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 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> Defaults to opencl.
opencl OpenCL compute kernel (see OpenCL).
amdgizcl Same as opencl except a different address space mapping is used (see Address Spaces).
amdgiz Same as opencl except a different address space mapping is used (see Address Spaces).
hcc AMD HC language compute kernel (see HCC).

Processors

Use the clang -mcpu <Processor> option to specify the AMD GPU 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      
redwood   r600 dGPU      
sumo   r600 dGPU      
juniper   r600 dGPU      
cypress   r600 dGPU      
Radeon HD 6000 Series (Northern Islands) [AMD-RADEON-HD-6000]
barts   r600 dGPU      
turks   r600 dGPU      
caicos   r600 dGPU      
cayman   r600 dGPU      
GCN GFX6 (Southern Islands (SI)) [AMD-GCN-GFX6]
gfx600
  • tahiti
 amdgcn dGPU      
gfx601
  • pitcairn
  • verde
  • oland
  • hainan
 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
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
  • tonga
  • iceland
 amdgcn dGPU
  • xnack [off]
 ROCm
  • FirePro S7150
  • FirePro S7100
  • FirePro W7100
  • Radeon R285
  • Radeon R9 380
  • Radeon R9 385
  • Mobile FirePro M7170
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
gfx810
  • stoney
 amdgcn APU
  • xnack [on]
    
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]
   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 AMD GPU 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.

Address Spaces

The AMDGPU backend uses the following address space mappings.

The memory space names used in the table, aside from the region memory space, is from the OpenCL standard.

LLVM Address Space number is used throughout LLVM (for example, in LLVM IR).

Address Space Mapping
LLVM Address Space Memory Space
Current Default amdgiz/amdgizcl hcc Future Default
0 Private (Scratch) Generic (Flat) Generic (Flat) Generic (Flat)
1 Global Global Global Global
2 Constant Constant Constant Region (GDS)
3 Local (group/LDS) Local (group/LDS) Local (group/LDS) Local (group/LDS)
4 Generic (Flat) Region (GDS) Region (GDS) Constant
5 Region (GDS) Private (Scratch) Private (Scratch) Private (Scratch)
Current Default
This is the current default address space mapping used for all languages except hcc. This will shortly be deprecated.
amdgiz/amdgizcl
This is the current address space mapping used when amdgiz or amdgizcl is specified as the target triple environment value.
hcc
This is the current address space mapping used when hcc is specified as the target triple environment value.This will shortly be deprecated.
Future Default
This will shortly be the only address space mapping for all languages using AMDGPU 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 synchonizes-with relation independent of scope, and synchonizes-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 workgroup.
  • 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 workgroup.
  • 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 workgroup.
  • 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).

AMDGPU Intrinsics

The AMDGPU backend implements the following intrinsics.

This section is WIP.

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 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

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.

The AMDGPU backend code object uses the following ELF note records in the .note section. The Description column specifies the layout of the note record’s desc field. All fields are consecutive bytes. Note records with variable size strings have a corresponding *_size field that specifies the number of bytes, including the terminating null character, in the string. The string(s) come immediately after the preceding fields.

Additional note records can be present.

AMDGPU ELF Note Records
Name Type Description
“AMD” NT_AMD_AMDGPU_HSA_METADATA <metadata null terminated string>
AMDGPU 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 Metadata for the syntax of the code object metadata string.

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
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.

All global symbols, whether defined in the compilation unit or external, are accessed by the machine code indirectly through a GOT table entry. This allows them to be preemptable. The GOT table is only supported when the target triple OS is amdhsa (see Target Triples).

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

DWARF

Standard DWARF [DWARF] Version 2 sections can be generated. These 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.

Address Space Mapping

The following address space mapping is used:

AMDGPU DWARF Address Space Mapping
DWARF Address Space Memory Space
1 Private (Scratch)
2 Local (group/LDS)
omitted Global
omitted Constant
omitted Generic (Flat)
not supported Region (GDS)

See Address Spaces for information on the memory space terminology used in the table.

An address_class attribute is generated on pointer type DIEs to specify the DWARF address space of the value of the pointer when it is in the private or local address space. Otherwise the attribute is omitted.

An XDEREF operation is generated in location list expressions for variables that are allocated in the private and local address space. Otherwise no XDREF is omitted.

Register Mapping

This section is WIP.

Source Text

This section is WIP.

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 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]. It is specified by the NT_AMD_AMDGPU_HSA_METADATA 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.

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 Metadata Mapping 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 Metadata Mapping
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 Kernel Metadata Mapping for the definition of the mapping.
AMDHSA Code Object Kernel Metadata Mapping
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 Kernel Attribute Metadata Mapping for the mapping definition.
“Args” sequence of mapping   Sequence of mappings of the kernel arguments. See AMDHSA Code Object Kernel Argument Metadata Mapping for the definition of the mapping.
“CodeProps” mapping   Mapping of properties related to the kernel code. See AMDHSA Code Object Kernel Code Properties Metadata Mapping for the mapping definition.
AMDHSA Code Object Kernel Attribute Metadata Mapping
String Key Value Type Required? Description
“ReqdWorkGroupSize” sequence of 3 integers  

The dispatch work-group size X, Y, Z must correspond to the specified values.

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 Kernel Argument Metadata Mapping
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.
“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.
“ValueType” string Required

Kernel argument value type. Only present if “ValueKind” is “ByValue”. For vector data types, the value is for the element type. Values include:

  • “Struct”
  • “I8”
  • “U8”
  • “I16”
  • “U16”
  • “F16”
  • “I32”
  • “U32”
  • “F32”
  • “I64”
  • “U64”
  • “F64”
“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 Kernel Code Properties Metadata Mapping
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-GFX9. 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.
“NumVGPRs” integer Required Number of vector registers used by each work-item for GFX6-GFX9
“MaxFlatWorkGroupSize” integer Required Maximum flat work-group size supported by the kernel in work-items. Must be >=1 and consistent with any non-0 values in FixedWorkGroupSize.
“FixedWorkGroupSize” sequence of 3 integers   Corresponds to the dispatch work-group size X, Y, Z. If omitted, defaults to 0, 0, 0. If an element is non-0 then the kernel must only be launched with a matching corresponding work-group size.
“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.

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 wave 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.

The generic address space uses the hardware flat address support available in GFX7-GFX9. This uses two fixed ranges of virtual addresses (the private and local appertures), 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 apperture 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 appertures 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 the appature base addresses are directly available as inline constant registers SRC_SHARED_BASE/LIMIT and SRC_PRIVATE_BASE/LIMIT. In 64 bit address mode the apperture 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-GFX9

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

Kernel Descriptor for GFX6-GFX9
Bits Size Field Name Description
31:0 4 bytes GroupSegmentFixedSize 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 PrivateSegmentFixedSize 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 KernelCodeEntryByteOffset Byte offset (possibly negative) from base address of kernel descriptor to kernel’s entry point instruction which must be 256 byte aligned.
223:192 4 bytes MaxFlatWorkGroupSize Maximum flat work-group size supported by the kernel in work-items. If an exact work-group size is required then must be omitted or 0 and ReqdWorkGroupSize* must be set to non-0.
239:224 2 bytes ReqdWorkGroupSizeX If present and non-0 then the kernel must be executed with the specified work-group size for X.
255:240 2 bytes ReqdWorkGroupSizeY If present and non-0 then the kernel must be executed with the specified work-group size for Y.
271:256 2 bytes ReqdWorkGroupSizeZ If present and non-0 then the kernel must be executed with the specified work-group size for Z.
383:272 14 bytes   Reserved, must be 0.
415:384 4 bytes ComputePgmRsrc1 Compute Shader (CS) program settings used by CP to set up COMPUTE_PGM_RSRC1 configuration register. See compute_pgm_rsrc1 for GFX6-GFX9.
447:416 4 bytes ComputePgmRsrc2 Compute Shader (CS) program settings used by CP to set up COMPUTE_PGM_RSRC2 configuration register. See compute_pgm_rsrc2 for GFX6-GFX9.
448 1 bit EnableSGPRPrivateSegmentBuffer

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 EnableSGPRDispatchPtr see above
450 1 bit EnableSGPRQueuePtr see above
451 1 bit EnableSGPRKernargSegmentPtr see above
452 1 bit EnableSGPRDispatchID see above
453 1 bit EnableSGPRFlatScratchInit see above
454 1 bit EnableSGPRPrivateSegmentSize see above
455 1 bit EnableSGPRGridWorkgroupCountX Not implemented in CP and should always be 0.
456 1 bit EnableSGPRGridWorkgroupCountY Not implemented in CP and should always be 0.
457 1 bit EnableSGPRGridWorkgroupCountZ Not implemented in CP and should always be 0.
463:458 6 bits   Reserved, must be 0.
511:464 6 bytes   Reserved, must be 0.
512 Total size 64 bytes.
compute_pgm_rsrc1 for GFX6-GFX9
Bits Size Field Name Description
5:0 6 bits GRANULATED_WORKITEM_VGPR_COUNT

Number of vector registers used by each work-item, granularity is device specific:

GFX6-GFX9
  • max_vgpr 1..256
  • roundup((max_vgpg + 1) / 4) - 1

Used by CP to set up COMPUTE_PGM_RSRC1.VGPRS.
9:6 4 bits GRANULATED_WAVEFRONT_SGPR_COUNT

Number of scalar registers used by a wavefront, granularity is device specific:

GFX6-GFX8
  • max_sgpr 1..112
  • roundup((max_sgpg + 1) / 8) - 1
GFX9
  • max_sgpr 1..112
  • roundup((max_sgpg + 1) / 16) - 1

Includes the special SGPRs for VCC, Flat Scratch (for GFX7 onwards) and XNACK (for GFX8 onwards). It does not include the 16 SGPR added if a trap handler is enabled.

Used by CP to set up COMPUTE_PGM_RSRC1.SGPRS.
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

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.
31:27 5 bits   Reserved, must be 0.
32 Total size 4 bytes
compute_pgm_rsrc2 for GFX6-GFX9
Bits Size Field Name Description
0 1 bit ENABLE_SGPR_PRIVATE_SEGMENT _WAVE_OFFSET

Enable the setup of the SGPR wave 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

Set to 1 if code contains a TRAP instruction which requires a trap handler to be enabled.

CP sets COMPUTE_PGM_RSRC2.TRAP_PRESENT if the runtime has installed a trap handler regardless of the setting of this field.
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 wave 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-GFX9:
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.
Floating Point Rounding Mode Enumeration Values
Enumeration Name Value Description
AMDGPU_FLOAT_ROUND_MODE_NEAR_EVEN 0 Round Ties To Even
AMDGPU_FLOAT_ROUND_MODE_PLUS_INFINITY 1 Round Toward +infinity
AMDGPU_FLOAT_ROUND_MODE_MINUS_INFINITY 2 Round Toward -infinity
AMDGPU_FLOAT_ROUND_MODE_ZERO 3 Round Toward 0
Floating Point Denorm Mode Enumeration Values
Enumeration Name Value Description
AMDGPU_FLOAT_DENORM_MODE_FLUSH_SRC_DST 0 Flush Source and Destination Denorms
AMDGPU_FLOAT_DENORM_MODE_FLUSH_DST 1 Flush Output Denorms
AMDGPU_FLOAT_DENORM_MODE_FLUSH_SRC 2 Flush Source Denorms
AMDGPU_FLOAT_DENORM_MODE_FLUSH_NONE 3 No Flush
System VGPR Work-Item ID Enumeration Values
Enumeration Name Value Description
AMDGPU_SYSTEM_VGPR_WORKITEM_ID_X 0 Set work-item X dimension ID.
AMDGPU_SYSTEM_VGPR_WORKITEM_ID_X_Y 1 Set work-item X and Y dimensions ID.
AMDGPU_SYSTEM_VGPR_WORKITEM_ID_X_Y_Z 2 Set work-item X, Y and Z dimensions ID.
AMDGPU_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 waves 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 wave 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 Wave 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 Wave 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 wave). 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 wave FLAT SCRATCH BASE in flat memory instructions that access the scratch apperture.

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
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 Wave 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 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_wave, 14’b0000, ordered_append_term[10:0], threadgroup_size_in_waves[5:0]}
then Scratch Wave Offset (enable_sgpr_private _segment_wave_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 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 Wave Offset is set by SPI in a per wave basis which is why its value cannot 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 SGRRs 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-GFX9), or global instructions (GFX9).

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

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
The M0 register is not used for range checking LDS accesses and so does not need to be initialized in the prolog.
Flat Scratch

If the kernel may use flat operations to access scratch memory, the prolog code must set up FLAT_SCRATCH register pair (FLAT_SCRATCH_LO/FLAT_SCRATCH_HI which are in SGPRn-4/SGPRn-3). Initialization uses Flat Scratch Init and Scratch Wave 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 Wave Offset to get the wave’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
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 Wave Offset and moved to the FLAT_SCRATCH pair for use as the flat scratch base in flat memory instructions.

Memory Model

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

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

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

The sequences specify the order of instructions that a single thread must execute. The s_waitcnt and 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 convervatively 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.

For GFX6-GFX9:

  • Each agent 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 waves 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 waves 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 waves 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 waves 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.

Private address space uses buffer_load/store using the scratch V# (GFX6-GFX8), or scratch_load/store (GFX9). 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 access 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.

AMDGPU backend only uses scalar memory operations 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 L1 cache to ensure it is coherent with the vector L1 cache. The scalar and vector L1 caches are invalidated between kernel dispatches by CP since constant address space data may change between kernel dispatch executions. See Memory Spaces.

The one execption 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 wave 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.

Scratch backing memory (which is used for the private address space) is accessed with MTYPE NC_NV (non-coherenent 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.

On dGPU the kernarg backing memory is accessed as 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 it is accessed as CC (cache coherent) and so the L2 cache will coherent with the CPU and other agents.

AMDHSA Memory Model Code Sequences GFX6-GFX9
LLVM Instr LLVM Memory Ordering LLVM Memory Sync Scope AMDGPU Address Space AMDGPU Machine Code
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_stote 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
  • generic
  1. buffer/global/flat_load
load atomic monotonic
  • singlethread
  • wavefront
  • workgroup
  • local
  1. ds_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/flat_load
load atomic acquire
  • workgroup
  • local
  1. ds_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 the load atomic value being acquired.
load atomic acquire
  • workgroup
  • generic
  1. 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 the load atomic value being acquired.
load atomic acquire
  • agent
  • system
  • global
  1. buffer/global/flat_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/flat_atomic
atomicrmw acquire
  • workgroup
  • local
  1. ds_atomic
  2. 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 atomicrmw value being acquired.
atomicrmw acquire
  • workgroup
  • generic
  1. flat_atomic
  2. 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 atomicrmw value being acquired.
atomicrmw acquire
  • agent
  • system
  • global
  1. buffer/global/flat_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
  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
  • 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 store.
  • Ensures that all memory operations to local have completed before performing the store that is being released.
  1. flat_store
store atomic 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 store.
  • Ensures that all memory operations to memory have completed before performing the store that is being released.
  1. buffer/global/ds/flat_store
atomicrmw release
  • singlethread
  • wavefront
  • global
  • local
  • generic
  1. buffer/global/ds/flat_atomic
atomicrmw release
  • 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/flat_atomic
atomicrmw release
  • workgroup
  • local
  1. ds_atomic
atomicrmw release
  • 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
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/ds/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/flat_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 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 the 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/flat_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 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.)
  • 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 waitcnt of the release, but there is nothing preventing a store release followed by load acquire from competing out of order.)
  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.
  • 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.)
  • 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 waitcnt of the release, but there is nothing preventing a store release followed by load acquire from competing out of order.)
  1. Following instructions same as corresponding load atomic acquire, except must generated all instructions even for OpenCL.
store atomic seq_cst
  • singlethread
  • wavefront
  • workgroup
  • global
  • local
  • generic
 Same as corresponding store atomic release, except must generated all instructions even for OpenCL.
store atomic seq_cst
  • agent
  • system
  • global
  • generic
 Same as corresponding store atomic release, except must generated all instructions even for OpenCL.
atomicrmw seq_cst
  • singlethread
  • wavefront
  • workgroup
  • global
  • local
  • generic
 Same as corresponding atomicrmw acq_rel, except must generated all instructions even for OpenCL.
atomicrmw seq_cst
  • agent
  • system
  • global
  • 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.

The memory order also adds the single thread optimization constrains defined in table AMDHSA Memory Model Single Thread Optimization Constraints GFX6-GFX9.

AMDHSA Memory Model Single Thread Optimization Constraints GFX6-GFX9
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.

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
SGPR0-1:
queue_ptr
 If debugger not installed handled same as llvm.trap.
debugger breakpoint s_trap 0x07   Reserved for debugger breakpoints.
debugger s_trap 0x08   Reserved for debugger.
debugger s_trap 0xfe   Reserved for debugger.
debugger s_trap 0xff   Reserved for debugger.

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 generating code for the OpenCL language the target triple environment should be opencl or amdgizcl (see Target Triples).

When the language is OpenCL the following differences occur:

  1. The OpenCL memory model is used (see Memory Model).
  2. The AMDGPU backend adds additional arguments to the kernel.
  3. Additional metadata is generated (Code Object Metadata).

HCC

When generating code for the OpenCL language the target triple environment should be hcc (see Target Triples).

When the language is OpenCL 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-GFX9.

This section describes general syntax for instructions and operands. 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] and [AMD-GCN-GFX9].

An instruction has the following syntax (register operands are normally comma-separated while extra operands are space-separated):

<opcode> <register_operand0>, … <extra_operand0> …

Operands

The following syntax for register operands is supported:

  • SGPR registers: s0, … or s[0], …
  • VGPR registers: v0, … or v[0], …
  • TTMP registers: ttmp0, … or ttmp[0], …
  • Special registers: exec (exec_lo, exec_hi), vcc (vcc_lo, vcc_hi), flat_scratch (flat_scratch_lo, flat_scratch_hi)
  • Special trap registers: tba (tba_lo, tba_hi), tma (tma_lo, tma_hi)
  • Register pairs, quads, etc: s[2:3], v[10:11], ttmp[5:6], s[4:7], v[12:15], ttmp[4:7], s[8:15], …
  • Register lists: [s0, s1], [ttmp0, ttmp1, ttmp2, ttmp3]
  • Register index expressions: v[2*2], s[1-1:2-1]
  • ‘off’ indicates that an operand is not enabled

The following extra operands are supported:

  • offset, offset0, offset1
  • idxen, offen bits
  • glc, slc, tfe bits
  • waitcnt: integer or combination of counter values
  • VOP3 modifiers:
    • abs (| |), neg (-)
  • DPP modifiers:
    • row_shl, row_shr, row_ror, row_rol
    • row_mirror, row_half_mirror, row_bcast
    • wave_shl, wave_shr, wave_ror, wave_rol, quad_perm
    • row_mask, bank_mask, bound_ctrl
  • SDWA modifiers:
    • dst_sel, src0_sel, src1_sel (BYTE_N, WORD_M, DWORD)
    • dst_unused (UNUSED_PAD, UNUSED_SEXT, UNUSED_PRESERVE)
    • abs, neg, sext

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”.

HSA Code Object Directives

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:

  • kernel_code_version_major defaults to 1.
  • machine_kind defaults to 1.
  • machine_version_major, machine_version_minor, and 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 to 6.
  • kernarg_segment_alignment, group_segment_alignment, and private_segment_alignment default to 4. Note that alignments are specified as a power of two, so a value of n means an alignment of 2^ n.

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.

Here is an example of a minimal amd_kernel_code_t specification:

.hsa_code_object_version 1,0
.hsa_code_object_isa

.hsatext
.globl  hello_world
.p2align 8
.amdgpu_hsa_kernel hello_world

hello_world:

   .amd_kernel_code_t
      enable_sgpr_kernarg_segment_ptr = 1
      is_ptr64 = 1
      compute_pgm_rsrc1_vgprs = 0
      compute_pgm_rsrc1_sgprs = 0
      compute_pgm_rsrc2_user_sgpr = 2
      kernarg_segment_byte_size = 8
      wavefront_sgpr_count = 2
      workitem_vgpr_count = 3
  .end_amd_kernel_code_t

  s_load_dwordx2 s[0:1], s[0:1] 0x0
  v_mov_b32 v0, 3.14159
  s_waitcnt lgkmcnt(0)
  v_mov_b32 v1, s0
  v_mov_b32 v2, s1
  flat_store_dword v[1:2], v0
  s_endpgm
.Lfunc_end0:
     .size   hello_world, .Lfunc_end0-hello_world