diff --git a/deps/build_ci.jl b/deps/build_ci.jl
index 831872c8..68ed8c9c 100644
--- a/deps/build_ci.jl
+++ b/deps/build_ci.jl
@@ -16,6 +16,7 @@ if isdir(joinpath(@__DIR__), "..", ".git")
     @info "Latest change to the wrappers: $(unix2datetime(deps_timestamp))"
 
     # find out which version of oneAPI_Support_jll we are using
+    Pkg.add(; url="https://github.com/leios/GPUArrays.jl/", rev="yoyoyo_rebase_time")
     Pkg.activate(joinpath(@__DIR__, ".."))
     Pkg.instantiate()
     deps = collect(values(Pkg.dependencies()))
diff --git a/src/gpuarrays.jl b/src/gpuarrays.jl
index 614e655d..1cc39492 100644
--- a/src/gpuarrays.jl
+++ b/src/gpuarrays.jl
@@ -1,5 +1,7 @@
 # GPUArrays.jl interface
 
+import KernelAbstractions
+import KernelAbstractions: Backend
 
 #
 # Device functionality
@@ -8,13 +10,16 @@
 
 ## execution
 
-struct oneArrayBackend <: AbstractGPUBackend end
+@inline function GPUArrays.launch_heuristic(::oneAPIBackend, obj::O, args::Vararg{Any,N};
+                                            elements::Int, elements_per_thread::Int) where {O,N}
+    ndrange = ceil(Int, elements / elements_per_thread)
+    ndrange, workgroupsize, iterspace, dynamic = KA.launch_config(obj, ndrange,
+                                                                  nothing)
 
-struct oneKernelContext <: AbstractKernelContext end
+    # this might not be the final context, since we may tune the workgroupsize
+    ctx = KA.mkcontext(obj, ndrange, iterspace)
 
-@inline function GPUArrays.launch_heuristic(::oneArrayBackend, f::F, args::Vararg{Any,N};
-                                             elements::Int, elements_per_thread::Int) where {F,N}
-    kernel = @oneapi launch=false f(oneKernelContext(), args...)
+    kernel = @oneapi launch=false obj.f(ctx, args...)
 
     items = launch_configuration(kernel)
     # XXX: how many groups is a good number? the API doesn't tell us.
@@ -23,48 +28,6 @@ struct oneKernelContext <: AbstractKernelContext end
     return (threads=items, blocks=32)
 end
 
-function GPUArrays.gpu_call(::oneArrayBackend, f, args, threads::Int, blocks::Int;
-                            name::Union{String,Nothing})
-    @oneapi items=threads groups=blocks name=name f(oneKernelContext(), args...)
-end
-
-
-## on-device
-
-# indexing
-
-GPUArrays.blockidx(ctx::oneKernelContext) = oneAPI.get_group_id(0)
-GPUArrays.blockdim(ctx::oneKernelContext) = oneAPI.get_local_size(0)
-GPUArrays.threadidx(ctx::oneKernelContext) = oneAPI.get_local_id(0)
-GPUArrays.griddim(ctx::oneKernelContext) = oneAPI.get_num_groups(0)
-
-# math
-
-@inline GPUArrays.cos(ctx::oneKernelContext, x) = oneAPI.cos(x)
-@inline GPUArrays.sin(ctx::oneKernelContext, x) = oneAPI.sin(x)
-@inline GPUArrays.sqrt(ctx::oneKernelContext, x) = oneAPI.sqrt(x)
-@inline GPUArrays.log(ctx::oneKernelContext, x) = oneAPI.log(x)
-
-# memory
-
-@inline function GPUArrays.LocalMemory(::oneKernelContext, ::Type{T}, ::Val{dims}, ::Val{id}
-                                      ) where {T, dims, id}
-    ptr = oneAPI.emit_localmemory(Val(id), T, Val(prod(dims)))
-    oneDeviceArray(dims, LLVMPtr{T, onePI.AS.Local}(ptr))
-end
-
-# synchronization
-
-@inline GPUArrays.synchronize_threads(::oneKernelContext) = oneAPI.barrier()
-
-
-
-#
-# Host abstractions
-#
-
-GPUArrays.backend(::Type{<:oneArray}) = oneArrayBackend()
-
 const GLOBAL_RNGs = Dict{ZeDevice,GPUArrays.RNG}()
 function GPUArrays.default_rng(::Type{<:oneArray})
     dev = device()
diff --git a/src/oneAPI.jl b/src/oneAPI.jl
index 259fea77..00d83593 100644
--- a/src/oneAPI.jl
+++ b/src/oneAPI.jl
@@ -17,6 +17,8 @@ using Core: LLVMPtr
 
 using SPIRV_LLVM_Translator_unified_jll, SPIRV_Tools_jll
 
+import KernelAbstractions as KA
+
 export oneL0
 
 # core library
@@ -67,14 +69,14 @@ end
 # integrations and specialized functionality
 include("broadcast.jl")
 include("mapreduce.jl")
+include("oneAPIKernels.jl")
+import .oneAPIKernels: oneAPIBackend, KA.launch_config
+export oneAPIBackend
+
 include("gpuarrays.jl")
 include("random.jl")
 include("utils.jl")
 
-include("oneAPIKernels.jl")
-import .oneAPIKernels: oneAPIBackend
-export oneAPIBackend
-
 function __init__()
     precompiling = ccall(:jl_generating_output, Cint, ()) != 0
     precompiling && return