Merge pull request #438 from JuliaGPU/tb/bump

Bump toolchain.

Author: maleadt

Project.toml

@@ -34,7 +34,7 @@ GPUArrays = "10"
 GPUCompiler = "0.23, 0.24, 0.25, 0.26"
 KernelAbstractions = "0.9.1"
 LLVM = "6"
-NEO_jll = "=24.09.28717"
+NEO_jll = "=24.13.29138"
 Preferences = "1"
 SPIRV_LLVM_Translator_unified_jll = "0.4"
 SpecialFunctions = "1.3, 2"

lib/level-zero/module.jl

@@ -237,7 +237,10 @@ function properties(kernel::ZeKernel)
     preferred_group_size_props_ref = Ref(ze_kernel_preferred_group_size_properties_t())
     link_extensions(props_ref, preferred_group_size_props_ref)
     if haskey(oneL0.extension_properties(kernel.mod.context.driver),
-              "ZE_extension_kernel_max_group_size_properties")
+              "ZE_extension_kernel_max_group_size_properties") ||
+       (!validation_layer[] &&  # intel/compute-runtime#733
+        properties(kernel.mod.device).vendorId == 0x8086 &&
+        properties(kernel.mod.context.driver).driverVersion >= v"1.3.29138")
         # TODO: memoize
         max_group_size_props_ref = Ref(ze_kernel_max_group_size_properties_ext_t())
         link_extensions(preferred_group_size_props_ref, max_group_size_props_ref)

lib/level-zero/oneL0.jl

@@ -101,6 +101,9 @@ include("residency.jl")
 
 const functional = Ref{Bool}(false)
 
+const validation_layer = Ref{Bool}()
+const parameter_validation = Ref{Bool}()
+
 function __init__()
     precompiling = ccall(:jl_generating_output, Cint, ()) != 0
     precompiling && return
@@ -132,6 +135,9 @@ function __init__()
         functional[] = false
         return
     end
+
+    validation_layer[] = parse(Bool, get(ENV, "ZE_ENABLE_VALIDATION_LAYER", "false"))
+    parameter_validation[] = parse(Bool, get(ENV, "ZE_ENABLE_PARAMETER_VALIDATION", "false"))
 end
 
 end

src/compiler/execution.jl

@@ -167,11 +167,11 @@ function launch_configuration(kernel::HostKernel{F,TT}) where {F,TT}
     # configurations, so roll our own version that behaves like CUDA's
     # occupancy API and assumes the kernel still does bounds checking.
 
-    # once the MAX_GROUP_SIZE extension is implemented, we can use it here
     kernel_props = oneL0.properties(kernel.fun)
     group_size = if kernel_props.maxGroupSize !== missing
         kernel_props.maxGroupSize
     else
+        # without the MAX_GROUP_SIZE extension, we need to be conservative
         dev = kernel.fun.mod.device
         compute_props = oneL0.compute_properties(dev)
         max_size = compute_props.maxTotalGroupSize