Merge #249

249: Stop overdubbing CUDA math functions r=vchuravy a=ali-ramadhan

I think with CUDA.jl v3 there is no longer a need to overdub/rewrite CUDA specific math functions.

If anything `CUDA.pow` doesn't exist anymore.

Co-authored-by: Ali Ramadhan <ali.hh.ramadhan@gmail.com>
Co-authored-by: ali.hh.ramadhan@gmail.com <ali.hh.ramadhan@gmail.com>

Author: bors[bot]

lib/CUDAKernels/src/CUDAKernels.jl

@@ -282,12 +282,6 @@ KernelAbstractions.generate_overdubs(@__MODULE__, CUDACtx)
 # CUDA specific method rewrites
 ###
 
-@inline Cassette.overdub(::CUDACtx, ::typeof(^), x::Float64, y::Float64) = CUDA.pow(x, y)
-@inline Cassette.overdub(::CUDACtx, ::typeof(^), x::Float32, y::Float32) = CUDA.pow(x, y)
-@inline Cassette.overdub(::CUDACtx, ::typeof(^), x::Float64, y::Int32)   = CUDA.pow(x, y)
-@inline Cassette.overdub(::CUDACtx, ::typeof(^), x::Float32, y::Int32)   = CUDA.pow(x, y)
-@inline Cassette.overdub(::CUDACtx, ::typeof(^), x::Union{Float32, Float64}, y::Int64) = CUDA.pow(x, y)
-
 # libdevice.jl
 const cudafuns = (:cos, :cospi, :sin, :sinpi, :tan,
           :acos, :asin, :atan,
@@ -302,7 +296,7 @@ const cudafuns = (:cos, :cospi, :sin, :sinpi, :tan,
 for f in cudafuns
     @eval function Cassette.overdub(ctx::CUDACtx, ::typeof(Base.$f), x::Union{Float32, Float64})
         @Base._inline_meta
-        return CUDA.$f(x)
+        return Base.$f(x)
     end
 end
 

test/compiler.jl

@@ -12,6 +12,10 @@ end
     A[1] = 2^11
 end
 
+@kernel function square(A, B)
+    A[1] = B[1]^2
+end
+
 @kernel function checked(A, a, b)
     A[1] = Base.Checked.checked_add(a, b)
 end
@@ -28,6 +32,11 @@ function compiler_testsuite()
         @test !any(stmt->(stmt isa Expr) && stmt.head == :invoke, CI.code)
     end
 
+    let (CI, rt) = @ka_code_typed square(CPU())(zeros(1), zeros(1), ndrange=1)
+        # test that there is no invoke of overdub
+        @test !any(stmt->(stmt isa Expr) && stmt.head == :invoke, CI.code)
+    end
+
     if VERSION >= v"1.5"
         let (CI, rt) = @ka_code_typed checked(CPU())(zeros(Int,1), 1, 2, ndrange=1)
             # test that there is no invoke of overdub