Agent Information

Agent Information#

Rocprofiler SDK Developer API: Agent Information

Rocprofiler SDK Developer API 0.4.0

ROCm Profiling API and tools

needs brief description More...

Data Structures
struct	rocprofiler_agent_cache_t
	Cache information for an agent. More...

struct	rocprofiler_agent_io_link_t
	IO link information for an agent. More...

struct	rocprofiler_agent_mem_bank_t
	Memory bank information for an agent. More...

struct	rocprofiler_agent_v0_t
	Stores the properties of an agent (CPU, GPU, etc.) More...

Typedefs
typedef rocprofiler_agent_v0_t	rocprofiler_agent_t

typedef rocprofiler_status_t(*	rocprofiler_query_available_agents_cb_t) (rocprofiler_agent_version_t version, const void *agents, unsigned long num_agents, void user_data)
	Callback function type for querying the available agents.

Enumerations
enum	rocprofiler_agent_version_t { ROCPROFILER_AGENT_INFO_VERSION_NONE = 0 , ROCPROFILER_AGENT_INFO_VERSION_0 = 1 , ROCPROFILER_AGENT_INFO_VERSION_LAST }
	Enumeration ID for version of the rocprofiler_agent_v*_t struct in rocprofiler_i. More...

Functions
rocprofiler_status_t	rocprofiler_query_available_agents (rocprofiler_agent_version_t version, rocprofiler_query_available_agents_cb_t callback, unsigned long agent_size, void *user_data)
	Receive synchronous callback with an array of available agents at moment of invocation.

Detailed Description

needs brief description

Data Structure Documentation

◆ rocprofiler_agent_cache_t

struct rocprofiler_agent_cache_t

Cache information for an agent.

Definition at line 55 of file agent.h.

Collaboration diagram for rocprofiler_agent_cache_t:

Data Fields
uint32_t	association	Cache Associativity.
uint32_t	cache_line_size	Cache line size in bytes.
uint32_t	cache_lines_per_tag	Cache lines per Cache Tag.
uint32_t	latency	Cache latency in ns.
uint32_t	level	Integer representing level: 1, 2, 3, 4, etc.
uint64_t	processor_id_low	Identifies the processor number.
uint64_t	size	Size of the cache.
HsaCacheType	type

◆ rocprofiler_agent_io_link_t

struct rocprofiler_agent_io_link_t

IO link information for an agent.

Definition at line 70 of file agent.h.

Collaboration diagram for rocprofiler_agent_io_link_t:

Data Fields
HSA_LINKPROPERTY	flags	override flags (may be active for specific platforms)
uint32_t	max_bandwidth	maximum interface Bandwidth in MB/s
uint32_t	max_latency	maximum cost of time to transfer (rounded to ns)
uint32_t	min_bandwidth	minimum interface Bandwidth in MB/s
uint32_t	min_latency	minimum cost of time to transfer (rounded to ns)
uint32_t	node_from	See rocprofiler_agent_id_t.
uint32_t	node_to	See rocprofiler_agent_id_t.
uint32_t	recommended_transfer_size	recommended transfer size to reach maximum bandwidth in bytes
HSA_IOLINKTYPE	type	Discoverable IoLink Properties (optional)
uint32_t	version_major	Bus interface version (optional)
uint32_t	version_minor	Bus interface version (optional)
uint32_t	weight	weight factor (derived from CDIT)

◆ rocprofiler_agent_mem_bank_t

struct rocprofiler_agent_mem_bank_t

Memory bank information for an agent.

Definition at line 90 of file agent.h.

Collaboration diagram for rocprofiler_agent_mem_bank_t:

Data Fields
HSA_MEMORYPROPERTY	flags
HSA_HEAPTYPE	heap_type
uint32_t	mem_clk_max	clock for the memory, this allows computing the available bandwidth to the memory when needed
uint64_t	size_in_bytes	physical memory size of the memory range in bytes
uint32_t	width	the number of parallel bits of the memoryinterface

◆ rocprofiler_agent_v0_t

struct rocprofiler_agent_v0_t

Stores the properties of an agent (CPU, GPU, etc.)

The node_id member is the KFD topology node id. It should be considered the "universal" indexing number. It is equivalent to the HSA-runtime HSA_AMD_AGENT_INFO_DRIVER_NODE_ID property of a hsa_agent_t. The const char* fields (name, vendor_name, etc.) are guaranteed to be valid pointers to null-terminated strings during tool finalization. Pointers to the agents via

See also: rocprofiler_query_available_agents are constant and will not be deallocated until after tool finalization. Making copies of the agent struct is also valid.

Definition at line 110 of file agent.h.

Collaboration diagram for rocprofiler_agent_v0_t:

Data Fields
uint32_t	array_count	Number of SIMD arrays.
const rocprofiler_agent_cache_t *	caches
uint32_t	caches_count	of discoverable cache affinity properties on this "H-NUMA" node.
HSA_CAPABILITY	capability	GPU only.
uint32_t	cpu_core_id_base	low value of the logical processor ID of the latency (= CPU) cores available on this node
uint32_t	cpu_cores_count	of latency (= CPU) cores present on this HSA node. This value is 0 for a HSA node with no such cores, e.g a "discrete HSA GPU"
uint32_t	cu_count	Number of compute units.
uint32_t	cu_per_engine	computed
uint32_t	cu_per_simd_array	Number of Compute Units (CU) per SIMD array.
uint16_t	device_id	GPU device id; 0 on latency (= CPU)-only nodes.
uint32_t	domain	PCI domain of the GPU.
uint32_t	drm_render_minor	DRM render device minor device number.
uint32_t	family_id	Family code.
HSA_ENGINE_ID	fw_version	GPU only. Identifier (rev) of the GPU uEngine or Firmware, may be 0.
uint32_t	gds_size_in_kb	Size of Global Data Store in Kilobytes shared across SIMD Wavefronts.
uint32_t	gfx_target_version	major_version=((value / 10000) % 100) minor_version=((value / 100) % 100) patch_version=(value % 100)
uint64_t	gpu_id	GPU only. KFD identifier.
rocprofiler_dim3_t	grid_max_dim	GPU only. Maximum number of work-items of each dimension of a grid.
uint32_t	grid_max_size	GPU only. Maximum number of fbarriers per work-group. Must be at least 32.
uint64_t	hive_id	XGMI Hive the GPU node belongs to in the system. It is an opaque and static number hash created by the PSP.
rocprofiler_agent_id_t	id	Internal opaque identifier.
const rocprofiler_agent_io_link_t *	io_links
uint32_t	io_links_count	of discoverable IO link affinity properties of this node connecting to other nodes.
uint32_t	lds_size_in_kb	Size of Local Data Store in Kilobytes per SIMD Wavefront.
uint64_t	local_mem_size	GPU only. Local memory size.
uint32_t	location_id	GPU BDF (Bus/Device/function number) - identifies the device location in the overall system.
int32_t	logical_node_id	Logical sequence number. This will always be [0..N) where N is the total number of agents.
int32_t	logical_node_type_id	Logical sequence number with respect to other agents of same type. This will always be [0..N) where N is the total number of X agents (where X is a rocprofiler_agent_type_t value). This field is intended to help with environment variable indexing used to mask GPUs at runtime (i.e. HIP_VISIBLE_DEVICES and ROCR_VISIBLE_DEVICES) which start at zero and only apply to GPUs, e.g., logical_node_type_id value for first GPU will be 0, second GPU will have value of 1, etc., regardless of however many agents of a different type preceeded (and thus increased the node_id or logical_node_id). Example: a system with 2 CPUs and 2 GPUs, where the node ids are 0=CPU, 1=GPU, 2=CPU, 3=GPU, then then CPU node_ids 0 and 2 would have logical_node_type_id values of 0 and 1, respectively, and GPU node_ids 1 and 3 would also have logical_node_type_id values of 0 and 1.
uint32_t	max_engine_clk_ccompute	maximum engine clocks for CPU, including any boost capabilities
uint32_t	max_engine_clk_fcompute	GPU only. Maximum engine clocks for GPU, including any boost capabilities.
uint32_t	max_slots_scratch_cu	Number of temp. memory ("scratch") wave slots available to access, may be 0 if HW has no restrictions.
uint32_t	max_waves_per_cu	computed
uint32_t	max_waves_per_simd	This identifies the max. number of launched waves per SIMD. If NumFComputeCores is 0, this value is ignored.
const rocprofiler_agent_mem_bank_t *	mem_banks
uint32_t	mem_banks_count	of discoverable memory bank affinity properties on this "H-NUMA" node.
const char *	model_name	GPU only. Will be something like vega20, mi200, etc.
const char *	name	Name of the agent. Will be identical to product name for CPU.
uint32_t	node_id	Node sequence number. This will be equivalent to the HSA-runtime HSA_AMD_AGENT_INFO_DRIVER_NODE_ID property.
uint32_t	num_cp_queues	number of Compute queues
uint32_t	num_gws	Number of GWS barriers.
uint32_t	num_sdma_engines	number of PCIe optimized SDMA engines
uint32_t	num_sdma_queues_per_engine	number of SDMA queue per one engine
uint32_t	num_sdma_xgmi_engines	number of XGMI optimized SDMA engines
uint32_t	num_shader_banks	Number of Shader Banks or Shader Engines, typical values are 1 or 2.
uint32_t	num_xcc	Number of XCC.
const char *	product_name	Marketing name.
int32_t	reserved_padding0	padding logical_node_id to 64 bytes
HSA_ENGINE_VERSION	sdma_fw_version	GPU only.
uint32_t	simd_arrays_per_engine	Number of SIMD arrays per engine.
uint32_t	simd_count	of HSA throughtput (= GPU) FCompute cores ("SIMD") present in a node. This value is 0 if no FCompute cores are present (e.g. pure "CPU node").
uint32_t	simd_id_base	low value of the logical processor ID of the throughput (= GPU) units available on this node
uint32_t	simd_per_cu	Number of SIMD representing a Compute Unit (CU)
uint64_t	size	set to sizeof(rocprofiler_agent_t) by rocprofiler. This can be used for versioning and compatibility handling
rocprofiler_agent_type_t	type	Enumeration for identifying the agent type (CPU, GPU, etc.)
uint16_t	vendor_id	GPU vendor id; 0 on latency (= CPU)-only nodes.
const char *	vendor_name	Vendor of agent (will be AMD)
uint32_t	wave_front_size	Number of SIMD cores per wavefront executed, typically 64, may be 32 or a different value for some HSA based architectures.
rocprofiler_dim3_t	workgroup_max_dim	GPU only. Maximum number of work-items of each dimension of a work-group.
uint32_t	workgroup_max_size	GPU only. Maximum total number of work-items in a work-group.

Typedef Documentation

◆ rocprofiler_agent_t

typedef rocprofiler_agent_v0_t rocprofiler_agent_t

Definition at line 212 of file agent.h.

◆ rocprofiler_query_available_agents_cb_t

typedef rocprofiler_status_t(* rocprofiler_query_available_agents_cb_t) (rocprofiler_agent_version_t version, const void **agents, unsigned long num_agents, void *user_data)

Callback function type for querying the available agents.

If callback is invoked, returns the rocprofiler_status_t value returned from callback

Parameters

[in]	version	Enum specifying the version of agent info
[in]	agents	Array of pointers to agents
[in]	num_agents	Number of agents in array
[in]	user_data	Data pointer passback

Returns: rocprofiler_status_t

Return values

ROCPROFILER_STATUS_ERROR_INCOMPATIBLE_ABI	size of the agent struct in application is larger than the agent struct for rocprofiler-sdk
ROCPROFILER_STATUS_ERROR_INVALID_ARGUMENT	Invalid rocprofiler_agent_version_t value

Definition at line 228 of file agent.h.

Enumeration Type Documentation

◆ rocprofiler_agent_version_t

enum rocprofiler_agent_version_t

Enumeration ID for version of the rocprofiler_agent_v*_t struct in rocprofiler_i.

Enumerator
ROCPROFILER_AGENT_INFO_VERSION_NONE
ROCPROFILER_AGENT_INFO_VERSION_0
ROCPROFILER_AGENT_INFO_VERSION_LAST

Definition at line 45 of file agent.h.

{
    ROCPROFILER_AGENT_INFO_VERSION_NONE = 0,
    ROCPROFILER_AGENT_INFO_VERSION_0    = 1,
    ROCPROFILER_AGENT_INFO_VERSION_LAST,
} rocprofiler_agent_version_t;

Function Documentation

◆ rocprofiler_query_available_agents()

rocprofiler_status_t rocprofiler_query_available_agents	(	rocprofiler_agent_version_t	version,
		rocprofiler_query_available_agents_cb_t	callback,
		unsigned long	agent_size,
		void *	user_data
	)

Receive synchronous callback with an array of available agents at moment of invocation.

Parameters

[in]	version	Enum value specifying the struct type of the agent info
[in]	callback	Callback function accepting list of agents
[in]	agent_size	Should be set to sizeof(rocprofiler_agent_t)
[in]	user_data	Data pointer provided to callback

Returns: rocprofiler_status_t

Agent Information

Agent Information#

Data Structures

Typedefs

Enumerations

Functions