Contextualization of kernels with Cassette

Author: jpsamaroo

Project.toml

@@ -5,6 +5,7 @@ version = "0.2.1"
 [deps]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
 BinaryProvider = "b99e7846-7c00-51b0-8f62-c81ae34c0232"
+Cassette = "7057c7e9-c182-5462-911a-8362d720325c"
 DataStructures = "864edb3b-99cc-5e75-8d2d-829cb0a9cfe8"
 HSARuntime = "2c364e2c-59fb-59c3-96f3-194112e690e0"
 InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"

src/compiler/common.jl

@@ -7,6 +7,8 @@ struct CompilerJob
     device::RuntimeDevice
     kernel::Bool
 
+    contextualize::Bool
+
     # optional properties
     minthreads::Union{Nothing,ROCDim}
     maxthreads::Union{Nothing,ROCDim}
@@ -16,8 +18,10 @@ struct CompilerJob
 
     CompilerJob(f, tt, device, kernel; name=nothing,
                 minthreads=nothing, maxthreads=nothing,
-                blocks_per_sm=nothing, maxregs=nothing) =
-        new(f, tt, device, kernel, minthreads, maxthreads, blocks_per_sm,
+                blocks_per_sm=nothing, maxregs=nothing,
+                contextualize=true) =
+        new(f, tt, device, kernel, contextualize,
+            minthreads, maxthreads, blocks_per_sm,
             maxregs, name)
 end
 

src/compiler/driver.jl

@@ -51,11 +51,12 @@ function codegen(target::Symbol, job::CompilerJob; libraries::Bool=true,
     @timeit to[] "validation" check_method(job)
 
     @timeit to[] "Julia front-end" begin
+        f = job.contextualize ? contextualize(job.f) : job.f
 
         # get the method instance
         world = typemax(UInt)
-        meth = which(job.f, job.tt)
-        sig = Base.signature_type(job.f, job.tt)::Type
+        meth = which(f, job.tt)
+        sig = Base.signature_type(f, job.tt)::Type
         (ti, env) = ccall(:jl_type_intersection_with_env, Any,
                           (Any, Any), sig, meth.sig)::Core.SimpleVector
         if VERSION >= v"1.2.0-DEV.320"

src/context.jl

@@ -0,0 +1,75 @@
+##
+# Implements contextual dispatch through Cassette.jl
+# Goals:
+# - Rewrite common CPU functions to appropriate GPU intrinsics
+#
+# TODO:
+# - error (erf, ...)
+# - pow
+# - min, max
+# - mod, rem
+# - gamma
+# - bessel
+# - distributions
+# - unsorted
+
+using Cassette
+
+function transform(ctx, ref)
+    CI = ref.code_info
+    noinline = any(@nospecialize(x) ->
+                       Core.Compiler.isexpr(x, :meta) &&
+                       x.args[1] == :noinline,
+                   CI.code)
+    CI.inlineable = !noinline
+
+    CI.ssavaluetypes = length(CI.code)
+    # Core.Compiler.validate_code(CI)
+    return CI
+end
+
+const InlinePass = Cassette.@pass transform
+
+Cassette.@context ROCCtx
+const rocctx = Cassette.disablehooks(ROCCtx(pass = InlinePass))
+
+###
+# Cassette fixes
+###
+
+# kwfunc fix
+Cassette.overdub(::ROCCtx, ::typeof(Core.kwfunc), f) = return Core.kwfunc(f)
+
+# the functions below are marked `@pure` and by rewritting them we hide that from
+# inference so we leave them alone (see https://github.com/jrevels/Cassette.jl/issues/108).
+@inline Cassette.overdub(::ROCCtx, ::typeof(Base.isimmutable), x)     = return Base.isimmutable(x)
+@inline Cassette.overdub(::ROCCtx, ::typeof(Base.isstructtype), t)    = return Base.isstructtype(t)
+@inline Cassette.overdub(::ROCCtx, ::typeof(Base.isprimitivetype), t) = return Base.isprimitivetype(t)
+@inline Cassette.overdub(::ROCCtx, ::typeof(Base.isbitstype), t)      = return Base.isbitstype(t)
+@inline Cassette.overdub(::ROCCtx, ::typeof(Base.isbits), x)          = return Base.isbits(x)
+
+@inline Cassette.overdub(::ROCCtx, ::typeof(datatype_align), ::Type{T}) where {T} = datatype_align(T)
+
+###
+# Rewrite functions
+###
+Cassette.overdub(ctx::ROCCtx, ::typeof(isdevice)) = true
+
+# libdevice.jl
+for f in (:cos, :cospi, :sin, :sinpi, :tan,
+          :acos, :asin, :atan,
+          :cosh, :sinh, :tanh,
+          :acosh, :asinh, :atanh,
+          :log, :log10, :log1p, :log2,
+          :exp, :exp2, :exp10, :expm1, :ldexp,
+          :isfinite, :isinf, :isnan,
+          :signbit, :abs,
+          :sqrt, :cbrt,
+          :ceil, :floor,)
+    @eval function Cassette.overdub(ctx::ROCCtx, ::typeof(Base.$f), x::Union{Float32, Float64})
+        @Base._inline_meta
+        return AMDGPUnative.$f(x)
+    end
+end
+
+contextualize(f::F) where F = (args...) -> Cassette.overdub(rocctx, f, args...)

src/execution.jl

@@ -12,7 +12,7 @@ end
 # affecting the compiler, kernel execution, or both.
 function split_kwargs(kwargs)
     # TODO: Alias groupsize and gridsize as threads and blocks, respectively
-    compiler_kws = [:device, :agent, :queue, :name]
+    compiler_kws = [:device, :agent, :queue, :name, :contextualize]
     call_kws     = [:groupsize, :gridsize, :device, :agent, :queue]
     compiler_kwargs = []
     call_kwargs = []
@@ -199,7 +199,8 @@ Low-level interface to compile a function invocation for the currently-active
 GPU, returning a callable kernel object. For a higher-level interface, use
 [`@roc`](@ref).
 
-Currently, no keyword arguments are implemented.
+The following keyword arguments are supported:
+- `contextualize`: whether to contextualize functions using Cassette (default: true)
 
 The output of this function is automatically cached, i.e. you can simply call
 `rocfunction` in a hot path without degrading performance. New code will be