Gbaraldi/att tool (#766)

* Enabling codeobj and thread trace samples

* Updating aqlprofile_v2 header

* Codeobj and thread trace samples with output log files

* Fixing clang format

* Cmake formatting

* Adding coverage to codeobj

* Comment trace sample

* Adding ATT Parser API

* Fixing forwarding to aqlprofile

* Clang formatting

* Clang tidy

* Adding option to print memory kernels

* Clang format

* Remove default from switch case

* Separating  client/main on codeobj sample for ASAn

* Formatting

* Gbaraldi/att tool rebase (#801)

* Enabling codeobj and thread trace samples

* Updating aqlprofile_v2 header

* Codeobj and thread trace samples with output log files

* Fixing clang format

* Cmake formatting

* Adding coverage to codeobj

* Comment trace sample

* Removing python from workflow

* Adding ATT Parser API

* Fixing forwarding to aqlprofile

* Clang formatting

* Clang tidy

* Adding option to print memory kernels

* Clang format

* Remove default from switch case

* Separating  client/main on codeobj sample for ASAn

* Formatting

* Enabling codeobj and thread trace samples

* Updating aqlprofile_v2 header

* Codeobj and thread trace samples with output log files

* Fixing clang format

* Cmake formatting

* Adding coverage to codeobj

* Comment trace sample

* Adding ATT Parser API

* Fixing forwarding to aqlprofile

* Clang formatting

* Clang tidy

* Adding option to print memory kernels

* Clang format

* Remove default from switch case

* Separating  client/main on codeobj sample for ASAn

* Formatting

* Fix codeobj library

* Allow thread trace in parallel with other service

* Zeroing the HSA signals

* Adding exception wrappers in ATT sample

* Removed force configure

* Remove force configure from ISA decode

* Removing codecov flag

* Gbaraldi/att tool tests (#828)

* Adding tests for codeobj ISA decode

* Adding ATT tests

* Adding ATT integration tests

* Formatting

* Changing codeobj binary extension

* Renaming codeobj library spaces

* Fixing samples

* Formatting

* Formatting

* Fixing int test

* Fixing linker error

* Fixing memory fault

* Moving kernel ot inside namespace

* ASAN linking fix

* Removing unecessary headers

* Formatting

* Fixing target_cu

* Remove codeobj binary

* Revert "Remove codeobj binary"

This reverts commit 7d286f89d8096bc36925cd79cd742a5e6d10d179.

* Enable memory snapshot

* adding comgr

---------

Co-authored-by: Ammar ELWazir <ammar.elwazir@amd.com>
Dieser Commit ist enthalten in:
Giovanni Lenzi Baraldi
2024-05-03 18:45:47 -03:00
committet von GitHub
Ursprung 6d3fbcffad
Commit 099ac7c72d
38 geänderte Dateien mit 1829 neuen und 753 gelöschten Zeilen
+129 -150
Datei anzeigen
@@ -36,12 +36,11 @@
#include <rocprofiler-sdk/fwd.h>
#include <rocprofiler-sdk/registration.h>
#include <rocprofiler-sdk/rocprofiler.h>
#include "lib/rocprofiler-sdk/aql/aql_profile_v2.h"
#include <rocprofiler-sdk-codeobj/code_printing.hpp>
#include <shared_mutex>
#include "code_object_track.hpp"
#include "common/defines.hpp"
#include "common/filesystem.hpp"
#include "lib/rocprofiler-sdk-codeobj/code_printing.hpp"
#include <cxxabi.h>
#include <atomic>
@@ -64,19 +63,28 @@
#include <unordered_map>
#include <vector>
#define TARGET_CU 1
#define SIMD_SELECT 0x3
#define BUFFER_SIZE 0x6000000
#define NUM_SE 2
constexpr bool COPY_MEMORY_CODEOBJ = false;
#define OUTPUT_OFSTREAM "advanced_thread_trace.log"
#define TARGET_CU 1
#define SIMD_SELECT 0x3
#define BUFFER_SIZE 0x6000000
#define SE_MASK 0x11
constexpr bool COPY_MEMORY_CODEOBJ = true;
template <>
struct std::hash<pcinfo_t>
{
uint64_t operator()(const pcinfo_t& info) const
#define C_API_BEGIN \
try \
{
return info.addr ^ (info.marker_id << 32ul) ^ (info.marker_id >> 32ul);
}
#define C_API_END \
} \
catch(std::exception & e) \
{ \
std::cerr << "Error in " << __FILE__ << ':' << __LINE__ << ' ' << e.what() << std::endl; \
} \
catch(...) { std::cerr << "Error in " << __FILE__ << ':' << __LINE__ << std::endl; }
struct pcinfo_t
{
uint64_t marker_id;
uint64_t addr;
};
bool
@@ -97,6 +105,10 @@ namespace client
using code_obj_load_data_t = rocprofiler_callback_tracing_code_object_load_data_t;
using kernel_symbol_data_t = rocprofiler_callback_tracing_code_object_kernel_symbol_register_data_t;
using Instruction = rocprofiler::codeobj::disassembly::Instruction;
using CodeobjAddressTranslate = rocprofiler::codeobj::disassembly::CodeobjAddressTranslate;
using SymbolInfo = rocprofiler::codeobj::disassembly::SymbolInfo;
rocprofiler_client_id_t* client_id = nullptr;
rocprofiler_context_id_t client_ctx = {};
@@ -109,13 +121,35 @@ struct isa_map_elem_t
struct ToolData
{
ToolData()
{
try
{
output_file.open(OUTPUT_OFSTREAM);
} catch(...)
{}
if(output_file.is_open())
std::cout << "Writing code-object-isa-decode log to: " << OUTPUT_OFSTREAM << std::endl;
else
std::cout << "Could not open log file: " << OUTPUT_OFSTREAM << ", writing to stdout\n";
};
std::shared_mutex isa_map_mut;
std::mutex output_mut;
CodeobjAddressTranslate codeobjTranslate;
std::map<pcinfo_t, std::unique_ptr<isa_map_elem_t>> isa_map;
std::unordered_map<uint64_t, SymbolInfo> kernels_in_codeobj = {};
std::unordered_map<uint64_t, std::string> kernel_object_to_kernel_name = {};
std::stringstream output;
int num_waves = 0;
std::ostream& output()
{
if(output_file.is_open())
return output_file;
else
return std::cout;
}
std::stringstream printKernel(uint64_t vaddr)
{
@@ -133,6 +167,9 @@ struct ToolData
}
return ss;
}
private:
std::ofstream output_file;
};
struct source_location
@@ -151,15 +188,12 @@ struct trace_data_t
ToolData* tool;
};
std::atomic<int> TRACE_DATA_ID{-1};
std::atomic<int> KERNEL_ADDR_ID{-1};
std::atomic<int> OCCUPANCY_ID{-1};
void
tool_codeobj_tracing_callback(rocprofiler_callback_tracing_record_t record,
rocprofiler_user_data_t* user_data,
void* callback_data)
{
C_API_BEGIN
if(record.kind != ROCPROFILER_CALLBACK_TRACING_CODE_OBJECT) return;
if(record.phase != ROCPROFILER_CALLBACK_PHASE_LOAD) return;
@@ -200,19 +234,18 @@ tool_codeobj_tracing_callback(rocprofiler_callback_tracing_record_t record,
(void) user_data;
(void) callback_data;
C_API_END
}
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wunused-parameter"
rocprofiler_att_control_flags_t
dispatch_callback(rocprofiler_queue_id_t queue_id,
const rocprofiler_agent_t* agent,
rocprofiler_correlation_id_t correlation_id,
dispatch_callback(rocprofiler_queue_id_t /* queue_id */,
const rocprofiler_agent_t* /* agent */,
rocprofiler_correlation_id_t /* correlation_id */,
const hsa_kernel_dispatch_packet_t* dispatch_packet,
uint64_t kernel_id,
void* userdata)
uint64_t /* kernel_id */,
void* userdata)
{
C_API_BEGIN
assert(userdata && "Dispatch callback passed null!");
ToolData& tool = *reinterpret_cast<ToolData*>(userdata);
@@ -236,82 +269,55 @@ dispatch_callback(rocprofiler_queue_id_t queue_id,
<< std::endl;
}
C_API_END
return ROCPROFILER_ATT_CONTROL_NONE;
}
void
iterate_trace_types(int id, const char* metadata, void*)
{
if(std::string_view(metadata).find("occupancy") == 0)
OCCUPANCY_ID.store(id);
else if(std::string_view(metadata).find("kernel_ids_addr") == 0)
KERNEL_ADDR_ID.store(id);
else if(std::string_view(metadata).find("tracedata") == 0)
TRACE_DATA_ID.store(id);
}
hsa_status_t
get_trace_data(int trace_type_id,
int correlation_id,
void* trace_events,
uint64_t trace_size,
void* userdata)
get_trace_data(rocprofiler_att_parser_data_type_t type, void* att_data, void* userdata)
{
C_API_BEGIN
assert(userdata && "ISA callback passed null!");
trace_data_t& trace_data = *reinterpret_cast<trace_data_t*>(userdata);
assert(trace_data.tool && "ISA callback passed null!");
ToolData& tool = *reinterpret_cast<ToolData*>(trace_data.tool);
std::stringstream ss;
std::shared_lock<std::shared_mutex> shared_lock(tool.isa_map_mut);
if(trace_type_id == OCCUPANCY_ID)
{
ss << "Num waves: " << trace_size / 2 << '\n';
// auto* occ = reinterpret_cast<att_occupancy_info_t*>(trace_events);
}
else if(trace_type_id == KERNEL_ADDR_ID)
{
ss << "Num KRN events: " << trace_size << std::hex << '\n';
auto* kaddr = reinterpret_cast<pcinfo_t*>(trace_events);
for(size_t i = 0; i < trace_size; i++)
if(kaddr[i].addr != 0)
{
ss << " - ADDR: " << kaddr[i].addr << ' ' << tool.printKernel(kaddr[i].addr).str()
<< '\n';
}
ss << std::dec;
}
else if(trace_type_id == TRACE_DATA_ID)
{
ss << "Trace Length: " << trace_size << '\n';
auto* tracedata = reinterpret_cast<att_trace_event_t*>(trace_events);
if(type == ROCPROFILER_ATT_PARSER_DATA_TYPE_OCCUPANCY) tool.num_waves++;
for(size_t i = 0; i < trace_size; i++)
if(type != ROCPROFILER_ATT_PARSER_DATA_TYPE_ISA) return;
auto& event = *reinterpret_cast<rocprofiler_att_data_type_isa_t*>(att_data);
pcinfo_t pc{event.marker_id, event.offset};
auto it = tool.isa_map.find(pc);
if(it == tool.isa_map.end())
{
shared_lock.unlock();
{
pcinfo_t pc = tracedata[i].pc;
auto it = tool.isa_map.find(pc);
if(it == tool.isa_map.end())
std::unique_lock<std::shared_mutex> unique_lock(tool.isa_map_mut);
auto ptr = std::make_unique<isa_map_elem_t>();
try
{
shared_lock.unlock();
{
std::unique_lock<std::shared_mutex> unique_lock(tool.isa_map_mut);
auto ptr = std::make_unique<isa_map_elem_t>();
ptr->code_line = tool.codeobjTranslate.get(pc.marker_id, pc.addr);
it = tool.isa_map.emplace(pc, std::move(ptr)).first;
}
shared_lock.lock();
ptr->code_line = tool.codeobjTranslate.get(pc.marker_id, pc.addr);
} catch(std::exception& e)
{
std::cerr << pc.marker_id << ":" << pc.addr << ' ' << e.what() << std::endl;
return;
} catch(...)
{
std::cerr << "Could not fetch: " << pc.marker_id << ':' << pc.addr << std::endl;
return;
}
it->second->hitcount.fetch_add(tracedata[i].hitcount, std::memory_order_relaxed);
it->second->latency.fetch_add(tracedata[i].latency, std::memory_order_relaxed);
it = tool.isa_map.emplace(pc, std::move(ptr)).first;
}
shared_lock.lock();
}
std::unique_lock<std::mutex> lk(tool.output_mut);
tool.output << ss.str();
return HSA_STATUS_SUCCESS;
it->second->hitcount.fetch_add(event.hitcount, std::memory_order_relaxed);
it->second->latency.fetch_add(event.latency, std::memory_order_relaxed);
C_API_END
}
uint64_t
@@ -325,16 +331,15 @@ copy_trace_data(int* seid, uint8_t** buffer, uint64_t* buffer_size, void* userda
return *buffer_size;
}
hsa_status_t
rocprofiler_status_t
isa_callback(char* isa_instruction,
char* source_reference,
uint64_t* isa_memory_size,
uint64_t* isa_size,
uint64_t* source_size,
uint64_t marker_id,
uint64_t offset,
void* userdata)
{
C_API_BEGIN
assert(userdata && "ISA callback passed null!");
trace_data_t& trace_data = *reinterpret_cast<trace_data_t*>(userdata);
assert(trace_data.tool && "ISA callback passed null!");
@@ -347,56 +352,47 @@ isa_callback(char* isa_instruction,
instruction = tool.codeobjTranslate.get(marker_id, offset);
}
if(!instruction.get()) return HSA_STATUS_ERROR_INVALID_ARGUMENT;
if(!instruction.get()) return ROCPROFILER_STATUS_ERROR_INVALID_ARGUMENT;
{
size_t tmp_isa_size = *isa_size;
size_t tmp_source_size = *source_size;
*isa_size = instruction->inst.size();
*source_size = instruction->comment.size();
size_t tmp_isa_size = *isa_size;
*isa_size = instruction->inst.size();
if(*isa_size > tmp_isa_size || *source_size > tmp_source_size)
return HSA_STATUS_ERROR_OUT_OF_RESOURCES;
if(*isa_size > tmp_isa_size) return ROCPROFILER_STATUS_ERROR_OUT_OF_RESOURCES;
}
memcpy(isa_instruction, instruction->inst.data(), *isa_size);
memcpy(source_reference, instruction->comment.data(), *source_size);
*isa_memory_size = instruction->size;
auto ptr = std::make_unique<isa_map_elem_t>();
ptr->code_line = std::move(instruction);
tool.isa_map.emplace(pcinfo_t{offset, marker_id}, std::move(ptr));
return HSA_STATUS_SUCCESS;
tool.isa_map.emplace(pcinfo_t{marker_id, offset}, std::move(ptr));
C_API_END
return ROCPROFILER_STATUS_SUCCESS;
}
void
shader_data_callback(int64_t se_id,
int64_t data_type_id,
const char* data_type_name,
void* se_data,
size_t data_size,
void* userdata)
shader_data_callback(int64_t se_id, void* se_data, size_t data_size, void* userdata)
{
C_API_BEGIN
assert(userdata && "Shader callback passed null!");
ToolData& tool = *reinterpret_cast<ToolData*>(userdata);
{
std::unique_lock<std::mutex> lk(tool.output_mut);
tool.output << "SE ID: " << se_id << " with size " << data_size << std::hex << '\n';
tool.output() << "SE ID: " << se_id << " with size " << data_size << std::hex << '\n';
}
trace_data_t data{.id = se_id, .data = (uint8_t*) se_data, .size = data_size, .tool = &tool};
auto status = aqlprofile_att_parse_data(copy_trace_data, get_trace_data, isa_callback, &data);
(void) status;
auto status = rocprofiler_att_parse_data(copy_trace_data, get_trace_data, isa_callback, &data);
if(status != ROCPROFILER_STATUS_SUCCESS)
std::cerr << "shader_data_callback failed with status " << status << std::endl;
C_API_END
}
#pragma GCC diagnostic pop
int
tool_init(rocprofiler_client_finalize_t fini_func, void* tool_data)
{
(void) fini_func;
aqlprofile_att_parser_iterate_event_list(iterate_trace_types, nullptr);
ROCPROFILER_CALL(rocprofiler_create_context(&client_ctx), "context creation");
ROCPROFILER_CALL(
@@ -408,22 +404,19 @@ tool_init(rocprofiler_client_finalize_t fini_func, void* tool_data)
tool_data),
"code object tracing service configure");
rocprofiler_att_parameters_t parameters{};
parameters.target_cu = TARGET_CU;
parameters.simd_select = SIMD_SELECT;
parameters.buffer_size = BUFFER_SIZE;
std::vector<rocprofiler_att_parameter_t> parameters;
parameters.push_back({ROCPROFILER_ATT_PARAMETER_TARGET_CU, TARGET_CU});
parameters.push_back({ROCPROFILER_ATT_PARAMETER_SIMD_SELECT, SIMD_SELECT});
parameters.push_back({ROCPROFILER_ATT_PARAMETER_BUFFER_SIZE, BUFFER_SIZE});
parameters.push_back({ROCPROFILER_ATT_PARAMETER_SHADER_ENGINE_MASK, SE_MASK});
std::vector<int> shaders;
for(size_t i = 0; i < NUM_SE; i++)
shaders.push_back(2 * i); // use shader engines 0, 2
parameters.shader_ids = shaders.data();
parameters.shader_num = shaders.size();
ROCPROFILER_CALL(
rocprofiler_configure_thread_trace_service(
client_ctx, parameters, dispatch_callback, shader_data_callback, tool_data),
"thread trace service configure");
ROCPROFILER_CALL(rocprofiler_configure_thread_trace_service(client_ctx,
parameters.data(),
parameters.size(),
dispatch_callback,
shader_data_callback,
tool_data),
"thread trace service configure");
int valid_ctx = 0;
ROCPROFILER_CALL(rocprofiler_context_is_valid(client_ctx, &valid_ctx),
@@ -476,9 +469,9 @@ tool_fini(void* tool_data)
size_t latency = line->latency.load(std::memory_order_relaxed);
auto& code_line = line->code_line->inst;
tool.output << std::hex << "0x" << addr.addr << std::dec << ' ' << code_line
<< empty_space.substr(0, max_inst_size - code_line.size())
<< " Hit: " << hitcount << " - Latency: " << latency << '\n';
tool.output() << std::hex << "0x" << addr.addr << std::dec << ' ' << code_line
<< empty_space.substr(0, max_inst_size - code_line.size())
<< " Hit: " << hitcount << " - Latency: " << latency << '\n';
if(code_line.find("s_waitcnt") == 0)
{
@@ -511,30 +504,16 @@ tool_fini(void* tool_data)
float vmc_fraction = 100 * vmc_latency / float(total_latency);
float lgk_fraction = 100 * lgk_latency / float(total_latency);
tool.output << "Total executed instructions: " << total_exec << '\n'
<< "Total executed vector instructions: " << vector_exec << " with average "
<< vector_latency / float(vector_exec) << " cycles.\n"
<< "Total executed scalar instructions: " << scalar_exec << " with average "
<< scalar_latency / float(scalar_exec) << " cycles.\n"
<< "Vector memory ops occupied: " << vmc_fraction << "% of cycles.\n"
<< "Scalar and LDS memory ops occupied: " << lgk_fraction << "% of cycles.\n";
std::cout << tool.output.str();
tool.output() << "Total executed instructions: " << total_exec << '\n'
<< "Total executed vector instructions: " << vector_exec << " with average "
<< vector_latency / float(vector_exec) << " cycles.\n"
<< "Total executed scalar instructions: " << scalar_exec << " with average "
<< scalar_latency / float(scalar_exec) << " cycles.\n"
<< "Vector memory ops occupied: " << vmc_fraction << "% of cycles.\n"
<< "Scalar and LDS memory ops occupied: " << lgk_fraction << "% of cycles.\n"
<< std::endl;
}
void
setup()
{
if(int status = 0;
rocprofiler_is_initialized(&status) == ROCPROFILER_STATUS_SUCCESS && status == 0)
{
ROCPROFILER_CALL(rocprofiler_force_configure(&rocprofiler_configure),
"force configuration");
}
}
// force configuration when library is loaded
bool cfg_on_load = (client::setup(), true);
} // namespace client
extern "C" rocprofiler_tool_configure_result_t*