position in the grid. Conceptually, each of these threads executes the same kernel code, with the same arguments.

The dispatch grid is organized as a three-dimensional collection of workgroups, where each workgroup is the same size (except for potential boundary partial workgroups). The workgroups form a three-dimensional collection of work-items. The work-items are the threads of execution. The position of a work-item is its zero-based three-dimensional position in a workgroup, termed its work-item ID, plus its workgroup's three-dimensional position in the dispatch grid, termed its workgroup ID. These three-dimensional IDs can also be expressed as a zero-based one-dimensional ID, termed a flat ID, by simply numbering the elements in a natural manner akin to linearizing a multi-dimensional array.

Consecutive work-items, in flat work-item ID order, of a workgroup are organized into fixed size wavefronts, or waves for short. Each work-item position in the wave is termed a lane, and has a zero-base lane ID. The hardware imposes an upper limit on the number of work-items in a workgroup but does not limit the number of workgroups in a dispatch grid. The hardware executes instructions for waves independently. But the lanes of a wave all execute the same instruction jointly. This is termed Single Instruction Multiple Thread (SIMT) execution.

Each hardware wave has a set of registers that are shared by all lanes of the wave, termed scalar registers. There is only one set of scalar registers for the whole wave. Instructions that act on the whole wave, which typically use scalar registers, are termed scalar instructions.

Additionally, each wave also has a set of vector registers that are replicated so each lane has its own copy. A set of vector registers can be viewed as a vector with each element of the vector belonging to the corresponding lane of the wave. Instructions that act on vector registers, which produce independent results for each lane, are termed vector instructions.

Each hardware wave has an execution mask that controls if the execution of a vector instruction should change the state of a particular lane. If the lane is masked off, no changes are made for that lane and the instruction is effectively ignored. The compiler generates code to update the execution mask which emulates independent work-item execution. However, the lanes of a wave do not execute instructions independently. If two subsets of lanes in a wave need to execute different code, the compiler will generate code to set the execution mask to execute the subset of lanes for one path, then generate instructions for that path. The compiler will then generate code to change the execution mask to enable the other subset of lanes, then generate code for those lanes. If both subsets of lanes execute the same code, the compiler will generate code to set the execution mask to include both subsets of lanes, then generate code as usual. When only a subset of lanes is enabled, they are said to be executing divergent control flow. When all lanes are enabled, they are said to be executing wave uniform control flow.

Not all MEs have the hardware to execute kernel functions. One such ME is used to execute the HWS microcode and to execute microcode that manages a service queue that is used to update GPU state. If the ME does support kernel function execution it uses fixed-function hardware to initiate the creation of waves. This is accomplished by sending requests to create workgroups to one or more Compute Units (CUs). Requests are sent to create all the workgroups of a dispatch grid. Each CU has resources to hold a fixed number of waves and has fixed-function hardware to schedule execution of these waves. The scheduler may execute multiple waves concurrently and will hide latency by switching between the waves that are ready to execute. At any point of time, a subset of the waves belonging to workgroups in a dispatch may be actively executing. As waves complete, the waves of subsequent workgroup requests are created.

Each CU has a fixed amount of memory from which it allocates vector and scalar registers. The kernel descriptor specifies how many registers to allocate for a wave. There is a tradeoff between how many waves can be created on a CU and the number of registers each can use.

The CU also has a fixed size Local Data Store (LDS). A dispatch packet specifies how much LDS each workgroup is allocated. All waves in a workgroup are created on the same CU. This allows the LDS to be used to share data between the waves of the same workgroup. There is a tradeoff between how much LDS a workgroup can allocate, and the number of workgroups that can fit on a CU. The address of a location in a workgroup LDS allocation is zero-based and is a different address space than the global virtual memory. There are specific instructions that take an LDS address to access it. There are also flat address instructions that map the LDS address range into an unused fixed aperture range of the global virtual address range. An LDS address can be converted to or from a flat address by offsetting by the base of the aperture. Note that a flat address in the LDS aperture only accesses the LDS workgroup allocation for the wave that uses it. The same address will access different LDS allocations if used by waves in different workgroups.

The dispatch packet specifies the amount of scratch memory that must be allocated for a work-item. This is used for work-item private memory. Fixed-function hardware in the CU manages per wave allocation of scratch memory from pre-allocated global virtual memory mapped to GPU device memory. Like an LDS address, a scratch address is zero-based, but is per work-item instead of per workgroup. It maps to an aperture in a flat address. The hardware swizzles this address so that adjacent lanes access adjacent DWORDs (4 bytes) in global memory for better cache performance.

For an AMD Radeon Instinct™ MI60 GPU the workgroup size limit is 1,024 work-items, the wave size is 64, and the CU count is 64. A CU can hold up to 40 waves (this is limited to 32 if using scratch memory). Therefore, a workgroup can comprise between 1 and 16 waves inclusive, and there can be up to 2,560 waves, making a maximum of 163,840 work-items. A CU is organized as 4 Execution Units (EUs) also referred to as Single Instruction Multiple Data units (SIMDs) that can each hold 10 waves. Each SIMD has 256 64-wide DWORD vector registers and each CU has 800 DWORD scalar registers. A single wave can access up to 256 64-wide vector registers and 112 scalar registers. A CU has 64KiB of LDS.

Supported AMD GPU Architectures

The following AMD GPU architectures are supported:

gfx900 (AMD Vega 10)
gfx906 (AMD Vega 7nm also referred to as AMD Vega 20)
gfx908 (AMD Instinct™ MI100 accelerator)
gfx90a (Aldebaran)
gfx940 (Aqua Vanjaram)
gfx941
gfx942
gfx1010 (Navi10)
gfx1011 (Navi12)
gfx1012 (Navi14)
gfx1030 (Sienna Cichlid)
gfx1031 (Navy Flounder)
gfx1032 (Dimgrey Cavefish)
gfx1100 (Plum Bonito)
gfx1101 (Wheat Nas)
gfx1102 (Hotpink Bonefish)
gfx1200
gfx1201

For more information about the AMD ROCm ecosystem, please refer to:

https://rocm.docs.amd.com/

Known Limitations and Restrictions

The AMD Debugger API library implementation currently has the following restrictions. Future releases aim to address these restrictions.

The following *_get_info queries are not yet implemented:
- AMD_DBGAPI_QUEUE_INFO_ERROR_REASON
- AMD_DBGAPI_QUEUE_INFO_STATE
On a AMD_DBGAPI_STATUS_FATAL error the library does fully reset the internal state and so subsequent functions may not operate correctly.
amd_dbgapi_process_next_pending_event returns AMD_DBGAPI_EVENT_KIND_WAVE_STOP events only for AQL queues. PM4 queues that launch wavefronts are not supported.
amd_dbgapi_queue_packet_list returns packets only for AQL queues.
Generation of the AMD_DBGAPI_EVENT_KIND_QUEUE_ERROR event, the AMD_DBGAPI_EVENT_INFO_QUEUE query, and the generation of AMD_DBGAPI_EVENT_KIND_WAVE_COMMAND_TERMINATED events for waves with pending single step requests when a queue enters the queue error state, have not been implemented.
By default, for some architectures, the AMD GPU device driver for Linux causes all wavefronts created when the library is not attached to the process by amd_dbgapi_process_attach to be unable to query the wavesfront's AMD_DBGAPI_WAVE_INFO_DISPATCH, AMD_DBGAPI_WAVE_INFO_WORKGROUP_COORD, or AMD_DBGAPI_WAVE_INFO_WAVE_NUMBER_IN_WORKGROUP, or workgroup's AMD_DBGAPI_WORKGROUP_INFO_DISPATCH, or AMD_DBGAPI_WORKGROUP_INFO_WORKGROUP_COORD. This does not affect wavefronts and workgroups created while the library is attached to the process which are always capable of reporting this information.

If the HSA_ENABLE_DEBUG environment variable is set to "1" when the inferior's runtime is successfully enabled (see AMD_DBGAPI_EVENT_KIND_RUNTIME), then this information will be available for all architecture even for wavefronts created when the library was not attached to the process. Setting this environment variable may very marginally reduce wavefront launch latency for some architectures for very short lived wavefronts.

See also
amd_dbgapi_wave_get_info
The AMD_DBGAPI_WAVE_STOP_REASON_FP_* and AMD_DBGAPI_WAVE_STOP_REASON_INT-* stop reasons (see amd_dbgapi_wave_stop_reasons_t) are not reported for enabled arithmetic exceptions if the DX10_CLAMP bit in the MODE register is set. This happens if the DX10_CLAMP kernel descriptor field is set.
The library does not support single root I/O virtualization (SR-IOV) on any AMD GPU architecture that supports it. That includes gfx1030, gfx1031, and gfx1032.
The library does not support debugging programs that use cooperative groups or CU masking for gfx1100, gfx1101, and gfx11102. A restriction will be reported when attaching to a process that has already created cooperative group queues or CU masked queues. Any attempt by the process to create a cooperative queue or CU masked queue when attached will fail.
On gfx1100, gfx1101 and gfx1102, the library cannot debug a program past a "s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)" instruction. If an exception is delivered to a wave in an attached process after the wave has executed this instruction, the wave is killed.

References

Advanced Micro Devices: www.amd.com
AMD ROCm Ecosystem: rocm.docs.amd.com
Bus:Device.Function (BDF) Notation: wiki.xen.org/wiki/Bus:Device.Function_(BDF)_Notation
HSA Platform System Architecture Specification: https://hsafoundation.com/wp-content/uploads/2021/02/HSA-SysArch-1.2.pdf
HSA Programmer's Reference Manual: https://hsafoundation.com/wp-content/uploads/2021/02/HSA-PRM-1.2.pdf
Semantic Versioning: semver.org
The LLVM Compiler Infrastructure: llvm.org
User Guide for AMDGPU LLVM Backend: llvm.org/docs/AMDGPUUsage.html

Legal Disclaimer and Copyright Information

AMD ROCm software is made available by Advanced Micro Devices, Inc. under the open source license identified in the top-level directory for the library in the repository on Github.com (Portions of AMD ROCm software are licensed under MITx11 and UIL/NCSA. For more information on the license, review the license.txt in the top-level directory for the library on Github.com). The additional terms and conditions below apply to your use of AMD ROCm technical documentation.

The information presented in this document is for informational purposes only and may contain technical inaccuracies, omissions, and typographical errors. The information contained herein is subject to change and may be rendered inaccurate for many reasons, including but not limited to product and roadmap changes, component and motherboard version changes, new model and/or product releases, product differences between differing manufacturers, software changes, BIOS flashes, firmware upgrades, or the like. Any computer system has risks of security vulnerabilities that cannot be completely prevented or mitigated. AMD assumes no obligation to update or otherwise correct or revise this information. However, AMD reserves the right to revise this information and to make changes from time to time to the content hereof without obligation of AMD to notify any person of such revisions or changes.

THIS INFORMATION IS PROVIDED "AS IS." AMD MAKES NO REPRESENTATIONS OR WARRANTIES WITH RESPECT TO THE CONTENTS HEREOF AND ASSUMES NO RESPONSIBILITY FOR ANY INACCURACIES, ERRORS, OR OMISSIONS THAT MAY APPEAR IN THIS INFORMATION. AMD SPECIFICALLY DISCLAIMS ANY IMPLIED WARRANTIES OF NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR ANY PARTICULAR PURPOSE. IN NO EVENT WILL AMD BE LIABLE TO ANY PERSON FOR ANY RELIANCE, DIRECT, INDIRECT, SPECIAL, OR OTHER CONSEQUENTIAL DAMAGES ARISING FROM THE USE OF ANY INFORMATION CONTAINED HEREIN, EVEN IF AMD IS EXPRESSLY ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.

AMD, the AMD Arrow logo, AMD Instinct™, Radeon™, AMD ROCm™, and combinations thereof are trademarks of Advanced Micro Devices, Inc. Linux® is the registered trademark of Linus Torvalds in the U.S. and other countries. PCIe® is a registered trademark of PCI-SIG Corporation. Other product names used in this publication are for identification purposes only and may be trademarks of their respective companies.

AMD DBG API

AMD DBG API#