Updated rocprofiler.h for v2 (#18)
* Update and rename rocprofiler.h to rocprofiler.h.in
- Removing Service IDs
- Fixing agent_id to be agent
* [0/N] New rocprofiler headers
- created rocprofiler/defines.h
- ppdef macros
- created rocprofiler/hip.h
- HIP specific types
- created rocprofiler/hsa.h
- HSA specific types
- created rocprofiler/marker.h
- Marker (ROCTx) specific types
- create version.h.in
- file containing version info
- updated source/lib/rocprofiler/CMakeLists.txt
- set DEFINE_SYMBOL
- compile defs provided by rocprofiler::rocprofiler-headers
* [1/N] Update rocprofiler.h
- pragma once
- removed some ppdefs (in version.h.in and defines.h)
- extern "C" after includes
- added *_NONE and *_LAST enum values to all enums
- provided some rocprofiler_status_t enums
- tweaked rocprofiler_agent_type_t enum fields
- tweaked rocprofiler_agent_info_t enum fields
- provided rocprofiler_tracer_activity_domain_t
- added missing rocprofiler_counter_instance_id_t typedef
- may not be correct
- provided rocprofiler_record_header_t struct
- provided rocprofiler_record_tracer_t struct
- add ROCPROFILER_NONNULL attribute where appropriate
- CMakeLists.txt: add subdirectories for hsa, hip, and marker
- defines.h: remove ROCPROFILER_CALL ppdef
- rocprofiler.h
- ROCPROFILER_STATUS_ERROR_NOT_IMPLEMENTED
- extend rocprofiler_agent_t
- modify rocprofiler_query_available_agents signature to callback
- rocprofiler_pc_sampling_config_array_t
- update rocprofiler_buffer_callback_t to include context id
- update rocprofiler_create_buffer to accept context
- rocprofiler_plugin.h
- non-const rocprofiler_record_header_t**
* [2/N] Update include/rocprofiler/rocprofiler_plugin.h
- change prototype of rocprofiler_plugin_write_buffer_records to resemble rocprofiler_buffer_callback_t
* [3/N] Update include/rocprofiler/hsa
- Update hsa.h
- Details in hsa subfolder
* [4/N] Update include/rocprofiler/hip
- Update hip.h
- Details in hip subfolder
* [5/N] Update include/rocprofiler/marker
- Update marker.h
- Details in marker subfolder
* [6/N] Update samples/pc_sampling
- fix issues with macros
- fix API changes
---------
Co-authored-by: Jonathan Madsen <jrmadsen@users.noreply.github.com>
[ROCm/rocprofiler-sdk commit: 39b209c2a7]
Tento commit je obsažen v:
@@ -0,0 +1,124 @@
|
||||
#ifndef PC_SAMPLING_COMMON_H
|
||||
#define PC_SAMPLING_COMMON_H
|
||||
|
||||
#include <rocprofiler/rocprofiler.h>
|
||||
|
||||
#include <cstdio>
|
||||
#include <cstdlib>
|
||||
#include <cstddef>
|
||||
#include <cstdint>
|
||||
|
||||
constexpr size_t BUFFER_SIZE_BYTES = 4096;
|
||||
constexpr size_t WATERMARK = (BUFFER_SIZE_BYTES / 2);
|
||||
|
||||
#define ROCPROFILER_CALL(result, msg) \
|
||||
{ \
|
||||
rocprofiler_status_t status = result; \
|
||||
if(status != 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 status = result; \
|
||||
if(status == ROCPROFILER_STATUS_SUCCESS) \
|
||||
{ \
|
||||
puts(#result " succeeded"); \
|
||||
} \
|
||||
}
|
||||
|
||||
static rocprofiler_status_t
|
||||
find_first_gpu_agent_impl(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 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;
|
||||
|
||||
ROCPROFILER_CALL(rocprofiler_query_available_agents(&find_first_gpu_agent_impl,
|
||||
sizeof(rocprofiler_agent_t),
|
||||
static_cast<void*>(&gpu_agent)),
|
||||
"Failed to find GPU agents");
|
||||
|
||||
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->kind == 0)
|
||||
{
|
||||
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
|
||||
Odkázat v novém úkolu
Zablokovat Uživatele