Separate agent cache from queue controller (#145)

* Update lib/rocprofiler/agent.{hpp,cpp}

- get_agents() function for internal access to agent pointers

* Update AgentCache

- make member variables and member functions distinguish b/t hsa agent and rocprofiler agent clear

* Change ctor of AgentCache

* Update lib/rocprofiler/hsa/queue_controller.cpp

- QueueController::init uses agent::get_agent_cache

* Update lib/rocprofiler/hsa/agent_cache.*

- member function to get index
- operator== for rocprofiler_agent_t and hsa_agent_t
- removed hsa_iterate_agents from ctor (now in agent.cpp)

* Update lib/rocprofiler/agent.*

- construct_agent_cache function
- functions for rocprofiler agent <-> HSA agent
- functions for getting agent cache

* Update lib/rocprofiler/registration.cpp

- invoke construct_agent_cache when HSA table is receieved

* Update lib/rocprofiler/agent.cpp

- loosen failure conditions
- handle spurious duplicate entry warning

* Update lib/rocprofiler/agent.cpp

- improve read_map diagnostics

* Update lib/rocprofiler/agent.cpp

- avoid infinite loop in read_map

* Update lib/rocprofiler/agent.cpp

- handle empty kfd node properties file

* Update lib/rocprofiler/agent.cpp

- check for permissions to read a node properties file

* Update lib/rocprofiler/agent.cpp

- more checks on file readability

* Update lib/rocprofiler/tests/agent.cpp

- print virtual kfd topology

* Update lib/rocprofiler/tests/agent.cpp

- verify id.handle == hsa_agent internal node id

* Update lib/rocprofiler/tests/agent.cpp

- check node_id
- check location id
- check device id
- update abi test

* Update include/rocprofiler/agent.h

- add node_id field
- add reserved0 field to ensure new field increases struct size

* Update lib/rocprofiler/agent.cpp

- node_id instead of id.handle

* Update lib/rocprofiler/agent_cache.cpp

- node_id instead of id.handle

* Update samples/pc_sampling

- node_id for agent instead of id.handle

* Update lib/rocprofiler/buffer.cpp

- remove debug prints

[ROCm/rocprofiler-sdk commit: 7f631de401]
Tento commit je obsažen v:
Jonathan R. Madsen
2023-10-19 19:04:02 -05:00
odevzdal GitHub
rodič 6c26870c8c
revize 178bb0c300
16 změnil soubory, kde provedl 434 přidání a 167 odebrání
+4
Zobrazit soubor
@@ -21,6 +21,7 @@
// SOFTWARE.
#include "lib/rocprofiler/registration.hpp"
#include "lib/rocprofiler/agent.hpp"
#include "lib/rocprofiler/context/context.hpp"
#include "lib/rocprofiler/hsa/hsa.hpp"
#include "lib/rocprofiler/hsa/queue.hpp"
@@ -541,6 +542,9 @@ rocprofiler_set_api_table(const char* name,
<< " rocprofiler expected HSA library to pass 1 API table, not " << num_tables;
auto* hsa_api_table = static_cast<HsaApiTable*>(*tables);
// need to construct agent mappings before initializing the queue controller
rocprofiler::agent::construct_agent_cache(hsa_api_table);
rocprofiler::hsa::queue_controller_init(hsa_api_table);
// any internal modifications to the HsaApiTable need to be done before we make the