/// <inheritdoc cref="IDisposable.Dispose"/>
public void Dispose()
{
_mapped?.Dispose();
_mapped = null;
_memory.Dispose();
_memory = null;
}
public void Dispose()
{
ThrowIfDisposed();
memory.Dispose();
disposed = true;
}
public void Resize(ref DeviceMemory <float> buffer, int length)
{
if (buffer.Length >= length)
{
return;
}
buffer.Dispose();
buffer = _worker.Malloc <float>(length);
}
public void Dispose()
{
Fence?.Dispose();
Memory.Dispose();
Buffer.Dispose();
if (WriteUsingStagingBuffer)
{
StagingMemory?.Dispose();
StagingBuffer?.Dispose();
StagingCommandBuffer?.Dispose();
}
}
protected virtual void Dispose(bool disposing)
{
if (disposing && !_disposed)
{
_blas.Dispose();
_cuda.Dispose();
_cache.Dispose();
//if(_solver.IsValueCreated)
// _solver.Value.Dispose();
_numerics.Dispose();
_disposed = true;
}
}
public static VulkanBuffer Vertex <T>(VulkanContext ctx, T[] vertices) where T : struct
{
long size = vertices.Length * Interop.SizeOf <T>();
// Create a staging buffer that is writable by host.
Buffer stagingBuffer = ctx.Device.CreateBuffer(new BufferCreateInfo(size, BufferUsages.TransferSrc));
MemoryRequirements stagingReq = stagingBuffer.GetMemoryRequirements();
int stagingMemoryTypeIndex = ctx.MemoryProperties.MemoryTypes.IndexOf(
stagingReq.MemoryTypeBits,
MemoryProperties.HostVisible | MemoryProperties.HostCoherent);
DeviceMemory stagingMemory = ctx.Device.AllocateMemory(new MemoryAllocateInfo(stagingReq.Size, stagingMemoryTypeIndex));
IntPtr vertexPtr = stagingMemory.Map(0, stagingReq.Size);
Interop.Write(vertexPtr, vertices);
stagingMemory.Unmap();
stagingBuffer.BindMemory(stagingMemory);
// Create a device local buffer where the vertex data will be copied and which will be used for rendering.
Buffer buffer = ctx.Device.CreateBuffer(new BufferCreateInfo(size, BufferUsages.VertexBuffer | BufferUsages.TransferDst));
MemoryRequirements req = buffer.GetMemoryRequirements();
int memoryTypeIndex = ctx.MemoryProperties.MemoryTypes.IndexOf(
req.MemoryTypeBits,
MemoryProperties.DeviceLocal);
DeviceMemory memory = ctx.Device.AllocateMemory(new MemoryAllocateInfo(req.Size, memoryTypeIndex));
buffer.BindMemory(memory);
// Copy the data from staging buffers to device local buffers.
CommandBuffer cmdBuffer = ctx.GraphicsCommandPool.AllocateBuffers(new CommandBufferAllocateInfo(CommandBufferLevel.Primary, 1))[0];
cmdBuffer.Begin(new CommandBufferBeginInfo(CommandBufferUsages.OneTimeSubmit));
cmdBuffer.CmdCopyBuffer(stagingBuffer, buffer, new BufferCopy(size));
cmdBuffer.End();
// Submit.
Fence fence = ctx.Device.CreateFence();
ctx.GraphicsQueue.Submit(new SubmitInfo(commandBuffers: new[] { cmdBuffer }), fence);
fence.Wait();
// Cleanup.
fence.Dispose();
cmdBuffer.Dispose();
stagingBuffer.Dispose();
stagingMemory.Dispose();
return(new VulkanBuffer(buffer, memory, vertices.Length));
}
internal static VKBuffer Index(Context ctx, int[] indices)
{
long size = indices.Length * sizeof(int);
// Create staging buffer.
VulkanCore.Buffer stagingBuffer = ctx.Device.CreateBuffer(new BufferCreateInfo(size, BufferUsages.TransferSrc));
MemoryRequirements stagingReq = stagingBuffer.GetMemoryRequirements();
int stagingMemoryTypeIndex = ctx.MemoryProperties.MemoryTypes.IndexOf(
stagingReq.MemoryTypeBits,
MemoryProperties.HostVisible | MemoryProperties.HostCoherent);
DeviceMemory stagingMemory = ctx.Device.AllocateMemory(new MemoryAllocateInfo(stagingReq.Size, stagingMemoryTypeIndex));
IntPtr indexPtr = stagingMemory.Map(0, stagingReq.Size);
Interop.Write(indexPtr, indices);
stagingMemory.Unmap();
stagingBuffer.BindMemory(stagingMemory);
// Create a device local buffer.
VulkanCore.Buffer buffer = ctx.Device.CreateBuffer(new BufferCreateInfo(size, BufferUsages.IndexBuffer | BufferUsages.TransferDst));
MemoryRequirements req = buffer.GetMemoryRequirements();
int memoryTypeIndex = ctx.MemoryProperties.MemoryTypes.IndexOf(
req.MemoryTypeBits,
MemoryProperties.DeviceLocal);
DeviceMemory memory = ctx.Device.AllocateMemory(new MemoryAllocateInfo(req.Size, memoryTypeIndex));
buffer.BindMemory(memory);
// Copy the data from staging buffer to device local buffer.
CommandBuffer cmdBuffer = ctx.GraphicsCommandPool.AllocateBuffers(new CommandBufferAllocateInfo(CommandBufferLevel.Primary, 1))[0];
cmdBuffer.Begin(new CommandBufferBeginInfo(CommandBufferUsages.OneTimeSubmit));
cmdBuffer.CmdCopyBuffer(stagingBuffer, buffer, new BufferCopy(size));
cmdBuffer.End();
// Submit.
Fence fence = ctx.Device.CreateFence();
ctx.GraphicsQueue.Submit(new SubmitInfo(commandBuffers: new[] { cmdBuffer }), fence);
fence.Wait();
// Cleanup.
fence.Dispose();
cmdBuffer.Dispose();
stagingBuffer.Dispose();
stagingMemory.Dispose();
return(new VKBuffer(ctx, buffer, memory, null, indices.Length, size));
}
public void Cleanup()
{
/*
* Clean up all Vulkan Resources.
*/
bufferMemory.Dispose();
buffer.Dispose();
computeShaderModule.Dispose();
descriptorPool.Dispose();
descriptorSetLayout.Dispose();
pipelineLayout.Dispose();
foreach (Pipeline pipeline in pipelines)
{
pipeline.Dispose();
}
commandPool.Dispose();
device.Dispose();
instance.Dispose();
}
public void Dispose()
{
if (Interlocked.Exchange(ref _disposed, 1) != 0)
{
return;
}
switch (MemoryType)
{
case CuMemoryType.Device:
DeviceMemory.Dispose();
DeviceMemory = CuDeviceMemory.Empty;
break;
case CuMemoryType.Host:
Marshal.FreeHGlobal(Bytes);
Bytes = IntPtr.Zero;
break;
}
}
double[] ExecuteGPU(int size)
{
// いったんメインメモリ上に変数を準備
List <double> a = new List <double>();
List <double> b = new List <double>();
for (int i = 0; i < size * size; i++)
{
a.Add(i + 1);
b.Add((i + 1) * 10);
}
// デバイス上にメモリを転送
DeviceMemory memory = new DeviceMemory();
memory.Add <double>("a", a);
memory.Add <double>("b", b);
memory.Alloc <double>("c", size * size);
// 関数の実行
CallMethod(
"matrixDot", size, size,
memory["a"],
memory["b"],
memory["c"],
size,
size
);
// 全てのスレッドが終了するまで待つ
context.Synchronize();
// 結果を取得して出力画面に表示
double[] result = memory.Read <double>("c", size * size);
// リソースを解放する
memory.Dispose();
return(result);
}
//[overrideFuctions]
protected override void Dispose(bool disposing)
{
foreach (var resource in _resources)
{
CUDAInterop.cuSafeCall(CUDAInterop.cuGraphicsUnregisterResource(resource));
}
foreach (var buffer in _buffers)
{
CUDAInterop.cuSafeCall(CUDAInterop.cuGLUnregisterBufferObject(buffer));
}
if (_buffers.Length > 0)
{
GL.DeleteBuffers(_buffers.Length, _buffers);
}
if (disposing)
{
_vel.Dispose();
_disposeSimulators();
_worker.Dispose();
}
base.Dispose(disposing);
}
public void Resize(ref DeviceMemory<float> buffer, int length)
{
if (buffer.Length >= length) return;
buffer.Dispose();
buffer = _worker.Malloc<float>(length);
}
public void Dispose()
{
DeviceResource.Dispose();
DeviceMemory.Dispose();
}
internal static VulkanImage Texture2D(VulkanContext ctx, TextureData tex2D)
{
Buffer stagingBuffer = ctx.Device.CreateBuffer(
new BufferCreateInfo(tex2D.Mipmaps[0].Size, BufferUsages.TransferSrc));
MemoryRequirements stagingMemReq = stagingBuffer.GetMemoryRequirements();
int heapIndex = ctx.MemoryProperties.MemoryTypes.IndexOf(
stagingMemReq.MemoryTypeBits, MemoryProperties.HostVisible);
DeviceMemory stagingMemory = ctx.Device.AllocateMemory(
new MemoryAllocateInfo(stagingMemReq.Size, heapIndex));
stagingBuffer.BindMemory(stagingMemory);
IntPtr ptr = stagingMemory.Map(0, stagingMemReq.Size);
Interop.Write(ptr, tex2D.Mipmaps[0].Data);
stagingMemory.Unmap();
// Setup buffer copy regions for each mip level.
var bufferCopyRegions = new BufferImageCopy[tex2D.Mipmaps.Length];
int offset = 0;
for (int i = 0; i < bufferCopyRegions.Length; i++)
{
bufferCopyRegions = new[]
{
new BufferImageCopy
{
ImageSubresource = new ImageSubresourceLayers(ImageAspects.Color, i, 0, 1),
ImageExtent = tex2D.Mipmaps[0].Extent,
BufferOffset = offset
}
};
offset += tex2D.Mipmaps[i].Size;
}
// Create optimal tiled target image.
Image image = ctx.Device.CreateImage(new ImageCreateInfo
{
ImageType = ImageType.Image2D,
Format = tex2D.Format,
MipLevels = tex2D.Mipmaps.Length,
ArrayLayers = 1,
Samples = SampleCounts.Count1,
Tiling = ImageTiling.Optimal,
SharingMode = SharingMode.Exclusive,
InitialLayout = ImageLayout.Undefined,
Extent = tex2D.Mipmaps[0].Extent,
Usage = ImageUsages.Sampled | ImageUsages.TransferDst
});
MemoryRequirements imageMemReq = image.GetMemoryRequirements();
int imageHeapIndex = ctx.MemoryProperties.MemoryTypes.IndexOf(
imageMemReq.MemoryTypeBits, MemoryProperties.DeviceLocal);
DeviceMemory memory = ctx.Device.AllocateMemory(new MemoryAllocateInfo(imageMemReq.Size, imageHeapIndex));
image.BindMemory(memory);
var subresourceRange = new ImageSubresourceRange(ImageAspects.Color, 0, tex2D.Mipmaps.Length, 0, 1);
// Copy the data from staging buffers to device local buffers.
CommandBuffer cmdBuffer = ctx.GraphicsCommandPool.AllocateBuffers(new CommandBufferAllocateInfo(CommandBufferLevel.Primary, 1))[0];
cmdBuffer.Begin(new CommandBufferBeginInfo(CommandBufferUsages.OneTimeSubmit));
cmdBuffer.CmdPipelineBarrier(PipelineStages.TopOfPipe, PipelineStages.TopOfPipe,
imageMemoryBarriers: new[]
{
new ImageMemoryBarrier(
image, subresourceRange,
0, Accesses.TransferWrite,
ImageLayout.Undefined, ImageLayout.TransferDstOptimal)
});
cmdBuffer.CmdCopyBufferToImage(stagingBuffer, image, ImageLayout.TransferDstOptimal, bufferCopyRegions);
cmdBuffer.CmdPipelineBarrier(PipelineStages.TopOfPipe, PipelineStages.TopOfPipe,
imageMemoryBarriers: new[]
{
new ImageMemoryBarrier(
image, subresourceRange,
Accesses.TransferWrite, Accesses.ShaderRead,
ImageLayout.TransferDstOptimal, ImageLayout.ShaderReadOnlyOptimal)
});
cmdBuffer.End();
// Submit.
Fence fence = ctx.Device.CreateFence();
ctx.GraphicsQueue.Submit(new SubmitInfo(commandBuffers: new[] { cmdBuffer }), fence);
fence.Wait();
// Cleanup staging resources.
fence.Dispose();
stagingMemory.Dispose();
stagingBuffer.Dispose();
// Create image view.
ImageView view = image.CreateView(new ImageViewCreateInfo(tex2D.Format, subresourceRange));
return(new VulkanImage(image, memory, view, tex2D.Format));
}
private void TestCompile()
{
// プログラムのコンパイル (cu から PTX へ)
RuntimeCompiler compiler = new RuntimeCompiler();
//compiler.AddHeader("curand_kernel.h", @"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0\include\curand_kernel.h");
compiler.AddOptions(
RuntimeCompiler.OPTION_TARGET_30,
RuntimeCompiler.OPTION_FMAD_FALSE,
RuntimeCompiler.OPTION_LINE_INFO,
RuntimeCompiler.OPTION_DEVICE_AS_DEFAULT_EXECUTION_SPACE
//RuntimeCompiler.OPTION_INCLUDE_PATH_ + @"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0\include\"
);
string ptx = compiler.Compile("addKernel.cu", addKernelString);
if (ptx == null)
{
Console.WriteLine("Compile Error:");
Console.WriteLine();
Console.WriteLine(compiler.Log);
return;
}
// コンパイル時のログを出力画面に表示
Console.WriteLine("----- <Compile Log>");
Console.WriteLine(compiler.Log);
Console.WriteLine("----- </Compile Log>");
// コンパイル済みプログラムを出力画面に表示
Console.WriteLine("----- <PTX>");
Console.WriteLine(ptx);
Console.WriteLine("----- </PTX>");
// プログラムの実行準備
Device device = new Device(0);
Context context = device.CreateContext();
Module module = new Module();
Console.WriteLine(device.Name);
Console.WriteLine(device.PCIBusId);
Console.WriteLine(device.TotalMem);
//Console.WriteLine(device.GetProperties());
Console.WriteLine(context.ApiVersion);
//return;
// PTX データをロード
module.LoadData(ptx);
// いったんメインメモリ上に変数を準備
const int arraySize = 5;
List <int> a = new List <int>();
List <int> b = new List <int>();
for (int i = 0; i < arraySize; i++)
{
a.Add(i + 1);
b.Add((i + 1) * 10);
}
// デバイス上にメモリを転送
DeviceMemory memory = new DeviceMemory();
memory.Add <int>("a", a);
memory.Add <int>("b", b);
memory.Alloc <int>("c", arraySize);
// 関数の実行
module.SetBlockCount(1, 1, 1);
module.SetThreadCount(arraySize, 1, 1);
module.Excecute(
"addKernel",
memory["c"],
memory["a"],
memory["b"]
);
// 全てのスレッドが終了するまで待つ
context.Synchronize();
// 結果を取得して出力画面に表示
int[] results = memory.Read <int>("c", arraySize);
Console.WriteLine("----- <Execute Log>");
for (int i = 0; i < arraySize; i++)
{
Console.WriteLine("{0} + {1} = {2}", a[i], b[i], results[i]);
}
Console.WriteLine("----- </Execute Log>");
// リソースを解放する
memory.Dispose();
module.Dispose();
context.Dispose();
}