Add Buffer Device Address support (#311)

Co-authored-by: Tim Besard <[email protected]>

Author: VarLad

.github/workflows/Test.yml

@@ -13,7 +13,7 @@ concurrency:
 
 jobs:
   test:
-    name: Julia ${{ matrix.version }} - ${{ matrix.os }} - ${{ matrix.arch }} - PoCL ${{ matrix.pocl }}
+    name: Julia ${{ matrix.version }} - ${{ matrix.os }} - ${{ matrix.arch }} - ${{ matrix.memory_backend }} - PoCL ${{ matrix.pocl }}
     runs-on: ${{ matrix.os }}
     timeout-minutes: 180
     permissions: # needed to allow julia-actions/cache to proactively delete old caches that it has created
@@ -26,7 +26,9 @@ jobs:
         os: [ubuntu-24.04, ubuntu-24.04-arm, macOS-13, macOS-15, windows-2025]
         arch: [x64, arm64]
         pocl: [jll, local]
+        memory_backend: [usm, svm, buffer]
         exclude:
+          # unsupported combinations
           - os: ubuntu-24.04
             arch: arm64
           - os: windows-2025
@@ -125,11 +127,13 @@ jobs:
                   run(```$(cmake()) --build $builddir --parallel $(Sys.CPU_THREADS) --target install```)
               end'
 
-            echo '[pocl_jll]' > test/LocalPreferences.toml
+            echo '[pocl_jll]' >> test/LocalPreferences.toml
             echo 'libpocl_path="${{ github.workspace }}/target/lib/libpocl.so"' >> test/LocalPreferences.toml
 
       - name: Setup OpenCL.jl
         run: |
+          echo '[OpenCL]' >> test/LocalPreferences.toml
+          echo 'default_memory_backend="${{ matrix.memory_backend }}"' >> test/LocalPreferences.toml
           julia --project -e '
             using Pkg
             Pkg.develop(path="lib/intrinsics")'

.gitignore

@@ -1,2 +1 @@
 Manifest.toml
-LocalPreferences.toml

LocalPreferences.toml

@@ -0,0 +1,7 @@
+[OpenCL]
+# Which memory back-end to use for unspecified CLArray allocations. This can be:
+# - "buffer": plain buffers (using pointers if `cl_ext_buffer_device_address` is available)
+# - "usm": Unified Shared Memory (requiring `cl_intel_unified_shared_memory`)
+# - "svm": Shared Virtual Memory (requiring coarse-grained SVM support)
+# If unspecified, the default will be used based on the platform and device capabilities.
+#default_memory_backend="..."

Project.toml

@@ -10,6 +10,7 @@ KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
 LLVM = "929cbde3-209d-540e-8aea-75f648917ca0"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 OpenCL_jll = "6cb37087-e8b6-5417-8430-1f242f1e46e4"
+Preferences = "21216c6a-2e73-6563-6e65-726566657250"
 Printf = "de0858da-6303-5e67-8744-51eddeeeb8d7"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 Reexport = "189a3867-3050-52da-a836-e630ba90ab69"
@@ -26,6 +27,7 @@ KernelAbstractions = "0.9.2"
 LLVM = "9.1"
 LinearAlgebra = "1"
 OpenCL_jll = "=2024.10.24"
+Preferences = "1"
 Printf = "1"
 Random = "1"
 Reexport = "1"

lib/cl/CL.jl

@@ -1,6 +1,8 @@
 module cl
 
+import ..OpenCL
 using Printf
+using Preferences
 
 include("pointer.jl")
 include("api.jl")
@@ -18,8 +20,7 @@ include("device.jl")
 include("context.jl")
 include("cmdqueue.jl")
 include("event.jl")
-include("memory/memory.jl")
-include("buffer.jl")
+include("memory.jl")
 include("program.jl")
 include("kernel.jl")
 

lib/cl/device.jl

@@ -190,10 +190,7 @@ function exec_capabilities(d::Device)
     )
 end
 
-function usm_supported(d::Device)
-    "cl_intel_unified_shared_memory" in d.extensions || return false
-    return true
-end
+usm_supported(d::Device) = "cl_intel_unified_shared_memory" in d.extensions
 
 function usm_capabilities(d::Device)
     usm_supported(d) || throw(ArgumentError("Unified Shared Memory not supported on this device"))
@@ -256,6 +253,8 @@ function svm_capabilities(d::Device)
     )
 end
 
+bda_supported(d::Device) = "cl_ext_buffer_device_address" in d.extensions
+
 function cl_device_type(dtype::Symbol)
     if dtype == :all
         cl_dtype = CL_DEVICE_TYPE_ALL

lib/cl/kernel.jl

@@ -69,7 +69,7 @@ function set_arg!(k::Kernel, idx::Integer, arg::CLPtr{T}) where {T}
 end
 
 # raw memory
-function set_arg!(k::Kernel, idx::Integer, arg::AbstractMemory)
+function set_arg!(k::Kernel, idx::Integer, arg::AbstractPointerMemory)
     # XXX: this assumes that the receiving argument is pointer-typed, which is not the case
     #      with Julia's `Ptr` ABI. Instead, one should reinterpret the pointer as a
     #      `Core.LLVMPtr`, which _is_ pointer-valued. We retain this handling for `Ptr` for
@@ -79,6 +79,8 @@ function set_arg!(k::Kernel, idx::Integer, arg::AbstractMemory)
         clSetKernelArgSVMPointer(k, idx - 1, pointer(arg))
     elseif arg isa UnifiedMemory
         clSetKernelArgMemPointerINTEL(k, idx - 1, pointer(arg))
+    elseif arg isa Buffer
+        clSetKernelArgDevicePointerEXT(k, idx - 1, pointer(arg))
     else
         error("Unknown memory type")
     end
@@ -191,6 +193,7 @@ function call(
     if !isempty(indirect_memory)
         svm_pointers = CLPtr{Cvoid}[]
         usm_pointers = CLPtr{Cvoid}[]
+        bda_pointers = CLPtr{Cvoid}[]
         device_access = host_access = shared_access = false
         for memory in indirect_memory
             ptr = pointer(memory)
@@ -200,6 +203,8 @@ function call(
 
             if memory isa SharedVirtualMemory
                 push!(svm_pointers, ptr)
+            elseif memory isa Buffer
+                push!(bda_pointers, ptr)
             elseif memory isa UnifiedDeviceMemory
                 device_access = true
                 push!(usm_pointers, ptr)
@@ -229,6 +234,9 @@ function call(
         if !isempty(svm_pointers)
             clSetKernelExecInfo(k, CL_KERNEL_EXEC_INFO_SVM_PTRS, sizeof(svm_pointers), svm_pointers)
         end
+        if !isempty(bda_pointers)
+            clSetKernelExecInfo(k, CL_KERNEL_EXEC_INFO_DEVICE_PTRS_EXT, sizeof(bda_pointers), bda_pointers)
+        end
         if !isempty(usm_pointers)
             clSetKernelExecInfo(k, CL_KERNEL_EXEC_INFO_USM_PTRS_INTEL, sizeof(usm_pointers), usm_pointers)
         end