implement NVTXT

src/KernelAbstractions.jl

@@ -430,6 +430,15 @@ end
     end
 end
 
+###
+# Extras
+# - LoopInfo
+# - Timeline
+###
+
+include("extras/extras.jl")
+import .Extras.Timeline
+
 ###
 # Backends/Implementation
 ###
@@ -442,11 +451,4 @@ include("backends/cpu.jl")
 @init @require CUDAnative="be33ccc6-a3ff-5ff2-a52e-74243cff1e17" begin
     include("backends/cuda.jl")
 end
-
-###
-# Extras
-# - LoopInfo
-###
-
-include("extras/extras.jl")
 end #module

src/backends/cpu.jl

@@ -8,8 +8,11 @@ end
 
 function Event(f, args...; dependencies=nothing, progress=nothing)
     T = Threads.@spawn begin
-        wait(MultiEvent(dependencies), progress)
-        f(args...)
+        Timeline.range "Event($(nameof(f))" begin
+            wait(MultiEvent(dependencies), progress)
+            Timeline.mark("Event($(nameof(f))) waiting done")
+            f(args...)
+        end
     end
     return CPUEvent(T)
 end
@@ -79,13 +82,15 @@ function __run(obj, ndrange, iterspace, args, ::Val{dynamic}) where dynamic
         Nthreads = N
         len, rem = 1, 0
     end
+    Timeline.@range string(nameof(obj.f)) begin
     if Nthreads == 1
         __thread_run(1, len, rem, obj, ndrange, iterspace, args, Val(dynamic))
     else
         @sync for tid in 1:Nthreads
             Threads.@spawn __thread_run(tid, len, rem, obj, ndrange, iterspace, args, Val(dynamic))
         end
     end
+    end # Timeline
     return nothing
 end
 

src/extras/extras.jl

@@ -4,4 +4,8 @@ include("loopinfo.jl")
 using .LoopInfo
 export @unroll
 
+include("timeline.jl")
+using .Timeline
+export Timeline
+
 end # module

src/extras/timeline.jl

@@ -0,0 +1,135 @@
+module Timeline
+
+using Requires
+export @range, mark
+
+module NVTXT
+    const LOG_FILE=Ref{IOStream}()
+    const SHOULD_LOG=Ref{Bool}(false)
+
+    function __init__()
+        if haskey(ENV, "KERNELABSTRACTIONS_TIMELINE")
+            SHOULD_LOG[] = true
+        else
+            SHOULD_LOG[] = false
+            return
+        end
+        pid = Libc.getpid()
+        LOG_FILE[] = open("ka-$pid.nvtxt", "w")
+        initialize()
+        atexit() do
+            close(LOG_FILE[])
+        end
+    end
+
+    function initialize()
+        SHOULD_LOG[] || return
+        io = LOG_FILE[]
+        pid = Libc.getpid()
+        print(io, """
+        SetFileDisplayName, KernelAbstractions
+        @RangeStartEnd, Start, End, ThreadId, Message
+        ProcessId = $pid
+        CategoryId = 1
+        Color = Blue
+        TimeBase = Manual
+        @RangePush, Time, ThreadId, Message
+        ProcessId = $pid
+        CategoryId = 1
+        Color = Blue
+        TimeBase = Manual
+        @RangePop, Time, ThreadId
+        ProcessId = $pid
+        TimeBase = Manual
+        @Marker, Time, ThreadId, Message
+        ProcessId = $pid
+        CategoryId = 1
+        Color = Blue
+        TimeBase = Manual
+        """)
+    end
+
+    function push_range(msg)
+        SHOULD_LOG[] || return
+        time = time_ns()
+        io = LOG_FILE[]
+        print(io, "RangePush, ")
+        print(io, time)
+        println(io, ", ", Base.Threads.threadid(), ", \"", msg, "\"")
+    end
+
+    function pop_range()
+        SHOULD_LOG[] || return
+        time = time_ns()
+        io = LOG_FILE[]
+        print(io, "RangePop, ")
+        print(io, time)
+        println(io, ", ", Base.Threads.threadid())
+    end
+
+    struct Range
+        start::UInt64
+        msg::String
+    end
+
+    start_range(msg::String) = Range(time_ns(), msg)
+    function end_range(r::Range)
+        SHOULD_LOG[] || return
+        time = time_ns()
+        io = LOG_FILE[]
+        print(io, "RangeStartEnd, ")
+        show(io, r.start)
+        print(io, ", ")
+        show(io, time)
+        println(io, ", ", Base.Threads.threadid(), ", \"", r.msg, "\"")
+    end
+
+    function mark(msg::String)
+        SHOULD_LOG[] || return
+        time = time_ns()
+        io = LOG_FILE[]
+        print(io, "Marker, ")
+        show(io, time)
+        println(io, ", ", Base.Threads.threadid(), ", \"", msg, "\"")
+    end
+end # NVTXT
+
+_mark(msg) = NVTXT.mark(msg)
+_push_range(msg) = NVTXT.push_range(msg)
+_pop_range() = NVTXT.pop_range()
+_start_range(msg) = NVTXT.start_range(msg)
+_end_range(r) = NVTXT.end_range(r)
+
+@init @require CUDAnative="be33ccc6-a3ff-5ff2-a52e-74243cff1e17" begin
+    # replace implementations
+    import CUDAnative.NVTX
+
+    _mark(msg) = NVTX.mark(msg)
+    _push_range(msg) = NVTX.push_range(msg)
+    _pop_range() = NVTX.pop_range()
+    _start_range(msg) = NVTX.start_range(msg)
+    _end_range(r) = NVTX.end_range(r)
+end
+
+import Base: invokelatest
+mark(msg) = invokelatest(_mark, msg)
+push_range(msg) = invokelatest(_push_range, msg)
+pop_range() = invokelatest(_pop_range)
+start_range(msg) = invokelatest(_start_range, msg)
+end_range(r) = invokelatest(_end_range, r)
+
+"""
+    @range "msg" ex
+Create a new range and execute `ex`. The range is popped automatically afterwards.
+See also: [`range`](@ref)
+"""
+macro range(msg, ex)
+    quote
+        local range = $start_range($(esc(msg)))
+        local ret = $(esc(ex))
+        $end_range(range)
+        ret
+    end
+end
+
+end