Implemented override for Cuda device instruction set.

Src/ILGPU/Runtime/Cuda/CudaAccelerator.cs

@@ -1,6 +1,6 @@
 // ---------------------------------------------------------------------------------------
 //                                        ILGPU
-//                        Copyright (c) 2016-2021 ILGPU Project
+//                        Copyright (c) 2016-2023 ILGPU Project
 //                                    www.ilgpu.net
 //
 // File: CudaAccelerator.cs
@@ -9,7 +9,6 @@
 // Source License. See LICENSE.txt for details.
 // ---------------------------------------------------------------------------------------
 
-using ILGPU.Backends;
 using ILGPU.Backends.IL;
 using ILGPU.Backends.PTX;
 using ILGPU.Resources;
@@ -188,7 +187,7 @@ internal CudaAccelerator(
                 Context,
                 Capabilities,
                 Architecture,
-                InstructionSet,
+                (CudaInstructionSet)Device.InstructionSet,
                 nvvmAPI));
         }
 
@@ -220,8 +219,7 @@ internal CudaAccelerator(
         /// <summary>
         /// Returns the PTX instruction set.
         /// </summary>
-        public CudaInstructionSet InstructionSet =>
-            (CudaInstructionSet)Device.InstructionSet;
+        public CudaInstructionSet InstructionSet => Backend.InstructionSet;
 
         /// <summary>
         /// Returns the clock rate.

Src/ILGPU/Runtime/Cuda/CudaContextExtensions.cs

@@ -1,6 +1,6 @@
 // ---------------------------------------------------------------------------------------
 //                                        ILGPU
-//                        Copyright (c) 2021-2022 ILGPU Project
+//                        Copyright (c) 2021-2023 ILGPU Project
 //                                    www.ilgpu.net
 //
 // File: CudaContextExtensions.cs
@@ -29,11 +29,7 @@ public static class CudaContextExtensions
         /// <param name="builder">The builder instance.</param>
         /// <returns>The updated builder instance.</returns>
         public static Context.Builder Cuda(this Context.Builder builder) =>
-            builder.Cuda(desc =>
-                desc.Architecture.HasValue &&
-                desc.InstructionSet.HasValue &&
-                PTXCodeGenerator.SupportedInstructionSets.Contains(
-                    desc.InstructionSet.Value));
+            builder.Cuda(@override => { });
 
         /// <summary>
         /// Enables all Cuda devices.
@@ -45,6 +41,42 @@ public static Context.Builder Cuda(this Context.Builder builder) =>
         /// <returns>The updated builder instance.</returns>
         public static Context.Builder Cuda(
             this Context.Builder builder,
+            Predicate<CudaDevice> predicate) =>
+            builder.CudaInternal(@override => { }, predicate);
+
+        /// <summary>
+        /// Enables and configures all Cuda devices.
+        /// </summary>
+        /// <param name="builder">The builder instance.</param>
+        /// <param name="configure">
+        /// The action to configure a given device.
+        /// </param>
+        /// <returns>The updated builder instance.</returns>
+        public static Context.Builder Cuda(
+            this Context.Builder builder,
+            Action<CudaDeviceOverride> configure) =>
+            builder.CudaInternal(
+                configure,
+                desc =>
+                    desc.Architecture.HasValue &&
+                    desc.InstructionSet.HasValue &&
+                    PTXCodeGenerator.SupportedInstructionSets.Contains(
+                        desc.InstructionSet.Value));
+
+        /// <summary>
+        /// Enables and configures all Cuda devices.
+        /// </summary>
+        /// <param name="builder">The builder instance.</param>
+        /// <param name="configure">
+        /// The action to configure a given device.
+        /// </param>
+        /// <param name="predicate">
+        /// The predicate to include a given device.
+        /// </param>
+        /// <returns>The updated builder instance.</returns>
+        public static Context.Builder CudaInternal(
+            this Context.Builder builder,
+            Action<CudaDeviceOverride> configure,
             Predicate<CudaDevice> predicate)
         {
             if (!Backend.RuntimePlatform.Is64Bit())
@@ -55,6 +87,7 @@ public static Context.Builder Cuda(
             }
 
             CudaDevice.GetDevices(
+                configure,
                 predicate,
                 builder.DeviceRegistry);
             return builder;

Src/ILGPU/Runtime/Cuda/CudaDevice.cs

@@ -1,6 +1,6 @@
 // ---------------------------------------------------------------------------------------
 //                                        ILGPU
-//                           Copyright (c) 2021 ILGPU Project
+//                        Copyright (c) 2021-2023 ILGPU Project
 //                                    www.ilgpu.net
 //
 // File: CudaDevice.cs
@@ -9,7 +9,6 @@
 // Source License. See LICENSE.txt for details.
 // ---------------------------------------------------------------------------------------
 
-using ILGPU.Backends;
 using ILGPU.Resources;
 using System;
 using System.Collections.Immutable;
@@ -39,14 +38,16 @@ public static ImmutableArray<Device> GetDevices(
             Predicate<CudaDevice> predicate)
         {
             var registry = new DeviceRegistry();
-            GetDevices(predicate, registry);
+            GetDevices(@override => { }, predicate, registry);
             return registry.ToImmutable();
         }
 
-
         /// <summary>
         /// Detects Cuda devices.
         /// </summary>
+        /// <param name="configure">
+        /// The action to configure a given device.
+        /// </param>
         /// <param name="predicate">
         /// The predicate to include a given device.
         /// </param>
@@ -56,17 +57,20 @@ public static ImmutableArray<Device> GetDevices(
             "CA1031:Do not catch general exception types",
             Justification = "We want to hide all exceptions at this level")]
         internal static void GetDevices(
+            Action<CudaDeviceOverride> configure,
             Predicate<CudaDevice> predicate,
             DeviceRegistry registry)
         {
+            if (configure is null)
+                throw new ArgumentNullException(nameof(configure));
             if (registry is null)
                 throw new ArgumentNullException(nameof(registry));
             if (predicate is null)
                 throw new ArgumentNullException(nameof(predicate));
 
             try
             {
-                GetDevicesInternal(predicate, registry);
+                GetDevicesInternal(configure, predicate, registry);
             }
             catch (Exception)
             {
@@ -77,11 +81,15 @@ internal static void GetDevices(
         /// <summary>
         /// Detects Cuda devices.
         /// </summary>
+        /// <param name="configure">
+        /// The action to configure a given device.
+        /// </param>
         /// <param name="predicate">
         /// The predicate to include a given device.
         /// </param>
         /// <param name="registry">The registry to add all devices to.</param>
         private static void GetDevicesInternal(
+            Action<CudaDeviceOverride> configure,
             Predicate<CudaDevice> predicate,
             DeviceRegistry registry)
         {
@@ -99,6 +107,10 @@ private static void GetDevicesInternal(
                     continue;
 
                 var desc = new CudaDevice(device);
+                var deviceOverride = new CudaDeviceOverride(desc);
+                configure(deviceOverride);
+                deviceOverride.ApplyOverrides();
+
                 if (predicate(desc))
                     registry.Register(desc);
             }
@@ -376,7 +388,7 @@ private void InitPCIInfo()
         /// <summary>
         /// Returns the PTX instruction set (if supported).
         /// </summary>
-        public CudaInstructionSet? InstructionSet { get; private set; }
+        public CudaInstructionSet? InstructionSet { get; internal set; }
 
         /// <summary>
         /// Returns the clock rate.

Src/ILGPU/Runtime/Cuda/CudaDeviceOverride.cs

@@ -0,0 +1,48 @@
+// ---------------------------------------------------------------------------------------
+//                                        ILGPU
+//                           Copyright (c) 2023 ILGPU Project
+//                                    www.ilgpu.net
+//
+// File: CudaDeviceOverride.cs
+//
+// This file is part of ILGPU and is distributed under the University of Illinois Open
+// Source License. See LICENSE.txt for details.
+// ---------------------------------------------------------------------------------------
+
+namespace ILGPU.Runtime.Cuda
+{
+    /// <summary>
+    /// Represents overridable settings of a Cuda device.
+    /// </summary>
+    public sealed class CudaDeviceOverride
+    {
+        /// <summary>
+        /// The Cuda device to configure.
+        /// </summary>
+        public CudaDevice Device { get; }
+
+        /// <summary>
+        /// Forces the Cuda device to use the specified Instruction Set.
+        /// </summary>
+        public CudaInstructionSet? InstructionSet { get; set; }
+
+        /// <summary>
+        /// Constructs a new instance with the overridable settings.
+        /// </summary>
+        /// <param name="device">The Cuda device.</param>
+        internal CudaDeviceOverride(CudaDevice device)
+        {
+            Device = device;
+            InstructionSet = device.InstructionSet;
+        }
+
+        /// <summary>
+        /// Applies all the overridden settings to the Cuda device.
+        /// </summary>
+        internal void ApplyOverrides()
+        {
+            if (InstructionSet.HasValue)
+                Device.InstructionSet = InstructionSet;
+        }
+    }
+}

Src/ILGPU/Runtime/Cuda/CudaInstructionSet.cs

@@ -148,7 +148,12 @@ namespace ILGPU.Runtime.Cuda
 
         #region Instance
 
-        private CudaInstructionSet(int major, int minor)
+        /// <summary>
+        /// Creates the instruction set from major/minor values.
+        /// </summary>
+        /// <param name="major">The major version.</param>
+        /// <param name="minor">The minor version.</param>
+        public CudaInstructionSet(int major, int minor)
         {
             Major = major;
             Minor = minor;