CudaEngine constructor

CudaEngine()

Implementation

CudaEngine() {
  _lib = ffi.DynamicLibrary.open('${Directory.current.path}/libmat_mul.so');
  createTensor = _lib.lookupFunction<_C_create, _D_create>('create_tensor');
  destroyTensor = _lib.lookupFunction<_C_destroy, _D_destroy>(
    'destroy_tensor',
  );
  getTensorData = _lib.lookupFunction<_C_copy, _D_copy>('get_tensor_data');
  getTensorGrad = _lib.lookupFunction<_C_copy, _D_copy>('get_tensor_grad');
  backward = _lib.lookupFunction<_C_destroy, _D_destroy>('backward');
  zeroGrad = _lib.lookupFunction<_C_destroy, _D_destroy>('zero_grad');
  tensorStep = _lib.lookupFunction<_C_step, _D_step>('tensor_step');
  addTensors = _lib.lookupFunction<_C_op2, _D_op2>('add_tensors');
  subTensors = _lib.lookupFunction<_C_op2, _D_op2>('sub_tensors');
  mulTensors = _lib.lookupFunction<_C_op2, _D_op2>('mul_tensors');
  divTensors = _lib.lookupFunction<_C_op2, _D_op2>('div_tensors');
  matmulTensors = _lib.lookupFunction<_C_op2, _D_op2>('matmul_tensors');
  powTensor = _lib.lookupFunction<_C_pow, _D_pow>('pow_tensor');
  reluTensor = _lib.lookupFunction<_C_op1, _D_op1>('relu_tensor');
  tanhTensor = _lib.lookupFunction<_C_op1, _D_op1>('tanh_tensor');
  sigmoidTensor = _lib.lookupFunction<_C_op1, _D_op1>('sigmoid_tensor');
  logTensor = _lib.lookupFunction<_C_op1, _D_op1>('log_tensor');
  aftForward = _lib.lookupFunction<_C_aft, _D_aft>('aft_forward');
  aftCrossForward = _lib.lookupFunction<_C_aft_cross, _D_aft_cross>(
    'aft_cross_forward',
  );
  concatTensors = _lib.lookupFunction<_C_concat, _D_concat>(
    'concat_tensors_gpu',
  );

  layernormForward = _lib.lookupFunction<_C_layernorm, _D_layernorm>(
    'layernorm_forward',
  );

  geluTensor = _lib.lookupFunction<_C_op1, _D_op1>('gelu_tensor');
  embeddingForward = _lib.lookupFunction<_C_embedding, _D_embedding>(
    'embedding_forward',
  );
  crossEntropyLoss = _lib.lookupFunction<_C_loss, _D_loss>(
    'cross_entropy_loss',
  );

  tensorToHost = _lib.lookupFunction<_C_to_host, _D_to_host>(
    'tensor_to_host',
  );

  adamStep = _lib.lookupFunction<_C_adam_step, _D_adam_step>('adam_step');
  sdgStep = _lib.lookupFunction<_C_adam_step, _D_adam_step>('sdg_step');
  // Ensure you also have zeroGrad defined
  zeroGrad = _lib
      .lookupFunction<
        ffi.Void Function(ffi.Pointer<ffi.Void>),
        void Function(ffi.Pointer<ffi.Void>)
      >('zero_grad');

  clipGradients = _lib.lookupFunction<_C_clip, _D_clip>('clip_gradients');

  setTensorData = _lib.lookupFunction<_C_set_data, _D_set_data>(
    'set_tensor_data',
  );

  sliceTensor = _lib.lookupFunction<_C_slice, _D_slice>('slice_tensor');

  abs_tensor = _lib.lookupFunction<UnaryOpFn, UnaryOpDart>('abs_tensor');

  softmax_forward = _lib.lookupFunction<UnaryOpFn, UnaryOpDart>(
    'softmax_forward',
  );

  _computeCostMatrix = _lib
      .lookup<ffi.NativeFunction<NativeComputeCost>>('compute_cost_matrix')
      .asFunction<DartComputeCost>();

  sumTensor = _lib.lookupFunction<_C_reduce, _D_reduce>('sum_tensor');

  // Mean reduction: returns a 1x1 Tensor pointer
  meanTensor = _lib.lookupFunction<_C_reduce, _D_reduce>('mean_tensor');

  _tensorXavierInit = _lib.lookupFunction<_XavierInitC, _XavierInitDart>(
    'tensor_xavier_init',
  );

  _tensorZeroInit = _lib.lookupFunction<_ZeroInitC, _ZeroInitDart>(
    'tensor_zero_init',
  );

  // l2Normalize = _lib.lookupFunction<_C_l2norm, _D_l2norm>(
  //   'l2_normalize_tensor',
  // );

  layerNorm = _lib.lookupFunction<_C_l2norm, _D_l2norm>('layer_norm_tensor');
  im2col = _lib.lookupFunction<_C_im2col, _D_im2col>('im2col_cuda');
  col2im = _lib.lookupFunction<_C_im2col, _D_im2col>('col2im_cuda');
}