a798a26227
* Agent information w/o using hsa-runtime
- remove lib/rocprofiler/hsa/agent.{hpp,cpp}
- update include/rocprofiler/agent.h
- basically all possible info from /sys/class/kfd/kfd/topology/nodes/*
* Print topology in rocprofiler_lib.agent test
- hack to help diagnose errors
* Update lib/rocprofiler/tests/details/agent.cpp
- use LOG_IF(WARNING, ...) instead of LOG_IF(FATAL, ...)
* Update lib/rocprofiler/tests/agent.cpp
- print rocminfo at beginning of test
- fix comparison of agent handle
- misc other checks
* Updte lib/rocprofiler/agent.cpp
- handle unreadable /sys/class/kfd/kfd/topology/nodes/<N>/properties file
* Update lib/tests/buffering/CMakeLists.txt
- increase timeout to 120
- buffering.parallel will timeout when thread sanitizing is enabled
* Update cmake: rocprofiler-drm
- find drm headers and libraries
* Update include/rocprofiler/agent.h
- add family_id field
* Update lib/rocprofiler/agent.cpp
- parse /proc/cpuinfo for name, family, apicid, etc.
- read_topology uses unique pointers to cleanup memory allocations
- implement name and gfxip
* Update lib/rocprofiler/tests/agent.cpp
- improved failure message
- check name/gfxip
- remove check against hsa_agent_t.handle
- this value is dependent on the address of C++ class
* Update lib/rocprofiler/tests/details/agent.cpp
- tweak gfxip_ variable which is broken for CPU
* Update lib/rocprofiler/agent.cpp
- update string handling for name and gfxip
* Update lib/rocprofiler/tests/agent.cpp
- minor output tweak
* Update lib/rocprofiler/registration.{hpp,cpp}
- registration::init_logging() function
* Update lib/rocprofiler/agent.cpp
- fix hex handling of GFX step version
* Update lib/rocprofiler/tests/details/agent.cpp
- fix format string when nearest CPUs not found
* Update lib/rocprofiler/tests/CMakeLists.txt
- exclude details/agent.cpp from being parsed for gtest tests
* Update include/rocprofiler/fwd.h
- add ROCPROFILER_STATUS_ERROR_INCOMPATIBLE_ABI status
* Update lib/rocprofiler/tests/details/agent.{hpp,cpp}
- replace with slightly modified implementation of rocminfo
- primary change was not printing
* Update lib/rocprofiler/tests/agent.cpp
- update test to use rocminfo data
* Update lib/rocprofiler/agent.cpp
- add pc_sampling_configs
- return error on incompatible ABI
* Update counters and counters tests
- rename test names for consistency
- fixed incorrect spelling of derived
* Add lib/rocprofiler/tests/{timestamp,version}.cpp
- add timestamp and version tests for rocprofiler_get_timestamp and rocprofiler_get_version, respectively
* Update lib/rocprofiler/tests/agent
- fix double free of name_str from isa_info_t
* Update include/rocprofiler/agent.h
- comments for rocprofiler_agent_mem_bank_t
- add rocprofiler_dim3_t
- comments for rocprofiler_agent_t
- add new fields to rocprofiler_agent_t
- cu_count
- workgroup_max_size
- workgroup_max_dim
- grid_max_size
- grid_max_dim
- vendor_name
- product_name
- change prototype of rocprofiler_available_agents_cb_t to be const agent**
* Update lib/rocprofiler/agent.cpp
- set size field
- implement:
- product_name
- vendor_name
- workgroup_max_size
- workgroup_max_dim
- grid_max_size
- grid_max_dim
- cu_count
* Update lib/rocprofiler/tests/agent.cpp
- changes for const agent*
* Update samples/pc_sampling
- updates for const agent*
* Update lib/rocprofiler/agent.cpp
- fix ABI compatibility check
- return incompatible if tool agent is larger than our agent
* Update include/rocprofiler/agent.h
- doxygen comments
- make size field of rocprofiler_agent_t uint64_t for consistency
- add gpu_id via /sys/class/kfd/kfd/.../<idx>/gpu_id
- add model_name via /sys/class/kfd/kfd/.../<idx>/name
* Update lib/rocprofiler/agent.cpp
- add read_file function (vector of strings)
- support enum in read_property
- assign model_name and gpu_id fields
- remove unique_id
* Update lib/rocprofiler/tests/details/agent.*
- support family id, ucode_version, sdma_ucode_version
* Update lib/rocprofiler/tests/agent.cpp
- Add rocprofiler_lib.agent_abi test
- Verify family_id, ucode_version, sdma_ucode_version
129 línte
5.1 KiB
C++
129 línte
5.1 KiB
C++
#ifndef PC_SAMPLING_COMMON_H
|
|
#define PC_SAMPLING_COMMON_H
|
|
|
|
#include <rocprofiler/rocprofiler.h>
|
|
|
|
#include <cstddef>
|
|
#include <cstdint>
|
|
#include <cstdio>
|
|
#include <cstdlib>
|
|
#include <optional>
|
|
#include <string>
|
|
#include <string_view>
|
|
|
|
constexpr size_t BUFFER_SIZE_BYTES = 4096;
|
|
constexpr size_t WATERMARK = (BUFFER_SIZE_BYTES / 2);
|
|
const std::string_view MI200_NAME = "gfx90a";
|
|
|
|
#define ROCPROFILER_CALL(result, msg) \
|
|
{ \
|
|
rocprofiler_status_t CHECKSTATUS = result; \
|
|
if(CHECKSTATUS != ROCPROFILER_STATUS_SUCCESS) \
|
|
{ \
|
|
puts(#result " failed"); \
|
|
} \
|
|
}
|
|
|
|
// We might want to test the calls that fails
|
|
// e.g. calling `rocprofiler_configure_pc_sampling_service `
|
|
// after previous initialization.
|
|
#define ROCPROFILER_CALL_FAILS(result, msg) \
|
|
{ \
|
|
rocprofiler_status_t CHECKSTATUS = result; \
|
|
if(CHECKSTATUS == ROCPROFILER_STATUS_SUCCESS) \
|
|
{ \
|
|
puts(#result " succeeded"); \
|
|
} \
|
|
}
|
|
|
|
static rocprofiler_status_t
|
|
find_first_gpu_agent_impl(const rocprofiler_agent_t** agents, size_t num_agents, void* data)
|
|
{
|
|
// data is required
|
|
if(!data) return ROCPROFILER_STATUS_ERROR;
|
|
|
|
auto* _out_agent = static_cast<rocprofiler_agent_t*>(data);
|
|
// find the first GPU agent
|
|
for(size_t i = 0; i < num_agents; i++)
|
|
{
|
|
if(agents[i]->type == ROCPROFILER_AGENT_TYPE_GPU)
|
|
{
|
|
*_out_agent = *agents[i];
|
|
printf("[%s] %s :: id=%zu, type=%i, num pc sample configs=%zu\n",
|
|
__FUNCTION__,
|
|
_out_agent->name,
|
|
_out_agent->id.handle,
|
|
_out_agent->type,
|
|
_out_agent->pc_sampling_configs.size);
|
|
return ROCPROFILER_STATUS_SUCCESS;
|
|
}
|
|
else
|
|
{
|
|
printf("[%s] %s :: id=%zu, type=%i, num pc sample configs=%zu\n",
|
|
__FUNCTION__,
|
|
agents[i]->name,
|
|
agents[i]->id.handle,
|
|
agents[i]->type,
|
|
agents[i]->pc_sampling_configs.size);
|
|
}
|
|
}
|
|
return ROCPROFILER_STATUS_ERROR;
|
|
}
|
|
|
|
static std::optional<rocprofiler_agent_t>
|
|
find_first_gpu_agent()
|
|
{
|
|
// This function returns the first gpu agent it encounters.
|
|
// TODO: write the better function querying information about the agent,
|
|
// and return if the agent is MI200.
|
|
rocprofiler_agent_t gpu_agent;
|
|
|
|
auto status = rocprofiler_query_available_agents(
|
|
&find_first_gpu_agent_impl, sizeof(rocprofiler_agent_t), static_cast<void*>(&gpu_agent));
|
|
|
|
if(status != ROCPROFILER_STATUS_SUCCESS) return std::nullopt;
|
|
|
|
return gpu_agent;
|
|
}
|
|
|
|
static void
|
|
rocprofiler_pc_sampling_callback(rocprofiler_context_id_t /*context_id*/,
|
|
rocprofiler_buffer_id_t /*buffer_id*/,
|
|
rocprofiler_record_header_t** headers,
|
|
size_t num_headers,
|
|
void* /*data*/,
|
|
uint64_t drop_count)
|
|
{
|
|
// Vladimir: I am not sure if this is the right way of iterating over PC sampling records.
|
|
printf(
|
|
"The number of delivered samples is: %zu, while the number of dropped samples is: %lu.\n",
|
|
num_headers,
|
|
drop_count);
|
|
|
|
for(size_t i = 0; i < num_headers; i++)
|
|
{
|
|
auto* cur_header = headers[i];
|
|
if(cur_header->category == ROCPROFILER_BUFFER_CATEGORY_PC_SAMPLING)
|
|
{
|
|
auto* pc_sample = static_cast<rocprofiler_pc_sampling_record_t*>(cur_header->payload);
|
|
printf("--- pc: %lx, dispatch_id: %lx, timestamp: %lu, hardware_id: %lu\n",
|
|
pc_sample->pc,
|
|
pc_sample->dispatch_id,
|
|
pc_sample->timestamp,
|
|
pc_sample->hardware_id);
|
|
// Vladimir: How to parse the remaining part of the `rocprofiler_pc_sampling_record_t`
|
|
// struct?
|
|
}
|
|
}
|
|
// Vladimr: We might want to add somewhere in the documentation that headars actually contain PC
|
|
// samples.
|
|
}
|
|
|
|
static void
|
|
run_HIP_app()
|
|
{
|
|
// TODO: provide the simple HIP app
|
|
}
|
|
|
|
#endif
|