ROCProfV3 PC sampling tests: Initial multi-agents test (#72)

Testing multi-agent host-trap PC sampling support in ROCProfV3.

[ROCm/rocprofiler-sdk commit: 00d4c179c6]
This commit is contained in:
Indic, Vladimir
2025-01-04 02:35:16 +01:00
committed by GitHub
parent 93421ab066
commit 5c0235ccc4
7 changed files with 319 additions and 0 deletions
@@ -3,3 +3,4 @@
#
add_subdirectory(exec-mask-manipulation)
add_subdirectory(transpose-multiple-agents)
@@ -0,0 +1,165 @@
#
# rocprofv3 tool test
#
cmake_minimum_required(VERSION 3.21.0 FATAL_ERROR)
project(
rocprofiler-tests-pc-sampling
LANGUAGES CXX
VERSION 0.0.0)
find_package(rocprofiler-sdk REQUIRED)
rocprofiler_configure_pytest_files(CONFIG pytest.ini COPY validate.py conftest.py
input.json input.yml)
# To ensure we sample all agents, use 16 threads each assigned to one agent in a round
# robin manner. To keep the job per device reasonable, each thread offloads 100 instances
# of a transpose kernel.
set(TRANSPOSE_NUM_THREADS 16)
set(TRANSPOSE_NUM_ITERATIONS 100)
add_test(
NAME rocprofv3-test-pc-sampling-host-trap-transpose-multiple-agents-input-cmd-execute
COMMAND
$<TARGET_FILE:rocprofiler-sdk::rocprofv3> --kernel-trace --pc-sampling-unit time
--pc-sampling-method host_trap --pc-sampling-interval 1 -d
${CMAKE_CURRENT_BINARY_DIR}/pc_sampling_cmd_input -o out --output-format csv --
$<TARGET_FILE:transpose> ${TRANSPOSE_NUM_THREADS} ${TRANSPOSE_NUM_ITERATIONS})
string(REPLACE "LD_PRELOAD=" "ROCPROF_PRELOAD=" PRELOAD_ENV
"${ROCPROFILER_MEMCHECK_PRELOAD_ENV}")
set(pc-sampling-env-host-trap-transpose-multiple-agents "${PRELOAD_ENV}")
set_tests_properties(
rocprofv3-test-pc-sampling-host-trap-transpose-multiple-agents-input-cmd-execute
PROPERTIES TIMEOUT
45
LABELS
"integration-tests;pc-sampling"
ENVIRONMENT
"${pc-sampling-env-host-trap-transpose-multiple-agents}"
FAIL_REGULAR_EXPRESSION
"${ROCPROFILER_DEFAULT_FAIL_REGEX}"
SKIP_REGULAR_EXPRESSION
"PC sampling unavailable")
add_test(
NAME rocprofv3-test-pc-sampling-host-trap-transpose-multiple-agents-input-json-execute
COMMAND
$<TARGET_FILE:rocprofiler-sdk::rocprofv3> -i
${CMAKE_CURRENT_BINARY_DIR}/input.json -d
${CMAKE_CURRENT_BINARY_DIR}/pc_sampling_json_input -- $<TARGET_FILE:transpose>
${TRANSPOSE_NUM_THREADS} ${TRANSPOSE_NUM_ITERATIONS})
set_tests_properties(
rocprofv3-test-pc-sampling-host-trap-transpose-multiple-agents-input-json-execute
PROPERTIES TIMEOUT
45
LABELS
"integration-tests;pc-sampling"
ENVIRONMENT
"${pc-sampling-env-host-trap-transpose-multiple-agents}"
FAIL_REGULAR_EXPRESSION
"${ROCPROFILER_DEFAULT_FAIL_REGEX}"
SKIP_REGULAR_EXPRESSION
"PC sampling unavailable")
add_test(
NAME rocprofv3-test-pc-sampling-host-trap-transpose-multiple-agents-input-yaml-execute
COMMAND
$<TARGET_FILE:rocprofiler-sdk::rocprofv3> -i
${CMAKE_CURRENT_BINARY_DIR}/input.yml -d
${CMAKE_CURRENT_BINARY_DIR}/pc_sampling_yaml_input -o out --output-format csv --
$<TARGET_FILE:transpose> ${TRANSPOSE_NUM_THREADS} ${TRANSPOSE_NUM_ITERATIONS})
set_tests_properties(
rocprofv3-test-pc-sampling-host-trap-transpose-multiple-agents-input-yaml-execute
PROPERTIES TIMEOUT
45
LABELS
"integration-tests;pc-sampling"
ENVIRONMENT
"${pc-sampling-env-host-trap-transpose-multiple-agents}"
FAIL_REGULAR_EXPRESSION
"${ROCPROFILER_DEFAULT_FAIL_REGEX}"
SKIP_REGULAR_EXPRESSION
"PC sampling unavailable")
# ========================= Validation tests
add_test(
NAME rocprofv3-test-pc-sampling-host-trap-transpose-multiple-agents-input-cmd-validate
COMMAND
${Python3_EXECUTABLE} ${CMAKE_CURRENT_BINARY_DIR}/validate.py -k
test_multi_agent_support --input-samples-csv
${CMAKE_CURRENT_BINARY_DIR}/pc_sampling_cmd_input/out_pc_sampling_host_trap.csv
--input-kernel-trace-csv
${CMAKE_CURRENT_BINARY_DIR}/pc_sampling_cmd_input/out_kernel_trace.csv
--input-agent-info-csv
${CMAKE_CURRENT_BINARY_DIR}/pc_sampling_cmd_input/out_agent_info.csv)
set_tests_properties(
rocprofv3-test-pc-sampling-host-trap-transpose-multiple-agents-input-cmd-validate
PROPERTIES
TIMEOUT
60
LABELS
"integration-tests;pc-sampling"
DEPENDS
"rocprofv3-test-pc-sampling-host-trap-transpose-multiple-agents-input-cmd-execute"
FAIL_REGULAR_EXPRESSION
"${ROCPROFILER_DEFAULT_FAIL_REGEX}"
SKIP_REGULAR_EXPRESSION
"PC sampling unavailable")
add_test(
NAME rocprofv3-test-pc-sampling-host-trap-transpose-multiple-agents-input-json-validate
COMMAND
${Python3_EXECUTABLE} ${CMAKE_CURRENT_BINARY_DIR}/validate.py -k
test_multi_agent_support --input-samples-csv
${CMAKE_CURRENT_BINARY_DIR}/pc_sampling_json_input/out_pc_sampling_host_trap.csv
--input-kernel-trace-csv
${CMAKE_CURRENT_BINARY_DIR}/pc_sampling_json_input/out_kernel_trace.csv
--input-agent-info-csv
${CMAKE_CURRENT_BINARY_DIR}/pc_sampling_json_input/out_agent_info.csv)
set_tests_properties(
rocprofv3-test-pc-sampling-host-trap-transpose-multiple-agents-input-json-validate
PROPERTIES
TIMEOUT
60
LABELS
"integration-tests;pc-sampling"
DEPENDS
"rocprofv3-test-pc-sampling-host-trap-transpose-multiple-agents-input-json-execute"
FAIL_REGULAR_EXPRESSION
"${ROCPROFILER_DEFAULT_FAIL_REGEX}"
SKIP_REGULAR_EXPRESSION
"PC sampling unavailable")
add_test(
NAME rocprofv3-test-pc-sampling-host-trap-transpose-multiple-agents-input-yaml-validate
COMMAND
${Python3_EXECUTABLE} ${CMAKE_CURRENT_BINARY_DIR}/validate.py -k
test_multi_agent_support --input-samples-csv
${CMAKE_CURRENT_BINARY_DIR}/pc_sampling_yaml_input/out_pc_sampling_host_trap.csv
--input-kernel-trace-csv
${CMAKE_CURRENT_BINARY_DIR}/pc_sampling_yaml_input/out_kernel_trace.csv
--input-agent-info-csv
${CMAKE_CURRENT_BINARY_DIR}/pc_sampling_yaml_input/out_agent_info.csv)
set_tests_properties(
rocprofv3-test-pc-sampling-host-trap-transpose-multiple-agents-input-yaml-validate
PROPERTIES
TIMEOUT
60
LABELS
"integration-tests;pc-sampling"
DEPENDS
"rocprofv3-test-pc-sampling-host-trap-transpose-multiple-agents-input-yaml-execute"
FAIL_REGULAR_EXPRESSION
"${ROCPROFILER_DEFAULT_FAIL_REGEX}"
SKIP_REGULAR_EXPRESSION
"PC sampling unavailable")
@@ -0,0 +1,65 @@
#!/usr/bin/env python3
import json
import os
import pytest
import pandas as pd
from rocprofiler_sdk.pytest_utils.dotdict import dotdict
from rocprofiler_sdk.pytest_utils import collapse_dict_list
def pytest_addoption(parser):
parser.addoption(
"--input-samples-csv",
action="store",
help="Path to CSV file containing PC samples.",
)
parser.addoption(
"--input-kernel-trace-csv",
action="store",
help="Path to CSV file containing kernel trace.",
)
parser.addoption(
"--input-agent-info-csv",
action="store",
help="Path to CSV file containing agents information.",
)
@pytest.fixture
def input_samples_csv(request):
filename = request.config.getoption("--input-samples-csv")
if not os.path.isfile(filename):
# The CSV file is not generated, because the dependency test
# responsible to generate this file was skipped or failed.
# Thus emit the message to skip this test as well.
print("PC sampling unavailable")
else:
with open(filename, "r") as inp:
return pd.read_csv(
inp,
na_filter=False, # parse empty fields as ""
keep_default_na=False, # parse empty fields as ""
dtype={
"Exec_Mask": "uint64",
"Instruction": str,
"Instruction_Comment": str,
},
)
@pytest.fixture
def input_kernel_trace_csv(request):
filename = request.config.getoption("--input-kernel-trace-csv")
with open(filename, "r") as inp:
return pd.read_csv(inp)
@pytest.fixture
def input_agent_info_csv(request):
filename = request.config.getoption("--input-agent-info-csv")
with open(filename, "r") as inp:
return pd.read_csv(inp)
@@ -0,0 +1,14 @@
{
"jobs": [
{
"kernel_trace": true,
"pc_sampling_unit": "time",
"pc_sampling_method": "host_trap",
"pc_sampling_interval": 1,
"output_file": "out",
"output_format": [
"csv"
]
}
]
}
@@ -0,0 +1,5 @@
jobs:
- kernel_trace: true
pc_sampling_unit: "time"
pc_sampling_method: "host_trap"
pc_sampling_interval: 1
@@ -0,0 +1,5 @@
[pytest]
addopts = --durations=20 -rA -s -vv
testpaths = validate.py
pythonpath = @ROCPROFILER_SDK_TESTS_BINARY_DIR@/pytest-packages
@@ -0,0 +1,64 @@
#!/usr/bin/env python3
import itertools
import sys
import pytest
import numpy as np
import pandas as pd
def test_multi_agent_support(
input_samples_csv: pd.DataFrame,
input_kernel_trace_csv: pd.DataFrame,
input_agent_info_csv: pd.DataFrame,
):
transpose_kernel_source_line_start = 137
transpose_kernel_source_line_end = 145
mi2xx_mi3xx_agents_df = input_agent_info_csv[
input_agent_info_csv["Name"].apply(
lambda name: name == "gfx90a" or name.startswith("gfx94")
)
]
# Extract samples that originates from know code object it
samples_df = input_samples_csv[input_samples_csv["Dispatch_Id"] != 0].copy()
# Determine the agent on which sample was generated
samples_df["Agent_Id"] = (
samples_df["Dispatch_Id"]
.map(input_kernel_trace_csv.set_index("Dispatch_Id")["Agent_Id"])
.astype(np.uint64)
)
sampled_agents = samples_df["Agent_Id"].unique()
sampled_agents_num = len(sampled_agents)
# all agents must be sampled
assert sampled_agents_num == len(mi2xx_mi3xx_agents_df)
# separate samples per agents
grouped_samples_per_agent = samples_df.groupby("Agent_Id")
for agent_id, agent_samples_df in grouped_samples_per_agent:
sampled_dispatches = agent_samples_df["Dispatch_Id"].unique()
# at least 1 sampled dispatch per agent
assert len(sampled_dispatches) >= 1
# extract decoded samples that are mapped to the transpose.cpp file
transpose_samples_df = samples_df[
samples_df["Instruction_Comment"].apply(
lambda comment: "transpose-all-agents.cpp" in comment
)
].copy()
# determine the line number for each sample
transpose_samples_df["Source_Line_Num"] = transpose_samples_df[
"Instruction_Comment"
].apply(lambda source_line: int(source_line.split(":")[-1]))
# assert that line belongs to a kernel range
assert (
(transpose_samples_df["Source_Line_Num"] >= transpose_kernel_source_line_start)
& (transpose_samples_df["Source_Line_Num"] <= transpose_kernel_source_line_end)
).all()
if __name__ == "__main__":
exit_code = pytest.main(["-x", __file__] + sys.argv[1:])
sys.exit(exit_code)