Improve NSight Systems activation by inspecting the session list.

[maleadt]

Fixes #2629

[github-actions]

Your PR requires formatting changes to meet the project's style guidelines.
Please consider running Runic (git runic master) to apply these changes.

Click here to view the suggested changes.

diff --git a/src/profile.jl b/src/profile.jl
index 15c98ddf1..cbfc38295 100644
--- a/src/profile.jl
+++ b/src/profile.jl
@@ -169,7 +169,7 @@ function find_nsys()
         return ENV["_"]
     else
         # look at a couple of environment variables that may point to NSight
-        nsys = nothing
+            nsys = nothing
         for var in ("LD_PRELOAD", "CUDA_INJECTION64_PATH", "NVTX_INJECTION64_PATH")
             haskey(ENV, var) || continue
             for val in split(ENV[var], Sys.iswindows() ? ';' : ':')
@@ -186,56 +186,56 @@ function find_nsys()
     error("Running under Nsight Systems, but could not find the `nsys` binary to start the profiler. Please specify using JULIA_CUDA_NSYS=path/to/nsys, and file an issue with the contents of ENV.")
 end
 
-const __nsys = Ref{Union{Nothing,String}}()
-function nsys()
-    if !isassigned(__nsys)
+    const __nsys = Ref{Union{Nothing, String}}()
+    function nsys()
+        if !isassigned(__nsys)
         # find the active Nsight Systems profiler
         if haskey(ENV, "NSYS_PROFILING_SESSION_ID") && ccall(:jl_generating_output, Cint, ()) == 0
-            __nsys[] = find_nsys()
+                __nsys[] = find_nsys()
         else
-            __nsys[] = nothing
+                __nsys[] = nothing
         end
     end
 
-    __nsys[]
+        return __nsys[]
 end
 
-function nsys_sessions()
-    sessions = Dict{Int,Dict{String,String}}()
-    open(`$(nsys()) sessions list`, "r") do io
-        header = Dict()
-        for line in eachline(io)
-            # parse the header
-            if isempty(header)
-                @assert startswith(line, r"\s+ID")
-                colnames = split(line)[1:end-1] # ignore the final left-aligned column
-                colranges = []
-                for column in colnames
-                    push!(colranges, findfirst(Regex("\\s+\\b$column\\b"), line))
-                end
-                for (name, range) in zip(colnames, colranges)
-                    header[name] = range
-                end
+    function nsys_sessions()
+        sessions = Dict{Int, Dict{String, String}}()
+        open(`$(nsys()) sessions list`, "r") do io
+            header = Dict()
+            for line in eachline(io)
+                # parse the header
+                if isempty(header)
+                    @assert startswith(line, r"\s+ID")
+                    colnames = split(line)[1:(end - 1)] # ignore the final left-aligned column
+                    colranges = []
+                    for column in colnames
+                        push!(colranges, findfirst(Regex("\\s+\\b$column\\b"), line))
+                    end
+                    for (name, range) in zip(colnames, colranges)
+                        header[name] = range
+                    end
 
-            # parse the data
-            else
-                session = Dict()
-                for (name, range) in header
-                    session[name] = lstrip(line[range])
-                end
+                    # parse the data
+                else
+                    session = Dict()
+                    for (name, range) in header
+                        session[name] = lstrip(line[range])
+                    end
 
-                id = parse(Int, session["ID"])
-                delete!(session, "ID")
-                sessions[id] = session
+                    id = parse(Int, session["ID"])
+                    delete!(session, "ID")
+                    sessions[id] = session
+                end
             end
         end
+        return sessions
     end
-    return sessions
-end
 
-nsys_session() = parse(Int, ENV["NSYS_PROFILING_SESSION_ID"])
+    nsys_session() = parse(Int, ENV["NSYS_PROFILING_SESSION_ID"])
 
-nsys_state() = nsys_sessions()[nsys_session()]["STATE"]
+    nsys_state() = nsys_sessions()[nsys_session()]["STATE"]
 
 
 
@@ -246,50 +246,50 @@ Enables profile collection by the active profiling tool for the current context.
 profiling is already enabled, then this call has no effect.
 """
 function start()
-    if nsys() !== nothing
-        # by default, running under NSight Systems does not activate the profiler API-based
-        # ranged collection; that's done by calling `nsys start --capture-range=cudaProfilerApi`.
-        # however, as of recent we cannot do this anymore when already running under the
-        # capturing `nsys profile`, so we need to detect the state and act accordingly.
-        try
-            state = nsys_state()
-
-            # `nsys profile`
-            if state == "Collection"
-                @warn """The application is already being profiled; starting the profiler is a no-op.
-
-                         If you meant to profile a specific region, make sure to start NSight Systems in
-                         delayed mode (`nsys profile --start-later=true --capture-range=cudaProfilerApi`)
-                         or simply switch to the interactive `nsys launch` command."""
-                return
-
-            # `nsys profile --start-later=true`
-            elseif state == "DelayedCollection"
-                @error """The application is running under a delayed profiling session which CUDA.jl cannot activate.
-
-                          If you want `CUDA.@profile` to enable the profiler, make sure
-                          to pass `--capture-range=cudaProfilerApi` to `nsys profile`."""
-                return
-
-            # `nsys profile --start-later=true --capture-range=cudaProfilerApi`
-            elseif state == "StartRange"
-
-            # `nsys launch`
-            elseif state == "Launched"
-                run(`$(nsys()) start --capture-range=cudaProfilerApi`)
+        if nsys() !== nothing
+            # by default, running under NSight Systems does not activate the profiler API-based
+            # ranged collection; that's done by calling `nsys start --capture-range=cudaProfilerApi`.
+            # however, as of recent we cannot do this anymore when already running under the
+            # capturing `nsys profile`, so we need to detect the state and act accordingly.
+            try
+                state = nsys_state()
+
+                # `nsys profile`
+                if state == "Collection"
+                    @warn """The application is already being profiled; starting the profiler is a no-op.
+
+                    If you meant to profile a specific region, make sure to start NSight Systems in
+                    delayed mode (`nsys profile --start-later=true --capture-range=cudaProfilerApi`)
+                    or simply switch to the interactive `nsys launch` command."""
+                    return
+
+                    # `nsys profile --start-later=true`
+                elseif state == "DelayedCollection"
+                    @error """The application is running under a delayed profiling session which CUDA.jl cannot activate.
+
+                    If you want `CUDA.@profile` to enable the profiler, make sure
+                    to pass `--capture-range=cudaProfilerApi` to `nsys profile`."""
+                    return
+
+                    # `nsys profile --start-later=true --capture-range=cudaProfilerApi`
+                elseif state == "StartRange"
+
+                    # `nsys launch`
+                elseif state == "Launched"
+                    run(`$(nsys()) start --capture-range=cudaProfilerApi`)
 
-            else
-                error("Unexpected state: $state")
+                else
+                    error("Unexpected state: $state")
+                end
+            catch err
+                @error "Failed to find the active profiling session ($(nsys_session())) in the session list:\n" * read(`$(nsys()) sessions list`, String) * "\n\nPlease file an issue." exception = (err, catch_backtrace())
             end
-        catch err
-            @error "Failed to find the active profiling session ($(nsys_session())) in the session list:\n" * read(`$(nsys()) sessions list`, String) * "\n\nPlease file an issue." exception=(err,catch_backtrace())
-        end
 
-        # it takes a while for the profiler to attach to our process
+            # it takes a while for the profiler to attach to our process
         sleep(0.01)
     end
 
-    # actually start the capture
+        # actually start the capture
     CUDA.cuProfilerStart()
 end
 
@@ -300,7 +300,7 @@ Disables profile collection by the active profiling tool for the current context
 profiling is already disabled, then this call has no effect.
 """
 function stop()
-    CUDA.cuProfilerStop()
+        return CUDA.cuProfilerStop()
 end
 
 

[codecov]

Codecov Report

Attention: Patch coverage is 10.86957% with 41 lines in your changes missing coverage. Please review.

Project coverage is 73.36%. Comparing base (24c236a) to head (a71b1bf).
Report is 3 commits behind head on master.

Files with missing lines	Patch %	Lines
src/profile.jl	10.86%	41 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2638      +/-   ##
==========================================
- Coverage   73.53%   73.36%   -0.18%     
==========================================
  Files         157      157              
  Lines       15238    15272      +34     
==========================================
- Hits        11205    11204       -1     
- Misses       4033     4068      +35     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[github-actions]

CUDA.jl Benchmarks

Benchmark suite	Current: a71b1bf	Previous: 24c236a	Ratio
latency/precompile	46377879057 ns	46329110276 ns	1.00
latency/ttfp	6994478571 ns	6970088608 ns	1.00
latency/import	3638981184 ns	3628117061 ns	1.00
integration/volumerhs	9624929 ns	9622596.5 ns	1.00
integration/byval/slices=1	147109 ns	147068 ns	1.00
integration/byval/slices=3	425287 ns	425521 ns	1.00
integration/byval/reference	145172 ns	144904 ns	1.00
integration/byval/slices=2	286178 ns	286164 ns	1.00
integration/cudadevrt	103567 ns	103422 ns	1.00
kernel/indexing	14350 ns	14042 ns	1.02
kernel/indexing_checked	14927 ns	14520 ns	1.03
kernel/occupancy	633.3550295857988 ns	633.7076023391813 ns	1.00
kernel/launch	2188.777777777778 ns	2101.9 ns	1.04
kernel/rand	14846 ns	14379 ns	1.03
array/reverse/1d	19733 ns	19719 ns	1.00
array/reverse/2d	25320 ns	25121 ns	1.01
array/reverse/1d_inplace	10574 ns	11160 ns	0.95
array/reverse/2d_inplace	12253 ns	13070 ns	0.94
array/copy	21532 ns	21024 ns	1.02
array/iteration/findall/int	157929.5 ns	155786 ns	1.01
array/iteration/findall/bool	137134.5 ns	134662.5 ns	1.02
array/iteration/findfirst/int	154486.5 ns	147148 ns	1.05
array/iteration/findfirst/bool	153973 ns	154167.5 ns	1.00
array/iteration/scalar	64196.5 ns	61499 ns	1.04
array/iteration/logical	208380 ns	203811.5 ns	1.02
array/iteration/findmin/1d	39462 ns	39639 ns	1.00
array/iteration/findmin/2d	94142 ns	94387 ns	1.00
array/reductions/reduce/1d	38671.5 ns	30507 ns	1.27
array/reductions/reduce/2d	51283.5 ns	51213.5 ns	1.00
array/reductions/mapreduce/1d	38292 ns	30459 ns	1.26
array/reductions/mapreduce/2d	51530.5 ns	51487 ns	1.00
array/broadcast	21140.5 ns	20729 ns	1.02
array/copyto!/gpu_to_gpu	12008 ns	11560 ns	1.04
array/copyto!/cpu_to_gpu	210181 ns	208930 ns	1.01
array/copyto!/gpu_to_cpu	243254.5 ns	241934 ns	1.01
array/accumulate/1d	109436 ns	108332 ns	1.01
array/accumulate/2d	80310 ns	80060 ns	1.00
array/construct	1340.7 ns	1258.3 ns	1.07
array/random/randn/Float32	44338.5 ns	42993 ns	1.03
array/random/randn!/Float32	27063 ns	26226 ns	1.03
array/random/rand!/Int64	27383 ns	26878 ns	1.02
array/random/rand!/Float32	8812.833333333332 ns	8569 ns	1.03
array/random/rand/Int64	38260.5 ns	29957 ns	1.28
array/random/rand/Float32	13367 ns	12962 ns	1.03
array/permutedims/4d	61382 ns	61080 ns	1.00
array/permutedims/2d	55577 ns	55180 ns	1.01
array/permutedims/3d	56451 ns	55780 ns	1.01
array/sorting/1d	2766078 ns	2774685 ns	1.00
array/sorting/by	3369788 ns	3367411.5 ns	1.00
array/sorting/2d	1084807 ns	1085055 ns	1.00
cuda/synchronization/stream/auto	1092.6 ns	1048.8 ns	1.04
cuda/synchronization/stream/nonblocking	6449.8 ns	6455.4 ns	1.00
cuda/synchronization/stream/blocking	830.5952380952381 ns	846.986301369863 ns	0.98
cuda/synchronization/context/auto	1225.9 ns	1230.1 ns	1.00
cuda/synchronization/context/nonblocking	6732.4 ns	6670.8 ns	1.01
cuda/synchronization/context/blocking	973.5833333333334 ns	928 ns	1.05

This comment was automatically generated by workflow using github-action-benchmark.

[luraess]

Thanks! This now execute as expected when profiling a block of code in non-interactive mode

nsys profile --start-later=true --capture-range=cudaProfilerApi --capture-range-end=stop --trace=cuda,nvtx julia --project test.jl