protected override NdArray NeedPreviousForwardGpu(NdArray x)
{
Real[] y = NoBias ? new Real[OutputCount * x.BatchCount] : GetBiasedValue(x.BatchCount);
using (ComputeBuffer <Real> gpuX = new ComputeBuffer <Real>(Weaver.Context, ComputeMemoryFlags.None, x.Data))
using (ComputeBuffer <Real> gpuY = new ComputeBuffer <Real>(Weaver.Context, ComputeMemoryFlags.None, y))
{
ForwardKernel.SetMemoryArgument(0, gpuX);
ForwardKernel.SetMemoryArgument(1, gpuW);
ForwardKernel.SetMemoryArgument(2, gpuY);
ForwardKernel.SetValueArgument(3, OutputCount);
ForwardKernel.SetValueArgument(4, InputCount);
Weaver.CommandQueue.Execute
(
ForwardKernel,
null,
new long[] { OutputCount, x.BatchCount },
null,
null
);
Weaver.CommandQueue.Flush();
//for less cpu use. this is 65% of computation time (10.4ms on 1080ti).
ASleep(6.5);
Weaver.CommandQueue.Finish();
Weaver.CommandQueue.ReadFromBuffer(gpuY, ref y, true, null);
}
return(NdArray.Convert(y, new[] { OutputCount }, x.BatchCount, this));
}
protected override NdArray NeedPreviousForwardGpu(NdArray x)
{
Real[] y = this.NoBias ? new Real[OutputCount * x.BatchCount] : GetBiasedValue(x.BatchCount);
using (ComputeBuffer <Real> gpuX = new ComputeBuffer <Real>(Weaver.Context, ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer, x.Data))
using (ComputeBuffer <Real> gpuW = new ComputeBuffer <Real>(Weaver.Context, ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer, this.Weight.Data))
using (ComputeBuffer <Real> gpuY = new ComputeBuffer <Real>(Weaver.Context, ComputeMemoryFlags.ReadWrite | ComputeMemoryFlags.CopyHostPointer, y))
{
ForwardKernel.SetMemoryArgument(0, gpuX);
ForwardKernel.SetMemoryArgument(1, gpuW);
ForwardKernel.SetMemoryArgument(2, gpuY);
ForwardKernel.SetValueArgument(3, this.OutputCount);
ForwardKernel.SetValueArgument(4, this.InputCount);
Weaver.CommandQueue.Execute
(
ForwardKernel,
null,
new long[] { OutputCount, x.BatchCount },
null,
null
);
Weaver.CommandQueue.Finish();
Weaver.CommandQueue.ReadFromBuffer(gpuY, ref y, true, null);
}
return(NdArray.Convert(y, new[] { OutputCount }, x.BatchCount, this));
}
public override NdArray SingleInputForward(NdArray x)
{
//フラグチェック
if (!IsParallel)
{
return(base.SingleInputForward(x));
}
Real[] y = this.NoBias ? new Real[OutputCount * x.BatchCount] : GetBiasedValue(x.BatchCount);
using (ComputeBuffer <Real> gpuX = new ComputeBuffer <Real>(OpenCL.Context, ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.UseHostPointer, x.Data))
using (ComputeBuffer <Real> gpuW = new ComputeBuffer <Real>(OpenCL.Context, ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.UseHostPointer, this.Weight.Data))
using (ComputeBuffer <Real> gpuY = new ComputeBuffer <Real>(OpenCL.Context, ComputeMemoryFlags.ReadWrite | ComputeMemoryFlags.UseHostPointer, y))
{
ForwardKernel.SetMemoryArgument(0, gpuX);
ForwardKernel.SetMemoryArgument(1, gpuW);
ForwardKernel.SetMemoryArgument(2, gpuY);
ForwardKernel.SetValueArgument(3, this.OutputCount);
ForwardKernel.SetValueArgument(4, this.InputCount);
OpenCL.CommandQueue.Execute
(
ForwardKernel,
null,
new long[] { OutputCount, x.BatchCount },
null,
null
);
OpenCL.CommandQueue.Finish();
OpenCL.CommandQueue.ReadFromBuffer(gpuY, ref y, true, null);
}
return(NdArray.Convert(y, new[] { OutputCount }, x.BatchCount, this));
}
protected override NdArray NeedPreviousForwardGpu(NdArray x)
{
var ytemp = NoBias ? new Real[OutputCount * x.BatchCount] : GetBiasedValue(x.BatchCount);
var y = GetArray("y", ytemp.Length, ComputeMemoryFlags.ReadWrite | ComputeMemoryFlags.CopyHostPointer);
y.Write(ytemp);
x.Data.Switch(Common.ComputeDeviceTypes.Gpu);
Weight.Data.Switch(Common.ComputeDeviceTypes.Gpu);
y.Switch(Common.ComputeDeviceTypes.Gpu);
var gpuX = x.Data.GetBuffer();
var gpuW = Weight.Data.GetBuffer();
var gpuY = y.GetBuffer();
ForwardKernel.SetMemoryArgument(0, gpuX);
ForwardKernel.SetMemoryArgument(1, gpuW);
ForwardKernel.SetMemoryArgument(2, gpuY);
ForwardKernel.SetValueArgument(3, OutputCount);
ForwardKernel.SetValueArgument(4, InputCount);
Weaver.CommandQueue.Execute
(
ForwardKernel,
null,
new long[] { OutputCount, x.BatchCount },
null,
null
);
Weaver.CommandQueue.Flush();
Weaver.CommandQueue.Finish();
return(new NdArray(y, GetArray("y.Grad", y.Length, ComputeMemoryFlags.ReadWrite | ComputeMemoryFlags.CopyHostPointer), new[] { OutputCount }, x.BatchCount, this));
}
public override NdArray SingleInputForward(NdArray input)
{
//フラグチェック
if (!IsParallel)
{
return(base.SingleInputForward(input));
}
int outputHeight = (input.Shape[1] - 1) * this.StrideY + this.KernelHeight - this.PadY * 2;
int outputWidth = (input.Shape[2] - 1) * this.StrideX + this.KernelWidth - this.PadX * 2;
Real[] result = new Real[input.BatchCount * this.OutputCount * outputWidth * outputHeight];
using (ComputeBuffer <Real> gpuX = new ComputeBuffer <Real>(OpenCL.Context, ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.UseHostPointer, input.Data))
using (ComputeBuffer <Real> gpuW = new ComputeBuffer <Real>(OpenCL.Context, ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.UseHostPointer, this.Weight.Data))
using (ComputeBuffer <Real> gpub = new ComputeBuffer <Real>(OpenCL.Context, ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.UseHostPointer, this.NoBias ? new Real[OutputCount] : this.Bias.Data))
using (ComputeBuffer <Real> gpuY = new ComputeBuffer <Real>(OpenCL.Context, ComputeMemoryFlags.WriteOnly | ComputeMemoryFlags.AllocateHostPointer, result.Length))
{
ForwardKernel.SetMemoryArgument(0, gpuX);
ForwardKernel.SetMemoryArgument(1, gpuW);
ForwardKernel.SetMemoryArgument(2, gpub);
ForwardKernel.SetMemoryArgument(3, gpuY);
ForwardKernel.SetValueArgument(4, input.Shape[1]);
ForwardKernel.SetValueArgument(5, input.Shape[2]);
ForwardKernel.SetValueArgument(6, input.Length);
ForwardKernel.SetValueArgument(7, outputWidth);
ForwardKernel.SetValueArgument(8, outputHeight);
ForwardKernel.SetValueArgument(9, this.StrideX);
ForwardKernel.SetValueArgument(10, this.StrideY);
ForwardKernel.SetValueArgument(11, this.PadX);
ForwardKernel.SetValueArgument(12, this.PadY);
ForwardKernel.SetValueArgument(13, this.KernelHeight);
ForwardKernel.SetValueArgument(14, this.KernelWidth);
ForwardKernel.SetValueArgument(15, this.OutputCount);
ForwardKernel.SetValueArgument(16, this.InputCount);
OpenCL.CommandQueue.Execute
(
ForwardKernel,
null,
new long[] { input.BatchCount *OutputCount, outputHeight, outputWidth },
null,
null
);
OpenCL.CommandQueue.Finish();
OpenCL.CommandQueue.ReadFromBuffer(gpuY, ref result, true, null);
}
return(NdArray.Convert(result, new[] { this.OutputCount, outputHeight, outputWidth }, input.BatchCount, this));
}
protected override NdArray NeedPreviousForwardGpu(NdArray input)
{
int outputHeight = (int)Math.Floor((input.Shape[1] - this._kHeight + this._padY * 2.0) / this._strideY) + 1;
int outputWidth = (int)Math.Floor((input.Shape[2] - this._kWidth + this._padX * 2.0) / this._strideX) + 1;
Real[] result = new Real[this.OutputCount * outputHeight * outputWidth * input.BatchCount];
using (ComputeBuffer <Real> gpuX = new ComputeBuffer <Real>(Weaver.Context, ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer, input.Data))
using (ComputeBuffer <Real> gpuW = new ComputeBuffer <Real>(Weaver.Context, ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer, this.Weight.Data))
//TODO
using (ComputeBuffer <Real> gpub = new ComputeBuffer <Real>(Weaver.Context, ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer, this.NoBias ? new Real[OutputCount] : (Real[])this.Bias.Data))
using (ComputeBuffer <Real> gpuY = new ComputeBuffer <Real>(Weaver.Context, ComputeMemoryFlags.WriteOnly | ComputeMemoryFlags.AllocateHostPointer, result.Length))
{
ForwardKernel.SetMemoryArgument(0, gpuX);
ForwardKernel.SetMemoryArgument(1, gpuW);
ForwardKernel.SetMemoryArgument(2, gpub);
ForwardKernel.SetMemoryArgument(3, gpuY);
ForwardKernel.SetValueArgument(4, input.Shape[1]);
ForwardKernel.SetValueArgument(5, input.Shape[2]);
ForwardKernel.SetValueArgument(6, input.Length);
ForwardKernel.SetValueArgument(7, outputWidth);
ForwardKernel.SetValueArgument(8, outputHeight);
ForwardKernel.SetValueArgument(9, this._strideX);
ForwardKernel.SetValueArgument(10, this._strideY);
ForwardKernel.SetValueArgument(11, this._padX);
ForwardKernel.SetValueArgument(12, this._padY);
ForwardKernel.SetValueArgument(13, this._kHeight);
ForwardKernel.SetValueArgument(14, this._kWidth);
ForwardKernel.SetValueArgument(15, this.OutputCount);
ForwardKernel.SetValueArgument(16, this.InputCount);
Weaver.CommandQueue.Execute
(
ForwardKernel,
null,
new long[] { input.BatchCount *OutputCount, outputHeight, outputWidth },
null,
null
);
Weaver.CommandQueue.Finish();
Weaver.CommandQueue.ReadFromBuffer(gpuY, ref result, true, null);
}
return(NdArray.Convert(result, new[] { this.OutputCount, outputHeight, outputWidth }, input.BatchCount, this));
}
protected override NdArray NeedPreviousForwardGpu([NotNull] NdArray input)
{
int outputHeight = (input.Shape[1] - 1) * _subSampleY + _kHeight - _trimY * 2;
int outputWidth = (input.Shape[2] - 1) * _subSampleX + _kWidth - _trimX * 2;
Real[] result = new Real[input.BatchCount * OutputCount * outputWidth * outputHeight];
using (ComputeBuffer <Real> gpuX = new ComputeBuffer <Real>(Weaver.Context, ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer, input.Data))
{
using (ComputeBuffer <Real> gpuW = new ComputeBuffer <Real>(Weaver.Context, ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer, Weight.Data))
{
using (ComputeBuffer <Real> gpub = new ComputeBuffer <Real>(Weaver.Context, ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer, NoBias ? new Real[OutputCount] : Bias.Data))
{
using (ComputeBuffer <Real> gpuY = new ComputeBuffer <Real>(Weaver.Context, ComputeMemoryFlags.WriteOnly | ComputeMemoryFlags.AllocateHostPointer, result.Length))
{
ForwardKernel.SetMemoryArgument(0, gpuX);
ForwardKernel.SetMemoryArgument(1, gpuW);
ForwardKernel.SetMemoryArgument(2, gpub);
ForwardKernel.SetMemoryArgument(3, gpuY);
ForwardKernel.SetValueArgument(4, input.Shape[1]);
ForwardKernel.SetValueArgument(5, input.Shape[2]);
ForwardKernel.SetValueArgument(6, input.Length);
ForwardKernel.SetValueArgument(7, outputWidth);
ForwardKernel.SetValueArgument(8, outputHeight);
ForwardKernel.SetValueArgument(9, _subSampleX);
ForwardKernel.SetValueArgument(10, _subSampleY);
ForwardKernel.SetValueArgument(11, _trimX);
ForwardKernel.SetValueArgument(12, _trimY);
ForwardKernel.SetValueArgument(13, _kHeight);
ForwardKernel.SetValueArgument(14, _kWidth);
ForwardKernel.SetValueArgument(15, OutputCount);
ForwardKernel.SetValueArgument(16, InputCount);
Weaver.CommandQueue.Execute(ForwardKernel, null, new long[] { input.BatchCount *OutputCount, outputHeight, outputWidth },
null, null);
Weaver.CommandQueue.Finish();
Weaver.CommandQueue.ReadFromBuffer(gpuY, ref result, true, null);
}
}
}
}
return(NdArray.Convert(result, new[] { OutputCount, outputHeight, outputWidth }, input.BatchCount, this));
}