Gbaraldi/threadtrace2 (#724)

* Added first ATT API

* Finalizing thread trace API

* Fixing more rebase conflicts

* Added codeobj disassembly sample

* Fixing merge issues with rebase [2]

* Adding ATT packets

* Implemented thread trace intercept

* Moved codeobj parser to same repo as rocprofiler

* Moved thread trace to new API

* Fixing merge conflicts

* Fixing more merge conflicts

* Adding thread trace packet reuse

* Merged aql_profile_v2 headers

* Linked ATT sample to aqlprofile

* Updated decoder to include non-loaded codeobjs

* Implemented ISA decoder into ATT sample

* Added marker_id to vaddr

* Updating aql_profile_v2 API to memcpy

* Updating thread trace API to include 64bit markers. Using the result of ISA matching.

* Added instruction type and cycles summary

* Updated sample with selection of kernel by kernel_object

* Added option to copy from memory kernels

* Moved tool_data in thread_trace to dynamic alloc

* Restoring hsa.cpp

* Fixed ATT sample crash. General improvements.

* Moved codeobj library to outside src/

* Updated license header

* Moved codeobj_capture to camelcase

* Solving some more merge conflicts

* Update samples/advanced_thread_trace/CMakeLists.txt

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* Update samples/advanced_thread_trace/CMakeLists.txt

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* Update samples/code_object_isa_decode/CMakeLists.txt

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* Update source/lib/rocprofiler-sdk/thread_trace/CMakeLists.txt

* Removing unused parameter check

* Adding const to isEmpty

* Removing unused warning

* Adding libdw-dev to requirements

* Running clang-format

* Commenting out new aql calls

* Clang format

* Unused variable fix

* Adding codeobj-decoder coverage

* Commenting out threadtrace

* Update samples/CMakeLists.txt

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* P

* WOverloaded

* Addressing clang-tidy

* Virtual destructor on ttracer class

* Corr id

* Fixing code source format

* Update CMakeLists.txt

* Build fixes

* Update source/lib/rocprofiler-sdk-codeobj/code_object_track.cpp

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* Fix shadowing

* Update CMakeLists.txt

* Update samples/CMakeLists.txt

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

---------

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Ammar ELWazir <ammar.elwazir@amd.com>
Co-authored-by: Ammar ELWazir <aelwazir@amd.com>
Co-authored-by: Benjamin Welton <bewelton@amd.com>
This commit is contained in:
Giovanni Lenzi Baraldi
2024-04-08 16:43:02 -03:00
committed by GitHub
parent d2bad14bf6
commit 69b8a43dc6
48 changed files with 4649 additions and 109 deletions
+61 -4
View File
@@ -21,14 +21,22 @@
// THE SOFTWARE.
#include "lib/rocprofiler-sdk/hsa/aql_packet.hpp"
#include <cstdlib>
#include <iostream>
#include <string>
#define CHECK_HSA(fn, message) \
if((fn) != HSA_STATUS_SUCCESS) \
{ \
std::cerr << __FILE__ << ':' << __LINE__ << ' ' << message; \
exit(1); \
}
namespace rocprofiler
{
namespace hsa
{
AQLPacket::~AQLPacket()
CounterAQLPacket::~CounterAQLPacket()
{
if(!profile.command_buffer.ptr)
{
@@ -36,7 +44,7 @@ AQLPacket::~AQLPacket()
}
else if(!command_buf_mallocd)
{
free_func(profile.command_buffer.ptr);
CHECK_HSA(free_func(profile.command_buffer.ptr), "freeing memory");
}
else
{
@@ -49,12 +57,61 @@ AQLPacket::~AQLPacket()
}
else if(!output_buffer_malloced)
{
free_func(profile.output_buffer.ptr);
CHECK_HSA(free_func(profile.output_buffer.ptr), "freeing memory");
}
else
{
::free(profile.output_buffer.ptr);
}
}
TraceAQLPacket::~TraceAQLPacket() = default;
/*
TraceAQLPacket::~TraceAQLPacket()
{
aqlprofile_att_delete_packets(this->handle);
}
*/
hsa_status_t
TraceAQLPacket::Alloc(void** ptr, size_t size, aqlprofile_buffer_desc_flags_t flags, void* data)
{
if(!data) return HSA_STATUS_ERROR;
if(!reinterpret_cast<TraceAQLPacket*>(data)->tracepool) return HSA_STATUS_ERROR;
auto& pool = *reinterpret_cast<TraceAQLPacket*>(data)->tracepool;
if(!pool.allocate_fn || !pool.free_fn || !pool.allow_access_fn) return HSA_STATUS_ERROR;
if(flags.host_access)
{
hsa_status_t status = pool.allocate_fn(pool.cpu_pool_, size, 0, ptr);
if(!flags.device_access || status != HSA_STATUS_SUCCESS) return status;
return pool.allow_access_fn(1, &pool.gpu_agent, nullptr, *ptr);
}
return pool.allocate_fn(pool.gpu_pool_, size, 0, ptr);
}
void
TraceAQLPacket::Free(void* ptr, void* data)
{
auto* pool = reinterpret_cast<TraceAQLPacket*>(data)->tracepool.get();
if(!pool || !pool->free_fn) return;
pool->free_fn(ptr);
}
hsa_status_t
TraceAQLPacket::Copy(void* dst, const void* src, size_t size, void* data)
{
auto* pool = reinterpret_cast<TraceAQLPacket*>(data)->tracepool.get();
if(!pool || !pool->api_copy_fn) return HSA_STATUS_ERROR;
return pool->api_copy_fn(dst, src, size);
}
TraceAQLPacket::TraceAQLPacket(std::shared_ptr<TraceMemoryPool>& _tracepool)
: tracepool(_tracepool){};
} // namespace hsa
} // namespace rocprofiler
+70 -16
View File
@@ -23,6 +23,7 @@
#pragma once
#include "lib/common/container/small_vector.hpp"
#include "lib/rocprofiler-sdk/aql/aql_profile_v2.h"
#include <hsa/hsa_ext_amd.h>
#include <hsa/hsa_ven_amd_aqlprofile.h>
@@ -30,6 +31,12 @@
namespace rocprofiler
{
namespace aql
{
class CounterPacketConstruct;
class ThreadTraceAQLPacketFactory;
} // namespace aql
namespace hsa
{
constexpr hsa_ext_amd_aql_pm4_packet_t null_amd_aql_pm4_packet = {
@@ -41,12 +48,11 @@ constexpr hsa_ext_amd_aql_pm4_packet_t null_amd_aql_pm4_packet = {
* Struct containing AQL packet information. Including start/stop/read
* packets along with allocated buffers
*/
struct AQLPacket
class AQLPacket
{
using memory_pool_free_func_t = decltype(::hsa_amd_memory_pool_free)*;
AQLPacket(memory_pool_free_func_t func);
~AQLPacket();
public:
AQLPacket() = default;
virtual ~AQLPacket() = default;
// Keep move constuctors (i.e. std::move())
AQLPacket(AQLPacket&& other) = default;
@@ -56,20 +62,68 @@ struct AQLPacket
AQLPacket(const AQLPacket&) = delete;
AQLPacket& operator=(const AQLPacket&) = delete;
bool command_buf_mallocd = false;
bool output_buffer_malloced = false;
bool empty = {true};
hsa_ven_amd_aqlprofile_profile_t profile = {};
hsa_ext_amd_aql_pm4_packet_t start = null_amd_aql_pm4_packet;
hsa_ext_amd_aql_pm4_packet_t stop = null_amd_aql_pm4_packet;
hsa_ext_amd_aql_pm4_packet_t read = null_amd_aql_pm4_packet;
memory_pool_free_func_t free_func = nullptr;
hsa_ven_amd_aqlprofile_profile_t profile = {};
hsa_ext_amd_aql_pm4_packet_t start = null_amd_aql_pm4_packet;
hsa_ext_amd_aql_pm4_packet_t stop = null_amd_aql_pm4_packet;
hsa_ext_amd_aql_pm4_packet_t read = null_amd_aql_pm4_packet;
common::container::small_vector<hsa_ext_amd_aql_pm4_packet_t, 3> before_krn_pkt = {};
common::container::small_vector<hsa_ext_amd_aql_pm4_packet_t, 2> after_krn_pkt = {};
bool isEmpty() const { return empty; }
bool empty = true;
};
class CounterAQLPacket : public AQLPacket
{
friend class rocprofiler::aql::CounterPacketConstruct;
using memory_pool_free_func_t = decltype(::hsa_amd_memory_pool_free)*;
public:
CounterAQLPacket(memory_pool_free_func_t func)
: free_func{func} {};
~CounterAQLPacket() override;
protected:
bool command_buf_mallocd = false;
bool output_buffer_malloced = false;
memory_pool_free_func_t free_func = nullptr;
};
struct TraceMemoryPool
{
hsa_agent_t gpu_agent;
hsa_amd_memory_pool_t cpu_pool_;
hsa_amd_memory_pool_t gpu_pool_;
decltype(hsa_amd_memory_pool_allocate)* allocate_fn;
decltype(hsa_amd_agents_allow_access)* allow_access_fn;
decltype(hsa_amd_memory_pool_free)* free_fn;
decltype(hsa_memory_copy)* api_copy_fn;
};
class TraceAQLPacket : public AQLPacket
{
friend class rocprofiler::aql::ThreadTraceAQLPacketFactory;
public:
TraceAQLPacket(std::shared_ptr<TraceMemoryPool>& _tracepool);
TraceMemoryPool& GetPool() const { return *tracepool; }
aqlprofile_handle_t GetHandle() const { return handle; }
uint64_t GetAgent() const { return tracepool->gpu_agent.handle; }
~TraceAQLPacket() override;
protected:
std::shared_ptr<TraceMemoryPool> tracepool;
aqlprofile_att_control_aql_packets_t packets;
aqlprofile_handle_t handle;
static hsa_status_t Alloc(void** ptr,
size_t size,
aqlprofile_buffer_desc_flags_t flags,
void* data);
static void Free(void* ptr, void* data);
static hsa_status_t Copy(void* dst, const void* src, size_t size, void* data);
};
inline AQLPacket::AQLPacket(memory_pool_free_func_t func)
: free_func{func}
{}
} // namespace hsa
} // namespace rocprofiler
@@ -672,10 +672,10 @@ code_object_load_callback(hsa_executable_t executable,
if(_status != HSA_STATUS_SUCCESS) return _status; \
}
auto& loader_table = get_loader_table();
auto code_obj_v = code_object{};
auto& data = code_obj_v.rocp_data;
int _storage_type = ROCPROFILER_CODE_OBJECT_STORAGE_TYPE_NONE;
auto& loader_table = get_loader_table();
auto code_obj_v = code_object{};
auto& data = code_obj_v.rocp_data;
uint32_t _storage_type = ROCPROFILER_CODE_OBJECT_STORAGE_TYPE_NONE;
LOG_IF(FATAL, data.size == 0) << "code object did not properly initialized the size field upon "
"construction (this is likely a compiler bug)";
-2
View File
@@ -342,8 +342,6 @@ WriteInterceptor(const void* packets,
{
hsa_barrier_and_packet_t barrier{};
barrier.header = HSA_PACKET_TYPE_BARRIER_AND << HSA_PACKET_HEADER_TYPE;
// barrier.header |= HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE;
// barrier.header |= HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE;
barrier.header |= 1 << HSA_PACKET_HEADER_BARRIER;
barrier.completion_signal = original_packet.completion_signal;
transformed_packets.emplace_back(barrier);
@@ -90,6 +90,9 @@ constexpr rocprofiler_agent_t default_agent =
void
QueueController::add_queue(hsa_queue_t* id, std::unique_ptr<Queue> queue)
{
for(auto& pre_initialize_fn : pre_initialize)
pre_initialize_fn(queue->get_agent(), get_core_table(), get_ext_table());
CHECK(queue);
_callback_cache.wlock([&](auto& callbacks) {
_queues.wlock([&](auto& map) {
@@ -111,6 +114,11 @@ void
QueueController::destroy_queue(hsa_queue_t* id)
{
if(!id) return;
_queues.wlock([&](auto& map) {
for(auto& deinitialize_fn : pre_deinitialize)
if(map.find(id) != map.end())
deinitialize_fn(map.at(id)->get_agent(), get_core_table(), get_ext_table());
});
const auto* queue = get_queue(*id);
@@ -195,7 +203,8 @@ QueueController::init(CoreApiTable& core_table, AmdExtTable& ext_table)
{
constexpr auto expected_context_size = 160UL;
static_assert(
sizeof(context::context) == expected_context_size,
sizeof(context::context) ==
expected_context_size + sizeof(std::shared_ptr<rocprofiler::ThreadTracer>),
"If you added a new field to context struct, make sure there is a check here if it "
"requires queue interception. Once you have done so, increment expected_context_size");
@@ -212,6 +221,20 @@ QueueController::init(CoreApiTable& core_table, AmdExtTable& ext_table)
break;
}
}
else if(itr->thread_trace)
{
enable_intercepter = true;
std::weak_ptr<rocprofiler::ThreadTracer> trace = itr->thread_trace;
pre_initialize.emplace_back(
[trace](const AgentCache& cache, const CoreApiTable& core, const AmdExtTable& ext) {
if(auto locked = trace.lock()) locked->resource_init(cache, core, ext);
});
pre_deinitialize.emplace_back(
[trace](const AgentCache& cache, const CoreApiTable&, const AmdExtTable&) {
if(auto locked = trace.lock()) locked->resource_deinit(cache);
});
break;
}
}
if(enable_intercepter)
@@ -103,6 +103,7 @@ public:
private:
using client_id_map_t = std::unordered_map<ClientID, agent_callback_tuple_t>;
using agent_cache_map_t = std::unordered_map<uint32_t, AgentCache>;
using resource_alloc_t = void(const AgentCache&, const CoreApiTable&, const AmdExtTable&);
CoreApiTable _core_table = {};
AmdExtTable _ext_table = {};
@@ -110,6 +111,9 @@ private:
common::Synchronized<client_id_map_t> _callback_cache = {};
agent_cache_map_t _supported_agents = {};
common::Synchronized<hsa::profiler_serializer> _profiler_serializer;
std::vector<std::function<resource_alloc_t>> pre_initialize;
std::vector<std::function<resource_alloc_t>> pre_deinitialize;
};
QueueController*