Code Object V2 Note Records (-mattr=-code-object-v3)¶

The AMDGPU backend code object uses the following ELF note record in the .note section.

Additional note records can be present.

AMDGPU Code Object V2 ELF Note Records¶

Name	Type	Description
“AMD”	NT_AMD_AMDGPU_HSA_METADATA	<metadata null terminated string>

AMDGPU Code Object V2 ELF Note Record Enumeration Values¶

Name	Value
reserved	0-9
NT_AMD_AMDGPU_HSA_METADATA	10
reserved	11

NT_AMD_AMDGPU_HSA_METADATA

Specifies extensible metadata associated with the code objects executed on HSA [HSA] compatible runtimes such as AMD’s ROCm [AMD-ROCm]. It is required when the target triple OS is amdhsa (see Target Triples). See Code Object V2 Metadata (-mattr=-code-object-v3) for the syntax of the code object metadata string.

Code Object V3 Note Records (-mattr=+code-object-v3)¶

The AMDGPU backend code object uses the following ELF note record in the .note section.

Additional note records can be present.

AMDGPU Code Object V3 ELF Note Records¶

Name	Type	Description
“AMDGPU”	NT_AMDGPU_METADATA	Metadata in Message Pack [MsgPack] binary format.

AMDGPU Code Object V3 ELF Note Record Enumeration Values¶

Name	Value
reserved	0-31
NT_AMDGPU_METADATA	32

NT_AMDGPU_METADATA

Specifies extensible metadata associated with an AMDGPU code object. It is encoded as a map in the Message Pack [MsgPack] binary data format. See Code Object V3 Metadata (-mattr=+code-object-v3) for the map keys defined for the amdhsa OS.

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¶

Source text for online-compiled programs (e.g. those compiled by the OpenCL runtime) may be embedded into the DWARF v5 line table using the clang -gembed-source option, described in table AMDGPU Debug Options.

For example:

-gembed-source

Enable the embedded source DWARF v5 extension.

-gno-embed-source

Disable the embedded source DWARF v5 extension.

AMDGPU Debug Options¶

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

This option enables one extended content types in the DWARF v5 Line Number Program Header, which is used to encode embedded source.

AMDGPU DWARF Line Number Program Header Extended Content Types¶

Content Type	Form
DW_LNCT_LLVM_source	DW_FORM_line_strp

The source field will contain the UTF-8 encoded, null-terminated source text with '\n' line endings. When the source field is present, consumers can use the embedded source instead of attempting to discover the source on disk. When the source field is absent, consumers can access the file to get the source text.

The above content type appears in the file_name_entry_format field of the line table prologue, and its corresponding value appear in the file_names field. The current encoding of the content type is documented in table AMDGPU DWARF Line Number Program Header Extended Content Types Encoding

AMDGPU DWARF Line Number Program Header Extended Content Types Encoding¶

Content Type	Value
DW_LNCT_LLVM_source	0x2001

Code Object Target Identification¶

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

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

Where:

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

For example:

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

Code Object Metadata¶

The code object metadata specifies extensible metadata associated with the code objects executed on HSA [HSA] compatible runtimes such as AMD’s ROCm [AMD-ROCm]. It is specified in a note record (see Note Records) and is required when the target triple OS is amdhsa (see Target Triples). It must contain the minimum information necessary to support the ROCM kernel queries. For example, the segment sizes needed in a dispatch packet. In addition, a high level language runtime may require other information to be included. For example, the AMD OpenCL runtime records kernel argument information.

Kernel Dispatch¶

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

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

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

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

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

Memory Spaces¶

The memory space properties are:

AMDHSA Memory Spaces¶

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

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

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

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

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

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

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

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.

Initial Kernel Execution State¶

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

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

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

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

SGPR Register Set Up Order¶

SGPR Order	Name (kernel descriptor enable field)	Number of SGPRs	Description
First	Private Segment Buffer (enable_sgpr_private _segment_buffer)	4	V# that can be used, together with Scratch Wavefront Offset as an offset, to access the private memory space using a segment address. CP uses the value provided by the runtime.
then	Dispatch Ptr (enable_sgpr_dispatch_ptr)	2	64 bit address of AQL dispatch packet for kernel dispatch actually executing.
then	Queue Ptr (enable_sgpr_queue_ptr)	2	64 bit address of amd_queue_t object for AQL queue on which the dispatch packet was queued.
then	Kernarg Segment Ptr (enable_sgpr_kernarg _segment_ptr)	2	64 bit address of Kernarg segment. This is directly copied from the kernarg_address in the kernel dispatch packet. Having CP load it once avoids loading it at the beginning of every wavefront.
then	Dispatch Id (enable_sgpr_dispatch_id)	2	64 bit Dispatch ID of the dispatch packet being executed.
then	Flat Scratch Init (enable_sgpr_flat_scratch _init)	2	This is 2 SGPRs: GFX6 Not supported. GFX7-GFX8 The first SGPR is a 32 bit byte offset from SH_HIDDEN_PRIVATE_BASE_VIMID to per SPI base of memory for scratch for the queue executing the kernel dispatch. CP obtains this from the runtime. (The Scratch Segment Buffer base address is SH_HIDDEN_PRIVATE_BASE_VIMID plus this offset.) The value of Scratch Wavefront Offset must be added to this offset by the kernel machine code, right shifted by 8, and moved to the FLAT_SCRATCH_HI SGPR register. FLAT_SCRATCH_HI corresponds to SGPRn-4 on GFX7, and SGPRn-6 on GFX8 (where SGPRn is the highest numbered SGPR allocated to the wavefront). FLAT_SCRATCH_HI is multiplied by 256 (as it is in units of 256 bytes) and added to SH_HIDDEN_PRIVATE_BASE_VIMID to calculate the per wavefront FLAT SCRATCH BASE in flat memory instructions that access the scratch 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 Wavefront Offset must be added by the kernel machine code and the result moved to the FLAT_SCRATCH SGPR which is SGPRn-6 and SGPRn-5. It is used as the FLAT SCRATCH BASE in flat memory instructions.
then	Private Segment Size	1	The 32 bit byte size of a (enable_sgpr_private single work-item’s scratch_segment_size) memory allocation. This is the value from the kernel dispatch packet Private Segment Byte Size rounded up by CP to a multiple of DWORD. Having CP load it once avoids loading it at the beginning of every wavefront. This is not used for GFX7-GFX8 since it is the same value as the second SGPR of Flat Scratch Init. However, it may be needed for GFX9 which changes the meaning of the Flat Scratch Init value.
then	Grid Work-Group Count X (enable_sgpr_grid _workgroup_count_X)	1	32 bit count of the number of work-groups in the X dimension for the grid being executed. Computed from the fields in the kernel dispatch packet as ((grid_size.x + workgroup_size.x - 1) / workgroup_size.x).
then	Grid Work-Group Count Y (enable_sgpr_grid _workgroup_count_Y && less than 16 previous SGPRs)	1	32 bit count of the number of work-groups in the Y dimension for the grid being executed. Computed from the fields in the kernel dispatch packet as ((grid_size.y + workgroup_size.y - 1) / workgroupSize.y). Only initialized if <16 previous SGPRs initialized.
then	Grid Work-Group Count Z (enable_sgpr_grid _workgroup_count_Z && less than 16 previous SGPRs)	1	32 bit count of the number of work-groups in the Z dimension for the grid being executed. Computed from the fields in the kernel dispatch packet as ((grid_size.z + workgroup_size.z - 1) / workgroupSize.z). Only initialized if <16 previous SGPRs initialized.
then	Work-Group Id X (enable_sgpr_workgroup_id _X)	1	32 bit work-group id in X dimension of grid for wavefront.
then	Work-Group Id Y (enable_sgpr_workgroup_id _Y)	1	32 bit work-group id in Y dimension of grid for wavefront.
then	Work-Group Id Z (enable_sgpr_workgroup_id _Z)	1	32 bit work-group id in Z dimension of grid for wavefront.
then	Work-Group Info (enable_sgpr_workgroup _info)	1	{first_wavefront, 14’b0000, ordered_append_term[10:0], threadgroup_size_in_wavefronts[5:0]}
then	Scratch Wavefront Offset (enable_sgpr_private _segment_wavefront_offset)	1	32 bit byte offset from base of scratch base of queue executing the kernel dispatch. Must be used as an offset with Private segment address when using Scratch Segment Buffer. It must be used to set up FLAT SCRATCH for flat addressing (see Flat Scratch).

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

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

VGPR Register Set Up Order¶

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

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

1. SGPRs before the Work-Group Ids are set by CP using the 16 User Data registers.
2. Work-group Id registers X, Y, Z are set by ADC which supports any combination including none.
3. Scratch Wavefront Offset is set by SPI in a per wavefront basis which is why its value cannot 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¶

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

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 wavefront that uses the same scratch area, or a function call that creates a frame at the same address, respectively. There is no need for a s_dcache_inv as all scalar writes are write-before-read in the same thread.

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 buffer/global/flat_load volatile & !nontemporal buffer/global/flat_load glc=1 nontemporal buffer/global/flat_load glc=1 slc=1
load	none	none	local	ds_load
store	none	none	global generic private constant	!nontemporal buffer/global/flat_store nontemporal buffer/global/flat_stote glc=1 slc=1
store	none	none	local	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	buffer/global/flat_load
load atomic	monotonic	singlethread wavefront workgroup	local	ds_load
load atomic	monotonic	agent system	global generic	buffer/global/flat_load glc=1
store atomic	monotonic	singlethread wavefront workgroup agent system	global generic	buffer/global/flat_store
store atomic	monotonic	singlethread wavefront workgroup	local	ds_store
atomicrmw	monotonic	singlethread wavefront workgroup agent system	global generic	buffer/global/flat_atomic
atomicrmw	monotonic	singlethread wavefront workgroup	local	ds_atomic
Acquire Atomic
load atomic	acquire	singlethread wavefront	global local generic	buffer/global/ds/flat_load
load atomic	acquire	workgroup	global	buffer/global/flat_load
load atomic	acquire	workgroup	local	ds_load 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	flat_load 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	buffer/global/flat_load glc=1 s_waitcnt vmcnt(0) Must happen before following buffer_wbinvl1_vol. Ensures the load has completed before invalidating the cache. 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	flat_load glc=1 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. 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	buffer/global/ds/flat_atomic
atomicrmw	acquire	workgroup	global	buffer/global/flat_atomic
atomicrmw	acquire	workgroup	local	ds_atomic 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	flat_atomic 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	buffer/global/flat_atomic s_waitcnt vmcnt(0) Must happen before following buffer_wbinvl1_vol. Ensures the atomicrmw has completed before invalidating the cache. 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	flat_atomic 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. 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	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	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. 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	buffer/global/ds/flat_store
store atomic	release	workgroup	global	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. buffer/global/flat_store
store atomic	release	workgroup	local	ds_store
store atomic	release	workgroup	generic	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. flat_store
store atomic	release	agent system	global generic	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. buffer/global/ds/flat_store
atomicrmw	release	singlethread wavefront	global local generic	buffer/global/ds/flat_atomic
atomicrmw	release	workgroup	global	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. buffer/global/flat_atomic
atomicrmw	release	workgroup	local	ds_atomic
atomicrmw	release	workgroup	generic	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. flat_atomic
atomicrmw	release	agent system	global generic	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. buffer/global/ds/flat_atomic
fence	release	singlethread wavefront	none	none
fence	release	workgroup	none	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	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	buffer/global/ds/flat_atomic
atomicrmw	acq_rel	workgroup	global	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. buffer/global/flat_atomic
atomicrmw	acq_rel	workgroup	local	ds_atomic 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	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. flat_atomic 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	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. buffer/global/flat_atomic s_waitcnt vmcnt(0) Must happen before following buffer_wbinvl1_vol. Ensures the atomicrmw has completed before invalidating the cache. 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	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. flat_atomic 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. 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	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	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. 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	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.) 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	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.) 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		If debugger not installed then behaves as a no-operation. The trap handler is entered and immediately returns to continue execution of the wavefront. If the debugger is installed, causes the debug trap to be reported by the debugger and the wavefront is put in the halt state until resumed by the debugger.
reserved	s_trap 0x04		Reserved.
reserved	s_trap 0x05		Reserved.
reserved	s_trap 0x06		Reserved.
debugger breakpoint	s_trap 0x07		Reserved for debugger breakpoints.
reserved	s_trap 0x08		Reserved.
reserved	s_trap 0xfe		Reserved.
reserved	s_trap 0xff		Reserved.

User Data¶

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

Compute User Data¶

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

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

PAL Compute Shader User Data Registers¶

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

Graphics User Data¶

Graphics pipelines support a much more flexible user data mapping:

PAL Graphics Shader User Data Registers¶

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

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

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

Global Internal Table¶

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

The following table illustrates the required format:

PAL Global Internal Table¶

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

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

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.

Instructions¶

An instruction has the following syntax:

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

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

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

See detailed instruction syntax description for GFX7, GFX8 and GFX9.

Note that features under development are not included in this description.

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

Operands¶

Detailed description of operands may be found here.

Modifiers¶

Detailed description of modifiers may be found here.

Instruction Examples¶

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]

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

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

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]

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]

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]

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

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

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

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

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

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 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

Predefined Symbols (-mattr=+code-object-v3)¶

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

Code Object Directives (-mattr=+code-object-v3)¶

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

Example HSA Source Code (-mattr=+code-object-v3)¶

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

.amdgcn_target "amdgcn-amd-amdhsa--gfx900+xnack" // optional

.text
.globl hello_world
.p2align 8
.type hello_world,@function
hello_world:
  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

.rodata
.p2align 6
.amdhsa_kernel hello_world
  .amdhsa_user_sgpr_kernarg_segment_ptr 1
  .amdhsa_next_free_vgpr .amdgcn.next_free_vgpr
  .amdhsa_next_free_sgpr .amdgcn.next_free_sgpr
.end_amdhsa_kernel

.amdgpu_metadata
---
amdhsa.version:
  - 1
  - 0
amdhsa.kernels:
  - .name: hello_world
    .symbol: hello_world.kd
    .kernarg_segment_size: 48
    .group_segment_fixed_size: 0
    .private_segment_fixed_size: 0
    .kernarg_segment_align: 4
    .wavefront_size: 64
    .sgpr_count: 2
    .vgpr_count: 3
    .max_flat_workgroup_size: 256
...
.end_amdgpu_metadata