Local backend support

Author: christiangnrd

Project.toml

@@ -40,7 +40,7 @@ LLVM = "9.4.1"
 LinearAlgebra = "1.6"
 MacroTools = "0.5"
 PrecompileTools = "1"
-SPIRVIntrinsics = "0.5"
+SPIRVIntrinsics = "0.5.7"
 SPIRV_LLVM_Backend_jll = "20"
 SPIRV_Tools_jll = "2024.4, 2025.1"
 SparseArrays = "<0.0.1, 1.6"

src/pocl/backend.jl

@@ -7,6 +7,8 @@ using ..POCL: device, clconvert, clfunction
 import KernelAbstractions as KA
 import KernelAbstractions.KernelIntrinsics as KI
 
+import SPIRVIntrinsics
+
 import StaticArrays
 
 import Adapt
@@ -174,10 +176,36 @@ end
 function KI.max_work_group_size(::POCLBackend)::Int
     return Int(device().max_work_group_size)
 end
+function KI.sub_group_size(::POCLBackend)::Int
+    sg_sizes = cl.device().sub_group_sizes
+    if 32 in sg_sizes
+        return 32
+    elseif 64 in sg_sizes
+        return 64
+    elseif 16 in sg_sizes
+        return 16
+    else
+        return 1
+    end
+end
 function KI.multiprocessor_count(::POCLBackend)::Int
     return Int(device().max_compute_units)
 end
 
+function KI.shfl_down_types(::POCLBackend)
+    res = copy(SPIRVIntrinsics.gentypes)
+
+    backend_extensions = cl.device().extensions
+    if "cl_khr_fp64" ∉ backend_extensions
+        res = setdiff(res, [Float64])
+    end
+    if "cl_khr_fp16" ∉ backend_extensions
+        res = setdiff(res, [Float16])
+    end
+
+    return res
+end
+
 ## Indexing Functions
 
 @device_override @inline function KI.get_local_id()
@@ -204,6 +232,16 @@ end
     return (; x = Int(get_global_size(1)), y = Int(get_global_size(2)), z = Int(get_global_size(3)))
 end
 
+@device_override KI.get_sub_group_size() = get_sub_group_size()
+
+@device_override KI.get_max_sub_group_size() = get_max_sub_group_size()
+
+@device_override KI.get_num_sub_groups() = get_num_sub_groups()
+
+@device_override KI.get_sub_group_id() = get_sub_group_id()
+
+@device_override KI.get_sub_group_local_id() = get_sub_group_local_id()
+
 @device_override @inline function KA.__validindex(ctx)
     if KA.__dynamic_checkbounds(ctx)
         I = @inbounds KA.expand(KA.__iterspace(ctx), get_group_id(1), get_local_id(1))
@@ -232,6 +270,14 @@ end
     work_group_barrier(POCL.LOCAL_MEM_FENCE | POCL.GLOBAL_MEM_FENCE)
 end
 
+@device_override @inline function KI.sub_group_barrier()
+    sub_group_barrier(POCL.LOCAL_MEM_FENCE | POCL.GLOBAL_MEM_FENCE)
+end
+
+@device_override function KI.shfl_down(val::T, offset::Integer) where {T}
+    sub_group_shuffle(val, get_sub_group_local_id() + offset)
+end
+
 @device_override @inline function KI._print(args...)
     POCL._print(args...)
 end

src/pocl/compiler/compilation.jl

@@ -1,6 +1,9 @@
 ## gpucompiler interface
 
-struct OpenCLCompilerParams <: AbstractCompilerParams end
+Base.@kwdef struct OpenCLCompilerParams <: AbstractCompilerParams
+    sub_group_size::Int
+end
+
 const OpenCLCompilerConfig = CompilerConfig{SPIRVCompilerTarget, OpenCLCompilerParams}
 const OpenCLCompilerJob = CompilerJob{SPIRVCompilerTarget, OpenCLCompilerParams}
 
@@ -19,7 +22,21 @@ GPUCompiler.isintrinsic(job::OpenCLCompilerJob, fn::String) =
     in(fn, known_intrinsics) ||
     contains(fn, "__spirv_")
 
+function GPUCompiler.finish_module!(
+        @nospecialize(job::OpenCLCompilerJob),
+        mod::LLVM.Module, entry::LLVM.Function
+    )
+    entry = invoke(
+        GPUCompiler.finish_module!,
+        Tuple{CompilerJob{SPIRVCompilerTarget}, LLVM.Module, LLVM.Function},
+        job, mod, entry
+    )
+
+    # Set the subgroup size
+    metadata(entry)["intel_reqd_sub_group_size"] = MDNode([ConstantInt(Int32(job.config.params.sub_group_size))])
 
+    return entry
+end
 ## compiler implementation (cache, configure, compile, and link)
 
 # cache of compilation caches, per context
@@ -45,14 +62,17 @@ function compiler_config(dev::cl.Device; kwargs...)
     end
     return config
 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...)
     supports_fp16 = "cl_khr_fp16" in dev.extensions
     supports_fp64 = "cl_khr_fp64" in dev.extensions
 
+    if sub_group_size ∉ dev.sub_group_sizes
+        @error("$sub_group_size is not a valid sub-group size for this device.")
+    end
 
     # create GPUCompiler objects
     target = SPIRVCompilerTarget(; supports_fp16, supports_fp64, kwargs...)
-    params = OpenCLCompilerParams()
+    params = OpenCLCompilerParams(; sub_group_size)
     return CompilerConfig(target, params; kernel, name, always_inline)
 end
 

src/pocl/compiler/execution.jl

@@ -4,7 +4,7 @@ export @opencl, clfunction, clconvert
 ## high-level @opencl 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 = [:global_size, :local_size, :queue]
 
 macro opencl(ex...)

src/pocl/nanoOpenCL.jl

@@ -390,6 +390,8 @@ const CL_KERNEL_EXEC_INFO_SVM_PTRS = 0x11b6
 
 const CL_KERNEL_EXEC_INFO_SVM_FINE_GRAIN_SYSTEM = 0x11b7
 
+const CL_DEVICE_SUB_GROUP_SIZES_INTEL = 0x4108
+
 struct CLError <: Exception
     code::Cint
 end
@@ -935,6 +937,14 @@ devices(p::Platform) = devices(p, CL_DEVICE_TYPE_ALL)
         return tuple([Int(r) for r in result]...)
     end
 
+    if s == :sub_group_sizes
+        res_size = Ref{Csize_t}()
+        clGetDeviceInfo(d, CL_DEVICE_SUB_GROUP_SIZES_INTEL, C_NULL, C_NULL, res_size)
+        result = Vector{Csize_t}(undef, res_size[] ÷ sizeof(Csize_t))
+        clGetDeviceInfo(d, CL_DEVICE_SUB_GROUP_SIZES_INTEL, sizeof(result), result, C_NULL)
+        return tuple([Int(r) for r in result]...)
+    end
+
     if s == :max_image2d_shape
         width = Ref{Csize_t}()
         height = Ref{Csize_t}()

src/pocl/pocl.jl

@@ -41,7 +41,7 @@ function queue()
 end
 
 using GPUCompiler
-import LLVM
+import LLVM: LLVM, MDNode, ConstantInt, metadata
 using Adapt
 
 ## device overrides