From f91eca5154d7447dd08ada756597e219c2447fae Mon Sep 17 00:00:00 2001
From: Christian <28689358+christiangnrd@users.noreply.github.com>
Date: Fri, 11 Sep 2026 19:26:46 -0300
Subject: [PATCH 1/9] Add KI device selection interface

---
 src/OpenCLKernels.jl | 31 +++++++++++++++++++++++++++++++
 1 file changed, 31 insertions(+)

diff --git a/src/OpenCLKernels.jl b/src/OpenCLKernels.jl
index f57c1642..9f16bafe 100644
--- a/src/OpenCLKernels.jl
+++ b/src/OpenCLKernels.jl
@@ -15,6 +15,12 @@ import Adapt
 export OpenCLBackend
 
 struct OpenCLBackend <: KA.GPU
+    platform::cl.Platform
+
+    function OpenCLBackend(; platform=cl.platform())
+        cl.platform!(platform)
+        new(platform)
+    end
 end
 
 function KA.allocate(::OpenCLBackend, ::Type{T}, dims::Tuple; unified::Bool = false) where T
@@ -48,6 +54,31 @@ Adapt.adapt_storage(::KA.CPU, a::CLArray) = convert(Array, a)
 # rather than for `CLArray`.
 Adapt.adapt_storage(::KA.ConstAdaptor, a::CLDeviceArray) = Base.Experimental.Const(a)
 
+## Device Selection
+
+# devices are numbered consecutively across all platforms, in enumeration order
+
+function KA.ndevices(b::OpenCLBackend)
+    length(cl.devices(b.platform))
+end
+
+function KA.device(b::OpenCLBackend)
+    current = cl.device()
+    i = 0
+    for d in cl.devices(b.platform)
+        i += 1
+        d == current && return i
+    end
+    error("Active OpenCL device $current not found in the current OpenCL platform \"$(b.platform.name)\".")
+end
+
+function KA.device!(b::OpenCLBackend, id::Int)
+    devs = cl.devices(b.platform)
+    id > length(devs) && throw(ArgumentError("Device id $id out of bounds."))
+
+    cl.device!(devs[id])
+    return nothing
+end
 
 ## Memory Operations
 

From e630e4b162592349c9d9acaf6088d4b58369f24a Mon Sep 17 00:00:00 2001
From: Christian <28689358+christiangnrd@users.noreply.github.com>
Date: Tue, 15 Sep 2026 12:04:59 -0300
Subject: [PATCH 2/9] Fix test

---
 src/OpenCLKernels.jl | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/OpenCLKernels.jl b/src/OpenCLKernels.jl
index 9f16bafe..b480aaf6 100644
--- a/src/OpenCLKernels.jl
+++ b/src/OpenCLKernels.jl
@@ -74,7 +74,7 @@ end
 
 function KA.device!(b::OpenCLBackend, id::Int)
     devs = cl.devices(b.platform)
-    id > length(devs) && throw(ArgumentError("Device id $id out of bounds."))
+    0 < id < length(devs) || throw(ArgumentError("Device id $id out of bounds."))
 
     cl.device!(devs[id])
     return nothing

From 544b2d39ee40c9b006b66f38e15b88d18cdeba3e Mon Sep 17 00:00:00 2001
From: Christian <28689358+christiangnrd@users.noreply.github.com>
Date: Tue, 15 Sep 2026 12:05:08 -0300
Subject: [PATCH 3/9] Style fix

---
 src/OpenCLKernels.jl | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/src/OpenCLKernels.jl b/src/OpenCLKernels.jl
index b480aaf6..53fac2eb 100644
--- a/src/OpenCLKernels.jl
+++ b/src/OpenCLKernels.jl
@@ -64,9 +64,7 @@ end
 
 function KA.device(b::OpenCLBackend)
     current = cl.device()
-    i = 0
-    for d in cl.devices(b.platform)
-        i += 1
+    for (i, d) in enumerate(cl.devices(b.platform))
         d == current && return i
     end
     error("Active OpenCL device $current not found in the current OpenCL platform \"$(b.platform.name)\".")

From 25d4399df0f25d90cea037a26228ccde6bce7914 Mon Sep 17 00:00:00 2001
From: Christian <28689358+christiangnrd@users.noreply.github.com>
Date: Tue, 15 Sep 2026 12:06:52 -0300
Subject: [PATCH 4/9] Fix test fix

---
 src/OpenCLKernels.jl | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/OpenCLKernels.jl b/src/OpenCLKernels.jl
index 53fac2eb..123ef2c1 100644
--- a/src/OpenCLKernels.jl
+++ b/src/OpenCLKernels.jl
@@ -72,7 +72,7 @@ end
 
 function KA.device!(b::OpenCLBackend, id::Int)
     devs = cl.devices(b.platform)
-    0 < id < length(devs) || throw(ArgumentError("Device id $id out of bounds."))
+    0 < id <= length(devs) || throw(ArgumentError("Device id $id out of bounds."))
 
     cl.device!(devs[id])
     return nothing

From fc5a516e4db923470505b4dbe2687b522f649717 Mon Sep 17 00:00:00 2001
From: Christian <28689358+christiangnrd@users.noreply.github.com>
Date: Tue, 15 Sep 2026 12:55:57 -0300
Subject: [PATCH 5/9] Fix comment

---
 src/OpenCLKernels.jl | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/OpenCLKernels.jl b/src/OpenCLKernels.jl
index 123ef2c1..f23850e4 100644
--- a/src/OpenCLKernels.jl
+++ b/src/OpenCLKernels.jl
@@ -56,7 +56,7 @@ Adapt.adapt_storage(::KA.ConstAdaptor, a::CLDeviceArray) = Base.Experimental.Con
 
 ## Device Selection
 
-# devices are numbered consecutively across all platforms, in enumeration order
+# devices are numbered consecutively within the backend's platform, in enumeration order
 
 function KA.ndevices(b::OpenCLBackend)
     length(cl.devices(b.platform))

From b181074ec21f451706734428d61344f097031827 Mon Sep 17 00:00:00 2001
From: Christian <28689358+christiangnrd@users.noreply.github.com>
Date: Tue, 15 Sep 2026 13:20:25 -0300
Subject: [PATCH 6/9] Add ndevices function to OpenCL and use it in the KA
 interface

---
 lib/cl/platform.jl   | 21 ++++++++++++++-------
 src/OpenCLKernels.jl |  4 ++--
 2 files changed, 16 insertions(+), 9 deletions(-)

diff --git a/lib/cl/platform.jl b/lib/cl/platform.jl
index 7ea5a27c..33bbd751 100644
--- a/lib/cl/platform.jl
+++ b/lib/cl/platform.jl
@@ -81,16 +81,23 @@ function num_platforms()
     return Int(nplatforms[])
 end
 
-function devices(p::Platform, dtype)
-    ndevices = Ref{Cuint}()
-    ret = unchecked_clGetDeviceIDs(p, dtype, 0, C_NULL, ndevices)
-    if ret == CL_DEVICE_NOT_FOUND || ndevices[] == 0
-        return Device[]
+function ndevices(p::Platform, dtype=CL_DEVICE_TYPE_ALL)
+    ndevs = Ref{Cuint}()
+    ret = unchecked_clGetDeviceIDs(p, dtype, 0, C_NULL, ndevs)
+    if ret == CL_DEVICE_NOT_FOUND
+        return Cuint(0)
     elseif ret != CL_SUCCESS
         throw(CLError(ret))
     end
-    result = Vector{cl_device_id}(undef, ndevices[])
-    clGetDeviceIDs(p, dtype, ndevices[], result, C_NULL)
+
+    return ndevs[]
+end
+
+function devices(p::Platform, dtype)
+    ndevs = ndevices(p, dtype)
+    ndevs > 0 || return Device[]
+    result = Vector{cl_device_id}(undef, ndevs)
+    clGetDeviceIDs(p, dtype, ndevs, result, C_NULL)
     devs = Device[Device(id) for id in result]
 
     # OpenCL does not guarantee a stable enumeration order, so sort deterministically
diff --git a/src/OpenCLKernels.jl b/src/OpenCLKernels.jl
index f23850e4..88d49fde 100644
--- a/src/OpenCLKernels.jl
+++ b/src/OpenCLKernels.jl
@@ -59,7 +59,7 @@ Adapt.adapt_storage(::KA.ConstAdaptor, a::CLDeviceArray) = Base.Experimental.Con
 # devices are numbered consecutively within the backend's platform, in enumeration order
 
 function KA.ndevices(b::OpenCLBackend)
-    length(cl.devices(b.platform))
+    Int(cl.ndevices(b.platform))
 end
 
 function KA.device(b::OpenCLBackend)
@@ -71,8 +71,8 @@ function KA.device(b::OpenCLBackend)
 end
 
 function KA.device!(b::OpenCLBackend, id::Int)
+    0 < id <= KA.ndevices(b) || throw(ArgumentError("Device id $id out of bounds."))
     devs = cl.devices(b.platform)
-    0 < id <= length(devs) || throw(ArgumentError("Device id $id out of bounds."))
 
     cl.device!(devs[id])
     return nothing

From 65ecbc9e8d58aa427cebe92b0f5a44b983699d6a Mon Sep 17 00:00:00 2001
From: Christian <28689358+christiangnrd@users.noreply.github.com>
Date: Tue, 15 Sep 2026 13:26:14 -0300
Subject: [PATCH 7/9] Clarify error message

---
 src/OpenCLKernels.jl | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/OpenCLKernels.jl b/src/OpenCLKernels.jl
index 88d49fde..752b1827 100644
--- a/src/OpenCLKernels.jl
+++ b/src/OpenCLKernels.jl
@@ -67,7 +67,7 @@ function KA.device(b::OpenCLBackend)
     for (i, d) in enumerate(cl.devices(b.platform))
         d == current && return i
     end
-    error("Active OpenCL device $current not found in the current OpenCL platform \"$(b.platform.name)\".")
+    error("Active OpenCL device $current not found in the OpenCLBackend's platform \"$(b.platform.name)\".")
 end
 
 function KA.device!(b::OpenCLBackend, id::Int)

From 4609f7c3e48730729e4b5737a547f1b8b7509460 Mon Sep 17 00:00:00 2001
From: Christian <28689358+christiangnrd@users.noreply.github.com>
Date: Tue, 15 Sep 2026 13:38:53 -0300
Subject: [PATCH 8/9] Don't change the platform when constructing backend, but
 check

---
 src/OpenCLKernels.jl | 17 ++++++++++-------
 1 file changed, 10 insertions(+), 7 deletions(-)

diff --git a/src/OpenCLKernels.jl b/src/OpenCLKernels.jl
index 752b1827..afd48af7 100644
--- a/src/OpenCLKernels.jl
+++ b/src/OpenCLKernels.jl
@@ -14,16 +14,17 @@ import Adapt
 
 export OpenCLBackend
 
-struct OpenCLBackend <: KA.GPU
-    platform::cl.Platform
+Base.@kwdef struct OpenCLBackend <: KA.GPU
+    platform::cl.Platform = cl.platform()
+end
 
-    function OpenCLBackend(; platform=cl.platform())
-        cl.platform!(platform)
-        new(platform)
-    end
+function check_platform(b::OpenCLBackend)
+    b.platform === cl.platform() || @warn "OpenCLBackend platform \"$(b.platform.name)\" is not the active platform \"$(cl.platform().name)\""
+    return nothing
 end
 
-function KA.allocate(::OpenCLBackend, ::Type{T}, dims::Tuple; unified::Bool = false) where T
+function KA.allocate(b::OpenCLBackend, ::Type{T}, dims::Tuple; unified::Bool = false) where T
+    check_platform(b)
     if unified
         memory_backend = cl.unified_memory_backend()
         if memory_backend === cl.USMBackend()
@@ -130,6 +131,8 @@ function threads_to_workgroupsize(threads, ndrange)
 end
 
 function (obj::KA.Kernel{OpenCLBackend})(args...; ndrange=nothing, workgroupsize=nothing)
+    check_platform(obj.backend)
+
     ndrange, workgroupsize, iterspace, dynamic =
         KA.launch_config(obj, ndrange, workgroupsize)
 

From 5350e44ba396496e8d8b264e112ba6643617a7bd Mon Sep 17 00:00:00 2001
From: Christian <28689358+christiangnrd@users.noreply.github.com>
Date: Wed, 16 Sep 2026 07:57:54 -0300
Subject: [PATCH 9/9] Inline check, outline warning

---
 src/OpenCLKernels.jl | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/src/OpenCLKernels.jl b/src/OpenCLKernels.jl
index afd48af7..b1d4eaa7 100644
--- a/src/OpenCLKernels.jl
+++ b/src/OpenCLKernels.jl
@@ -18,13 +18,13 @@ Base.@kwdef struct OpenCLBackend <: KA.GPU
     platform::cl.Platform = cl.platform()
 end
 
-function check_platform(b::OpenCLBackend)
-    b.platform === cl.platform() || @warn "OpenCLBackend platform \"$(b.platform.name)\" is not the active platform \"$(cl.platform().name)\""
+@noinline function platform_mismatch_warning(expected::cl.Platform, active::cl.Platform)
+    @warn "OpenCLBackend platform \"$(expected.name)\" is not the active platform \"$(active.name)\""
     return nothing
 end
 
 function KA.allocate(b::OpenCLBackend, ::Type{T}, dims::Tuple; unified::Bool = false) where T
-    check_platform(b)
+    b.platform === cl.platform() || platform_mismatch_warning(b.platform, cl.platform())
     if unified
         memory_backend = cl.unified_memory_backend()
         if memory_backend === cl.USMBackend()
@@ -131,7 +131,7 @@ function threads_to_workgroupsize(threads, ndrange)
 end
 
 function (obj::KA.Kernel{OpenCLBackend})(args...; ndrange=nothing, workgroupsize=nothing)
-    check_platform(obj.backend)
+    obj.backend.platform === cl.platform() || platform_mismatch_warning(obj.backend.platform, cl.platform())
 
     ndrange, workgroupsize, iterspace, dynamic =
         KA.launch_config(obj, ndrange, workgroupsize)