KernelInterface

Project.toml

@@ -13,6 +13,7 @@ GPUArrays = "0c68f7d7-f131-5f86-a1c3-88cf8149b2d7"
 GPUCompiler = "61eb1bfa-7361-4325-ad38-22787b887f55"
 GPUToolbox = "096a3bc2-3ced-46d0-87f4-dd12716f4bfc"
 KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
+KernelInterface = "4ee993da-d684-4d17-a7dd-4e58e78d92bf"
 LLVM = "929cbde3-209d-540e-8aea-75f648917ca0"
 Libdl = "8f399da3-3557-5675-b5ff-fb832c97cbdb"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
@@ -42,6 +43,7 @@ GPUArrays = "11.2.1"
 GPUCompiler = "2 - 2.2.1"
 GPUToolbox = "0.1, 0.2, 0.3, 1, 3"
 KernelAbstractions = "0.9.39"
+KernelInterface = "0.1"
 LLVM = "6, 7, 8, 9"
 NEO_jll = "=26.18.38308"
 PrecompileTools = "1"

src/compiler/compilation.jl

@@ -1,6 +1,9 @@
 ## gpucompiler interface implementation
 
-struct oneAPICompilerParams <: AbstractCompilerParams end
+Base.@kwdef struct oneAPICompilerParams <: AbstractCompilerParams
+    sub_group_size::Union{Nothing,Int} = nothing
+end
+
 const oneAPICompilerConfig = CompilerConfig{SPIRVCompilerTarget, oneAPICompilerParams}
 const oneAPICompilerJob = CompilerJob{SPIRVCompilerTarget,oneAPICompilerParams}
 
@@ -48,6 +51,11 @@ function GPUCompiler.finish_module!(job::oneAPICompilerJob, mod::LLVM.Module,
                    Tuple{CompilerJob{SPIRVCompilerTarget}, typeof(mod), typeof(entry)},
                    job, mod, entry)
 
+    # Set the subgroup size
+    if job.config.params.sub_group_size !== nothing
+        metadata(entry)["intel_reqd_sub_group_size"] = MDNode([ConstantInt(Int32(job.config.params.sub_group_size))])
+    end
+
     # OpenCL 2.0
     push!(metadata(mod)["opencl.ocl.version"],
           MDNode([ConstantInt(Int32(2)),
@@ -323,11 +331,16 @@ function _driver_supports_bfloat16_spirv(dev=device())
     end
 end
 
-@noinline function _compiler_config(dev; kernel=true, name=nothing, always_inline=false, kwargs...)
+@noinline function _compiler_config(dev; kernel=true, name=nothing, always_inline=false, sub_group_size=32, kwargs...)
     properties = oneL0.module_properties(dev)
     supports_fp16 = properties.fp16flags & oneL0.ZE_DEVICE_MODULE_FLAG_FP16 == oneL0.ZE_DEVICE_MODULE_FLAG_FP16
     supports_fp64 = properties.fp64flags & oneL0.ZE_DEVICE_MODULE_FLAG_FP64 == oneL0.ZE_DEVICE_MODULE_FLAG_FP64
 
+    if sub_group_size ∉ oneL0.compute_properties(dev).subGroupSizes
+        @error("$sub_group_size is not a valid sub-group size for this device.")
+    end
+
+
     # SPIR-V codegen path. The Aurora LTS NEO/IGC runtime only accepts SPIR-V from the
     # Khronos translator; the rolling stack uses the LLVM SPIR-V back-end. GPUCompiler picks
     # the tool from the target's `backend` field and loads the JLL lazily, so both can be
@@ -360,7 +373,7 @@ end
 
     # create GPUCompiler objects
     target = SPIRVCompilerTarget(; backend, extensions = extensions_str, supports_fp16, supports_fp64, supports_bfloat16, kwargs...)
-    params = oneAPICompilerParams()
+    params = oneAPICompilerParams(; sub_group_size)
     CompilerConfig(target, params; kernel, name, always_inline)
 end
 

src/compiler/execution.jl

@@ -4,7 +4,7 @@ export @oneapi, zefunction, kernel_convert
 ## high-level @oneapi interface
 
 const MACRO_KWARGS = [:launch]
-const COMPILER_KWARGS = [:kernel, :name, :always_inline]
+const COMPILER_KWARGS = [:kernel, :name, :always_inline, :sub_group_size]
 const LAUNCH_KWARGS = [:groups, :items, :queue]
 
 """

src/oneAPI.jl

@@ -12,6 +12,7 @@ using SpecialFunctions
 import Preferences
 
 import KernelAbstractions: KernelAbstractions
+import KernelInterface
 
 using LLVM
 using LLVM.Interop
@@ -76,12 +77,18 @@ include("gpuarrays.jl")
 include("random.jl")
 include("utils.jl")
 
-include("oneAPIKernels.jl")
+# KernelAbstractions
+include("oneAPIKernelsOld.jl")
 import .oneAPIKernels: oneAPIBackend
+export oneAPIBackend
+
+# KernelInterface
+include("oneAPIKernels.jl")
+import .oneAPIInterface
+
 include("accumulate.jl")
 include("sorting.jl")
 include("indexing.jl")
-export oneAPIBackend
 
 # precompilation workload (warms up the SPIR-V compilation pipeline)
 include("compiler/precompile.jl")