Use gersemi for CMake formatting (#257)

* Replace `cmake-format` with `gersemi`

Signed-off-by: David Galiffi <David.Galiffi@amd.com>

* Remove .cmake-format.yaml

Signed-off-by: David Galiffi <David.Galiffi@amd.com>

* Update workflow to use gersemi

Signed-off-by: David Galiffi <David.Galiffi@amd.com>

* Update CONTRIBUTING.md

* Update helper scripts

* Don't include `*/external/*` in workflows

---------

Signed-off-by: David Galiffi <David.Galiffi@amd.com>
This commit is contained in:
David Galiffi
2025-06-22 10:44:33 -04:00
committato da GitHub
parent 1e13b590e7
commit 122623a929
106 ha cambiato i file con 5116 aggiunte e 3156 eliminazioni
+25 -12
Vedi File
@@ -26,15 +26,20 @@
#
# -------------------------------------------------------------------------------------- #
if(ROCPROFSYS_USE_PAPI
AND (rocprofiler_systems_perf_event_paranoid LESS_EQUAL 3
if(
ROCPROFSYS_USE_PAPI
AND (
rocprofiler_systems_perf_event_paranoid LESS_EQUAL 3
OR rocprofiler_systems_cap_sys_admin EQUAL 0
OR rocprofiler_systems_cap_perfmon EQUAL 0))
OR rocprofiler_systems_cap_perfmon EQUAL 0
)
)
set(_annotate_environment
"${_base_environment}"
"ROCPROFSYS_TIMEMORY_COMPONENTS=thread_cpu_clock papi_array"
"ROCPROFSYS_PAPI_EVENTS=perf::PERF_COUNT_SW_CPU_CLOCK"
"ROCPROFSYS_USE_SAMPLING=OFF")
"ROCPROFSYS_USE_SAMPLING=OFF"
)
rocprofiler_systems_add_test(
SKIP_BASELINE SKIP_RUNTIME
@@ -58,23 +63,28 @@ if(ROCPROFSYS_USE_PAPI
functions
--print-instructions
ENVIRONMENT "${_annotate_environment}"
LABELS "annotate;papi")
LABELS "annotate;papi"
)
rocprofiler_systems_add_validation_test(
NAME annotate-binary-rewrite
PERFETTO_FILE "perfetto-trace.proto"
LABELS "annotate;papi"
ARGS --key-names perf::PERF_COUNT_SW_CPU_CLOCK thread_cpu_clock --key-counts 8 8)
ARGS --key-names perf::PERF_COUNT_SW_CPU_CLOCK thread_cpu_clock --key-counts 8 8
)
rocprofiler_systems_add_validation_test(
NAME annotate-sampling
PERFETTO_FILE "perfetto-trace.proto"
LABELS "papi"
ARGS --key-names thread_cpu_clock --key-counts 6)
ARGS --key-names thread_cpu_clock --key-counts 6
)
else()
set(_annotate_environment
"${_base_environment}" "ROCPROFSYS_TIMEMORY_COMPONENTS=thread_cpu_clock"
"ROCPROFSYS_USE_SAMPLING=OFF")
"${_base_environment}"
"ROCPROFSYS_TIMEMORY_COMPONENTS=thread_cpu_clock"
"ROCPROFSYS_USE_SAMPLING=OFF"
)
rocprofiler_systems_add_test(
SKIP_BASELINE SKIP_RUNTIME
@@ -98,17 +108,20 @@ else()
functions
--print-instructions
ENVIRONMENT "${_annotate_environment}"
LABELS "annotate")
LABELS "annotate"
)
rocprofiler_systems_add_validation_test(
NAME annotate-binary-rewrite
PERFETTO_FILE "perfetto-trace.proto"
LABELS "annotate"
ARGS --key-names thread_cpu_clock --key-counts 8)
ARGS --key-names thread_cpu_clock --key-counts 8
)
rocprofiler_systems_add_validation_test(
NAME annotate-sampling
PERFETTO_FILE "perfetto-trace.proto"
LABELS "annotate"
ARGS --key-names thread_cpu_clock --key-counts 6)
ARGS --key-names thread_cpu_clock --key-counts 6
)
endif()
+15 -14
Vedi File
@@ -38,7 +38,7 @@ else()
rocprofiler_systems_message(
AUTHOR_WARNING
"Disabling attach tests. Run 'echo 0 | sudo tee /proc/sys/kernel/yama/ptrace_scope' to enable attaching to process"
)
)
endif()
if(NOT _VALID_PTRACE_SCOPE)
@@ -55,21 +55,22 @@ add_test(
${CMAKE_CURRENT_LIST_DIR}/run-rocprof-sys-pid.sh
$<TARGET_FILE:rocprofiler-systems-instrument> -ME "\.c$" -E fib -e -v 1 --label
return args file -l -- $<TARGET_FILE:parallel-overhead> 30 8 1000
WORKING_DIRECTORY ${PROJECT_BINARY_DIR})
WORKING_DIRECTORY ${PROJECT_BINARY_DIR}
)
set(_parallel_overhead_attach_environ
"${_attach_environment}" "ROCPROFSYS_OUTPUT_PATH=rocprof-sys-tests-output"
"ROCPROFSYS_OUTPUT_PREFIX=parallel-overhead-attach/")
"${_attach_environment}"
"ROCPROFSYS_OUTPUT_PATH=rocprof-sys-tests-output"
"ROCPROFSYS_OUTPUT_PREFIX=parallel-overhead-attach/"
)
set_tests_properties(
parallel-overhead-attach
PROPERTIES ENVIRONMENT
"${_parallel_overhead_attach_environ}"
TIMEOUT
300
LABELS
"parallel-overhead;attach"
PASS_REGULAR_EXPRESSION
"Outputting.*(perfetto-trace.proto).*Outputting.*(wall_clock.txt)"
FAIL_REGULAR_EXPRESSION
"Dyninst was unable to attach to the specified process")
PROPERTIES
ENVIRONMENT "${_parallel_overhead_attach_environ}"
TIMEOUT 300
LABELS "parallel-overhead;attach"
PASS_REGULAR_EXPRESSION
"Outputting.*(perfetto-trace.proto).*Outputting.*(wall_clock.txt)"
FAIL_REGULAR_EXPRESSION "Dyninst was unable to attach to the specified process"
)
+21 -15
Vedi File
@@ -33,7 +33,7 @@ rocprofiler_systems_add_causal_test(
CAUSAL_MODE "function"
CAUSAL_PASS_REGEX
"Starting causal experiment #1(.*)causal/experiments.json(.*)causal/experiments.coz"
)
)
rocprofiler_systems_add_causal_test(
SKIP_BASELINE
@@ -43,7 +43,7 @@ rocprofiler_systems_add_causal_test(
CAUSAL_MODE "function"
CAUSAL_PASS_REGEX
"Starting causal experiment #1(.*)causal/experiments.json(.*)causal/experiments.coz"
)
)
rocprofiler_systems_add_causal_test(
SKIP_BASELINE
@@ -53,7 +53,7 @@ rocprofiler_systems_add_causal_test(
CAUSAL_MODE "line"
CAUSAL_PASS_REGEX
"Starting causal experiment #1(.*)causal/experiments.json(.*)causal/experiments.coz"
)
)
rocprofiler_systems_add_causal_test(
NAME both-rocprofsys-func
@@ -79,7 +79,7 @@ rocprofiler_systems_add_causal_test(
ENVIRONMENT "ROCPROFSYS_STRICT_CONFIG=OFF"
CAUSAL_PASS_REGEX
"Starting causal experiment #1(.*)causal/experiments.json(.*)causal/experiments.coz"
)
)
rocprofiler_systems_add_causal_test(
NAME lulesh-func
@@ -89,7 +89,7 @@ rocprofiler_systems_add_causal_test(
CAUSAL_ARGS -s 0,10,25,50,75
CAUSAL_PASS_REGEX
"Starting causal experiment #1(.*)causal/experiments.json(.*)causal/experiments.coz"
)
)
rocprofiler_systems_add_causal_test(
SKIP_BASELINE
@@ -100,7 +100,7 @@ rocprofiler_systems_add_causal_test(
CAUSAL_ARGS -s 0,10,25,50,75
CAUSAL_PASS_REGEX
"Starting causal experiment #1(.*)causal/experiments.json(.*)causal/experiments.coz"
)
)
rocprofiler_systems_add_causal_test(
SKIP_BASELINE
@@ -111,13 +111,14 @@ rocprofiler_systems_add_causal_test(
CAUSAL_ARGS -s 0,10,25,50,75 -S lulesh.cc
CAUSAL_PASS_REGEX
"Starting causal experiment #1(.*)causal/experiments.json(.*)causal/experiments.coz"
)
)
# set(_causal_e2e_exe_args 80 100 432525 100000000) set(_causal_e2e_exe_args 80 12 432525
# 500000000)
set(_causal_e2e_exe_args 80 50 432525 100000000)
set(_causal_common_args
"-n 5 -e -s 0 10 20 30 -B $<TARGET_FILE_BASE_NAME:causal-cpu-rocprofsys>")
"-n 5 -e -s 0 10 20 30 -B $<TARGET_FILE_BASE_NAME:causal-cpu-rocprofsys>"
)
macro(
causal_e2e_args_and_validation
@@ -129,14 +130,14 @@ macro(
_V20
_V30
_TOL # tolerance for virtual speedup
)
)
# arguments to rocprofiler-systems-causal
set(${_NAME}_args "${_causal_common_args} ${_MODE} ${_EXPER}")
# arguments to validate-causal-json.py
set(${_NAME}_valid
"-n 0 -i rocprof-sys-tests-output/causal-cpu-rocprofsys-${_TEST}-e2e/causal/experiments.json -v ${_EXPER} $<TARGET_FILE_BASE_NAME:causal-cpu-rocprofsys> 10 ${_V10} ${_TOL} ${_EXPER} $<TARGET_FILE_BASE_NAME:causal-cpu-rocprofsys> 20 ${_V20} ${_TOL} ${_EXPER} $<TARGET_FILE_BASE_NAME:causal-cpu-rocprofsys> 30 ${_V30} ${_TOL}"
)
)
# patch string for command-line
string(REPLACE " " ";" ${_NAME}_args "${${_NAME}_args}")
@@ -144,7 +145,8 @@ macro(
endmacro()
causal_e2e_args_and_validation(_causal_slow_func slow-func "-F" "cpu_slow_func" 10 20 20
5)
5
)
causal_e2e_args_and_validation(_causal_fast_func fast-func "-F" "cpu_fast_func" 0 0 0 5)
causal_e2e_args_and_validation(_causal_line_100 line-100 "-S" "causal.cpp:100" 10 20 20 5)
causal_e2e_args_and_validation(_causal_line_110 line-110 "-S" "causal.cpp:110" 0 0 0 5)
@@ -167,7 +169,8 @@ rocprofiler_systems_add_causal_test(
CAUSAL_PASS_REGEX
"Starting causal experiment #1(.*)causal/experiments.json(.*)causal/experiments.coz"
ENVIRONMENT "${_causal_e2e_environment}"
PROPERTIES PROCESSORS 2 PROCESSOR_AFFINITY OFF)
PROPERTIES PROCESSORS 2 PROCESSOR_AFFINITY OFF
)
rocprofiler_systems_add_causal_test(
SKIP_BASELINE
@@ -181,7 +184,8 @@ rocprofiler_systems_add_causal_test(
CAUSAL_PASS_REGEX
"Starting causal experiment #1(.*)causal/experiments.json(.*)causal/experiments.coz"
ENVIRONMENT "${_causal_e2e_environment}"
PROPERTIES PROCESSORS 2 PROCESSOR_AFFINITY OFF)
PROPERTIES PROCESSORS 2 PROCESSOR_AFFINITY OFF
)
rocprofiler_systems_add_causal_test(
SKIP_BASELINE
@@ -195,7 +199,8 @@ rocprofiler_systems_add_causal_test(
CAUSAL_PASS_REGEX
"Starting causal experiment #1(.*)causal/experiments.json(.*)causal/experiments.coz"
ENVIRONMENT "${_causal_e2e_environment}"
PROPERTIES PROCESSORS 2 PROCESSOR_AFFINITY OFF)
PROPERTIES PROCESSORS 2 PROCESSOR_AFFINITY OFF
)
rocprofiler_systems_add_causal_test(
SKIP_BASELINE
@@ -209,4 +214,5 @@ rocprofiler_systems_add_causal_test(
CAUSAL_PASS_REGEX
"Starting causal experiment #1(.*)causal/experiments.json(.*)causal/experiments.coz"
ENVIRONMENT "${_causal_e2e_environment}"
PROPERTIES PROCESSORS 2 PROCESSOR_AFFINITY OFF)
PROPERTIES PROCESSORS 2 PROCESSOR_AFFINITY OFF
)
@@ -63,7 +63,8 @@ rocprofiler_systems_add_test(
RUN_ARGS 10 ${NUM_THREADS} 1000
ENVIRONMENT "${_base_environment}"
RUNTIME_PASS_REGEX "(\\\[[0-9]+\\\]) code coverage :: 66.67%"
REWRITE_RUN_PASS_REGEX "(\\\[[0-9]+\\\]) code coverage :: 66.67%")
REWRITE_RUN_PASS_REGEX "(\\\[[0-9]+\\\]) code coverage :: 66.67%"
)
rocprofiler_systems_add_test(
SKIP_BASELINE SKIP_SAMPLING
@@ -90,7 +91,8 @@ rocprofiler_systems_add_test(
RUN_ARGS 10 ${NUM_THREADS} 1000
ENVIRONMENT "${_base_environment}"
RUNTIME_PASS_REGEX "(\\\[[0-9]+\\\]) code coverage :: 66.67%"
REWRITE_RUN_PASS_REGEX "(\\\[[0-9]+\\\]) code coverage :: 66.67%")
REWRITE_RUN_PASS_REGEX "(\\\[[0-9]+\\\]) code coverage :: 66.67%"
)
rocprofiler_systems_add_test(
SKIP_BASELINE SKIP_SAMPLING
@@ -129,7 +131,8 @@ rocprofiler_systems_add_test(
RUN_ARGS 10 ${NUM_THREADS} 1000
ENVIRONMENT "${_base_environment}"
RUNTIME_PASS_REGEX "(\\\[[0-9]+\\\]) function coverage :: 66.67%"
REWRITE_RUN_PASS_REGEX "(\\\[[0-9]+\\\]) function coverage :: 66.67%")
REWRITE_RUN_PASS_REGEX "(\\\[[0-9]+\\\]) function coverage :: 66.67%"
)
rocprofiler_systems_add_test(
SKIP_BASELINE SKIP_SAMPLING
@@ -156,4 +159,5 @@ rocprofiler_systems_add_test(
RUN_ARGS 10 ${NUM_THREADS} 1000
ENVIRONMENT "${_base_environment}"
RUNTIME_PASS_REGEX "(\\\[[0-9]+\\\]) function coverage :: 66.67%"
REWRITE_RUN_PASS_REGEX "(\\\[[0-9]+\\\]) function coverage :: 66.67%")
REWRITE_RUN_PASS_REGEX "(\\\[[0-9]+\\\]) function coverage :: 66.67%"
)
+18 -18
Vedi File
@@ -27,11 +27,13 @@
# -------------------------------------------------------------------------------------- #
file(
WRITE ${CMAKE_CURRENT_BINARY_DIR}/invalid.cfg
WRITE
${CMAKE_CURRENT_BINARY_DIR}/invalid.cfg
"
ROCPROFSYS_CONFIG_FILE =
FOOBAR = ON
")
"
)
if(TARGET parallel-overhead)
set(_CONFIG_TEST_EXE $<TARGET_FILE:parallel-overhead>)
@@ -42,33 +44,31 @@ endif()
add_test(
NAME rocprofiler-systems-invalid-config
COMMAND $<TARGET_FILE:rocprofiler-systems-instrument> -- ${_CONFIG_TEST_EXE}
WORKING_DIRECTORY ${PROJECT_BINARY_DIR})
WORKING_DIRECTORY ${PROJECT_BINARY_DIR}
)
set_tests_properties(
rocprofiler-systems-invalid-config
PROPERTIES
ENVIRONMENT
"ROCPROFSYS_CONFIG_FILE=${CMAKE_CURRENT_BINARY_DIR}/invalid.cfg;ROCPROFSYS_CI=ON;ROCPROFSYS_CI_TIMEOUT=120"
TIMEOUT
120
LABELS
"config"
WILL_FAIL
ON)
"ROCPROFSYS_CONFIG_FILE=${CMAKE_CURRENT_BINARY_DIR}/invalid.cfg;ROCPROFSYS_CI=ON;ROCPROFSYS_CI_TIMEOUT=120"
TIMEOUT 120
LABELS "config"
WILL_FAIL ON
)
add_test(
NAME rocprofiler-systems-missing-config
COMMAND $<TARGET_FILE:rocprofiler-systems-instrument> -- ${_CONFIG_TEST_EXE}
WORKING_DIRECTORY ${PROJECT_BINARY_DIR})
WORKING_DIRECTORY ${PROJECT_BINARY_DIR}
)
set_tests_properties(
rocprofiler-systems-missing-config
PROPERTIES
ENVIRONMENT
"ROCPROFSYS_CONFIG_FILE=${CMAKE_CURRENT_BINARY_DIR}/missing.cfg;ROCPROFSYS_CI=ON;ROCPROFSYS_CI_TIMEOUT=120"
TIMEOUT
120
LABELS
"config"
WILL_FAIL
ON)
"ROCPROFSYS_CONFIG_FILE=${CMAKE_CURRENT_BINARY_DIR}/missing.cfg;ROCPROFSYS_CI=ON;ROCPROFSYS_CI_TIMEOUT=120"
TIMEOUT 120
LABELS "config"
WILL_FAIL ON
)
+12 -6
Vedi File
@@ -30,12 +30,14 @@ set(_video_decode_environment
"${_base_environment}"
"ROCPROFSYS_ROCM_DOMAINS=hip_runtime_api,kernel_dispatch,memory_copy,rocdecode_api"
"ROCPROFSYS_AMD_SMI_METRICS=busy,temp,power,vcn_activity,mem_usage"
"ROCPROFSYS_SAMPLING_CPUS=none")
"ROCPROFSYS_SAMPLING_CPUS=none"
)
set(_jpeg_decode_environment
"${_base_environment}"
"ROCPROFSYS_ROCM_DOMAINS=hip_runtime_api,kernel_dispatch,memory_copy,rocjpeg_api"
"ROCPROFSYS_AMD_SMI_METRICS=busy,temp,power,jpeg_activity,mem_usage"
"ROCPROFSYS_SAMPLING_CPUS=none")
"ROCPROFSYS_SAMPLING_CPUS=none"
)
rocprofiler_systems_get_gfx_archs(MI300_DETECTED GFX_MATCH "gfx9[4-9][A-Fa-f0-9]" ECHO)
@@ -54,14 +56,16 @@ rocprofiler_systems_add_test(
GPU ON
ENVIRONMENT "${_video_decode_environment}"
RUN_ARGS -i ${PROJECT_BINARY_DIR}/videos -t 1
LABELS "decode")
LABELS "decode"
)
rocprofiler_systems_add_validation_test(
NAME video-decode-sampling
PERFETTO_METRIC "rocm_rocdecode_api"
PERFETTO_FILE "perfetto-trace.proto"
LABELS "decode"
ARGS -l rocDecCreateVideoParser -c 2 -d 1 ${VCN_COUNTER_NAMES_ARG} -p)
ARGS -l rocDecCreateVideoParser -c 2 -d 1 ${VCN_COUNTER_NAMES_ARG} -p
)
# -------------------------------------------------------------------------------------- #
#
@@ -76,11 +80,13 @@ rocprofiler_systems_add_test(
GPU ON
ENVIRONMENT "${_jpeg_decode_environment}"
RUN_ARGS -i ${PROJECT_BINARY_DIR}/images -b 32
LABELS "decode")
LABELS "decode"
)
rocprofiler_systems_add_validation_test(
NAME jpeg-decode-sampling
PERFETTO_METRIC "rocm_rocjpeg_api"
PERFETTO_FILE "perfetto-trace.proto"
LABELS "decode"
ARGS -l rocJpegCreate -c 1 -d 1 ${JPEG_COUNTER_NAMES_ARG} -p)
ARGS -l rocJpegCreate -c 1 -d 1 ${JPEG_COUNTER_NAMES_ARG} -p
)
+2 -1
Vedi File
@@ -38,4 +38,5 @@ rocprofiler_systems_add_test(
REWRITE_RUN_PASS_REGEX "fork.. called on PID"
SAMPLING_FAIL_REGEX "(${ROCPROFSYS_ABORT_FAIL_REGEX})"
RUNTIME_FAIL_REGEX "(${ROCPROFSYS_ABORT_FAIL_REGEX})"
REWRITE_RUN_FAIL_REGEX "(${ROCPROFSYS_ABORT_FAIL_REGEX})")
REWRITE_RUN_FAIL_REGEX "(${ROCPROFSYS_ABORT_FAIL_REGEX})"
)
+5 -3
Vedi File
@@ -45,7 +45,8 @@ rocprofiler_systems_add_test(
ENVIRONMENT "${_base_environment};ROCPROFSYS_COUT_OUTPUT=ON"
BASELINE_PASS_REGEX "number of calls made = 17"
REWRITE_PASS_REGEX "\\[function\\]\\[Forcing\\] caller-include-regex :: 'outer'"
REWRITE_RUN_PASS_REGEX ">>> ._outer ([ \\|]+) 17")
REWRITE_RUN_PASS_REGEX ">>> ._outer ([ \\|]+) 17"
)
rocprofiler_systems_add_test(
NAME parallel-overhead
@@ -62,7 +63,8 @@ rocprofiler_systems_add_test(
return
args
RUN_ARGS 10 ${NUM_THREADS} 1000
ENVIRONMENT "${_base_environment}")
ENVIRONMENT "${_base_environment}"
)
rocprofiler_systems_add_test(
SKIP_BASELINE SKIP_RUNTIME
@@ -73,4 +75,4 @@ rocprofiler_systems_add_test(
RUN_ARGS 10 4 1000
ENVIRONMENT
"${_lock_environment};ROCPROFSYS_FLAT_PROFILE=ON;ROCPROFSYS_PROFILE=OFF;ROCPROFSYS_TRACE=ON;ROCPROFSYS_SAMPLING_KEEP_INTERNAL=OFF"
)
)
+12 -6
Vedi File
@@ -50,7 +50,8 @@ rocprofiler_systems_add_test(
ENVIRONMENT
"${_base_environment};ROCPROFSYS_USE_KOKKOSP=ON;ROCPROFSYS_COUT_OUTPUT=ON;ROCPROFSYS_SAMPLING_FREQ=50;ROCPROFSYS_KOKKOSP_PREFIX=[kokkos];KOKKOS_TOOLS_LIBS=librocprof-sys-dl.so"
REWRITE_RUN_PASS_REGEX "\\|_\\[kokkos\\] [a-zA-Z]"
RUNTIME_PASS_REGEX "\\|_\\[kokkos\\] [a-zA-Z]")
RUNTIME_PASS_REGEX "\\|_\\[kokkos\\] [a-zA-Z]"
)
rocprofiler_systems_add_test(
SKIP_RUNTIME SKIP_REWRITE
@@ -63,7 +64,8 @@ rocprofiler_systems_add_test(
RUN_ARGS -i 10 -s 20 -p
ENVIRONMENT
"${_base_environment};ROCPROFSYS_USE_KOKKOSP=ON;ROCPROFSYS_COUT_OUTPUT=ON;ROCPROFSYS_SAMPLING_FREQ=50;ROCPROFSYS_KOKKOSP_PREFIX=[kokkos];KOKKOS_TOOLS_LIBS=librocprof-sys.so"
BASELINE_PASS_REGEX "\\|_\\[kokkos\\] [a-zA-Z]")
BASELINE_PASS_REGEX "\\|_\\[kokkos\\] [a-zA-Z]"
)
rocprofiler_systems_add_test(
SKIP_RUNTIME SKIP_REWRITE
@@ -76,7 +78,8 @@ rocprofiler_systems_add_test(
RUN_ARGS -i 10 -s 20 -p
ENVIRONMENT
"${_base_environment};ROCPROFSYS_USE_KOKKOSP=ON;ROCPROFSYS_COUT_OUTPUT=ON;ROCPROFSYS_SAMPLING_FREQ=50;ROCPROFSYS_KOKKOSP_PREFIX=[kokkos];KOKKOS_TOOLS_LIBS=librocprof-sys-dl.so"
BASELINE_PASS_REGEX "\\|_\\[kokkos\\] [a-zA-Z]")
BASELINE_PASS_REGEX "\\|_\\[kokkos\\] [a-zA-Z]"
)
rocprofiler_systems_add_test(
SKIP_BASELINE
@@ -99,7 +102,8 @@ rocprofiler_systems_add_test(
-ME
[==[lib(gomp|m-)]==]
RUN_ARGS -i 10 -s 20 -p
ENVIRONMENT "${_base_environment};ROCPROFSYS_USE_KOKKOSP=ON")
ENVIRONMENT "${_base_environment};ROCPROFSYS_USE_KOKKOSP=ON"
)
rocprofiler_systems_add_test(
SKIP_BASELINE
@@ -121,7 +125,8 @@ rocprofiler_systems_add_test(
-ME
[==[libgomp]==]
RUN_ARGS -i 10 -s 20 -p
ENVIRONMENT "${_perfetto_environment};ROCPROFSYS_USE_KOKKOSP=OFF")
ENVIRONMENT "${_perfetto_environment};ROCPROFSYS_USE_KOKKOSP=OFF"
)
rocprofiler_systems_add_test(
NAME lulesh-timemory
@@ -143,4 +148,5 @@ rocprofiler_systems_add_test(
ROCPROFSYS_TIMEMORY_COMPONENTS="wall_clock peak_rss"
RUN_ARGS -i 10 -s 20 -p
ENVIRONMENT "${_timemory_environment};ROCPROFSYS_USE_KOKKOSP=OFF"
REWRITE_FAIL_REGEX "0 instrumented loops in procedure")
REWRITE_FAIL_REGEX "0 instrumented loops in procedure"
)
+19 -7
Vedi File
@@ -52,7 +52,7 @@ rocprofiler_systems_add_test(
"(/[A-Za-z-]+/perfetto-trace-0.proto).*(/[A-Za-z-]+/wall_clock-0.txt')"
REWRITE_RUN_FAIL_REGEX
"(perfetto-trace|trip_count|sampling_percent|sampling_cpu_clock|sampling_wall_clock|wall_clock)-[0-9][0-9]+.(json|txt|proto)|ROCPROFSYS_ABORT_FAIL_REGEX"
)
)
rocprofiler_systems_add_test(
SKIP_RUNTIME
@@ -75,7 +75,7 @@ rocprofiler_systems_add_test(
"${_flat_environment};ROCPROFSYS_USE_SAMPLING=OFF;ROCPROFSYS_STRICT_CONFIG=OFF;ROCPROFSYS_USE_MPIP=ON"
REWRITE_RUN_PASS_REGEX
">>> mpi-flat-mpip.inst(.*\n.*)>>> MPI_Init_thread(.*\n.*)>>> pthread_create(.*\n.*)>>> MPI_Comm_size(.*\n.*)>>> MPI_Comm_rank(.*\n.*)>>> MPI_Barrier(.*\n.*)>>> MPI_Alltoall"
)
)
rocprofiler_systems_add_test(
SKIP_RUNTIME
@@ -97,7 +97,7 @@ rocprofiler_systems_add_test(
ENVIRONMENT "${_flat_environment};ROCPROFSYS_USE_SAMPLING=OFF"
REWRITE_RUN_PASS_REGEX
">>> mpi-flat.inst(.*\n.*)>>> MPI_Init_thread(.*\n.*)>>> pthread_create(.*\n.*)>>> MPI_Comm_size(.*\n.*)>>> MPI_Comm_rank(.*\n.*)>>> MPI_Barrier(.*\n.*)>>> MPI_Alltoall"
)
)
set(_mpip_environment
"ROCPROFSYS_TRACE=ON"
@@ -111,7 +111,8 @@ set(_mpip_environment
"ROCPROFSYS_VERBOSE=2"
"ROCPROFSYS_DL_VERBOSE=2"
"${_test_openmp_env}"
"${_test_library_path}")
"${_test_library_path}"
)
set(_mpip_all2all_environment
"ROCPROFSYS_TRACE=ON"
@@ -125,9 +126,19 @@ set(_mpip_all2all_environment
"ROCPROFSYS_VERBOSE=3"
"ROCPROFSYS_DL_VERBOSE=3"
"${_test_openmp_env}"
"${_test_library_path}")
"${_test_library_path}"
)
foreach(_EXAMPLE all2all allgather allreduce bcast reduce scatter-gather send-recv)
foreach(
_EXAMPLE
all2all
allgather
allreduce
bcast
reduce
scatter-gather
send-recv
)
if("${_mpip_${_EXAMPLE}_environment}" STREQUAL "")
set(_mpip_${_EXAMPLE}_environment "${_mpip_environment}")
endif()
@@ -140,5 +151,6 @@ foreach(_EXAMPLE all2all allgather allreduce bcast reduce scatter-gather send-re
LABELS "mpip"
REWRITE_ARGS -e -v 2 --label file line --min-instructions 0
RUN_ARGS 30
ENVIRONMENT "${_mpip_${_EXAMPLE}_environment}")
ENVIRONMENT "${_mpip_${_EXAMPLE}_environment}"
)
endforeach()
+27 -21
Vedi File
@@ -30,7 +30,8 @@
execute_process(
COMMAND "${CMAKE_SOURCE_DIR}/tests/get_default_nic.sh"
OUTPUT_STRIP_TRAILING_WHITESPACE
OUTPUT_VARIABLE _network_interface)
OUTPUT_VARIABLE _network_interface
)
message(STATUS "Default network interface is ${_network_interface}")
@@ -46,23 +47,28 @@ set(_nic_perf_environment
"ROCPROFSYS_TIMEMORY_COMPONENTS=wall_clock,papi_array,network_stats"
"ROCPROFSYS_NETWORK_INTERFACE=${_network_interface}"
"ROCPROFSYS_PAPI_EVENTS=net:::${_network_interface}:tx:byte net:::${_network_interface}:rx:byte net:::${_network_interface}:rx:packet net:::${_network_interface}:tx:packet"
"ROCPROFSYS_SAMPLING_DELAY=0.05")
"ROCPROFSYS_SAMPLING_DELAY=0.05"
)
# Set _download_url to a large file that will give rocprof-sys-sample time to collect NIC
# performance data (but not too large, because each test will time out after 2 min).
set(_download_url
"https://github.com/ROCm/rocprofiler-systems/releases/download/rocm-6.4.1/rocprofiler-systems-1.0.1-ubuntu-22.04-ROCm-60400-PAPI-OMPT-Python3.sh"
)
)
# Run the NIC performance test
add_test(
NAME nic-performance
COMMAND $<TARGET_FILE:rocprofiler-systems-sample> -- wget --no-check-certificate
${_download_url} -O /tmp/rocprofiler-systems.test.bin
WORKING_DIRECTORY ${PROJECT_BINARY_DIR})
COMMAND
$<TARGET_FILE:rocprofiler-systems-sample> -- wget --no-check-certificate
${_download_url} -O /tmp/rocprofiler-systems.test.bin
WORKING_DIRECTORY ${PROJECT_BINARY_DIR}
)
set_tests_properties(nic-performance PROPERTIES ENVIRONMENT "${_nic_perf_environment}"
TIMEOUT 120 LABELS "network")
set_tests_properties(
nic-performance
PROPERTIES ENVIRONMENT "${_nic_perf_environment}" TIMEOUT 120 LABELS "network"
)
# Validate the perfetto file generated from NIC performance test output
add_test(
@@ -73,22 +79,22 @@ add_test(
${PROJECT_BINARY_DIR}/rocprof-sys-tests-output/nic-performance/perfetto-trace.proto
--counter-names rx:byte rx:packet tx:byte tx:packet -t
/opt/trace_processor/bin/trace_processor_shell -p
WORKING_DIRECTORY ${PROJECT_BINARY_DIR})
WORKING_DIRECTORY ${PROJECT_BINARY_DIR}
)
set(_test_pass_regex
"rocprof-sys-tests-output/nic-performance/perfetto-trace.proto validated")
"rocprof-sys-tests-output/nic-performance/perfetto-trace.proto validated"
)
set(_test_fail_regex
"Failure validating rocprof-sys-tests-output/nic-performance/perfetto-trace.proto")
"Failure validating rocprof-sys-tests-output/nic-performance/perfetto-trace.proto"
)
set_tests_properties(
validate-nic-performance-perfetto
PROPERTIES TIMEOUT
30
LABELS
"network"
DEPENDS
nic-performance
PASS_REGULAR_EXPRESSION
${_test_pass_regex}
FAIL_REGULAR_EXPRESSION
${_test_fail_regex})
PROPERTIES
TIMEOUT 30
LABELS "network"
DEPENDS nic-performance
PASS_REGULAR_EXPRESSION ${_test_pass_regex}
FAIL_REGULAR_EXPRESSION ${_test_fail_regex}
)
+21 -12
Vedi File
@@ -44,7 +44,8 @@ rocprofiler_systems_add_test(
"${_ompt_environment};ROCPROFSYS_USE_SAMPLING=OFF;ROCPROFSYS_COUT_OUTPUT=ON"
REWRITE_RUN_PASS_REGEX "${_OMPT_PASS_REGEX}"
RUNTIME_PASS_REGEX "${_OMPT_PASS_REGEX}"
REWRITE_FAIL_REGEX "0 instrumented loops in procedure")
REWRITE_FAIL_REGEX "0 instrumented loops in procedure"
)
rocprofiler_systems_add_test(
SKIP_RUNTIME
@@ -58,7 +59,8 @@ rocprofiler_systems_add_test(
ENVIRONMENT
"${_ompt_environment};ROCPROFSYS_USE_SAMPLING=ON;ROCPROFSYS_SAMPLING_FREQ=50;ROCPROFSYS_COUT_OUTPUT=ON"
REWRITE_RUN_PASS_REGEX "${_OMPT_PASS_REGEX}"
REWRITE_FAIL_REGEX "0 instrumented loops in procedure")
REWRITE_FAIL_REGEX "0 instrumented loops in procedure"
)
rocprofiler_systems_add_test(
SKIP_RUNTIME SKIP_REWRITE
@@ -67,7 +69,8 @@ rocprofiler_systems_add_test(
GPU ON
LABELS "openmp;openmp-target"
ENVIRONMENT
"${_ompt_environment};ROCPROFSYS_ROCM_DOMAINS=hip_runtime_api,kernel_dispatch")
"${_ompt_environment};ROCPROFSYS_ROCM_DOMAINS=hip_runtime_api,kernel_dispatch"
)
rocprofiler_systems_add_validation_test(
NAME openmp-target-sampling
@@ -86,7 +89,8 @@ rocprofiler_systems_add_validation_test(
0
0
0
-p)
-p
)
set(_ompt_sampling_environ
"${_ompt_environment}"
@@ -101,7 +105,8 @@ set(_ompt_sampling_environ
"ROCPROFSYS_SAMPLING_REALTIME=ON"
"ROCPROFSYS_SAMPLING_CPUTIME_FREQ=1000"
"ROCPROFSYS_SAMPLING_REALTIME_FREQ=500"
"ROCPROFSYS_MONOCHROME=ON")
"ROCPROFSYS_MONOCHROME=ON"
)
set(_ompt_sample_no_tmpfiles_environ
"${_ompt_environment}"
@@ -113,17 +118,18 @@ set(_ompt_sample_no_tmpfiles_environ
"ROCPROFSYS_SAMPLING_REALTIME=OFF"
"ROCPROFSYS_SAMPLING_CPUTIME_FREQ=700"
"ROCPROFSYS_USE_TEMPORARY_FILES=OFF"
"ROCPROFSYS_MONOCHROME=ON")
"ROCPROFSYS_MONOCHROME=ON"
)
set(_ompt_sampling_samp_regex
"Sampler for thread 0 will be triggered 1000.0x per second of CPU-time(.*)Sampler for thread 0 will be triggered 500.0x per second of wall-time(.*)Sampling will be disabled after 0.250000 seconds(.*)Sampling duration of 0.250000 seconds has elapsed. Shutting down sampling"
)
)
set(_ompt_sampling_file_regex
"sampling-duration-sampling/sampling_percent.(json|txt)(.*)sampling-duration-sampling/sampling_cpu_clock.(json|txt)(.*)sampling-duration-sampling/sampling_wall_clock.(json|txt)"
)
)
set(_notmp_sampling_file_regex
"sampling-no-tmp-files-sampling/sampling_percent.(json|txt)(.*)sampling-no-tmp-files-sampling/sampling_cpu_clock.(json|txt)(.*)sampling-no-tmp-files-sampling/sampling_wall_clock.(json|txt)"
)
)
rocprofiler_systems_add_test(
SKIP_BASELINE SKIP_RUNTIME SKIP_REWRITE
@@ -131,7 +137,8 @@ rocprofiler_systems_add_test(
TARGET openmp-cg
LABELS "openmp;sampling-duration"
ENVIRONMENT "${_ompt_sampling_environ}"
SAMPLING_PASS_REGEX "${_ompt_sampling_samp_regex}(.*)${_ompt_sampling_file_regex}")
SAMPLING_PASS_REGEX "${_ompt_sampling_samp_regex}(.*)${_ompt_sampling_file_regex}"
)
rocprofiler_systems_add_test(
SKIP_BASELINE SKIP_RUNTIME SKIP_REWRITE
@@ -139,7 +146,8 @@ rocprofiler_systems_add_test(
TARGET openmp-lu
LABELS "openmp;sampling-duration"
ENVIRONMENT "${_ompt_sampling_environ}"
SAMPLING_PASS_REGEX "${_ompt_sampling_samp_regex}(.*)${_ompt_sampling_file_regex}")
SAMPLING_PASS_REGEX "${_ompt_sampling_samp_regex}(.*)${_ompt_sampling_file_regex}"
)
rocprofiler_systems_add_test(
SKIP_BASELINE SKIP_RUNTIME SKIP_REWRITE
@@ -147,4 +155,5 @@ rocprofiler_systems_add_test(
TARGET openmp-cg
LABELS "openmp;no-tmp-files"
ENVIRONMENT "${_ompt_sample_no_tmpfiles_environ}"
SAMPLING_PASS_REGEX "${_notmp_sampling_file_regex}")
SAMPLING_PASS_REGEX "${_notmp_sampling_file_regex}"
)
+9 -5
Vedi File
@@ -34,11 +34,14 @@ set(_overflow_environment
"ROCPROFSYS_SAMPLING_OVERFLOW=ON"
"ROCPROFSYS_SAMPLING_OVERFLOW_EVENT=PERF_COUNT_SW_CPU_CLOCK"
"ROCPROFSYS_SAMPLING_OVERFLOW_FREQ=10000"
"ROCPROFSYS_DEBUG_THREADING_GET_ID=ON")
"ROCPROFSYS_DEBUG_THREADING_GET_ID=ON"
)
if(rocprofiler_systems_perf_event_paranoid LESS_EQUAL 3
OR rocprofiler_systems_cap_sys_admin EQUAL 0
OR rocprofiler_systems_cap_perfmon EQUAL 0)
if(
rocprofiler_systems_perf_event_paranoid LESS_EQUAL 3
OR rocprofiler_systems_cap_sys_admin EQUAL 0
OR rocprofiler_systems_cap_perfmon EQUAL 0
)
rocprofiler_systems_add_test(
SKIP_BASELINE
NAME overflow
@@ -50,5 +53,6 @@ if(rocprofiler_systems_perf_event_paranoid LESS_EQUAL 3
LABELS "perf;overflow"
SAMPLING_PASS_REGEX "sampling_wall_clock.txt"
RUNTIME_PASS_REGEX "sampling_wall_clock.txt"
REWRITE_RUN_PASS_REGEX "sampling_wall_clock.txt")
REWRITE_RUN_PASS_REGEX "sampling_wall_clock.txt"
)
endif()
+2 -2
Vedi File
@@ -39,7 +39,7 @@ rocprofiler_systems_add_test(
"wall_clock .*\\|_pthread_create .* 4 .*\\|_pthread_mutex_lock .* 1000 .*\\|_pthread_mutex_unlock .* 1000 .*\\|_pthread_mutex_lock .* 1000 .*\\|_pthread_mutex_unlock .* 1000 .*\\|_pthread_mutex_lock .* 1000 .*\\|_pthread_mutex_unlock .* 1000 .*\\|_pthread_mutex_lock .* 1000 .*\\|_pthread_mutex_unlock .* 1000"
RUNTIME_PASS_REGEX
"wall_clock .*\\|_pthread_create .* 4 .*\\|_pthread_mutex_lock .* 1000 .*\\|_pthread_mutex_unlock .* 1000 .*\\|_pthread_mutex_lock .* 1000 .*\\|_pthread_mutex_unlock .* 1000 .*\\|_pthread_mutex_lock .* 1000 .*\\|_pthread_mutex_unlock .* 1000 .*\\|_pthread_mutex_lock .* 1000 .*\\|_pthread_mutex_unlock .* 1000"
)
)
rocprofiler_systems_add_test(
SKIP_RUNTIME
@@ -53,4 +53,4 @@ rocprofiler_systems_add_test(
"${_lock_environment};ROCPROFSYS_FLAT_PROFILE=ON;ROCPROFSYS_PROFILE=ON;ROCPROFSYS_TRACE=OFF;ROCPROFSYS_SAMPLING_KEEP_INTERNAL=OFF"
REWRITE_RUN_PASS_REGEX
"start_thread (.*) 4 (.*) pthread_mutex_lock (.*) 4000 (.*) pthread_mutex_unlock (.*) 4000"
)
)
+51 -28
Vedi File
@@ -38,7 +38,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
rocprofiler_systems_find_python(
_PYTHON
ROOT_DIR "${_PYTHON_ROOT_DIR}"
COMPONENTS Interpreter)
COMPONENTS Interpreter
)
# ---------------------------------------------------------------------------------- #
# python tests
@@ -50,7 +51,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
FILE ${CMAKE_SOURCE_DIR}/examples/python/external.py
PROFILE_ARGS "--label" "file"
RUN_ARGS -v 10 -n 5
ENVIRONMENT "${_python_environment}")
ENVIRONMENT "${_python_environment}"
)
rocprofiler_systems_add_python_test(
NAME python-external-exclude-inefficient
@@ -59,7 +61,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
FILE ${CMAKE_SOURCE_DIR}/examples/python/external.py
PROFILE_ARGS -E "^inefficient$"
RUN_ARGS -v 10 -n 5
ENVIRONMENT "${_python_environment}")
ENVIRONMENT "${_python_environment}"
)
rocprofiler_systems_add_python_test(
NAME python-builtin
@@ -68,7 +71,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
FILE ${CMAKE_SOURCE_DIR}/examples/python/builtin.py
PROFILE_ARGS "-b" "--label" "file" "line"
RUN_ARGS -v 10 -n 5
ENVIRONMENT "${_python_environment}")
ENVIRONMENT "${_python_environment}"
)
rocprofiler_systems_add_python_test(
NAME python-builtin-noprofile
@@ -77,7 +81,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
FILE ${CMAKE_SOURCE_DIR}/examples/python/noprofile.py
PROFILE_ARGS "-b" "--label" "file"
RUN_ARGS -v 15 -n 5
ENVIRONMENT "${_python_environment}")
ENVIRONMENT "${_python_environment}"
)
rocprofiler_systems_add_python_test(
STANDALONE
@@ -86,7 +91,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
PYTHON_VERSION ${_VERSION}
FILE ${CMAKE_SOURCE_DIR}/examples/python/source.py
RUN_ARGS -v 5 -n 5 -s 3
ENVIRONMENT "${_python_environment}")
ENVIRONMENT "${_python_environment}"
)
rocprofiler_systems_add_python_test(
STANDALONE
@@ -104,16 +110,14 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
code-coverage-basic-blocks-binary-rewrite-run
code-coverage-basic-blocks-hybrid-runtime-instrument
LABELS "code-coverage"
ENVIRONMENT "${_python_environment}")
ENVIRONMENT "${_python_environment}"
)
# ---------------------------------------------------------------------------------- #
# python output tests
# ---------------------------------------------------------------------------------- #
if(CMAKE_VERSION VERSION_LESS "3.18.0")
find_program(
ROCPROFSYS_CAT_EXE
NAMES cat
PATH_SUFFIXES bin)
find_program(ROCPROFSYS_CAT_EXE NAMES cat PATH_SUFFIXES bin)
if(ROCPROFSYS_CAT_EXE)
set(ROCPROFSYS_CAT_COMMAND ${ROCPROFSYS_CAT_EXE})
@@ -131,7 +135,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
PASS_REGEX
"(\\\[compile\\\]).*(\\\| \\\|0>>> \\\[run\\\]\\\[external.py\\\]).*(\\\| \\\|0>>> \\\|_\\\[fib\\\]\\\[external.py\\\]).*(\\\| \\\|0>>> \\\|_\\\[inefficient\\\]\\\[external.py\\\])"
DEPENDS python-external-${_VERSION}
ENVIRONMENT "${_python_environment}")
ENVIRONMENT "${_python_environment}"
)
rocprofiler_systems_add_python_test(
NAME python-external-exclude-inefficient-check
@@ -140,7 +145,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
FILE rocprof-sys-tests-output/python-external-exclude-inefficient/${_VERSION}/trip_count.txt
FAIL_REGEX "(\\\|_inefficient).*(\\\|_sum)|ROCPROFSYS_ABORT_FAIL_REGEX"
DEPENDS python-external-exclude-inefficient-${_VERSION}
ENVIRONMENT "${_python_environment}")
ENVIRONMENT "${_python_environment}"
)
rocprofiler_systems_add_python_test(
NAME python-builtin-check
@@ -149,7 +155,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
FILE rocprof-sys-tests-output/python-builtin/${_VERSION}/trip_count.txt
PASS_REGEX "\\\[inefficient\\\]\\\[builtin.py:14\\\]"
DEPENDS python-builtin-${_VERSION}
ENVIRONMENT "${_python_environment}")
ENVIRONMENT "${_python_environment}"
)
rocprofiler_systems_add_python_test(
NAME python-builtin-noprofile-check
@@ -159,17 +166,23 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
PASS_REGEX ".(run)..(noprofile.py)."
FAIL_REGEX ".(fib|inefficient)..(noprofile.py).|ROCPROFSYS_ABORT_FAIL_REGEX"
DEPENDS python-builtin-noprofile-${_VERSION}
ENVIRONMENT "${_python_environment}")
ENVIRONMENT "${_python_environment}"
)
else()
rocprofiler_systems_message(
WARNING
"Neither 'cat' nor 'cmake -E cat' are available. Python source checks are disabled"
)
)
endif()
function(ROCPROFILER_SYSTEMS_ADD_PYTHON_VALIDATION_TEST)
cmake_parse_arguments(TEST "" "NAME;TIMEMORY_METRIC;TIMEMORY_FILE;PERFETTO_FILE"
"ARGS;PERFETTO_METRIC" ${ARGN})
cmake_parse_arguments(
TEST
""
"NAME;TIMEMORY_METRIC;TIMEMORY_FILE;PERFETTO_FILE"
"ARGS;PERFETTO_METRIC"
${ARGN}
)
rocprofiler_systems_add_python_test(
NAME ${TEST_NAME}-validate-timemory
@@ -181,7 +194,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
DEPENDS ${TEST_NAME}-${_VERSION}
PASS_REGEX
"rocprof-sys-tests-output/${TEST_NAME}/${_VERSION}/${TEST_TIMEMORY_FILE} validated"
ENVIRONMENT "${_python_environment}")
ENVIRONMENT "${_python_environment}"
)
rocprofiler_systems_add_python_test(
NAME ${TEST_NAME}-validate-perfetto
@@ -194,7 +208,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
DEPENDS ${TEST_NAME}-${_VERSION}
PASS_REGEX
"rocprof-sys-tests-output/${TEST_NAME}/${_VERSION}/${TEST_PERFETTO_FILE} validated"
ENVIRONMENT "${_python_environment}")
ENVIRONMENT "${_python_environment}"
)
endfunction()
set(python_source_labels
@@ -206,7 +221,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
fib
fib
inefficient
_sum)
_sum
)
set(python_source_count
5
3
@@ -216,7 +232,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
18
6
3
3)
3
)
set(python_source_depth
0
1
@@ -226,7 +243,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
5
6
2
3)
3
)
set(python_source_categories python user)
@@ -237,7 +255,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
PERFETTO_FILE "perfetto-trace.proto"
PERFETTO_METRIC ${python_source_categories}
ARGS -l ${python_source_labels} -c ${python_source_count} -d
${python_source_depth})
${python_source_depth}
)
set(python_builtin_labels
[run][builtin.py:28]
@@ -251,7 +270,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
[fib][builtin.py:10]
[fib][builtin.py:10]
[fib][builtin.py:10]
[inefficient][builtin.py:14])
[inefficient][builtin.py:14]
)
set(python_builtin_count
5
5
@@ -264,7 +284,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
220
80
10
5)
5
)
set(python_builtin_depth
0
1
@@ -277,7 +298,8 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
8
9
10
1)
1
)
rocprofiler_systems_add_python_validation_test(
NAME python-builtin
@@ -286,6 +308,7 @@ foreach(_VERSION ${ROCPROFSYS_PYTHON_VERSIONS})
PERFETTO_METRIC "python"
PERFETTO_FILE "perfetto-trace.proto"
ARGS -l ${python_builtin_labels} -c ${python_builtin_count} -d
${python_builtin_depth})
${python_builtin_depth}
)
math(EXPR _INDEX "${_INDEX} + 1")
endforeach()
+4 -2
Vedi File
@@ -81,12 +81,14 @@ foreach(_TARGET ${RCCL_TEST_TARGETS})
-z
-s
1
ENVIRONMENT "${_rccl_environment}")
ENVIRONMENT "${_rccl_environment}"
)
rocprofiler_systems_add_validation_test(
NAME rccl-test-${_NAME}-sampling
PERFETTO_METRIC "rocm_rccl_api"
PERFETTO_FILE "perfetto-trace.proto"
LABELS "rccl-tests;rcclp"
ARGS --counter-names "RCCL Comm" -p)
ARGS --counter-names "RCCL Comm" -p
)
endforeach()
+25 -11
Vedi File
@@ -45,7 +45,8 @@ rocprofiler_systems_add_test(
-E
uniform_int_distribution
ENVIRONMENT "${_base_environment}"
RUNTIME_TIMEOUT 480)
RUNTIME_TIMEOUT 480
)
rocprofiler_systems_add_test(
SKIP_REWRITE SKIP_RUNTIME
@@ -55,7 +56,8 @@ rocprofiler_systems_add_test(
GPU ON
NUM_PROCS 1
RUN_ARGS 1 2 2
ENVIRONMENT "${_base_environment}")
ENVIRONMENT "${_base_environment}"
)
rocprofiler_systems_add_test(
SKIP_BASELINE SKIP_RUNTIME
@@ -79,7 +81,8 @@ rocprofiler_systems_add_test(
uniform_int_distribution
RUN_ARGS 2 100 50
ENVIRONMENT "${_base_environment}"
REWRITE_FAIL_REGEX "0 instrumented loops in procedure transpose")
REWRITE_FAIL_REGEX "0 instrumented loops in procedure transpose"
)
if(ROCPROFSYS_USE_ROCM)
set(NAVI_REGEX "gfx(10|11|12)[A-Fa-f0-9][A-Fa-f0-9]")
@@ -91,12 +94,20 @@ if(ROCPROFSYS_USE_ROCM)
set(ROCPROFSYS_COUNTER_NAMES_ARG "SQ_WAVES")
else()
set(ROCPROFSYS_ROCM_EVENTS_TEST
"GRBM_COUNT,SQ_WAVES,SQ_INSTS_VALU,TA_TA_BUSY:device=0")
"GRBM_COUNT,SQ_WAVES,SQ_INSTS_VALU,TA_TA_BUSY:device=0"
)
set(ROCPROFSYS_FILE_CHECKS
"rocprof-device-0-GRBM_COUNT.txt" "rocprof-device-0-SQ_WAVES.txt"
"rocprof-device-0-SQ_INSTS_VALU.txt" "rocprof-device-0-TA_TA_BUSY.txt")
set(ROCPROFSYS_COUNTER_NAMES_ARG "GRBM_COUNT" "SQ_WAVES" "SQ_INSTS_VALU"
"TA_TA_BUSY")
"rocprof-device-0-GRBM_COUNT.txt"
"rocprof-device-0-SQ_WAVES.txt"
"rocprof-device-0-SQ_INSTS_VALU.txt"
"rocprof-device-0-TA_TA_BUSY.txt"
)
set(ROCPROFSYS_COUNTER_NAMES_ARG
"GRBM_COUNT"
"SQ_WAVES"
"SQ_INSTS_VALU"
"TA_TA_BUSY"
)
endif()
rocprofiler_systems_add_test(
@@ -111,19 +122,22 @@ if(ROCPROFSYS_USE_ROCM)
ENVIRONMENT
"${_base_environment};ROCPROFSYS_ROCM_EVENTS=${ROCPROFSYS_ROCM_EVENTS_TEST}"
REWRITE_RUN_PASS_REGEX "${_ROCP_PASS_REGEX}"
SAMPLING_PASS_REGEX "${_ROCP_PASS_REGEX}")
SAMPLING_PASS_REGEX "${_ROCP_PASS_REGEX}"
)
rocprofiler_systems_add_validation_test(
NAME transpose-rocprofiler-sampling
PERFETTO_FILE "perfetto-trace.proto"
ARGS --counter-names ${ROCPROFSYS_COUNTER_NAMES_ARG} -p
EXIST_FILES ${ROCPROFSYS_FILE_CHECKS}
LABELS "rocprofiler")
LABELS "rocprofiler"
)
rocprofiler_systems_add_validation_test(
NAME transpose-rocprofiler-binary-rewrite
PERFETTO_FILE "perfetto-trace.proto"
ARGS --counter-names ${ROCPROFSYS_COUNTER_NAMES_ARG} -p
EXIST_FILES ${ROCPROFSYS_FILE_CHECKS}
LABELS "rocprofiler")
LABELS "rocprofiler"
)
endif()
File diff soppresso perché troppo grande Carica Diff
+11 -6
Vedi File
@@ -33,7 +33,8 @@ rocprofiler_systems_add_test(
REWRITE_ARGS -e -v 2 --caller-include inner -i 4096
RUNTIME_ARGS -e -v 1 --caller-include inner -i 4096
LABELS "time-window"
ENVIRONMENT "${_window_environment};ROCPROFSYS_TRACE_DURATION=1.25")
ENVIRONMENT "${_window_environment};ROCPROFSYS_TRACE_DURATION=1.25"
)
rocprofiler_systems_add_validation_test(
NAME trace-time-window-binary-rewrite
@@ -58,7 +59,8 @@ rocprofiler_systems_add_validation_test(
1
1
1
-p)
-p
)
rocprofiler_systems_add_validation_test(
NAME trace-time-window-runtime-instrument
@@ -83,7 +85,8 @@ rocprofiler_systems_add_validation_test(
1
1
1
-p)
-p
)
rocprofiler_systems_add_test(
SKIP_BASELINE SKIP_SAMPLING ${_TRACE_WINDOW_SKIP}
@@ -94,7 +97,7 @@ rocprofiler_systems_add_test(
LABELS "time-window"
ENVIRONMENT
"${_window_environment};ROCPROFSYS_TRACE_DELAY=0.75;ROCPROFSYS_TRACE_DURATION=0.75"
)
)
rocprofiler_systems_add_validation_test(
NAME trace-time-window-delay-binary-rewrite
@@ -112,7 +115,8 @@ rocprofiler_systems_add_validation_test(
-d
0
0
-p)
-p
)
rocprofiler_systems_add_validation_test(
NAME trace-time-window-delay-runtime-instrument
@@ -130,4 +134,5 @@ rocprofiler_systems_add_validation_test(
-d
0
0
-p)
-p
)
+2 -1
Vedi File
@@ -50,4 +50,5 @@ rocprofiler_systems_add_test(
RUNTIME_PASS_REGEX "Pushing custom region :: run.10. x 1000"
SAMPLING_PASS_REGEX "Pushing custom region :: run.10. x 1000"
BASELINE_FAIL_REGEX "Pushing custom region"
REWRITE_FAIL_REGEX "0 instrumented loops in procedure")
REWRITE_FAIL_REGEX "0 instrumented loops in procedure"
)
+6 -3
Vedi File
@@ -38,14 +38,16 @@ set(_thread_limit_environment
"ROCPROFSYS_USE_SAMPLING=ON"
"ROCPROFSYS_SAMPLING_FREQ=250"
"ROCPROFSYS_VERBOSE=2"
"ROCPROFSYS_TIMEMORY_COMPONENTS=wall_clock,peak_rss,page_rss")
"ROCPROFSYS_TIMEMORY_COMPONENTS=wall_clock,peak_rss,page_rss"
)
math(EXPR THREAD_LIMIT_TEST_VALUE "${ROCPROFSYS_MAX_THREADS} + 24")
math(EXPR THREAD_LIMIT_TEST_VALUE_PLUS_ONE "${THREAD_LIMIT_TEST_VALUE} + 1")
set(_thread_limit_pass_regex "\\|${THREAD_LIMIT_TEST_VALUE}>>>")
set(_thread_limit_fail_regex
"\\|${THREAD_LIMIT_TEST_VALUE_PLUS_ONE}>>>|ROCPROFSYS_ABORT_FAIL_REGEX")
"\\|${THREAD_LIMIT_TEST_VALUE_PLUS_ONE}>>>|ROCPROFSYS_ABORT_FAIL_REGEX"
)
rocprofiler_systems_add_test(
SKIP_BASELINE
@@ -63,4 +65,5 @@ rocprofiler_systems_add_test(
RUNTIME_FAIL_REGEX "${_thread_limit_fail_regex}"
SAMPLING_FAIL_REGEX "${_thread_limit_fail_regex}"
REWRITE_RUN_FAIL_REGEX "${_thread_limit_fail_regex}"
ENVIRONMENT "${_thread_limit_environment}")
ENVIRONMENT "${_thread_limit_environment}"
)