Detect device-side exceptions on the host.

lib/cl/cmdqueue.jl

@@ -62,6 +62,7 @@ end
 
 function finish(q::CmdQueue)
     clFinish(q)
+    OpenCL.check_exceptions()
     return q
 end
 

src/OpenCL.jl

@@ -31,6 +31,7 @@ include("array.jl")
 
 # compiler implementation
 include("compiler/compilation.jl")
+include("compiler/exceptions.jl")
 include("compiler/execution.jl")
 include("compiler/reflection.jl")
 

src/compiler/compilation.jl

@@ -9,6 +9,8 @@ GPUCompiler.runtime_module(::CompilerJob{<:Any,OpenCLCompilerParams}) = OpenCL
 GPUCompiler.method_table_view(job::OpenCLCompilerJob) =
     GPUCompiler.StackedMethodTable(job.world, method_table, SPIRVIntrinsics.method_table)
 
+GPUCompiler.kernel_state_type(job::OpenCLCompilerJob) = KernelState
+
 # filter out OpenCL built-ins
 # TODO: eagerly lower these using the translator API
 GPUCompiler.isintrinsic(job::OpenCLCompilerJob, fn::String) =

src/compiler/exceptions.jl

@@ -0,0 +1,52 @@
+# support for device-side exceptions
+
+## exception type
+
+struct KernelException <: Exception
+    devs::Vector{cl.Device}
+end
+
+function Base.showerror(io::IO, err::KernelException)
+    print(io, "KernelException: exception thrown during kernel execution on device(s) $(join(map(dev->dev.name, err.devs), ", "))")
+end
+
+
+## exception handling
+
+const exception_infos = Dict{cl.Context, Union{Nothing, cl.AbstractPointerMemory}}()
+
+# create a CPU/GPU exception flag for error signalling
+function create_exceptions!(ctx::cl.Context, dev::cl.Device)
+    mem = get!(exception_infos, ctx) do
+        if cl.svm_capabilities(cl.device()).fine_grain_buffer
+            cl.svm_alloc(sizeof(ExceptionInfo_st); fine_grained=true)
+        elseif cl.usm_supported(dev) && cl.usm_capabilities(dev).host.access
+            cl.host_alloc(sizeof(ExceptionInfo_st))
+        else
+            nothing
+        end
+    end
+    if mem === nothing
+        return convert(ExceptionInfo, C_NULL)
+    end
+
+    exception_info = convert(ExceptionInfo, mem)
+    unsafe_store!(exception_info, ExceptionInfo_st())
+    return exception_info
+end
+
+# check the exception flags on every API call
+function check_exceptions()
+    for (ctx, mem) in exception_infos
+        mem === nothing && continue
+        exception_info = convert(ExceptionInfo, mem)
+        if exception_info.status != 0
+            # restore the structure
+            unsafe_store!(exception_info, ExceptionInfo_st())
+
+            # throw host-side
+            throw(KernelException(ctx.devices))
+        end
+    end
+    return
+end

src/compiler/execution.jl

@@ -143,6 +143,10 @@ abstract type AbstractKernel{F, TT} end
     call_t =                  Type[x[1] for x in zip(sig.parameters,  to_pass) if x[2]]
     call_args = Union{Expr,Symbol}[x[1] for x in zip(args, to_pass)            if x[2]]
 
+    # add the kernel state as the first argument
+    pushfirst!(call_t, KernelState)
+    pushfirst!(call_args, :(kernel.state))
+
     # replace non-isbits arguments (they should be unused, or compilation would have failed)
     for (i,dt) in enumerate(call_t)
         if !isbitstype(dt)
@@ -156,6 +160,16 @@ abstract type AbstractKernel{F, TT} end
 
     quote
         indirect_memory = cl.AbstractMemory[]
+
+        # add exception info buffer to indirect memory
+        # XXX: this is too expensive
+        if kernel.state.exception_info != C_NULL
+            ctx = cl.context()
+            if haskey(exception_infos, ctx)
+                push!(indirect_memory, exception_infos[ctx])
+            end
+        end
+
         clcall(kernel.fun, $call_tt, $(call_args...); indirect_memory, call_kwargs...)
     end
 end
@@ -167,6 +181,7 @@ end
 struct HostKernel{F,TT} <: AbstractKernel{F,TT}
     f::F
     fun::cl.Kernel
+    state::KernelState
 end
 
 
@@ -191,7 +206,9 @@ function clfunction(f::F, tt::TT=Tuple{}; kwargs...) where {F,TT}
         kernel = get(_kernel_instances, h, nothing)
         if kernel === nothing
             # create the kernel state object
-            kernel = HostKernel{F,tt}(f, fun)
+            exception_info = create_exceptions!(ctx, dev)
+            state = KernelState(exception_info)
+            kernel = HostKernel{F,tt}(f, fun, state)
             _kernel_instances[h] = kernel
         end
         return kernel::HostKernel{F,tt}

src/device/runtime.jl

@@ -1,7 +1,56 @@
 # reset the runtime cache from global scope, so that any change triggers recompilation
 GPUCompiler.reset_runtime()
 
-signal_exception() = return
+## exception handling
+
+struct ExceptionInfo_st
+    # whether an exception has been encountered (0 -> 1)
+    status::Int32
+
+    ExceptionInfo_st() = new(0)
+end
+
+# to simplify use of this struct, which is passed by-reference, use property overloading
+const ExceptionInfo = Ptr{ExceptionInfo_st}
+@inline function Base.getproperty(info::ExceptionInfo, sym::Symbol)
+    if sym === :status
+        unsafe_load(convert(Ptr{Int32}, info))
+    else
+        getfield(info, sym)
+    end
+end
+@inline function Base.setproperty!(info::ExceptionInfo, sym::Symbol, value)
+    if sym === :status
+        unsafe_store!(convert(Ptr{Int32}, info), value)
+    else
+        setfield!(info, sym, value)
+    end
+end
+
+## kernel state
+
+struct KernelState
+    exception_info::ExceptionInfo
+
+    # XXX: Intel's SPIR-V compiler does not support array-valued kernel arguments, and Julia
+    #      emits homogeneous structs as arrays. Work around this by including a dummy field.
+    dummy::UInt32
+end
+KernelState(exception_info::ExceptionInfo) = KernelState(exception_info, 42)
+
+@inline @generated kernel_state() = GPUCompiler.kernel_state_value(KernelState)
+
+function signal_exception()
+    info = kernel_state().exception_info
+
+    # inform the host
+    if info != C_NULL
+        info.status = 1
+        write_mem_fence(OpenCL.GLOBAL_MEM_FENCE)
+    end
+
+    return
+end
 
 malloc(sz) = C_NULL