feat: profile macro to provide profiling data in the terminal

CondaPkg.toml

@@ -5,3 +5,4 @@ python = "<=3.13,>=3.9,<4"
 jax = ">= 0.6"
 tensorflow = ">= 2.17"
 numpy = ">= 2"
+xprof = ">= 2.20"

Project.toml

@@ -13,13 +13,15 @@ EnzymeCore = "f151be2c-9106-41f4-ab19-57ee4f262869"
 Functors = "d9f16b24-f501-4c13-a1f2-28368ffc5196"
 GPUArraysCore = "46192b85-c4d5-4398-a991-12ede77f4527"
 HTTP = "cd3eb016-35fb-5094-929b-558a96fad6f3"
+JSON3 = "0f8b85d8-7281-11e9-16c2-39a750bddbf1"
 LLVM = "929cbde3-209d-540e-8aea-75f648917ca0"
 LLVMOpenMP_jll = "1d63c593-3942-5779-bab2-d838dc0a180e"
 Libdl = "8f399da3-3557-5675-b5ff-fb832c97cbdb"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 OrderedCollections = "bac558e1-5e72-5ebc-8fee-abe8a469f55d"
 PrecompileTools = "aea7be01-6a6a-4083-8856-8a6e6704d82a"
 Preferences = "21216c6a-2e73-6563-6e65-726566657250"
+PrettyTables = "08abe8d2-0d0c-5749-adfa-8a2ac140af0d"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 ReactantCore = "a3311ec8-5e00-46d5-b541-4f83e724a433"
 Reactant_jll = "0192cb87-2b54-54ad-80e0-3be72ad8a3c0"
@@ -89,6 +91,7 @@ Functors = "0.5"
 GPUArraysCore = "0.2"
 GPUCompiler = "1.3"
 HTTP = "1.10.15"
+JSON3 = "1.14.3"
 KernelAbstractions = "0.9.30"
 LLVM = "9.1"
 LLVMOpenMP_jll = "18.1.7"
@@ -101,6 +104,7 @@ OneHotArrays = "0.2.10"
 OrderedCollections = "1"
 PrecompileTools = "1.2"
 Preferences = "1.4.3"
+PrettyTables = "3.1.0"
 PythonCall = "0.9.25"
 Random = "1.10"
 Random123 = "1.7"

ext/ReactantPythonCallExt/ReactantPythonCallExt.jl

@@ -15,6 +15,11 @@ const npptr = Ref{Py}()
 
 const SAVED_MODEL_EXPORT_SUPPORTED = Ref{Bool}(false)
 
+const xprofptr = Ref{Py}()
+const xprofconvertptr = Ref{Py}()
+
+const XPROF_PROFILER_SUPPORTED = Ref{Bool}(false)
+
 const NUMPY_SIMPLE_TYPES = Dict(
     Bool => :bool,
     Int8 => :int8,
@@ -54,11 +59,23 @@ function __init__()
                tensorflow SavedModel will not be \
                supported." exception = (err, catch_backtrace())
     end
+
+    try
+        xprofptr[] = pyimport("xprof")
+        xprofconvertptr[] = pyimport("xprof.convert.raw_to_tool_data")
+        XPROF_PROFILER_SUPPORTED[] = true
+    catch err
+        @warn "Failed to import xprof. Xprof support will not be available." exception = (
+            err, catch_backtrace()
+        )
+    end
+
     return nothing
 end
 
 include("overlays.jl")
 include("pycall.jl")
 include("saved_model.jl")
+include("xprof.jl")
 
 end

ext/ReactantPythonCallExt/xprof.jl

@@ -0,0 +1,14 @@
+# Currently prototyping with xprof via python. we should instead add this into
+# the C++ API.
+function xspace_to_tools_data(filename::String, tool_name::String)
+    if !XPROF_PROFILER_SUPPORTED[]
+        error("xprof is not supported...")
+    end
+
+    return String(
+        pyconvert(
+            Vector{UInt8},
+            xprofconvertptr[].xspace_to_tool_data(pylist([filename]), tool_name, pydict())[0],
+        ),
+    )
+end

src/Profiler.jl

@@ -2,6 +2,8 @@ module Profiler
 
 import ..Reactant
 using Sockets: Sockets
+using PrettyTables: PrettyTables
+using JSON3: JSON3
 
 """
     with_profiler(f, trace_output_dir::String; trace_device=true, trace_host=true, create_perfetto_link=false)
@@ -195,4 +197,160 @@ mutable struct ProfileServer
     end
 end
 
+function wrap_string(s; width=20)
+    s_str = string(s)
+    return join([s_str[i:min(i + width - 1, end)] for i in 1:width:length(s_str)], "\n")
+end
+
+struct KernelStatsProfileResults
+    data
+end
+
+function KernelStatsProfileResults(data::JSON3.Object)
+    cols = data["cols"]
+    rows = data["rows"]
+    keys = Tuple(Symbol.(get.(cols, "id")))
+    table = Vector{NamedTuple}(undef, length(rows))
+
+    for (i, row) in enumerate(rows)
+        vals = get.(row["c"], "v")
+        table[i] = NamedTuple{keys}(vals)
+    end
+
+    return KernelStatsProfileResults(table)
+end
+
+function Base.show(io::IO, r::KernelStatsProfileResults)
+    tbl = r.data
+
+    println(io, "╔════════════════╗")
+    println(io, "║  Kernel Stats  ║")
+    println(io, "╚════════════════╝")
+
+    isempty(tbl) && return nothing
+
+    fields = fieldnames(typeof(tbl[1]))
+    wrapped = split.(wrap_string.(fields; width=10), "\n")
+    nrows = maximum(length.(wrapped))
+    column_labels = [[get(wrapped[j], i, "") for j in 1:length(wrapped)] for i in 1:nrows]
+
+    PrettyTables.pretty_table(
+        io,
+        tbl;
+        line_breaks=true,
+        maximum_data_column_widths=10,
+        auto_wrap=true,
+        column_labels,
+    )
+    return nothing
+end
+
+struct FrameworkStatsProfileResults
+    data
+end
+
+function FrameworkStatsProfileResults(data::JSON3.Array{JSON3.Object})
+    results = Vector{Vector{NamedTuple}}()
+
+    for table in data
+        local_result = Vector{NamedTuple}()
+
+        # Extract column information
+        cols = table["cols"]
+        col_ids = [Symbol(col["id"]) for col in cols]
+
+        # Extract rows
+        rows = table["rows"]
+
+        # Parse each row into a NamedTuple
+        for row in rows
+            values = [cell["v"] for cell in row["c"]]
+            nt = NamedTuple{Tuple(col_ids)}(Tuple(values))
+            push!(local_result, nt)
+        end
+
+        push!(results, local_result)
+    end
+
+    return FrameworkStatsProfileResults(results)
+end
+
+function Base.show(io::IO, r::FrameworkStatsProfileResults)
+    println(io, "╔══════════════════════════════╗")
+    println(io, "║  FrameworkOpStatsResults     ║")
+    println(io, "╚══════════════════════════════╝")
+
+    isempty(r.data) && return nothing
+
+    for tbl in r.data
+        fields = fieldnames(typeof(tbl[1]))
+        wrapped = split.(wrap_string.(fields; width=10), "\n")
+        nrows = maximum(length.(wrapped))
+        column_labels = [
+            [get(wrapped[j], i, "") for j in 1:length(wrapped)] for i in 1:nrows
+        ]
+
+        PrettyTables.pretty_table(
+            io,
+            tbl;
+            line_breaks=true,
+            auto_wrap=true,
+            maximum_data_column_widths=10,
+            column_labels,
+        )
+    end
+
+    return nothing
+end
+
+struct ReactantProfileResults
+    kernel_stats::KernelStatsProfileResults
+    framework_stats::FrameworkStatsProfileResults
+end
+
+function Base.show(io::IO, r::ReactantProfileResults)
+    println(io, "╔═══════════════════════════════════════════════════════╗")
+    println(io, "║ Reactant Profile Results                              ║")
+    println(io, "╚═══════════════════════════════════════════════════════╝")
+    show(io, r.kernel_stats)
+    println(io)
+    show(io, r.framework_stats)
+    return nothing
+end
+
+function parse_xprof_profile_data(data)
+    extmod = Base.get_extension(Reactant, :ReactantPythonCallExt)
+    if extmod === nothing
+        error("Currently we require `PythonCall` to be loaded to parse xprof data.")
+    end
+    kernel_stats = KernelStatsProfileResults(
+        JSON3.read(extmod.xspace_to_tools_data(data, "kernel_stats"))
+    )
+    framework_stats = FrameworkStatsProfileResults(
+        JSON3.read(extmod.xspace_to_tools_data(data, "framework_op_stats"))
+    )
+    return ReactantProfileResults(kernel_stats, framework_stats)
+    return nothing
+end
+
+macro profile(ex)
+    profile_dir = joinpath(tempdir(), "reactant_profile")
+    mkpath(profile_dir)
+
+    quote
+        # TODO: optionally compile the code first and profile
+
+        Reactant.Profiler.with_profiler($(esc(profile_dir))) do
+            $(esc(ex))
+        end
+
+        trace_output_dir = joinpath($(esc(profile_dir)), "plugins", "profile")
+        date = maximum(readdir(trace_output_dir))
+        traces_path = joinpath(trace_output_dir, date)
+
+        filename = first(f for f in readdir(traces_path) if endswith(f, ".xplane.pb"))
+        data = $(parse_xprof_profile_data)(joinpath(traces_path, filename))
+    end
+end
+
 end # module Profiler