/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-hip/checkouts/ROCR-Runtime/src/inc/hsa_ext_amd.h File Reference#
#include "hsa.h"
#include "hsa_ext_image.h"
#include "hsa_ven_amd_pc_sampling.h"
Go to the source code of this file.
Data Structures | |
struct | hsa_amd_vendor_packet_header_t |
AMD vendor specific AQL packet header. More... | |
struct | hsa_amd_barrier_value_packet_t |
AMD barrier value packet. Halts packet processing and waits for (signal_value & mask) cond value to be satisfied, where signal_value is the value of the signal signal. More... | |
struct | hsa_amd_hdp_flush_t |
struct | hsa_amd_profiling_dispatch_time_t |
Structure containing profiling dispatch time information. More... | |
struct | hsa_amd_profiling_async_copy_time_t |
Structure containing profiling async copy time information. More... | |
struct | hsa_amd_memory_pool_t |
A memory pool encapsulates physical storage on an agent along with a memory access model. More... | |
struct | hsa_pitched_ptr_t |
struct | hsa_amd_memory_pool_link_info_t |
Link properties when accessing the memory pool from the specified agent. More... | |
struct | hsa_amd_image_descriptor_t |
Encodes an opaque vendor specific image format. The length of data depends on the underlying format. This structure must not be copied as its true length can not be determined. More... | |
struct | hsa_amd_pointer_info_t |
Describes a memory allocation known to ROCr. Within a ROCr major version this structure can only grow. More... | |
struct | hsa_amd_ipc_memory_t |
256-bit process independent identifier for a ROCr shared memory allocation. More... | |
struct | hsa_amd_gpu_memory_fault_info_t |
AMD GPU memory fault event data. More... | |
struct | hsa_amd_gpu_hw_exception_info_t |
AMD GPU HW Exception event data. More... | |
struct | hsa_amd_event_t |
AMD GPU event data passed to event handler. More... | |
struct | hsa_amd_svm_attribute_pair_t |
struct | hsa_amd_vmem_alloc_handle_t |
Struct containing an opaque handle to a memory allocation handle. More... | |
struct | hsa_amd_memory_access_desc_t |
Macros | |
#define | HSA_AMD_INTERFACE_VERSION_MAJOR 1 |
#define | HSA_AMD_INTERFACE_VERSION_MINOR 6 |
Typedefs | |
typedef uint32_t | hsa_signal_condition32_t |
A fixed-size type used to represent ::hsa_signal_condition_t constants. | |
typedef uint8_t | hsa_amd_packet_type8_t |
A fixed-size type used to represent hsa_amd_packet_type_t constants. | |
typedef bool(* | hsa_amd_signal_handler) (hsa_signal_value_t value, void *arg) |
Asyncronous signal handler function type. | |
typedef hsa_amd_ipc_memory_t | hsa_amd_ipc_signal_t |
256-bit process independent identifier for a ROCr IPC signal. | |
typedef hsa_status_t(* | hsa_amd_system_event_callback_t) (const hsa_amd_event_t *event, void *data) |
typedef void(* | hsa_amd_deallocation_callback_t) (void *ptr, void *user_data) |
Deallocation notifier function type. | |
Functions | |
return ((uint8_t *) value)[index] &(1<< subBit) | |
hsa_status_t HSA_API | hsa_amd_coherency_get_type (hsa_agent_t agent, hsa_amd_coherency_type_t *type) |
Get the coherency type of the fine grain region of an agent. | |
hsa_status_t HSA_API | hsa_amd_coherency_set_type (hsa_agent_t agent, hsa_amd_coherency_type_t type) |
Set the coherency type of the fine grain region of an agent. Deprecated. This is supported on KV platforms. For backward compatibility other platforms will spuriously succeed. | |
hsa_status_t HSA_API | hsa_amd_profiling_set_profiler_enabled (hsa_queue_t *queue, int enable) |
Enable or disable profiling capability of a queue. | |
hsa_status_t HSA_API | hsa_amd_profiling_async_copy_enable (bool enable) |
Enable or disable asynchronous memory copy profiling. | |
hsa_status_t HSA_API | hsa_amd_profiling_get_dispatch_time (hsa_agent_t agent, hsa_signal_t signal, hsa_amd_profiling_dispatch_time_t *time) |
Retrieve packet processing time stamps. | |
hsa_status_t HSA_API | hsa_amd_profiling_get_async_copy_time (hsa_signal_t signal, hsa_amd_profiling_async_copy_time_t *time) |
Retrieve asynchronous copy timestamps. | |
hsa_status_t HSA_API | hsa_amd_profiling_convert_tick_to_system_domain (hsa_agent_t agent, uint64_t agent_tick, uint64_t *system_tick) |
Computes the frequency ratio and offset between the agent clock and HSA system clock and converts the agent's tick to HSA system domain tick. | |
hsa_status_t HSA_API | hsa_amd_signal_create (hsa_signal_value_t initial_value, uint32_t num_consumers, const hsa_agent_t *consumers, uint64_t attributes, hsa_signal_t *signal) |
Create a signal with specific attributes. | |
hsa_status_t | hsa_amd_signal_value_pointer (hsa_signal_t signal, volatile hsa_signal_value_t **value_ptr) |
Returns a pointer to the value of a signal. | |
hsa_status_t HSA_API | hsa_amd_signal_async_handler (hsa_signal_t signal, hsa_signal_condition_t cond, hsa_signal_value_t value, hsa_amd_signal_handler handler, void *arg) |
Register asynchronous signal handler function. | |
hsa_status_t HSA_API | hsa_amd_async_function (void(*callback)(void *arg), void *arg) |
Call a function asynchronously. | |
uint32_t HSA_API | hsa_amd_signal_wait_any (uint32_t signal_count, hsa_signal_t *signals, hsa_signal_condition_t *conds, hsa_signal_value_t *values, uint64_t timeout_hint, hsa_wait_state_t wait_hint, hsa_signal_value_t *satisfying_value) |
Wait for any signal-condition pair to be satisfied. | |
hsa_status_t HSA_API | hsa_amd_image_get_info_max_dim (hsa_agent_t agent, hsa_agent_info_t attribute, void *value) |
Query image limits. | |
hsa_status_t HSA_API | hsa_amd_queue_cu_set_mask (const hsa_queue_t *queue, uint32_t num_cu_mask_count, const uint32_t *cu_mask) |
Set a queue's CU affinity mask. | |
hsa_status_t HSA_API | hsa_amd_queue_cu_get_mask (const hsa_queue_t *queue, uint32_t num_cu_mask_count, uint32_t *cu_mask) |
Retrieve a queue's CU affinity mask. | |
hsa_status_t HSA_API | hsa_amd_memory_pool_get_info (hsa_amd_memory_pool_t memory_pool, hsa_amd_memory_pool_info_t attribute, void *value) |
Get the current value of an attribute of a memory pool. | |
hsa_status_t HSA_API | hsa_amd_agent_iterate_memory_pools (hsa_agent_t agent, hsa_status_t(*callback)(hsa_amd_memory_pool_t memory_pool, void *data), void *data) |
Iterate over the memory pools associated with a given agent, and invoke an application-defined callback on every iteration. | |
hsa_status_t HSA_API | hsa_amd_memory_pool_allocate (hsa_amd_memory_pool_t memory_pool, size_t size, uint32_t flags, void **ptr) |
Allocate a block of memory (or buffer) in the specified pool. | |
hsa_status_t HSA_API | hsa_amd_memory_pool_free (void *ptr) |
Deallocate a block of memory previously allocated using hsa_amd_memory_pool_allocate. | |
hsa_status_t HSA_API | hsa_amd_memory_async_copy (void *dst, hsa_agent_t dst_agent, const void *src, hsa_agent_t src_agent, size_t size, uint32_t num_dep_signals, const hsa_signal_t *dep_signals, hsa_signal_t completion_signal) |
Asynchronously copy a block of memory from the location pointed to by src on the src_agent to the memory block pointed to by dst on the dst_agent . Because the DMA engines used may not be in the same coherency domain, the caller must ensure that buffers are system-level coherent. In general this requires the sending device to have released the buffer to system scope prior to executing the copy API and the receiving device must execute a system scope acquire fence prior to use of the destination buffer. | |
hsa_status_t HSA_API | hsa_amd_memory_async_copy_on_engine (void *dst, hsa_agent_t dst_agent, const void *src, hsa_agent_t src_agent, size_t size, uint32_t num_dep_signals, const hsa_signal_t *dep_signals, hsa_signal_t completion_signal, hsa_amd_sdma_engine_id_t engine_id, bool force_copy_on_sdma) |
Asynchronously copy a block of memory from the location pointed to by src on the src_agent to the memory block pointed to by dst on the dst_agent on engine_id. | |
hsa_status_t HSA_API | hsa_amd_memory_copy_engine_status (hsa_agent_t dst_agent, hsa_agent_t src_agent, uint32_t *engine_ids_mask) |
Reports the availability of SDMA copy engines. | |
hsa_status_t HSA_API | hsa_amd_memory_async_copy_rect (const hsa_pitched_ptr_t *dst, const hsa_dim3_t *dst_offset, const hsa_pitched_ptr_t *src, const hsa_dim3_t *src_offset, const hsa_dim3_t *range, hsa_agent_t copy_agent, hsa_amd_copy_direction_t dir, uint32_t num_dep_signals, const hsa_signal_t *dep_signals, hsa_signal_t completion_signal) |
hsa_status_t HSA_API | hsa_amd_agent_memory_pool_get_info (hsa_agent_t agent, hsa_amd_memory_pool_t memory_pool, hsa_amd_agent_memory_pool_info_t attribute, void *value) |
Get the current value of an attribute of the relationship between an agent and a memory pool. | |
hsa_status_t HSA_API | hsa_amd_agents_allow_access (uint32_t num_agents, const hsa_agent_t *agents, const uint32_t *flags, const void *ptr) |
Enable direct access to a buffer from a given set of agents. | |
hsa_status_t HSA_API | hsa_amd_memory_pool_can_migrate (hsa_amd_memory_pool_t src_memory_pool, hsa_amd_memory_pool_t dst_memory_pool, bool *result) |
Query if buffers currently located in some memory pool can be relocated to a destination memory pool. | |
hsa_status_t HSA_API | hsa_amd_memory_migrate (const void *ptr, hsa_amd_memory_pool_t memory_pool, uint32_t flags) |
Relocate a buffer to a new memory pool. | |
hsa_status_t HSA_API | hsa_amd_memory_lock (void *host_ptr, size_t size, hsa_agent_t *agents, int num_agent, void **agent_ptr) |
Pin a host pointer allocated by C/C++ or OS allocator (i.e. ordinary system DRAM) and return a new pointer accessible by the agents . If the host_ptr overlaps with previously locked memory, then the overlap area is kept locked (i.e multiple mappings are permitted). In this case, the same input host_ptr may give different locked agent_ptr and when it does, they are not necessarily coherent (i.e. accessing either agent_ptr is not equivalent). Accesses to agent_ptr are coarse grained. | |
hsa_status_t HSA_API | hsa_amd_memory_lock_to_pool (void *host_ptr, size_t size, hsa_agent_t *agents, int num_agent, hsa_amd_memory_pool_t pool, uint32_t flags, void **agent_ptr) |
Pin a host pointer allocated by C/C++ or OS allocator (i.e. ordinary system DRAM) and return a new pointer accessible by the agents . If the host_ptr overlaps with previously locked memory, then the overlap area is kept locked (i.e. multiple mappings are permitted). In this case, the same input host_ptr may give different locked agent_ptr and when it does, they are not necessarily coherent (i.e. accessing either agent_ptr is not equivalent). Acesses to the memory via agent_ptr have the same access properties as memory allocated from pool as determined by hsa_amd_memory_pool_get_info and hsa_amd_agent_memory_pool_get_info (ex. coarse/fine grain, platform atomic support, link info). Physical composition and placement of the memory (ex. page size, NUMA binding) is not changed. | |
hsa_status_t HSA_API | hsa_amd_memory_unlock (void *host_ptr) |
Unpin the host pointer previously pinned via hsa_amd_memory_lock or hsa_amd_memory_lock_to_pool. | |
hsa_status_t HSA_API | hsa_amd_memory_fill (void *ptr, uint32_t value, size_t count) |
Sets the first count of uint32_t of the block of memory pointed by ptr to the specified value . | |
hsa_status_t HSA_API | hsa_amd_interop_map_buffer (uint32_t num_agents, hsa_agent_t *agents, int interop_handle, uint32_t flags, size_t *size, void **ptr, size_t *metadata_size, const void **metadata) |
Maps an interop object into the HSA flat address space and establishes memory residency. The metadata pointer is valid during the lifetime of the map (until hsa_amd_interop_unmap_buffer is called). Multiple calls to hsa_amd_interop_map_buffer with the same interop_handle result in multiple mappings with potentially different addresses and different metadata pointers. Concurrent operations on these addresses are not coherent. Memory must be fenced to system scope to ensure consistency, between mappings and with any views of this buffer in the originating software stack. | |
hsa_status_t HSA_API | hsa_amd_interop_unmap_buffer (void *ptr) |
Removes a previously mapped interop object from HSA's flat address space. Ends lifetime for the mapping's associated metadata pointer. | |
hsa_status_t HSA_API | hsa_amd_image_create (hsa_agent_t agent, const hsa_ext_image_descriptor_t *image_descriptor, const hsa_amd_image_descriptor_t *image_layout, const void *image_data, hsa_access_permission_t access_permission, hsa_ext_image_t *image) |
Creates an image from an opaque vendor specific image format. Does not modify data at image_data. Intended initially for accessing interop images. | |
hsa_status_t HSA_API | hsa_amd_pointer_info (const void *ptr, hsa_amd_pointer_info_t *info, void *(*alloc)(size_t), uint32_t *num_agents_accessible, hsa_agent_t **accessible) |
Retrieves information about the allocation referenced by the given pointer. Optionally returns the number and list of agents which can directly access the allocation. In case this virtual address is unknown, the pointer type returned will be HSA_EXT_POINTER_TYPE_UNKNOWN and the only fields that are valid after hsa_amd_pointer_info returns are size and type. | |
hsa_status_t HSA_API | hsa_amd_pointer_info_set_userdata (const void *ptr, void *userdata) |
Associates an arbitrary pointer with an allocation known to ROCr. The pointer can be fetched by hsa_amd_pointer_info in the userData field. | |
hsa_status_t HSA_API | hsa_amd_ipc_memory_create (void *ptr, size_t len, hsa_amd_ipc_memory_t *handle) |
Prepares an allocation for interprocess sharing and creates a handle of type hsa_amd_ipc_memory_t uniquely identifying the allocation. A handle is valid while the allocation it references remains accessible in any process. In general applications should confirm that a shared memory region has been attached (via hsa_amd_ipc_memory_attach) in the remote process prior to releasing that memory in the local process. Repeated calls for the same allocation may, but are not required to, return unique handles. The allocation needs to be on memory on an agent of type HSA_DEVICE_TYPE_GPU. | |
hsa_status_t HSA_API | hsa_amd_ipc_memory_attach (const hsa_amd_ipc_memory_t *handle, size_t len, uint32_t num_agents, const hsa_agent_t *mapping_agents, void **mapped_ptr) |
Imports shared memory into the local process and makes it accessible by the given agents. If a shared memory handle is attached multiple times in a process each attach may return a different address. Each returned address is refcounted and requires a matching number of calls to hsa_amd_ipc_memory_detach to release the shared memory mapping. | |
hsa_status_t HSA_API | hsa_amd_ipc_memory_detach (void *mapped_ptr) |
Decrements the reference count for the shared memory mapping and releases access to shared memory imported with hsa_amd_ipc_memory_attach. | |
hsa_status_t HSA_API | hsa_amd_ipc_signal_create (hsa_signal_t signal, hsa_amd_ipc_signal_t *handle) |
Obtains an interprocess sharing handle for a signal. The handle is valid while the signal it references remains valid in any process. In general applications should confirm that the signal has been attached (via hsa_amd_ipc_signal_attach) in the remote process prior to destroying that signal in the local process. Repeated calls for the same signal may, but are not required to, return unique handles. | |
hsa_status_t HSA_API | hsa_amd_ipc_signal_attach (const hsa_amd_ipc_signal_t *handle, hsa_signal_t *signal) |
Imports an IPC capable signal into the local process. If an IPC signal handle is attached multiple times in a process each attach may return a different signal handle. Each returned signal handle is refcounted and requires a matching number of calls to hsa_signal_destroy to release the shared signal. | |
hsa_status_t HSA_API | hsa_amd_register_system_event_handler (hsa_amd_system_event_callback_t callback, void *data) |
Register AMD GPU event handler. | |
hsa_status_t HSA_API | hsa_amd_queue_set_priority (hsa_queue_t *queue, hsa_amd_queue_priority_t priority) |
Modifies the dispatch and wavefront scheduling prioirty for a given compute queue. The default is HSA_AMD_QUEUE_PRIORITY_NORMAL. | |
hsa_status_t HSA_API | hsa_amd_register_deallocation_callback (void *ptr, hsa_amd_deallocation_callback_t callback, void *user_data) |
Registers a deallocation notifier monitoring for release of agent accessible address ptr . If successful, callback will be invoked when ptr is removed from accessibility from all agents. | |
hsa_status_t HSA_API | hsa_amd_deregister_deallocation_callback (void *ptr, hsa_amd_deallocation_callback_t callback) |
Removes a deallocation notifier previously registered with hsa_amd_register_deallocation_callback. Arguments must be identical to those given in hsa_amd_register_deallocation_callback. | |
hsa_status_t | hsa_amd_svm_attributes_set (void *ptr, size_t size, hsa_amd_svm_attribute_pair_t *attribute_list, size_t attribute_count) |
Sets SVM memory attributes. | |
hsa_status_t | hsa_amd_svm_attributes_get (void *ptr, size_t size, hsa_amd_svm_attribute_pair_t *attribute_list, size_t attribute_count) |
Gets SVM memory attributes. | |
hsa_status_t | hsa_amd_svm_prefetch_async (void *ptr, size_t size, hsa_agent_t agent, uint32_t num_dep_signals, const hsa_signal_t *dep_signals, hsa_signal_t completion_signal) |
Asynchronously migrates memory to an agent. | |
hsa_status_t | hsa_amd_spm_acquire (hsa_agent_t preferred_agent) |
Acquire Stream Performance Monitor on an agent. | |
hsa_status_t | hsa_amd_spm_release (hsa_agent_t preferred_agent) |
Release Stream Performance Monitor on an agent. | |
hsa_status_t | hsa_amd_spm_set_dest_buffer (hsa_agent_t preferred_agent, size_t size_in_bytes, uint32_t *timeout, uint32_t *size_copied, void *dest, bool *is_data_loss) |
Set up the current destination user mode buffer for stream performance counter data. KFD will start writing SPM data into the destination buffer. KFD will continue to copy data into the current destination buffer until any of the following functions are called. | |
hsa_status_t | hsa_amd_portable_export_dmabuf (const void *ptr, size_t size, int *dmabuf, uint64_t *offset) |
Obtains an OS specific, vendor neutral, handle to a memory allocation. | |
hsa_status_t | hsa_amd_portable_close_dmabuf (int dmabuf) |
Closes an OS specific, vendor neutral, handle to a memory allocation. | |
hsa_status_t | hsa_amd_vmem_address_reserve (void **va, size_t size, uint64_t address, uint64_t flags) |
Allocate a reserved address range. | |
hsa_status_t | hsa_amd_vmem_address_reserve_align (void **va, size_t size, uint64_t address, uint64_t alignment, uint64_t flags) |
Allocate a reserved address range. | |
hsa_status_t | hsa_amd_vmem_address_free (void *va, size_t size) |
Free a reserved address range. | |
hsa_status_t | hsa_amd_vmem_handle_create (hsa_amd_memory_pool_t pool, size_t size, hsa_amd_memory_type_t type, uint64_t flags, hsa_amd_vmem_alloc_handle_t *memory_handle) |
Create a virtual memory handle. | |
hsa_status_t | hsa_amd_vmem_handle_release (hsa_amd_vmem_alloc_handle_t memory_handle) |
Release a virtual memory handle. | |
hsa_status_t | hsa_amd_vmem_map (void *va, size_t size, size_t in_offset, hsa_amd_vmem_alloc_handle_t memory_handle, uint64_t flags) |
Map a virtual memory handle. | |
hsa_status_t | hsa_amd_vmem_unmap (void *va, size_t size) |
Unmap a virtual memory handle. | |
hsa_status_t | hsa_amd_vmem_set_access (void *va, size_t size, const hsa_amd_memory_access_desc_t *desc, size_t desc_cnt) |
Make a memory mapping accessible. | |
hsa_status_t | hsa_amd_vmem_get_access (void *va, hsa_access_permission_t *perms, hsa_agent_t agent_handle) |
Get current access permissions for memory mapping. | |
hsa_status_t | hsa_amd_vmem_export_shareable_handle (int *dmabuf_fd, hsa_amd_vmem_alloc_handle_t handle, uint64_t flags) |
Get an exportable shareable handle. | |
hsa_status_t | hsa_amd_vmem_import_shareable_handle (int dmabuf_fd, hsa_amd_vmem_alloc_handle_t *handle) |
Import a shareable handle. | |
hsa_status_t | hsa_amd_vmem_retain_alloc_handle (hsa_amd_vmem_alloc_handle_t *memory_handle, void *addr) |
Returns memory handle for mapped memory. | |
hsa_status_t | hsa_amd_vmem_get_alloc_properties_from_handle (hsa_amd_vmem_alloc_handle_t memory_handle, hsa_amd_memory_pool_t *pool, hsa_amd_memory_type_t *type) |
Returns the current allocation properties of a handle. | |
hsa_status_t HSA_API | hsa_amd_agent_set_async_scratch_limit (hsa_agent_t agent, size_t threshold) |
Set the asynchronous scratch limit threshold on all the queues for this agent. Dispatches that are enqueued on HW queues on this agent that are smaller than threshold will not result in a scratch use-once method. | |
hsa_status_t | hsa_amd_queue_get_info (hsa_queue_t *queue, hsa_queue_info_attribute_t attribute, void *value) |
Variables | |
static __inline__ uint32_t | bit |
unsigned int | subBit = bit % 8 |
Macro Definition Documentation
◆ HSA_AMD_INTERFACE_VERSION_MAJOR
#define HSA_AMD_INTERFACE_VERSION_MAJOR 1 |
- 1.0 - initial version
- 1.1 - dmabuf export
- 1.2 - hsa_amd_memory_async_copy_on_engine
- 1.3 - HSA_AMD_MEMORY_POOL_GLOBAL_FLAG_EXTENDED_SCOPE_FINE_GRAINED pool
- 1.4 - Virtual Memory API
- 1.5 - hsa_amd_agent_info: HSA_AMD_AGENT_INFO_MEMORY_PROPERTIES
- 1.6 - Virtual Memory API: hsa_amd_vmem_address_reserve_align
◆ HSA_AMD_INTERFACE_VERSION_MINOR
#define HSA_AMD_INTERFACE_VERSION_MINOR 6 |
Typedef Documentation
◆ hsa_amd_deallocation_callback_t
typedef void(* hsa_amd_deallocation_callback_t) (void *ptr, void *user_data) |
Deallocation notifier function type.
◆ hsa_amd_ipc_signal_t
256-bit process independent identifier for a ROCr IPC signal.
◆ hsa_amd_signal_handler
typedef bool(* hsa_amd_signal_handler) (hsa_signal_value_t value, void *arg) |
Asyncronous signal handler function type.
Type definition of callback function to be used with hsa_amd_signal_async_handler. This callback is invoked if the associated signal and condition are met. The callback receives the value of the signal which satisfied the associated wait condition and a user provided value. If the callback returns true then the callback will be called again if the associated signal and condition are satisfied again. If the callback returns false then it will not be called again.
- Parameters
-
[in] value Contains the value of the signal observed by hsa_amd_signal_async_handler which caused the signal handler to be invoked. [in] arg Contains the user provided value given when the signal handler was registered with hsa_amd_signal_async_handler
- Return values
-
true resumes monitoring the signal with this handler (as if calling hsa_amd_signal_async_handler again with identical parameters) false stops monitoring the signal with this handler (handler will not be called again for this signal)
◆ hsa_amd_system_event_callback_t
typedef hsa_status_t(* hsa_amd_system_event_callback_t) (const hsa_amd_event_t *event, void *data) |
Enumeration Type Documentation
◆ anonymous enum
anonymous enum |
Enumeration constants added to ::hsa_status_t.
- Remarks
- Additions to hsa_status_t
◆ hsa_amd_agent_info_t
enum hsa_amd_agent_info_t |
Agent attributes.
Enumerator | |
---|---|
HSA_AMD_AGENT_INFO_CHIP_ID | Chip identifier. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_CACHELINE_SIZE | Size of a cacheline in bytes. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_COMPUTE_UNIT_COUNT | The number of compute unit available in the agent. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_MAX_CLOCK_FREQUENCY | The maximum clock frequency of the agent in MHz. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_DRIVER_NODE_ID | Internal driver node identifier. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_MAX_ADDRESS_WATCH_POINTS | Max number of watch points on memory address ranges to generate exception events when the watched addresses are accessed. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_BDFID | Agent BDF_ID, named LocationID in thunk. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_MEMORY_WIDTH | Memory Interface width, the return value type is uint32_t. This attribute is deprecated. |
HSA_AMD_AGENT_INFO_MEMORY_MAX_FREQUENCY | Max Memory Clock, the return value type is uint32_t. |
HSA_AMD_AGENT_INFO_PRODUCT_NAME | Board name of Agent - populated from MarketingName of Kfd Node The value is an Ascii string of 64 chars. |
HSA_AMD_AGENT_INFO_MAX_WAVES_PER_CU | Maximum number of waves possible in a Compute Unit. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_NUM_SIMDS_PER_CU | Number of SIMD's per compute unit CU The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_NUM_SHADER_ENGINES | Number of Shader Engines (SE) in Gpu The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_NUM_SHADER_ARRAYS_PER_SE | Number of Shader Arrays Per Shader Engines in Gpu The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_HDP_FLUSH | Address of the HDP flush registers. Use of these registers does not conform to the HSA memory model and should be treated with caution. The type of this attribute is hsa_amd_hdp_flush_t. |
HSA_AMD_AGENT_INFO_DOMAIN | PCIe domain for the agent. Pairs with HSA_AMD_AGENT_INFO_BDFID to give the full physical location of the Agent. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_COOPERATIVE_QUEUES | Queries for support of cooperative queues. See ::HSA_QUEUE_TYPE_COOPERATIVE. The type of this attribute is bool. |
HSA_AMD_AGENT_INFO_UUID | Queries UUID of an agent. The value is an Ascii string with a maximum of 21 chars including NUL. The string value consists of two parts: header and body. The header identifies device type (GPU, CPU, DSP) while body encodes UUID as a 16 digit hex string Agents that do not support UUID will return the string "GPU-XX" or "CPU-XX" or "DSP-XX" depending upon their device type ::hsa_device_type_t |
HSA_AMD_AGENT_INFO_ASIC_REVISION | Queries for the ASIC revision of an agent. The value is an integer that increments for each revision. This can be used by user-level software to change how it operates, depending on the hardware version. This allows selective workarounds for hardware errata. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_SVM_DIRECT_HOST_ACCESS | Queries whether or not the host can directly access SVM memory that is physically resident in the agent's local memory. The type of this attribute is bool. |
HSA_AMD_AGENT_INFO_COOPERATIVE_COMPUTE_UNIT_COUNT | Some processors support more CUs than can reliably be used in a cooperative dispatch. This queries the count of CUs which are fully enabled for cooperative dispatch. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_MEMORY_AVAIL | Queries the amount of memory available in bytes accross all global pools owned by the agent. The type of this attribute is uint64_t. |
HSA_AMD_AGENT_INFO_TIMESTAMP_FREQUENCY | Timestamp value increase rate, in Hz. The timestamp (clock) frequency is in the range 1-400MHz. The type of this attribute is uint64_t. |
HSA_AMD_AGENT_INFO_ASIC_FAMILY_ID | Queries for the ASIC family ID of an agent. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_UCODE_VERSION | Queries for the Packet Processor(CP Firmware) ucode version of an agent. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_SDMA_UCODE_VERSION | Queries for the SDMA engine ucode of an agent. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_NUM_SDMA_ENG | Queries the number of SDMA engines. If HSA_AMD_AGENT_INFO_NUM_SDMA_XGMI_ENG query returns non-zero, this query returns the the number of SDMA engines optimized for host to device bidirectional traffic. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_NUM_SDMA_XGMI_ENG | Queries the number of additional SDMA engines optimized for D2D xGMI copies. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_IOMMU_SUPPORT | Queries for version of IOMMU supported by agent. The type of this attribute is hsa_amd_iommu_version_t. |
HSA_AMD_AGENT_INFO_NUM_XCC | Queries for number of XCCs within the agent. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_DRIVER_UID | Queries for driver unique identifier. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_INFO_NEAREST_CPU | Returns the hsa_agent_t of the nearest CPU agent The type of this attribute is hsa_agent_t. |
HSA_AMD_AGENT_INFO_MEMORY_PROPERTIES | Bit-mask indicating memory properties of this agent. A memory property is set if the flag bit is set at that position. User may use the hsa_flag_isset64 macro to verify whether a flag is set. The type of this attribute is uint8_t[8]. |
HSA_AMD_AGENT_INFO_AQL_EXTENSIONS | Bit-mask indicating AQL Extensions supported by this agent. An AQL extension is set if the flag bit is set at that position. User may use the hsa_flag_isset64 macro to verify whether a flag is set. The type of this attribute is uint8_t[8]. |
◆ hsa_amd_agent_memory_pool_info_t
Properties of the relationship between an agent a memory pool.
Enumerator | |
---|---|
HSA_AMD_AGENT_MEMORY_POOL_INFO_ACCESS | Access to buffers located in the memory pool. The type of this attribute is hsa_amd_memory_pool_access_t. An agent can always directly access buffers currently located in a memory pool that is associated (the memory_pool is one of the values returned by hsa_amd_agent_iterate_memory_pools on the agent) with that agent. If the buffer is currently located in a memory pool that is not associated with the agent, and the value returned by this function for the given combination of agent and memory pool is not HSA_AMD_MEMORY_POOL_ACCESS_NEVER_ALLOWED, the application still needs to invoke hsa_amd_agents_allow_access in order to gain direct access to the buffer. If the given agent can directly access buffers the pool, the result is not HSA_AMD_MEMORY_POOL_ACCESS_NEVER_ALLOWED. If the memory pool is associated with the agent, or it is of fined-grained type, the result must not be HSA_AMD_MEMORY_POOL_ACCESS_NEVER_ALLOWED. If the memory pool is not associated with the agent, and does not reside in the global segment, the result must be HSA_AMD_MEMORY_POOL_ACCESS_NEVER_ALLOWED. |
HSA_AMD_AGENT_MEMORY_POOL_INFO_NUM_LINK_HOPS | Number of links to hop when accessing the memory pool from the specified agent. The value of this attribute is zero if the memory pool is associated with the agent, or if the access type is HSA_AMD_MEMORY_POOL_ACCESS_NEVER_ALLOWED. The type of this attribute is uint32_t. |
HSA_AMD_AGENT_MEMORY_POOL_INFO_LINK_INFO | Details of each link hop when accessing the memory pool starting from the specified agent. The type of this attribute is an array size of HSA_AMD_AGENT_MEMORY_POOL_INFO_NUM_LINK_HOPS with each element containing hsa_amd_memory_pool_link_info_t. |
◆ hsa_amd_agent_memory_properties_t
◆ hsa_amd_coherency_type_t
◆ hsa_amd_copy_direction_t
◆ hsa_amd_event_type_t
enum hsa_amd_event_type_t |
◆ hsa_amd_hw_exception_reset_cause_t
◆ hsa_amd_hw_exception_reset_type_t
◆ hsa_amd_iommu_version_t
◆ hsa_amd_link_info_type_t
Properties of the relationship between an agent a memory pool.
◆ hsa_amd_memory_fault_reason_t
◆ hsa_amd_memory_pool_access_t
Type of accesses to a memory pool from a given agent.
Enumerator | |
---|---|
HSA_AMD_MEMORY_POOL_ACCESS_NEVER_ALLOWED | The agent cannot directly access any buffer in the memory pool. |
HSA_AMD_MEMORY_POOL_ACCESS_ALLOWED_BY_DEFAULT | The agent can directly access a buffer located in the pool; the application does not need to invoke hsa_amd_agents_allow_access. |
HSA_AMD_MEMORY_POOL_ACCESS_DISALLOWED_BY_DEFAULT | The agent can directly access a buffer located in the pool, but only if the application has previously requested access to that buffer using hsa_amd_agents_allow_access. |
◆ hsa_amd_memory_pool_flag_t
Memory pool flag used to specify allocation directives.
◆ hsa_amd_memory_pool_global_flag_t
Enumerator | |
---|---|
HSA_AMD_MEMORY_POOL_GLOBAL_FLAG_KERNARG_INIT | The application can use allocations in the memory pool to store kernel arguments, and provide the values for the kernarg segment of a kernel dispatch. |
HSA_AMD_MEMORY_POOL_GLOBAL_FLAG_FINE_GRAINED | Updates to memory in this pool conform to HSA memory consistency model. If this flag is set, then HSA_AMD_MEMORY_POOL_GLOBAL_FLAG_COARSE_GRAINED must not be set. |
HSA_AMD_MEMORY_POOL_GLOBAL_FLAG_COARSE_GRAINED | Writes to memory in this pool can be performed by a single agent at a time. |
HSA_AMD_MEMORY_POOL_GLOBAL_FLAG_EXTENDED_SCOPE_FINE_GRAINED | Updates to memory in this memory pool have extended scope, acting as system-scope atomics for variables in memory regions of this type. Note: On non-compliant systems, device-specific actions may be required for system-scope coherence. |
◆ hsa_amd_memory_pool_info_t
Memory pool features.
Enumerator | |
---|---|
HSA_AMD_MEMORY_POOL_INFO_SEGMENT | Segment where the memory pool resides. The type of this attribute is hsa_amd_segment_t. |
HSA_AMD_MEMORY_POOL_INFO_GLOBAL_FLAGS | Flag mask. The value of this attribute is undefined if the value of HSA_AMD_MEMORY_POOL_INFO_SEGMENT is not HSA_AMD_SEGMENT_GLOBAL. The type of this attribute is uint32_t, a bit-field of hsa_amd_memory_pool_global_flag_t values. |
HSA_AMD_MEMORY_POOL_INFO_SIZE | Size of this pool, in bytes. The type of this attribute is size_t. |
HSA_AMD_MEMORY_POOL_INFO_RUNTIME_ALLOC_ALLOWED | Indicates whether memory in this pool can be allocated using hsa_amd_memory_pool_allocate. The type of this attribute is bool. The value of this flag is always false for memory pools in the group and private segments. |
HSA_AMD_MEMORY_POOL_INFO_RUNTIME_ALLOC_GRANULE | Allocation granularity of buffers allocated by hsa_amd_memory_pool_allocate in this memory pool. The size of a buffer allocated in this pool is a multiple of the value of this attribute. While this is the minimum size of allocation allowed, it is recommened to use HSA_AMD_MEMORY_POOL_INFO_RUNTIME_ALLOC_REC_GRANULE to obtain the recommended allocation granularity size for this pool. The value of this attribute is only defined if HSA_AMD_MEMORY_POOL_INFO_RUNTIME_ALLOC_ALLOWED is true for this pool. The type of this attribute is size_t. |
HSA_AMD_MEMORY_POOL_INFO_RUNTIME_ALLOC_ALIGNMENT | Alignment of buffers allocated by hsa_amd_memory_pool_allocate in this pool. The value of this attribute is only defined if HSA_AMD_MEMORY_POOL_INFO_RUNTIME_ALLOC_ALLOWED is true for this pool, and must be a power of 2. The type of this attribute is size_t. |
HSA_AMD_MEMORY_POOL_INFO_ACCESSIBLE_BY_ALL | This memory_pool can be made directly accessible by all the agents in the system (hsa_amd_agent_memory_pool_get_info does not return HSA_AMD_MEMORY_POOL_ACCESS_NEVER_ALLOWED for any agent). The type of this attribute is bool. |
HSA_AMD_MEMORY_POOL_INFO_ALLOC_MAX_SIZE | Maximum aggregate allocation size in bytes. The type of this attribute is size_t. |
HSA_AMD_MEMORY_POOL_INFO_LOCATION | Location of this memory pool. The type of this attribute is hsa_amd_memory_pool_location_t. |
HSA_AMD_MEMORY_POOL_INFO_RUNTIME_ALLOC_REC_GRANULE | Internal block size for allocations. This would also be the recommended granularity size for allocations as this prevents internal fragmentation. The value of this attribute is only defined if HSA_AMD_MEMORY_POOL_INFO_RUNTIME_ALLOC_ALLOWED is true for this pool. The size of this attribute is size_t. |
◆ hsa_amd_memory_pool_location_t
◆ hsa_amd_memory_type_t
◆ hsa_amd_pointer_type_t
◆ hsa_amd_queue_priority_t
◆ hsa_amd_region_info_t
Region attributes.
◆ hsa_amd_sdma_engine_id_t
SDMA engine IDs unique by single set bit position.
◆ hsa_amd_segment_t
enum hsa_amd_segment_t |
Memory segments associated with a memory pool.
◆ hsa_amd_signal_attribute_t
Signal attribute flags.
◆ hsa_amd_svm_attribute_t
◆ hsa_amd_svm_model_t
enum hsa_amd_svm_model_t |
◆ hsa_queue_info_attribute_t
Function Documentation
◆ hsa_amd_agent_iterate_memory_pools()
hsa_status_t HSA_API hsa_amd_agent_iterate_memory_pools | ( | hsa_agent_t | agent, |
hsa_status_t(*)(hsa_amd_memory_pool_t memory_pool, void *data) | callback, | ||
void * | data | ||
) |
Iterate over the memory pools associated with a given agent, and invoke an application-defined callback on every iteration.
An agent can directly access buffers located in some memory pool, or be enabled to access them by the application (see hsa_amd_agents_allow_access), yet that memory pool may not be returned by this function for that given agent.
A memory pool of fine-grained type must be associated only with the host.
- Parameters
-
[in] agent A valid agent. [in] callback Callback to be invoked on the same thread that called hsa_amd_agent_iterate_memory_pools, serially, once per memory pool that is associated with the agent. The HSA runtime passes two arguments to the callback: the memory pool, and the application data. If callback
returns a status other than ::HSA_STATUS_SUCCESS for a particular iteration, the traversal stops and hsa_amd_agent_iterate_memory_pools returns that status value.[in] data Application data that is passed to callback
on every iteration. May be NULL.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_AGENT The agent is invalid. ::HSA_STATUS_ERROR_INVALID_ARGUMENT callback
is NULL.
◆ hsa_amd_agent_memory_pool_get_info()
hsa_status_t HSA_API hsa_amd_agent_memory_pool_get_info | ( | hsa_agent_t | agent, |
hsa_amd_memory_pool_t | memory_pool, | ||
hsa_amd_agent_memory_pool_info_t | attribute, | ||
void * | value | ||
) |
Get the current value of an attribute of the relationship between an agent and a memory pool.
- Parameters
-
[in] agent Agent. [in] memory_pool Memory pool. [in] attribute Attribute to query. [out] value Pointer to a application-allocated buffer where to store the value of the attribute. If the buffer passed by the application is not large enough to hold the value of attribute
, the behavior is undefined.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully.
◆ hsa_amd_agent_set_async_scratch_limit()
hsa_status_t HSA_API hsa_amd_agent_set_async_scratch_limit | ( | hsa_agent_t | agent, |
size_t | threshold | ||
) |
Set the asynchronous scratch limit threshold on all the queues for this agent. Dispatches that are enqueued on HW queues on this agent that are smaller than threshold will not result in a scratch use-once method.
Increasing this threshold will only increase the internal limit and not cause immediate allocation of additional scratch memory. Decreasing this threshold will result in a release in scratch memory on queues where the current amount of allocated scratch exceeds the new limit.
This API is only supported on devices that support asynchronous scratch reclaim.
- Parameters
-
[in] agent A valid agent. [in] threshold Threshold size in bytes
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_AGENT The agent is invalid. ::HSA_STATUS_ERROR_INVALID_ARGUMENT This agent does not support asynchronous scratch reclaim
◆ hsa_amd_agents_allow_access()
hsa_status_t HSA_API hsa_amd_agents_allow_access | ( | uint32_t | num_agents, |
const hsa_agent_t * | agents, | ||
const uint32_t * | flags, | ||
const void * | ptr | ||
) |
Enable direct access to a buffer from a given set of agents.
Upon return, only the listed agents and the agent associated with the buffer's memory pool have direct access to the ptr
.
Any agent that has access to the buffer before and after the call to hsa_amd_agents_allow_access will also have access while hsa_amd_agents_allow_access is in progress.
The caller is responsible for ensuring that each agent in the list must be able to access the memory pool containing ptr
(using hsa_amd_agent_memory_pool_get_info with HSA_AMD_AGENT_MEMORY_POOL_INFO_ACCESS attribute), otherwise error code is returned.
- Parameters
-
[in] num_agents Size of agents
.[in] agents List of agents. If num_agents
is 0, this argument is ignored.[in] flags A list of bit-field that is used to specify access information in a per-agent basis. This is currently reserved and must be NULL. [in] ptr A buffer previously allocated using hsa_amd_memory_pool_allocate.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_ARGUMENT num_agents
is 0, oragents
is NULL,flags
is not NULL, or attempting to enable access to agent(s) becauseptr
is allocated from an inaccessible pool.
◆ hsa_amd_async_function()
hsa_status_t HSA_API hsa_amd_async_function | ( | void(*)(void *arg) | callback, |
void * | arg | ||
) |
Call a function asynchronously.
Provides access to the runtime's asynchronous event handling thread for general asynchronous functions. Functions queued this way are executed in the same manner as if they were a signal handler who's signal is satisfied.
- Parameters
-
[in] callback asynchronous function to be invoked [in] arg user provided value which is provided to handler when handler is invoked
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_ARGUMENT handler is invalid (NULL) ::HSA_STATUS_ERROR_OUT_OF_RESOURCES The HSA runtime is out of resources or blocking signals are not supported by the HSA driver component.
◆ hsa_amd_coherency_get_type()
hsa_status_t HSA_API hsa_amd_coherency_get_type | ( | hsa_agent_t | agent, |
hsa_amd_coherency_type_t * | type | ||
) |
Get the coherency type of the fine grain region of an agent.
- Parameters
-
[in] agent A valid agent. [out] type Pointer to a memory location where the HSA runtime will store the coherency type of the fine grain region.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_AGENT The agent is invalid. ::HSA_STATUS_ERROR_INVALID_ARGUMENT type
is NULL.
◆ hsa_amd_coherency_set_type()
hsa_status_t HSA_API hsa_amd_coherency_set_type | ( | hsa_agent_t | agent, |
hsa_amd_coherency_type_t | type | ||
) |
Set the coherency type of the fine grain region of an agent. Deprecated. This is supported on KV platforms. For backward compatibility other platforms will spuriously succeed.
- Parameters
-
[in] agent A valid agent. [in] type The coherency type to be set.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_AGENT The agent is invalid. ::HSA_STATUS_ERROR_INVALID_ARGUMENT type
is invalid.
◆ hsa_amd_deregister_deallocation_callback()
hsa_status_t HSA_API hsa_amd_deregister_deallocation_callback | ( | void * | ptr, |
hsa_amd_deallocation_callback_t | callback | ||
) |
Removes a deallocation notifier previously registered with hsa_amd_register_deallocation_callback. Arguments must be identical to those given in hsa_amd_register_deallocation_callback.
- Parameters
-
[in] ptr Agent accessible address which was monitored for deallocation. [in] callback Notifier to be removed.
- Return values
-
::HSA_STATUS_SUCCESS The notifier has been removed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_ARGUMENT The given notifier was not registered.
◆ hsa_amd_image_create()
hsa_status_t HSA_API hsa_amd_image_create | ( | hsa_agent_t | agent, |
const hsa_ext_image_descriptor_t * | image_descriptor, | ||
const hsa_amd_image_descriptor_t * | image_layout, | ||
const void * | image_data, | ||
hsa_access_permission_t | access_permission, | ||
hsa_ext_image_t * | image | ||
) |
Creates an image from an opaque vendor specific image format. Does not modify data at image_data. Intended initially for accessing interop images.
- Parameters
-
agent[in] Agent on which to create the image [in] image_descriptor[in] Vendor specific image format [in] image_data Pointer to image backing store [in] access_permission Access permissions for the image object [out] image Created image object.
- Return values
-
HSA_STATUS_SUCCESS Image created successfully HSA_STATUS_ERROR_NOT_INITIALIZED if HSA is not initialized HSA_STATUS_ERROR_OUT_OF_RESOURCES if there is a failure in allocating necessary resources HSA_STATUS_ERROR_INVALID_ARGUMENT Bad or mismatched descriptor, null image_data, or mismatched access_permission.
◆ hsa_amd_image_get_info_max_dim()
hsa_status_t HSA_API hsa_amd_image_get_info_max_dim | ( | hsa_agent_t | agent, |
hsa_agent_info_t | attribute, | ||
void * | value | ||
) |
Query image limits.
- Parameters
-
[in] agent A valid agent. [in] attribute HSA image info attribute to query. [out] value Pointer to an application-allocated buffer where to store the value of the attribute. If the buffer passed by the application is not large enough to hold the value of attribute
, the behavior is undefined.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_QUEUE value
is NULL orattribute
< HSA_EXT_AGENT_INFO_IMAGE_1D_MAX_ELEMENTS orattribute
> HSA_EXT_AGENT_INFO_IMAGE_ARRAY_MAX_LAYERS.
◆ hsa_amd_interop_map_buffer()
hsa_status_t HSA_API hsa_amd_interop_map_buffer | ( | uint32_t | num_agents, |
hsa_agent_t * | agents, | ||
int | interop_handle, | ||
uint32_t | flags, | ||
size_t * | size, | ||
void ** | ptr, | ||
size_t * | metadata_size, | ||
const void ** | metadata | ||
) |
Maps an interop object into the HSA flat address space and establishes memory residency. The metadata pointer is valid during the lifetime of the map (until hsa_amd_interop_unmap_buffer is called). Multiple calls to hsa_amd_interop_map_buffer with the same interop_handle result in multiple mappings with potentially different addresses and different metadata pointers. Concurrent operations on these addresses are not coherent. Memory must be fenced to system scope to ensure consistency, between mappings and with any views of this buffer in the originating software stack.
- Parameters
-
[in] num_agents Number of agents which require access to the memory [in] agents List of accessing agents. [in] interop_handle Handle of interop buffer (dmabuf handle in Linux) [in] flags Reserved, must be 0 [out] size Size in bytes of the mapped object [out] ptr Base address of the mapped object [out] metadata_size Size of metadata in bytes, may be NULL [out] metadata Pointer to metadata, may be NULL
- Return values
-
HSA_STATUS_SUCCESS if successfully mapped HSA_STATUS_ERROR_NOT_INITIALIZED if HSA is not initialized HSA_STATUS_ERROR_OUT_OF_RESOURCES if there is a failure in allocating necessary resources HSA_STATUS_ERROR_INVALID_ARGUMENT all other errors
◆ hsa_amd_interop_unmap_buffer()
hsa_status_t HSA_API hsa_amd_interop_unmap_buffer | ( | void * | ptr | ) |
Removes a previously mapped interop object from HSA's flat address space. Ends lifetime for the mapping's associated metadata pointer.
◆ hsa_amd_ipc_memory_attach()
hsa_status_t HSA_API hsa_amd_ipc_memory_attach | ( | const hsa_amd_ipc_memory_t * | handle, |
size_t | len, | ||
uint32_t | num_agents, | ||
const hsa_agent_t * | mapping_agents, | ||
void ** | mapped_ptr | ||
) |
Imports shared memory into the local process and makes it accessible by the given agents. If a shared memory handle is attached multiple times in a process each attach may return a different address. Each returned address is refcounted and requires a matching number of calls to hsa_amd_ipc_memory_detach to release the shared memory mapping.
- Parameters
-
[in] handle Pointer to the identifier for the shared memory. [in] len Length of the shared memory to import. Reserved. Must be the full length of the shared allocation in this version. [in] num_agents Count of agents in mapping_agents
. May be zero if all agents are to be allowed access.[in] mapping_agents List of agents to access the shared memory. Ignored if num_agents
is zero.[out] mapped_ptr Recieves a process local pointer to the shared memory.
- Return values
-
HSA_STATUS_SUCCESS if memory is successfully imported. HSA_STATUS_ERROR_NOT_INITIALIZED if HSA is not initialized HSA_STATUS_ERROR_OUT_OF_RESOURCES if there is a failure in allocating necessary resources HSA_STATUS_ERROR_INVALID_ARGUMENT handle
is not valid,len
is incorrect,mapped_ptr
is NULL, or some agent for which access was requested can not access the shared memory.
◆ hsa_amd_ipc_memory_create()
hsa_status_t HSA_API hsa_amd_ipc_memory_create | ( | void * | ptr, |
size_t | len, | ||
hsa_amd_ipc_memory_t * | handle | ||
) |
Prepares an allocation for interprocess sharing and creates a handle of type hsa_amd_ipc_memory_t uniquely identifying the allocation. A handle is valid while the allocation it references remains accessible in any process. In general applications should confirm that a shared memory region has been attached (via hsa_amd_ipc_memory_attach) in the remote process prior to releasing that memory in the local process. Repeated calls for the same allocation may, but are not required to, return unique handles. The allocation needs to be on memory on an agent of type HSA_DEVICE_TYPE_GPU.
- Parameters
-
[in] ptr Pointer to device memory allocated via ROCr APIs to prepare for sharing. [in] len Length in bytes of the allocation to share. [out] handle Process independent identifier referencing the shared allocation.
- Return values
-
HSA_STATUS_SUCCESS allocation is prepared for interprocess sharing. HSA_STATUS_ERROR_NOT_INITIALIZED if HSA is not initialized HSA_STATUS_ERROR_OUT_OF_RESOURCES if there is a failure in allocating necessary resources HSA_STATUS_ERROR_INVALID_ARGUMENT ptr
does not point to the first byte of an allocation made through ROCr, or len is not the full length of the allocation or handle is NULL.
◆ hsa_amd_ipc_memory_detach()
hsa_status_t HSA_API hsa_amd_ipc_memory_detach | ( | void * | mapped_ptr | ) |
Decrements the reference count for the shared memory mapping and releases access to shared memory imported with hsa_amd_ipc_memory_attach.
- Parameters
-
[in] mapped_ptr Pointer to the first byte of a shared allocation imported with hsa_amd_ipc_memory_attach.
- Return values
-
HSA_STATUS_SUCCESS if mapped_ptr
was imported with hsa_amd_ipc_memory_attach.HSA_STATUS_ERROR_NOT_INITIALIZED if HSA is not initialized HSA_STATUS_ERROR_INVALID_ARGUMENT mapped_ptr
was not imported with hsa_amd_ipc_memory_attach.
◆ hsa_amd_ipc_signal_attach()
hsa_status_t HSA_API hsa_amd_ipc_signal_attach | ( | const hsa_amd_ipc_signal_t * | handle, |
hsa_signal_t * | signal | ||
) |
Imports an IPC capable signal into the local process. If an IPC signal handle is attached multiple times in a process each attach may return a different signal handle. Each returned signal handle is refcounted and requires a matching number of calls to hsa_signal_destroy to release the shared signal.
- Parameters
-
[in] handle Pointer to the identifier for the shared signal. [out] signal Recieves a process local signal handle to the shared signal.
- Return values
-
HSA_STATUS_SUCCESS if the signal is successfully imported. HSA_STATUS_ERROR_NOT_INITIALIZED if HSA is not initialized HSA_STATUS_ERROR_OUT_OF_RESOURCES if there is a failure in allocating necessary resources HSA_STATUS_ERROR_INVALID_ARGUMENT handle
is not valid.
◆ hsa_amd_ipc_signal_create()
hsa_status_t HSA_API hsa_amd_ipc_signal_create | ( | hsa_signal_t | signal, |
hsa_amd_ipc_signal_t * | handle | ||
) |
Obtains an interprocess sharing handle for a signal. The handle is valid while the signal it references remains valid in any process. In general applications should confirm that the signal has been attached (via hsa_amd_ipc_signal_attach) in the remote process prior to destroying that signal in the local process. Repeated calls for the same signal may, but are not required to, return unique handles.
- Parameters
-
[in] signal Signal created with attribute HSA_AMD_SIGNAL_IPC. [out] handle Process independent identifier referencing the shared signal.
- Return values
-
HSA_STATUS_SUCCESS handle
is ready to use for interprocess sharing.HSA_STATUS_ERROR_NOT_INITIALIZED if HSA is not initialized HSA_STATUS_ERROR_OUT_OF_RESOURCES if there is a failure in allocating necessary resources HSA_STATUS_ERROR_INVALID_ARGUMENT signal
is not a valid signal created with attribute HSA_AMD_SIGNAL_IPC or handle is NULL.
◆ hsa_amd_memory_async_copy()
hsa_status_t HSA_API hsa_amd_memory_async_copy | ( | void * | dst, |
hsa_agent_t | dst_agent, | ||
const void * | src, | ||
hsa_agent_t | src_agent, | ||
size_t | size, | ||
uint32_t | num_dep_signals, | ||
const hsa_signal_t * | dep_signals, | ||
hsa_signal_t | completion_signal | ||
) |
Asynchronously copy a block of memory from the location pointed to by src
on the src_agent
to the memory block pointed to by dst
on the dst_agent
. Because the DMA engines used may not be in the same coherency domain, the caller must ensure that buffers are system-level coherent. In general this requires the sending device to have released the buffer to system scope prior to executing the copy API and the receiving device must execute a system scope acquire fence prior to use of the destination buffer.
- Parameters
-
[out] dst Buffer where the content is to be copied. [in] dst_agent Agent associated with the dst
. The agent must be able to directly access both the source and destination buffers in their current locations. May be zero in which case the runtime will attempt to discover the destination agent. Discovery may have variable and/or high latency.[in] src A valid pointer to the source of data to be copied. The source buffer must not overlap with the destination buffer, otherwise the copy will succeed but contents of dst
is undefined.[in] src_agent Agent associated with the src
. The agent must be able to directly access both the source and destination buffers in their current locations. May be zero in which case the runtime will attempt to discover the destination agent. Discovery may have variable and/or high latency.[in] size Number of bytes to copy. If size
is 0, no copy is performed and the function returns success. Copying a number of bytes larger than the size of the buffers pointed bydst
orsrc
results in undefined behavior.[in] num_dep_signals Number of dependent signals. Can be 0. [in] dep_signals List of signals that must be waited on before the copy operation starts. The copy will start after every signal has been observed with the value 0. The dependent signal should not include completion signal from hsa_amd_memory_async_copy operation to be issued in future as that can result in a deadlock. If num_dep_signals
is 0, this argument is ignored.[in] completion_signal Signal used to indicate completion of the copy operation. When the copy operation is finished, the value of the signal is decremented. The runtime indicates that an error has occurred during the copy operation by setting the value of the completion signal to a negative number. The signal handle must not be 0.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. The application is responsible for checking for asynchronous error conditions (see the description of completion_signal
).::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_AGENT An agent is invalid or no discovered agent has access. ::HSA_STATUS_ERROR_INVALID_SIGNAL completion_signal
is invalid.::HSA_STATUS_ERROR_INVALID_ARGUMENT The source or destination pointers are NULL, or the completion signal is 0.
◆ hsa_amd_memory_async_copy_on_engine()
hsa_status_t HSA_API hsa_amd_memory_async_copy_on_engine | ( | void * | dst, |
hsa_agent_t | dst_agent, | ||
const void * | src, | ||
hsa_agent_t | src_agent, | ||
size_t | size, | ||
uint32_t | num_dep_signals, | ||
const hsa_signal_t * | dep_signals, | ||
hsa_signal_t | completion_signal, | ||
hsa_amd_sdma_engine_id_t | engine_id, | ||
bool | force_copy_on_sdma | ||
) |
Asynchronously copy a block of memory from the location pointed to by src
on the src_agent
to the memory block pointed to by dst
on the dst_agent
on engine_id.
WARNING: Concurrent use of this call with hsa_amd_memory_async_copy can result in resource conflicts as HSA runtime will auto assign engines with the latter call. Approach using both calls concurrently with caution.
All param definitions are identical to hsa_amd_memory_async_copy with the exception of engine_id and force_copy_on_sdma.
- Parameters
-
[in] - engine_id Target engine defined by hsa_amd_sdma_engine_id_t. Client should use hsa_amd_memory_copy_engine_status first to get the ID availability. [in] - force_copy_on_sdma By default, blit kernel copies are used when dst_agent == src_agent. Setting this to true will force the copy over SDMA1.
All return definitions are identical to hsa_amd_memory_async_copy with the following ammendments:
- Return values
-
::HSA_STATUS_ERROR_INVALID_ARGUMENT The source or destination pointers are NULL, or the completion signal is 0 or engine_id is improperly bounded.
◆ hsa_amd_memory_async_copy_rect()
hsa_status_t HSA_API hsa_amd_memory_async_copy_rect | ( | const hsa_pitched_ptr_t * | dst, |
const hsa_dim3_t * | dst_offset, | ||
const hsa_pitched_ptr_t * | src, | ||
const hsa_dim3_t * | src_offset, | ||
const hsa_dim3_t * | range, | ||
hsa_agent_t | copy_agent, | ||
hsa_amd_copy_direction_t | dir, | ||
uint32_t | num_dep_signals, | ||
const hsa_signal_t * | dep_signals, | ||
hsa_signal_t | completion_signal | ||
) |
◆ hsa_amd_memory_copy_engine_status()
hsa_status_t HSA_API hsa_amd_memory_copy_engine_status | ( | hsa_agent_t | dst_agent, |
hsa_agent_t | src_agent, | ||
uint32_t * | engine_ids_mask | ||
) |
Reports the availability of SDMA copy engines.
- Parameters
-
[in] dst_agent Destination agent of copy status direction. [in] src_agent Source agent of copy status direction. [out] engine_ids_mask returns available SDMA engine IDs that can be masked with hsa_amd_sdma_engine_id_t.
- Return values
-
::HSA_STATUS_SUCCESS Agent has available SDMA engines. ::HSA_STATUS_ERROR_OUT_OF_RESOURCES Agent does not have available SDMA engines. ::HSA_STATUS_ERROR_INVALID_AGENT dst_agent and src_agent are the same as dst_agent == src_agent is generally used for shader copies.
◆ hsa_amd_memory_fill()
hsa_status_t HSA_API hsa_amd_memory_fill | ( | void * | ptr, |
uint32_t | value, | ||
size_t | count | ||
) |
Sets the first count
of uint32_t of the block of memory pointed by ptr
to the specified value
.
- Parameters
-
[in] ptr Pointer to the block of memory to fill. [in] value Value to be set. [in] count Number of uint32_t element to be set to the value.
- Return values
-
HSA_STATUS_SUCCESS The function has been executed successfully. HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. HSA_STATUS_ERROR_INVALID_ARGUMENT ptr
is NULL or not 4 bytes alignedHSA_STATUS_ERROR_INVALID_ALLOCATION if the given memory region was not allocated with HSA runtime APIs.
◆ hsa_amd_memory_lock()
hsa_status_t HSA_API hsa_amd_memory_lock | ( | void * | host_ptr, |
size_t | size, | ||
hsa_agent_t * | agents, | ||
int | num_agent, | ||
void ** | agent_ptr | ||
) |
Pin a host pointer allocated by C/C++ or OS allocator (i.e. ordinary system DRAM) and return a new pointer accessible by the agents
. If the host_ptr
overlaps with previously locked memory, then the overlap area is kept locked (i.e multiple mappings are permitted). In this case, the same input host_ptr
may give different locked agent_ptr
and when it does, they are not necessarily coherent (i.e. accessing either agent_ptr
is not equivalent). Accesses to agent_ptr
are coarse grained.
- Parameters
-
[in] host_ptr A buffer allocated by C/C++ or OS allocator. [in] size The size to be locked. [in] agents Array of agent handle to gain access to the host_ptr
. If this parameter is NULL and thenum_agent
is 0, all agents in the platform will gain access to thehost_ptr
.[out] agent_ptr Pointer to the location where to store the new address.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_OUT_OF_RESOURCES There is a failure in allocating the necessary resources. ::HSA_STATUS_ERROR_INVALID_AGENT One or more agent in agents
is invalid.::HSA_STATUS_ERROR_INVALID_ARGUMENT size
is 0 orhost_ptr
oragent_ptr
is NULL oragents
not NULL butnum_agent
is 0 oragents
is NULL butnum_agent
is not 0.
◆ hsa_amd_memory_lock_to_pool()
hsa_status_t HSA_API hsa_amd_memory_lock_to_pool | ( | void * | host_ptr, |
size_t | size, | ||
hsa_agent_t * | agents, | ||
int | num_agent, | ||
hsa_amd_memory_pool_t | pool, | ||
uint32_t | flags, | ||
void ** | agent_ptr | ||
) |
Pin a host pointer allocated by C/C++ or OS allocator (i.e. ordinary system DRAM) and return a new pointer accessible by the agents
. If the host_ptr
overlaps with previously locked memory, then the overlap area is kept locked (i.e. multiple mappings are permitted). In this case, the same input host_ptr
may give different locked agent_ptr
and when it does, they are not necessarily coherent (i.e. accessing either agent_ptr
is not equivalent). Acesses to the memory via agent_ptr
have the same access properties as memory allocated from pool
as determined by hsa_amd_memory_pool_get_info and hsa_amd_agent_memory_pool_get_info (ex. coarse/fine grain, platform atomic support, link info). Physical composition and placement of the memory (ex. page size, NUMA binding) is not changed.
- Parameters
-
[in] host_ptr A buffer allocated by C/C++ or OS allocator. [in] size The size to be locked. [in] agents Array of agent handle to gain access to the host_ptr
. If this parameter is NULL and thenum_agent
is 0, all agents in the platform will gain access to thehost_ptr
.[in] pool Global memory pool owned by a CPU agent. [in] flags A bit-field that is used to specify allocation directives. Reserved parameter, must be 0. [out] agent_ptr Pointer to the location where to store the new address.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_OUT_OF_RESOURCES There is a failure in allocating the necessary resources. ::HSA_STATUS_ERROR_INVALID_AGENT One or more agent in agents
is invalid or can not accesspool
.HSA_STATUS_ERROR_INVALID_MEMORY_POOL pool
is invalid or not owned by a CPU agent.::HSA_STATUS_ERROR_INVALID_ARGUMENT size
is 0 orhost_ptr
oragent_ptr
is NULL oragents
not NULL butnum_agent
is 0 oragents
is NULL butnum_agent
is not 0 or flags is not 0.
◆ hsa_amd_memory_migrate()
hsa_status_t HSA_API hsa_amd_memory_migrate | ( | const void * | ptr, |
hsa_amd_memory_pool_t | memory_pool, | ||
uint32_t | flags | ||
) |
Relocate a buffer to a new memory pool.
When a buffer is migrated, its virtual address remains the same but its physical contents are moved to the indicated memory pool.
After migration, only the agent associated with the destination pool will have access.
The caller is also responsible for ensuring that the allocation in the source memory pool where the buffer is currently located can be migrated to the specified destination memory pool (using hsa_amd_memory_pool_can_migrate returns a value of true for the source and destination memory pools), otherwise behavior is undefined.
The caller must ensure that the buffer is not accessed while it is migrated.
- Parameters
-
[in] ptr Buffer to be relocated. The buffer must have been released to system prior to call this API. The buffer will be released to system upon completion. [in] memory_pool Memory pool where to place the buffer. [in] flags A bit-field that is used to specify migration information. Must be zero.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. HSA_STATUS_ERROR_INVALID_MEMORY_POOL The destination memory pool is invalid. ::HSA_STATUS_ERROR_OUT_OF_RESOURCES There is a failure in allocating the necessary resources. ::HSA_STATUS_ERROR_INVALID_ARGUMENT flags
is not 0.
◆ hsa_amd_memory_pool_allocate()
hsa_status_t HSA_API hsa_amd_memory_pool_allocate | ( | hsa_amd_memory_pool_t | memory_pool, |
size_t | size, | ||
uint32_t | flags, | ||
void ** | ptr | ||
) |
Allocate a block of memory (or buffer) in the specified pool.
- Parameters
-
[in] memory_pool Memory pool where to allocate memory from. The memory pool must have the HSA_AMD_MEMORY_POOL_INFO_RUNTIME_ALLOC_ALLOWED flag set. [in] size Allocation size, in bytes. Must not be zero. This value is rounded up to the nearest multiple of HSA_AMD_MEMORY_POOL_INFO_RUNTIME_ALLOC_GRANULE in memory_pool
.[in] flags A bit-field that is used to specify allocation directives. [out] ptr Pointer to the location where to store the base virtual address of the allocated block. The returned base address is aligned to the value of HSA_AMD_MEMORY_POOL_INFO_RUNTIME_ALLOC_ALIGNMENT in memory_pool
. If the allocation fails, the returned value is undefined.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_OUT_OF_RESOURCES No memory is available. HSA_STATUS_ERROR_INVALID_MEMORY_POOL The memory pool is invalid. ::HSA_STATUS_ERROR_INVALID_ALLOCATION The host is not allowed to allocate memory in memory_pool
, orsize
is greater than the value of HSA_AMD_MEMORY_POOL_INFO_ALLOC_MAX_SIZE inmemory_pool
.::HSA_STATUS_ERROR_INVALID_ARGUMENT ptr
is NULL, orsize
is 0, or flags is not 0.
◆ hsa_amd_memory_pool_can_migrate()
hsa_status_t HSA_API hsa_amd_memory_pool_can_migrate | ( | hsa_amd_memory_pool_t | src_memory_pool, |
hsa_amd_memory_pool_t | dst_memory_pool, | ||
bool * | result | ||
) |
Query if buffers currently located in some memory pool can be relocated to a destination memory pool.
If the returned value is non-zero, a migration of a buffer to dst_memory_pool
using hsa_amd_memory_migrate may nevertheless fail due to resource limitations.
- Parameters
-
[in] src_memory_pool Source memory pool. [in] dst_memory_pool Destination memory pool. [out] result Pointer to a memory location where the result of the query is stored. Must not be NULL. If buffers currently located in src_memory_pool
can be relocated todst_memory_pool
, the result is true.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. HSA_STATUS_ERROR_INVALID_MEMORY_POOL One of the memory pools is invalid. ::HSA_STATUS_ERROR_INVALID_ARGUMENT result
is NULL.
◆ hsa_amd_memory_pool_free()
hsa_status_t HSA_API hsa_amd_memory_pool_free | ( | void * | ptr | ) |
Deallocate a block of memory previously allocated using hsa_amd_memory_pool_allocate.
- Parameters
-
[in] ptr Pointer to a memory block. If ptr
does not match a value previously returned by hsa_amd_memory_pool_allocate, the behavior is undefined.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized.
◆ hsa_amd_memory_pool_get_info()
hsa_status_t HSA_API hsa_amd_memory_pool_get_info | ( | hsa_amd_memory_pool_t | memory_pool, |
hsa_amd_memory_pool_info_t | attribute, | ||
void * | value | ||
) |
Get the current value of an attribute of a memory pool.
- Parameters
-
[in] memory_pool A valid memory pool. [in] attribute Attribute to query. [out] value Pointer to a application-allocated buffer where to store the value of the attribute. If the buffer passed by the application is not large enough to hold the value of attribute
, the behavior is undefined.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully.
◆ hsa_amd_memory_unlock()
hsa_status_t HSA_API hsa_amd_memory_unlock | ( | void * | host_ptr | ) |
Unpin the host pointer previously pinned via hsa_amd_memory_lock or hsa_amd_memory_lock_to_pool.
The behavior is undefined if the host pointer being unpinned does not match previous pinned address or if the host pointer was already deallocated.
- Parameters
-
[in] host_ptr A buffer allocated by C/C++ or OS allocator that was pinned previously via hsa_amd_memory_lock or hsa_amd_memory_lock_to_pool.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized.
◆ hsa_amd_pointer_info()
hsa_status_t HSA_API hsa_amd_pointer_info | ( | const void * | ptr, |
hsa_amd_pointer_info_t * | info, | ||
void *(*)(size_t) | alloc, | ||
uint32_t * | num_agents_accessible, | ||
hsa_agent_t ** | accessible | ||
) |
Retrieves information about the allocation referenced by the given pointer. Optionally returns the number and list of agents which can directly access the allocation. In case this virtual address is unknown, the pointer type returned will be HSA_EXT_POINTER_TYPE_UNKNOWN and the only fields that are valid after hsa_amd_pointer_info returns are size and type.
- Parameters
-
[in] ptr Pointer which references the allocation to retrieve info for. [in,out] info Pointer to structure to be filled with allocation info. Data member size must be set to the size of the structure prior to calling hsa_amd_pointer_info. On return size will be set to the size of the pointer info structure supported by the runtime, if smaller. Members beyond the returned value of size will not be updated by the API. Must not be NULL. [in] alloc Function pointer to an allocator used to allocate the accessible
array. If NULLaccessible
will not be returned.[out] num_agents_accessible Recieves the count of agents in accessible
. If NULLaccessible
will not be returned.[out] accessible Recieves a pointer to the array, allocated by alloc
, holding the list of agents which may directly access the allocation. May be NULL.
- Return values
-
HSA_STATUS_SUCCESS Info retrieved successfully HSA_STATUS_ERROR_NOT_INITIALIZED if HSA is not initialized HSA_STATUS_ERROR_OUT_OF_RESOURCES if there is a failure in allocating necessary resources HSA_STATUS_ERROR_INVALID_ARGUMENT NULL in ptr
orinfo
.
◆ hsa_amd_pointer_info_set_userdata()
hsa_status_t HSA_API hsa_amd_pointer_info_set_userdata | ( | const void * | ptr, |
void * | userdata | ||
) |
Associates an arbitrary pointer with an allocation known to ROCr. The pointer can be fetched by hsa_amd_pointer_info in the userData field.
- Parameters
-
[in] ptr Pointer to the first byte of an allocation known to ROCr with which to associate userdata
.[in] userdata Abitrary pointer to associate with the allocation.
- Return values
-
HSA_STATUS_SUCCESS userdata
successfully stored.HSA_STATUS_ERROR_NOT_INITIALIZED if HSA is not initialized HSA_STATUS_ERROR_OUT_OF_RESOURCES if there is a failure in allocating necessary resources HSA_STATUS_ERROR_INVALID_ARGUMENT ptr
is not known to ROCr.
◆ hsa_amd_portable_close_dmabuf()
hsa_status_t hsa_amd_portable_close_dmabuf | ( | int | dmabuf | ) |
Closes an OS specific, vendor neutral, handle to a memory allocation.
Closes an OS specific handle to GPU agent memory.
Applications should close a handle after imports are complete. The handle is not required to remain open for the lifetime of imported mappings. The referenced allocation will remain valid until all handles and mappings are closed.
- Parameters
-
[in] dmabuf Handle to be closed.
- Return values
-
::HSA_STATUS_SUCCESS Handle closed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_RESOURCE_FREE A generic error was encountered when closing the handle. The handle may have been closed already or an async IO error may have occured.
◆ hsa_amd_portable_export_dmabuf()
hsa_status_t hsa_amd_portable_export_dmabuf | ( | const void * | ptr, |
size_t | size, | ||
int * | dmabuf, | ||
uint64_t * | offset | ||
) |
Obtains an OS specific, vendor neutral, handle to a memory allocation.
Obtains an OS specific handle to GPU agent memory. The memory must be part of a single allocation from an hsa_amd_memory_pool_t exposed by a GPU Agent. The handle may be used with other APIs (e.g. Vulkan) to obtain shared access to the allocation.
Shared access to the memory is not guaranteed to be fine grain coherent even if the allocation exported is from a fine grain pool. The shared memory consistency model will be no stronger than the model exported from, consult the importing API to determine the final consistency model.
The allocation's memory remains valid as long as the handle and any mapping of the handle remains valid. When the handle and all mappings are closed the backing memory will be released for reuse.
- Parameters
-
[in] ptr Pointer to the allocation being exported. [in] size Size in bytes to export following ptr
. The entire range being exported must be contained within a single allocation.[out] dmabuf Pointer to a dma-buf file descriptor holding a reference to the allocation. Contents will not be altered in the event of failure. [out] offset Offset in bytes into the memory referenced by the dma-buf object at which ptr
resides. Contents will not be altered in the event of failure.
- Return values
-
::HSA_STATUS_SUCCESS Export completed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_ARGUMENT One or more arguments is NULL. ::HSA_STATUS_ERROR_INVALID_ALLOCATION The address range described by ptr
andsize
are not contained within a single allocation.::HSA_STATUS_ERROR_INVALID_AGENT The allocation described by ptr
andsize
was allocated on a device which can not export memory.::HSA_STATUS_ERROR_OUT_OF_RESOURCES The return file descriptor, dmabuf
, could not be created.
◆ hsa_amd_profiling_async_copy_enable()
hsa_status_t HSA_API hsa_amd_profiling_async_copy_enable | ( | bool | enable | ) |
Enable or disable asynchronous memory copy profiling.
The runtime will provide the copy processing start timestamp and completion timestamp of each call to hsa_amd_memory_async_copy if the async copy profiling is enabled prior to the call to hsa_amd_memory_async_copy. The completion signal object is used to hold the last async copy start and end timestamp. The client can retrieve these timestamps via call to hsa_amd_profiling_get_async_copy_time.
- Parameters
-
[in] enable True to enable profiling. False to disable profiling.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_OUT_OF_RESOURCES Failed on allocating resources needed to profile the asynchronous copy.
◆ hsa_amd_profiling_convert_tick_to_system_domain()
hsa_status_t HSA_API hsa_amd_profiling_convert_tick_to_system_domain | ( | hsa_agent_t | agent, |
uint64_t | agent_tick, | ||
uint64_t * | system_tick | ||
) |
Computes the frequency ratio and offset between the agent clock and HSA system clock and converts the agent's tick to HSA system domain tick.
- Parameters
-
[in] agent The agent used to retrieve the agent_tick. It is user's responsibility to make sure the tick number is from this agent, otherwise, the behavior is undefined. [in] agent_tick The tick count retrieved from the specified agent
.[out] system_tick The translated HSA system domain clock counter tick.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_AGENT The agent is invalid. ::HSA_STATUS_ERROR_INVALID_ARGUMENT system_tick
is NULL;
◆ hsa_amd_profiling_get_async_copy_time()
hsa_status_t HSA_API hsa_amd_profiling_get_async_copy_time | ( | hsa_signal_t | signal, |
hsa_amd_profiling_async_copy_time_t * | time | ||
) |
Retrieve asynchronous copy timestamps.
Async copy profiling is enabled via call to hsa_amd_profiling_async_copy_enable.
- Parameters
-
[in] signal A signal used as the completion signal of the call to hsa_amd_memory_async_copy. [out] time Async copy processing timestamps in the HSA system clock domain.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_SIGNAL The signal is invalid. ::HSA_STATUS_ERROR_INVALID_ARGUMENT time
is NULL.
◆ hsa_amd_profiling_get_dispatch_time()
hsa_status_t HSA_API hsa_amd_profiling_get_dispatch_time | ( | hsa_agent_t | agent, |
hsa_signal_t | signal, | ||
hsa_amd_profiling_dispatch_time_t * | time | ||
) |
Retrieve packet processing time stamps.
- Parameters
-
[in] agent The agent with which the signal was last used. For instance, if the profiled dispatch packet is dispatched onto queue Q, which was created on agent A, then this parameter must be A. [in] signal A signal used as the completion signal of the dispatch packet to retrieve time stamps from. This dispatch packet must have been issued to a queue with profiling enabled and have already completed. Also the signal must not have yet been used in any other packet following the completion of the profiled dispatch packet. [out] time Packet processing timestamps in the HSA system clock domain.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_AGENT The agent is invalid. ::HSA_STATUS_ERROR_INVALID_SIGNAL The signal is invalid. ::HSA_STATUS_ERROR_INVALID_ARGUMENT time
is NULL.
◆ hsa_amd_profiling_set_profiler_enabled()
hsa_status_t HSA_API hsa_amd_profiling_set_profiler_enabled | ( | hsa_queue_t * | queue, |
int | enable | ||
) |
Enable or disable profiling capability of a queue.
- Parameters
-
[in] queue A valid queue. [in] enable 1 to enable profiling. 0 to disable profiling.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_QUEUE The queue is invalid. ::HSA_STATUS_ERROR_INVALID_ARGUMENT queue
is NULL.
◆ hsa_amd_queue_cu_get_mask()
hsa_status_t HSA_API hsa_amd_queue_cu_get_mask | ( | const hsa_queue_t * | queue, |
uint32_t | num_cu_mask_count, | ||
uint32_t * | cu_mask | ||
) |
Retrieve a queue's CU affinity mask.
Returns the first num_cu_mask_count bits of a queue's CU mask. Ensure that num_cu_mask_count is at least as large as HSA_AMD_AGENT_INFO_COMPUTE_UNIT_COUNT to retrieve the entire mask.
- Parameters
-
[in] queue A pointer to HSA queue. [in] num_cu_mask_count Size of CUMask bit array passed in, in bits. [out] cu_mask Bit-vector representing the CU mask.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_QUEUE queue
is NULL or invalid.::HSA_STATUS_ERROR_INVALID_ARGUMENT num_cu_mask_count
is 0, not a multiple of 32 orcu_mask
is NULL.
◆ hsa_amd_queue_cu_set_mask()
hsa_status_t HSA_API hsa_amd_queue_cu_set_mask | ( | const hsa_queue_t * | queue, |
uint32_t | num_cu_mask_count, | ||
const uint32_t * | cu_mask | ||
) |
Set a queue's CU affinity mask.
Enables the queue to run on only selected CUs. The given mask is combined by bitwise AND with any device wide mask in HSA_CU_MASK before being applied. If num_cu_mask_count is 0 then the request is interpreted as a request to enable all CUs and no cu_mask array need be given.
- Parameters
-
[in] queue A pointer to HSA queue. [in] num_cu_mask_count Size of CUMask bit array passed in, in bits. [in] cu_mask Bit-vector representing the CU mask.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. HSA_STATUS_CU_MASK_REDUCED The function was successfully executed but the given mask attempted to enable a CU which was disabled by HSA_CU_MASK. CUs disabled by HSA_CU_MASK remain disabled. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_QUEUE queue
is NULL or invalid.::HSA_STATUS_ERROR_INVALID_ARGUMENT num_cu_mask_count
is not a multiple of 32 ornum_cu_mask_count
is not 0 and cu_mask is NULL. Devices with work group processors must even-index contiguous pairwise CU enable e.g. 0x33(b'110011) is valid while 0x5(0x101) and 0x6(b'0110) are invalid.
◆ hsa_amd_queue_get_info()
hsa_status_t hsa_amd_queue_get_info | ( | hsa_queue_t * | queue, |
hsa_queue_info_attribute_t | attribute, | ||
void * | value | ||
) |
◆ hsa_amd_queue_set_priority()
hsa_status_t HSA_API hsa_amd_queue_set_priority | ( | hsa_queue_t * | queue, |
hsa_amd_queue_priority_t | priority | ||
) |
Modifies the dispatch and wavefront scheduling prioirty for a given compute queue. The default is HSA_AMD_QUEUE_PRIORITY_NORMAL.
- Parameters
-
[in] queue Compute queue to apply new priority to. [in] priority Priority to associate with queue.
- Return values
-
HSA_STATUS_SUCCESS if priority was changed successfully. HSA_STATUS_ERROR_INVALID_QUEUE if queue is not a valid compute queue handle. HSA_STATUS_ERROR_INVALID_ARGUMENT if priority is not a valid value from hsa_amd_queue_priority_t.
◆ hsa_amd_register_deallocation_callback()
hsa_status_t HSA_API hsa_amd_register_deallocation_callback | ( | void * | ptr, |
hsa_amd_deallocation_callback_t | callback, | ||
void * | user_data | ||
) |
Registers a deallocation notifier monitoring for release of agent accessible address ptr
. If successful, callback
will be invoked when ptr
is removed from accessibility from all agents.
Notification callbacks are automatically deregistered when they are invoked.
Note: The current version supports notifications of address release originating from hsa_amd_memory_pool_free. Support for other address release APIs will follow.
- Parameters
-
[in] ptr Agent accessible address to monitor for deallocation. Passed to callback
.[in] callback Notifier to be invoked when ptr
is released from agent accessibility.[in] user_data User provided value passed to callback
. May be NULL.
- Return values
-
::HSA_STATUS_SUCCESS The notifier registered successfully ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_ALLOCATION ptr
does not refer to a valid agent accessible address.::HSA_STATUS_ERROR_INVALID_ARGUMENT callback
is NULL orptr
is NULL.::HSA_STATUS_ERROR_OUT_OF_RESOURCES if there is a failure in allocating necessary resources
◆ hsa_amd_register_system_event_handler()
hsa_status_t HSA_API hsa_amd_register_system_event_handler | ( | hsa_amd_system_event_callback_t | callback, |
void * | data | ||
) |
Register AMD GPU event handler.
- Parameters
-
[in] callback Callback to be invoked when an event is triggered. The HSA runtime passes two arguments to the callback: event
is defined per event by the HSA runtime, anddata
is the user data.[in] data User data that is passed to callback
. May be NULL.
- Return values
-
HSA_STATUS_SUCCESS The handler has been registered successfully. HSA_STATUS_ERROR An event handler has already been registered. HSA_STATUS_ERROR_INVALID_ARGUMENT event
is invalid.
◆ hsa_amd_signal_async_handler()
hsa_status_t HSA_API hsa_amd_signal_async_handler | ( | hsa_signal_t | signal, |
hsa_signal_condition_t | cond, | ||
hsa_signal_value_t | value, | ||
hsa_amd_signal_handler | handler, | ||
void * | arg | ||
) |
Register asynchronous signal handler function.
Allows registering a callback function and user provided value with a signal and wait condition. The callback will be invoked if the associated signal and wait condition are satisfied. Callbacks will be invoked serially but in an arbitrary order so callbacks should be independent of each other. After being invoked a callback may continue to wait for its associated signal and condition and, possibly, be invoked again. Or the callback may stop waiting. If the callback returns true then it will continue waiting and may be called again. If false then the callback will not wait again and will not be called again for the associated signal and condition. It is possible to register the same callback multiple times with the same or different signals and/or conditions. Each registration of the callback will be treated entirely independently.
- Parameters
-
[in] signal hsa signal to be asynchronously monitored [in] cond condition value to monitor for [in] value signal value used in condition expression [in] handler asynchronous signal handler invoked when signal's condition is met [in] arg user provided value which is provided to handler when handler is invoked
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_SIGNAL signal is not a valid hsa_signal_t ::HSA_STATUS_ERROR_INVALID_ARGUMENT handler is invalid (NULL) ::HSA_STATUS_ERROR_OUT_OF_RESOURCES The HSA runtime is out of resources or blocking signals are not supported by the HSA driver component.
◆ hsa_amd_signal_create()
hsa_status_t HSA_API hsa_amd_signal_create | ( | hsa_signal_value_t | initial_value, |
uint32_t | num_consumers, | ||
const hsa_agent_t * | consumers, | ||
uint64_t | attributes, | ||
hsa_signal_t * | signal | ||
) |
Create a signal with specific attributes.
- Parameters
-
[in] initial_value Initial value of the signal. [in] num_consumers Size of consumers
. A value of 0 indicates that any agent might wait on the signal.[in] consumers List of agents that might consume (wait on) the signal. If num_consumers
is 0, this argument is ignored; otherwise, the HSA runtime might use the list to optimize the handling of the signal object. If an agent not listed inconsumers
waits on the returned signal, the behavior is undefined. The memory associated withconsumers
can be reused or freed after the function returns.[in] attributes Requested signal attributes. Multiple signal attributes may be requested by combining them with bitwise OR. Requesting no attributes ( attributes
== 0) results in the same signal as would have been obtained via hsa_signal_create.[out] signal Pointer to a memory location where the HSA runtime will store the newly created signal handle. Must not be NULL.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_OUT_OF_RESOURCES The HSA runtime failed to allocate the required resources. ::HSA_STATUS_ERROR_INVALID_ARGUMENT signal
is NULL,num_consumers
is greater than 0 butconsumers
is NULL, orconsumers
contains duplicates.
◆ hsa_amd_signal_value_pointer()
hsa_status_t hsa_amd_signal_value_pointer | ( | hsa_signal_t | signal, |
volatile hsa_signal_value_t ** | value_ptr | ||
) |
Returns a pointer to the value of a signal.
Use of this API does not modify the lifetime of ::signal and any hsa_signal_value_t retrieved by this API has lifetime equal to that of ::signal.
This API is intended for partial interoperability with non-HSA compatible devices and should not be used where HSA interfaces are available.
Use of the signal value must comply with use restritions of ::signal. Use may result in data races if the operations performed are not platform atomic. Use with HSA_AMD_SIGNAL_AMD_GPU_ONLY or HSA_AMD_SIGNAL_IPC attributed signals is required.
- Parameters
-
[in] Signal handle to extract the signal value pointer from. [out] Location where the extracted signal value pointer will be placed.
- Return values
-
::HSA_STATUS_SUCCESS The function has been executed successfully. ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_SIGNAL signal is not a valid hsa_signal_t ::HSA_STATUS_ERROR_INVALID_ARGUMENT value_ptr is NULL.
◆ hsa_amd_signal_wait_any()
uint32_t HSA_API hsa_amd_signal_wait_any | ( | uint32_t | signal_count, |
hsa_signal_t * | signals, | ||
hsa_signal_condition_t * | conds, | ||
hsa_signal_value_t * | values, | ||
uint64_t | timeout_hint, | ||
hsa_wait_state_t | wait_hint, | ||
hsa_signal_value_t * | satisfying_value | ||
) |
Wait for any signal-condition pair to be satisfied.
Allows waiting for any of several signal and conditions pairs to be satisfied. The function returns the index into the list of signals of the first satisfying signal-condition pair. The value of the satisfying signal's value is returned in satisfying_value unless satisfying_value is NULL. This function provides only relaxed memory semantics.
◆ hsa_amd_spm_acquire()
hsa_status_t hsa_amd_spm_acquire | ( | hsa_agent_t | preferred_agent | ) |
Acquire Stream Performance Monitor on an agent.
Acquire exclusive use of SPM on preferred_agent
. See hsa_amd_spm_set_dest_buffer to provide a destination buffer to KFD to start recording and retrieve this data.
- Parameters
-
[in] preferred_agent Agent on which to acquire SPM
◆ hsa_amd_spm_release()
hsa_status_t hsa_amd_spm_release | ( | hsa_agent_t | preferred_agent | ) |
Release Stream Performance Monitor on an agent.
Release exclusive use of SPM on preferred_agent
. This will stop KFD writing SPM data. If a destination buffer is set, then data in the destination buffer is available to user when this function returns.
- Parameters
-
[in] preferred_agent Agent on which to release SPM
◆ hsa_amd_spm_set_dest_buffer()
hsa_status_t hsa_amd_spm_set_dest_buffer | ( | hsa_agent_t | preferred_agent, |
size_t | size_in_bytes, | ||
uint32_t * | timeout, | ||
uint32_t * | size_copied, | ||
void * | dest, | ||
bool * | is_data_loss | ||
) |
Set up the current destination user mode buffer for stream performance counter data. KFD will start writing SPM data into the destination buffer. KFD will continue to copy data into the current destination buffer until any of the following functions are called.
- hsa_amd_spm_release
- hsa_amd_spm_set_dest_buffer with dest set to NULL
- hsa_amd_spm_set_dest_buffer with dest set to a new buffer
if timeout
is non-0, the call will wait for up to timeout
ms for the previous buffer to be filled. If previous buffer to be filled before timeout, the timeout
will be updated value with the time remaining. If the timeout is exceeded, the function copies any partial data available into the previous user buffer and returns success. User should not access destination data while KFD is copying data. If the previous destination buffer was full, then is_data_loss
flag is set. dest
is CPU accessible memory. It could be malloc'ed memory or host allocated memory
- Parameters
-
[in] preferred_agent Agent on which to set the dest buffer [in] size_in_bytes size of the buffer [in/out] timeout timeout in milliseconds [out] size_copied number of bytes copied [in] dest destination address. Set to NULL to stop copy on previous buffer [out] is_data_loss true is data was lost
◆ hsa_amd_svm_attributes_get()
hsa_status_t hsa_amd_svm_attributes_get | ( | void * | ptr, |
size_t | size, | ||
hsa_amd_svm_attribute_pair_t * | attribute_list, | ||
size_t | attribute_count | ||
) |
Gets SVM memory attributes.
Attributes HSA_AMD_SVM_ATTRIB_AGENT_ACCESSIBLE, HSA_AMD_SVM_ATTRIB_AGENT_ACCESSIBLE_IN_PLACE and HSA_AMD_SVM_ATTRIB_PREFETCH_LOCATION may not be used with this API.
Note that attribute HSA_AMD_SVM_ATTRIB_ACCESS_QUERY takes as input an hsa_agent_t and returns the current access type through its attribute field.
- Parameters
-
[in] ptr Will be aligned down to nearest page boundary. [in] size Will be aligned up to nearest page boundary. [in] attribute_list List of attributes to set for the address range. [in] attribute_count Length of attribute_list
.
◆ hsa_amd_svm_attributes_set()
hsa_status_t hsa_amd_svm_attributes_set | ( | void * | ptr, |
size_t | size, | ||
hsa_amd_svm_attribute_pair_t * | attribute_list, | ||
size_t | attribute_count | ||
) |
Sets SVM memory attributes.
If HSA_AMD_SYSTEM_INFO_SVM_ACCESSIBLE_BY_DEFAULT returns false then enabling access to an Agent via this API (setting HSA_AMD_SVM_ATTRIB_AGENT_ACCESSIBLE or HSA_AMD_SVM_ATTRIB_AGENT_ACCESSIBLE_IN_PLACE) is required prior to SVM memory access by that Agent.
Attributes HSA_AMD_SVM_ATTRIB_ACCESS_QUERY and HSA_AMD_SVM_ATTRIB_PREFETCH_LOCATION may not be used with this API.
- Parameters
-
[in] ptr Will be aligned down to nearest page boundary. [in] size Will be aligned up to nearest page boundary. [in] attribute_list List of attributes to set for the address range. [in] attribute_count Length of attribute_list
.
◆ hsa_amd_svm_prefetch_async()
hsa_status_t hsa_amd_svm_prefetch_async | ( | void * | ptr, |
size_t | size, | ||
hsa_agent_t | agent, | ||
uint32_t | num_dep_signals, | ||
const hsa_signal_t * | dep_signals, | ||
hsa_signal_t | completion_signal | ||
) |
Asynchronously migrates memory to an agent.
Schedules memory migration to agent
when dep_signals
have been observed equal to zero. completion_signal
will decrement when the migration is complete.
- Parameters
-
[in] ptr Will be aligned down to nearest page boundary. [in] size Will be aligned up to nearest page boundary. [in] agent Agent to migrate to. [in] num_dep_signals Number of dependent signals. Can be 0. [in] dep_signals List of signals that must be waited on before the migration operation starts. The migration will start after every signal has been observed with the value 0. If num_dep_signals
is 0, this argument is ignored.[in] completion_signal Signal used to indicate completion of the migration operation. When the migration operation is finished, the value of the signal is decremented. The runtime indicates that an error has occurred during the copy operation by setting the value of the completion signal to a negative number. If no completion signal is required this handle may be null.
◆ hsa_amd_vmem_address_free()
hsa_status_t hsa_amd_vmem_address_free | ( | void * | va, |
size_t | size | ||
) |
Free a reserved address range.
Free a previously allocated address range. The size must match the size of a previously allocated address range.
- Parameters
-
[out] va virtual address to be freed [in] size of address range
- Return values
-
::HSA_STATUS_SUCCESS Address range released successfully ::HSA_STATUS_ERROR_INVALID_ALLOCATION Invalid va specified ::HSA_STATUS_ERROR_INVALID_ARGUMENT Invalid size specified ::HSA_STATUS_ERROR_RESOURCE_FREE Address range is still in use ::HSA_STATUS_ERROR Internal unexpected error
◆ hsa_amd_vmem_address_reserve()
hsa_status_t hsa_amd_vmem_address_reserve | ( | void ** | va, |
size_t | size, | ||
uint64_t | address, | ||
uint64_t | flags | ||
) |
Allocate a reserved address range.
Reserve a virtual address range. The size must be a multiple of the system page size. If it is not possible to allocate the address specified by address
, then va
will be a different address range. Address range should be released by calling hsa_amd_vmem_address_free.
- Parameters
-
[out] va virtual address allocated [in] size of address range requested [in] address requested [in] flags currently unsupported
- Return values
-
::HSA_STATUS_SUCCESS Address range allocated successfully ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_OUT_OF_RESOURCES Insufficient resources to allocate an address range of this size.
Note that this API will be deprecated in a future release and replaced by hsa_amd_vmem_address_reserve_align
◆ hsa_amd_vmem_address_reserve_align()
hsa_status_t hsa_amd_vmem_address_reserve_align | ( | void ** | va, |
size_t | size, | ||
uint64_t | address, | ||
uint64_t | alignment, | ||
uint64_t | flags | ||
) |
Allocate a reserved address range.
Reserve a virtual address range. The size must be a multiple of the system page size. If it is not possible to allocate the address specified by address
, then va
will be a different address range. Address range should be released by calling hsa_amd_vmem_address_free.
- Parameters
-
[out] va virtual address allocated [in] size of address range requested [in] address requested [in] alignment requested. 0 for default. Must be >= page-size and a power of 2 [in] flags currently unsupported
- Return values
-
::HSA_STATUS_SUCCESS Address range allocated successfully ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_OUT_OF_RESOURCES Insufficient resources to allocate an address range of this size.
◆ hsa_amd_vmem_export_shareable_handle()
hsa_status_t hsa_amd_vmem_export_shareable_handle | ( | int * | dmabuf_fd, |
hsa_amd_vmem_alloc_handle_t | handle, | ||
uint64_t | flags | ||
) |
Get an exportable shareable handle.
Get an exportable shareable handle for a memory_handle. This shareabl handle can then be used to re-create a virtual memory handle using hsa_amd_vmem_import_shareable_handle. The shareable handle can be transferred using mechanisms that support posix file descriptors Once all shareable handles are closed, the memory_handle is released.
- Parameters
-
[out] dmabuf_fd shareable handle [in] handle previously allocated virtual memory handle [in] flags Currently unsupported
- Return values
-
::HSA_STATUS_SUCCESS ::HSA_STATUS_ERROR_INVALID_ALLOCATION Invalid memory handle ::HSA_STATUS_ERROR_OUT_OF_RESOURCES Out of resources ::HSA_STATUS_ERROR Unexpected internal error
◆ hsa_amd_vmem_get_access()
hsa_status_t hsa_amd_vmem_get_access | ( | void * | va, |
hsa_access_permission_t * | perms, | ||
hsa_agent_t | agent_handle | ||
) |
Get current access permissions for memory mapping.
Get access permissions for memory mapping for specific agent.
- Parameters
-
[in] va previously mapped virtual address [in] perms current permissions [in] agent_handle agent
- Return values
-
::HSA_STATUS_SUCCESS ::HSA_STATUS_ERROR_INVALID_AGENT Invalid agent ::HSA_STATUS_ERROR_INVALID_ALLOCATION va is not mapped or permissions never set for this agent ::HSA_STATUS_ERROR Unexpected internal error
◆ hsa_amd_vmem_get_alloc_properties_from_handle()
hsa_status_t hsa_amd_vmem_get_alloc_properties_from_handle | ( | hsa_amd_vmem_alloc_handle_t | memory_handle, |
hsa_amd_memory_pool_t * | pool, | ||
hsa_amd_memory_type_t * | type | ||
) |
Returns the current allocation properties of a handle.
Returns the allocation properties of an existing handle
- Parameters
-
[in] memory_handle memory handle to be queried [out] pool memory pool that owns this handle [out] memory type
- Return values
-
::HSA_STATUS_SUCCESS ::HSA_STATUS_ERROR_INVALID_ALLOCATION Invalid memory_handle
◆ hsa_amd_vmem_handle_create()
hsa_status_t hsa_amd_vmem_handle_create | ( | hsa_amd_memory_pool_t | pool, |
size_t | size, | ||
hsa_amd_memory_type_t | type, | ||
uint64_t | flags, | ||
hsa_amd_vmem_alloc_handle_t * | memory_handle | ||
) |
Create a virtual memory handle.
Create a virtual memory handle within this pool size
must be a aligned to allocation granule size for this memory pool, see HSA_AMD_MEMORY_POOL_INFO_RUNTIME_ALLOC_GRANULE To minimize internal memory fragmentation, align the size to the recommended allocation granule size, see HSA_AMD_MEMORY_POOL_INFO_RUNTIME_ALLOC_REC_GRANULE
- Parameters
-
[in] pool memory to use [in] size of the memory allocation [in] type of memory [in] flags - currently unsupported [out] memory_handle - handle for the allocation
- Return values
-
::HSA_STATUS_SUCCESS memory allocated successfully ::HSA_STATUS_ERROR_NOT_INITIALIZED The HSA runtime has not been initialized. ::HSA_STATUS_ERROR_INVALID_ARGUMENT Invalid arguments ::HSA_STATUS_ERROR_INVALID_ALLOCATION This memory pool does not support allocations ::HSA_STATUS_ERROR_OUT_OF_RESOURCES Insufficient resources to allocate this memory
◆ hsa_amd_vmem_handle_release()
hsa_status_t hsa_amd_vmem_handle_release | ( | hsa_amd_vmem_alloc_handle_t | memory_handle | ) |
Release a virtual memory handle.
- Parameters
-
[in] memory handle that was previously allocated
- Return values
-
::HSA_STATUS_SUCCESS Address range allocated successfully ::HSA_STATUS_ERROR_INVALID_ALLOCATION Invalid memory handle
◆ hsa_amd_vmem_import_shareable_handle()
hsa_status_t hsa_amd_vmem_import_shareable_handle | ( | int | dmabuf_fd, |
hsa_amd_vmem_alloc_handle_t * | handle | ||
) |
Import a shareable handle.
Import a shareable handle for a memory handle. Importing a shareable handle that has been closed and released results in undefined behavior.
- Parameters
-
[in] dmabuf_fd shareable handle exported with hsa_amd_vmem_export_shareable_handle [out] handle virtual memory handle
- Return values
-
::HSA_STATUS_SUCCESS ::HSA_STATUS_ERROR_INVALID_ALLOCATION Invalid memory handle ::HSA_STATUS_ERROR_OUT_OF_RESOURCES Out of resources ::HSA_STATUS_ERROR Unexpected internal error
◆ hsa_amd_vmem_map()
hsa_status_t hsa_amd_vmem_map | ( | void * | va, |
size_t | size, | ||
size_t | in_offset, | ||
hsa_amd_vmem_alloc_handle_t | memory_handle, | ||
uint64_t | flags | ||
) |
Map a virtual memory handle.
Map a virtual memory handle to a reserved address range. The virtual address requested must be within a previously reserved address range. va
and (va
+ size) must be must be within (va + size) of the previous allocated address range. size
must be equal to size of the memory_handle
hsa_amd_vmem_set_access needs to be called to make the memory accessible to specific agents
- Parameters
-
[in] va virtual address range where memory will be mapped [in] size of memory mapping [in] in_offset offset into memory. Currently unsupported [in] memory_handle virtual memory handle to be mapped [in] flags. Currently unsupported
- Return values
-
::HSA_STATUS_SUCCESS Memory mapped successfully ::HSA_STATUS_ERROR_INVALID_ARGUMENT va, size or memory_handle are invalid ::HSA_STATUS_ERROR_OUT_OF_RESOURCES Insufficient resources ::HSA_STATUS_ERROR Unexpected internal error
◆ hsa_amd_vmem_retain_alloc_handle()
hsa_status_t hsa_amd_vmem_retain_alloc_handle | ( | hsa_amd_vmem_alloc_handle_t * | memory_handle, |
void * | addr | ||
) |
Returns memory handle for mapped memory.
Return a memory handle for previously mapped memory. The handle will be the same value of handle used to map the memory. The returned handle must be released with corresponding number of calls to hsa_amd_vmem_handle_release.
- Parameters
-
[out] memory_handle memory handle for this mapped address [in] mapped address
- Return values
-
::HSA_STATUS_SUCCESS ::HSA_STATUS_ERROR_INVALID_ALLOCATION Invalid address
◆ hsa_amd_vmem_set_access()
hsa_status_t hsa_amd_vmem_set_access | ( | void * | va, |
size_t | size, | ||
const hsa_amd_memory_access_desc_t * | desc, | ||
size_t | desc_cnt | ||
) |
Make a memory mapping accessible.
Make previously mapped virtual address accessible to specific agents. size
must be equal to size of previously mapped virtual memory handle. Calling hsa_amd_vmem_set_access multiple times on the same va
will overwrite previous permissions for all agents
- Parameters
-
[in] va previously mapped virtual address [in] size of memory mapping [in] desc list of access permissions for each agent [in] desc_cnt number of elements in desc
- Return values
-
::HSA_STATUS_SUCCESS ::HSA_STATUS_ERROR_INVALID_ARGUMENT va, size or memory_handle are invalid ::HSA_STATUS_ERROR_INVALID_ALLOCATION memory_handle is invalid ::HSA_STATUS_ERROR_OUT_OF_RESOURCES Insufficient resources ::HSA_STATUS_ERROR_INVALID_AGENT Invalid agent in desc ::HSA_STATUS_ERROR Unexpected internal error
◆ hsa_amd_vmem_unmap()
hsa_status_t hsa_amd_vmem_unmap | ( | void * | va, |
size_t | size | ||
) |
Unmap a virtual memory handle.
Unmap previously mapped virtual address range
- Parameters
-
[in] va virtual address range where memory will be mapped [in] size of memory mapping
- Return values
-
::HSA_STATUS_SUCCESS Memory backing unmapped successfully ::HSA_STATUS_ERROR_INVALID_ALLOCATION memory_handle is invalid ::HSA_STATUS_ERROR_INVALID_ARGUMENT size is invalid ::HSA_STATUS_ERROR Unexpected internal error