internal static void SpecializedExplicitValueKernel <T>(
ArrayView1D <T, Stride1D.Dense> data,
SpecializedValue <T> value)
where T : unmanaged, IEquatable <T>
{
data[0] = value;
}
public Camera(Camera camera, Vec3 movement, Vec3 turn)
{
this.width = camera.width;
this.height = camera.height;
this.maxBounces = camera.maxBounces;
Vector4 temp = camera.lookAt - camera.origin;
if (turn.y != 0)
{
temp += Vector4.Transform(temp, Matrix4x4.CreateFromAxisAngle(Vec3.cross(Vec3.cross(camera.up, (camera.lookAt - camera.origin)), (camera.lookAt - camera.origin)), (float)turn.y));
}
if (turn.x != 0)
{
temp += Vector4.Transform(temp, Matrix4x4.CreateFromAxisAngle(Vec3.cross(camera.up, (camera.lookAt - camera.origin)), (float)turn.x));
}
lookAt = camera.origin + Vec3.unitVector(temp);
this.origin = camera.origin + movement;
this.lookAt += movement;
this.up = camera.up;
axis = OrthoNormalBasis.fromZY(Vec3.unitVector(lookAt - origin), up);
aspectRatio = ((float)width / (float)height);
cameraPlaneDist = 1.0f / XMath.Tan(camera.verticalFov * XMath.PI / 360.0f);
this.verticalFov = camera.verticalFov;
reciprocalHeight = 1.0f / height;
reciprocalWidth = 1.0f / width;
}
internal static void SpecializedImplicitValueKernel <T>(
Index1 _,
ArrayView <T> data,
SpecializedValue <T> value)
where T : unmanaged, IEquatable <T>
{
data[0] = value;
}
// The specialization also works with generic kernels
static void SpecializedGenericKernel <TValue>(
ArrayView <TValue> view,
SpecializedValue <TValue> specialized)
where TValue : unmanaged, IEquatable <TValue>
{
var globalIndex = Grid.GlobalIndex.X;
view[globalIndex] = specialized;
}
private static void IlGpuKernelConstants(
ArrayView2D <Real> mSquaredDistances,
ArrayView <Real> mCoordinates,
SpecializedValue <int> c,
int n)
{
// Same as CudaKernelOptimised2, but the number of coordinates is given as a meta-constant.
// Also, we write the results as float2.
var shared = SharedMemory.GetDynamic <Real>();
var coordinatesI = shared.GetSubView(0, c * Group.DimX);
var coordinatesJ = shared.GetSubView(c * Group.DimX);
var bI = Grid.IdxY * Group.DimX;
var bJ = Grid.IdxX * Group.DimX;
for (int k = 0; k != c; ++k)
{
if (bI + Group.IdxX < n)
{
coordinatesI[k * Group.DimX + Group.IdxX] = mCoordinates[k * n + bI + Group.IdxX];
}
if (bJ + Group.IdxX < n)
{
coordinatesJ[k * Group.DimX + Group.IdxX] = mCoordinates[k * n + bJ + Group.IdxX];
}
}
Group.Barrier();
var line = Group.IdxX / (Group.DimX / 2);
var tid = Group.IdxX % (Group.DimX / 2);
if (bJ + tid * 2 < n)
{
var coordinatesJ2 = coordinatesJ.Cast <IlReal2>();
for (int i = line; i < Group.DimX & bI + i < n; i += 2)
{
var dist = default(IlReal2);
for (int k = 0; k != c; ++k)
{
var coord1 = coordinatesI[k * Group.DimX + i];
var coord2 = coordinatesJ2[(k * Group.DimX / 2) + tid];
var diff = new IlReal2(coord1 - coord2.X, coord1 - coord2.Y);
dist += diff * diff;
}
var dst = mSquaredDistances.Cast <IlReal2>();
dst[bJ / 2 + tid, bI + i] = dist;
}
}
}
// A simple kernel with a specialized kernel parameter
public static void SpecializedKernel(
ArrayView <int> view,
// This parameter will be replaced by a 'constant' value during
// the first kernel call
SpecializedValue <int> specialized)
{
var globalIndex = Grid.GlobalIndex.X;
view[globalIndex] = specialized;
}
static void Main()
{
using (var context = new Context())
{
// For each available accelerator...
foreach (var acceleratorId in Accelerator.Accelerators)
{
// Create default accelerator for the given accelerator id
using (var accelerator = Accelerator.Create(context, acceleratorId))
{
Console.WriteLine($"Performing operations on {accelerator}");
int groupSize = accelerator.MaxNumThreadsPerGroup;
// Scenario 1: simple version
using (var buffer = accelerator.Allocate <int>(groupSize))
{
var kernel = accelerator.LoadStreamKernel <
ArrayView <int>,
SpecializedValue <int> >(SpecializedKernel);
kernel((1, groupSize), buffer.View, SpecializedValue.New(2));
kernel((1, groupSize), buffer.View, SpecializedValue.New(23));
kernel((1, groupSize), buffer.View, SpecializedValue.New(42));
}
// Scenario 2: custom structure
using (var buffer = accelerator.Allocate <int>(groupSize))
{
var kernel = accelerator.LoadStreamKernel <
ArrayView <int>,
SpecializedValue <CustomStruct> >(SpecializedCustomStructKernel);
kernel(
(1, groupSize),
buffer.View,
SpecializedValue.New(
new CustomStruct(1, 7)));
kernel(
(1, groupSize),
buffer.View,
SpecializedValue.New(
new CustomStruct(23, 42)));
}
// Scenario 3: generic kernel
using (var buffer = accelerator.Allocate <long>(groupSize))
{
var kernel = accelerator.LoadStreamKernel <
ArrayView <long>,
SpecializedValue <long> >(SpecializedGenericKernel);
kernel((1, groupSize), buffer.View, SpecializedValue.New(23L));
kernel((1, groupSize), buffer.View, SpecializedValue.New(42L));
}
}
}
}
}
private static void IlGpuKernelLocalMemory(
ArrayView2D <Real> mSquaredDistances,
ArrayView <Real> mCoordinates,
SpecializedValue <int> dimX,
SpecializedValue <int> c,
int n)
{
// Same as KernelConstants, but use both local and shared memory to increase the effective shared memory.
var coordinatesI = SharedMemory.Allocate <Real>(c * dimX);
var coordinatesJ = new IlReal2[c.Value];
var bI = Grid.IdxY * dimX;
var bJ = Grid.IdxX * dimX;
var line = Group.IdxX / (dimX / 2);
var tid = Group.IdxX % (dimX / 2);
var isActive = bJ + tid * 2 < n;
for (int k = 0; k != c.Value; ++k)
{
if (bI + Group.IdxX < n)
{
coordinatesI[k * dimX + Group.IdxX] = mCoordinates[k * n + bI + Group.IdxX];
}
if (isActive)
{
var mCoordinates2 = mCoordinates.Cast <IlReal2>();
coordinatesJ[k] = mCoordinates2[(k * n + bJ) / 2 + tid];
}
}
Group.Barrier();
if (isActive)
{
for (int i = line; i < dimX && bI + i < n; i += 2)
{
var dist = default(IlReal2);
for (int k = 0; k != c.Value; ++k)
{
var coord1 = coordinatesI[k * dimX + i];
var coord2 = coordinatesJ[k];
var diff = new IlReal2(coord1 - coord2.X, coord1 - coord2.Y);
dist += diff * diff;
}
var dst = mSquaredDistances.Cast <IlReal2>();
dst[bJ / 2 + tid, bI + i] = dist;
}
}
}
/// <summary>
/// Kernel that multiplies two UInt64 values to produce a UInt64 value.
/// </summary>
public static void MultiplyUInt128Kernel(
Index1D index,
ArrayView <UInt128> buffer,
SpecializedValue <ulong> constant)
{
// NB: Need to convert index.X to ulong, so that %2 will use the correct PTX register type.
ulong multiplier = (ulong)index.X;
CudaAsm.Emit("mul.hi.u64 %0, %1, %2;", out ulong high, constant.Value, multiplier);
CudaAsm.Emit("mul.lo.u64 %0, %1, %2;", out ulong low, constant.Value, multiplier);
buffer[index] = new UInt128(high, low);
}
// The specialization functionality supports user-defined types, as long as they
// are value types and implement the IEquatable interface (and have useful
// GetHashCode and Equals implementations).
public static void SpecializedCustomStructKernel(
ArrayView <int> view,
SpecializedValue <CustomStruct> specialized)
{
var globalIndex = Grid.GlobalIndex.X;
// Note that an implicit conversion from a specialized value to
// a non-specialized value is possible. But: not the other way around ;)
CustomStruct customValue = specialized;
// The value is specialized and the additional optimization passes will
// perform constant propagation to create an 'optimized' store with a single constant
// value (in this case)
view[globalIndex] = customValue.Value1 + customValue.Value2;
}
public Camera(Vec3 origin, Vec3 lookAt, Vec3 up, int width, int height, float verticalFov)
{
this.width = new SpecializedValue <int>(width);
this.height = new SpecializedValue <int>(height);
this.verticalFov = verticalFov;
this.origin = origin;
this.lookAt = lookAt;
this.up = up;
axis = OrthoNormalBasis.fromZY(Vec3.unitVector(lookAt - origin), up);
aspectRatio = ((float)width / (float)height);
cameraPlaneDist = 1.0f / XMath.Tan(verticalFov * XMath.PI / 360.0f);
reciprocalHeight = 1.0f / height;
reciprocalWidth = 1.0f / width;
}
/// <summary>
/// Demonstrates using the mul.hi.u64 and mul.lo.u64 inline PTX instructions to
/// multiply two UInt64 values to produce a UInt128 value.
/// </summary>
static void MultiplyUInt128(CudaAccelerator accelerator)
{
using var buffer = accelerator.Allocate1D <UInt128>(1024);
var kernel = accelerator.LoadAutoGroupedStreamKernel <Index1D, ArrayView <UInt128>, SpecializedValue <ulong> >(MultiplyUInt128Kernel);
kernel(
(int)buffer.Length,
buffer.View,
SpecializedValue.New(ulong.MaxValue));
var results = buffer.GetAsArray1D();
for (var i = 0; i < results.Length; i++)
{
Console.WriteLine($"[{i}] = {results[i]}");
}
}
private static void IlGpuOptimisedImpl(
CudaAccelerator gpu,
Real[] mSquaredDistances,
Real[] mCoordinates,
int c,
int n,
string name,
Action <ArrayView2D <Real>, ArrayView <Real>, SpecializedValue <int>, SpecializedValue <int>, int> kernelFunc)
{
using var cudaSquaredDistance = gpu.Allocate <Real>(n, n);
using var cudaCoordinates = gpu.Allocate(mCoordinates);
var timer = Stopwatch.StartNew();
const int blockSize = 128;
var gridSize = Util.DivUp(n, blockSize);
var lp = ((gridSize, gridSize, 1), (blockSize, 1, 1));
gpu.Launch(kernelFunc, gpu.DefaultStream, lp, cudaSquaredDistance.View, cudaCoordinates.View, SpecializedValue.New(blockSize), SpecializedValue.New(c), n);
gpu.Synchronize();
Util.PrintPerformance(timer, name, n, c, n);
cudaSquaredDistance.CopyTo(mSquaredDistances, (0, 0), 0, (n, n));
}
/// <summary>
/// The actual unique kernel implementation.
/// </summary>
/// <typeparam name="T">The element type.</typeparam>
/// <typeparam name="TComparisonOperation">The comparison operation.</typeparam>
/// <param name="input">The input view.</param>
/// <param name="output">The output view to store the new length.</param>
/// <param name="sequentialGroupExecutor">
/// The sequential group executor to use.
/// </param>
/// <param name="tileSize">The tile size.</param>
/// <param name="numIterationsPerGroup">
/// The number of iterations per group.
/// </param>
internal static void UniqueKernel <T, TComparisonOperation>(
ArrayView <T> input,
ArrayView <long> output,
SequentialGroupExecutor sequentialGroupExecutor,
SpecializedValue <int> tileSize,
Index1D numIterationsPerGroup)
where T : unmanaged
where TComparisonOperation : struct, IComparisonOperation <T>
{
TComparisonOperation comparison = default;
var isFirstGrid = Grid.IdxX == 0;
var tileInfo = new TileInfo(input.IntLength, numIterationsPerGroup);
// Sync groups and wait for the current one to become active
sequentialGroupExecutor.Wait();
var temp = SharedMemory.Allocate <bool>(tileSize);
var startIdx = Grid.ComputeGlobalIndex(Grid.IdxX, 0);
for (
int i = tileInfo.StartIndex;
i < tileInfo.MaxLength;
i += Group.DimX)
{
if (Group.IsFirstThread && i == tileInfo.StartIndex && isFirstGrid)
{
temp[0] = true;
}
else
{
var currIdx = i;
var prevIdx = Group.IsFirstThread && i == tileInfo.StartIndex
? output[0] - 1
: currIdx - 1;
temp[currIdx - startIdx] =
comparison.Compare(input[currIdx], input[prevIdx]) != 0;
}
}
Group.Barrier();
if (Group.IsFirstThread)
{
var offset = isFirstGrid ? 0 : output[0];
var maxLength =
XMath.Min(startIdx + temp.IntLength, tileInfo.MaxLength) - startIdx;
for (var i = 0; i < maxLength; i++)
{
if (temp[i])
{
input[offset++] = input[startIdx + i];
}
}
output[0] = offset;
}
MemoryFence.DeviceLevel();
Group.Barrier();
sequentialGroupExecutor.Release();
}
/// <summary>
/// Performs the first radix-sort pass.
/// </summary>
/// <typeparam name="T">The element type.</typeparam>
/// <typeparam name="TOperation">The radix-sort operation.</typeparam>
/// <typeparam name="TSpecialization">The specialization type.</typeparam>
/// <param name="view">The input view to use.</param>
/// <param name="counter">The global counter view.</param>
/// <param name="groupSize">The number of threads in the group.</param>
/// <param name="numGroups">The number of virtually launched groups.</param>
/// <param name="paddedLength">The padded length of the input view.</param>
/// <param name="shift">The bit shift to use.</param>
internal static void RadixSortKernel1 <T, TOperation, TSpecialization>(
ArrayView <T> view,
ArrayView <int> counter,
SpecializedValue <int> groupSize,
int numGroups,
int paddedLength,
int shift)
where T : unmanaged
where TOperation : struct, IRadixSortOperation <T>
where TSpecialization : struct, IRadixSortSpecialization
{
TSpecialization specialization = default;
var scanMemory = SharedMemory.Allocate <int>(
groupSize * specialization.UnrollFactor);
int gridIdx = Grid.IdxX;
for (
int i = Grid.GlobalIndex.X;
i < paddedLength;
i += GridExtensions.GridStrideLoopStride)
{
bool inRange = i < view.Length;
// Read value from global memory
TOperation operation = default;
T value = operation.DefaultValue;
if (inRange)
{
value = view[i];
}
var bits = operation.ExtractRadixBits(
value,
shift,
specialization.UnrollFactor - 1);
for (int j = 0; j < specialization.UnrollFactor; ++j)
{
scanMemory[Group.IdxX + groupSize * j] = 0;
}
if (inRange)
{
scanMemory[Group.IdxX + groupSize * bits] = 1;
}
Group.Barrier();
for (int j = 0; j < specialization.UnrollFactor; ++j)
{
var address = Group.IdxX + groupSize * j;
scanMemory[address] =
GroupExtensions.ExclusiveScan <int, AddInt32>(scanMemory[address]);
}
Group.Barrier();
if (Group.IdxX == Group.DimX - 1)
{
// Write counters to global memory
for (int j = 0; j < specialization.UnrollFactor; ++j)
{
ref var newOffset = ref scanMemory[Group.IdxX + groupSize * j];
newOffset += Utilities.Select(inRange & j == bits, 1, 0);
counter[j * numGroups + gridIdx] = newOffset;
}
}
Group.Barrier();
var gridSize = gridIdx * Group.DimX;
Index1 pos = gridSize + scanMemory[Group.IdxX + groupSize * bits] -
Utilities.Select(inRange & Group.IdxX == Group.DimX - 1, 1, 0);
for (int j = 1; j <= bits; ++j)
{
pos += scanMemory[groupSize * j - 1] +
Utilities.Select(j - 1 == bits, 1, 0);
}
// Pre-sort the current value into the corresponding segment
if (inRange)
{
view[pos] = value;
}
Group.Barrier();
gridIdx += Grid.DimX;
}
static void ScalarConsecutiveOperationKernel(Index1 index, ArrayView <float> OutPut, ArrayView <float> Input, float Scalar, SpecializedValue <int> operation)
{
switch ((Operations)operation.Value)
{
case Operations.multiplication:
OutPut[index] = Input[index] * Scalar;
break;
case Operations.addition:
OutPut[index] = Input[index] + Scalar;
break;
case Operations.subtraction:
OutPut[index] = Input[index] - Scalar;
break;
case Operations.flipSubtraction:
OutPut[index] = Scalar - Input[index];
break;
case Operations.division:
OutPut[index] = Input[index] / Scalar;
break;
case Operations.inverseDivision:
OutPut[index] = Scalar / Input[index];
break;
case Operations.power:
OutPut[index] = XMath.Pow(Input[index], Scalar);
break;
case Operations.powerFlipped:
OutPut[index] = XMath.Pow(Scalar, Input[index]);
break;
case Operations.squareOfDiffs:
OutPut[index] = XMath.Pow((Input[index] - Scalar), 2f);
break;
}
}
