/// <summary>
/// Multiplies two dense matrices and returns the resultant matrix (using tiling).
/// </summary>
/// <param name="accelerator">The Accelerator to run the multiplication on</param>
/// <param name="a">A dense MxK matrix</param>
/// <param name="b">A dense KxN matrix</param>
/// <returns>A dense MxN matrix</returns>
static float[,] MatrixMultiplyTiled(Accelerator accelerator, float[,] a, float[,] b)
{
var m = a.GetLength(0);
var ka = a.GetLength(1);
var kb = b.GetLength(0);
var n = b.GetLength(1);
if (ka != kb)
{
throw new ArgumentException($"Cannot multiply {m}x{ka} matrix by {n}x{kb} matrix", nameof(b));
}
var kernel = accelerator.LoadStreamKernel <GroupedIndex2, ArrayView2D <float>, ArrayView2D <float>, ArrayView2D <float> >(MatrixMultiplyTiledKernel);
var groupSize = new Index2(TILE_SIZE, TILE_SIZE);
var numGroups = new Index2((m + TILE_SIZE - 1) / TILE_SIZE, (n + TILE_SIZE - 1) / TILE_SIZE);
var launchDimension = new GroupedIndex2(numGroups, groupSize);
using (var aBuffer = accelerator.Allocate <float>(m, ka))
using (var bBuffer = accelerator.Allocate <float>(ka, n))
using (var cBuffer = accelerator.Allocate <float>(m, n))
{
aBuffer.CopyFrom(a, Index2.Zero, Index2.Zero, aBuffer.Extent);
bBuffer.CopyFrom(b, Index2.Zero, Index2.Zero, bBuffer.Extent);
kernel(launchDimension, aBuffer, bBuffer, cBuffer);
accelerator.Synchronize();
return(cBuffer.GetAs2DArray());
}
}
public void GroupedIndex2EntryPoint(int length)
{
var end = (int)Math.Sqrt(Accelerator.MaxNumThreadsPerGroup);
for (int i = 1; i <= end; i <<= 1)
{
var stride = new Index2(i, i);
var extent = new GroupedIndex2(
new Index2(length, length),
stride);
using var buffer = Accelerator.Allocate <int>(extent.Size);
buffer.MemSetToZero(Accelerator.DefaultStream);
Execute(extent, buffer.View, stride, extent.GridIdx);
var expected = new int[extent.Size];
for (int j = 0; j < length * length; ++j)
{
var gridIdx = Index2.ReconstructIndex(j, extent.GridIdx);
for (int k = 0; k < i * i; ++k)
{
var groupIdx = Index2.ReconstructIndex(k, extent.GroupIdx);
var idx = (gridIdx * stride + groupIdx).ComputeLinearIndex(extent.GridIdx);
expected[idx] = idx;
}
}
Verify(buffer, expected);
}
}
internal static void GroupedIndex2EntryPointKernel(
GroupedIndex2 index, ArrayView <int> output, Index2 stride, Index2 extent)
{
var idx1 = index.GridIdx.X * stride.X + index.GroupIdx.X;
var idx2 = index.GridIdx.Y * stride.Y + index.GroupIdx.Y;
var idx = idx2 * extent.X + idx1;
output[idx] = idx;
}
/// <summary>
/// The tiled matrix multiplication kernel that runs on the accelerated device.
/// </summary>
/// <param name="index">Current matrix index</param>
/// <param name="aView">An input matrix of size MxK</param>
/// <param name="bView">An input matrix of size KxN</param>
/// <param name="cView">An output matrix of size MxN</param>
static void MatrixMultiplyTiledKernel(GroupedIndex2 index, ArrayView2D <float> aView, ArrayView2D <float> bView, ArrayView2D <float> cView)
{
var global = index.ComputeGlobalIndex();
var x = index.GroupIdx.X;
var y = index.GroupIdx.Y;
var aTile = SharedMemory.Allocate2D <float>(TILE_SIZE, TILE_SIZE);
var bTile = SharedMemory.Allocate2D <float>(TILE_SIZE, TILE_SIZE);
var sum = 0.0f;
for (var i = 0; i < aView.Width; i += TILE_SIZE)
{
if (global.X < aView.Width && y + i < aView.Height)
{
aTile[x, y] = aView[global.X, y + i];
}
else
{
aTile[x, y] = 0;
}
if (x + i < bView.Width && global.Y < bView.Height)
{
bTile[x, y] = bView[x + i, global.Y];
}
else
{
bTile[x, y] = 0;
}
Group.Barrier();
for (var k = 0; k < TILE_SIZE; k++)
{
sum += aTile[new Index2(x, k)] * bTile[new Index2(k, y)];
}
Group.Barrier();
}
if (global.X < cView.Width && global.Y < cView.Height)
{
cView[global] = sum;
}
}
static void ShadowKernel1(
GroupedIndex2 index,
ArrayView <float> points,
ArrayView <float> matrices,
ArrayView <int> horizon,
ArrayView <float> test_array,
[SharedMemory(1440)]
ArrayView <int> horizon_shared)
{
var target_line = index.GridIdx.Y;
var target_sample = index.GridIdx.X;
var caster_line = index.GroupIdx.Y;
Debug.Assert(index.GroupIdx.X == 1);
// Copy horizon for a[target_line,target_sample] into shared memory
{
var dim = Group.Dimension.Y;
var len = horizon_shared.Length;
var passes = (len + (dim - 1)) / dim;
var offset = (target_line * TerrainPatch.DefaultSize + target_sample) * len;
for (var pass = 0; pass < passes; pass++)
{
var ptr = pass * dim + caster_line;
if (ptr < len) // divergence
{
horizon_shared[ptr] = horizon[ptr + offset];
}
}
}
Group.Barrier();
// Copy the matrix into registers
var pos = (target_line * TerrainPatch.DefaultSize + target_sample) * 12;
var row0x = matrices[pos++];
var row1x = matrices[pos++];
var row2x = matrices[pos++];
var row3x = matrices[pos++];
var row0y = matrices[pos++];
var row1y = matrices[pos++];
var row2y = matrices[pos++];
var row3y = matrices[pos++];
var row0z = matrices[pos++];
var row1z = matrices[pos++];
var row2z = matrices[pos++];
var row3z = matrices[pos];
for (var caster_sample = 0; caster_sample < TerrainPatch.DefaultSize; caster_sample++)
{
// Fetch the other point in local frame
var points_offset = (caster_line * TerrainPatch.DefaultSize + caster_sample) * 3;
var x_patch = points[points_offset];
var y_patch = points[points_offset + 1];
var z_patch = points[points_offset + 2];
// Transform the point to the local frame
var x = x_patch * row0x + y_patch * row1x + z_patch * row2x + row3x;
var y = x_patch * row0y + y_patch * row1y + z_patch * row2y + row3y;
var z = x_patch * row0z + y_patch * row1z + z_patch * row2z + row3z;
// Adjust for solar array height (this is temporary, and I'm not sure we want this in the final version)
z -= ObserverHeight; // meters
var azimuth = GPUMath.Atan2(y, x) + GPUMath.PI; // [0,2 PI]
var alen = GPUMath.Sqrt(x * x + y * y);
var slope = z / alen;
var slopem = slope > 2f ? 2f : slope;
slopem = slopem < -2f ? -2f : slopem;
slopem = slopem / 4f;
var slopei = (int)(slopem * 1000000);
var horizon_index = (int)(0.5f + 1439 * (azimuth / (2f * GPUMath.PI)));
Atomic.Max(horizon_shared.GetVariableView(horizon_index), slopei);
if (caster_sample == 0 && caster_line == 0 && target_line == 0 && target_sample == 0)
{
test_array[0] = x_patch;
test_array[1] = y_patch;
test_array[2] = z_patch;
test_array[3] = x;
test_array[4] = y;
test_array[5] = z;
test_array[6] = slope;
test_array[7] = slopem;
test_array[8] = slopei;
test_array[9] = row3x;
test_array[10] = row3y;
test_array[11] = row3z;
}
}
Group.Barrier();
{
var dim = Group.Dimension.Y;
var len = horizon_shared.Length;
var passes = (len + (dim - 1)) / dim;
var offset = (target_line * TerrainPatch.DefaultSize + target_sample) * len;
for (var pass = 0; pass < passes; pass++)
{
var ptr = pass * dim + caster_line;
if (ptr < len) // divergence
{
horizon[ptr + offset] = horizon_shared[ptr];
}
}
}
}
/// <summary>
/// Update the horizons of a patch based on a list of shadow casters.
/// The horizons will be in slope, not angle, format
/// </summary>
/// <param name="target"></param>
/// <param name="casters"></param>
public void UpdateHorizons(TerrainPatch target, List <TerrainPatch> casters)
{
Debug.Assert(Terrain != null);
if (casters.Count < 1)
{
return;
}
using (var context = new Context())
{
AcceleratorId aid = Accelerator.Accelerators.Where(id => id.AcceleratorType == AcceleratorType.Cuda).FirstOrDefault();
if (aid.AcceleratorType != AcceleratorType.Cuda)
{
Console.WriteLine(@"There is no CUDA accelerator present. Doing nothing.");
return;
}
using (var accelerator = Accelerator.Create(context, aid))
{
target.FillPoints(Terrain);
target.FillMatricesRelativeToPoint(Terrain, target.Points[0][0]);
// Matrices
var cpu_matrices_size = target.Height * target.Width * 12;
var basePoint = target.Points[0][0];
var cpu_matrices = MakeCPUMatrices(target);
// Horizon (load from target)
var cpu_horizon_size = target.Height * target.Width * Horizon.HorizonSamples;
var cpu_horizon = new int[cpu_horizon_size];
for (var line = 0; line < TerrainPatch.DefaultSize; line++)
{
for (var sample = 0; sample < TerrainPatch.DefaultSize; sample++)
{
var offset = (line * TerrainPatch.DefaultSize + sample) * Horizon.HorizonSamples;
var buffer = target.Horizons[line][sample].Buffer;
for (var i = 0; i < Horizon.HorizonSamples; i++)
{
cpu_horizon[i + offset] = SlopeToEncoding(buffer[i]);
}
}
}
// Caster points
var cpu_caster_points_size = casters[0].Width * casters[0].Height * 3;
var cpu_caster_points = new float[cpu_caster_points_size];
// test array
var cpu_test_array = new float[20];
using (var gpu_matrices = accelerator.Allocate <float>(cpu_matrices_size))
using (var gpu_horizon = accelerator.Allocate <int>(cpu_horizon_size))
using (var gpu_caster_points = accelerator.Allocate <float>(cpu_caster_points_size))
using (var gpu_test_array = accelerator.Allocate <float>(cpu_test_array.Length))
{
gpu_matrices.CopyFrom(cpu_matrices, 0, 0, cpu_matrices_size);
gpu_horizon.CopyFrom(cpu_horizon, 0, 0, cpu_horizon_size);
var groupSize = accelerator.MaxNumThreadsPerGroup;
var launchDimension = new GroupedIndex2(
new Index2(128, 128), // (data.Length + groupSize - 1) / groupSize, // Compute the number of groups (round up)
new Index2(1, 128));
var kernel1 = accelerator.LoadSharedMemoryStreamKernel1 <GroupedIndex2, ArrayView <float>, ArrayView <float>, ArrayView <int>, ArrayView <float>, ArrayView <int> >(ShadowKernel1);
//var stopwatch = new Stopwatch();
//stopwatch.Start();
foreach (var caster in casters)
{
caster.FillPoints(Terrain);
CopyPointsToCpuArray(caster, basePoint, cpu_caster_points);
gpu_caster_points.CopyFrom(cpu_caster_points, 0, 0, cpu_caster_points_size);
kernel1(launchDimension, gpu_caster_points, gpu_matrices, gpu_horizon, gpu_test_array);
accelerator.Synchronize();
}
// Copy out data
gpu_horizon.CopyTo(cpu_horizon, 0, 0, cpu_horizon_size);
gpu_test_array.CopyTo(cpu_test_array, 0, 0, cpu_test_array.Length);
//stopwatch.Stop();
//Console.WriteLine($"kernel time={stopwatch.Elapsed} cpu_horizon.Max()={cpu_horizon.Max()} cpu_horizon[0]={cpu_horizon[0]}");
// Update the horizons
for (var line = 0; line < TerrainPatch.DefaultSize; line++)
{
for (var sample = 0; sample < TerrainPatch.DefaultSize; sample++)
{
var offset = (line * TerrainPatch.DefaultSize + sample) * Horizon.HorizonSamples;
var buffer = target.Horizons[line][sample].Buffer;
for (var i = 0; i < Horizon.HorizonSamples; i++)
{
buffer[i] = EncodingToSlope(cpu_horizon[i + offset]);
}
}
}
//Console.WriteLine($" max slope={cpu_horizon.Select(EncodingToSlope).Max()}");
}
}
}
}
/// <summary>
/// Computes the global index of a grouped index (gridIdx, groupIdx).
/// </summary>
/// <param name="index">The grouped index.</param>
/// <returns>The computes global index.</returns>
public static Index2 ComputeGlobalIndex(GroupedIndex2 index)
{
return(ComputeGlobalIndex(index.GridIdx, index.GroupIdx));
}
