User Guide for AMDGPU Backend

Introduction

The AMDGPU backend provides ISA code generation for AMD GPUs, starting with the R600 family up until the current GCN families. It lives in the llvm/lib/Target/AMDGPU directory.

LLVM

Target Triples

Use the clang -target <Architecture>-<Vendor>-<OS>-<Environment> option to specify the target triple:

AMDGPU Architectures
Architecture Description
r600 AMD GPUs HD2XXX-HD6XXX for graphics and compute shaders.
amdgcn AMD GPUs GCN GFX6 onwards for graphics and compute shaders.
AMDGPU Vendors
Vendor Description
amd Can be used for all AMD GPU usage.
mesa3d Can be used if the OS is mesa3d.
AMDGPU Operating Systems
OS Description
<empty> Defaults to the unknown OS.
amdhsa Compute kernels executed on HSA [HSA] compatible runtimes such as AMD’s ROCm [AMD-ROCm].
amdpal Graphic shaders and compute kernels executed on AMD PAL runtime.
mesa3d Graphic shaders and compute kernels executed on Mesa 3D runtime.
AMDGPU Environments
Environment Description
<empty> Default.

Processors

Use the clang -mcpu <Processor> option to specify the AMDGPU processor. The names from both the Processor and Alternative Processor can be used.

AMDGPU Processors
Processor Alternative Processor Target Triple Architecture dGPU/ APU Target Features Supported [Default] ROCm Support Example Products
Radeon HD 2000/3000 Series (R600) [AMD-RADEON-HD-2000-3000]
r600   r600 dGPU      
r630   r600 dGPU      
rs880   r600 dGPU      
rv670   r600 dGPU      
Radeon HD 4000 Series (R700) [AMD-RADEON-HD-4000]
rv710   r600 dGPU      
rv730   r600 dGPU      
rv770   r600 dGPU      
Radeon HD 5000 Series (Evergreen) [AMD-RADEON-HD-5000]
cedar   r600 dGPU      
cypress   r600 dGPU      
juniper   r600 dGPU      
redwood   r600 dGPU      
sumo   r600 dGPU      
Radeon HD 6000 Series (Northern Islands) [AMD-RADEON-HD-6000]
barts   r600 dGPU      
caicos   r600 dGPU      
cayman   r600 dGPU      
turks   r600 dGPU      
GCN GFX6 (Southern Islands (SI)) [AMD-GCN-GFX6]
gfx600
  • tahiti
 amdgcn dGPU      
gfx601
  • pitcairn
  • verde
 amdgcn dGPU      
gfx602
  • hainan
  • oland
 amdgcn dGPU      
GCN GFX7 (Sea Islands (CI)) [AMD-GCN-GFX7]
gfx700
  • kaveri
 amdgcn APU    
  • A6-7000
  • A6 Pro-7050B
  • A8-7100
  • A8 Pro-7150B
  • A10-7300
  • A10 Pro-7350B
  • FX-7500
  • A8-7200P
  • A10-7400P
  • FX-7600P
gfx701
  • hawaii
 amdgcn dGPU   ROCm
  • FirePro W8100
  • FirePro W9100
  • FirePro S9150
  • FirePro S9170
gfx702   amdgcn dGPU   ROCm
  • Radeon R9 290
  • Radeon R9 290x
  • Radeon R390
  • Radeon R390x
gfx703
  • kabini
  • mullins
 amdgcn APU    
  • E1-2100
  • E1-2200
  • E1-2500
  • E2-3000
  • E2-3800
  • A4-5000
  • A4-5100
  • A6-5200
  • A4 Pro-3340B
gfx704
  • bonaire
 amdgcn dGPU    
  • Radeon HD 7790
  • Radeon HD 8770
  • R7 260
  • R7 260X
gfx705   amdgcn APU     TBA
GCN GFX8 (Volcanic Islands (VI)) [AMD-GCN-GFX8]
gfx801
  • carrizo
 amdgcn APU
  • xnack [on]
  
  • A6-8500P
  • Pro A6-8500B
  • A8-8600P
  • Pro A8-8600B
  • FX-8800P
  • Pro A12-8800B
  amdgcn APU
  • xnack [on]
 ROCm
  • A10-8700P
  • Pro A10-8700B
  • A10-8780P
  amdgcn APU
  • xnack [on]
  
  • A10-9600P
  • A10-9630P
  • A12-9700P
  • A12-9730P
  • FX-9800P
  • FX-9830P
  amdgcn APU
  • xnack [on]
  
  • E2-9010
  • A6-9210
  • A9-9410
gfx802
  • iceland
  • tonga
 amdgcn dGPU
  • xnack [off]
 ROCm
  • Radeon R285
  • Radeon R9 380
  • Radeon R9 385
gfx803
  • fiji
 amdgcn dGPU
  • xnack [off]
 ROCm
  • Radeon R9 Nano
  • Radeon R9 Fury
  • Radeon R9 FuryX
  • Radeon Pro Duo
  • FirePro S9300x2
  • Radeon Instinct MI8
  • polaris10
 amdgcn dGPU
  • xnack [off]
 ROCm
  • Radeon RX 470
  • Radeon RX 480
  • Radeon Instinct MI6
  • polaris11
 amdgcn dGPU
  • xnack [off]
 ROCm
  • Radeon RX 460
gfx805
  • tongapro
 amdgcn dGPU
  • xnack [off]
 ROCm
  • FirePro S7150
  • FirePro S7100
  • FirePro W7100
  • Mobile FirePro M7170
gfx810
  • stoney
 amdgcn APU
  • xnack [on]
   TBA
GCN GFX9 [AMD-GCN-GFX9]
gfx900   amdgcn dGPU
  • xnack [off]
 ROCm
  • Radeon Vega Frontier Edition
  • Radeon RX Vega 56
  • Radeon RX Vega 64
  • Radeon RX Vega 64 Liquid
  • Radeon Instinct MI25
gfx902   amdgcn APU
  • xnack [on]
  
  • Ryzen 3 2200G
  • Ryzen 5 2400G
gfx904   amdgcn dGPU
  • xnack [off]
   TBA
gfx906   amdgcn dGPU
  • xnack [off]
  • sram-ecc [off]
  
  • Radeon Instinct MI50
  • Radeon Instinct MI60
  • Radeon VII
  • Radeon Pro VII
gfx908   amdgcn dGPU
  • xnack [off]
  • sram-ecc [on]
   TBA
gfx909   amdgcn APU
  • xnack [off]
   TBA
gfx90c   amdgcn APU
  • xnack [on]
  
  • Ryzen 7 4700G
  • Ryzen 7 4700GE
  • Ryzen 7 4700G
  • Ryzen 7 4700GE
  • Ryzen 5 4600G
  • Ryzen 5 4600GE
  • Ryzen 3 4300G
  • Ryzen 3 4300GE
  • Ryzen Pro 4000G
  • Ryzen 7 Pro 4700G
  • Ryzen 7 Pro 4750GE
  • Ryzen 5 Pro 4650G
  • Ryzen 5 Pro 4650GE
  • Ryzen 3 Pro 4350G
  • Ryzen 3 Pro 4350GE
GCN GFX10 [AMD-GCN-GFX10]
gfx1010   amdgcn dGPU
  • xnack [off]
  • wavefrontsize64 [off]
  • cumode [off]
  
  • Radeon RX 5700
  • Radeon RX 5700 XT
  • Radeon Pro 5600 XT
  • Radeon Pro 5600M
gfx1011   amdgcn dGPU
  • xnack [off]
  • wavefrontsize64 [off]
  • cumode [off]
   TBA
gfx1012   amdgcn dGPU
  • xnack [off]
  • wavefrontsize64 [off]
  • cumode [off]
  
  • Radeon RX 5500
  • Radeon RX 5500 XT
gfx1030   amdgcn dGPU
  • wavefrontsize64 [off]
  • cumode [off]
   TBA
gfx1031   amdgcn dGPU
  • wavefrontsize64 [off]
  • cumode [off]
   TBA
gfx1032   amdgcn dGPU
  • wavefrontsize64 [off]
  • cumode [off]
   TBA
gfx1033   amdgcn APU
  • wavefrontsize64 [off]
  • cumode [off]
   TBA

Target Features

Target features control how code is generated to support certain processor specific features. Not all target features are supported by all processors. The runtime must ensure that the features supported by the device used to execute the code match the features enabled when generating the code. A mismatch of features may result in incorrect execution, or a reduction in performance.

The target features supported by each processor, and the default value used if not specified explicitly, is listed in AMDGPU Processors.

Use the clang -m[no-]<TargetFeature> option to specify the AMDGPU target features.

For example:

-mxnack
Enable the xnack feature.
-mno-xnack

Disable the xnack feature.

AMDGPU Target Features
Target Feature Description
-m[no-]xnack

Enable/disable generating code that has memory clauses that are compatible with having XNACK replay enabled.

This is used for demand paging and page migration. If XNACK replay is enabled in the device, then if a page fault occurs the code may execute incorrectly if the xnack feature is not enabled. Executing code that has the feature enabled on a device that does not have XNACK replay enabled will execute correctly but may be less performant than code with the feature disabled.
-m[no-]sram-ecc Enable/disable generating code that assumes SRAM ECC is enabled/disabled.
-m[no-]wavefrontsize64 Control the default wavefront size used when generating code for kernels. When disabled native wavefront size 32 is used, when enabled wavefront size 64 is used.
-m[no-]cumode Control the default wavefront execution mode used when generating code for kernels. When disabled native WGP wavefront execution mode is used, when enabled CU wavefront execution mode is used (see Memory Model).

Address Spaces

The AMDGPU architecture supports a number of memory address spaces. The address space names use the OpenCL standard names, with some additions.

The AMDGPU address spaces correspond to target architecture specific LLVM address space numbers used in LLVM IR.

The AMDGPU address spaces are described in AMDGPU Address Spaces. Only 64-bit process address spaces are supported for the amdgcn target.

AMDGPU Address Spaces
      64-Bit Process Address Space
Address Space Name LLVM IR Address Space Number HSA Segment Name Hardware Name Address Size NULL Value
Generic 0 flat flat 64 0x0000000000000000
Global 1 global global 64 0x0000000000000000
Region 2 N/A GDS 32 not implemented for AMDHSA
Local 3 group LDS 32 0xFFFFFFFF
Constant 4 constant same as global 64 0x0000000000000000
Private 5 private scratch 32 0xFFFFFFFF
Constant 32-bit 6 TODO     0x00000000
Buffer Fat Pointer (experimental) 7 TODO      
Generic

The generic address space uses the hardware flat address support available in GFX7-GFX10. This uses two fixed ranges of virtual addresses (the private and local apertures), that are outside the range of addressable global memory, to map from a flat address to a private or local address.

FLAT instructions can take a flat address and access global, private (scratch), and group (LDS) memory depending on if the address is within one of the aperture ranges. Flat access to scratch requires hardware aperture setup and setup in the kernel prologue (see Flat Scratch). Flat access to LDS requires hardware aperture setup and M0 (GFX7-GFX8) register setup (see M0).

To convert between a private or group address space address (termed a segment address) and a flat address the base address of the corresponding aperture can be used. For GFX7-GFX8 these are available in the HSA AQL Queue the address of which can be obtained with Queue Ptr SGPR (see Initial Kernel Execution State). For GFX9-GFX10 the aperture base addresses are directly available as inline constant registers SRC_SHARED_BASE/LIMIT and SRC_PRIVATE_BASE/LIMIT. In 64-bit address mode the aperture sizes are 2^32 bytes and the base is aligned to 2^32 which makes it easier to convert from flat to segment or segment to flat.

A global address space address has the same value when used as a flat address so no conversion is needed.

Global and Constant

The global and constant address spaces both use global virtual addresses, which are the same virtual address space used by the CPU. However, some virtual addresses may only be accessible to the CPU, some only accessible by the GPU, and some by both.

Using the constant address space indicates that the data will not change during the execution of the kernel. This allows scalar read instructions to be used. The vector and scalar L1 caches are invalidated of volatile data before each kernel dispatch execution to allow constant memory to change values between kernel dispatches.

Region
The region address space uses the hardware Global Data Store (GDS). All wavefronts executing on the same device will access the same memory for any given region address. However, the same region address accessed by wavefronts executing on different devices will access different memory. It is higher performance than global memory. It is allocated by the runtime. The data store (DS) instructions can be used to access it.
Local
The local address space uses the hardware Local Data Store (LDS) which is automatically allocated when the hardware creates the wavefronts of a work-group, and freed when all the wavefronts of a work-group have terminated. All wavefronts belonging to the same work-group will access the same memory for any given local address. However, the same local address accessed by wavefronts belonging to different work-groups will access different memory. It is higher performance than global memory. The data store (DS) instructions can be used to access it.
Private

The private address space uses the hardware scratch memory support which automatically allocates memory when it creates a wavefront and frees it when a wavefronts terminates. The memory accessed by a lane of a wavefront for any given private address will be different to the memory accessed by another lane of the same or different wavefront for the same private address.

If a kernel dispatch uses scratch, then the hardware allocates memory from a pool of backing memory allocated by the runtime for each wavefront. The lanes of the wavefront access this using dword (4 byte) interleaving. The mapping used from private address to backing memory address is:

wavefront-scratch-base + ((private-address / 4) * wavefront-size * 4) + (wavefront-lane-id * 4) + (private-address % 4)

If each lane of a wavefront accesses the same private address, the interleaving results in adjacent dwords being accessed and hence requires fewer cache lines to be fetched.

There are different ways that the wavefront scratch base address is determined by a wavefront (see Initial Kernel Execution State).

Scratch memory can be accessed in an interleaved manner using buffer instructions with the scratch buffer descriptor and per wavefront scratch offset, by the scratch instructions, or by flat instructions. Multi-dword access is not supported except by flat and scratch instructions in GFX9-GFX10.

Constant 32-bit
TODO
Buffer Fat Pointer
The buffer fat pointer is an experimental address space that is currently unsupported in the backend. It exposes a non-integral pointer that is in the future intended to support the modelling of 128-bit buffer descriptors plus a 32-bit offset into the buffer (in total encapsulating a 160-bit pointer), allowing normal LLVM load/store/atomic operations to be used to model the buffer descriptors used heavily in graphics workloads targeting the backend.

Memory Scopes

This section provides LLVM memory synchronization scopes supported by the AMDGPU backend memory model when the target triple OS is amdhsa (see Memory Model and Target Triples).

The memory model supported is based on the HSA memory model [HSA] which is based in turn on HRF-indirect with scope inclusion [HRF]. The happens-before relation is transitive over the synchronizes-with relation independent of scope and synchronizes-with allows the memory scope instances to be inclusive (see table AMDHSA LLVM Sync Scopes).

This is different to the OpenCL [OpenCL] memory model which does not have scope inclusion and requires the memory scopes to exactly match. However, this is conservatively correct for OpenCL.

AMDHSA LLVM Sync Scopes
LLVM Sync Scope Description
none

The default: system.

Synchronizes with, and participates in modification and seq_cst total orderings with, other operations (except image operations) for all address spaces (except private, or generic that accesses private) provided the other operation’s sync scope is:

  • system.
  • agent and executed by a thread on the same agent.
  • workgroup and executed by a thread in the same work-group.
  • wavefront and executed by a thread in the same wavefront.
agent

Synchronizes with, and participates in modification and seq_cst total orderings with, other operations (except image operations) for all address spaces (except private, or generic that accesses private) provided the other operation’s sync scope is:

  • system or agent and executed by a thread on the same agent.
  • workgroup and executed by a thread in the same work-group.
  • wavefront and executed by a thread in the same wavefront.
workgroup

Synchronizes with, and participates in modification and seq_cst total orderings with, other operations (except image operations) for all address spaces (except private, or generic that accesses private) provided the other operation’s sync scope is:

  • system, agent or workgroup and executed by a thread in the same work-group.
  • wavefront and executed by a thread in the same wavefront.
wavefront

Synchronizes with, and participates in modification and seq_cst total orderings with, other operations (except image operations) for all address spaces (except private, or generic that accesses private) provided the other operation’s sync scope is:

  • system, agent, workgroup or wavefront and executed by a thread in the same wavefront.
singlethread Only synchronizes with and participates in modification and seq_cst total orderings with, other operations (except image operations) running in the same thread for all address spaces (for example, in signal handlers).
one-as Same as system but only synchronizes with other operations within the same address space.
agent-one-as Same as agent but only synchronizes with other operations within the same address space.
workgroup-one-as Same as workgroup but only synchronizes with other operations within the same address space.
wavefront-one-as Same as wavefront but only synchronizes with other operations within the same address space.
singlethread-one-as Same as singlethread but only synchronizes with other operations within the same address space.

LLVM IR Intrinsics

The AMDGPU backend implements the following LLVM IR intrinsics.

This section is WIP.

LLVM IR Attributes

The AMDGPU backend supports the following LLVM IR attributes.

AMDGPU LLVM IR Attributes
LLVM Attribute Description
“amdgpu-flat-work-group-size”=”min,max” Specify the minimum and maximum flat work group sizes that will be specified when the kernel is dispatched. Generated by the amdgpu_flat_work_group_size CLANG attribute [CLANG-ATTR].
“amdgpu-implicitarg-num-bytes”=”n” Number of kernel argument bytes to add to the kernel argument block size for the implicit arguments. This varies by OS and language (for OpenCL see OpenCL kernel implicit arguments appended for AMDHSA OS).
“amdgpu-num-sgpr”=”n” Specifies the number of SGPRs to use. Generated by the amdgpu_num_sgpr CLANG attribute [CLANG-ATTR].
“amdgpu-num-vgpr”=”n” Specifies the number of VGPRs to use. Generated by the amdgpu_num_vgpr CLANG attribute [CLANG-ATTR].
“amdgpu-waves-per-eu”=”m,n” Specify the minimum and maximum number of waves per execution unit. Generated by the amdgpu_waves_per_eu CLANG attribute [CLANG-ATTR].
“amdgpu-ieee” true/false. Specify whether the function expects the IEEE field of the mode register to be set on entry. Overrides the default for the calling convention.
“amdgpu-dx10-clamp” true/false. Specify whether the function expects the DX10_CLAMP field of the mode register to be set on entry. Overrides the default for the calling convention.

ELF Code Object

The AMDGPU backend generates a standard ELF [ELF] relocatable code object that can be linked by lld to produce a standard ELF shared code object which can be loaded and executed on an AMDGPU target.

Sections

An AMDGPU target ELF code object has the standard ELF sections which include:

AMDGPU ELF Sections
Name Type Attributes
.bss SHT_NOBITS SHF_ALLOC + SHF_WRITE
.data SHT_PROGBITS SHF_ALLOC + SHF_WRITE
.debug_* SHT_PROGBITS none
.dynamic SHT_DYNAMIC SHF_ALLOC
.dynstr SHT_PROGBITS SHF_ALLOC
.dynsym SHT_PROGBITS SHF_ALLOC
.got SHT_PROGBITS SHF_ALLOC + SHF_WRITE
.hash SHT_HASH SHF_ALLOC
.note SHT_NOTE none
.relaname SHT_RELA none
.rela.dyn SHT_RELA none
.rodata SHT_PROGBITS SHF_ALLOC
.shstrtab SHT_STRTAB none
.strtab SHT_STRTAB none
.symtab SHT_SYMTAB none
.text SHT_PROGBITS SHF_ALLOC + SHF_EXECINSTR

These sections have their standard meanings (see [ELF]) and are only generated if needed.

.debug*
The standard DWARF sections. See DWARF Debug Information for information on the DWARF produced by the AMDGPU backend.
.dynamic, .dynstr, .dynsym, .hash
The standard sections used by a dynamic loader.
.note
See Note Records for the note records supported by the AMDGPU backend.
.relaname, .rela.dyn

For relocatable code objects, name is the name of the section that the relocation records apply. For example, .rela.text is the section name for relocation records associated with the .text section.

For linked shared code objects, .rela.dyn contains all the relocation records from each of the relocatable code object’s .relaname sections.

See Relocation Records for the relocation records supported by the AMDGPU backend.

.text
The executable machine code for the kernels and functions they call. Generated as position independent code. See Code Conventions for information on conventions used in the isa generation.

Note Records

The AMDGPU backend code object contains ELF note records in the .note section. The set of generated notes and their semantics depend on the code object version; see Code Object V2 Note Records (–amdhsa-code-object-version=2) and Code Object V3 Note Records (–amdhsa-code-object-version=3).

As required by ELFCLASS32 and ELFCLASS64, minimal zero-byte padding must be generated after the name field to ensure the desc field is 4 byte aligned. In addition, minimal zero-byte padding must be generated to ensure the desc field size is a multiple of 4 bytes. The sh_addralign field of the .note section must be at least 4 to indicate at least 8 byte alignment.

Code Object V2 Note Records (–amdhsa-code-object-version=2)

Warning

Code Object V2 is not the default code object version emitted by this version of LLVM. For a description of the notes generated with the default configuration (Code Object V3) see Code Object V3 Note Records (–amdhsa-code-object-version=3).

The AMDGPU backend code object uses the following ELF note record in the .note section when compiling for Code Object V2 (–amdhsa-code-object-version=2).

Additional note records may be present, but any which are not documented here are deprecated and should not be used.

AMDGPU Code Object V2 ELF Note Records
Name Type Description
“AMD” NT_AMD_AMDGPU_HSA_METADATA <metadata null terminated string>
AMDGPU Code Object V2 ELF Note Record Enumeration Values
Name Value
reserved 0-9
NT_AMD_AMDGPU_HSA_METADATA 10
reserved 11
NT_AMD_AMDGPU_HSA_METADATA
Specifies extensible metadata associated with the code objects executed on HSA [HSA] compatible runtimes such as AMD’s ROCm [AMD-ROCm]. It is required when the target triple OS is amdhsa (see Target Triples). See Code Object V2 Metadata (–amdhsa-code-object-version=2) for the syntax of the code object metadata string.

Code Object V3 Note Records (–amdhsa-code-object-version=3)

The AMDGPU backend code object uses the following ELF note record in the .note section when compiling for Code Object V3 (–amdhsa-code-object-version=3).

Additional note records may be present, but any which are not documented here are deprecated and should not be used.

AMDGPU Code Object V3 ELF Note Records
Name Type Description
“AMDGPU” NT_AMDGPU_METADATA Metadata in Message Pack [MsgPack] binary format.
AMDGPU Code Object V3 ELF Note Record Enumeration Values
Name Value
reserved 0-31
NT_AMDGPU_METADATA 32
NT_AMDGPU_METADATA
Specifies extensible metadata associated with an AMDGPU code object. It is encoded as a map in the Message Pack [MsgPack] binary data format. See Code Object V3 Metadata (–amdhsa-code-object-version=3) for the map keys defined for the amdhsa OS.

Symbols

Symbols include the following:

AMDGPU ELF Symbols
Name Type Section Description
link-name STT_OBJECT
  • .data
  • .rodata
  • .bss
 Global variable
link-name.kd STT_OBJECT
  • .rodata
 Kernel descriptor
link-name STT_FUNC
  • .text
 Kernel entry point
link-name STT_OBJECT
  • SHN_AMDGPU_LDS
 Global variable in LDS
Global variable

Global variables both used and defined by the compilation unit.

If the symbol is defined in the compilation unit then it is allocated in the appropriate section according to if it has initialized data or is readonly.

If the symbol is external then its section is STN_UNDEF and the loader will resolve relocations using the definition provided by another code object or explicitly defined by the runtime.

If the symbol resides in local/group memory (LDS) then its section is the special processor specific section name SHN_AMDGPU_LDS, and the st_value field describes alignment requirements as it does for common symbols.

Kernel descriptor
Every HSA kernel has an associated kernel descriptor. It is the address of the kernel descriptor that is used in the AQL dispatch packet used to invoke the kernel, not the kernel entry point. The layout of the HSA kernel descriptor is defined in Kernel Descriptor.
Kernel entry point
Every HSA kernel also has a symbol for its machine code entry point.

Relocation Records

AMDGPU backend generates Elf64_Rela relocation records. Supported relocatable fields are:

word32
This specifies a 32-bit field occupying 4 bytes with arbitrary byte alignment. These values use the same byte order as other word values in the AMDGPU architecture.
word64
This specifies a 64-bit field occupying 8 bytes with arbitrary byte alignment. These values use the same byte order as other word values in the AMDGPU architecture.

Following notations are used for specifying relocation calculations:

A
Represents the addend used to compute the value of the relocatable field.
G
Represents the offset into the global offset table at which the relocation entry’s symbol will reside during execution.
GOT
Represents the address of the global offset table.
P
Represents the place (section offset for et_rel or address for et_dyn) of the storage unit being relocated (computed using r_offset).
S
Represents the value of the symbol whose index resides in the relocation entry. Relocations not using this must specify a symbol index of STN_UNDEF.
B
Represents the base address of a loaded executable or shared object which is the difference between the ELF address and the actual load address. Relocations using this are only valid in executable or shared objects.

The following relocation types are supported:

AMDGPU ELF Relocation Records
Relocation Type Kind Value Field Calculation
R_AMDGPU_NONE   0 none none
R_AMDGPU_ABS32_LO Static, Dynamic 1 word32 (S + A) & 0xFFFFFFFF
R_AMDGPU_ABS32_HI Static, Dynamic 2 word32 (S + A) >> 32
R_AMDGPU_ABS64 Static, Dynamic 3 word64 S + A
R_AMDGPU_REL32 Static 4 word32 S + A - P
R_AMDGPU_REL64 Static 5 word64 S + A - P
R_AMDGPU_ABS32 Static, Dynamic 6 word32 S + A
R_AMDGPU_GOTPCREL Static 7 word32 G + GOT + A - P
R_AMDGPU_GOTPCREL32_LO Static 8 word32 (G + GOT + A - P) & 0xFFFFFFFF
R_AMDGPU_GOTPCREL32_HI Static 9 word32 (G + GOT + A - P) >> 32
R_AMDGPU_REL32_LO Static 10 word32 (S + A - P) & 0xFFFFFFFF
R_AMDGPU_REL32_HI Static 11 word32 (S + A - P) >> 32
reserved   12    
R_AMDGPU_RELATIVE64 Dynamic 13 word64 B + A

R_AMDGPU_ABS32_LO and R_AMDGPU_ABS32_HI are only supported by the mesa3d OS, which does not support R_AMDGPU_ABS64.

There is no current OS loader support for 32-bit programs and so R_AMDGPU_ABS32 is not used.

Loaded Code Object Path Uniform Resource Identifier (URI)

The AMD GPU code object loader represents the path of the ELF shared object from which the code object was loaded as a textual Unifom Resource Identifier (URI). Note that the code object is the in memory loaded relocated form of the ELF shared object. Multiple code objects may be loaded at different memory addresses in the same process from the same ELF shared object.

The loaded code object path URI syntax is defined by the following BNF syntax:

code_object_uri ::== file_uri | memory_uri
file_uri        ::== "file://" file_path [ range_specifier ]
memory_uri      ::== "memory://" process_id range_specifier
range_specifier ::== [ "#" | "?" ] "offset=" number "&" "size=" number
file_path       ::== URI_ENCODED_OS_FILE_PATH
process_id      ::== DECIMAL_NUMBER
number          ::== HEX_NUMBER | DECIMAL_NUMBER | OCTAL_NUMBER
number
Is a C integral literal where hexadecimal values are prefixed by “0x” or “0X”, and octal values by “0”.
file_path
Is the file’s path specified as a URI encoded UTF-8 string. In URI encoding, every character that is not in the regular expression [a-zA-Z0-9/_.~-] is encoded as two uppercase hexadecimal digits proceeded by “%”. Directories in the path are separated by “/”.
offset
Is a 0-based byte offset to the start of the code object. For a file URI, it is from the start of the file specified by the file_path, and if omitted defaults to 0. For a memory URI, it is the memory address and is required.
size
Is the number of bytes in the code object. For a file URI, if omitted it defaults to the size of the file. It is required for a memory URI.
process_id
Is the identity of the process owning the memory. For Linux it is the C unsigned integral decimal literal for the process ID (PID).

For example:

file:///dir1/dir2/file1
file:///dir3/dir4/file2#offset=0x2000&size=3000
memory://1234#offset=0x20000&size=3000

DWARF Debug Information

Warning

This section describes provisional support for AMDGPU DWARF [DWARF] that is not currently fully implemented and is subject to change.

AMDGPU generates DWARF [DWARF] debugging information ELF sections (see ELF Code Object) which contain information that maps the code object executable code and data to the source language constructs. It can be used by tools such as debuggers and profilers. It uses features defined in DWARF Extensions For Heterogeneous Debugging that are made available in DWARF Version 4 and DWARF Version 5 as an LLVM vendor extension.

This section defines the AMDGPU target architecture specific DWARF mappings.

Register Identifier

This section defines the AMDGPU target architecture register numbers used in DWARF operation expressions (see DWARF Version 5 section 2.5 and DWARF Operation Expressions) and Call Frame Information instructions (see DWARF Version 5 section 6.4 and Call Frame Information).

A single code object can contain code for kernels that have different wavefront sizes. The vector registers and some scalar registers are based on the wavefront size. AMDGPU defines distinct DWARF registers for each wavefront size. This simplifies the consumer of the DWARF so that each register has a fixed size, rather than being dynamic according to the wavefront size mode. Similarly, distinct DWARF registers are defined for those registers that vary in size according to the process address size. This allows a consumer to treat a specific AMDGPU processor as a single architecture regardless of how it is configured at run time. The compiler explicitly specifies the DWARF registers that match the mode in which the code it is generating will be executed.

DWARF registers are encoded as numbers, which are mapped to architecture registers. The mapping for AMDGPU is defined in AMDGPU DWARF Register Mapping. All AMDGPU targets use the same mapping.

AMDGPU DWARF Register Mapping
DWARF Register AMDGPU Register Bit Size Description
0 PC_32 32 Program Counter (PC) when executing in a 32-bit process address space. Used in the CFI to describe the PC of the calling frame.
1 EXEC_MASK_32 32 Execution Mask Register when executing in wavefront 32 mode.
2-15 Reserved   Reserved for highly accessed registers using DWARF shortcut.
16 PC_64 64 Program Counter (PC) when executing in a 64-bit process address space. Used in the CFI to describe the PC of the calling frame.
17 EXEC_MASK_64 64 Execution Mask Register when executing in wavefront 64 mode.
18-31 Reserved   Reserved for highly accessed registers using DWARF shortcut.
32-95 SGPR0-SGPR63 32 Scalar General Purpose Registers.
96-127 Reserved   Reserved for frequently accessed registers using DWARF 1-byte ULEB.
128 STATUS 32 Status Register.
129-511 Reserved   Reserved for future Scalar Architectural Registers.
512 VCC_32 32 Vector Condition Code Register when executing in wavefront 32 mode.
513-1023 Reserved   Reserved for future Vector Architectural Registers when executing in wavefront 32 mode.
768 VCC_64 64 Vector Condition Code Register when executing in wavefront 64 mode.
769-1023 Reserved   Reserved for future Vector Architectural Registers when executing in wavefront 64 mode.
1024-1087 Reserved   Reserved for padding.
1088-1129 SGPR64-SGPR105 32 Scalar General Purpose Registers.
1130-1535 Reserved   Reserved for future Scalar General Purpose Registers.
1536-1791 VGPR0-VGPR255 32*32 Vector General Purpose Registers when executing in wavefront 32 mode.
1792-2047 Reserved   Reserved for future Vector General Purpose Registers when executing in wavefront 32 mode.
2048-2303 AGPR0-AGPR255 32*32 Vector Accumulation Registers when executing in wavefront 32 mode.
2304-2559 Reserved   Reserved for future Vector Accumulation Registers when executing in wavefront 32 mode.
2560-2815 VGPR0-VGPR255 64*32 Vector General Purpose Registers when executing in wavefront 64 mode.
2816-3071 Reserved   Reserved for future Vector General Purpose Registers when executing in wavefront 64 mode.
3072-3327 AGPR0-AGPR255 64*32 Vector Accumulation Registers when executing in wavefront 64 mode.
3328-3583 Reserved   Reserved for future Vector Accumulation Registers when executing in wavefront 64 mode.

The vector registers are represented as the full size for the wavefront. They are organized as consecutive dwords (32-bits), one per lane, with the dword at the least significant bit position corresponding to lane 0 and so forth. DWARF location expressions involving the DW_OP_LLVM_offset and DW_OP_LLVM_push_lane operations are used to select the part of the vector register corresponding to the lane that is executing the current thread of execution in languages that are implemented using a SIMD or SIMT execution model.

If the wavefront size is 32 lanes then the wavefront 32 mode register definitions are used. If the wavefront size is 64 lanes then the wavefront 64 mode register definitions are used. Some AMDGPU targets support executing in both wavefront 32 and wavefront 64 mode. The register definitions corresponding to the wavefront mode of the generated code will be used.

If code is generated to execute in a 32-bit process address space, then the 32-bit process address space register definitions are used. If code is generated to execute in a 64-bit process address space, then the 64-bit process address space register definitions are used. The amdgcn target only supports the 64-bit process address space.

Address Class Identifier

The DWARF address class represents the source language memory space. See DWARF Version 5 section 2.12 which is updated by the DWARF Extensions For Heterogeneous Debugging section Segmented Addresses.

The DWARF address class mapping used for AMDGPU is defined in AMDGPU DWARF Address Class Mapping.

AMDGPU DWARF Address Class Mapping
DWARF AMDGPU
Address Class Name Value Address Space
DW_ADDR_none 0x0000 Generic (Flat)
DW_ADDR_LLVM_global 0x0001 Global
DW_ADDR_LLVM_constant 0x0002 Global
DW_ADDR_LLVM_group 0x0003 Local (group/LDS)
DW_ADDR_LLVM_private 0x0004 Private (Scratch)
DW_ADDR_AMDGPU_region 0x8000 Region (GDS)

The DWARF address class values defined in the DWARF Extensions For Heterogeneous Debugging section Segmented Addresses are used.

In addition, DW_ADDR_AMDGPU_region is encoded as a vendor extension. This is available for use for the AMD extension for access to the hardware GDS memory which is scratchpad memory allocated per device.

For AMDGPU if no DW_AT_address_class attribute is present, then the default address class of DW_ADDR_none is used.

See Address Space Identifier for information on the AMDGPU mapping of DWARF address classes to DWARF address spaces, including address size and NULL value.

Address Space Identifier

DWARF address spaces correspond to target architecture specific linear addressable memory areas. See DWARF Version 5 section 2.12 and DWARF Extensions For Heterogeneous Debugging section Segmented Addresses.

The DWARF address space mapping used for AMDGPU is defined in AMDGPU DWARF Address Space Mapping.

AMDGPU DWARF Address Space Mapping
DWARF     AMDGPU Notes
Address Space Name Value Address Bit Size Address Space  
  64-bit process address space 32-bit process address space    
DW_ASPACE_none 0x00 64 32 Global default address space
DW_ASPACE_AMDGPU_generic 0x01 64 32 Generic (Flat)  
DW_ASPACE_AMDGPU_region 0x02 32 32 Region (GDS)  
DW_ASPACE_AMDGPU_local 0x03 32 32 Local (group/LDS)  
Reserved 0x04        
DW_ASPACE_AMDGPU_private_lane 0x05 32 32 Private (Scratch) focused lane
DW_ASPACE_AMDGPU_private_wave 0x06 32 32 Private (Scratch) unswizzled wavefront

See Address Spaces for information on the AMDGPU address spaces including address size and NULL value.

The DW_ASPACE_none address space is the default target architecture address space used in DWARF operations that do not specify an address space. It therefore has to map to the global address space so that the DW_OP_addr* and related operations can refer to addresses in the program code.

The DW_ASPACE_AMDGPU_generic address space allows location expressions to specify the flat address space. If the address corresponds to an address in the local address space, then it corresponds to the wavefront that is executing the focused thread of execution. If the address corresponds to an address in the private address space, then it corresponds to the lane that is executing the focused thread of execution for languages that are implemented using a SIMD or SIMT execution model.

Note

CUDA-like languages such as HIP that do not have address spaces in the language type system, but do allow variables to be allocated in different address spaces, need to explicitly specify the DW_ASPACE_AMDGPU_generic address space in the DWARF expression operations as the default address space is the global address space.

The DW_ASPACE_AMDGPU_local address space allows location expressions to specify the local address space corresponding to the wavefront that is executing the focused thread of execution.

The DW_ASPACE_AMDGPU_private_lane address space allows location expressions to specify the private address space corresponding to the lane that is executing the focused thread of execution for languages that are implemented using a SIMD or SIMT execution model.

The DW_ASPACE_AMDGPU_private_wave address space allows location expressions to specify the unswizzled private address space corresponding to the wavefront that is executing the focused thread of execution. The wavefront view of private memory is the per wavefront unswizzled backing memory layout defined in Address Spaces, such that address 0 corresponds to the first location for the backing memory of the wavefront (namely the address is not offset by wavefront-scratch-base). The following formula can be used to convert from a DW_ASPACE_AMDGPU_private_lane address to a DW_ASPACE_AMDGPU_private_wave address:

private-address-wavefront =
  ((private-address-lane / 4) * wavefront-size * 4) +
  (wavefront-lane-id * 4) + (private-address-lane % 4)

If the DW_ASPACE_AMDGPU_private_lane address is dword aligned, and the start of the dwords for each lane starting with lane 0 is required, then this simplifies to:

private-address-wavefront =
  private-address-lane * wavefront-size

A compiler can use the DW_ASPACE_AMDGPU_private_wave address space to read a complete spilled vector register back into a complete vector register in the CFI. The frame pointer can be a private lane address which is dword aligned, which can be shifted to multiply by the wavefront size, and then used to form a private wavefront address that gives a location for a contiguous set of dwords, one per lane, where the vector register dwords are spilled. The compiler knows the wavefront size since it generates the code. Note that the type of the address may have to be converted as the size of a DW_ASPACE_AMDGPU_private_lane address may be smaller than the size of a DW_ASPACE_AMDGPU_private_wave address.

Lane identifier

DWARF lane identifies specify a target architecture lane position for hardware that executes in a SIMD or SIMT manner, and on which a source language maps its threads of execution onto those lanes. The DWARF lane identifier is pushed by the DW_OP_LLVM_push_lane DWARF expression operation. See DWARF Version 5 section 2.5 which is updated by DWARF Extensions For Heterogeneous Debugging section DWARF Operation Expressions.

For AMDGPU, the lane identifier corresponds to the hardware lane ID of a wavefront. It is numbered from 0 to the wavefront size minus 1.

Operation Expressions

DWARF expressions are used to compute program values and the locations of program objects. See DWARF Version 5 section 2.5 and DWARF Operation Expressions.

DWARF location descriptions describe how to access storage which includes memory and registers. When accessing storage on AMDGPU, bytes are ordered with least significant bytes first, and bits are ordered within bytes with least significant bits first.

For AMDGPU CFI expressions, DW_OP_LLVM_select_bit_piece is used to describe unwinding vector registers that are spilled under the execution mask to memory: the zero-single location description is the vector register, and the one-single location description is the spilled memory location description. The DW_OP_LLVM_form_aspace_address is used to specify the address space of the memory location description.

In AMDGPU expressions, DW_OP_LLVM_select_bit_piece is used by the DW_AT_LLVM_lane_pc attribute expression where divergent control flow is controlled by the execution mask. An undefined location description together with DW_OP_LLVM_extend is used to indicate the lane was not active on entry to the subprogram. See DW_AT_LLVM_lane_pc for an example.

Debugger Information Entry Attributes

This section describes how certain debugger information entry attributes are used by AMDGPU. See the sections in DWARF Version 5 section 2 which are updated by DWARF Extensions For Heterogeneous Debugging section Debugging Information Entry Attributes.

DW_AT_LLVM_lane_pc

For AMDGPU, the DW_AT_LLVM_lane_pc attribute is used to specify the program location of the separate lanes of a SIMT thread.

If the lane is an active lane then this will be the same as the current program location.

If the lane is inactive, but was active on entry to the subprogram, then this is the program location in the subprogram at which execution of the lane is conceptual positioned.

If the lane was not active on entry to the subprogram, then this will be the undefined location. A client debugger can check if the lane is part of a valid work-group by checking that the lane is in the range of the associated work-group within the grid, accounting for partial work-groups. If it is not, then the debugger can omit any information for the lane. Otherwise, the debugger may repeatedly unwind the stack and inspect the DW_AT_LLVM_lane_pc of the calling subprogram until it finds a non-undefined location. Conceptually the lane only has the call frames that it has a non-undefined DW_AT_LLVM_lane_pc.

The following example illustrates how the AMDGPU backend can generate a DWARF location list expression for the nested IF/THEN/ELSE structures of the following subprogram pseudo code for a target with 64 lanes per wavefront.

 1 SUBPROGRAM X
 2 BEGIN
 3   a;
 4   IF (c1) THEN
 5     b;
 6     IF (c2) THEN
 7       c;
 8     ELSE
 9       d;
10     ENDIF
11     e;
12   ELSE
13     f;
14   ENDIF
15   g;
16 END

The AMDGPU backend may generate the following pseudo LLVM MIR to manipulate the execution mask (EXEC) to linearize the control flow. The condition is evaluated to make a mask of the lanes for which the condition evaluates to true. First the THEN region is executed by setting the EXEC mask to the logical AND of the current EXEC mask with the condition mask. Then the ELSE region is executed by negating the EXEC mask and logical AND of the saved EXEC mask at the start of the region. After the IF/THEN/ELSE region the EXEC mask is restored to the value it had at the beginning of the region. This is shown below. Other approaches are possible, but the basic concept is the same.

 1 $lex_start:
 2   a;
 3   %1 = EXEC
 4   %2 = c1
 5 $lex_1_start:
 6   EXEC = %1 & %2
 7 $if_1_then:
 8     b;
 9     %3 = EXEC
10     %4 = c2
11 $lex_1_1_start:
12     EXEC = %3 & %4
13 $lex_1_1_then:
14       c;
15     EXEC = ~EXEC & %3
16 $lex_1_1_else:
17       d;
18     EXEC = %3
19 $lex_1_1_end:
20     e;
21   EXEC = ~EXEC & %1
22 $lex_1_else:
23     f;
24   EXEC = %1
25 $lex_1_end:
26   g;
27 $lex_end:

To create the DWARF location list expression that defines the location description of a vector of lane program locations, the LLVM MIR DBG_VALUE pseudo instruction can be used to annotate the linearized control flow. This can be done by defining an artificial variable for the lane PC. The DWARF location list expression created for it is used as the value of the DW_AT_LLVM_lane_pc attribute on the subprogram’s debugger information entry.

A DWARF procedure is defined for each well nested structured control flow region which provides the conceptual lane program location for a lane if it is not active (namely it is divergent). The DWARF operation expression for each region conceptually inherits the value of the immediately enclosing region and modifies it according to the semantics of the region.

For an IF/THEN/ELSE region the divergent program location is at the start of the region for the THEN region since it is executed first. For the ELSE region the divergent program location is at the end of the IF/THEN/ELSE region since the THEN region has completed.

The lane PC artificial variable is assigned at each region transition. It uses the immediately enclosing region’s DWARF procedure to compute the program location for each lane assuming they are divergent, and then modifies the result by inserting the current program location for each lane that the EXEC mask indicates is active.

By having separate DWARF procedures for each region, they can be reused to define the value for any nested region. This reduces the total size of the DWARF operation expressions.

The following provides an example using pseudo LLVM MIR.

  1 $lex_start:
  2   DEFINE_DWARF %__uint_64 = DW_TAG_base_type[
  3     DW_AT_name = "__uint64";
  4     DW_AT_byte_size = 8;
  5     DW_AT_encoding = DW_ATE_unsigned;
  6   ];
  7   DEFINE_DWARF %__active_lane_pc = DW_TAG_dwarf_procedure[
  8     DW_AT_name = "__active_lane_pc";
  9     DW_AT_location = [
 10       DW_OP_regx PC;
 11       DW_OP_LLVM_extend 64, 64;
 12       DW_OP_regval_type EXEC, %uint_64;
 13       DW_OP_LLVM_select_bit_piece 64, 64;
 14     ];
 15   ];
 16   DEFINE_DWARF %__divergent_lane_pc = DW_TAG_dwarf_procedure[
 17     DW_AT_name = "__divergent_lane_pc";
 18     DW_AT_location = [
 19       DW_OP_LLVM_undefined;
 20       DW_OP_LLVM_extend 64, 64;
 21     ];
 22   ];
 23   DBG_VALUE $noreg, $noreg, %DW_AT_LLVM_lane_pc, DIExpression[
 24     DW_OP_call_ref %__divergent_lane_pc;
 25     DW_OP_call_ref %__active_lane_pc;
 26   ];
 27   a;
 28   %1 = EXEC;
 29   DBG_VALUE %1, $noreg, %__lex_1_save_exec;
 30   %2 = c1;
 31 $lex_1_start:
 32   EXEC = %1 & %2;
 33 $lex_1_then:
 34     DEFINE_DWARF %__divergent_lane_pc_1_then = DW_TAG_dwarf_procedure[
 35       DW_AT_name = "__divergent_lane_pc_1_then";
 36       DW_AT_location = DIExpression[
 37         DW_OP_call_ref %__divergent_lane_pc;
 38         DW_OP_addrx &lex_1_start;
 39         DW_OP_stack_value;
 40         DW_OP_LLVM_extend 64, 64;
 41         DW_OP_call_ref %__lex_1_save_exec;
 42         DW_OP_deref_type 64, %__uint_64;
 43         DW_OP_LLVM_select_bit_piece 64, 64;
 44       ];
 45     ];
 46     DBG_VALUE $noreg, $noreg, %DW_AT_LLVM_lane_pc, DIExpression[
 47       DW_OP_call_ref %__divergent_lane_pc_1_then;
 48       DW_OP_call_ref %__active_lane_pc;
 49     ];
 50     b;
 51     %3 = EXEC;
 52     DBG_VALUE %3, %__lex_1_1_save_exec;
 53     %4 = c2;
 54 $lex_1_1_start:
 55     EXEC = %3 & %4;
 56 $lex_1_1_then:
 57       DEFINE_DWARF %__divergent_lane_pc_1_1_then = DW_TAG_dwarf_procedure[
 58         DW_AT_name = "__divergent_lane_pc_1_1_then";
 59         DW_AT_location = DIExpression[
 60           DW_OP_call_ref %__divergent_lane_pc_1_then;
 61           DW_OP_addrx &lex_1_1_start;
 62           DW_OP_stack_value;
 63           DW_OP_LLVM_extend 64, 64;
 64           DW_OP_call_ref %__lex_1_1_save_exec;
 65           DW_OP_deref_type 64, %__uint_64;
 66           DW_OP_LLVM_select_bit_piece 64, 64;
 67         ];
 68       ];
 69       DBG_VALUE $noreg, $noreg, %DW_AT_LLVM_lane_pc, DIExpression[
 70         DW_OP_call_ref %__divergent_lane_pc_1_1_then;
 71         DW_OP_call_ref %__active_lane_pc;
 72       ];
 73       c;
 74     EXEC = ~EXEC & %3;
 75 $lex_1_1_else:
 76       DEFINE_DWARF %__divergent_lane_pc_1_1_else = DW_TAG_dwarf_procedure[
 77         DW_AT_name = "__divergent_lane_pc_1_1_else";
 78         DW_AT_location = DIExpression[
 79           DW_OP_call_ref %__divergent_lane_pc_1_then;
 80           DW_OP_addrx &lex_1_1_end;
 81           DW_OP_stack_value;
 82           DW_OP_LLVM_extend 64, 64;
 83           DW_OP_call_ref %__lex_1_1_save_exec;
 84           DW_OP_deref_type 64, %__uint_64;
 85           DW_OP_LLVM_select_bit_piece 64, 64;
 86         ];
 87       ];
 88       DBG_VALUE $noreg, $noreg, %DW_AT_LLVM_lane_pc, DIExpression[
 89         DW_OP_call_ref %__divergent_lane_pc_1_1_else;
 90         DW_OP_call_ref %__active_lane_pc;
 91       ];
 92       d;
 93     EXEC = %3;
 94 $lex_1_1_end:
 95     DBG_VALUE $noreg, $noreg, %DW_AT_LLVM_lane_pc, DIExpression[
 96       DW_OP_call_ref %__divergent_lane_pc;
 97       DW_OP_call_ref %__active_lane_pc;
 98     ];
 99     e;
100   EXEC = ~EXEC & %1;
101 $lex_1_else:
102     DEFINE_DWARF %__divergent_lane_pc_1_else = DW_TAG_dwarf_procedure[
103       DW_AT_name = "__divergent_lane_pc_1_else";
104       DW_AT_location = DIExpression[
105         DW_OP_call_ref %__divergent_lane_pc;
106         DW_OP_addrx &lex_1_end;
107         DW_OP_stack_value;
108         DW_OP_LLVM_extend 64, 64;
109         DW_OP_call_ref %__lex_1_save_exec;
110         DW_OP_deref_type 64, %__uint_64;
111         DW_OP_LLVM_select_bit_piece 64, 64;
112       ];
113     ];
114     DBG_VALUE $noreg, $noreg, %DW_AT_LLVM_lane_pc, DIExpression[
115       DW_OP_call_ref %__divergent_lane_pc_1_else;
116       DW_OP_call_ref %__active_lane_pc;
117     ];
118     f;
119   EXEC = %1;
120 $lex_1_end:
121   DBG_VALUE $noreg, $noreg, %DW_AT_LLVM_lane_pc DIExpression[
122     DW_OP_call_ref %__divergent_lane_pc;
123     DW_OP_call_ref %__active_lane_pc;
124   ];
125   g;
126 $lex_end:

The DWARF procedure %__active_lane_pc is used to update the lane pc elements that are active, with the current program location.

Artificial variables %__lex_1_save_exec and %__lex_1_1_save_exec are created for the execution masks saved on entry to a region. Using the DBG_VALUE pseudo instruction, location list entries will be created that describe where the artificial variables are allocated at any given program location. The compiler may allocate them to registers or spill them to memory.

The DWARF procedures for each region use the values of the saved execution mask artificial variables to only update the lanes that are active on entry to the region. All other lanes retain the value of the enclosing region where they were last active. If they were not active on entry to the subprogram, then will have the undefined location description.

Other structured control flow regions can be handled similarly. For example, loops would set the divergent program location for the region at the end of the loop. Any lanes active will be in the loop, and any lanes not active must have exited the loop.

An IF/THEN/ELSEIF/ELSEIF/... region can be treated as a nest of IF/THEN/ELSE regions.

The DWARF procedures can use the active lane artificial variable described in DW_AT_LLVM_active_lane rather than the actual EXEC mask in order to support whole or quad wavefront mode.

DW_AT_LLVM_active_lane

The DW_AT_LLVM_active_lane attribute on a subprogram debugger information entry is used to specify the lanes that are conceptually active for a SIMT thread.

The execution mask may be modified to implement whole or quad wavefront mode operations. For example, all lanes may need to temporarily be made active to execute a whole wavefront operation. Such regions would save the EXEC mask, update it to enable the necessary lanes, perform the operations, and then restore the EXEC mask from the saved value. While executing the whole wavefront region, the conceptual execution mask is the saved value, not the EXEC value.

This is handled by defining an artificial variable for the active lane mask. The active lane mask artificial variable would be the actual EXEC mask for normal regions, and the saved execution mask for regions where the mask is temporarily updated. The location list expression created for this artificial variable is used to define the value of the DW_AT_LLVM_active_lane attribute.

DW_AT_LLVM_augmentation

For AMDGPU, the DW_AT_LLVM_augmentation attribute of a compilation unit debugger information entry has the following value for the augmentation string:

[amdgpu:v0.0]

The “vX.Y” specifies the major X and minor Y version number of the AMDGPU extensions used in the DWARF of the compilation unit. The version number conforms to [SEMVER].

Call Frame Information

DWARF Call Frame Information (CFI) describes how a consumer can virtually unwind call frames in a running process or core dump. See DWARF Version 5 section 6.4 and Call Frame Information.

For AMDGPU, the Common Information Entry (CIE) fields have the following values:

  1. augmentation string contains the following null-terminated UTF-8 string:

    [amd:v0.0]

    The vX.Y specifies the major X and minor Y version number of the AMDGPU extensions used in this CIE or to the FDEs that use it. The version number conforms to [SEMVER].

  2. address_size for the Global address space is defined in Address Space Identifier.

  3. segment_selector_size is 0 as AMDGPU does not use a segment selector.

  4. code_alignment_factor is 4 bytes.

  5. data_alignment_factor is 4 bytes.

  6. return_address_register is PC_32 for 32-bit processes and PC_64 for 64-bit processes defined in Register Identifier.

  7. initial_instructions Since a subprogram X with fewer registers can be called from subprogram Y that has more allocated, X will not change any of the extra registers as it cannot access them. Therefore, the default rule for all columns is same value.

For AMDGPU the register number follows the numbering defined in Register Identifier.

For AMDGPU the instructions are variable size. A consumer can subtract 1 from the return address to get the address of a byte within the call site instructions. See DWARF Version 5 section 6.4.4.

Accelerated Access

See DWARF Version 5 section 6.1.

Lookup By Name Section Header

See DWARF Version 5 section 6.1.1.4.1 and Lookup By Name.

For AMDGPU the lookup by name section header table:

augmentation_string_size (uword)

Set to the length of the augmentation_string value which is always a multiple of 4.

augmentation_string (sequence of UTF-8 characters)

Contains the following UTF-8 string null padded to a multiple of 4 bytes:

[amdgpu:v0.0]

The “vX.Y” specifies the major X and minor Y version number of the AMDGPU extensions used in the DWARF of this index. The version number conforms to [SEMVER].

Note

This is different to the DWARF Version 5 definition that requires the first 4 characters to be the vendor ID. But this is consistent with the other augmentation strings and does allow multiple vendor contributions. However, backwards compatibility may be more desirable.

Lookup By Address Section Header

See DWARF Version 5 section 6.1.2.

For AMDGPU the lookup by address section header table:

address_size (ubyte)

Match the address size for the Global address space defined in Address Space Identifier.

segment_selector_size (ubyte)

AMDGPU does not use a segment selector so this is 0. The entries in the .debug_aranges do not have a segment selector.

Line Number Information

See DWARF Version 5 section 6.2 and Line Number Information.

AMDGPU does not use the isa state machine registers and always sets it to 0. The instruction set must be obtained from the ELF file header e_flags field in the EF_AMDGPU_MACH bit position (see ELF Header). See DWARF Version 5 section 6.2.2.

For AMDGPU the line number program header fields have the following values (see DWARF Version 5 section 6.2.4):

address_size (ubyte)
Matches the address size for the Global address space defined in Address Space Identifier.
segment_selector_size (ubyte)
AMDGPU does not use a segment selector so this is 0.
minimum_instruction_length (ubyte)
For GFX9-GFX10 this is 4.
maximum_operations_per_instruction (ubyte)
For GFX9-GFX10 this is 1.

Source text for online-compiled programs (for example, those compiled by the OpenCL language runtime) may be embedded into the DWARF Version 5 line table. See DWARF Version 5 section 6.2.4.1 which is updated by DWARF Extensions For Heterogeneous Debugging section DW_LNCT_LLVM_source.

The Clang option used to control source embedding in AMDGPU is defined in AMDGPU Clang Debug Options.

AMDGPU Clang Debug Options
Debug Flag Description
-g[no-]embed-source Enable/disable embedding source text in DWARF debug sections. Useful for environments where source cannot be written to disk, such as when performing online compilation.

For example:

-gembed-source
Enable the embedded source.
-gno-embed-source
Disable the embedded source.

32-Bit and 64-Bit DWARF Formats

See DWARF Version 5 section 7.4 and 32-Bit and 64-Bit DWARF Formats.

For AMDGPU:

  • For the amdgcn target architecture only the 64-bit process address space is supported.
  • The producer can generate either 32-bit or 64-bit DWARF format. LLVM generates the 32-bit DWARF format.

Unit Headers

For AMDGPU the following values apply for each of the unit headers described in DWARF Version 5 sections 7.5.1.1, 7.5.1.2, and 7.5.1.3:

address_size (ubyte)
Matches the address size for the Global address space defined in Address Space Identifier.

Code Conventions

This section provides code conventions used for each supported target triple OS (see Target Triples).

AMDHSA

This section provides code conventions used when the target triple OS is amdhsa (see Target Triples).

Code Object Target Identification

The AMDHSA OS uses the following syntax to specify the code object target as a single string:

<Architecture>-<Vendor>-<OS>-<Environment>-<Processor><Target Features>

Where:

  • <Architecture>, <Vendor>, <OS> and <Environment> are the same as the Target Triple (see Target Triples).
  • <Processor> is the same as the Processor (see Processors).
  • <Target Features> is a list of the enabled Target Features (see Target Features), each prefixed by a plus, that apply to Processor. The list must be in the same order as listed in the table AMDGPU Target Features. Note that Target Features must be included in the list if they are enabled even if that is the default for Processor.

For example:

"amdgcn-amd-amdhsa--gfx902+xnack"

Code Object Metadata

The code object metadata specifies extensible metadata associated with the code objects executed on HSA [HSA] compatible runtimes such as AMD’s ROCm [AMD-ROCm]. The encoding and semantics of this metadata depends on the code object version; see Code Object V2 Metadata (–amdhsa-code-object-version=2) and Code Object V3 Metadata (–amdhsa-code-object-version=3).

Code object metadata is specified in a note record (see Note Records) and is required when the target triple OS is amdhsa (see Target Triples). It must contain the minimum information necessary to support the ROCM kernel queries. For example, the segment sizes needed in a dispatch packet. In addition, a high-level language runtime may require other information to be included. For example, the AMD OpenCL runtime records kernel argument information.

Code Object V2 Metadata (–amdhsa-code-object-version=2)

Warning

Code Object V2 is not the default code object version emitted by this version of LLVM. For a description of the metadata generated with the default configuration (Code Object V3) see Code Object V3 Metadata (–amdhsa-code-object-version=3).

Code object V2 metadata is specified by the NT_AMD_AMDGPU_METADATA note record (see Code Object V2 Note Records (–amdhsa-code-object-version=2)).

The metadata is specified as a YAML formatted string (see [YAML] and YAML I/O).

The metadata is represented as a single YAML document comprised of the mapping defined in table AMDHSA Code Object V2 Metadata Map and referenced tables.

For boolean values, the string values of false and true are used for false and true respectively.

Additional information can be added to the mappings. To avoid conflicts, any non-AMD key names should be prefixed by “vendor-name.”.

AMDHSA Code Object V2 Metadata Map
String Key Value Type Required? Description
“Version” sequence of 2 integers Required
  • The first integer is the major version. Currently 1.
  • The second integer is the minor version. Currently 0.
“Printf” sequence of strings  

Each string is encoded information about a printf function call. The encoded information is organized as fields separated by colon (‘:’):

ID:N:S[0]:S[1]:...:S[N-1]:FormatString

where:

ID
A 32-bit integer as a unique id for each printf function call
N
A 32-bit integer equal to the number of arguments of printf function call minus 1
S[i] (where i = 0, 1, … , N-1)
32-bit integers for the size in bytes of the i-th FormatString argument of the printf function call
FormatString
The format string passed to the printf function call.
“Kernels” sequence of mapping Required Sequence of the mappings for each kernel in the code object. See AMDHSA Code Object V2 Kernel Metadata Map for the definition of the mapping.
AMDHSA Code Object V2 Kernel Metadata Map
String Key Value Type Required? Description
“Name” string Required Source name of the kernel.
“SymbolName” string Required Name of the kernel descriptor ELF symbol.
“Language” string  

Source language of the kernel. Values include:

  • “OpenCL C”
  • “OpenCL C++”
  • “HCC”
  • “OpenMP”
“LanguageVersion” sequence of 2 integers  
  • The first integer is the major version.
  • The second integer is the minor version.
“Attrs” mapping   Mapping of kernel attributes. See AMDHSA Code Object V2 Kernel Attribute Metadata Map for the mapping definition.
“Args” sequence of mapping   Sequence of mappings of the kernel arguments. See AMDHSA Code Object V2 Kernel Argument Metadata Map for the definition of the mapping.
“CodeProps” mapping   Mapping of properties related to the kernel code. See AMDHSA Code Object V2 Kernel Code Properties Metadata Map for the mapping definition.
AMDHSA Code Object V2 Kernel Attribute Metadata Map
String Key Value Type Required? Description
“ReqdWorkGroupSize” sequence of 3 integers  

If not 0, 0, 0 then all values must be >=1 and the dispatch work-group size X, Y, Z must correspond to the specified values. Defaults to 0, 0, 0.

Corresponds to the OpenCL reqd_work_group_size attribute.
“WorkGroupSizeHint” sequence of 3 integers  

The dispatch work-group size X, Y, Z is likely to be the specified values.

Corresponds to the OpenCL work_group_size_hint attribute.
“VecTypeHint” string  

The name of a scalar or vector type.

Corresponds to the OpenCL vec_type_hint attribute.
“RuntimeHandle” string   The external symbol name associated with a kernel. OpenCL runtime allocates a global buffer for the symbol and saves the kernel’s address to it, which is used for device side enqueueing. Only available for device side enqueued kernels.
AMDHSA Code Object V2 Kernel Argument Metadata Map
String Key Value Type Required? Description
“Name” string   Kernel argument name.
“TypeName” string   Kernel argument type name.
“Size” integer Required Kernel argument size in bytes.
“Align” integer Required Kernel argument alignment in bytes. Must be a power of two.
“ValueKind” string Required

Kernel argument kind that specifies how to set up the corresponding argument. Values include:

“ByValue”
The argument is copied directly into the kernarg.
“GlobalBuffer”
A global address space pointer to the buffer data is passed in the kernarg.
“DynamicSharedPointer”
A group address space pointer to dynamically allocated LDS is passed in the kernarg.
“Sampler”
A global address space pointer to a S# is passed in the kernarg.
“Image”
A global address space pointer to a T# is passed in the kernarg.
“Pipe”
A global address space pointer to an OpenCL pipe is passed in the kernarg.
“Queue”
A global address space pointer to an OpenCL device enqueue queue is passed in the kernarg.
“HiddenGlobalOffsetX”
The OpenCL grid dispatch global offset for the X dimension is passed in the kernarg.
“HiddenGlobalOffsetY”
The OpenCL grid dispatch global offset for the Y dimension is passed in the kernarg.
“HiddenGlobalOffsetZ”
The OpenCL grid dispatch global offset for the Z dimension is passed in the kernarg.
“HiddenNone”
An argument that is not used by the kernel. Space needs to be left for it, but it does not need to be set up.
“HiddenPrintfBuffer”
A global address space pointer to the runtime printf buffer is passed in kernarg.
“HiddenHostcallBuffer”
A global address space pointer to the runtime hostcall buffer is passed in kernarg.
“HiddenDefaultQueue”
A global address space pointer to the OpenCL device enqueue queue that should be used by the kernel by default is passed in the kernarg.
“HiddenCompletionAction”
A global address space pointer to help link enqueued kernels into the ancestor tree for determining when the parent kernel has finished.
“HiddenMultiGridSyncArg”
A global address space pointer for multi-grid synchronization is passed in the kernarg.
“ValueType” string   Unused and deprecated. This should no longer be emitted, but is accepted for compatibility.
“PointeeAlign” integer   Alignment in bytes of pointee type for pointer type kernel argument. Must be a power of 2. Only present if “ValueKind” is “DynamicSharedPointer”.
“AddrSpaceQual” string  

Kernel argument address space qualifier. Only present if “ValueKind” is “GlobalBuffer” or “DynamicSharedPointer”. Values are:

  • “Private”
  • “Global”
  • “Constant”
  • “Local”
  • “Generic”
  • “Region”
“AccQual” string  

Kernel argument access qualifier. Only present if “ValueKind” is “Image” or “Pipe”. Values are:

  • “ReadOnly”
  • “WriteOnly”
  • “ReadWrite”
“ActualAccQual” string  

The actual memory accesses performed by the kernel on the kernel argument. Only present if “ValueKind” is “GlobalBuffer”, “Image”, or “Pipe”. This may be more restrictive than indicated by “AccQual” to reflect what the kernel actual does. If not present then the runtime must assume what is implied by “AccQual” and “IsConst”. Values are:

  • “ReadOnly”
  • “WriteOnly”
  • “ReadWrite”
“IsConst” boolean   Indicates if the kernel argument is const qualified. Only present if “ValueKind” is “GlobalBuffer”.
“IsRestrict” boolean   Indicates if the kernel argument is restrict qualified. Only present if “ValueKind” is “GlobalBuffer”.
“IsVolatile” boolean   Indicates if the kernel argument is volatile qualified. Only present if “ValueKind” is “GlobalBuffer”.
“IsPipe” boolean   Indicates if the kernel argument is pipe qualified. Only present if “ValueKind” is “Pipe”.
AMDHSA Code Object V2 Kernel Code Properties Metadata Map
String Key Value Type Required? Description
“KernargSegmentSize” integer Required The size in bytes of the kernarg segment that holds the values of the arguments to the kernel.
“GroupSegmentFixedSize” integer Required The amount of group segment memory required by a work-group in bytes. This does not include any dynamically allocated group segment memory that may be added when the kernel is dispatched.
“PrivateSegmentFixedSize” integer Required The amount of fixed private address space memory required for a work-item in bytes. If the kernel uses a dynamic call stack then additional space must be added to this value for the call stack.
“KernargSegmentAlign” integer Required The maximum byte alignment of arguments in the kernarg segment. Must be a power of 2.
“WavefrontSize” integer Required Wavefront size. Must be a power of 2.
“NumSGPRs” integer Required Number of scalar registers used by a wavefront for GFX6-GFX10. This includes the special SGPRs for VCC, Flat Scratch (GFX7-GFX10) and XNACK (for GFX8-GFX10). It does not include the 16 SGPR added if a trap handler is enabled. It is not rounded up to the allocation granularity.
“NumVGPRs” integer Required Number of vector registers used by each work-item for GFX6-GFX10
“MaxFlatWorkGroupSize” integer Required Maximum flat work-group size supported by the kernel in work-items. Must be >=1 and consistent with ReqdWorkGroupSize if not 0, 0, 0.
“NumSpilledSGPRs” integer   Number of stores from a scalar register to a register allocator created spill location.
“NumSpilledVGPRs” integer   Number of stores from a vector register to a register allocator created spill location.
Code Object V3 Metadata (–amdhsa-code-object-version=3)

Code object V3 metadata is specified by the NT_AMDGPU_METADATA note record (see Code Object V3 Note Records (–amdhsa-code-object-version=3)).

The metadata is represented as Message Pack formatted binary data (see [MsgPack]). The top level is a Message Pack map that includes the keys defined in table AMDHSA Code Object V3 Metadata Map and referenced tables.

Additional information can be added to the maps. To avoid conflicts, any key names should be prefixed by “vendor-name.” where vendor-name can be the name of the vendor and specific vendor tool that generates the information. The prefix is abbreviated to simply “.” when it appears within a map that has been added by the same vendor-name.

AMDHSA Code Object V3 Metadata Map
String Key Value Type Required? Description
“amdhsa.version” sequence of 2 integers Required
  • The first integer is the major version. Currently 1.
  • The second integer is the minor version. Currently 0.
“amdhsa.printf” sequence of strings  

Each string is encoded information about a printf function call. The encoded information is organized as fields separated by colon (‘:’):

ID:N:S[0]:S[1]:...:S[N-1]:FormatString

where:

ID
A 32-bit integer as a unique id for each printf function call
N
A 32-bit integer equal to the number of arguments of printf function call minus 1
S[i] (where i = 0, 1, … , N-1)
32-bit integers for the size in bytes of the i-th FormatString argument of the printf function call
FormatString
The format string passed to the printf function call.
“amdhsa.kernels” sequence of map Required Sequence of the maps for each kernel in the code object. See AMDHSA Code Object V3 Kernel Metadata Map for the definition of the keys included in that map.
AMDHSA Code Object V3 Kernel Metadata Map
String Key Value Type Required? Description
“.name” string Required Source name of the kernel.
“.symbol” string Required Name of the kernel descriptor ELF symbol.
“.language” string  

Source language of the kernel. Values include:

  • “OpenCL C”
  • “OpenCL C++”
  • “HCC”
  • “HIP”
  • “OpenMP”
  • “Assembler”
“.language_version” sequence of 2 integers  
  • The first integer is the major version.
  • The second integer is the minor version.
“.args” sequence of map   Sequence of maps of the kernel arguments. See AMDHSA Code Object V3 Kernel Argument Metadata Map for the definition of the keys included in that map.
“.reqd_workgroup_size” sequence of 3 integers  

If not 0, 0, 0 then all values must be >=1 and the dispatch work-group size X, Y, Z must correspond to the specified values. Defaults to 0, 0, 0.

Corresponds to the OpenCL reqd_work_group_size attribute.
“.workgroup_size_hint” sequence of 3 integers  

The dispatch work-group size X, Y, Z is likely to be the specified values.

Corresponds to the OpenCL work_group_size_hint attribute.
“.vec_type_hint” string  

The name of a scalar or vector type.

Corresponds to the OpenCL vec_type_hint attribute.
“.device_enqueue_symbol” string   The external symbol name associated with a kernel. OpenCL runtime allocates a global buffer for the symbol and saves the kernel’s address to it, which is used for device side enqueueing. Only available for device side enqueued kernels.
“.kernarg_segment_size” integer Required The size in bytes of the kernarg segment that holds the values of the arguments to the kernel.
“.group_segment_fixed_size” integer Required The amount of group segment memory required by a work-group in bytes. This does not include any dynamically allocated group segment memory that may be added when the kernel is dispatched.
“.private_segment_fixed_size” integer Required The amount of fixed private address space memory required for a work-item in bytes. If the kernel uses a dynamic call stack then additional space must be added to this value for the call stack.
“.kernarg_segment_align” integer Required The maximum byte alignment of arguments in the kernarg segment. Must be a power of 2.
“.wavefront_size” integer Required Wavefront size. Must be a power of 2.
“.sgpr_count” integer Required Number of scalar registers required by a wavefront for GFX6-GFX9. A register is required if it is used explicitly, or if a higher numbered register is used explicitly. This includes the special SGPRs for VCC, Flat Scratch (GFX7-GFX9) and XNACK (for GFX8-GFX9). It does not include the 16 SGPR added if a trap handler is enabled. It is not rounded up to the allocation granularity.
“.vgpr_count” integer Required Number of vector registers required by each work-item for GFX6-GFX9. A register is required if it is used explicitly, or if a higher numbered register is used explicitly.
“.max_flat_workgroup_size” integer Required Maximum flat work-group size supported by the kernel in work-items. Must be >=1 and consistent with ReqdWorkGroupSize if not 0, 0, 0.
“.sgpr_spill_count” integer   Number of stores from a scalar register to a register allocator created spill location.
“.vgpr_spill_count” integer   Number of stores from a vector register to a register allocator created spill location.
AMDHSA Code Object V3 Kernel Argument Metadata Map
String Key Value Type Required? Description
“.name” string   Kernel argument name.
“.type_name” string   Kernel argument type name.
“.size” integer Required Kernel argument size in bytes.
“.offset” integer Required Kernel argument offset in bytes. The offset must be a multiple of the alignment required by the argument.
“.value_kind” string Required

Kernel argument kind that specifies how to set up the corresponding argument. Values include:

“by_value”
The argument is copied directly into the kernarg.
“global_buffer”
A global address space pointer to the buffer data is passed in the kernarg.
“dynamic_shared_pointer”
A group address space pointer to dynamically allocated LDS is passed in the kernarg.
“sampler”
A global address space pointer to a S# is passed in the kernarg.
“image”
A global address space pointer to a T# is passed in the kernarg.
“pipe”
A global address space pointer to an OpenCL pipe is passed in the kernarg.
“queue”
A global address space pointer to an OpenCL device enqueue queue is passed in the kernarg.
“hidden_global_offset_x”
The OpenCL grid dispatch global offset for the X dimension is passed in the kernarg.
“hidden_global_offset_y”
The OpenCL grid dispatch global offset for the Y dimension is passed in the kernarg.
“hidden_global_offset_z”
The OpenCL grid dispatch global offset for the Z dimension is passed in the kernarg.
“hidden_none”
An argument that is not used by the kernel. Space needs to be left for it, but it does not need to be set up.
“hidden_printf_buffer”
A global address space pointer to the runtime printf buffer is passed in kernarg.
“hidden_hostcall_buffer”
A global address space pointer to the runtime hostcall buffer is passed in kernarg.
“hidden_default_queue”
A global address space pointer to the OpenCL device enqueue queue that should be used by the kernel by default is passed in the kernarg.
“hidden_completion_action”
A global address space pointer to help link enqueued kernels into the ancestor tree for determining when the parent kernel has finished.
“hidden_multigrid_sync_arg”
A global address space pointer for multi-grid synchronization is passed in the kernarg.
“.value_type” string   Unused and deprecated. This should no longer be emitted, but is accepted for compatibility.
“.pointee_align” integer   Alignment in bytes of pointee type for pointer type kernel argument. Must be a power of 2. Only present if “.value_kind” is “dynamic_shared_pointer”.
“.address_space” string  

Kernel argument address space qualifier. Only present if “.value_kind” is “global_buffer” or “dynamic_shared_pointer”. Values are:

  • “private”
  • “global”
  • “constant”
  • “local”
  • “generic”
  • “region”
“.access” string  

Kernel argument access qualifier. Only present if “.value_kind” is “image” or “pipe”. Values are:

  • “read_only”
  • “write_only”
  • “read_write”
“.actual_access” string  

The actual memory accesses performed by the kernel on the kernel argument. Only present if “.value_kind” is “global_buffer”, “image”, or “pipe”. This may be more restrictive than indicated by “.access” to reflect what the kernel actual does. If not present then the runtime must assume what is implied by “.access” and “.is_const” . Values are:

  • “read_only”
  • “write_only”
  • “read_write”
“.is_const” boolean   Indicates if the kernel argument is const qualified. Only present if “.value_kind” is “global_buffer”.
“.is_restrict” boolean   Indicates if the kernel argument is restrict qualified. Only present if “.value_kind” is “global_buffer”.
“.is_volatile” boolean   Indicates if the kernel argument is volatile qualified. Only present if “.value_kind” is “global_buffer”.
“.is_pipe” boolean   Indicates if the kernel argument is pipe qualified. Only present if “.value_kind” is “pipe”.

Kernel Dispatch

The HSA architected queuing language (AQL) defines a user space memory interface that can be used to control the dispatch of kernels, in an agent independent way. An agent can have zero or more AQL queues created for it using the ROCm runtime, in which AQL packets (all of which are 64 bytes) can be placed. See the HSA Platform System Architecture Specification [HSA] for the AQL queue mechanics and packet layouts.

The packet processor of a kernel agent is responsible for detecting and dispatching HSA kernels from the AQL queues associated with it. For AMD GPUs the packet processor is implemented by the hardware command processor (CP), asynchronous dispatch controller (ADC) and shader processor input controller (SPI).

The ROCm runtime can be used to allocate an AQL queue object. It uses the kernel mode driver to initialize and register the AQL queue with CP.

To dispatch a kernel the following actions are performed. This can occur in the CPU host program, or from an HSA kernel executing on a GPU.

  1. A pointer to an AQL queue for the kernel agent on which the kernel is to be executed is obtained.
  2. A pointer to the kernel descriptor (see Kernel Descriptor) of the kernel to execute is obtained. It must be for a kernel that is contained in a code object that that was loaded by the ROCm runtime on the kernel agent with which the AQL queue is associated.
  3. Space is allocated for the kernel arguments using the ROCm runtime allocator for a memory region with the kernarg property for the kernel agent that will execute the kernel. It must be at least 16-byte aligned.
  4. Kernel argument values are assigned to the kernel argument memory allocation. The layout is defined in the HSA Programmer’s Language Reference [HSA]. For AMDGPU the kernel execution directly accesses the kernel argument memory in the same way constant memory is accessed. (Note that the HSA specification allows an implementation to copy the kernel argument contents to another location that is accessed by the kernel.)
  5. An AQL kernel dispatch packet is created on the AQL queue. The ROCm runtime api uses 64-bit atomic operations to reserve space in the AQL queue for the packet. The packet must be set up, and the final write must use an atomic store release to set the packet kind to ensure the packet contents are visible to the kernel agent. AQL defines a doorbell signal mechanism to notify the kernel agent that the AQL queue has been updated. These rules, and the layout of the AQL queue and kernel dispatch packet is defined in the HSA System Architecture Specification [HSA].
  6. A kernel dispatch packet includes information about the actual dispatch, such as grid and work-group size, together with information from the code object about the kernel, such as segment sizes. The ROCm runtime queries on the kernel symbol can be used to obtain the code object values which are recorded in the Code Object Metadata.
  7. CP executes micro-code and is responsible for detecting and setting up the GPU to execute the wavefronts of a kernel dispatch.
  8. CP ensures that when the a wavefront starts executing the kernel machine code, the scalar general purpose registers (SGPR) and vector general purpose registers (VGPR) are set up as required by the machine code. The required setup is defined in the Kernel Descriptor. The initial register state is defined in Initial Kernel Execution State.
  9. The prolog of the kernel machine code (see Kernel Prolog) sets up the machine state as necessary before continuing executing the machine code that corresponds to the kernel.
  10. When the kernel dispatch has completed execution, CP signals the completion signal specified in the kernel dispatch packet if not 0.

Memory Spaces

The memory space properties are:

AMDHSA Memory Spaces
Memory Space Name HSA Segment Name Hardware Name Address Size NULL Value
Private private scratch 32 0x00000000
Local group LDS 32 0xFFFFFFFF
Global global global 64 0x0000000000000000
Constant constant same as global 64 0x0000000000000000
Generic flat flat 64 0x0000000000000000
Region N/A GDS 32 not implemented for AMDHSA

The global and constant memory spaces both use global virtual addresses, which are the same virtual address space used by the CPU. However, some virtual addresses may only be accessible to the CPU, some only accessible by the GPU, and some by both.

Using the constant memory space indicates that the data will not change during the execution of the kernel. This allows scalar read instructions to be used. The vector and scalar L1 caches are invalidated of volatile data before each kernel dispatch execution to allow constant memory to change values between kernel dispatches.

The local memory space uses the hardware Local Data Store (LDS) which is automatically allocated when the hardware creates work-groups of wavefronts, and freed when all the wavefronts of a work-group have terminated. The data store (DS) instructions can be used to access it.

The private memory space uses the hardware scratch memory support. If the kernel uses scratch, then the hardware allocates memory that is accessed using wavefront lane dword (4 byte) interleaving. The mapping used from private address to physical address is:

wavefront-scratch-base + (private-address * wavefront-size * 4) + (wavefront-lane-id * 4)

There are different ways that the wavefront scratch base address is determined by a wavefront (see Initial Kernel Execution State). This memory can be accessed in an interleaved manner using buffer instruction with the scratch buffer descriptor and per wavefront scratch offset, by the scratch instructions, or by flat instructions. If each lane of a wavefront accesses the same private address, the interleaving results in adjacent dwords being accessed and hence requires fewer cache lines to be fetched. Multi-dword access is not supported except by flat and scratch instructions in GFX9-GFX10.

The generic address space uses the hardware flat address support available in GFX7-GFX10. This uses two fixed ranges of virtual addresses (the private and local apertures), that are outside the range of addressible global memory, to map from a flat address to a private or local address.

FLAT instructions can take a flat address and access global, private (scratch) and group (LDS) memory depending in if the address is within one of the aperture ranges. Flat access to scratch requires hardware aperture setup and setup in the kernel prologue (see Flat Scratch). Flat access to LDS requires hardware aperture setup and M0 (GFX7-GFX8) register setup (see M0).

To convert between a segment address and a flat address the base address of the apertures address can be used. For GFX7-GFX8 these are available in the HSA AQL Queue the address of which can be obtained with Queue Ptr SGPR (see Initial Kernel Execution State). For GFX9-GFX10 the aperture base addresses are directly available as inline constant registers SRC_SHARED_BASE/LIMIT and SRC_PRIVATE_BASE/LIMIT. In 64 bit address mode the aperture sizes are 2^32 bytes and the base is aligned to 2^32 which makes it easier to convert from flat to segment or segment to flat.

Image and Samplers

Image and sample handles created by the ROCm runtime are 64-bit addresses of a hardware 32-byte V# and 48 byte S# object respectively. In order to support the HSA query_sampler operations two extra dwords are used to store the HSA BRIG enumeration values for the queries that are not trivially deducible from the S# representation.

HSA Signals

HSA signal handles created by the ROCm runtime are 64-bit addresses of a structure allocated in memory accessible from both the CPU and GPU. The structure is defined by the ROCm runtime and subject to change between releases (see [AMD-ROCm-github]).

HSA AQL Queue

The HSA AQL queue structure is defined by the ROCm runtime and subject to change between releases (see [AMD-ROCm-github]). For some processors it contains fields needed to implement certain language features such as the flat address aperture bases. It also contains fields used by CP such as managing the allocation of scratch memory.

Kernel Descriptor

A kernel descriptor consists of the information needed by CP to initiate the execution of a kernel, including the entry point address of the machine code that implements the kernel.

Kernel Descriptor for GFX6-GFX10

CP microcode requires the Kernel descriptor to be allocated on 64-byte alignment.

Kernel Descriptor for GFX6-GFX10
Bits Size Field Name Description
31:0 4 bytes GROUP_SEGMENT_FIXED_SIZE The amount of fixed local address space memory required for a work-group in bytes. This does not include any dynamically allocated local address space memory that may be added when the kernel is dispatched.
63:32 4 bytes PRIVATE_SEGMENT_FIXED_SIZE The amount of fixed private address space memory required for a work-item in bytes. If is_dynamic_callstack is 1 then additional space must be added to this value for the call stack.
127:64 8 bytes   Reserved, must be 0.
191:128 8 bytes KERNEL_CODE_ENTRY_BYTE_OFFSET Byte offset (possibly negative) from base address of kernel descriptor to kernel’s entry point instruction which must be 256 byte aligned.
351:272 20 bytes   Reserved, must be 0.
383:352 4 bytes COMPUTE_PGM_RSRC3
GFX6-9
Reserved, must be 0.
GFX10
Compute Shader (CS) program settings used by CP to set up COMPUTE_PGM_RSRC3 configuration register. See compute_pgm_rsrc3 for GFX10.
415:384 4 bytes COMPUTE_PGM_RSRC1 Compute Shader (CS) program settings used by CP to set up COMPUTE_PGM_RSRC1 configuration register. See compute_pgm_rsrc1 for GFX6-GFX10.
447:416 4 bytes COMPUTE_PGM_RSRC2 Compute Shader (CS) program settings used by CP to set up COMPUTE_PGM_RSRC2 configuration register. See compute_pgm_rsrc2 for GFX6-GFX10.
448 1 bit ENABLE_SGPR_PRIVATE_SEGMENT _BUFFER

Enable the setup of the SGPR user data registers (see Initial Kernel Execution State).

The total number of SGPR user data registers requested must not exceed 16 and match value in compute_pgm_rsrc2.user_sgpr.user_sgpr_count. Any requests beyond 16 will be ignored.
449 1 bit ENABLE_SGPR_DISPATCH_PTR see above
450 1 bit ENABLE_SGPR_QUEUE_PTR see above
451 1 bit ENABLE_SGPR_KERNARG_SEGMENT_PTR see above
452 1 bit ENABLE_SGPR_DISPATCH_ID see above
453 1 bit ENABLE_SGPR_FLAT_SCRATCH_INIT see above
454 1 bit ENABLE_SGPR_PRIVATE_SEGMENT _SIZE see above
457:455 3 bits   Reserved, must be 0.
458 1 bit ENABLE_WAVEFRONT_SIZE32
GFX6-9
Reserved, must be 0.
GFX10
  • If 0 execute in wavefront size 64 mode.
  • If 1 execute in native wavefront size 32 mode.
463:459 5 bits   Reserved, must be 0.
511:464 6 bytes   Reserved, must be 0.
512 Total size 64 bytes.
compute_pgm_rsrc1 for GFX6-GFX10
Bits Size Field Name Description
5:0 6 bits GRANULATED_WORKITEM_VGPR_COUNT

Number of vector register blocks used by each work-item; granularity is device specific:

GFX6-GFX9
  • vgprs_used 0..256
  • max(0, ceil(vgprs_used / 4) - 1)
GFX10 (wavefront size 64)
  • max_vgpr 1..256
  • max(0, ceil(vgprs_used / 4) - 1)
GFX10 (wavefront size 32)
  • max_vgpr 1..256
  • max(0, ceil(vgprs_used / 8) - 1)

Where vgprs_used is defined as the highest VGPR number explicitly referenced plus one.

Used by CP to set up COMPUTE_PGM_RSRC1.VGPRS.

The Assembler calculates this automatically for the selected processor from values provided to the .amdhsa_kernel directive by the .amdhsa_next_free_vgpr nested directive (see AMDHSA Kernel Assembler Directives).
9:6 4 bits GRANULATED_WAVEFRONT_SGPR_COUNT

Number of scalar register blocks used by a wavefront; granularity is device specific:

GFX6-GFX8
  • sgprs_used 0..112
  • max(0, ceil(sgprs_used / 8) - 1)
GFX9
  • sgprs_used 0..112
  • 2 * max(0, ceil(sgprs_used / 16) - 1)
GFX10
Reserved, must be 0. (128 SGPRs always allocated.)

Where sgprs_used is defined as the highest SGPR number explicitly referenced plus one, plus a target specific number of additional special SGPRs for VCC, FLAT_SCRATCH (GFX7+) and XNACK_MASK (GFX8+), and any additional target specific limitations. It does not include the 16 SGPRs added if a trap handler is enabled.

The target specific limitations and special SGPR layout are defined in the hardware documentation, which can be found in the Processors table.

Used by CP to set up COMPUTE_PGM_RSRC1.SGPRS.

The Assembler calculates this automatically for the selected processor from values provided to the .amdhsa_kernel directive by the .amdhsa_next_free_sgpr and .amdhsa_reserve_* nested directives (see AMDHSA Kernel Assembler Directives).
11:10 2 bits PRIORITY

Must be 0.

Start executing wavefront at the specified priority.

CP is responsible for filling in COMPUTE_PGM_RSRC1.PRIORITY.
13:12 2 bits FLOAT_ROUND_MODE_32

Wavefront starts execution with specified rounding mode for single (32 bit) floating point precision floating point operations.

Floating point rounding mode values are defined in Floating Point Rounding Mode Enumeration Values.

Used by CP to set up COMPUTE_PGM_RSRC1.FLOAT_MODE.
15:14 2 bits FLOAT_ROUND_MODE_16_64

Wavefront starts execution with specified rounding denorm mode for half/double (16 and 64-bit) floating point precision floating point operations.

Floating point rounding mode values are defined in Floating Point Rounding Mode Enumeration Values.

Used by CP to set up COMPUTE_PGM_RSRC1.FLOAT_MODE.
17:16 2 bits FLOAT_DENORM_MODE_32

Wavefront starts execution with specified denorm mode for single (32 bit) floating point precision floating point operations.

Floating point denorm mode values are defined in Floating Point Denorm Mode Enumeration Values.

Used by CP to set up COMPUTE_PGM_RSRC1.FLOAT_MODE.
19:18 2 bits FLOAT_DENORM_MODE_16_64

Wavefront starts execution with specified denorm mode for half/double (16 and 64-bit) floating point precision floating point operations.

Floating point denorm mode values are defined in Floating Point Denorm Mode Enumeration Values.

Used by CP to set up COMPUTE_PGM_RSRC1.FLOAT_MODE.
20 1 bit PRIV

Must be 0.

Start executing wavefront in privilege trap handler mode.

CP is responsible for filling in COMPUTE_PGM_RSRC1.PRIV.
21 1 bit ENABLE_DX10_CLAMP

Wavefront starts execution with DX10 clamp mode enabled. Used by the vector ALU to force DX10 style treatment of NaN’s (when set, clamp NaN to zero, otherwise pass NaN through).

Used by CP to set up COMPUTE_PGM_RSRC1.DX10_CLAMP.
22 1 bit DEBUG_MODE

Must be 0.

Start executing wavefront in single step mode.

CP is responsible for filling in COMPUTE_PGM_RSRC1.DEBUG_MODE.
23 1 bit ENABLE_IEEE_MODE

Wavefront starts execution with IEEE mode enabled. Floating point opcodes that support exception flag gathering will quiet and propagate signaling-NaN inputs per IEEE 754-2008. Min_dx10 and max_dx10 become IEEE 754-2008 compliant due to signaling-NaN propagation and quieting.

Used by CP to set up COMPUTE_PGM_RSRC1.IEEE_MODE.
24 1 bit BULKY

Must be 0.

Only one work-group allowed to execute on a compute unit.

CP is responsible for filling in COMPUTE_PGM_RSRC1.BULKY.
25 1 bit CDBG_USER

Must be 0.

Flag that can be used to control debugging code.

CP is responsible for filling in COMPUTE_PGM_RSRC1.CDBG_USER.
26 1 bit FP16_OVFL
GFX6-GFX8
Reserved, must be 0.
GFX9-GFX10

Wavefront starts execution with specified fp16 overflow mode.

  • If 0, fp16 overflow generates +/-INF values.
  • If 1, fp16 overflow that is the result of an +/-INF input value or divide by 0 produces a +/-INF, otherwise clamps computed overflow to +/-MAX_FP16 as appropriate.

Used by CP to set up COMPUTE_PGM_RSRC1.FP16_OVFL.
28:27 2 bits   Reserved, must be 0.
29 1 bit WGP_MODE
GFX6-GFX9
Reserved, must be 0.
GFX10
  • If 0 execute work-groups in CU wavefront execution mode.
  • If 1 execute work-groups on in WGP wavefront execution mode.

See Memory Model.

Used by CP to set up COMPUTE_PGM_RSRC1.WGP_MODE.
30 1 bit MEM_ORDERED
GFX6-9
Reserved, must be 0.
GFX10

Controls the behavior of the s_waitcnt’s vmcnt and vscnt counters.

  • If 0 vmcnt reports completion of load and atomic with return out of order with sample instructions, and the vscnt reports the completion of store and atomic without return in order.
  • If 1 vmcnt reports completion of load, atomic with return and sample instructions in order, and the vscnt reports the completion of store and atomic without return in order.

Used by CP to set up COMPUTE_PGM_RSRC1.MEM_ORDERED.
31 1 bit FWD_PROGRESS
GFX6-9
Reserved, must be 0.
GFX10
  • If 0 execute SIMD wavefronts using oldest first policy.
  • If 1 execute SIMD wavefronts to ensure wavefronts will make some forward progress.

Used by CP to set up COMPUTE_PGM_RSRC1.FWD_PROGRESS.
32 Total size 4 bytes
compute_pgm_rsrc2 for GFX6-GFX10
Bits Size Field Name Description
0 1 bit ENABLE_SGPR_PRIVATE_SEGMENT _WAVEFRONT_OFFSET

Enable the setup of the SGPR wavefront scratch offset system register (see Initial Kernel Execution State).

Used by CP to set up COMPUTE_PGM_RSRC2.SCRATCH_EN.
5:1 5 bits USER_SGPR_COUNT

The total number of SGPR user data registers requested. This number must match the number of user data registers enabled.

Used by CP to set up COMPUTE_PGM_RSRC2.USER_SGPR.
6 1 bit ENABLE_TRAP_HANDLER

Must be 0.

This bit represents COMPUTE_PGM_RSRC2.TRAP_PRESENT, which is set by the CP if the runtime has installed a trap handler.
7 1 bit ENABLE_SGPR_WORKGROUP_ID_X

Enable the setup of the system SGPR register for the work-group id in the X dimension (see Initial Kernel Execution State).

Used by CP to set up COMPUTE_PGM_RSRC2.TGID_X_EN.
8 1 bit ENABLE_SGPR_WORKGROUP_ID_Y

Enable the setup of the system SGPR register for the work-group id in the Y dimension (see Initial Kernel Execution State).

Used by CP to set up COMPUTE_PGM_RSRC2.TGID_Y_EN.
9 1 bit ENABLE_SGPR_WORKGROUP_ID_Z

Enable the setup of the system SGPR register for the work-group id in the Z dimension (see Initial Kernel Execution State).

Used by CP to set up COMPUTE_PGM_RSRC2.TGID_Z_EN.
10 1 bit ENABLE_SGPR_WORKGROUP_INFO

Enable the setup of the system SGPR register for work-group information (see Initial Kernel Execution State).

Used by CP to set up COMPUTE_PGM_RSRC2.TGID_SIZE_EN.
12:11 2 bits ENABLE_VGPR_WORKITEM_ID

Enable the setup of the VGPR system registers used for the work-item ID. System VGPR Work-Item ID Enumeration Values defines the values.

Used by CP to set up COMPUTE_PGM_RSRC2.TIDIG_CMP_CNT.
13 1 bit ENABLE_EXCEPTION_ADDRESS_WATCH

Must be 0.

Wavefront starts execution with address watch exceptions enabled which are generated when L1 has witnessed a thread access an address of interest.

CP is responsible for filling in the address watch bit in COMPUTE_PGM_RSRC2.EXCP_EN_MSB according to what the runtime requests.
14 1 bit ENABLE_EXCEPTION_MEMORY

Must be 0.

Wavefront starts execution with memory violation exceptions exceptions enabled which are generated when a memory violation has occurred for this wavefront from L1 or LDS (write-to-read-only-memory, mis-aligned atomic, LDS address out of range, illegal address, etc.).

CP sets the memory violation bit in COMPUTE_PGM_RSRC2.EXCP_EN_MSB according to what the runtime requests.
23:15 9 bits GRANULATED_LDS_SIZE

Must be 0.

CP uses the rounded value from the dispatch packet, not this value, as the dispatch may contain dynamically allocated group segment memory. CP writes directly to COMPUTE_PGM_RSRC2.LDS_SIZE.

Amount of group segment (LDS) to allocate for each work-group. Granularity is device specific:

GFX6:
roundup(lds-size / (64 * 4))
GFX7-GFX10:
roundup(lds-size / (128 * 4))
24 1 bit ENABLE_EXCEPTION_IEEE_754_FP _INVALID_OPERATION

Wavefront starts execution with specified exceptions enabled.

Used by CP to set up COMPUTE_PGM_RSRC2.EXCP_EN (set from bits 0..6).

IEEE 754 FP Invalid Operation
25 1 bit ENABLE_EXCEPTION_FP_DENORMAL _SOURCE FP Denormal one or more input operands is a denormal number
26 1 bit ENABLE_EXCEPTION_IEEE_754_FP _DIVISION_BY_ZERO IEEE 754 FP Division by Zero
27 1 bit ENABLE_EXCEPTION_IEEE_754_FP _OVERFLOW IEEE 754 FP FP Overflow
28 1 bit ENABLE_EXCEPTION_IEEE_754_FP _UNDERFLOW IEEE 754 FP Underflow
29 1 bit ENABLE_EXCEPTION_IEEE_754_FP _INEXACT IEEE 754 FP Inexact
30 1 bit ENABLE_EXCEPTION_INT_DIVIDE_BY _ZERO Integer Division by Zero (rcp_iflag_f32 instruction only)
31 1 bit   Reserved, must be 0.
32 Total size 4 bytes.
compute_pgm_rsrc3 for GFX10
Bits Size Field Name Description
3:0 4 bits SHARED_VGPR_COUNT Number of shared VGPRs for wavefront size 64. Granularity 8. Value 0-120. compute_pgm_rsrc1.vgprs + shared_vgpr_cnt cannot exceed 64.
31:4 28 bits   Reserved, must be 0.
32 Total size 4 bytes.
Floating Point Rounding Mode Enumeration Values
Enumeration Name Value Description
FLOAT_ROUND_MODE_NEAR_EVEN 0 Round Ties To Even
FLOAT_ROUND_MODE_PLUS_INFINITY 1 Round Toward +infinity
FLOAT_ROUND_MODE_MINUS_INFINITY 2 Round Toward -infinity
FLOAT_ROUND_MODE_ZERO 3 Round Toward 0
Floating Point Denorm Mode Enumeration Values
Enumeration Name Value Description
FLOAT_DENORM_MODE_FLUSH_SRC_DST 0 Flush Source and Destination Denorms
FLOAT_DENORM_MODE_FLUSH_DST 1 Flush Output Denorms
FLOAT_DENORM_MODE_FLUSH_SRC 2 Flush Source Denorms
FLOAT_DENORM_MODE_FLUSH_NONE 3 No Flush
System VGPR Work-Item ID Enumeration Values
Enumeration Name Value Description
SYSTEM_VGPR_WORKITEM_ID_X 0 Set work-item X dimension ID.
SYSTEM_VGPR_WORKITEM_ID_X_Y 1 Set work-item X and Y dimensions ID.
SYSTEM_VGPR_WORKITEM_ID_X_Y_Z 2 Set work-item X, Y and Z dimensions ID.
SYSTEM_VGPR_WORKITEM_ID_UNDEFINED 3 Undefined.

Initial Kernel Execution State

This section defines the register state that will be set up by the packet processor prior to the start of execution of every wavefront. This is limited by the constraints of the hardware controllers of CP/ADC/SPI.

The order of the SGPR registers is defined, but the compiler can specify which ones are actually setup in the kernel descriptor using the enable_sgpr_* bit fields (see Kernel Descriptor). The register numbers used for enabled registers are dense starting at SGPR0: the first enabled register is SGPR0, the next enabled register is SGPR1 etc.; disabled registers do not have an SGPR number.

The initial SGPRs comprise up to 16 User SRGPs that are set by CP and apply to all wavefronts of the grid. It is possible to specify more than 16 User SGPRs using the enable_sgpr_* bit fields, in which case only the first 16 are actually initialized. These are then immediately followed by the System SGPRs that are set up by ADC/SPI and can have different values for each wavefront of the grid dispatch.

SGPR register initial state is defined in SGPR Register Set Up Order.

SGPR Register Set Up Order
SGPR Order Name (kernel descriptor enable field) Number of SGPRs Description
First Private Segment Buffer (enable_sgpr_private _segment_buffer) 4

V# that can be used, together with Scratch Wavefront Offset as an offset, to access the private memory space using a segment address.

CP uses the value provided by the runtime.
then Dispatch Ptr (enable_sgpr_dispatch_ptr) 2 64-bit address of AQL dispatch packet for kernel dispatch actually executing.
then Queue Ptr (enable_sgpr_queue_ptr) 2 64-bit address of amd_queue_t object for AQL queue on which the dispatch packet was queued.
then Kernarg Segment Ptr (enable_sgpr_kernarg _segment_ptr) 2

64-bit address of Kernarg segment. This is directly copied from the kernarg_address in the kernel dispatch packet.

Having CP load it once avoids loading it at the beginning of every wavefront.
then Dispatch Id (enable_sgpr_dispatch_id) 2 64-bit Dispatch ID of the dispatch packet being executed.
then Flat Scratch Init (enable_sgpr_flat_scratch _init) 2

This is 2 SGPRs:

GFX6
Not supported.
GFX7-GFX8

The first SGPR is a 32-bit byte offset from SH_HIDDEN_PRIVATE_BASE_VIMID to per SPI base of memory for scratch for the queue executing the kernel dispatch. CP obtains this from the runtime. (The Scratch Segment Buffer base address is SH_HIDDEN_PRIVATE_BASE_VIMID plus this offset.) The value of Scratch Wavefront Offset must be added to this offset by the kernel machine code, right shifted by 8, and moved to the FLAT_SCRATCH_HI SGPR register. FLAT_SCRATCH_HI corresponds to SGPRn-4 on GFX7, and SGPRn-6 on GFX8 (where SGPRn is the highest numbered SGPR allocated to the wavefront). FLAT_SCRATCH_HI is multiplied by 256 (as it is in units of 256 bytes) and added to SH_HIDDEN_PRIVATE_BASE_VIMID to calculate the per wavefront FLAT SCRATCH BASE in flat memory instructions that access the scratch aperture.

The second SGPR is 32-bit byte size of a single work-item’s scratch memory usage. CP obtains this from the runtime, and it is always a multiple of DWORD. CP checks that the value in the kernel dispatch packet Private Segment Byte Size is not larger and requests the runtime to increase the queue’s scratch size if necessary. The kernel code must move it to FLAT_SCRATCH_LO which is SGPRn-3 on GFX7 and SGPRn-5 on GFX8. FLAT_SCRATCH_LO is used as the FLAT SCRATCH SIZE in flat memory instructions. Having CP load it once avoids loading it at the beginning of every wavefront.

GFX9-GFX10
This is the 64-bit base address of the per SPI scratch backing memory managed by SPI for the queue executing the kernel dispatch. CP obtains this from the runtime (and divides it if there are multiple Shader Arrays each with its own SPI). The value of Scratch Wavefront Offset must be added by the kernel machine code and the result moved to the FLAT_SCRATCH SGPR which is SGPRn-6 and SGPRn-5. It is used as the FLAT SCRATCH BASE in flat memory instructions.
then Private Segment Size 1

The 32-bit byte size of a (enable_sgpr_private single work-item’s scratch_segment_size) memory allocation. This is the value from the kernel dispatch packet Private Segment Byte Size rounded up by CP to a multiple of DWORD.

Having CP load it once avoids loading it at the beginning of every wavefront.

This is not used for GFX7-GFX8 since it is the same value as the second SGPR of Flat Scratch Init. However, it may be needed for GFX9-GFX10 which changes the meaning of the Flat Scratch Init value.
then Grid Work-Group Count X (enable_sgpr_grid _workgroup_count_X) 1 32-bit count of the number of work-groups in the X dimension for the grid being executed. Computed from the fields in the kernel dispatch packet as ((grid_size.x + workgroup_size.x - 1) / workgroup_size.x).
then Grid Work-Group Count Y (enable_sgpr_grid _workgroup_count_Y && less than 16 previous SGPRs) 1

32-bit count of the number of work-groups in the Y dimension for the grid being executed. Computed from the fields in the kernel dispatch packet as ((grid_size.y + workgroup_size.y - 1) / workgroupSize.y).

Only initialized if <16 previous SGPRs initialized.
then Grid Work-Group Count Z (enable_sgpr_grid _workgroup_count_Z && less than 16 previous SGPRs) 1

32-bit count of the number of work-groups in the Z dimension for the grid being executed. Computed from the fields in the kernel dispatch packet as ((grid_size.z + workgroup_size.z - 1) / workgroupSize.z).

Only initialized if <16 previous SGPRs initialized.
then Work-Group Id X (enable_sgpr_workgroup_id _X) 1 32-bit work-group id in X dimension of grid for wavefront.
then Work-Group Id Y (enable_sgpr_workgroup_id _Y) 1 32-bit work-group id in Y dimension of grid for wavefront.
then Work-Group Id Z (enable_sgpr_workgroup_id _Z) 1 32-bit work-group id in Z dimension of grid for wavefront.
then Work-Group Info (enable_sgpr_workgroup _info) 1 {first_wavefront, 14’b0000, ordered_append_term[10:0], threadgroup_size_in_wavefronts[5:0]}
then Scratch Wavefront Offset (enable_sgpr_private _segment_wavefront_offset) 1 32-bit byte offset from base of scratch base of queue executing the kernel dispatch. Must be used as an offset with Private segment address when using Scratch Segment Buffer. It must be used to set up FLAT SCRATCH for flat addressing (see Flat Scratch).

The order of the VGPR registers is defined, but the compiler can specify which ones are actually setup in the kernel descriptor using the enable_vgpr* bit fields (see Kernel Descriptor). The register numbers used for enabled registers are dense starting at VGPR0: the first enabled register is VGPR0, the next enabled register is VGPR1 etc.; disabled registers do not have a VGPR number.

VGPR register initial state is defined in VGPR Register Set Up Order.

VGPR Register Set Up Order
VGPR Order Name (kernel descriptor enable field) Number of VGPRs Description
First Work-Item Id X (Always initialized) 1 32-bit work-item id in X dimension of work-group for wavefront lane.
then Work-Item Id Y (enable_vgpr_workitem_id > 0) 1 32-bit work-item id in Y dimension of work-group for wavefront lane.
then Work-Item Id Z (enable_vgpr_workitem_id > 1) 1 32-bit work-item id in Z dimension of work-group for wavefront lane.

The setting of registers is done by GPU CP/ADC/SPI hardware as follows:

  1. SGPRs before the Work-Group Ids are set by CP using the 16 User Data registers.
  2. Work-group Id registers X, Y, Z are set by ADC which supports any combination including none.
  3. Scratch Wavefront Offset is set by SPI in a per wavefront basis which is why its value cannot be included with the flat scratch init value which is per queue.
  4. The VGPRs are set by SPI which only supports specifying either (X), (X, Y) or (X, Y, Z).

Flat Scratch register pair are adjacent SGPRs so they can be moved as a 64-bit value to the hardware required SGPRn-3 and SGPRn-4 respectively.

The global segment can be accessed either using buffer instructions (GFX6 which has V# 64-bit address support), flat instructions (GFX7-GFX10), or global instructions (GFX9-GFX10).

If buffer operations are used, then the compiler can generate a V# with the following properties:

  • base address of 0
  • no swizzle
  • ATC: 1 if IOMMU present (such as APU)
  • ptr64: 1
  • MTYPE set to support memory coherence that matches the runtime (such as CC for APU and NC for dGPU).

Kernel Prolog

The compiler performs initialization in the kernel prologue depending on the target and information about things like stack usage in the kernel and called functions. Some of this initialization requires the compiler to request certain User and System SGPRs be present in the Initial Kernel Execution State via the Kernel Descriptor.

CFI
  1. The CFI return address is undefined.
  2. The CFI CFA is defined using an expression which evaluates to a location description that comprises one memory location description for the DW_ASPACE_AMDGPU_private_lane address space address 0.
M0
GFX6-GFX8
The M0 register must be initialized with a value at least the total LDS size if the kernel may access LDS via DS or flat operations. Total LDS size is available in dispatch packet. For M0, it is also possible to use maximum possible value of LDS for given target (0x7FFF for GFX6 and 0xFFFF for GFX7-GFX8).
GFX9-GFX10
The M0 register is not used for range checking LDS accesses and so does not need to be initialized in the prolog.
Stack Pointer

If the kernel has function calls it must set up the ABI stack pointer described in Non-Kernel Functions by setting SGPR32 to the unswizzled scratch offset of the address past the last local allocation.

Frame Pointer

If the kernel needs a frame pointer for the reasons defined in SIFrameLowering then SGPR33 is used and is always set to 0 in the kernel prolog. If a frame pointer is not required then all uses of the frame pointer are replaced with immediate 0 offsets.

Flat Scratch

If the kernel or any function it calls may use flat operations to access scratch memory, the prolog code must set up the FLAT_SCRATCH register pair (FLAT_SCRATCH_LO/FLAT_SCRATCH_HI which are in SGPRn-4/SGPRn-3). Initialization uses Flat Scratch Init and Scratch Wavefront Offset SGPR registers (see Initial Kernel Execution State):

GFX6
Flat scratch is not supported.

GFX7-GFX8

  1. The low word of Flat Scratch Init is 32-bit byte offset from SH_HIDDEN_PRIVATE_BASE_VIMID to the base of scratch backing memory being managed by SPI for the queue executing the kernel dispatch. This is the same value used in the Scratch Segment Buffer V# base address. The prolog must add the value of Scratch Wavefront Offset to get the wavefront’s byte scratch backing memory offset from SH_HIDDEN_PRIVATE_BASE_VIMID. Since FLAT_SCRATCH_LO is in units of 256 bytes, the offset must be right shifted by 8 before moving into FLAT_SCRATCH_LO.
  2. The second word of Flat Scratch Init is 32-bit byte size of a single work-items scratch memory usage. This is directly loaded from the kernel dispatch packet Private Segment Byte Size and rounded up to a multiple of DWORD. Having CP load it once avoids loading it at the beginning of every wavefront. The prolog must move it to FLAT_SCRATCH_LO for use as FLAT SCRATCH SIZE.
GFX9-GFX10
The Flat Scratch Init is the 64-bit address of the base of scratch backing memory being managed by SPI for the queue executing the kernel dispatch. The prolog must add the value of Scratch Wavefront Offset and moved to the FLAT_SCRATCH pair for use as the flat scratch base in flat memory instructions.
Private Segment Buffer

A set of four SGPRs beginning at a four-aligned SGPR index are always selected to serve as the scratch V# for the kernel as follows:

  • If it is known during instruction selection that there is stack usage, SGPR0-3 is reserved for use as the scratch V#. Stack usage is assumed if optimizations are disabled (-O0), if stack objects already exist (for locals, etc.), or if there are any function calls.

  • Otherwise, four high numbered SGPRs beginning at a four-aligned SGPR index are reserved for the tentative scratch V#. These will be used if it is determined that spilling is needed.

    • If no use is made of the tentative scratch V#, then it is unreserved, and the register count is determined ignoring it.
    • If use is made of the tentative scratch V#, then its register numbers are shifted to the first four-aligned SGPR index after the highest one allocated by the register allocator, and all uses are updated. The register count includes them in the shifted location.
    • In either case, if the processor has the SGPR allocation bug, the tentative allocation is not shifted or unreserved in order to ensure the register count is higher to workaround the bug.

    Note

    This approach of using a tentative scratch V# and shifting the register numbers if used avoids having to perform register allocation a second time if the tentative V# is eliminated. This is more efficient and avoids the problem that the second register allocation may perform spilling which will fail as there is no longer a scratch V#.

When the kernel prolog code is being emitted it is known whether the scratch V# described above is actually used. If it is, the prolog code must set it up by copying the Private Segment Buffer to the scratch V# registers and then adding the Private Segment Wavefront Offset to the queue base address in the V#. The result is a V# with a base address pointing to the beginning of the wavefront scratch backing memory.

The Private Segment Buffer is always requested, but the Private Segment Wavefront Offset is only requested if it is used (see Initial Kernel Execution State).

Memory Model

This section describes the mapping of the LLVM memory model onto AMDGPU machine code (see Memory Model for Concurrent Operations).

The AMDGPU backend supports the memory synchronization scopes specified in Memory Scopes.

The code sequences used to implement the memory model specify the order of instructions that a single thread must execute. The s_waitcnt and cache management instructions such as buffer_wbinvl1_vol are defined with respect to other memory instructions executed by the same thread. This allows them to be moved earlier or later which can allow them to be combined with other instances of the same instruction, or hoisted/sunk out of loops to improve performance. Only the instructions related to the memory model are given; additional s_waitcnt instructions are required to ensure registers are defined before being used. These may be able to be combined with the memory model s_waitcnt instructions as described above.

The AMDGPU backend supports the following memory models:

HSA Memory Model [HSA]
The HSA memory model uses a single happens-before relation for all address spaces (see Address Spaces).
OpenCL Memory Model [OpenCL]
The OpenCL memory model which has separate happens-before relations for the global and local address spaces. Only a fence specifying both global and local address space, and seq_cst instructions join the relationships. Since the LLVM memfence instruction does not allow an address space to be specified the OpenCL fence has to conservatively assume both local and global address space was specified. However, optimizations can often be done to eliminate the additional s_waitcnt instructions when there are no intervening memory instructions which access the corresponding address space. The code sequences in the table indicate what can be omitted for the OpenCL memory. The target triple environment is used to determine if the source language is OpenCL (see OpenCL).

ds/flat_load/store/atomic instructions to local memory are termed LDS operations.

buffer/global/flat_load/store/atomic instructions to global memory are termed vector memory operations.

Private address space uses buffer_load/store using the scratch V# (GFX6-GFX8), or scratch_load/store (GFX9-GFX10). Since only a single thread is accessing the memory, atomic memory orderings are not meaningful, and all accesses are treated as non-atomic.

Constant address space uses buffer/global_load instructions (or equivalent scalar memory instructions). Since the constant address space contents do not change during the execution of a kernel dispatch it is not legal to perform stores, and atomic memory orderings are not meaningful, and all accesses are treated as non-atomic.

A memory synchronization scope wider than work-group is not meaningful for the group (LDS) address space and is treated as work-group.

The memory model does not support the region address space which is treated as non-atomic.

Acquire memory ordering is not meaningful on store atomic instructions and is treated as non-atomic.

Release memory ordering is not meaningful on load atomic instructions and is treated a non-atomic.

Acquire-release memory ordering is not meaningful on load or store atomic instructions and is treated as acquire and release respectively.

The memory order also adds the single thread optimization constraints defined in table AMDHSA Memory Model Single Thread Optimization Constraints.

AMDHSA Memory Model Single Thread Optimization Constraints
LLVM Memory Optimization Constraints
Ordering  
unordered none
monotonic none
acquire
  • If a load atomic/atomicrmw then no following load/load atomic/store/store atomic/atomicrmw/fence instruction can be moved before the acquire.
  • If a fence then same as load atomic, plus no preceding associated fence-paired-atomic can be moved after the fence.
release
  • If a store atomic/atomicrmw then no preceding load/load atomic/store/store atomic/atomicrmw/fence instruction can be moved after the release.
  • If a fence then same as store atomic, plus no following associated fence-paired-atomic can be moved before the fence.
acq_rel Same constraints as both acquire and release.
seq_cst
  • If a load atomic then same constraints as acquire, plus no preceding sequentially consistent load atomic/store atomic/atomicrmw/fence instruction can be moved after the seq_cst.
  • If a store atomic then the same constraints as release, plus no following sequentially consistent load atomic/store atomic/atomicrmw/fence instruction can be moved before the seq_cst.
  • If an atomicrmw/fence then same constraints as acq_rel.

The code sequences used to implement the memory model are defined in the following sections:

Memory Model GFX6-GFX9

For GFX6-GFX9:

  • Each agent has multiple shader arrays (SA).
  • Each SA has multiple compute units (CU).
  • Each CU has multiple SIMDs that execute wavefronts.
  • The wavefronts for a single work-group are executed in the same CU but may be executed by different SIMDs.
  • Each CU has a single LDS memory shared by the wavefronts of the work-groups executing on it.
  • All LDS operations of a CU are performed as wavefront wide operations in a global order and involve no caching. Completion is reported to a wavefront in execution order.
  • The LDS memory has multiple request queues shared by the SIMDs of a CU. Therefore, the LDS operations performed by different wavefronts of a work-group can be reordered relative to each other, which can result in reordering the visibility of vector memory operations with respect to LDS operations of other wavefronts in the same work-group. A s_waitcnt lgkmcnt(0) is required to ensure synchronization between LDS operations and vector memory operations between wavefronts of a work-group, but not between operations performed by the same wavefront.
  • The vector memory operations are performed as wavefront wide operations and completion is reported to a wavefront in execution order. The exception is that for GFX7-GFX9 flat_load/store/atomic instructions can report out of vector memory order if they access LDS memory, and out of LDS operation order if they access global memory.
  • The vector memory operations access a single vector L1 cache shared by all SIMDs a CU. Therefore, no special action is required for coherence between the lanes of a single wavefront, or for coherence between wavefronts in the same work-group. A buffer_wbinvl1_vol is required for coherence between wavefronts executing in different work-groups as they may be executing on different CUs.
  • The scalar memory operations access a scalar L1 cache shared by all wavefronts on a group of CUs. The scalar and vector L1 caches are not coherent. However, scalar operations are used in a restricted way so do not impact the memory model. See Memory Spaces.
  • The vector and scalar memory operations use an L2 cache shared by all CUs on the same agent.
  • The L2 cache has independent channels to service disjoint ranges of virtual addresses.
  • Each CU has a separate request queue per channel. Therefore, the vector and scalar memory operations performed by wavefronts executing in different work-groups (which may be executing on different CUs) of an agent can be reordered relative to each other. A s_waitcnt vmcnt(0) is required to ensure synchronization between vector memory operations of different CUs. It ensures a previous vector memory operation has completed before executing a subsequent vector memory or LDS operation and so can be used to meet the requirements of acquire and release.
  • The L2 cache can be kept coherent with other agents on some targets, or ranges of virtual addresses can be set up to bypass it to ensure system coherence.

Scalar memory operations are only used to access memory that is proven to not change during the execution of the kernel dispatch. This includes constant address space and global address space for program scope const variables. Therefore, the kernel machine code does not have to maintain the scalar cache to ensure it is coherent with the vector caches. The scalar and vector caches are invalidated between kernel dispatches by CP since constant address space data may change between kernel dispatch executions. See Memory Spaces.

The one exception is if scalar writes are used to spill SGPR registers. In this case the AMDGPU backend ensures the memory location used to spill is never accessed by vector memory operations at the same time. If scalar writes are used then a s_dcache_wb is inserted before the s_endpgm and before a function return since the locations may be used for vector memory instructions by a future wavefront that uses the same scratch area, or a function call that creates a frame at the same address, respectively. There is no need for a s_dcache_inv as all scalar writes are write-before-read in the same thread.

For kernarg backing memory:

  • CP invalidates the L1 cache at the start of each kernel dispatch.
  • On dGPU the kernarg backing memory is allocated in host memory accessed as MTYPE UC (uncached) to avoid needing to invalidate the L2 cache. This also causes it to be treated as non-volatile and so is not invalidated by *_vol.
  • On APU the kernarg backing memory it is accessed as MTYPE CC (cache coherent) and so the L2 cache will be coherent with the CPU and other agents.

Scratch backing memory (which is used for the private address space) is accessed with MTYPE NC_NV (non-coherent non-volatile). Since the private address space is only accessed by a single thread, and is always write-before-read, there is never a need to invalidate these entries from the L1 cache. Hence all cache invalidates are done as *_vol to only invalidate the volatile cache lines.

The code sequences used to implement the memory model for GFX6-GFX9 are defined in table AMDHSA Memory Model Code Sequences GFX6-GFX9.

AMDHSA Memory Model Code Sequences GFX6-GFX9
LLVM Instr LLVM Memory Ordering LLVM Memory Sync Scope AMDGPU Address Space AMDGPU Machine Code GFX6-9
Non-Atomic
load none none
  • global
  • generic
  • private
  • constant
  • !volatile & !nontemporal
    1. buffer/global/flat_load
  • volatile & !nontemporal
    1. buffer/global/flat_load glc=1
  • nontemporal
    1. buffer/global/flat_load glc=1 slc=1
load none none
  • local
  1. ds_load
store none none
  • global
  • generic
  • private
  • constant
  • !nontemporal
    1. buffer/global/flat_store
  • nontemporal
    1. buffer/global/flat_store glc=1 slc=1
store none none
  • local
  1. ds_store
Unordered Atomic
load atomic unordered any any Same as non-atomic.
store atomic unordered any any Same as non-atomic.
atomicrmw unordered any any Same as monotonic atomic.
Monotonic Atomic
load atomic monotonic
  • singlethread
  • wavefront
  • workgroup
  • global
  • local
  • generic
  1. buffer/global/ds/flat_load
load atomic monotonic
  • agent
  • system
  • global
  • generic
  1. buffer/global/flat_load glc=1
store atomic monotonic
  • singlethread
  • wavefront
  • workgroup
  • agent
  • system
  • global
  • generic
  1. buffer/global/flat_store
store atomic monotonic
  • singlethread
  • wavefront
  • workgroup
  • local
  1. ds_store
atomicrmw monotonic
  • singlethread
  • wavefront
  • workgroup
  • agent
  • system
  • global
  • generic
  1. buffer/global/flat_atomic
atomicrmw monotonic
  • singlethread
  • wavefront
  • workgroup
  • local
  1. ds_atomic
Acquire Atomic
load atomic acquire
  • singlethread
  • wavefront
  • global
  • local
  • generic
  1. buffer/global/ds/flat_load
load atomic acquire
  • workgroup
  • global
  1. buffer/global_load
load atomic acquire
  • workgroup
  • local
  • generic
  1. ds/flat_load
  2. s_waitcnt lgkmcnt(0)
  • If OpenCL, omit.
  • Must happen before any following global/generic load/load atomic/store/store atomic/atomicrmw.
  • Ensures any following global data read is no older than a local load atomic value being acquired.
load atomic acquire
  • agent
  • system
  • global
  1. buffer/global_load glc=1
  2. s_waitcnt vmcnt(0)
  • Must happen before following buffer_wbinvl1_vol.
  • Ensures the load has completed before invalidating the cache.
  1. buffer_wbinvl1_vol
  • Must happen before any following global/generic load/load atomic/atomicrmw.
  • Ensures that following loads will not see stale global data.
load atomic acquire
  • agent
  • system
  • generic
  1. flat_load glc=1
  2. s_waitcnt vmcnt(0) & lgkmcnt(0)
  • If OpenCL omit lgkmcnt(0).
  • Must happen before following buffer_wbinvl1_vol.
  • Ensures the flat_load has completed before invalidating the cache.
  1. buffer_wbinvl1_vol
  • Must happen before any following global/generic load/load atomic/atomicrmw.
  • Ensures that following loads will not see stale global data.
atomicrmw acquire
  • singlethread
  • wavefront
  • global
  • local
  • generic
  1. buffer/global/ds/flat_atomic
atomicrmw acquire
  • workgroup
  • global
  1. buffer/global_atomic
atomicrmw acquire
  • workgroup
  • local
  • generic
  1. ds/flat_atomic
  2. s_waitcnt lgkmcnt(0)
  • If OpenCL, omit.
  • Must happen before any following global/generic load/load atomic/store/store atomic/atomicrmw.
  • Ensures any following global data read is no older than a local atomicrmw value being acquired.
atomicrmw acquire
  • agent
  • system
  • global
  1. buffer/global_atomic
  2. s_waitcnt vmcnt(0)
  • Must happen before following buffer_wbinvl1_vol.
  • Ensures the atomicrmw has completed before invalidating the cache.
  1. buffer_wbinvl1_vol
  • Must happen before any following global/generic load/load atomic/atomicrmw.
  • Ensures that following loads will not see stale global data.
atomicrmw acquire
  • agent
  • system
  • generic
  1. flat_atomic
  2. s_waitcnt vmcnt(0) & lgkmcnt(0)
  • If OpenCL, omit lgkmcnt(0).
  • Must happen before following buffer_wbinvl1_vol.
  • Ensures the atomicrmw has completed before invalidating the cache.
  1. buffer_wbinvl1_vol
  • Must happen before any following global/generic load/load atomic/atomicrmw.
  • Ensures that following loads will not see stale global data.
fence acquire
  • singlethread
  • wavefront
 none none
fence acquire
  • workgroup
 none
  1. s_waitcnt lgkmcnt(0)
  • If OpenCL and address space is not generic, omit.
  • However, since LLVM currently has no address space on the fence need to conservatively always generate. If fence had an address space then set to address space of OpenCL fence flag, or to generic if both local and global flags are specified.
  • Must happen after any preceding local/generic load atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the fence-paired-atomic).
  • Must happen before any following global/generic load/load atomic/store/store atomic/atomicrmw.
  • Ensures any following global data read is no older than the value read by the fence-paired-atomic.
fence acquire
  • agent
  • system
 none
  1. s_waitcnt lgkmcnt(0) & vmcnt(0)
  • If OpenCL and address space is not generic, omit lgkmcnt(0).
  • However, since LLVM currently has no address space on the fence need to conservatively always generate (see comment for previous fence).
  • Could be split into separate s_waitcnt vmcnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the fence-paired-atomic).
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the fence-paired-atomic).
  • Must happen before the following buffer_wbinvl1_vol.
  • Ensures that the fence-paired atomic has completed before invalidating the cache. Therefore any following locations read must be no older than the value read by the fence-paired-atomic.
  1. buffer_wbinvl1_vol
  • Must happen before any following global/generic load/load atomic/store/store atomic/atomicrmw.
  • Ensures that following loads will not see stale global data.
Release Atomic
store atomic release
  • singlethread
  • wavefront
  • global
  • local
  • generic
  1. buffer/global/ds/flat_store
store atomic release
  • workgroup
  • global
  • generic
  1. s_waitcnt lgkmcnt(0)
  • If OpenCL, omit.
  • Must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following store.
  • Ensures that all memory operations to local have completed before performing the store that is being released.
  1. buffer/global/flat_store
store atomic release
  • workgroup
  • local
  1. ds_store
store atomic release
  • agent
  • system
  • global
  • generic
  1. s_waitcnt lgkmcnt(0) & vmcnt(0)
  • If OpenCL and address space is not generic, omit lgkmcnt(0).
  • Could be split into separate s_waitcnt vmcnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/store/load atomic/store atomic/atomicrmw.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following store.
  • Ensures that all memory operations to memory have completed before performing the store that is being released.
  1. buffer/global/flat_store
atomicrmw release
  • singlethread
  • wavefront
  • global
  • local
  • generic
  1. buffer/global/ds/flat_atomic
atomicrmw release
  • workgroup
  • global
  • generic
  1. s_waitcnt lgkmcnt(0)
  • If OpenCL, omit.
  • Must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following atomicrmw.
  • Ensures that all memory operations to local have completed before performing the atomicrmw that is being released.
  1. buffer/global/flat_atomic
atomicrmw release
  • workgroup
  • local
  1. ds_atomic
atomicrmw release
  • agent
  • system
  • global
  • generic
  1. s_waitcnt lgkmcnt(0) & vmcnt(0)
  • If OpenCL, omit lgkmcnt(0).
  • Could be split into separate s_waitcnt vmcnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/store/load atomic/store atomic/atomicrmw.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following atomicrmw.
  • Ensures that all memory operations to global and local have completed before performing the atomicrmw that is being released.
  1. buffer/global/flat_atomic
fence release
  • singlethread
  • wavefront
 none none
fence release
  • workgroup
 none
  1. s_waitcnt lgkmcnt(0)
  • If OpenCL and address space is not generic, omit.
  • However, since LLVM currently has no address space on the fence need to conservatively always generate. If fence had an address space then set to address space of OpenCL fence flag, or to generic if both local and global flags are specified.
  • Must happen after any preceding local/generic load/load atomic/store/store atomic/atomicrmw.
  • Must happen before any following store atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the fence-paired-atomic).
  • Ensures that all memory operations to local have completed before performing the following fence-paired-atomic.
fence release
  • agent
  • system
 none
  1. s_waitcnt lgkmcnt(0) & vmcnt(0)
  • If OpenCL and address space is not generic, omit lgkmcnt(0).
  • If OpenCL and address space is local, omit vmcnt(0).
  • However, since LLVM currently has no address space on the fence need to conservatively always generate. If fence had an address space then set to address space of OpenCL fence flag, or to generic if both local and global flags are specified.
  • Could be split into separate s_waitcnt vmcnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/store/load atomic/store atomic/atomicrmw.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before any following store atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the fence-paired-atomic).
  • Ensures that all memory operations have completed before performing the following fence-paired-atomic.
Acquire-Release Atomic
atomicrmw acq_rel
  • singlethread
  • wavefront
  • global
  • local
  • generic
  1. buffer/global/ds/flat_atomic
atomicrmw acq_rel
  • workgroup
  • global
  1. s_waitcnt lgkmcnt(0)
  • If OpenCL, omit.
  • Must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following atomicrmw.
  • Ensures that all memory operations to local have completed before performing the atomicrmw that is being released.
  1. buffer/global_atomic
atomicrmw acq_rel
  • workgroup
  • local
  1. ds_atomic
  2. s_waitcnt lgkmcnt(0)
  • If OpenCL, omit.
  • Must happen before any following global/generic load/load atomic/store/store atomic/atomicrmw.
  • Ensures any following global data read is no older than the local load atomic value being acquired.
atomicrmw acq_rel
  • workgroup
  • generic
  1. s_waitcnt lgkmcnt(0)
  • If OpenCL, omit.
  • Must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following atomicrmw.
  • Ensures that all memory operations to local have completed before performing the atomicrmw that is being released.
  1. flat_atomic
  2. s_waitcnt lgkmcnt(0)
  • If OpenCL, omit.
  • Must happen before any following global/generic load/load atomic/store/store atomic/atomicrmw.
  • Ensures any following global data read is no older than a local load atomic value being acquired.
atomicrmw acq_rel
  • agent
  • system
  • global
  1. s_waitcnt lgkmcnt(0) & vmcnt(0)
  • If OpenCL, omit lgkmcnt(0).
  • Could be split into separate s_waitcnt vmcnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/store/load atomic/store atomic/atomicrmw.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following atomicrmw.
  • Ensures that all memory operations to global have completed before performing the atomicrmw that is being released.
  1. buffer/global_atomic
  2. s_waitcnt vmcnt(0)
  • Must happen before following buffer_wbinvl1_vol.
  • Ensures the atomicrmw has completed before invalidating the cache.
  1. buffer_wbinvl1_vol
  • Must happen before any following global/generic load/load atomic/atomicrmw.
  • Ensures that following loads will not see stale global data.
atomicrmw acq_rel
  • agent
  • system
  • generic
  1. s_waitcnt lgkmcnt(0) & vmcnt(0)
  • If OpenCL, omit lgkmcnt(0).
  • Could be split into separate s_waitcnt vmcnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/store/load atomic/store atomic/atomicrmw.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following atomicrmw.
  • Ensures that all memory operations to global have completed before performing the atomicrmw that is being released.
  1. flat_atomic
  2. s_waitcnt vmcnt(0) & lgkmcnt(0)
  • If OpenCL, omit lgkmcnt(0).
  • Must happen before following buffer_wbinvl1_vol.
  • Ensures the atomicrmw has completed before invalidating the cache.
  1. buffer_wbinvl1_vol
  • Must happen before any following global/generic load/load atomic/atomicrmw.
  • Ensures that following loads will not see stale global data.
fence acq_rel
  • singlethread
  • wavefront
 none none
fence acq_rel
  • workgroup
 none
  1. s_waitcnt lgkmcnt(0)
  • If OpenCL and address space is not generic, omit.
  • However, since LLVM currently has no address space on the fence need to conservatively always generate (see comment for previous fence).
  • Must happen after any preceding local/generic load/load atomic/store/store atomic/atomicrmw.
  • Must happen before any following global/generic load/load atomic/store/store atomic/atomicrmw.
  • Ensures that all memory operations to local have completed before performing any following global memory operations.
  • Ensures that the preceding local/generic load atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the acquire-fence-paired-atomic) has completed before following global memory operations. This satisfies the requirements of acquire.
  • Ensures that all previous memory operations have completed before a following local/generic store atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the release-fence-paired-atomic). This satisfies the requirements of release.
fence acq_rel
  • agent
  • system
 none
  1. s_waitcnt lgkmcnt(0) & vmcnt(0)
  • If OpenCL and address space is not generic, omit lgkmcnt(0).
  • However, since LLVM currently has no address space on the fence need to conservatively always generate (see comment for previous fence).
  • Could be split into separate s_waitcnt vmcnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/store/load atomic/store atomic/atomicrmw.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following buffer_wbinvl1_vol.
  • Ensures that the preceding global/local/generic load atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the acquire-fence-paired-atomic) has completed before invalidating the cache. This satisfies the requirements of acquire.
  • Ensures that all previous memory operations have completed before a following global/local/generic store atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the release-fence-paired-atomic). This satisfies the requirements of release.
  1. buffer_wbinvl1_vol
  • Must happen before any following global/generic load/load atomic/store/store atomic/atomicrmw.
  • Ensures that following loads will not see stale global data. This satisfies the requirements of acquire.
Sequential Consistent Atomic
load atomic seq_cst
  • singlethread
  • wavefront
  • global
  • local
  • generic
 Same as corresponding load atomic acquire, except must generated all instructions even for OpenCL.
load atomic seq_cst
  • workgroup
  • global
  • generic
  1. s_waitcnt lgkmcnt(0)
  • Must happen after preceding local/generic load atomic/store atomic/atomicrmw with memory ordering of seq_cst and with equal or wider sync scope. (Note that seq_cst fences have their own s_waitcnt lgkmcnt(0) and so do not need to be considered.)
  • Ensures any preceding sequential consistent local memory instructions have completed before executing this sequentially consistent instruction. This prevents reordering a seq_cst store followed by a seq_cst load. (Note that seq_cst is stronger than acquire/release as the reordering of load acquire followed by a store release is prevented by the s_waitcnt of the release, but there is nothing preventing a store release followed by load acquire from completing out of order. The s_waitcnt could be placed after seq_store or before the seq_load. We choose the load to make the s_waitcnt be as late as possible so that the store may have already completed.)
  1. Following instructions same as corresponding load atomic acquire, except must generated all instructions even for OpenCL.
load atomic seq_cst
  • workgroup
  • local
 Same as corresponding load atomic acquire, except must generated all instructions even for OpenCL.
load atomic seq_cst
  • agent
  • system
  • global
  • generic
  1. s_waitcnt lgkmcnt(0) & vmcnt(0)
  • Could be split into separate s_waitcnt vmcnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt lgkmcnt(0) must happen after preceding global/generic load atomic/store atomic/atomicrmw with memory ordering of seq_cst and with equal or wider sync scope. (Note that seq_cst fences have their own s_waitcnt lgkmcnt(0) and so do not need to be considered.)
  • s_waitcnt vmcnt(0) must happen after preceding global/generic load atomic/store atomic/atomicrmw with memory ordering of seq_cst and with equal or wider sync scope. (Note that seq_cst fences have their own s_waitcnt vmcnt(0) and so do not need to be considered.)
  • Ensures any preceding sequential consistent global memory instructions have completed before executing this sequentially consistent instruction. This prevents reordering a seq_cst store followed by a seq_cst load. (Note that seq_cst is stronger than acquire/release as the reordering of load acquire followed by a store release is prevented by the s_waitcnt of the release, but there is nothing preventing a store release followed by load acquire from completing out of order. The s_waitcnt could be placed after seq_store or before the seq_load. We choose the load to make the s_waitcnt be as late as possible so that the store may have already completed.)
  1. Following instructions same as corresponding load atomic acquire, except must generated all instructions even for OpenCL.
store atomic seq_cst
  • singlethread
  • wavefront
  • workgroup
  • agent
  • system
  • global
  • local
  • generic
 Same as corresponding store atomic release, except must generated all instructions even for OpenCL.
atomicrmw seq_cst
  • singlethread
  • wavefront
  • workgroup
  • agent
  • system
  • global
  • local
  • generic
 Same as corresponding atomicrmw acq_rel, except must generated all instructions even for OpenCL.
fence seq_cst
  • singlethread
  • wavefront
  • workgroup
  • agent
  • system
 none Same as corresponding fence acq_rel, except must generated all instructions even for OpenCL.
Memory Model GFX10

For GFX10:

  • Each agent has multiple shader arrays (SA).
  • Each SA has multiple work-group processors (WGP).
  • Each WGP has multiple compute units (CU).
  • Each CU has multiple SIMDs that execute wavefronts.
  • The wavefronts for a single work-group are executed in the same WGP. In CU wavefront execution mode the wavefronts may be executed by different SIMDs in the same CU. In WGP wavefront execution mode the wavefronts may be executed by different SIMDs in different CUs in the same WGP.
  • Each WGP has a single LDS memory shared by the wavefronts of the work-groups executing on it.
  • All LDS operations of a WGP are performed as wavefront wide operations in a global order and involve no caching. Completion is reported to a wavefront in execution order.
  • The LDS memory has multiple request queues shared by the SIMDs of a WGP. Therefore, the LDS operations performed by different wavefronts of a work-group can be reordered relative to each other, which can result in reordering the visibility of vector memory operations with respect to LDS operations of other wavefronts in the same work-group. A s_waitcnt lgkmcnt(0) is required to ensure synchronization between LDS operations and vector memory operations between wavefronts of a work-group, but not between operations performed by the same wavefront.
  • The vector memory operations are performed as wavefront wide operations. Completion of load/store/sample operations are reported to a wavefront in execution order of other load/store/sample operations performed by that wavefront.
  • The vector memory operations access a vector L0 cache. There is a single L0 cache per CU. Each SIMD of a CU accesses the same L0 cache. Therefore, no special action is required for coherence between the lanes of a single wavefront. However, a buffer_gl0_inv is required for coherence between wavefronts executing in the same work-group as they may be executing on SIMDs of different CUs that access different L0s. A buffer_gl0_inv is also required for coherence between wavefronts executing in different work-groups as they may be executing on different WGPs.
  • The scalar memory operations access a scalar L0 cache shared by all wavefronts on a WGP. The scalar and vector L0 caches are not coherent. However, scalar operations are used in a restricted way so do not impact the memory model. See Memory Spaces.
  • The vector and scalar memory L0 caches use an L1 cache shared by all WGPs on the same SA. Therefore, no special action is required for coherence between the wavefronts of a single work-group. However, a buffer_gl1_inv is required for coherence between wavefronts executing in different work-groups as they may be executing on different SAs that access different L1s.
  • The L1 caches have independent quadrants to service disjoint ranges of virtual addresses.
  • Each L0 cache has a separate request queue per L1 quadrant. Therefore, the vector and scalar memory operations performed by different wavefronts, whether executing in the same or different work-groups (which may be executing on different CUs accessing different L0s), can be reordered relative to each other. A s_waitcnt vmcnt(0) & vscnt(0) is required to ensure synchronization between vector memory operations of different wavefronts. It ensures a previous vector memory operation has completed before executing a subsequent vector memory or LDS operation and so can be used to meet the requirements of acquire, release and sequential consistency.
  • The L1 caches use an L2 cache shared by all SAs on the same agent.
  • The L2 cache has independent channels to service disjoint ranges of virtual addresses.
  • Each L1 quadrant of a single SA accesses a different L2 channel. Each L1 quadrant has a separate request queue per L2 channel. Therefore, the vector and scalar memory operations performed by wavefronts executing in different work-groups (which may be executing on different SAs) of an agent can be reordered relative to each other. A s_waitcnt vmcnt(0) & vscnt(0) is required to ensure synchronization between vector memory operations of different SAs. It ensures a previous vector memory operation has completed before executing a subsequent vector memory and so can be used to meet the requirements of acquire, release and sequential consistency.
  • The L2 cache can be kept coherent with other agents on some targets, or ranges of virtual addresses can be set up to bypass it to ensure system coherence.

Scalar memory operations are only used to access memory that is proven to not change during the execution of the kernel dispatch. This includes constant address space and global address space for program scope const variables. Therefore, the kernel machine code does not have to maintain the scalar cache to ensure it is coherent with the vector caches. The scalar and vector caches are invalidated between kernel dispatches by CP since constant address space data may change between kernel dispatch executions. See Memory Spaces.

The one exception is if scalar writes are used to spill SGPR registers. In this case the AMDGPU backend ensures the memory location used to spill is never accessed by vector memory operations at the same time. If scalar writes are used then a s_dcache_wb is inserted before the s_endpgm and before a function return since the locations may be used for vector memory instructions by a future wavefront that uses the same scratch area, or a function call that creates a frame at the same address, respectively. There is no need for a s_dcache_inv as all scalar writes are write-before-read in the same thread.

For kernarg backing memory:

  • CP invalidates the L0 and L1 caches at the start of each kernel dispatch.
  • On dGPU the kernarg backing memory is accessed as MTYPE UC (uncached) to avoid needing to invalidate the L2 cache.
  • On APU the kernarg backing memory is accessed as MTYPE CC (cache coherent) and so the L2 cache will be coherent with the CPU and other agents.

Scratch backing memory (which is used for the private address space) is accessed with MTYPE NC (non-coherent). Since the private address space is only accessed by a single thread, and is always write-before-read, there is never a need to invalidate these entries from the L0 or L1 caches.

Wavefronts are executed in native mode with in-order reporting of loads and sample instructions. In this mode vmcnt reports completion of load, atomic with return and sample instructions in order, and the vscnt reports the completion of store and atomic without return in order. See MEM_ORDERED field in compute_pgm_rsrc1 for GFX6-GFX10.

Wavefronts can be executed in WGP or CU wavefront execution mode:

  • In WGP wavefront execution mode the wavefronts of a work-group are executed on the SIMDs of both CUs of the WGP. Therefore, explicit management of the per CU L0 caches is required for work-group synchronization. Also accesses to L1 at work-group scope need to be explicitly ordered as the accesses from different CUs are not ordered.
  • In CU wavefront execution mode the wavefronts of a work-group are executed on the SIMDs of a single CU of the WGP. Therefore, all global memory access by the work-group access the same L0 which in turn ensures L1 accesses are ordered and so do not require explicit management of the caches for work-group synchronization.

See WGP_MODE field in compute_pgm_rsrc1 for GFX6-GFX10 and Target Features.

The code sequences used to implement the memory model for GFX10 are defined in table AMDHSA Memory Model Code Sequences GFX10.

AMDHSA Memory Model Code Sequences GFX10
LLVM Instr LLVM Memory Ordering LLVM Memory Sync Scope AMDGPU Address Space AMDGPU Machine Code GFX10
Non-Atomic
load none none
  • global
  • generic
  • private
  • constant
  • !volatile & !nontemporal
    1. buffer/global/flat_load
  • volatile & !nontemporal
    1. buffer/global/flat_load glc=1 dlc=1
  • nontemporal
    1. buffer/global/flat_load slc=1
load none none
  • local
  1. ds_load
store none none
  • global
  • generic
  • private
  • constant

  • !nontemporal

    1. buffer/global/flat_store

  • nontemporal

    1. buffer/global/flat_store slc=1
store none none
  • local
  1. ds_store
Unordered Atomic
load atomic unordered any any Same as non-atomic.
store atomic unordered any any Same as non-atomic.
atomicrmw unordered any any Same as monotonic atomic.
Monotonic Atomic
load atomic monotonic
  • singlethread
  • wavefront
  • global
  • generic
  1. buffer/global/flat_load
load atomic monotonic
  • workgroup
  • global
  • generic
  1. buffer/global/flat_load glc=1
  • If CU wavefront execution mode, omit glc=1.
load atomic monotonic
  • singlethread
  • wavefront
  • workgroup
  • local
  1. ds_load
load atomic monotonic
  • agent
  • system
  • global
  • generic
  1. buffer/global/flat_load glc=1 dlc=1
store atomic monotonic
  • singlethread
  • wavefront
  • workgroup
  • agent
  • system
  • global
  • generic
  1. buffer/global/flat_store
store atomic monotonic
  • singlethread
  • wavefront
  • workgroup
  • local
  1. ds_store
atomicrmw monotonic
  • singlethread
  • wavefront
  • workgroup
  • agent
  • system
  • global
  • generic
  1. buffer/global/flat_atomic
atomicrmw monotonic
  • singlethread
  • wavefront
  • workgroup
  • local
  1. ds_atomic
Acquire Atomic
load atomic acquire
  • singlethread
  • wavefront
  • global
  • local
  • generic
  1. buffer/global/ds/flat_load
load atomic acquire
  • workgroup
  • global
  1. buffer/global_load glc=1
  • If CU wavefront execution mode, omit glc=1.
  1. s_waitcnt vmcnt(0)
  • If CU wavefront execution mode, omit.
  • Must happen before the following buffer_gl0_inv and before any following global/generic load/load atomic/store/store atomic/atomicrmw.
  1. buffer_gl0_inv
  • If CU wavefront execution mode, omit.
  • Ensures that following loads will not see stale data.
load atomic acquire
  • workgroup
  • local
  1. ds_load
  2. s_waitcnt lgkmcnt(0)
  • If OpenCL, omit.
  • Must happen before the following buffer_gl0_inv and before any following global/generic load/load atomic/store/store atomic/atomicrmw.
  • Ensures any following global data read is no older than the local load atomic value being acquired.
  1. buffer_gl0_inv
  • If CU wavefront execution mode, omit.
  • If OpenCL, omit.
  • Ensures that following loads will not see stale data.
load atomic acquire
  • workgroup
  • generic
  1. flat_load glc=1
  • If CU wavefront execution mode, omit glc=1.
  1. s_waitcnt lgkmcnt(0) & vmcnt(0)
  • If CU wavefront execution mode, omit vmcnt(0).
  • If OpenCL, omit lgkmcnt(0).
  • Must happen before the following buffer_gl0_inv and any following global/generic load/load atomic/store/store atomic/atomicrmw.
  • Ensures any following global data read is no older than a local load atomic value being acquired.
  1. buffer_gl0_inv
  • If CU wavefront execution mode, omit.
  • Ensures that following loads will not see stale data.
load atomic acquire
  • agent
  • system
  • global
  1. buffer/global_load glc=1 dlc=1
  2. s_waitcnt vmcnt(0)
  • Must happen before following buffer_gl*_inv.
  • Ensures the load has completed before invalidating the caches.
  1. buffer_gl0_inv; buffer_gl1_inv
  • Must happen before any following global/generic load/load atomic/atomicrmw.
  • Ensures that following loads will not see stale global data.
load atomic acquire
  • agent
  • system
  • generic
  1. flat_load glc=1 dlc=1
  2. s_waitcnt vmcnt(0) & lgkmcnt(0)
  • If OpenCL omit lgkmcnt(0).
  • Must happen before following buffer_gl*_invl.
  • Ensures the flat_load has completed before invalidating the caches.
  1. buffer_gl0_inv; buffer_gl1_inv
  • Must happen before any following global/generic load/load atomic/atomicrmw.
  • Ensures that following loads will not see stale global data.
atomicrmw acquire
  • singlethread
  • wavefront
  • global
  • local
  • generic
  1. buffer/global/ds/flat_atomic
atomicrmw acquire
  • workgroup
  • global
  1. buffer/global_atomic
  2. s_waitcnt vm/vscnt(0)
  • If CU wavefront execution mode, omit.
  • Use vmcnt(0) if atomic with return and vscnt(0) if atomic with no-return.
  • Must happen before the following buffer_gl0_inv and before any following global/generic load/load atomic/store/store atomic/atomicrmw.
  1. buffer_gl0_inv
  • If CU wavefront execution mode, omit.
  • Ensures that following loads will not see stale data.
atomicrmw acquire
  • workgroup
  • local
  1. ds_atomic
  2. s_waitcnt lgkmcnt(0)
  • If OpenCL, omit.
  • Must happen before the following buffer_gl0_inv.
  • Ensures any following global data read is no older than the local atomicrmw value being acquired.
  1. buffer_gl0_inv
  • If OpenCL omit.
  • Ensures that following loads will not see stale data.
atomicrmw acquire
  • workgroup
  • generic
  1. flat_atomic
  2. s_waitcnt lgkmcnt(0) & vm/vscnt(0)
  • If CU wavefront execution mode, omit vm/vscnt(0).
  • If OpenCL, omit lgkmcnt(0).
  • Use vmcnt(0) if atomic with return and vscnt(0) if atomic with no-return.
  • Must happen before the following buffer_gl0_inv.
  • Ensures any following global data read is no older than a local atomicrmw value being acquired.
  1. buffer_gl0_inv
  • If CU wavefront execution mode, omit.
  • Ensures that following loads will not see stale data.
atomicrmw acquire
  • agent
  • system
  • global
  1. buffer/global_atomic
  2. s_waitcnt vm/vscnt(0)
  • Use vmcnt(0) if atomic with return and vscnt(0) if atomic with no-return.
  • Must happen before following buffer_gl*_inv.
  • Ensures the atomicrmw has completed before invalidating the caches.
  1. buffer_gl0_inv; buffer_gl1_inv
  • Must happen before any following global/generic load/load atomic/atomicrmw.
  • Ensures that following loads will not see stale global data.
atomicrmw acquire
  • agent
  • system
  • generic
  1. flat_atomic
  2. s_waitcnt vm/vscnt(0) & lgkmcnt(0)
  • If OpenCL, omit lgkmcnt(0).
  • Use vmcnt(0) if atomic with return and vscnt(0) if atomic with no-return.
  • Must happen before following buffer_gl*_inv.
  • Ensures the atomicrmw has completed before invalidating the caches.
  1. buffer_gl0_inv; buffer_gl1_inv
  • Must happen before any following global/generic load/load atomic/atomicrmw.
  • Ensures that following loads will not see stale global data.
fence acquire
  • singlethread
  • wavefront
 none none
fence acquire
  • workgroup
 none
  1. s_waitcnt lgkmcnt(0) & vmcnt(0) & vscnt(0)
  • If CU wavefront execution mode, omit vmcnt(0) and vscnt(0).
  • If OpenCL and address space is not generic, omit lgkmcnt(0).
  • If OpenCL and address space is local, omit vmcnt(0) and vscnt(0).
  • However, since LLVM currently has no address space on the fence need to conservatively always generate. If fence had an address space then set to address space of OpenCL fence flag, or to generic if both local and global flags are specified.
  • Could be split into separate s_waitcnt vmcnt(0), s_waitcnt vscnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load atomic/ atomicrmw-with-return-value with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the fence-paired-atomic).
  • s_waitcnt vscnt(0) must happen after any preceding global/generic atomicrmw-no-return-value with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the fence-paired-atomic).
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the fence-paired-atomic).
  • Must happen before the following buffer_gl0_inv.
  • Ensures that the fence-paired atomic has completed before invalidating the cache. Therefore any following locations read must be no older than the value read by the fence-paired-atomic.
  1. buffer_gl0_inv
  • If CU wavefront execution mode, omit.
  • Ensures that following loads will not see stale data.
fence acquire
  • agent
  • system
 none
  1. s_waitcnt lgkmcnt(0) & vmcnt(0) & vscnt(0)
  • If OpenCL and address space is not generic, omit lgkmcnt(0).
  • If OpenCL and address space is local, omit vmcnt(0) and vscnt(0).
  • However, since LLVM currently has no address space on the fence need to conservatively always generate (see comment for previous fence).
  • Could be split into separate s_waitcnt vmcnt(0), s_waitcnt vscnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load atomic/ atomicrmw-with-return-value with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the fence-paired-atomic).
  • s_waitcnt vscnt(0) must happen after any preceding global/generic atomicrmw-no-return-value with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the fence-paired-atomic).
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the fence-paired-atomic).
  • Must happen before the following buffer_gl*_inv.
  • Ensures that the fence-paired atomic has completed before invalidating the caches. Therefore any following locations read must be no older than the value read by the fence-paired-atomic.
  1. buffer_gl0_inv; buffer_gl1_inv
  • Must happen before any following global/generic load/load atomic/store/store atomic/atomicrmw.
  • Ensures that following loads will not see stale global data.
Release Atomic
store atomic release
  • singlethread
  • wavefront
  • global
  • local
  • generic
  1. buffer/global/ds/flat_store
store atomic release
  • workgroup
  • global
  • generic
  1. s_waitcnt lgkmcnt(0) & vmcnt(0) & vscnt(0)
  • If CU wavefront execution mode, omit vmcnt(0) and vscnt(0).
  • If OpenCL, omit lgkmcnt(0).
  • Could be split into separate s_waitcnt vmcnt(0), s_waitcnt vscnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/load atomic/ atomicrmw-with-return-value.
  • s_waitcnt vscnt(0) must happen after any preceding global/generic store/store atomic/ atomicrmw-no-return-value.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following store.
  • Ensures that all memory operations have completed before performing the store that is being released.
  1. buffer/global/flat_store
store atomic release
  • workgroup
  • local
  1. s_waitcnt vmcnt(0) & vscnt(0)
  • If CU wavefront execution mode, omit.
  • If OpenCL, omit.
  • Could be split into separate s_waitcnt vmcnt(0) and s_waitcnt vscnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/load atomic/ atomicrmw-with-return-value.
  • s_waitcnt vscnt(0) must happen after any preceding global/generic store/store atomic/ atomicrmw-no-return-value.
  • Must happen before the following store.
  • Ensures that all global memory operations have completed before performing the store that is being released.
  1. ds_store
store atomic release
  • agent
  • system
  • global
  • generic
  1. s_waitcnt lgkmcnt(0) & vmcnt(0) & vscnt(0)
  • If OpenCL and address space is not generic, omit lgkmcnt(0).
  • Could be split into separate s_waitcnt vmcnt(0), s_waitcnt vscnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/load atomic/ atomicrmw-with-return-value.
  • s_waitcnt vscnt(0) must happen after any preceding global/generic store/store atomic/ atomicrmw-no-return-value.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following store.
  • Ensures that all memory operations have completed before performing the store that is being released.
  1. buffer/global/flat_store
atomicrmw release
  • singlethread
  • wavefront
  • global
  • local
  • generic
  1. buffer/global/ds/flat_atomic
atomicrmw release
  • workgroup
  • global
  • generic
  1. s_waitcnt lgkmcnt(0) & vmcnt(0) & vscnt(0)
  • If CU wavefront execution mode, omit vmcnt(0) and vscnt(0).
  • If OpenCL, omit lgkmcnt(0).
  • Could be split into separate s_waitcnt vmcnt(0), s_waitcnt vscnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/load atomic/ atomicrmw-with-return-value.
  • s_waitcnt vscnt(0) must happen after any preceding global/generic store/store atomic/ atomicrmw-no-return-value.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following atomicrmw.
  • Ensures that all memory operations have completed before performing the atomicrmw that is being released.
  1. buffer/global/flat_atomic
atomicrmw release
  • workgroup
  • local
  1. s_waitcnt vmcnt(0) & vscnt(0)
  • If CU wavefront execution mode, omit.
  • If OpenCL, omit.
  • Could be split into separate s_waitcnt vmcnt(0) and s_waitcnt vscnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/load atomic/ atomicrmw-with-return-value.
  • s_waitcnt vscnt(0) must happen after any preceding global/generic store/store atomic/ atomicrmw-no-return-value.
  • Must happen before the following store.
  • Ensures that all global memory operations have completed before performing the store that is being released.
  1. ds_atomic
atomicrmw release
  • agent
  • system
  • global
  • generic
  1. s_waitcnt lgkmcnt(0) &
    vmcnt(0) & vscnt(0)
  • If OpenCL, omit lgkmcnt(0).
  • Could be split into separate s_waitcnt vmcnt(0), s_waitcnt vscnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/load atomic/ atomicrmw-with-return-value.
  • s_waitcnt vscnt(0) must happen after any preceding global/generic store/store atomic/ atomicrmw-no-return-value.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following atomicrmw.
  • Ensures that all memory operations to global and local have completed before performing the atomicrmw that is being released.
  1. buffer/global/flat_atomic
fence release
  • singlethread
  • wavefront
 none none
fence release
  • workgroup
 none
  1. s_waitcnt lgkmcnt(0) & vmcnt(0) & vscnt(0)
  • If CU wavefront execution mode, omit vmcnt(0) and vscnt(0).
  • If OpenCL and address space is not generic, omit lgkmcnt(0).
  • If OpenCL and address space is local, omit vmcnt(0) and vscnt(0).
  • However, since LLVM currently has no address space on the fence need to conservatively always generate. If fence had an address space then set to address space of OpenCL fence flag, or to generic if both local and global flags are specified.
  • Could be split into separate s_waitcnt vmcnt(0), s_waitcnt vscnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/load atomic/ atomicrmw-with-return-value.
  • s_waitcnt vscnt(0) must happen after any preceding global/generic store/store atomic/ atomicrmw-no-return-value.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/ atomicrmw.
  • Must happen before any following store atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the fence-paired-atomic).
  • Ensures that all memory operations have completed before performing the following fence-paired-atomic.
fence release
  • agent
  • system
 none
  1. s_waitcnt lgkmcnt(0) & vmcnt(0) & vscnt(0)
  • If OpenCL and address space is not generic, omit lgkmcnt(0).
  • If OpenCL and address space is local, omit vmcnt(0) and vscnt(0).
  • However, since LLVM currently has no address space on the fence need to conservatively always generate. If fence had an address space then set to address space of OpenCL fence flag, or to generic if both local and global flags are specified.
  • Could be split into separate s_waitcnt vmcnt(0), s_waitcnt vscnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/load atomic/ atomicrmw-with-return-value.
  • s_waitcnt vscnt(0) must happen after any preceding global/generic store/store atomic/ atomicrmw-no-return-value.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before any following store atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the fence-paired-atomic).
  • Ensures that all memory operations have completed before performing the following fence-paired-atomic.
Acquire-Release Atomic
atomicrmw acq_rel
  • singlethread
  • wavefront
  • global
  • local
  • generic
  1. buffer/global/ds/flat_atomic
atomicrmw acq_rel
  • workgroup
  • global
  1. s_waitcnt lgkmcnt(0) & vmcnt(0) & vscnt(0)
  • If CU wavefront execution mode, omit vmcnt(0) and vscnt(0).
  • If OpenCL, omit lgkmcnt(0).
  • Must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Could be split into separate s_waitcnt vmcnt(0), s_waitcnt vscnt(0), and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/load atomic/ atomicrmw-with-return-value.
  • s_waitcnt vscnt(0) must happen after any preceding global/generic store/store atomic/ atomicrmw-no-return-value.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following atomicrmw.
  • Ensures that all memory operations have completed before performing the atomicrmw that is being released.
  1. buffer/global_atomic
  2. s_waitcnt vm/vscnt(0)
  • If CU wavefront execution mode, omit.
  • Use vmcnt(0) if atomic with return and vscnt(0) if atomic with no-return.
  • Must happen before the following buffer_gl0_inv.
  • Ensures any following global data read is no older than the atomicrmw value being acquired.
  1. buffer_gl0_inv
  • If CU wavefront execution mode, omit.
  • Ensures that following loads will not see stale data.
atomicrmw acq_rel
  • workgroup
  • local
  1. s_waitcnt vmcnt(0) & vscnt(0)
  • If CU wavefront execution mode, omit.
  • If OpenCL, omit.
  • Could be split into separate s_waitcnt vmcnt(0) and s_waitcnt vscnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/load atomic/ atomicrmw-with-return-value.
  • s_waitcnt vscnt(0) must happen after any preceding global/generic store/store atomic/ atomicrmw-no-return-value.
  • Must happen before the following store.
  • Ensures that all global memory operations have completed before performing the store that is being released.
  1. ds_atomic
  2. s_waitcnt lgkmcnt(0)
  • If OpenCL, omit.
  • Must happen before the following buffer_gl0_inv.
  • Ensures any following global data read is no older than the local load atomic value being acquired.
  1. buffer_gl0_inv
  • If CU wavefront execution mode, omit.
  • If OpenCL omit.
  • Ensures that following loads will not see stale data.
atomicrmw acq_rel
  • workgroup
  • generic
  1. s_waitcnt lgkmcnt(0) & vmcnt(0) & vscnt(0)
  • If CU wavefront execution mode, omit vmcnt(0) and vscnt(0).
  • If OpenCL, omit lgkmcnt(0).
  • Could be split into separate s_waitcnt vmcnt(0), s_waitcnt vscnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/load atomic/ atomicrmw-with-return-value.
  • s_waitcnt vscnt(0) must happen after any preceding global/generic store/store atomic/ atomicrmw-no-return-value.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following atomicrmw.
  • Ensures that all memory operations have completed before performing the atomicrmw that is being released.
  1. flat_atomic
  2. s_waitcnt lgkmcnt(0) & vmcnt(0) & vscnt(0)
  • If CU wavefront execution mode, omit vmcnt(0) and vscnt(0).
  • If OpenCL, omit lgkmcnt(0).
  • Must happen before the following buffer_gl0_inv.
  • Ensures any following global data read is no older than the load atomic value being acquired.
  1. buffer_gl0_inv
  • If CU wavefront execution mode, omit.
  • Ensures that following loads will not see stale data.
atomicrmw acq_rel
  • agent
  • system
  • global
  1. s_waitcnt lgkmcnt(0) & vmcnt(0) & vscnt(0)
  • If OpenCL, omit lgkmcnt(0).
  • Could be split into separate s_waitcnt vmcnt(0), s_waitcnt vscnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/load atomic/ atomicrmw-with-return-value.
  • s_waitcnt vscnt(0) must happen after any preceding global/generic store/store atomic/ atomicrmw-no-return-value.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following atomicrmw.
  • Ensures that all memory operations to global have completed before performing the atomicrmw that is being released.
  1. buffer/global_atomic
  2. s_waitcnt vm/vscnt(0)
  • Use vmcnt(0) if atomic with return and vscnt(0) if atomic with no-return.
  • Must happen before following buffer_gl*_inv.
  • Ensures the atomicrmw has completed before invalidating the caches.
  1. buffer_gl0_inv; buffer_gl1_inv
  • Must happen before any following global/generic load/load atomic/atomicrmw.
  • Ensures that following loads will not see stale global data.
atomicrmw acq_rel
  • agent
  • system
  • generic
  1. s_waitcnt lgkmcnt(0) & vmcnt(0) & vscnt(0)
  • If OpenCL, omit lgkmcnt(0).
  • Could be split into separate s_waitcnt vmcnt(0), s_waitcnt vscnt(0), and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/load atomic atomicrmw-with-return-value.
  • s_waitcnt vscnt(0) must happen after any preceding global/generic store/store atomic/ atomicrmw-no-return-value.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following atomicrmw.
  • Ensures that all memory operations have completed before performing the atomicrmw that is being released.
  1. flat_atomic
  2. s_waitcnt vm/vscnt(0) & lgkmcnt(0)
  • If OpenCL, omit lgkmcnt(0).
  • Use vmcnt(0) if atomic with return and vscnt(0) if atomic with no-return.
  • Must happen before following buffer_gl*_inv.
  • Ensures the atomicrmw has completed before invalidating the caches.
  1. buffer_gl0_inv; buffer_gl1_inv
  • Must happen before any following global/generic load/load atomic/atomicrmw.
  • Ensures that following loads will not see stale global data.
fence acq_rel
  • singlethread
  • wavefront
 none none
fence acq_rel
  • workgroup
 none
  1. s_waitcnt lgkmcnt(0) & vmcnt(0) & vscnt(0)
  • If CU wavefront execution mode, omit vmcnt(0) and vscnt(0).
  • If OpenCL and address space is not generic, omit lgkmcnt(0).
  • If OpenCL and address space is local, omit vmcnt(0) and vscnt(0).
  • However, since LLVM currently has no address space on the fence need to conservatively always generate (see comment for previous fence).
  • Could be split into separate s_waitcnt vmcnt(0), s_waitcnt vscnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/load atomic/ atomicrmw-with-return-value.
  • s_waitcnt vscnt(0) must happen after any preceding global/generic store/store atomic/ atomicrmw-no-return-value.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/ atomicrmw.
  • Must happen before any following global/generic load/load atomic/store/store atomic/atomicrmw.
  • Ensures that all memory operations have completed before performing any following global memory operations.
  • Ensures that the preceding local/generic load atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the acquire-fence-paired-atomic) has completed before following global memory operations. This satisfies the requirements of acquire.
  • Ensures that all previous memory operations have completed before a following local/generic store atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the release-fence-paired-atomic). This satisfies the requirements of release.
  • Must happen before the following buffer_gl0_inv.
  • Ensures that the acquire-fence-paired atomic has completed before invalidating the cache. Therefore any following locations read must be no older than the value read by the acquire-fence-paired-atomic.
  1. buffer_gl0_inv
  • If CU wavefront execution mode, omit.
  • Ensures that following loads will not see stale data.
fence acq_rel
  • agent
  • system
 none
  1. s_waitcnt lgkmcnt(0) & vmcnt(0) & vscnt(0)
  • If OpenCL and address space is not generic, omit lgkmcnt(0).
  • If OpenCL and address space is local, omit vmcnt(0) and vscnt(0).
  • However, since LLVM currently has no address space on the fence need to conservatively always generate (see comment for previous fence).
  • Could be split into separate s_waitcnt vmcnt(0), s_waitcnt vscnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) must happen after any preceding global/generic load/load atomic/ atomicrmw-with-return-value.
  • s_waitcnt vscnt(0) must happen after any preceding global/generic store/store atomic/ atomicrmw-no-return-value.
  • s_waitcnt lgkmcnt(0) must happen after any preceding local/generic load/store/load atomic/store atomic/atomicrmw.
  • Must happen before the following buffer_gl*_inv.
  • Ensures that the preceding global/local/generic load atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the acquire-fence-paired-atomic) has completed before invalidating the caches. This satisfies the requirements of acquire.
  • Ensures that all previous memory operations have completed before a following global/local/generic store atomic/atomicrmw with an equal or wider sync scope and memory ordering stronger than unordered (this is termed the release-fence-paired-atomic). This satisfies the requirements of release.
  1. buffer_gl0_inv; buffer_gl1_inv
  • Must happen before any following global/generic load/load atomic/store/store atomic/atomicrmw.
  • Ensures that following loads will not see stale global data. This satisfies the requirements of acquire.
Sequential Consistent Atomic
load atomic seq_cst
  • singlethread
  • wavefront
  • global
  • local
  • generic
 Same as corresponding load atomic acquire, except must generated all instructions even for OpenCL.
load atomic seq_cst
  • workgroup
  • global
  • generic
  1. s_waitcnt lgkmcnt(0) & vmcnt(0) & vscnt(0)
  • If CU wavefront execution mode, omit vmcnt(0) and vscnt(0).
  • Could be split into separate s_waitcnt vmcnt(0), s_waitcnt vscnt(0), and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt lgkmcnt(0) must happen after preceding local/generic load atomic/store atomic/atomicrmw with memory ordering of seq_cst and with equal or wider sync scope. (Note that seq_cst fences have their own s_waitcnt lgkmcnt(0) and so do not need to be considered.)
  • s_waitcnt vmcnt(0) must happen after preceding global/generic load atomic/ atomicrmw-with-return-value with memory ordering of seq_cst and with equal or wider sync scope. (Note that seq_cst fences have their own s_waitcnt vmcnt(0) and so do not need to be considered.)
  • s_waitcnt vscnt(0) Must happen after preceding global/generic store atomic/ atomicrmw-no-return-value with memory ordering of seq_cst and with equal or wider sync scope. (Note that seq_cst fences have their own s_waitcnt vscnt(0) and so do not need to be considered.)
  • Ensures any preceding sequential consistent global/local memory instructions have completed before executing this sequentially consistent instruction. This prevents reordering a seq_cst store followed by a seq_cst load. (Note that seq_cst is stronger than acquire/release as the reordering of load acquire followed by a store release is prevented by the s_waitcnt of the release, but there is nothing preventing a store release followed by load acquire from completing out of order. The s_waitcnt could be placed after seq_store or before the seq_load. We choose the load to make the s_waitcnt be as late as possible so that the store may have already completed.)
  1. Following instructions same as corresponding load atomic acquire, except must generated all instructions even for OpenCL.
load atomic seq_cst
  • workgroup
  • local
  1. s_waitcnt vmcnt(0) & vscnt(0)
  • If CU wavefront execution mode, omit.
  • Could be split into separate s_waitcnt vmcnt(0) and s_waitcnt vscnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt vmcnt(0) Must happen after preceding global/generic load atomic/ atomicrmw-with-return-value with memory ordering of seq_cst and with equal or wider sync scope. (Note that seq_cst fences have their own s_waitcnt vmcnt(0) and so do not need to be considered.)
  • s_waitcnt vscnt(0) Must happen after preceding global/generic store atomic/ atomicrmw-no-return-value with memory ordering of seq_cst and with equal or wider sync scope. (Note that seq_cst fences have their own s_waitcnt vscnt(0) and so do not need to be considered.)
  • Ensures any preceding sequential consistent global memory instructions have completed before executing this sequentially consistent instruction. This prevents reordering a seq_cst store followed by a seq_cst load. (Note that seq_cst is stronger than acquire/release as the reordering of load acquire followed by a store release is prevented by the s_waitcnt of the release, but there is nothing preventing a store release followed by load acquire from completing out of order. The s_waitcnt could be placed after seq_store or before the seq_load. We choose the load to make the s_waitcnt be as late as possible so that the store may have already completed.)
  1. Following instructions same as corresponding load atomic acquire, except must generated all instructions even for OpenCL.
load atomic seq_cst
  • agent
  • system
  • global
  • generic
  1. s_waitcnt lgkmcnt(0) & vmcnt(0) & vscnt(0)
  • Could be split into separate s_waitcnt vmcnt(0), s_waitcnt vscnt(0) and s_waitcnt lgkmcnt(0) to allow them to be independently moved according to the following rules.
  • s_waitcnt lgkmcnt(0) must happen after preceding local load atomic/store atomic/atomicrmw with memory ordering of seq_cst and with equal or wider sync scope. (Note that seq_cst fences have their own s_waitcnt lgkmcnt(0) and so do not need to be considered.)
  • s_waitcnt vmcnt(0) must happen after preceding global/generic load atomic/ atomicrmw-with-return-value with memory ordering of seq_cst and with equal or wider sync scope. (Note that seq_cst fences have their own s_waitcnt vmcnt(0) and so do not need to be considered.)
  • s_waitcnt vscnt(0) Must happen after preceding global/generic store atomic/ atomicrmw-no-return-value with memory ordering of seq_cst and with equal or wider sync scope. (Note that seq_cst fences have their own s_waitcnt vscnt(0) and so do not need to be considered.)
  • Ensures any preceding sequential consistent global memory instructions have completed before executing this sequentially consistent instruction. This prevents reordering a seq_cst store followed by a seq_cst load. (Note that seq_cst is stronger than acquire/release as the reordering of load acquire followed by a store release is prevented by the s_waitcnt of the release, but there is nothing preventing a store release followed by load acquire from completing out of order. The s_waitcnt could be placed after seq_store or before the seq_load. We choose the load to make the s_waitcnt be as late as possible so that the store may have already completed.)
  1. Following instructions same as corresponding load atomic acquire, except must generated all instructions even for OpenCL.
store atomic seq_cst
  • singlethread
  • wavefront
  • workgroup
  • agent
  • system
  • global
  • local
  • generic
 Same as corresponding store atomic release, except must generated all instructions even for OpenCL.
atomicrmw seq_cst
  • singlethread
  • wavefront
  • workgroup
  • agent
  • system
  • global
  • local
  • generic
 Same as corresponding atomicrmw acq_rel, except must generated all instructions even for OpenCL.
fence seq_cst
  • singlethread
  • wavefront
  • workgroup
  • agent
  • system
 none Same as corresponding fence acq_rel, except must generated all instructions even for OpenCL.

Trap Handler ABI

For code objects generated by AMDGPU backend for HSA [HSA] compatible runtimes (such as ROCm [AMD-ROCm]), the runtime installs a trap handler that supports the s_trap instruction with the following usage:

AMDGPU Trap Handler for AMDHSA OS
Usage Code Sequence Trap Handler Inputs Description
reserved s_trap 0x00   Reserved by hardware.
debugtrap(arg) s_trap 0x01
SGPR0-1:
queue_ptr
VGPR0:
arg
 Reserved for HSA debugtrap intrinsic (not implemented).
llvm.trap s_trap 0x02
SGPR0-1:
queue_ptr
 Causes dispatch to be terminated and its associated queue put into the error state.
llvm.debugtrap s_trap 0x03  
  • If debugger not installed then behaves as a no-operation. The trap handler is entered and immediately returns to continue execution of the wavefront.
  • If the debugger is installed, causes the debug trap to be reported by the debugger and the wavefront is put in the halt state until resumed by the debugger.
reserved s_trap 0x04   Reserved.
reserved s_trap 0x05   Reserved.
reserved s_trap 0x06   Reserved.
debugger breakpoint s_trap 0x07   Reserved for debugger breakpoints.
reserved s_trap 0x08   Reserved.
reserved s_trap 0xfe   Reserved.
reserved s_trap 0xff   Reserved.

Call Convention

Note

This section is currently incomplete and has inaccuracies. It is WIP that will be updated as information is determined.

See Address Space Identifier for information on swizzled addresses. Unswizzled addresses are normal linear addresses.

Kernel Functions

This section describes the call convention ABI for the outer kernel function.

See Initial Kernel Execution State for the kernel call convention.

The following is not part of the AMDGPU kernel calling convention but describes how the AMDGPU implements function calls:

  1. Clang decides the kernarg layout to match the HSA Programmer’s Language Reference [HSA].
    • All structs are passed directly.
    • Lambda values are passed TBA.
  1. The kernel performs certain setup in its prolog, as described in Kernel Prolog.
Non-Kernel Functions

This section describes the call convention ABI for functions other than the outer kernel function.

If a kernel has function calls then scratch is always allocated and used for the call stack which grows from low address to high address using the swizzled scratch address space.

On entry to a function:

  1. SGPR0-3 contain a V# with the following properties (see Private Segment Buffer):

    • Base address pointing to the beginning of the wavefront scratch backing memory.
    • Swizzled with dword element size and stride of wavefront size elements.

  2. The FLAT_SCRATCH register pair is setup. See Flat Scratch.

  3. GFX6-8: M0 register set to the size of LDS in bytes. See M0.

  4. The EXEC register is set to the lanes active on entry to the function.

  5. MODE register: TBD

  6. VGPR0-31 and SGPR4-29 are used to pass function input arguments as described below.

  7. SGPR30-31 return address (RA). The code address that the function must return to when it completes. The value is undefined if the function is no return.

  8. SGPR32 is used for the stack pointer (SP). It is an unswizzled scratch offset relative to the beginning of the wavefront scratch backing memory.

    The unswizzled SP can be used with buffer instructions as an unswizzled SGPR offset with the scratch V# in SGPR0-3 to access the stack in a swizzled manner.

    The unswizzled SP value can be converted into the swizzled SP value by:

    swizzled SP = unswizzled SP / wavefront size

    This may be used to obtain the private address space address of stack objects and to convert this address to a flat address by adding the flat scratch aperture base address.

    The swizzled SP value is always 4 bytes aligned for the r600 architecture and 16 byte aligned for the amdgcn architecture.

    Note

    The amdgcn value is selected to avoid dynamic stack alignment for the OpenCL language which has the largest base type defined as 16 bytes.

    On entry, the swizzled SP value is the address of the first function argument passed on the stack. Other stack passed arguments are positive offsets from the entry swizzled SP value.

    The function may use positive offsets beyond the last stack passed argument for stack allocated local variables and register spill slots. If necessary, the function may align these to greater alignment than 16 bytes. After these the function may dynamically allocate space for such things as runtime sized alloca local allocations.

    If the function calls another function, it will place any stack allocated arguments after the last local allocation and adjust SGPR32 to the address after the last local allocation.

  9. All other registers are unspecified.

  10. Any necessary s_waitcnt has been performed to ensure memory is available to the function.

On exit from a function:

  1. VGPR0-31 and SGPR4-29 are used to pass function result arguments as described below. Any registers used are considered clobbered registers.

  2. The following registers are preserved and have the same value as on entry:

    • FLAT_SCRATCH

    • EXEC

    • GFX6-8: M0

    • All SGPR registers except the clobbered registers of SGPR4-31.

    • VGPR40-47 VGPR56-63 VGPR72-79 VGPR88-95 VGPR104-111 VGPR120-127 VGPR136-143 VGPR152-159 VGPR168-175 VGPR184-191 VGPR200-207 VGPR216-223 VGPR232-239 VGPR248-255

      Except the argument registers, the VGPR clobbered and the preserved registers are intermixed at regular intervals in order to get a better occupancy.

      For the AMDGPU backend, an inter-procedural register allocation (IPRA) optimization may mark some of clobbered SGPR and VGPR registers as preserved if it can be determined that the called function does not change their value.

  1. The PC is set to the RA provided on entry.
  2. MODE register: TBD.
  3. All other registers are clobbered.
  4. Any necessary s_waitcnt has been performed to ensure memory accessed by function is available to the caller.

The function input arguments are made up of the formal arguments explicitly declared by the source language function plus the implicit input arguments used by the implementation.

The source language input arguments are:

  1. Any source language implicit this or self argument comes first as a pointer type.
  2. Followed by the function formal arguments in left to right source order.

The source language result arguments are:

  1. The function result argument.

The source language input or result struct type arguments that are less than or equal to 16 bytes, are decomposed recursively into their base type fields, and each field is passed as if a separate argument. For input arguments, if the called function requires the struct to be in memory, for example because its address is taken, then the function body is responsible for allocating a stack location and copying the field arguments into it. Clang terms this direct struct.

The source language input struct type arguments that are greater than 16 bytes, are passed by reference. The caller is responsible for allocating a stack location to make a copy of the struct value and pass the address as the input argument. The called function is responsible to perform the dereference when accessing the input argument. Clang terms this by-value struct.

A source language result struct type argument that is greater than 16 bytes, is returned by reference. The caller is responsible for allocating a stack location to hold the result value and passes the address as the last input argument (before the implicit input arguments). In this case there are no result arguments. The called function is responsible to perform the dereference when storing the result value. Clang terms this structured return (sret).

TODO: correct the ``sret`` definition.

Lambda argument types are treated as struct types with an implementation defined set of fields.

For AMDGPU backend all source language arguments (including the decomposed struct type arguments) are passed in VGPRs unless marked inreg in which case they are passed in SGPRs.

The AMDGPU backend walks the function call graph from the leaves to determine which implicit input arguments are used, propagating to each caller of the function. The used implicit arguments are appended to the function arguments after the source language arguments in the following order:

  1. Work-Item ID (1 VGPR)

    The X, Y and Z work-item ID are packed into a single VGRP with the following layout. Only fields actually used by the function are set. The other bits are undefined.

    The values come from the initial kernel execution state. See VGPR Register Set Up Order.

    Work-item implicit argument layout
    Bits Size Field Name
    9:0 10 bits X Work-Item ID
    19:10 10 bits Y Work-Item ID
    29:20 10 bits Z Work-Item ID
    31:30 2 bits Unused

  2. Dispatch Ptr (2 SGPRs)

    The value comes from the initial kernel execution state. See SGPR Register Set Up Order.

  3. Queue Ptr (2 SGPRs)

    The value comes from the initial kernel execution state. See SGPR Register Set Up Order.

  4. Kernarg Segment Ptr (2 SGPRs)

    The value comes from the initial kernel execution state. See SGPR Register Set Up Order.

  5. Dispatch id (2 SGPRs)

    The value comes from the initial kernel execution state. See SGPR Register Set Up Order.

  6. Work-Group ID X (1 SGPR)

    The value comes from the initial kernel execution state. See SGPR Register Set Up Order.

  7. Work-Group ID Y (1 SGPR)

    The value comes from the initial kernel execution state. See SGPR Register Set Up Order.

  8. Work-Group ID Z (1 SGPR)

    The value comes from the initial kernel execution state. See SGPR Register Set Up Order.

  9. Implicit Argument Ptr (2 SGPRs)

    The value is computed by adding an offset to Kernarg Segment Ptr to get the global address space pointer to the first kernarg implicit argument.

The input and result arguments are assigned in order in the following manner:

Note

There are likely some errors and omissions in the following description that need correction.

  • VGPR arguments are assigned to consecutive VGPRs starting at VGPR0 up to VGPR31.

    If there are more arguments than will fit in these registers, the remaining arguments are allocated on the stack in order on naturally aligned addresses.

  • SGPR arguments are assigned to consecutive SGPRs starting at SGPR0 up to SGPR29.

    If there are more arguments than will fit in these registers, the remaining arguments are allocated on the stack in order on naturally aligned addresses.

Note that decomposed struct type arguments may have some fields passed in registers and some in memory.

The following is not part of the AMDGPU function calling convention but describes how the AMDGPU implements function calls:

  1. SGPR33 is used as a frame pointer (FP) if necessary. Like the SP it is an unswizzled scratch address. It is only needed if runtime sized alloca are used, or for the reasons defined in SIFrameLowering.

  2. Runtime stack alignment is supported. SGPR34 is used as a base pointer (BP) to access the incoming stack arguments in the function. The BP is needed only when the function requires the runtime stack alignment.

  3. Allocating SGPR arguments on the stack are not supported.

  4. No CFI is currently generated. See Call Frame Information.

    Note

    CFI will be generated that defines the CFA as the unswizzled address relative to the wave scratch base in the unswizzled private address space of the lowest address stack allocated local variable.

    DW_AT_frame_base will be defined as the swizzled address in the swizzled private address space by dividing the CFA by the wavefront size (since CFA is always at least dword aligned which matches the scratch swizzle element size).

    If no dynamic stack alignment was performed, the stack allocated arguments are accessed as negative offsets relative to DW_AT_frame_base, and the local variables and register spill slots are accessed as positive offsets relative to DW_AT_frame_base.

  5. Function argument passing is implemented by copying the input physical registers to virtual registers on entry. The register allocator can spill if necessary. These are copied back to physical registers at call sites. The net effect is that each function call can have these values in entirely distinct locations. The IPRA can help avoid shuffling argument registers.

  6. Call sites are implemented by setting up the arguments at positive offsets from SP. Then SP is incremented to account for the known frame size before the call and decremented after the call.

    Note

    The CFI will reflect the changed calculation needed to compute the CFA from SP.

  7. 4 byte spill slots are used in the stack frame. One slot is allocated for an emergency spill slot. Buffer instructions are used for stack accesses and not the flat_scratch instruction.

AMDPAL

This section provides code conventions used when the target triple OS is amdpal (see Target Triples) for passing runtime parameters from the application/runtime to each invocation of a hardware shader. These parameters include both generic, application-controlled parameters called user data as well as system-generated parameters that are a product of the draw or dispatch execution.

User Data

Each hardware stage has a set of 32-bit user data registers which can be written from a command buffer and then loaded into SGPRs when waves are launched via a subsequent dispatch or draw operation. This is the way most arguments are passed from the application/runtime to a hardware shader.

Compute User Data

Compute shader user data mappings are simpler than graphics shaders and have a fixed mapping.

Note that there are always 10 available user data entries in registers - entries beyond that limit must be fetched from memory (via the spill table pointer) by the shader.

PAL Compute Shader User Data Registers
User Register Description
0 Global Internal Table (32-bit pointer)
1 Per-Shader Internal Table (32-bit pointer)
2 - 11 Application-Controlled User Data (10 32-bit values)
12 Spill Table (32-bit pointer)
13 - 14 Thread Group Count (64-bit pointer)
15 GDS Range

Graphics User Data

Graphics pipelines support a much more flexible user data mapping:

PAL Graphics Shader User Data Registers
User Register Description
0 Global Internal Table (32-bit pointer)
Per-Shader Internal Table (32-bit pointer)
  • 1-15
 Application Controlled User Data (1-15 Contiguous 32-bit Values in Registers)
Spill Table (32-bit pointer)
Draw Index (First Stage Only)
Vertex Offset (First Stage Only)
Instance Offset (First Stage Only)

The placement of the global internal table remains fixed in the first user data SGPR register. Otherwise all parameters are optional, and can be mapped to any desired user data SGPR register, with the following restrictions:

  • Draw Index, Vertex Offset, and Instance Offset can only be used by the first active hardware stage in a graphics pipeline (i.e. where the API vertex shader runs).
  • Application-controlled user data must be mapped into a contiguous range of user data registers.
  • The application-controlled user data range supports compaction remapping, so only entries that are actually consumed by the shader must be assigned to corresponding registers. Note that in order to support an efficient runtime implementation, the remapping must pack registers in the same order as entries, with unused entries removed.

Global Internal Table

The global internal table is a table of shader resource descriptors (SRDs) that define how certain engine-wide, runtime-managed resources should be accessed from a shader. The majority of these resources have HW-defined formats, and it is up to the compiler to write/read data as required by the target hardware.

The following table illustrates the required format:

PAL Global Internal Table
Offset Description
0-3 Graphics Scratch SRD
4-7 Compute Scratch SRD
8-11 ES/GS Ring Output SRD
12-15 ES/GS Ring Input SRD
16-19 GS/VS Ring Output #0
20-23 GS/VS Ring Output #1
24-27 GS/VS Ring Output #2
28-31 GS/VS Ring Output #3
32-35 GS/VS Ring Input SRD
36-39 Tessellation Factor Buffer SRD
40-43 Off-Chip LDS Buffer SRD
44-47 Off-Chip Param Cache Buffer SRD
48-51 Sample Position Buffer SRD
52 vaRange::ShadowDescriptorTable High Bits

The pointer to the global internal table passed to the shader as user data is a 32-bit pointer. The top 32 bits should be assumed to be the same as the top 32 bits of the pipeline, so the shader may use the program counter’s top 32 bits.

Unspecified OS

This section provides code conventions used when the target triple OS is empty (see Target Triples).

Trap Handler ABI

For code objects generated by AMDGPU backend for non-amdhsa OS, the runtime does not install a trap handler. The llvm.trap and llvm.debugtrap instructions are handled as follows:

AMDGPU Trap Handler for Non-AMDHSA OS
Usage Code Sequence Description
llvm.trap s_endpgm Causes wavefront to be terminated.
llvm.debugtrap none Compiler warning given that there is no trap handler installed.

Source Languages

OpenCL

When the language is OpenCL the following differences occur:

  1. The OpenCL memory model is used (see Memory Model).
  2. The AMDGPU backend appends additional arguments to the kernel’s explicit arguments for the AMDHSA OS (see OpenCL kernel implicit arguments appended for AMDHSA OS).
  3. Additional metadata is generated (see Code Object Metadata).
OpenCL kernel implicit arguments appended for AMDHSA OS
Position Byte Size Byte Alignment Description
1 8 8 OpenCL Global Offset X
2 8 8 OpenCL Global Offset Y
3 8 8 OpenCL Global Offset Z
4 8 8 OpenCL address of printf buffer
5 8 8 OpenCL address of virtual queue used by enqueue_kernel.
6 8 8 OpenCL address of AqlWrap struct used by enqueue_kernel.
7 8 8 Pointer argument used for Multi-gird synchronization.

HCC

When the language is HCC the following differences occur:

  1. The HSA memory model is used (see Memory Model).

Assembler

AMDGPU backend has LLVM-MC based assembler which is currently in development. It supports AMDGCN GFX6-GFX10.

This section describes general syntax for instructions and operands.

Instructions

An instruction has the following syntax:

<opcode> <operand0>, <operand1>,... <modifier0> <modifier1>...

Operands are comma-separated while modifiers are space-separated.

The order of operands and modifiers is fixed. Most modifiers are optional and may be omitted.

Links to detailed instruction syntax description may be found in the following table. Note that features under development are not included in this description.

For more information about instructions, their semantics and supported combinations of operands, refer to one of instruction set architecture manuals [AMD-GCN-GFX6], [AMD-GCN-GFX7], [AMD-GCN-GFX8], [AMD-GCN-GFX9] and [AMD-GCN-GFX10].

Operands

Detailed description of operands may be found here.

Modifiers

Detailed description of modifiers may be found here.

Instruction Examples

DS
ds_add_u32 v2, v4 offset:16
ds_write_src2_b64 v2 offset0:4 offset1:8
ds_cmpst_f32 v2, v4, v6
ds_min_rtn_f64 v[8:9], v2, v[4:5]

For full list of supported instructions, refer to “LDS/GDS instructions” in ISA Manual.

FLAT
flat_load_dword v1, v[3:4]
flat_store_dwordx3 v[3:4], v[5:7]
flat_atomic_swap v1, v[3:4], v5 glc
flat_atomic_cmpswap v1, v[3:4], v[5:6] glc slc
flat_atomic_fmax_x2 v[1:2], v[3:4], v[5:6] glc

For full list of supported instructions, refer to “FLAT instructions” in ISA Manual.

MUBUF
buffer_load_dword v1, off, s[4:7], s1
buffer_store_dwordx4 v[1:4], v2, ttmp[4:7], s1 offen offset:4 glc tfe
buffer_store_format_xy v[1:2], off, s[4:7], s1
buffer_wbinvl1
buffer_atomic_inc v1, v2, s[8:11], s4 idxen offset:4 slc

For full list of supported instructions, refer to “MUBUF Instructions” in ISA Manual.

SMRD/SMEM
s_load_dword s1, s[2:3], 0xfc
s_load_dwordx8 s[8:15], s[2:3], s4
s_load_dwordx16 s[88:103], s[2:3], s4
s_dcache_inv_vol
s_memtime s[4:5]

For full list of supported instructions, refer to “Scalar Memory Operations” in ISA Manual.

SOP1
s_mov_b32 s1, s2
s_mov_b64 s[0:1], 0x80000000
s_cmov_b32 s1, 200
s_wqm_b64 s[2:3], s[4:5]
s_bcnt0_i32_b64 s1, s[2:3]
s_swappc_b64 s[2:3], s[4:5]
s_cbranch_join s[4:5]

For full list of supported instructions, refer to “SOP1 Instructions” in ISA Manual.

SOP2
s_add_u32 s1, s2, s3
s_and_b64 s[2:3], s[4:5], s[6:7]
s_cselect_b32 s1, s2, s3
s_andn2_b32 s2, s4, s6
s_lshr_b64 s[2:3], s[4:5], s6
s_ashr_i32 s2, s4, s6
s_bfm_b64 s[2:3], s4, s6
s_bfe_i64 s[2:3], s[4:5], s6
s_cbranch_g_fork s[4:5], s[6:7]

For full list of supported instructions, refer to “SOP2 Instructions” in ISA Manual.

SOPC
s_cmp_eq_i32 s1, s2
s_bitcmp1_b32 s1, s2
s_bitcmp0_b64 s[2:3], s4
s_setvskip s3, s5

For full list of supported instructions, refer to “SOPC Instructions” in ISA Manual.

SOPP
s_barrier
s_nop 2
s_endpgm
s_waitcnt 0 ; Wait for all counters to be 0
s_waitcnt vmcnt(0) & expcnt(0) & lgkmcnt(0) ; Equivalent to above
s_waitcnt vmcnt(1) ; Wait for vmcnt counter to be 1.
s_sethalt 9
s_sleep 10
s_sendmsg 0x1
s_sendmsg sendmsg(MSG_INTERRUPT)
s_trap 1

For full list of supported instructions, refer to “SOPP Instructions” in ISA Manual.

Unless otherwise mentioned, little verification is performed on the operands of SOPP Instructions, so it is up to the programmer to be familiar with the range or acceptable values.

VALU

For vector ALU instruction opcodes (VOP1, VOP2, VOP3, VOPC, VOP_DPP, VOP_SDWA), the assembler will automatically use optimal encoding based on its operands. To force specific encoding, one can add a suffix to the opcode of the instruction:

  • _e32 for 32-bit VOP1/VOP2/VOPC
  • _e64 for 64-bit VOP3
  • _dpp for VOP_DPP
  • _sdwa for VOP_SDWA

VOP1/VOP2/VOP3/VOPC examples:

v_mov_b32 v1, v2
v_mov_b32_e32 v1, v2
v_nop
v_cvt_f64_i32_e32 v[1:2], v2
v_floor_f32_e32 v1, v2
v_bfrev_b32_e32 v1, v2
v_add_f32_e32 v1, v2, v3
v_mul_i32_i24_e64 v1, v2, 3
v_mul_i32_i24_e32 v1, -3, v3
v_mul_i32_i24_e32 v1, -100, v3
v_addc_u32 v1, s[0:1], v2, v3, s[2:3]
v_max_f16_e32 v1, v2, v3

VOP_DPP examples:

v_mov_b32 v0, v0 quad_perm:[0,2,1,1]
v_sin_f32 v0, v0 row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0
v_mov_b32 v0, v0 wave_shl:1
v_mov_b32 v0, v0 row_mirror
v_mov_b32 v0, v0 row_bcast:31
v_mov_b32 v0, v0 quad_perm:[1,3,0,1] row_mask:0xa bank_mask:0x1 bound_ctrl:0
v_add_f32 v0, v0, |v0| row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0
v_max_f16 v1, v2, v3 row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0

VOP_SDWA examples:

v_mov_b32 v1, v2 dst_sel:BYTE_0 dst_unused:UNUSED_PRESERVE src0_sel:DWORD
v_min_u32 v200, v200, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:DWORD
v_sin_f32 v0, v0 dst_unused:UNUSED_PAD src0_sel:WORD_1
v_fract_f32 v0, |v0| dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
v_cmpx_le_u32 vcc, v1, v2 src0_sel:BYTE_2 src1_sel:WORD_0

For full list of supported instructions, refer to “Vector ALU instructions”.

Code Object V2 Predefined Symbols (–amdhsa-code-object-version=2)

Warning

Code Object V2 is not the default code object version emitted by this version of LLVM. For a description of the predefined symbols available with the default configuration (Code Object V3) see Code Object V3 Predefined Symbols (–amdhsa-code-object-version=3).

The AMDGPU assembler defines and updates some symbols automatically. These symbols do not affect code generation.

.option.machine_version_major

Set to the GFX major generation number of the target being assembled for. For example, when assembling for a “GFX9” target this will be set to the integer value “9”. The possible GFX major generation numbers are presented in Processors.

.option.machine_version_minor

Set to the GFX minor generation number of the target being assembled for. For example, when assembling for a “GFX810” target this will be set to the integer value “1”. The possible GFX minor generation numbers are presented in Processors.

.option.machine_version_stepping

Set to the GFX stepping generation number of the target being assembled for. For example, when assembling for a “GFX704” target this will be set to the integer value “4”. The possible GFX stepping generation numbers are presented in Processors.

.kernel.vgpr_count

Set to zero each time a .amdgpu_hsa_kernel (name) directive is encountered. At each instruction, if the current value of this symbol is less than or equal to the maximum VGPR number explicitly referenced within that instruction then the symbol value is updated to equal that VGPR number plus one.

.kernel.sgpr_count

Set to zero each time a .amdgpu_hsa_kernel (name) directive is encountered. At each instruction, if the current value of this symbol is less than or equal to the maximum VGPR number explicitly referenced within that instruction then the symbol value is updated to equal that SGPR number plus one.

Code Object V2 Directives (–amdhsa-code-object-version=2)

Warning

Code Object V2 is not the default code object version emitted by this version of LLVM. For a description of the directives supported with the default configuration (Code Object V3) see Code Object V3 Directives (–amdhsa-code-object-version=3).

AMDGPU ABI defines auxiliary data in output code object. In assembly source, one can specify them with assembler directives.

.hsa_code_object_version major, minor

major and minor are integers that specify the version of the HSA code object that will be generated by the assembler.

.hsa_code_object_isa [major, minor, stepping, vendor, arch]

major, minor, and stepping are all integers that describe the instruction set architecture (ISA) version of the assembly program.

vendor and arch are quoted strings. vendor should always be equal to “AMD” and arch should always be equal to “AMDGPU”.

By default, the assembler will derive the ISA version, vendor, and arch from the value of the -mcpu option that is passed to the assembler.

.amdgpu_hsa_kernel (name)

This directives specifies that the symbol with given name is a kernel entry point (label) and the object should contain corresponding symbol of type STT_AMDGPU_HSA_KERNEL.

.amd_kernel_code_t

This directive marks the beginning of a list of key / value pairs that are used to specify the amd_kernel_code_t object that will be emitted by the assembler. The list must be terminated by the .end_amd_kernel_code_t directive. For any amd_kernel_code_t values that are unspecified a default value will be used. The default value for all keys is 0, with the following exceptions:

  • amd_code_version_major defaults to 1.
  • amd_kernel_code_version_minor defaults to 2.
  • amd_machine_kind defaults to 1.
  • amd_machine_version_major, machine_version_minor, and amd_machine_version_stepping are derived from the value of the -mcpu option that is passed to the assembler.
  • kernel_code_entry_byte_offset defaults to 256.
  • wavefront_size defaults 6 for all targets before GFX10. For GFX10 onwards defaults to 6 if target feature wavefrontsize64 is enabled, otherwise 5. Note that wavefront size is specified as a power of two, so a value of n means a size of 2^ n.
  • call_convention defaults to -1.
  • kernarg_segment_alignment, group_segment_alignment, and private_segment_alignment default to 4. Note that alignments are specified as a power of 2, so a value of n means an alignment of 2^ n.
  • enable_wgp_mode defaults to 1 if target feature cumode is disabled for GFX10 onwards.
  • enable_mem_ordered defaults to 1 for GFX10 onwards.

The .amd_kernel_code_t directive must be placed immediately after the function label and before any instructions.

For a full list of amd_kernel_code_t keys, refer to AMDGPU ABI document, comments in lib/Target/AMDGPU/AmdKernelCodeT.h and test/CodeGen/AMDGPU/hsa.s.

Code Object V2 Example Source Code (–amdhsa-code-object-version=2)

Warning

Code Object V2 is not the default code object version emitted by this version of LLVM. For a description of the directives supported with the default configuration (Code Object V3) see Code Object V3 Example Source Code (–amdhsa-code-object-version=3).

Here is an example of a minimal assembly source file, defining one HSA kernel:

 1 .hsa_code_object_version 1,0
 2 .hsa_code_object_isa
 3 
 4 .hsatext
 5 .globl  hello_world
 6 .p2align 8
 7 .amdgpu_hsa_kernel hello_world
 8 
 9 hello_world:
10 
11    .amd_kernel_code_t
12       enable_sgpr_kernarg_segment_ptr = 1
13       is_ptr64 = 1
14       compute_pgm_rsrc1_vgprs = 0
15       compute_pgm_rsrc1_sgprs = 0
16       compute_pgm_rsrc2_user_sgpr = 2
17       compute_pgm_rsrc1_wgp_mode = 0
18       compute_pgm_rsrc1_mem_ordered = 0
19       compute_pgm_rsrc1_fwd_progress = 1
20   .end_amd_kernel_code_t
21 
22   s_load_dwordx2 s[0:1], s[0:1] 0x0
23   v_mov_b32 v0, 3.14159
24   s_waitcnt lgkmcnt(0)
25   v_mov_b32 v1, s0
26   v_mov_b32 v2, s1
27   flat_store_dword v[1:2], v0
28   s_endpgm
29 .Lfunc_end0:
30      .size   hello_world, .Lfunc_end0-hello_world

Code Object V3 Predefined Symbols (–amdhsa-code-object-version=3)

The AMDGPU assembler defines and updates some symbols automatically. These symbols do not affect code generation.

.amdgcn.gfx_generation_number

Set to the GFX major generation number of the target being assembled for. For example, when assembling for a “GFX9” target this will be set to the integer value “9”. The possible GFX major generation numbers are presented in Processors.

.amdgcn.gfx_generation_minor

Set to the GFX minor generation number of the target being assembled for. For example, when assembling for a “GFX810” target this will be set to the integer value “1”. The possible GFX minor generation numbers are presented in Processors.

.amdgcn.gfx_generation_stepping

Set to the GFX stepping generation number of the target being assembled for. For example, when assembling for a “GFX704” target this will be set to the integer value “4”. The possible GFX stepping generation numbers are presented in Processors.

.amdgcn.next_free_vgpr

Set to zero before assembly begins. At each instruction, if the current value of this symbol is less than or equal to the maximum VGPR number explicitly referenced within that instruction then the symbol value is updated to equal that VGPR number plus one.

May be used to set the .amdhsa_next_free_vgpr directive in AMDHSA Kernel Assembler Directives.

May be set at any time, e.g. manually set to zero at the start of each kernel.

.amdgcn.next_free_sgpr

Set to zero before assembly begins. At each instruction, if the current value of this symbol is less than or equal the maximum SGPR number explicitly referenced within that instruction then the symbol value is updated to equal that SGPR number plus one.

May be used to set the .amdhsa_next_free_spgr directive in AMDHSA Kernel Assembler Directives.

May be set at any time, e.g. manually set to zero at the start of each kernel.

Code Object V3 Directives (–amdhsa-code-object-version=3)

Directives which begin with .amdgcn are valid for all amdgcn architecture processors, and are not OS-specific. Directives which begin with .amdhsa are specific to amdgcn architecture processors when the amdhsa OS is specified. See Target Triples and Processors.

.amdgcn_target <target>

Optional directive which declares the target supported by the containing assembler source file. Valid values are described in Code Object Target Identification. Used by the assembler to validate command-line options such as -triple, -mcpu, and those which specify target features.

.amdhsa_kernel <name>

Creates a correctly aligned AMDHSA kernel descriptor and a symbol, <name>.kd, in the current location of the current section. Only valid when the OS is amdhsa. <name> must be a symbol that labels the first instruction to execute, and does not need to be previously defined.

Marks the beginning of a list of directives used to generate the bytes of a kernel descriptor, as described in Kernel Descriptor. Directives which may appear in this list are described in AMDHSA Kernel Assembler Directives. Directives may appear in any order, must be valid for the target being assembled for, and cannot be repeated. Directives support the range of values specified by the field they reference in Kernel Descriptor. If a directive is not specified, it is assumed to have its default value, unless it is marked as “Required”, in which case it is an error to omit the directive. This list of directives is terminated by an .end_amdhsa_kernel directive.

AMDHSA Kernel Assembler Directives
Directive Default Supported On Description
.amdhsa_group_segment_fixed_size 0 GFX6-GFX10 Controls GROUP_SEGMENT_FIXED_SIZE in Kernel Descriptor for GFX6-GFX10.
.amdhsa_private_segment_fixed_size 0 GFX6-GFX10 Controls PRIVATE_SEGMENT_FIXED_SIZE in Kernel Descriptor for GFX6-GFX10.
.amdhsa_user_sgpr_private_segment_buffer 0 GFX6-GFX10 Controls ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER in Kernel Descriptor for GFX6-GFX10.
.amdhsa_user_sgpr_dispatch_ptr 0 GFX6-GFX10 Controls ENABLE_SGPR_DISPATCH_PTR in Kernel Descriptor for GFX6-GFX10.
.amdhsa_user_sgpr_queue_ptr 0 GFX6-GFX10 Controls ENABLE_SGPR_QUEUE_PTR in Kernel Descriptor for GFX6-GFX10.
.amdhsa_user_sgpr_kernarg_segment_ptr 0 GFX6-GFX10 Controls ENABLE_SGPR_KERNARG_SEGMENT_PTR in Kernel Descriptor for GFX6-GFX10.
.amdhsa_user_sgpr_dispatch_id 0 GFX6-GFX10 Controls ENABLE_SGPR_DISPATCH_ID in Kernel Descriptor for GFX6-GFX10.
.amdhsa_user_sgpr_flat_scratch_init 0 GFX6-GFX10 Controls ENABLE_SGPR_FLAT_SCRATCH_INIT in Kernel Descriptor for GFX6-GFX10.
.amdhsa_user_sgpr_private_segment_size 0 GFX6-GFX10 Controls ENABLE_SGPR_PRIVATE_SEGMENT_SIZE in Kernel Descriptor for GFX6-GFX10.
.amdhsa_wavefront_size32 Target Feature Specific (-wavefrontsize64) GFX10 Controls ENABLE_WAVEFRONT_SIZE32 in Kernel Descriptor for GFX6-GFX10.
.amdhsa_system_sgpr_private_segment_wavefront_offset 0 GFX6-GFX10 Controls ENABLE_SGPR_PRIVATE_SEGMENT_WAVEFRONT_OFFSET in compute_pgm_rsrc2 for GFX6-GFX10.
.amdhsa_system_sgpr_workgroup_id_x 1 GFX6-GFX10 Controls ENABLE_SGPR_WORKGROUP_ID_X in compute_pgm_rsrc2 for GFX6-GFX10.
.amdhsa_system_sgpr_workgroup_id_y 0 GFX6-GFX10 Controls ENABLE_SGPR_WORKGROUP_ID_Y in compute_pgm_rsrc2 for GFX6-GFX10.
.amdhsa_system_sgpr_workgroup_id_z 0 GFX6-GFX10 Controls ENABLE_SGPR_WORKGROUP_ID_Z in compute_pgm_rsrc2 for GFX6-GFX10.
.amdhsa_system_sgpr_workgroup_info 0 GFX6-GFX10 Controls ENABLE_SGPR_WORKGROUP_INFO in compute_pgm_rsrc2 for GFX6-GFX10.
.amdhsa_system_vgpr_workitem_id 0 GFX6-GFX10 Controls ENABLE_VGPR_WORKITEM_ID in compute_pgm_rsrc2 for GFX6-GFX10. Possible values are defined in System VGPR Work-Item ID Enumeration Values.
.amdhsa_next_free_vgpr Required GFX6-GFX10 Maximum VGPR number explicitly referenced, plus one. Used to calculate GRANULATED_WORKITEM_VGPR_COUNT in compute_pgm_rsrc1 for GFX6-GFX10.
.amdhsa_next_free_sgpr Required GFX6-GFX10 Maximum SGPR number explicitly referenced, plus one. Used to calculate GRANULATED_WAVEFRONT_SGPR_COUNT in compute_pgm_rsrc1 for GFX6-GFX10.
.amdhsa_reserve_vcc 1 GFX6-GFX10 Whether the kernel may use the special VCC SGPR. Used to calculate GRANULATED_WAVEFRONT_SGPR_COUNT in compute_pgm_rsrc1 for GFX6-GFX10.
.amdhsa_reserve_flat_scratch 1 GFX7-GFX10 Whether the kernel may use flat instructions to access scratch memory. Used to calculate GRANULATED_WAVEFRONT_SGPR_COUNT in compute_pgm_rsrc1 for GFX6-GFX10.
.amdhsa_reserve_xnack_mask Target Feature Specific (+xnack) GFX8-GFX10 Whether the kernel may trigger XNACK replay. Used to calculate GRANULATED_WAVEFRONT_SGPR_COUNT in compute_pgm_rsrc1 for GFX6-GFX10.
.amdhsa_float_round_mode_32 0 GFX6-GFX10 Controls FLOAT_ROUND_MODE_32 in compute_pgm_rsrc1 for GFX6-GFX10. Possible values are defined in Floating Point Rounding Mode Enumeration Values.
.amdhsa_float_round_mode_16_64 0 GFX6-GFX10 Controls FLOAT_ROUND_MODE_16_64 in compute_pgm_rsrc1 for GFX6-GFX10. Possible values are defined in Floating Point Rounding Mode Enumeration Values.
.amdhsa_float_denorm_mode_32 0 GFX6-GFX10 Controls FLOAT_DENORM_MODE_32 in compute_pgm_rsrc1 for GFX6-GFX10. Possible values are defined in Floating Point Denorm Mode Enumeration Values.
.amdhsa_float_denorm_mode_16_64 3 GFX6-GFX10 Controls FLOAT_DENORM_MODE_16_64 in compute_pgm_rsrc1 for GFX6-GFX10. Possible values are defined in Floating Point Denorm Mode Enumeration Values.
.amdhsa_dx10_clamp 1 GFX6-GFX10 Controls ENABLE_DX10_CLAMP in compute_pgm_rsrc1 for GFX6-GFX10.
.amdhsa_ieee_mode 1 GFX6-GFX10 Controls ENABLE_IEEE_MODE in compute_pgm_rsrc1 for GFX6-GFX10.
.amdhsa_fp16_overflow 0 GFX9-GFX10 Controls FP16_OVFL in compute_pgm_rsrc1 for GFX6-GFX10.
.amdhsa_workgroup_processor_mode Target Feature Specific (-cumode) GFX10 Controls ENABLE_WGP_MODE in Kernel Descriptor for GFX6-GFX10.
.amdhsa_memory_ordered 1 GFX10 Controls MEM_ORDERED in compute_pgm_rsrc1 for GFX6-GFX10.
.amdhsa_forward_progress 0 GFX10 Controls FWD_PROGRESS in compute_pgm_rsrc1 for GFX6-GFX10.
.amdhsa_exception_fp_ieee_invalid_op 0 GFX6-GFX10 Controls ENABLE_EXCEPTION_IEEE_754_FP_INVALID_OPERATION in compute_pgm_rsrc2 for GFX6-GFX10.
.amdhsa_exception_fp_denorm_src 0 GFX6-GFX10 Controls ENABLE_EXCEPTION_FP_DENORMAL_SOURCE in compute_pgm_rsrc2 for GFX6-GFX10.
.amdhsa_exception_fp_ieee_div_zero 0 GFX6-GFX10 Controls ENABLE_EXCEPTION_IEEE_754_FP_DIVISION_BY_ZERO in compute_pgm_rsrc2 for GFX6-GFX10.
.amdhsa_exception_fp_ieee_overflow 0 GFX6-GFX10 Controls ENABLE_EXCEPTION_IEEE_754_FP_OVERFLOW in compute_pgm_rsrc2 for GFX6-GFX10.
.amdhsa_exception_fp_ieee_underflow 0 GFX6-GFX10 Controls ENABLE_EXCEPTION_IEEE_754_FP_UNDERFLOW in compute_pgm_rsrc2 for GFX6-GFX10.
.amdhsa_exception_fp_ieee_inexact 0 GFX6-GFX10 Controls ENABLE_EXCEPTION_IEEE_754_FP_INEXACT in compute_pgm_rsrc2 for GFX6-GFX10.
.amdhsa_exception_int_div_zero 0 GFX6-GFX10 Controls ENABLE_EXCEPTION_INT_DIVIDE_BY_ZERO in compute_pgm_rsrc2 for GFX6-GFX10.
.amdgpu_metadata

Optional directive which declares the contents of the NT_AMDGPU_METADATA note record (see AMDGPU Code Object V3 ELF Note Records).

The contents must be in the [YAML] markup format, with the same structure and semantics described in Code Object V3 Metadata (–amdhsa-code-object-version=3).

This directive is terminated by an .end_amdgpu_metadata directive.

Code Object V3 Example Source Code (–amdhsa-code-object-version=3)

Here is an example of a minimal assembly source file, defining one HSA kernel:

 1 .amdgcn_target "amdgcn-amd-amdhsa--gfx900+xnack" // optional
 2 
 3 .text
 4 .globl hello_world
 5 .p2align 8
 6 .type hello_world,@function
 7 hello_world:
 8   s_load_dwordx2 s[0:1], s[0:1] 0x0
 9   v_mov_b32 v0, 3.14159
10   s_waitcnt lgkmcnt(0)
11   v_mov_b32 v1, s0
12   v_mov_b32 v2, s1
13   flat_store_dword v[1:2], v0
14   s_endpgm
15 .Lfunc_end0:
16   .size   hello_world, .Lfunc_end0-hello_world
17 
18 .rodata
19 .p2align 6
20 .amdhsa_kernel hello_world
21   .amdhsa_user_sgpr_kernarg_segment_ptr 1
22   .amdhsa_next_free_vgpr .amdgcn.next_free_vgpr
23   .amdhsa_next_free_sgpr .amdgcn.next_free_sgpr
24 .end_amdhsa_kernel
25 
26 .amdgpu_metadata
27 ---
28 amdhsa.version:
29   - 1
30   - 0
31 amdhsa.kernels:
32   - .name: hello_world
33     .symbol: hello_world.kd
34     .kernarg_segment_size: 48
35     .group_segment_fixed_size: 0
36     .private_segment_fixed_size: 0
37     .kernarg_segment_align: 4
38     .wavefront_size: 64
39     .sgpr_count: 2
40     .vgpr_count: 3
41     .max_flat_workgroup_size: 256
42 ...
43 .end_amdgpu_metadata

If an assembly source file contains multiple kernels and/or functions, the .amdgcn.next_free_vgpr and .amdgcn.next_free_sgpr symbols may be reset using the .set <symbol>, <expression> directive. For example, in the case of two kernels, where function1 is only called from kernel1 it is sufficient to group the function with the kernel that calls it and reset the symbols between the two connected components:

 1 .amdgcn_target "amdgcn-amd-amdhsa--gfx900+xnack" // optional
 2 
 3 // gpr tracking symbols are implicitly set to zero
 4 
 5 .text
 6 .globl kern0
 7 .p2align 8
 8 .type kern0,@function
 9 kern0:
10   // ...
11   s_endpgm
12 .Lkern0_end:
13   .size   kern0, .Lkern0_end-kern0
14 
15 .rodata
16 .p2align 6
17 .amdhsa_kernel kern0
18   // ...
19   .amdhsa_next_free_vgpr .amdgcn.next_free_vgpr
20   .amdhsa_next_free_sgpr .amdgcn.next_free_sgpr
21 .end_amdhsa_kernel
22 
23 // reset symbols to begin tracking usage in func1 and kern1
24 .set .amdgcn.next_free_vgpr, 0
25 .set .amdgcn.next_free_sgpr, 0
26 
27 .text
28 .hidden func1
29 .global func1
30 .p2align 2
31 .type func1,@function
32 func1:
33   // ...
34   s_setpc_b64 s[30:31]
35 .Lfunc1_end:
36 .size func1, .Lfunc1_end-func1
37 
38 .globl kern1
39 .p2align 8
40 .type kern1,@function
41 kern1:
42   // ...
43   s_getpc_b64 s[4:5]
44   s_add_u32 s4, s4, func1@rel32@lo+4
45   s_addc_u32 s5, s5, func1@rel32@lo+4
46   s_swappc_b64 s[30:31], s[4:5]
47   // ...
48   s_endpgm
49 .Lkern1_end:
50   .size   kern1, .Lkern1_end-kern1
51 
52 .rodata
53 .p2align 6
54 .amdhsa_kernel kern1
55   // ...
56   .amdhsa_next_free_vgpr .amdgcn.next_free_vgpr
57   .amdhsa_next_free_sgpr .amdgcn.next_free_sgpr
58 .end_amdhsa_kernel

These symbols cannot identify connected components in order to automatically track the usage for each kernel. However, in some cases careful organization of the kernels and functions in the source file means there is minimal additional effort required to accurately calculate GPR usage.