Use GPUToolbox's Float64-free math overrides on devices without fp64

Project.toml

@@ -40,7 +40,7 @@ CEnum = "0.4, 0.5"
 ExprTools = "0.1"
 GPUArrays = "11.5.14"
 GPUCompiler = "2.8"
-GPUToolbox = "0.1, 0.2, 0.3, 1, 3"
+GPUToolbox = "3.1"
 KernelAbstractions = "0.9.39"
 LLVM = "6, 7, 8, 9"
 NEO_jll = "=26.18.38308"

src/compiler/compilation.jl

@@ -32,8 +32,16 @@ GPUCompiler.runtime_module(::oneAPICompilerJob) = oneAPI
 
 GPUCompiler.kernel_state_type(::oneAPICompilerJob) = KernelState
 
-GPUCompiler.method_table_view(job::oneAPICompilerJob) =
-    GPUCompiler.StackedMethodTable(job.world, method_table, SPIRVIntrinsics.method_table)
+function GPUCompiler.method_table_view(job::oneAPICompilerJob)
+    if job.config.target.supports_fp64
+        parent = SPIRVIntrinsics.method_table
+    else
+        # keep single-precision math that Base computes in Float64 out of double precision
+        parent = GPUCompiler.StackedMethodTable(job.world, SPIRVIntrinsics.method_table,
+                                                GPUToolbox.Overlays.float64_overrides)
+    end
+    GPUCompiler.StackedMethodTable(job.world, method_table, parent)
+end
 
 # filter out OpenCL built-ins
 # TODO: eagerly lower these using the translator API

src/device/quirks.jl

@@ -65,8 +65,4 @@ end
         @inline
         return checkindex(Bool, eachindex(IndexLinear(), v), i)
     end
-
-    # Less accurate division for Float32 than Base Julia which relies on Float64
-    # https://github.com/JuliaLang/julia/pull/49637
-    @device_override Base.div(x::Float32, y::Float32) = trunc(x / y)
 end

src/oneAPI.jl

@@ -4,6 +4,7 @@ using GPUArrays
 using Adapt
 
 using GPUCompiler
+import GPUToolbox
 
 import ExprTools