Use stacked method tables (#615)

Author: vchuravy

Project.toml

@@ -34,13 +34,13 @@ SparseArraysExt = "SparseArrays"
 Adapt = "0.4, 1.0, 2.0, 3.0, 4"
 Atomix = "0.1, 1"
 EnzymeCore = "0.7, 0.8.1"
-GPUCompiler = "1.2"
+GPUCompiler = "1.6"
 InteractiveUtils = "1.6"
 LLVM = "9.4.1"
 LinearAlgebra = "1.6"
 MacroTools = "0.5"
 PrecompileTools = "1"
-SPIRVIntrinsics = "0.3"
+SPIRVIntrinsics = "0.4"
 SPIRV_LLVM_Backend_jll = "20"
 SPIRV_Tools_jll = "2024.4, 2025.1"
 SparseArrays = "<0.0.1, 1.6"

src/pocl/backend.jl

@@ -1,7 +1,7 @@
 module POCLKernels
 
 using ..POCL
-using ..POCL: @device_override, SPIRVIntrinsics, cl
+using ..POCL: @device_override, SPIRVIntrinsics, cl, method_table
 using ..POCL: device
 
 import KernelAbstractions as KA

src/pocl/compiler/compilation.jl

@@ -6,7 +6,7 @@ const OpenCLCompilerJob = CompilerJob{SPIRVCompilerTarget, OpenCLCompilerParams}
 
 GPUCompiler.runtime_module(::CompilerJob{<:Any, OpenCLCompilerParams}) = POCL
 
-GPUCompiler.method_table(::OpenCLCompilerJob) = method_table
+GPUCompiler.method_table_view(job::OpenCLCompilerJob) = GPUCompiler.StackedMethodTable(job.world, method_table, SPIRVIntrinsics.method_table)
 
 # filter out OpenCL built-ins
 # TODO: eagerly lower these using the translator API
@@ -50,7 +50,7 @@ end
 
 
     # create GPUCompiler objects
-    target = SPIRVCompilerTarget(; supports_fp16, supports_fp64, version = v"1.2", kwargs...)
+    target = SPIRVCompilerTarget(; supports_fp16, supports_fp64, kwargs...)
     params = OpenCLCompilerParams()
     return CompilerConfig(target, params; kernel, name, always_inline)
 end

src/pocl/pocl.jl

@@ -44,9 +44,12 @@ using GPUCompiler
 import LLVM
 using Adapt
 
+## device overrides
 import SPIRVIntrinsics
 SPIRVIntrinsics.@import_all
 SPIRVIntrinsics.@reexport_public
+# local method table for device functions
+Base.Experimental.@MethodTable(method_table)
 
 include("compiler/compilation.jl")
 include("compiler/execution.jl")