Warrior_EA/AI/NeuronOCLConvPool.mqh

//+------------------------------------------------------------------+
//|                                            NeuronOCLConvPool.mqh |
//|                                                     AnimateDread |
//|                                             https://www.mql5.com |
//+------------------------------------------------------------------+
//| CNeuronConvOCL/CNeuronPoolOCL - the GPU-accelerated (OpenCL +     |
//| DirectML) convolution and max-pooling layers. Both derive from   |
//| CNeuronBaseOCL (AI\Network.mqh, must already be declared) and use|
//| CBufferDouble (AI\BufferDouble.mqh). Included from Network.mqh at|
//| the exact point these classes used to sit (right after            |
//| CNeuronBaseOCL's own declaration, before CNeuronLSTMOCL), so      |
//| ordering matches the original file. Extracted verbatim (SOLID    |
//| cleanup) - no logic changes.                                       |
//+------------------------------------------------------------------+
#include "BufferDouble.mqh"
//| GPU-accelerated convolution layer (OpenCL + DirectML). Ported    |
//| from the NeuroNet_DNG reference library's CNeuronConvOCL/         |
//| CNeuronProofOCL, adapted to this project's double-precision       |
//| CBufferDouble/COpenCLMy/CDirectMLMy conventions. A single         |
//| (window+1)*window_out weight block is shared across every        |
//| sliding position - unlike CNeuronBaseOCL, where every output has  |
//| its own private weight vector.                                    |
//+------------------------------------------------------------------+
class CNeuronConvOCL : public CNeuronBaseOCL
  {
protected:
   uint              iWindow;
   uint              iStep;
   uint              iWindowOut;
   CBufferDouble     *WeightsConv;
   CBufferDouble     *DeltaWeightsConv;
   CBufferDouble     *FirstMomentumConv;
   CBufferDouble     *SecondMomentumConv;
   //--- Mini-batch accumulator for the CONVOLUTION KERNEL block. Distinct from the base class's
   //--- GradAccum, which shadows the base Weights - i.e. the outgoing dense matrix that the layer ABOVE
   //--- accumulates into. A conv neuron owns both tensors, so it needs both accumulators; sharing one
   //--- slot would have the two layers writing each other's gradients.
   CBufferDouble     *GradAccumConv;
   //---
   virtual bool      feedForward(CNeuronBaseOCL *NeuronOCL);
   virtual bool      feedForwardCPU(CNeuronBaseOCL *NeuronOCL); // pure-MQL5 mirror of FeedForwardConv
   virtual bool      updateInputWeights(CNeuronBaseOCL *NeuronOCL);
   virtual bool      accumulateInputWeightGrads(CNeuronBaseOCL *NeuronOCL);

public:
                     CNeuronConvOCL(void) : iWindow(1), iStep(1), iWindowOut(1)
     {
      WeightsConv = NULL;
      DeltaWeightsConv = NULL;
      FirstMomentumConv = NULL;
      SecondMomentumConv = NULL;
      GradAccumConv = NULL;
     }
                    ~CNeuronConvOCL(void);
   virtual bool      Init(uint numOutputs, uint myIndex, COpenCLMy *open_cl, uint window_in, uint step, uint window_out, uint units_count, ENUM_OPTIMIZATION optimization_type);
   virtual bool      Init(uint numOutputs, uint myIndex, CDirectMLMy *direct_ml, uint window_in, uint step, uint window_out, uint units_count, ENUM_OPTIMIZATION optimization_type);
   virtual bool      calcInputGradients(CNeuronBaseOCL *NeuronOCL);
   virtual bool      Save(int const file_handle);
   virtual bool      Load(int const file_handle);
   virtual int       Type(void)   const   {  return defNeuronConvOCL;  }
   // Shape accessors. A .nnw persists the window/step it was BUILT with, so a model saved by an older
   // build keeps that shape forever on load - these let EnforceTopologyContract() detect a stale
   // receptive field instead of training on silently. See AI\Network.mqh's CNet::FirstConvWindow.
   uint              Window(void)    const   {  return iWindow;     }
   uint              Step(void)      const   {  return iStep;       }
   uint              WindowOut(void) const   {  return iWindowOut;  }
   // See CNeuronBaseOCL::getWeights/setWeights - same pair, targeting WeightsConv instead of the
   // base class's Weights, for CNet::BlendWeightsFrom()'s EMA shadow-weight deployment.
   virtual int       getWeightsConv(double &values[])     {  return (CheckPointer(WeightsConv) == POINTER_INVALID ? 0 : WeightsConv.GetData(values));  }
   virtual bool      setWeightsConv(double &values[])
     {
      if(CheckPointer(WeightsConv) == POINTER_INVALID)
         return false;
      if(!WeightsConv.AssignArray(values))
         return false;
      return WeightsConv.BufferWrite();
     }
   //--- The conv kernel block keeps its own moment/momentum buffers beside the base class's - see
   //--- CNet::ResetOptimizerState.
   virtual bool      ResetOptimizerState(void)
     {
      bool ok = CNeuronBaseOCL::ResetOptimizerState();
      ok = ZeroOptimizerBuffer(FirstMomentumConv) && ok;
      ok = ZeroOptimizerBuffer(SecondMomentumConv) && ok;
      ok = ZeroOptimizerBuffer(DeltaWeightsConv) && ok;
      return ok;
     }
   //--- Both accumulators - the base one for the outgoing dense matrix, this class's for the conv
   //--- kernel. See GradAccumConv's declaration for why they cannot share a slot.
   virtual bool      BeginGradAccum(void)
     {
      bool ok = CNeuronBaseOCL::BeginGradAccum();
      if(CheckPointer(GradAccumConv) != POINTER_INVALID && GradAccumConv.Total() > 0)
         ok = ZeroOptimizerBuffer(GradAccumConv) && ok;
      return ok;
     }
   virtual bool      ApplyAccumulatedGradients(double scale)
     {
      //--- Both blocks step on the SAME batch, so t advances once here rather than once per block.
      bool ok = ApplyAccumToBlock(Weights, GradAccum, FirstMomentum, SecondMomentum, DeltaWeights, scale);
      ok = ApplyAccumToBlock(WeightsConv, GradAccumConv, FirstMomentumConv, SecondMomentumConv,
                             DeltaWeightsConv, scale) && ok;
      if(optimization == ADAM)
         t++;
      return ok;
     }
  };
//+------------------------------------------------------------------+
//|                                                                  |
//+------------------------------------------------------------------+
CNeuronConvOCL::~CNeuronConvOCL(void)
  {
   if(CheckPointer(WeightsConv) != POINTER_INVALID)
      delete WeightsConv;
   if(CheckPointer(DeltaWeightsConv) != POINTER_INVALID)
      delete DeltaWeightsConv;
   if(CheckPointer(FirstMomentumConv) != POINTER_INVALID)
      delete FirstMomentumConv;
   if(CheckPointer(SecondMomentumConv) != POINTER_INVALID)
      delete SecondMomentumConv;
   if(CheckPointer(GradAccumConv) != POINTER_INVALID)
      delete GradAccumConv;
  }
//+------------------------------------------------------------------+
//|                                                                  |
//+------------------------------------------------------------------+
bool CNeuronConvOCL::Init(uint numOutputs, uint myIndex, COpenCLMy *open_cl, uint window_in, uint step, uint window_out, uint units_count, ENUM_OPTIMIZATION optimization_type)
  {
   if(window_out <= 0)
      return false;
   if(!CNeuronBaseOCL::Init(numOutputs, myIndex, open_cl, units_count * window_out, optimization_type))
      return false;
//---
   iWindow = window_in;
   iStep = step;
   iWindowOut = (uint)fmax(window_out, 1);
//---
   int count = (int)((iWindow + 1) * iWindowOut);
   if(CheckPointer(WeightsConv) == POINTER_INVALID)
     {
      WeightsConv = new CBufferDouble();
      if(CheckPointer(WeightsConv) == POINTER_INVALID)
         return false;
     }
   if(!WeightsConv.Reserve(count))
      return false;
   // Fan-in-scaled (LeCun-uniform) init - see CNeuronBaseOCL::Init's OpenCL overload for the full
   // rationale; fan-in here is the conv window size.
   double weighScale = 1.0 / MathSqrt((double)iWindow + 1.0);
   for(int i = 0; i < count; i++)
     {
      double weigh = ((MathRand() + 1) / 32768.0 - 0.5) * 2.0 * weighScale;
      if(weigh == 0)
         weigh = 0.001;
      if(!WeightsConv.Add(weigh))
         return false;
     }
   if(!WeightsConv.BufferCreate(OpenCL))
      return false;
//---
   if(optimization == SGD)
     {
      if(CheckPointer(DeltaWeightsConv) == POINTER_INVALID)
        {
         DeltaWeightsConv = new CBufferDouble();
         if(CheckPointer(DeltaWeightsConv) == POINTER_INVALID)
            return false;
        }
      if(!DeltaWeightsConv.BufferInit(count, 0))
         return false;
      if(!DeltaWeightsConv.BufferCreate(OpenCL))
         return false;
     }
   else
     {
      if(CheckPointer(FirstMomentumConv) == POINTER_INVALID)
        {
         FirstMomentumConv = new CBufferDouble();
         if(CheckPointer(FirstMomentumConv) == POINTER_INVALID)
            return false;
        }
      if(!FirstMomentumConv.BufferInit(count, 0))
         return false;
      if(!FirstMomentumConv.BufferCreate(OpenCL))
         return false;
      //---
      if(CheckPointer(SecondMomentumConv) == POINTER_INVALID)
        {
         SecondMomentumConv = new CBufferDouble();
         if(CheckPointer(SecondMomentumConv) == POINTER_INVALID)
            return false;
        }
      if(!SecondMomentumConv.BufferInit(count, 0))
         return false;
      if(!SecondMomentumConv.BufferCreate(OpenCL))
         return false;
     }
//---
   return true;
  }
//+------------------------------------------------------------------+
//| DirectML/D3D12 tier equivalent of Init(COpenCLMy*) above.        |
//+------------------------------------------------------------------+
bool CNeuronConvOCL::Init(uint numOutputs, uint myIndex, CDirectMLMy *direct_ml, uint window_in, uint step, uint window_out, uint units_count, ENUM_OPTIMIZATION optimization_type)
  {
   if(window_out <= 0)
      return false;
   if(!CNeuronBaseOCL::Init(numOutputs, myIndex, direct_ml, units_count * window_out, optimization_type))
      return false;
//---
   iWindow = window_in;
   iStep = step;
   iWindowOut = (uint)fmax(window_out, 1);
//---
   int count = (int)((iWindow + 1) * iWindowOut);
   if(CheckPointer(WeightsConv) == POINTER_INVALID)
     {
      WeightsConv = new CBufferDouble();
      if(CheckPointer(WeightsConv) == POINTER_INVALID)
         return false;
     }
   if(!WeightsConv.Reserve(count))
      return false;
   // Fan-in-scaled (LeCun-uniform) init - see the matching OpenCL Init() overload above.
   double weighScale = 1.0 / MathSqrt((double)iWindow + 1.0);
   for(int i = 0; i < count; i++)
     {
      double weigh = ((MathRand() + 1) / 32768.0 - 0.5) * 2.0 * weighScale;
      if(weigh == 0)
         weigh = 0.001;
      if(!WeightsConv.Add(weigh))
         return false;
     }
   if(!WeightsConv.BufferCreate(DirectML))
      return false;
//---
   if(optimization == SGD)
     {
      if(CheckPointer(DeltaWeightsConv) == POINTER_INVALID)
        {
         DeltaWeightsConv = new CBufferDouble();
         if(CheckPointer(DeltaWeightsConv) == POINTER_INVALID)
            return false;
        }
      if(!DeltaWeightsConv.BufferInit(count, 0))
         return false;
      if(!DeltaWeightsConv.BufferCreate(DirectML))
         return false;
     }
   else
     {
      if(CheckPointer(FirstMomentumConv) == POINTER_INVALID)
        {
         FirstMomentumConv = new CBufferDouble();
         if(CheckPointer(FirstMomentumConv) == POINTER_INVALID)
            return false;
        }
      if(!FirstMomentumConv.BufferInit(count, 0))
         return false;
      if(!FirstMomentumConv.BufferCreate(DirectML))
         return false;
      //---
      if(CheckPointer(SecondMomentumConv) == POINTER_INVALID)
        {
         SecondMomentumConv = new CBufferDouble();
         if(CheckPointer(SecondMomentumConv) == POINTER_INVALID)
            return false;
        }
      if(!SecondMomentumConv.BufferInit(count, 0))
         return false;
      if(!SecondMomentumConv.BufferCreate(DirectML))
         return false;
     }
//---
   return true;
  }
//+------------------------------------------------------------------+
//|                                                                  |
//+------------------------------------------------------------------+
bool CNeuronConvOCL::feedForward(CNeuronBaseOCL *NeuronOCL)
  {
   if(CheckPointer(NeuronOCL) == POINTER_INVALID)
      return false;
   int positions = Output.Total() / (int)iWindowOut;
   if(CheckPointer(DirectML) != POINTER_INVALID)
     {
      if(!DirectML.FeedForwardConv(WeightsConv.GetIndex(), NeuronOCL.getOutputIndex(), Output.GetIndex(),
                                    NeuronOCL.Neurons(), (int)iStep, (int)iWindow, (int)iWindowOut, NativeActivationCode(activation), positions))
        {
         Print(__FUNCTION__ + ": " + DirectML.BackendName() + " FeedForwardConv failed, error " + IntegerToString(DirectML.LastError()));
         return false;
        }
      return Output.BufferRead();
     }
   if(CheckPointer(OpenCL) == POINTER_INVALID)
      return false;
   uint global_work_offset[1] = {0};
   uint global_work_size[1];
   global_work_size[0] = (uint)positions;
   OpenCL.SetArgumentBuffer(def_k_FeedForwardConv, def_k_ffc_matrix_w, WeightsConv.GetIndex());
   OpenCL.SetArgumentBuffer(def_k_FeedForwardConv, def_k_ffc_matrix_i, NeuronOCL.getOutputIndex());
   OpenCL.SetArgumentBuffer(def_k_FeedForwardConv, def_k_ffc_matrix_o, Output.GetIndex());
   OpenCL.SetArgument(def_k_FeedForwardConv, def_k_ffc_inputs, NeuronOCL.Neurons());
   OpenCL.SetArgument(def_k_FeedForwardConv, def_k_ffc_step, (int)iStep);
   OpenCL.SetArgument(def_k_FeedForwardConv, def_k_ffc_window_in, (int)iWindow);
   OpenCL.SetArgument(def_k_FeedForwardConv, def_k_ffc_window_out, (int)iWindowOut);
   OpenCL.SetArgument(def_k_FeedForwardConv, def_k_ffc_activation, NativeActivationCode(activation));
   if(!OpenCL.Execute(def_k_FeedForwardConv, 1, global_work_offset, global_work_size))
     {
      printf("Error of execution kernel FeedForwardConv: %d", GetLastError());
      return false;
     }
   //--- Output stays GPU-resident; see the note in CNeuronBaseOCL::feedForward().
   return true;
  }
//+------------------------------------------------------------------+
//| Pure-MQL5 double-precision mirror of Network.cl's FeedForwardConv |
//| kernel (host buffers only) - the CPU inference path. Filters live |
//| in this layer's own WeightsConv, [window_out][window_in+1] with   |
//| the +1 bias last; inputs are the previous layer's Output.         |
//+------------------------------------------------------------------+
bool CNeuronConvOCL::feedForwardCPU(CNeuronBaseOCL *NeuronOCL)
  {
   if(CheckPointer(NeuronOCL) == POINTER_INVALID || CheckPointer(Output) == POINTER_INVALID || CheckPointer(WeightsConv) == POINTER_INVALID)
      return false;
   int inputs     = NeuronOCL.Neurons();
   int window_in  = (int)iWindow;
   int step       = (int)iStep;
   int window_out = (int)iWindowOut;
   if(window_out <= 0)
      return false;
   int positions  = Output.Total() / window_out;
   int wTotal     = WeightsConv.Total();
   for(int i = 0; i < positions; i++)
     {
      int shift_out = window_out * i;
      int shift_in  = step * i;
      for(int out = 0; out < window_out; out++)
        {
         int shift = (window_in + 1) * out;
         if(shift + window_in >= wTotal)
            return false;
         int stop = (window_in <= (inputs - shift_in)) ? window_in : (inputs - shift_in);
         double sum = 0.0;
         for(int k = 0; k < stop; k++)
            sum += NeuronOCL.OutputHost(shift_in + k) * WeightsConv.At(shift + k);
         sum += WeightsConv.At(shift + window_in); // bias
         switch(activation)
           {
            case TANH:
               sum = tanh(sum);
               break;
            case SIGMOID:
               sum = 1.0 / (1.0 + exp(-MathMax(-50.0, MathMin(50.0, sum))));
               break;
            case PRELU:
               if(sum < 0.0)
                  sum *= 0.01;
               break;
           }
         if(!Output.Update(out + shift_out, sum))
            return false;
        }
     }
   return true;
  }
//+------------------------------------------------------------------+
//| Writes the gradient into NeuronOCL (the EARLIER/input-side layer) |
//| - opposite call direction from the dense calcHiddenGradients, but |
//| matches the reference library's own CNeuronConvOCL convention.    |
//+------------------------------------------------------------------+
bool CNeuronConvOCL::calcInputGradients(CNeuronBaseOCL *NeuronOCL)
  {
   if(CheckPointer(NeuronOCL) == POINTER_INVALID)
      return false;
   int outputs = Neurons();
   if(CheckPointer(DirectML) != POINTER_INVALID)
     {
      //--- NativeActivationCode, NOT a raw enum cast: the backends number activations differently
      //--- from ENUM_ACTIVATION (see NativeActivationCode's declaration comment). The raw cast sent
      //--- NONE(0) into the kernels' tanh branch (which clamps and damps a BN layer's unbounded
      //--- z-score outputs), TANH(1) into the sigmoid branch, and PRELU(3) past every branch (no
      //--- derivative at all). Dormant in current presets only because the conv sits at layer 1
      //--- with nothing trainable below it - any deeper conv placement activates it silently.
      if(!DirectML.CalcHiddenGradientConv(WeightsConv.GetIndex(), getGradientIndex(), NeuronOCL.getOutputIndex(), NeuronOCL.getGradientIndex(),
                                           outputs, (int)iStep, (int)iWindow, (int)iWindowOut, NativeActivationCode(NeuronOCL.Activation()), NeuronOCL.Neurons()))
        {
         Print(__FUNCTION__ + ": " + DirectML.BackendName() + " CalcHiddenGradientConv failed, error " + IntegerToString(DirectML.LastError()));
         return false;
        }
      double temp[];
      return NeuronOCL.getGradient(temp) > 0;
     }
   if(CheckPointer(OpenCL) == POINTER_INVALID)
      return false;
   uint global_work_offset[1] = {0};
   uint global_work_size[1];
   global_work_size[0] = NeuronOCL.Neurons();
   OpenCL.SetArgumentBuffer(def_k_CalcHiddenGradientConv, def_k_chgc_matrix_w, WeightsConv.GetIndex());
   OpenCL.SetArgumentBuffer(def_k_CalcHiddenGradientConv, def_k_chgc_matrix_g, getGradientIndex());
   OpenCL.SetArgumentBuffer(def_k_CalcHiddenGradientConv, def_k_chgc_matrix_o, NeuronOCL.getOutputIndex());
   OpenCL.SetArgumentBuffer(def_k_CalcHiddenGradientConv, def_k_chgc_matrix_ig, NeuronOCL.getGradientIndex());
   OpenCL.SetArgument(def_k_CalcHiddenGradientConv, def_k_chgc_outputs, outputs);
   OpenCL.SetArgument(def_k_CalcHiddenGradientConv, def_k_chgc_step, (int)iStep);
   OpenCL.SetArgument(def_k_CalcHiddenGradientConv, def_k_chgc_window_in, (int)iWindow);
   OpenCL.SetArgument(def_k_CalcHiddenGradientConv, def_k_chgc_window_out, (int)iWindowOut);
   //--- NativeActivationCode, NOT a raw enum cast - see the DirectML branch's comment above.
   OpenCL.SetArgument(def_k_CalcHiddenGradientConv, def_k_chgc_activation, NativeActivationCode(NeuronOCL.Activation()));
   if(!OpenCL.Execute(def_k_CalcHiddenGradientConv, 1, global_work_offset, global_work_size))
     {
      printf("Error of execution kernel CalcHiddenGradientConv: %d", GetLastError());
      return false;
     }
   //--- NeuronOCL's Gradient stays GPU-resident; its own calcHiddenGradients/calcInputGradients
   //--- reads it via getGradientIndex(). The old getGradient(temp) call was a discarded-result
   //--- sync (GetData() BufferRead()s internally) with no consumer of the read - pure overhead.
   return true;
  }
//+------------------------------------------------------------------+
//|                                                                  |
//+------------------------------------------------------------------+
bool CNeuronConvOCL::updateInputWeights(CNeuronBaseOCL *NeuronOCL)
  {
   if(CheckPointer(NeuronOCL) == POINTER_INVALID)
      return false;
   int inputs = NeuronOCL.Neurons();
   if(CheckPointer(DirectML) != POINTER_INVALID)
     {
      if(optimization == SGD)
        {
         if(!DirectML.UpdateWeightsConvMomentum(WeightsConv.GetIndex(), getGradientIndex(), NeuronOCL.getOutputIndex(), DeltaWeightsConv.GetIndex(),
               inputs, eta, alpha, (int)iWindow, (int)iWindowOut, (int)iStep, 0))
           {
            Print(__FUNCTION__ + ": " + DirectML.BackendName() + " UpdateWeightsConvMomentum failed, error " + IntegerToString(DirectML.LastError()));
            return false;
           }
        }
      else
        {
         double lt = eta * sqrt(1 - pow(b2, t)) / (1 - pow(b1, t));
         if(!DirectML.UpdateWeightsConvAdam(WeightsConv.GetIndex(), getGradientIndex(), NeuronOCL.getOutputIndex(),
               FirstMomentumConv.GetIndex(), SecondMomentumConv.GetIndex(), inputs, lt, b1, b2, (int)iWindow, (int)iWindowOut, (int)iStep))
           {
            Print(__FUNCTION__ + ": " + DirectML.BackendName() + " UpdateWeightsConvAdam failed, error " + IntegerToString(DirectML.LastError()));
            return false;
           }
         t++;
        }
      //--- WeightsConv stays DLL-resident; feedForward reads it via GetIndex() (same as the OpenCL
      //--- branch below). Save()/BlendWeightsFrom() BufferRead() on demand.
      return true;
     }
   if(CheckPointer(OpenCL) == POINTER_INVALID)
      return false;
   uint global_work_offset[1] = {0};
   uint global_work_size[1];
   if(optimization == SGD)
     {
      global_work_size[0] = WeightsConv.Total();
      OpenCL.SetArgumentBuffer(def_k_UpdateWeightsConvMomentum, def_k_uwcm_matrix_w, WeightsConv.GetIndex());
      OpenCL.SetArgumentBuffer(def_k_UpdateWeightsConvMomentum, def_k_uwcm_matrix_g, getGradientIndex());
      OpenCL.SetArgumentBuffer(def_k_UpdateWeightsConvMomentum, def_k_uwcm_matrix_i, NeuronOCL.getOutputIndex());
      OpenCL.SetArgumentBuffer(def_k_UpdateWeightsConvMomentum, def_k_uwcm_matrix_dw, DeltaWeightsConv.GetIndex());
      OpenCL.SetArgument(def_k_UpdateWeightsConvMomentum, def_k_uwcm_inputs, inputs);
      OpenCL.SetArgument(def_k_UpdateWeightsConvMomentum, def_k_uwcm_learning_rates, (float)eta);
      OpenCL.SetArgument(def_k_UpdateWeightsConvMomentum, def_k_uwcm_momentum, (float)alpha);
      OpenCL.SetArgument(def_k_UpdateWeightsConvMomentum, def_k_uwcm_window_in, (int)iWindow);
      OpenCL.SetArgument(def_k_UpdateWeightsConvMomentum, def_k_uwcm_window_out, (int)iWindowOut);
      OpenCL.SetArgument(def_k_UpdateWeightsConvMomentum, def_k_uwcm_step, (int)iStep);
      OpenCL.SetArgument(def_k_UpdateWeightsConvMomentum, def_k_uwcm_optimizer, 0);
      ResetLastError();
      if(!OpenCL.Execute(def_k_UpdateWeightsConvMomentum, 1, global_work_offset, global_work_size))
        {
         printf("Error of execution kernel UpdateWeightsConvMomentum: %d", GetLastError());
         return false;
        }
     }
   else
     {
      global_work_size[0] = iWindow + 1;
      double lt = eta * sqrt(1 - pow(b2, t)) / (1 - pow(b1, t));
      OpenCL.SetArgumentBuffer(def_k_UpdateWeightsConvAdam, def_k_uwca_matrix_w, WeightsConv.GetIndex());
      OpenCL.SetArgumentBuffer(def_k_UpdateWeightsConvAdam, def_k_uwca_matrix_g, getGradientIndex());
      OpenCL.SetArgumentBuffer(def_k_UpdateWeightsConvAdam, def_k_uwca_matrix_i, NeuronOCL.getOutputIndex());
      OpenCL.SetArgumentBuffer(def_k_UpdateWeightsConvAdam, def_k_uwca_matrix_m, FirstMomentumConv.GetIndex());
      OpenCL.SetArgumentBuffer(def_k_UpdateWeightsConvAdam, def_k_uwca_matrix_v, SecondMomentumConv.GetIndex());
      OpenCL.SetArgument(def_k_UpdateWeightsConvAdam, def_k_uwca_inputs, inputs);
      OpenCL.SetArgument(def_k_UpdateWeightsConvAdam, def_k_uwca_l, (float)lt);
      OpenCL.SetArgument(def_k_UpdateWeightsConvAdam, def_k_uwca_b1, (float)b1);
      OpenCL.SetArgument(def_k_UpdateWeightsConvAdam, def_k_uwca_b2, (float)b2);
      OpenCL.SetArgument(def_k_UpdateWeightsConvAdam, def_k_uwca_window_in, (int)iWindow);
      OpenCL.SetArgument(def_k_UpdateWeightsConvAdam, def_k_uwca_window_out, (int)iWindowOut);
      OpenCL.SetArgument(def_k_UpdateWeightsConvAdam, def_k_uwca_step, (int)iStep);
      ResetLastError();
      if(!OpenCL.Execute(def_k_UpdateWeightsConvAdam, 1, global_work_offset, global_work_size))
        {
         printf("Error of execution kernel UpdateWeightsConvAdam: %d", GetLastError());
         return false;
        }
      t++;
     }
   //--- WeightsConv stays GPU-resident; see the note in CNeuronBaseOCL::updateInputWeights().
   return true;
  }
//+------------------------------------------------------------------+
//| MINI-BATCH ACCUMULATE (conv kernel block). Batched counterpart of |
//| updateInputWeights above - identical dispatch, identical operands,|
//| but it only ADDS into GradAccumConv. The base class's own         |
//| accumulator (the outgoing dense matrix) is filled separately by   |
//| the layer ABOVE this one calling its accumulate on this neuron.   |
//+------------------------------------------------------------------+
bool CNeuronConvOCL::accumulateInputWeightGrads(CNeuronBaseOCL *NeuronOCL)
  {
   if(CheckPointer(NeuronOCL) == POINTER_INVALID || CheckPointer(WeightsConv) == POINTER_INVALID)
      return false;
   if(!EnsureGradAccumFor(GradAccumConv, WeightsConv))
      return false;
   int inputs = NeuronOCL.Neurons();
   if(CheckPointer(DirectML) != POINTER_INVALID)
     {
      if(!DirectML.AccumulateWeightGradConv(GradAccumConv.GetIndex(), getGradientIndex(), NeuronOCL.getOutputIndex(),
                                            inputs, (int)iWindow, (int)iWindowOut, (int)iStep))
        {
         Print(__FUNCTION__ + ": " + DirectML.BackendName() + " AccumulateWeightGradConv failed, error " + IntegerToString(DirectML.LastError()));
         return false;
        }
      return true;
     }
   if(CheckPointer(OpenCL) == POINTER_INVALID)
      return false;
   uint global_work_offset[1] = {0};
   uint global_work_size[1];
   //--- Flat over the whole kernel block, matching AccumulateWeightGradConv's indexing (and
   //--- UpdateWeightsConvMomentum's), NOT the Adam kernel's (window_in+1) shape.
   global_work_size[0] = WeightsConv.Total();
   if(!OpenCL.SetArgumentBuffer(def_k_AccumulateWeightGradConv, def_k_awgc_matrix_acc, GradAccumConv.GetIndex()))
      return false;
   if(!OpenCL.SetArgumentBuffer(def_k_AccumulateWeightGradConv, def_k_awgc_matrix_g, getGradientIndex()))
      return false;
   if(!OpenCL.SetArgumentBuffer(def_k_AccumulateWeightGradConv, def_k_awgc_matrix_i, NeuronOCL.getOutputIndex()))
      return false;
   if(!OpenCL.SetArgument(def_k_AccumulateWeightGradConv, def_k_awgc_inputs, inputs))
      return false;
   if(!OpenCL.SetArgument(def_k_AccumulateWeightGradConv, def_k_awgc_window_in, (int)iWindow))
      return false;
   if(!OpenCL.SetArgument(def_k_AccumulateWeightGradConv, def_k_awgc_window_out, (int)iWindowOut))
      return false;
   if(!OpenCL.SetArgument(def_k_AccumulateWeightGradConv, def_k_awgc_step, (int)iStep))
      return false;
   ResetLastError();
   if(!OpenCL.Execute(def_k_AccumulateWeightGradConv, 1, global_work_offset, global_work_size))
     {
      printf("Error of execution kernel AccumulateWeightGradConv: %d", GetLastError());
      return false;
     }
   return true;
  }
//+------------------------------------------------------------------+
//|                                                                  |
//+------------------------------------------------------------------+
bool CNeuronConvOCL::Save(const int file_handle)
  {
   if(!CNeuronBaseOCL::Save(file_handle))
      return false;
   if(FileWriteInteger(file_handle, (int)iWindow, INT_VALUE) < INT_VALUE)
      return false;
   if(FileWriteInteger(file_handle, (int)iStep, INT_VALUE) < INT_VALUE)
      return false;
   if(FileWriteInteger(file_handle, (int)iWindowOut, INT_VALUE) < INT_VALUE)
      return false;
   if(CheckPointer(WeightsConv) == POINTER_INVALID || !WeightsConv.BufferRead() || !WeightsConv.Save(file_handle))
      return false;
   if(optimization == SGD)
     {
      if(CheckPointer(DeltaWeightsConv) == POINTER_INVALID || !DeltaWeightsConv.BufferRead() || !DeltaWeightsConv.Save(file_handle))
         return false;
     }
   else
     {
      if(CheckPointer(FirstMomentumConv) == POINTER_INVALID || !FirstMomentumConv.BufferRead() || !FirstMomentumConv.Save(file_handle))
         return false;
      if(CheckPointer(SecondMomentumConv) == POINTER_INVALID || !SecondMomentumConv.BufferRead() || !SecondMomentumConv.Save(file_handle))
         return false;
     }
//---
   return true;
  }
//+------------------------------------------------------------------+
//|                                                                  |
//+------------------------------------------------------------------+
bool CNeuronConvOCL::Load(const int file_handle)
  {
   if(!CNeuronBaseOCL::Load(file_handle))
      return false;
   iWindow = (uint)FileReadInteger(file_handle, INT_VALUE);
   iStep = (uint)FileReadInteger(file_handle, INT_VALUE);
   iWindowOut = (uint)FileReadInteger(file_handle, INT_VALUE);
//---
   if(CheckPointer(WeightsConv) == POINTER_INVALID)
     {
      WeightsConv = new CBufferDouble();
      if(CheckPointer(WeightsConv) == POINTER_INVALID)
         return false;
     }
   if(WeightsConv.GetIndex() >= 0)
      WeightsConv.BufferFree();
   if(!WeightsConv.Load(file_handle))
      return false;
   if(!BackendBufferCreate(WeightsConv))
      return false;
//---
   if(optimization == SGD)
     {
      if(CheckPointer(DeltaWeightsConv) == POINTER_INVALID)
        {
         DeltaWeightsConv = new CBufferDouble();
         if(CheckPointer(DeltaWeightsConv) == POINTER_INVALID)
            return false;
        }
      if(DeltaWeightsConv.GetIndex() >= 0)
         DeltaWeightsConv.BufferFree();
      if(!DeltaWeightsConv.Load(file_handle))
         return false;
      if(!BackendBufferCreate(DeltaWeightsConv))
         return false;
     }
   else
     {
      if(CheckPointer(FirstMomentumConv) == POINTER_INVALID)
        {
         FirstMomentumConv = new CBufferDouble();
         if(CheckPointer(FirstMomentumConv) == POINTER_INVALID)
            return false;
        }
      if(FirstMomentumConv.GetIndex() >= 0)
         FirstMomentumConv.BufferFree();
      if(!FirstMomentumConv.Load(file_handle))
         return false;
      if(!BackendBufferCreate(FirstMomentumConv))
         return false;
      //---
      if(CheckPointer(SecondMomentumConv) == POINTER_INVALID)
        {
         SecondMomentumConv = new CBufferDouble();
         if(CheckPointer(SecondMomentumConv) == POINTER_INVALID)
            return false;
        }
      if(SecondMomentumConv.GetIndex() >= 0)
         SecondMomentumConv.BufferFree();
      if(!SecondMomentumConv.Load(file_handle))
         return false;
      if(!BackendBufferCreate(SecondMomentumConv))
         return false;
     }
//---
   return true;
  }
//+------------------------------------------------------------------+
//| GPU-accelerated max-pooling layer (OpenCL + DirectML). No weights,|
//| so no updateInputWeights work - just a sliding max. Ported from   |
//| the NeuroNet_DNG reference's CNeuronProofOCL kernels.              |
//+------------------------------------------------------------------+
class CNeuronPoolOCL : public CNeuronBaseOCL
  {
protected:
   uint              iWindow;
   uint              iStep;
   //---
   virtual bool      feedForward(CNeuronBaseOCL *NeuronOCL);
   virtual bool      feedForwardCPU(CNeuronBaseOCL *NeuronOCL); // pure-MQL5 mirror of FeedForwardProof
   virtual bool      updateInputWeights(CNeuronBaseOCL *NeuronOCL) { return true; }

public:
                     CNeuronPoolOCL(void) : iWindow(2), iStep(1) {}
                    ~CNeuronPoolOCL(void) {}
   virtual bool      Init(uint numOutputs, uint myIndex, COpenCLMy *open_cl, uint window, uint step, uint units_count, ENUM_OPTIMIZATION optimization_type);
   virtual bool      Init(uint numOutputs, uint myIndex, CDirectMLMy *direct_ml, uint window, uint step, uint units_count, ENUM_OPTIMIZATION optimization_type);
   virtual bool      calcInputGradients(CNeuronBaseOCL *NeuronOCL);
   virtual bool      Save(int const file_handle);
   virtual bool      Load(int const file_handle);
   virtual int       Type(void)   const   {  return defNeuronPoolOCL;  }
  };
//+------------------------------------------------------------------+
//|                                                                  |
//+------------------------------------------------------------------+
bool CNeuronPoolOCL::Init(uint numOutputs, uint myIndex, COpenCLMy *open_cl, uint window, uint step, uint units_count, ENUM_OPTIMIZATION optimization_type)
  {
   if(!CNeuronBaseOCL::Init(numOutputs, myIndex, open_cl, units_count, optimization_type))
      return false;
   iWindow = window;
   iStep = step;
   return true;
  }
//+------------------------------------------------------------------+
//|                                                                  |
//+------------------------------------------------------------------+
bool CNeuronPoolOCL::Init(uint numOutputs, uint myIndex, CDirectMLMy *direct_ml, uint window, uint step, uint units_count, ENUM_OPTIMIZATION optimization_type)
  {
   if(!CNeuronBaseOCL::Init(numOutputs, myIndex, direct_ml, units_count, optimization_type))
      return false;
   iWindow = window;
   iStep = step;
   return true;
  }
//+------------------------------------------------------------------+
//|                                                                  |
//+------------------------------------------------------------------+
bool CNeuronPoolOCL::feedForward(CNeuronBaseOCL *NeuronOCL)
  {
   if(CheckPointer(NeuronOCL) == POINTER_INVALID)
      return false;
   int outputs = Output.Total();
   if(CheckPointer(DirectML) != POINTER_INVALID)
     {
      if(!DirectML.FeedForwardProof(NeuronOCL.getOutputIndex(), Output.GetIndex(), NeuronOCL.Neurons(), (int)iWindow, (int)iStep, outputs))
        {
         Print(__FUNCTION__ + ": " + DirectML.BackendName() + " FeedForwardProof failed, error " + IntegerToString(DirectML.LastError()));
         return false;
        }
      return Output.BufferRead();
     }
   if(CheckPointer(OpenCL) == POINTER_INVALID)
      return false;
   uint offset1[1] = {0};
   uint size1[1] = {(uint)outputs};
   OpenCL.SetArgumentBuffer(def_k_FeedForwardProof, def_k_ffp_matrix_i, NeuronOCL.getOutputIndex());
   OpenCL.SetArgumentBuffer(def_k_FeedForwardProof, def_k_ffp_matrix_o, Output.GetIndex());
   OpenCL.SetArgument(def_k_FeedForwardProof, def_k_ffp_inputs, NeuronOCL.Neurons());
   OpenCL.SetArgument(def_k_FeedForwardProof, def_k_ffp_window, (int)iWindow);
   OpenCL.SetArgument(def_k_FeedForwardProof, def_k_ffp_step, (int)iStep);
   if(!OpenCL.Execute(def_k_FeedForwardProof, 1, offset1, size1))
     {
      printf("Error of execution kernel FeedForwardProof: %d", GetLastError());
      return false;
     }
   //--- Output stays GPU-resident; see the note in CNeuronBaseOCL::feedForward().
   return true;
  }
//+------------------------------------------------------------------+
//| Pure-MQL5 mirror of Network.cl's FeedForwardProof kernel (sliding |
//| max-pool over the previous layer's Output). Host buffers only.    |
//+------------------------------------------------------------------+
bool CNeuronPoolOCL::feedForwardCPU(CNeuronBaseOCL *NeuronOCL)
  {
   if(CheckPointer(NeuronOCL) == POINTER_INVALID || CheckPointer(Output) == POINTER_INVALID)
      return false;
   int outputs = Output.Total();
   int inputs  = NeuronOCL.Neurons();
   int window  = (int)iWindow;
   int step    = (int)iStep;
   for(int i = 0; i < outputs; i++)
     {
      int pos = i * step;
      if(pos >= inputs)
         return false;
      double result = NeuronOCL.OutputHost(pos);
      for(int k = 1; k < window; k++)
        {
         int shift = k + pos;
         if(shift >= inputs)
            break;
         result = MathMax(result, NeuronOCL.OutputHost(shift));
        }
      if(!Output.Update(i, result))
         return false;
     }
   return true;
  }
//+------------------------------------------------------------------+
//| Inverted-call convention, same as Conv/LSTM's calcInputGradients. |
//+------------------------------------------------------------------+
bool CNeuronPoolOCL::calcInputGradients(CNeuronBaseOCL *NeuronOCL)
  {
   if(CheckPointer(NeuronOCL) == POINTER_INVALID)
      return false;
   int outputs = Neurons();
   int inputs = NeuronOCL.Neurons();
   if(CheckPointer(DirectML) != POINTER_INVALID)
     {
      if(!DirectML.CalcInputGradientProof(NeuronOCL.getOutputIndex(), getGradientIndex(), getOutputIndex(), NeuronOCL.getGradientIndex(),
                                           outputs, (int)iWindow, (int)iStep, inputs))
        {
         Print(__FUNCTION__ + ": " + DirectML.BackendName() + " CalcInputGradientProof failed, error " + IntegerToString(DirectML.LastError()));
         return false;
        }
      double temp[];
      return NeuronOCL.getGradient(temp) > 0;
     }
   if(CheckPointer(OpenCL) == POINTER_INVALID)
      return false;
   uint offset1[1] = {0};
   uint size1[1] = {(uint)inputs};
   OpenCL.SetArgumentBuffer(def_k_CalcInputGradientProof, def_k_cigp_matrix_i, NeuronOCL.getOutputIndex());
   OpenCL.SetArgumentBuffer(def_k_CalcInputGradientProof, def_k_cigp_matrix_g, getGradientIndex());
   OpenCL.SetArgumentBuffer(def_k_CalcInputGradientProof, def_k_cigp_matrix_o, getOutputIndex());
   OpenCL.SetArgumentBuffer(def_k_CalcInputGradientProof, def_k_cigp_matrix_ig, NeuronOCL.getGradientIndex());
   OpenCL.SetArgument(def_k_CalcInputGradientProof, def_k_cigp_outputs, outputs);
   OpenCL.SetArgument(def_k_CalcInputGradientProof, def_k_cigp_window, (int)iWindow);
   OpenCL.SetArgument(def_k_CalcInputGradientProof, def_k_cigp_step, (int)iStep);
   if(!OpenCL.Execute(def_k_CalcInputGradientProof, 1, offset1, size1))
     {
      printf("Error of execution kernel CalcInputGradientProof: %d", GetLastError());
      return false;
     }
   //--- NeuronOCL's Gradient stays GPU-resident; see the note in
   //--- CNeuronConvOCL::calcInputGradients().
   return true;
  }
//+------------------------------------------------------------------+
//|                                                                  |
//+------------------------------------------------------------------+
bool CNeuronPoolOCL::Save(const int file_handle)
  {
   if(!CNeuronBaseOCL::Save(file_handle))
      return false;
   if(FileWriteInteger(file_handle, (int)iWindow, INT_VALUE) < INT_VALUE)
      return false;
   if(FileWriteInteger(file_handle, (int)iStep, INT_VALUE) < INT_VALUE)
      return false;
   return true;
  }
//+------------------------------------------------------------------+
//|                                                                  |
//+------------------------------------------------------------------+
bool CNeuronPoolOCL::Load(const int file_handle)
  {
   if(!CNeuronBaseOCL::Load(file_handle))
      return false;
   iWindow = (uint)FileReadInteger(file_handle, INT_VALUE);
   iStep = (uint)FileReadInteger(file_handle, INT_VALUE);
   return true;
  }