/// <summary>
/// Generates an intrinsic reduce.
/// </summary>
/// <typeparam name="T">The element type.</typeparam>
/// <typeparam name="TReduction">The type of the reduction logic.</typeparam>
/// <param name="backend">The current backend.</param>
/// <param name="codeGenerator">The code generator.</param>
/// <param name="value">The value to generate code for.</param>
public static void GenerateReduce <T, TReduction>(
CLBackend backend,
CLCodeGenerator codeGenerator,
Value value)
where T : unmanaged
where TReduction : struct, IScanReduceOperation <T> =>
GenerateAllReduce <T, TReduction>(
backend,
codeGenerator,
value);
/// <summary>
/// Generates intrinsic math instructions for the following kinds:
/// Rcp
/// </summary>
/// <param name="backend">The current backend.</param>
/// <param name="codeGenerator">The code generator.</param>
/// <param name="value">The value to generate code for.</param>
public static void GenerateMathIntrinsic(
CLBackend backend,
CLCodeGenerator codeGenerator,
Value value)
{
// Manually generate code for "1.0 / argument"
var arithmeticValue = value as UnaryArithmeticValue;
var argument = codeGenerator.Load(arithmeticValue.Value);
var target = codeGenerator.Allocate(arithmeticValue);
var operation = CLInstructions.GetArithmeticOperation(
BinaryArithmeticKind.Div,
arithmeticValue.BasicValueType.IsFloat(),
out var isFunction);
using var statement = codeGenerator.BeginStatement(target);
statement.AppendCast(arithmeticValue.ArithmeticBasicValueType);
if (isFunction)
{
statement.AppendCommand(operation);
statement.BeginArguments();
}
else
{
statement.OpenParen();
}
statement.AppendCast(arithmeticValue.ArithmeticBasicValueType);
if (arithmeticValue.BasicValueType == BasicValueType.Float32)
{
statement.AppendConstant(1.0f);
}
else
{
statement.AppendConstant(1.0);
}
if (!isFunction)
{
statement.AppendCommand(operation);
}
statement.AppendArgument();
statement.AppendCast(arithmeticValue.ArithmeticBasicValueType);
statement.Append(argument);
if (isFunction)
{
statement.EndArguments();
}
else
{
statement.CloseParen();
}
}
/// <summary>
/// Generates an intrinsic scan.
/// </summary>
/// <typeparam name="T">The element type.</typeparam>
/// <typeparam name="TScanOperation">The type of the warp scan logic.</typeparam>
/// <param name="backend">The current backend.</param>
/// <param name="codeGenerator">The code generator.</param>
/// <param name="value">The value to generate code for.</param>
public static void GenerateInclusiveScan <T, TScanOperation>(
CLBackend backend,
CLCodeGenerator codeGenerator,
Value value)
where T : unmanaged
where TScanOperation : struct, IScanReduceOperation <T> =>
CLContext.GenerateScanReduce <T, TScanOperation>(
backend,
codeGenerator,
value,
"work_group_scan_inclusive_");
/// <summary>
/// Generates an intrinsic reduce.
/// </summary>
/// <typeparam name="T">The element type.</typeparam>
/// <typeparam name="TReduction">The type of the reduction logic.</typeparam>
/// <param name="backend">The current backend.</param>
/// <param name="codeGenerator">The code generator.</param>
/// <param name="value">The value to generate code for.</param>
public static void GenerateAllReduce <T, TReduction>(
CLBackend backend,
CLCodeGenerator codeGenerator,
Value value)
where T : unmanaged
where TReduction : struct, IScanReduceOperation <T> =>
CLContext.GenerateScanReduce <T, TReduction>(
backend,
codeGenerator,
value,
"work_group_reduce_");
/// <summary>
/// Generates an intrinsic scan.
/// </summary>
/// <typeparam name="T">The element type.</typeparam>
/// <typeparam name="TScanOperation">The type of the warp scan logic.</typeparam>
/// <param name="backend">The current backend.</param>
/// <param name="codeGenerator">The code generator.</param>
/// <param name="value">The value to generate code for.</param>
public static void GenerateExclusiveScan <T, TScanOperation>(
CLBackend backend,
CLCodeGenerator codeGenerator,
Value value)
where T : struct
where TScanOperation : struct, IScanReduceOperation <T> =>
CLContext.GenerateScanReduce <T, TScanOperation>(
backend,
codeGenerator,
value,
"sub_group_scan_exclusive_");
/// <summary>
/// The OpenCL implementation.
/// </summary>
/// <remarks>
/// Note that this function signature corresponds to the OpenCL-backend specific
/// delegate type <see cref="CLIntrinsic.Handler"/>.
/// </remarks>
static void GenerateCLCode(
CLBackend backend,
CLCodeGenerator codeGenerator,
Value value)
{
// The passed value will be the call node in this case
// Load X parameter variable (first argument)
var xVariable = codeGenerator.Load(value[0]);
// Allocate target variable to write our result to
var target = codeGenerator.Allocate(value);
// Emit our desired instructions
using var statement = codeGenerator.BeginStatement(target);
statement.Append(xVariable);
statement.AppendCommand(
CLInstructions.GetArithmeticOperation(
BinaryArithmeticKind.Mul,
false,
out var _));
statement.AppendConstant(2);
}
public static void Main2()
{
using (var context = new Context())
{
foreach (var acceleratorId in Accelerator.Accelerators)
{
if (acceleratorId.AcceleratorType == AcceleratorType.CPU)
{
continue;
}
using (var accelerator = Accelerator.Create(context, acceleratorId))
{
CompiledKernel compiledKernel;
using (Backend b = new CLBackend(context, ILGPU.Runtime.OpenCL.CLAcceleratorVendor.AMD))
{
MethodInfo methodInfo = typeof(GPU).GetMethod("PixelKernel");
KernelSpecialization spec = KernelSpecialization.Empty;
compiledKernel = b.Compile(EntryPointDescription.FromImplicitlyGroupedKernel(methodInfo), spec);
// debug: check kernel.Source for source text
}
var kernel = accelerator.LoadAutoGroupedKernel(compiledKernel);
// var kernel = accelerator.LoadAutoGroupedStreamKernel<Index2, ArrayView2D<FSMUnit>>(MathKernel);
// kernel = accelerator.LoadAutoGroupedStreamKernel<Index2, ArrayView2D<Color3>, ArrayView<byte>, ArrayView2D<Neuron>>(PixelKernel);
MemoryBuffer2D <Color3> buffer = accelerator.Allocate <Color3>(pixelMap.GetLength(0), pixelMap.GetLength(1));
MemoryBuffer <byte> buffer2 = accelerator.Allocate <byte>(imageBytes.Length);
MemoryBuffer2D <Neuron> buffer3 = accelerator.Allocate <Neuron>(nrn.GetLength(0), nrn.GetLength(1));
buffer3.CopyFrom(nrn, new LongIndex2(0, 0), new LongIndex2(0, 0), new LongIndex2(nrn.GetLength(0), nrn.GetLength(1)));
while (running == true)
{
Stopwatch sw = new Stopwatch();
sw.Start();
Index2 gridSize = new Index2(pixelMap.GetLength(0), pixelMap.GetLength(1));
//kernel(gridSize, buffer.View, buffer2.View, buffer3.View);
sw.OutputDelta("Kernel");
accelerator.Synchronize();
sw.OutputDelta("Sync");
// imageBytes = buffer2.GetAsArray();
buffer2.CopyTo(imageBytes, 0, 0, imageBytes.Length);
sw.OutputDelta("Copy ImageBytes");
// Resolve and verify data
//pixelMap = buffer.GetAs2DArray();
// buffer.CopyTo(pixelMap, new LongIndex2(0, 0), new LongIndex2(0, 0), new LongIndex2(pixelMap.GetLength(0), pixelMap.GetLength(1)));
// Color3[] pixelMap1D = buffer.GetAsArray();
//Copy1DTo2DArray(pixelMap1D, pixelMap); // ~36ms, a bit faster
//Array.Copy(pixelMap1D, imageBytes, pixelMap1D.Length); // fails
//Buffer.BlockCopy(pixelMap1D, 0, pixelMap, 0, pixelMap1D.Length * Marshal.SizeOf(typeof(Color3))); // fails
// pixelMap = Make2DArray(pixelMap1D, pixelMap.GetLength(0), pixelMap.GetLength(1)); // still slow
//sw.OutputDelta("Copy PixelMap");
// MainForm.form.DrawPixels(pixelMap);
MainForm.form.DrawPixels(imageBytes, pixelMap.GetLength(0), pixelMap.GetLength(1));
Application.DoEvents();
//Debugger.Break();
sw.OutputDelta("DrawPixels");
}
buffer.Dispose();
buffer2.Dispose();
buffer3.Dispose();
Application.Exit();
//Debugger.Break();
}
}
}
}