use DISCARD for fill_policy since default RING_BUFFER fails when wrap around (most result table empty); add README.md and merge-trace.jl

[ROCm/rocprofiler-systems commit: 7793a1f331]

projects/rocprofiler-systems/README.md

@@ -0,0 +1,63 @@
+# hosttrace: application tracing with static/dynamic binary instrumentation and perfetto
+
+
+## 1. load necessary modules
+```
+module load julia
+module load dyninst
+module load rocm/VERSION
+```
+
+## 2. to install julia packages, only need it the first time.
+```
+$ julia
+type ']' to get into pkg mode, then 
+(@v1.6) pkg> add JSON, DataFrames, Dates, CSV, Chain, PrettyTables
+```
+
+## 3. hosttrace usage
+```
+export PATH=$PATH:$HTRACE_PATH/bin
+hosttrace --help
+```
+
+## 4. instrument the binaries
+```
+hosttrace -L $HTRACE_PATH/bin/libhosttrace.so -o app.inst -- path_to_your_app
+
+hosttrace -L $HTRACE_PATH/bin/libhosttrace.so -E 'hipApiName|hipGetCmdName' -o libamdhip64.so.4 --  /opt/rocm-VERSION/lib/libamdhip64.so.4
+
+hosttrace -L $HTRACE_PATH/bin/libhosttrace.so -E 'rocr::atomic|rocr::core|rocr::HSA' -o libhsa-runtime64.so.1 --  /opt/rocm-VERSION/lib/libhsa-runtime64.so.1
+```
+## 5. run the app
+```
+export LD_LIBRARY_PATH=.:$LD_LIBRARY_PATH
+rocprof --hip-trace --roctx-trace --stats ./app.inst
+```
+
+## 6. you may need to increase the buffer size if the hosttrace.perfetto-trace file is close to 1GB. for example to set it to 10GB. 
+```
+export HOSTTRACE_BUFFER_SIZE_KB=10240000
+```
+
+## 7. merge the traces from rocprof and hosttrace
+```
+julia $HTRACE_PATH/bin/merge-trace.jl results.json hosttrace.perfetto-trace*
+```
+
+## 8. another mode of Perfetto tracing is to use system backend. To do it:
+
+### in a separate window
+```
+    pkill traced; traced --background; perfetto --out ./htrace.out --txt -c $HTRACE/roctrace.cfg
+```
+
+### then in the app running window do this before running rocprof or other app
+```
+    export HOSTTRACE_BACKEND_SYSTEM=1
+```
+
+### for the merge use the htrace.out
+```
+julia $HTRACE_PATH/bin/merge-trace.jl results.json htrace.out
+```

projects/rocprofiler-systems/scripts/merge-trace.jl

@@ -0,0 +1,37 @@
+using Base: String
+using JSON, DataFrames, Dates, CSV
+
+function check_input_args()
+    message = "Needs only one json file and at least one trace data file(s)."
+    @assert length(ARGS) >= 2 && occursin(".json", join(ARGS)) message
+    json_file, trace_files = nothing, String[]
+    for arg in ARGS
+        @assert Base.Filesystem.ispath(arg) "'$(arg)' not found!"
+        if endswith(arg, ".json")
+            json_file = arg
+        else
+            push!(trace_files, arg)
+        end
+    end
+    json_file, trace_files
+end
+
+json_file, trace_files = check_input_args()
+
+out = JSON.parsefile(json_file)
+for trace in trace_files
+    @info "processing $(trace) ..."
+    f = tempname() * ".json"
+    run(`trace_to_text json $trace $f`)
+    todo = JSON.parsefile(f)
+    out["traceEvents"] = vcat(
+        out["traceEvents"],
+        todo["traceEvents"])
+end
+
+output_file = replace("$(string(now())).json", ":" => "-")
+open(output_file, "w") do f
+    write(f, JSON.json(out))
+    @info "output file: $(output_file)"
+end
+

projects/rocprofiler-systems/src/library.cpp

@@ -221,7 +221,10 @@ hosttrace_init_perfetto()
     perfetto::TraceConfig                   cfg{};
     perfetto::protos::gen::TrackEventConfig track_event_cfg{};
 
-    cfg.add_buffers()->set_size_kb(buffer_size);
+    auto *buffer_config = cfg.add_buffers();
+    buffer_config->set_size_kb(buffer_size);
+    buffer_config->set_fill_policy(perfetto::protos::gen::TraceConfig_BufferConfig_FillPolicy_DISCARD);
+
     auto* ds_cfg = cfg.add_data_sources()->mutable_config();
     ds_cfg->set_name("track_event");
     ds_cfg->set_track_event_config_raw(track_event_cfg.SerializeAsString());