#include "ccv_nnc.h"

#include "ccv_nnc_easy.h"

#include "ccv_nnc_internal.h"

#include "ccv_nnc_easy.h"

#include "ccv_internal.h"

#include "_ccv_nnc_dynamic_graph.h"

// MARK - Level-4 API

ccv_nnc_dynamic_graph_t* ccv_nnc_dynamic_graph_new(void)

{

  ccv_nnc_dynamic_graph_t* graph = ccmalloc(sizeof(ccv_nnc_dynamic_graph_t));

  graph->no_grad = 0;

  graph->reuse_var = -1;

  graph->vars = ccv_array_new(sizeof(ccv_nnc_tensor_variable_t), 1, 0);

  graph->binds = ccv_array_new(sizeof(ccv_nnc_tensor_variable_graph_bind_t), 1, 0);

  graph->tape = ccv_nnc_symbolic_graph_new();

  graph->xpu_alloc.mp_hdr = -1;

  graph->xpu_alloc.freed = kh_init(dy_str);

  graph->xpu_alloc.allocd = kh_init(dy_alloc);

  // These may not be used as frequent, init as needed.

  graph->stateful_execs = 0;

  graph->reuse_stateful_exec = -1;

  graph->stream_map = 0;

  graph->ws = 0;

  return graph;

}

static void _ccv_nnc_tensor_variable_free(ccv_nnc_dynamic_graph_t* const graph, const ccv_nnc_tensor_variable_t tensor_variable, const int zeroing)

{

  const int index = tensor_variable->index;

  if (tensor_variable->tensor_view)

  {

    if (tensor_variable->destructor_hook.func)

      tensor_variable->destructor_hook.func(graph, (ccv_nnc_tensor_t*)tensor_variable->tensor_view, tensor_variable->destructor_hook.context);

    if (!CCV_NNC_IS_EXTERN_TENSOR_VIEW(tensor_variable->tensor_view))

    {

      if (CCV_IS_TENSOR_VIEW(tensor_variable->tensor_view))

        ccv_nnc_tensor_view_free(tensor_variable->tensor_view);

      else {

        if (!tensor_variable->alias_index_ref && // Return this memory to the graph.

          
CCV_TENSOR_GET_MEMORY14.2k
(tensor_variable->tensor_view->info.type) == CCV_TENSOR_GPU_MEMORY14.2k
 && 
tensor_variable->tensor_view->data.u8419
)

          ccv_nnc_xpu_free(&graph->xpu_alloc, tensor_variable->tensor_view->data.u8);

        ccv_nnc_tensor_free((ccv_nnc_tensor_t*)tensor_variable->tensor_view);

      }

    }

  }

  ccfree(tensor_variable);

  if (zeroing)

    *(ccv_nnc_tensor_variable_t*)ccv_array_get(graph->vars, index) = 0;

  int i;

  for (i = graph->vars->rnum - 1; i >= 0; 
i--32.1k
)

    if (*(ccv_nnc_tensor_variable_t*)ccv_array_get(graph->vars, i) != 0)

    {

      graph->vars->rnum = i + 1;

      break;

    }

  if (index < graph->vars->rnum &&

    
(23.3k
index < graph->reuse_var23.3k
 || 
graph->reuse_var < 017.3k
))

    graph->reuse_var = index;

  else if (graph->reuse_var >= graph->vars->rnum)

    graph->reuse_var = -1;

}

static void _ccv_nnc_tensor_variable_graph_bind_free(ccv_nnc_dynamic_graph_t* const graph, ccv_nnc_tensor_variable_graph_bind_t* const bind, const int zeroing)

{

  bind->index = CCV_NNC_TENSOR_NO_VARIABLE;

  if (bind->sources)

    ccv_array_free(bind->sources);

  if (bind->destinations)

    ccv_array_free(bind->destinations);

  if (bind->tensor_view)

  {

    if (bind->destructor_hook.func)

      bind->destructor_hook.func(graph, (ccv_nnc_tensor_t*)bind->tensor_view, bind->destructor_hook.context);

    if (!CCV_NNC_IS_EXTERN_TENSOR_VIEW(bind->tensor_view))

    {

      if (CCV_IS_TENSOR_VIEW(bind->tensor_view))

        ccv_nnc_tensor_view_free(bind->tensor_view);

      else {

        if (!bind->alias_ref && // Return this memory to the graph.

          
CCV_TENSOR_GET_MEMORY16.4k
(bind->tensor_view->info.type) == CCV_TENSOR_GPU_MEMORY16.4k
 && 
bind->tensor_view->data.u838
)

          ccv_nnc_xpu_free(&graph->xpu_alloc, bind->tensor_view->data.u8);

        ccv_nnc_tensor_free((ccv_nnc_tensor_t*)bind->tensor_view);

      }

    }

  }

  if (zeroing)

  {

    bind->sources = 0;

    bind->destinations = 0;

    bind->tensor_view = 0;

    bind->destructor_hook.func = 0;

    bind->destructor_hook.context = 0;

  }

}

void ccv_nnc_dynamic_graph_free(ccv_nnc_dynamic_graph_t* const graph)

{

  int i;

  for (i = 0; i < graph->vars->rnum; 
i++302
)

  {

    ccv_nnc_tensor_variable_t tensor_variable = *(ccv_nnc_tensor_variable_t*)ccv_array_get(graph->vars, i);

    if (tensor_variable)

      _ccv_nnc_tensor_variable_free(graph, tensor_variable, 0);

  }

  ccv_array_free(graph->vars);

  for (i = 0; i < graph->binds->rnum; 
i++361
)

    _ccv_nnc_tensor_variable_graph_bind_free(graph, (ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(graph->binds, i), 0);

  ccv_array_free(graph->binds);

  ccv_nnc_symbolic_graph_free(graph->tape);

  if (graph->ws)

    ccv_array_free(graph->ws);

  if (graph->stateful_execs)

  {

    for (i = 0; i < graph->stateful_execs->rnum; 
i++26
)

    {

      ccv_nnc_stateful_exec_t* const stateful_exec = *(ccv_nnc_stateful_exec_t**)ccv_array_get(graph->stateful_execs, i);

      if (stateful_exec)

        ccfree(stateful_exec);

    }

    ccv_array_free(graph->stateful_execs);

  }

  if (graph->stream_map)

  {

    khiter_t k;

    for (k = 
kh_begin10
(graph->stream_map); k != kh_end(graph->stream_map); 
++k48
)

    {

      if (!kh_exist(graph->stream_map, k))

        continue;

      ccv_nnc_stream_context_t* const stream = kh_val(graph->stream_map, k);

      ccv_nnc_stream_context_free(stream);

    }

    kh_destroy(stream_map, graph->stream_map);

  }

  ccv_nnc_xpu_alloc_destroy(&graph->xpu_alloc);

  ccfree(graph);

}

void ccv_nnc_tensor_variable_set(ccv_nnc_dynamic_graph_t* const graph, const ccv_nnc_tensor_variable_t tensor_variable, ccv_nnc_tensor_t* const tensor)

{

  assert(!tensor_variable->alias_index_ref);

  if (tensor_variable->tensor_view && 
!2
CCV_NNC_IS_EXTERN_TENSOR_VIEW2
(tensor_variable->tensor_view))

  {

    assert(!CCV_IS_TENSOR_VIEW(tensor_variable->tensor_view));

    ccv_nnc_tensor_free((ccv_nnc_tensor_t*)tensor_variable->tensor_view);

  }

  tensor_variable->info = tensor->info;

  tensor_variable->tensor_view = (ccv_nnc_tensor_view_t*)((uintptr_t)tensor | 1);

}

void ccv_nnc_tensor_variable_destructor_hook(ccv_nnc_dynamic_graph_t* const graph, const ccv_nnc_tensor_variable_t tensor_variable, ccv_nnc_tensor_variable_destructor_f func, void* const context)

{

  tensor_variable->destructor_hook.func = func;

  tensor_variable->destructor_hook.context = context;

}

inline static void _ccv_nnc_tensor_variable_init(ccv_nnc_dynamic_graph_t* const graph, ccv_nnc_tensor_variable_t tensor_variable, const ccv_nnc_tensor_param_t info)

{

  tensor_variable->alias_index_ref = 0;

  tensor_variable->alias_off = 0;

  tensor_variable->destructor_hook.func = 0;

  tensor_variable->destructor_hook.context = 0;

  tensor_variable->info = info;

  tensor_variable->symbol = NO_TENSOR_SYMBOL;

  tensor_variable->tensor_view = 0;

  if (graph->reuse_var >= 0)

  {

    const int reuse_var = graph->reuse_var;

    assert(reuse_var < graph->vars->rnum);

    tensor_variable->index = reuse_var;

    *(ccv_nnc_tensor_variable_t*)ccv_array_get(graph->vars, reuse_var) = tensor_variable;

    int i;

    graph->reuse_var = -1;

    for (i = reuse_var + 1; i < graph->vars->rnum && 
graph->reuse_var < 01.02k
; 
i++631
)

      if (*(ccv_nnc_tensor_variable_t*)ccv_array_get(graph->vars, i) == 0)

        graph->reuse_var = i;

  } else {

    tensor_variable->index = graph->vars->rnum;

    ccv_array_push(graph->vars, &tensor_variable);

  }

}

ccv_nnc_tensor_variable_t ccv_nnc_tensor_variable_new_impl(ccv_nnc_dynamic_graph_t* const graph, const ccv_nnc_tensor_param_t info)

{

  ccv_nnc_tensor_variable_t tensor_variable = ccmalloc(sizeof(struct ccv_nnc_tensor_variable_s));

  tensor_variable->type = CCV_NNC_TENSOR_VARIABLE;

  _ccv_nnc_tensor_variable_init(graph, tensor_variable, info);

  return tensor_variable;

}

ccv_nnc_tensor_variable_t ccv_nnc_tensor_constant_new_impl(ccv_nnc_dynamic_graph_t* const graph, const ccv_nnc_tensor_param_t info)

{

  ccv_nnc_tensor_variable_t tensor_variable = ccmalloc(sizeof(struct ccv_nnc_tensor_variable_s));

  tensor_variable->type = CCV_NNC_TENSOR_CONSTANT;

  _ccv_nnc_tensor_variable_init(graph, tensor_variable, info);

  return tensor_variable;

}

int ccv_nnc_tensor_variable_is_constant(const ccv_nnc_dynamic_graph_t* const graph, const ccv_nnc_tensor_variable_t tensor_variable)

{

  return tensor_variable->type == CCV_NNC_TENSOR_CONSTANT;

}

ccv_nnc_tensor_param_t ccv_nnc_tensor_variable_params(ccv_nnc_dynamic_graph_t* const graph, const ccv_nnc_tensor_variable_t tensor_variable)

{

  return tensor_variable->info;

}

ccv_nnc_tensor_variable_t ccv_nnc_tensor_variable_alias_new(ccv_nnc_dynamic_graph_t* const graph, const ccv_nnc_tensor_variable_t tensor_variable, const int ofs[CCV_NNC_MAX_DIM_ALLOC], const int stride[CCV_NNC_MAX_DIM_ALLOC], const ccv_nnc_tensor_param_t info)

{

  ccv_nnc_tensor_variable_t variable_alias = ccmalloc(sizeof(struct ccv_nnc_tensor_variable_s));

  variable_alias->type = tensor_variable->type;

  // If the tensor variable is an alias itself, we point directly to its original.

  if (tensor_variable->alias_index_ref)

  {

    variable_alias->alias_index_ref = tensor_variable->alias_index_ref;

    // The tensor variable need to be fully specified if I am doing alias an alias.

    assert(!ccv_nnc_is_tensor_auto(tensor_variable->info));

    int i;

    int no_stride = 1;

    for (i = 0; no_stride && 
i < 1
CCV_NNC_MAX_DIM_ALLOC1
; 
i++1
)

      no_stride = (tensor_variable->stride[i] == 0);

    int stride_from_dim[CCV_NNC_MAX_DIM_ALLOC];

    int* to_stride;

    if (no_stride)

    {

      ccv_nnc_tensor_get_stride(tensor_variable->info.dim, stride_from_dim);

      to_stride = stride_from_dim;

    } else

      to_stride = tensor_variable->stride;

    // If we provide stride, or reshape to a different size, assert the tensor variable itself is contiguous (otherwise we cannot satisfy the reshape requirements).

    const int different_dim = ccv_nnc_tensor_nd(info.dim) != ccv_nnc_tensor_nd(tensor_variable->info.dim);

    if (different_dim || 
(0
stride[0] != 00
 && 
memcmp(stride, to_stride, sizeof(int) * 0
CCV_NNC_MAX_DIM_ALLOC0
) != 0))

      { assert(ccv_nnc_tensor_view_is_contiguous(tensor_variable->info.dim, to_stride)); }

    // Need to compute alias off, that is the alias off of the tensor variable plus its ofs.

    const off_t off = ccv_nnc_tensor_view_offset(tensor_variable->info.datatype, to_stride, tensor_variable->ofs);

    variable_alias->alias_off = tensor_variable->alias_off + off;

    // If we don't provide stride, copy the stride from previous variable.

    if (stride[0] == 0)

    {

      if (different_dim)

        ccv_nnc_tensor_get_stride(info.dim, variable_alias->stride);

      else

        memcpy(variable_alias->stride, to_stride, sizeof(int) * CCV_NNC_MAX_DIM_ALLOC);

    } else

      memcpy(variable_alias->stride, stride, sizeof(int) * CCV_NNC_MAX_DIM_ALLOC);

  } else {

    variable_alias->alias_index_ref = tensor_variable->index + 1;

    variable_alias->alias_off = 0;

    memcpy(variable_alias->stride, stride, sizeof(int) * CCV_NNC_MAX_DIM_ALLOC);

  }

  variable_alias->info = info;

  variable_alias->symbol = NO_TENSOR_SYMBOL;

  variable_alias->destructor_hook.func = 0;

  variable_alias->destructor_hook.context = 0;

  variable_alias->tensor_view = 0;

  memcpy(variable_alias->ofs, ofs, sizeof(int) * CCV_NNC_MAX_DIM_ALLOC);

  if (graph->reuse_var >= 0)

  {

    const int reuse_var = graph->reuse_var;

    assert(reuse_var < graph->vars->rnum);

    variable_alias->index = reuse_var;

    *(ccv_nnc_tensor_variable_t*)ccv_array_get(graph->vars, reuse_var) = variable_alias;

    int i;

    graph->reuse_var = -1;

    for (i = reuse_var + 1; i < graph->vars->rnum && graph->reuse_var < 0; i++)

      if (*(ccv_nnc_tensor_variable_t*)ccv_array_get(graph->vars, i) == 0)

        graph->reuse_var = i;

  } else {

    variable_alias->index = graph->vars->rnum;

    ccv_array_push(graph->vars, &variable_alias);

  }

  return variable_alias;

}

int ccv_nnc_tensor_variable_alias_params(const ccv_nnc_dynamic_graph_t* const graph, const ccv_nnc_tensor_variable_t tensor_variable, int ofs[CCV_NNC_MAX_DIM_ALLOC], int stride[CCV_NNC_MAX_DIM_ALLOC])

{

  if (!tensor_variable->alias_index_ref)

    return -1;

  if (ofs)

    memcpy(ofs, tensor_variable->ofs, sizeof(int) * CCV_NNC_MAX_DIM_ALLOC);

  if (stride)

    memcpy(stride, tensor_variable->stride, sizeof(int) * CCV_NNC_MAX_DIM_ALLOC);

  return 0;

}

ccv_nnc_tensor_t* ccv_nnc_tensor_from_variable_impl(ccv_nnc_dynamic_graph_t* const graph, const ccv_nnc_tensor_variable_t tensor_variable, ccv_nnc_stream_context_t* const stream_context)

{

  if (tensor_variable->tensor_view)

  {

    if (tensor_variable->alias_index_ref)

    {

      const int alias_index = tensor_variable->alias_index_ref - 1;

      assert(alias_index >= 0);

      ccv_nnc_tensor_variable_t variable_to = *(ccv_nnc_tensor_variable_t*)ccv_array_get(graph->vars, alias_index);

      if (CCV_IS_TENSOR_VIEW(tensor_variable->tensor_view))

      {

        ccv_nnc_tensor_view_t* const tv = tensor_variable->tensor_view;

        // We cannot have an alias with custom set tensor, otherwise the pointer update is invalid.

        assert(!CCV_NNC_IS_EXTERN_TENSOR_VIEW(tv));

        // Update the tensor_view pointer every time access it, because the underlying variable it alias to have changed.

        ccv_nnc_tensor_data(CCV_NNC_TENSOR_VIEW(variable_to->tensor_view)->info, CCV_NNC_TENSOR_VIEW(variable_to->tensor_view)->data.u8, tv->off + tensor_variable->alias_off, &tv->data, &tv->dataof);

      } else {

        ccv_nnc_tensor_t* const tv = (ccv_nnc_tensor_t*)tensor_variable->tensor_view;

        // We cannot have an alias with custom set tensor, otherwise the pointer update is invalid.

        assert(!CCV_NNC_IS_EXTERN_TENSOR_VIEW(tv));

        // Update the tensor_view pointer every time access it, because the underlying variable it alias to have changed.

        ccv_nnc_tensor_data(CCV_NNC_TENSOR_VIEW(variable_to->tensor_view)->info, CCV_NNC_TENSOR_VIEW(variable_to->tensor_view)->data.u8, tensor_variable->alias_off, &tv->data, &tv->dataof);

      }

    }

    return (ccv_nnc_tensor_t*)CCV_NNC_TENSOR_VIEW(tensor_variable->tensor_view);

  }

  if (!tensor_variable->alias_index_ref)

  {

    // If we haven't allocated tensor_variable, we cannot allocate them now (because no shape specified), return 0.

    if (ccv_nnc_is_tensor_auto(tensor_variable->info))

      return 0;

    void* ptr = 0;

    const size_t data_size = ccv_nnc_tensor_data_size(tensor_variable->info);

    if (CCV_TENSOR_GET_MEMORY(tensor_variable->info.type) == CCV_TENSOR_GPU_MEMORY && 
data_size > 0457
)

      ptr = ccv_nnc_xpu_alloc(&graph->xpu_alloc, CCV_TENSOR_GET_DEVICE_ID(tensor_variable->info.type), stream_context, data_size);

    tensor_variable->tensor_view = (ccv_nnc_tensor_view_t*)ccv_nnc_tensor_new(ptr, tensor_variable->info, 0);

    if (tensor_variable->info.dim[0] > 0)

      { assert(tensor_variable->tensor_view->data.u8); }

    return (ccv_nnc_tensor_t*)tensor_variable->tensor_view;

  }

  const int alias_index = tensor_variable->alias_index_ref - 1;

  assert(alias_index >= 0);

  ccv_nnc_tensor_variable_t variable_to = *(ccv_nnc_tensor_variable_t*)ccv_array_get(graph->vars, alias_index);

  assert(!variable_to->alias_index_ref);

  if (!variable_to->tensor_view)

  {

    // If we haven't allocated variable_to, we cannot allocate them now (because no shape specified), return 0.

    if (ccv_nnc_is_tensor_auto(variable_to->info))

      return 0;

    void* ptr = 0;

    assert(variable_to->info.type == tensor_variable->info.type);

    const size_t data_size = ccv_nnc_tensor_data_size(variable_to->info);

    if (CCV_TENSOR_GET_MEMORY(variable_to->info.type) == CCV_TENSOR_GPU_MEMORY && 
data_size > 00
)

      ptr = ccv_nnc_xpu_alloc(&graph->xpu_alloc, CCV_TENSOR_GET_DEVICE_ID(variable_to->info.type), stream_context, data_size);

    variable_to->tensor_view = (ccv_nnc_tensor_view_t*)ccv_nnc_tensor_new(ptr, variable_to->info, 0);

    assert(variable_to->tensor_view->data.u8);

  }

  int i;

  int no_ofs = 1;

  for (i = 0; no_ofs && 
i < 13.5k
CCV_NNC_MAX_DIM_ALLOC13.5k
; 
i++12.4k
)

    no_ofs = (tensor_variable->ofs[i] == 0);

  int no_stride = 1;

  for (i = 0; no_stride && 
i < 1.45k
CCV_NNC_MAX_DIM_ALLOC1.45k
; 
i++1.42k
)

    no_stride = (tensor_variable->stride[i] == 0);

  int stride_is_packed = no_stride;

  if (!no_stride) // We have stride, now if it is packed.

    stride_is_packed = ccv_nnc_is_tensor_stride_packed(tensor_variable->stride, tensor_variable->info.dim);

  assert(CCV_GET_DATA_TYPE_SIZE(tensor_variable->info.datatype) * ccv_nnc_tensor_count(tensor_variable->info) + tensor_variable->alias_off <= CCV_GET_DATA_TYPE_SIZE(variable_to->info.datatype) * ccv_nnc_tensor_count(variable_to->info));

  // Allowing vector type to be normal tensor, rather than a tensor view. We cannot have any offset though.

  if (no_ofs && 
!stride_is_packed1.03k
)

    stride_is_packed = ccv_nnc_tensor_view_is_contiguous(tensor_variable->info.dim, tensor_variable->stride);

  if (no_ofs && 
stride_is_packed1.03k
)

    tensor_variable->tensor_view = (ccv_nnc_tensor_view_t*)ccv_nnc_tensor_new(CCV_NNC_TENSOR_VIEW(variable_to->tensor_view)->data.u8, tensor_variable->info, 0);

  else {

    if (no_stride)

      ccv_nnc_tensor_get_stride(tensor_variable->info.dim, tensor_variable->stride);

    tensor_variable->tensor_view = ccv_nnc_tensor_view_new((ccv_nnc_tensor_t*)CCV_NNC_TENSOR_VIEW(variable_to->tensor_view), tensor_variable->info, tensor_variable->ofs, tensor_variable->stride);

  }

  if  (tensor_variable->alias_off)

    ccv_nnc_tensor_data_add(tensor_variable->tensor_view->info, tensor_variable->alias_off, &tensor_variable->tensor_view->data, &tensor_variable->tensor_view->dataof);

  return (ccv_nnc_tensor_t*)tensor_variable->tensor_view;

}

static void _ccv_nnc_tensor_symbol_extra_new(ccv_nnc_dynamic_graph_t* const graph, const ccv_nnc_tensor_variable_t tensor_variable, const ccv_nnc_tensor_symbol_t symbol)

{

  if (symbol.d >= graph->binds->rnum)

  {

    const int rnum = graph->binds->rnum;

    ccv_array_resize(graph->binds, symbol.d + 1);

    int i;

    for (i = rnum; i < graph->binds->rnum; 
i++361
)

      ((ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(graph->binds, i))->index = CCV_NNC_TENSOR_NO_VARIABLE;

  }

  ccv_nnc_tensor_variable_graph_bind_t* const bind = (ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(graph->binds, symbol.d);

  bind->type = tensor_variable->type;

  bind->index = tensor_variable->index;

  if (tensor_variable->alias_index_ref)

  {

    const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(graph->tape, (ccv_nnc_tensor_symbol_t){

      .d = symbol.d,

      .graph = graph->tape

    });

    assert(alias_to.d >= 0 && alias_to.d < graph->binds->rnum);

    bind->alias_ref = alias_to.d + 1;

  } else

    bind->alias_ref = 0;

  if (bind->sources)

    ccv_array_free(bind->sources);

  bind->sources = 0;

  if (bind->destinations)

    ccv_array_free(bind->destinations);

  bind->destinations = 0;

  bind->destructor_hook.func = 0;

  bind->destructor_hook.context = 0;

  bind->tensor_view = 0;

}

static ccv_nnc_tensor_symbol_t _ccv_nnc_tensor_symbol_from_variable(ccv_nnc_dynamic_graph_t* const graph, const ccv_nnc_tensor_variable_t tensor_variable)

{

  if (tensor_variable->symbol.d >= 0)

    return tensor_variable->symbol;

  if (!tensor_variable->alias_index_ref)

  {

    const ccv_nnc_tensor_symbol_t symbol = tensor_variable->symbol = ccv_nnc_tensor_symbol_new(graph->tape, tensor_variable->info, 0);

    _ccv_nnc_tensor_symbol_extra_new(graph, tensor_variable, symbol);

    return symbol;

  }

  const int alias_index = tensor_variable->alias_index_ref - 1;

  assert(alias_index >= 0);

  ccv_nnc_tensor_variable_t variable_to = *(ccv_nnc_tensor_variable_t*)ccv_array_get(graph->vars, alias_index);

  assert(!variable_to->alias_index_ref);

  int no_stride = 1;

  int i;

  for (i = 0; no_stride && 
i < 1.42k
CCV_NNC_MAX_DIM_ALLOC1.42k
; 
i++1.39k
)

    no_stride = (tensor_variable->stride[i] == 0);

  if (no_stride)

    ccv_nnc_tensor_get_stride(tensor_variable->info.dim, tensor_variable->stride);

  const ccv_nnc_tensor_symbol_t symbol = tensor_variable->symbol = ccv_nnc_tensor_symbol_alias_new(graph->tape, _ccv_nnc_tensor_symbol_from_variable(graph, variable_to), tensor_variable->ofs, tensor_variable->stride, tensor_variable->info, 0);

  _ccv_nnc_tensor_symbol_extra_new(graph, tensor_variable, symbol);

  return symbol;

}

// Return the tensor variable that is old (the provided tensor variable will have a new setting).

ccv_nnc_tensor_variable_t ccv_nnc_tensor_variable_exchange_new(ccv_nnc_dynamic_graph_t* const graph, ccv_nnc_tensor_variable_t tensor_variable)

{

  struct ccv_nnc_tensor_variable_s x = *tensor_variable;

  ccv_nnc_tensor_variable_t new_variable;

  // Need to handle alias.

  if (x.alias_index_ref)

    new_variable = ccv_nnc_tensor_variable_alias_new(graph, *(ccv_nnc_tensor_variable_t*)ccv_array_get(graph->vars, x.alias_index_ref - 1), x.ofs, x.stride, x.info);

  else

    new_variable = ccv_nnc_tensor_variable_new(graph, x.info);

  *tensor_variable = *new_variable;

  *new_variable = x;

  // The index should be the same though.

  const int index = new_variable->index;

  new_variable->index = tensor_variable->index;

  if (new_variable->symbol.d != CCV_NNC_NO_TENSOR_SYMBOL)

  {

    ccv_nnc_tensor_variable_graph_bind_t* const bind = (ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(graph->binds, new_variable->symbol.d);

    bind->index = new_variable->index;

  }

  tensor_variable->index = index;

  return new_variable;

}

void ccv_nnc_dynamic_graph_set_max_concurrency(ccv_nnc_dynamic_graph_t* const dynamic_graph, const int max_stream_count)

{

  dynamic_graph->max_stream_count = max_stream_count;

}

int ccv_nnc_dynamic_graph_set_no_grad(ccv_nnc_dynamic_graph_t* const dynamic_graph, const int no_grad)

{

  if (dynamic_graph->no_grad == no_grad)

    return -1;

  dynamic_graph->no_grad = no_grad;

  return 0;

}

static ccv_nnc_stream_context_t* _ccv_nnc_dynamic_graph_get_stream(ccv_nnc_dynamic_graph_t* const graph, const int type)

{

  if (!graph->stream_map)

    graph->stream_map = kh_init(stream_map);

  int ret = 0;

  khiter_t k = kh_put(stream_map, graph->stream_map, type, &ret);

  assert(ret >= 0);

  ccv_nnc_stream_context_t* stream = kh_val(graph->stream_map, k);

  // If ret == 0, the key already exist, we can return directly, otherwise, create and return.

  if (ret != 0)

  {

    stream = ccv_nnc_stream_context_new(type);

    kh_val(graph->stream_map, k) = stream;

  }

  return stream;

}

typedef struct {

  ccv_nnc_dynamic_graph_t* graph;

  int stream_type;

} ccv_nnc_dynamic_graph_neighbor_context_discovery_t;

static ccv_nnc_stream_context_t* _ccv_nnc_dynamic_graph_neighbor_context_discovery(const int device_id, void* const context)

{

  ccv_nnc_dynamic_graph_neighbor_context_discovery_t* const discovery = (ccv_nnc_dynamic_graph_neighbor_context_discovery_t*)context;

  int type = discovery->stream_type;

  CCV_STREAM_SET_DEVICE_ID(type, device_id);

  return _ccv_nnc_dynamic_graph_get_stream(discovery->graph, type);

}

void ccv_nnc_dynamic_graph_exec_ret(ccv_nnc_dynamic_graph_t* const graph, const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, const ccv_nnc_tensor_variable_t* const inputs, const int input_size, ccv_nnc_tensor_variable_t* const outputs, const int output_size, const int parallel, ccv_nnc_stream_context_t* const stream_context, ccv_nnc_graph_exec_symbol_t* const graph_execs)

{

  int i, j;

  for (i = 0; i < input_size; 
i++28.1k
)

    if (inputs[i] && 
!inputs[i]->alias_index_ref28.1k
)

      { assert(inputs[i]->tensor_view); }

  ccv_nnc_tensor_t* input_tensors[ccv_max(1, input_size)];

  for (i = 0; i < input_size; 
i++28.1k
)

    input_tensors[i] = inputs[i] ? ccv_nnc_tensor_from_variable(graph, inputs[i], stream_context) : 
02
;

  ccv_nnc_tensor_symbol_t input_symbols[ccv_max(1, input_size)];

  for (i = 0; i < input_size; 
i++28.1k
)

    input_symbols[i] = inputs[i] ? 
_ccv_nnc_tensor_symbol_from_variable(graph, inputs[i])28.1k
 : 
NO_TENSOR_SYMBOL2
;

  ccv_array_t* input_sources[ccv_max(1, input_size)];

  ccv_array_t* input_alias_sources[ccv_max(1, input_size)];

  for (i = 0; i < input_size; 
i++28.1k
)

  {

    input_sources[i] = input_symbols[i].d != CCV_NNC_NO_TENSOR_SYMBOL ? 
((ccv_nnc_tensor_variable_graph_bind_t*)28.1k
ccv_array_get28.1k
(graph->binds, input_symbols[i].d))->sources : 
02
;

    if (inputs[i] && 
inputs[i]->alias_index_ref28.1k
)

    {

      const int alias_index_ref = inputs[i]->alias_index_ref - 1;

      assert(alias_index_ref >= 0);

      ccv_nnc_tensor_variable_t variable_to = *(ccv_nnc_tensor_variable_t*)ccv_array_get(graph->vars, alias_index_ref);

      input_alias_sources[i] = ((ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(graph->binds, variable_to->symbol.d))->sources;

    } else

      input_alias_sources[i] = 0;

  }

  const int parallel_count = ccv_max(1, parallel);

  assert(input_size % parallel_count == 0);

  const int per_input_size = input_size / parallel_count;

  assert(output_size % parallel_count == 0);

  const int per_output_size = output_size / parallel_count;

  int output_auto = 0;

  for (i = 0; !output_auto && 
i < output_size16.3k
; 
i++15.6k
)

    output_auto = outputs[i] ? 
ccv_nnc_is_tensor_auto(outputs[i]->info)15.4k
 : 
0200
;

  // One extra step, infer the parameters for outputs.

  if (output_auto)

  {

    ccv_nnc_tensor_param_t input_params[ccv_max(1, per_input_size)];

    ccv_nnc_tensor_param_t output_params[ccv_max(1, per_output_size)];

    for (i = 0; i < parallel_count; 
i++14.7k
)

    {

      for (j = 0; j < per_input_size; 
j++27.2k
)

        input_params[j] = inputs[j + i * per_input_size] ? 
inputs[j + i * per_input_size]->info27.2k
 : 
ccv_nnc_tensor_auto2
;

      for (j = 0; j < per_output_size; 
j++14.9k
)

        output_params[j] = outputs[j + i * per_output_size] ? 
outputs[j + i * per_output_size]->info14.7k
 : 
ccv_nnc_tensor_auto207
;

      ccv_nnc_hint_tensor_auto(cmd, input_params, per_input_size, hint, output_params, per_output_size);

      for (j = 0; j < per_output_size; 
j++14.9k
)

        if (outputs[j + i * per_output_size])

          outputs[j + i * per_output_size]->info = output_params[j];

    }

  }

  int freeable_size = 0;

  ccv_nnc_tensor_variable_t freeables[ccv_max(1, output_size)];

  // Refresh the symbol if it is binded to an existing exec. Otherwise we cannot keep the SSA guarantee.

  for (i = 0; i < output_size; 
i++15.6k
)

  {

    // First, go over to see whether there is enforce inplace.

    int enforce_idx = -1;

    for (j = 0; enforce_idx < 0 && 
j < input_size44.7k
; 
j++29.1k
)

      if (inputs[j] && 
ccv_nnc_cmd_enforce_inplace(cmd, j, input_size, i, output_size)29.1k
)

        enforce_idx = j;

    if (enforce_idx >= 0)

      { assert(outputs[i] == inputs[enforce_idx] && outputs[i]->symbol.d != CCV_NNC_NO_TENSOR_SYMBOL); }

    // We don't allow or check "allow inplace" yet. That logic will be at odds with backward logic.

    if (outputs[i] && 
outputs[i]->symbol.d != CCV_NNC_NO_TENSOR_SYMBOL15.4k
)

    {

      const ccv_nnc_tensor_variable_graph_bind_t* const bind = (ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(graph->binds, outputs[i]->symbol.d);

      if (enforce_idx >= 0)

        { assert(!bind->destinations || bind->destinations->rnum == 0); }

      if (bind->sources && 
bind->sources->rnum > 0415
)

      {

        const ccv_nnc_tensor_variable_t old_var = freeables[freeable_size++] = ccv_nnc_tensor_variable_exchange_new(graph, outputs[i]);

        // If this is enforce output, make sure the tensor view is taken by the output.

        if (enforce_idx >= 0)

        {

          outputs[i]->destructor_hook = old_var->destructor_hook;

          outputs[i]->tensor_view = old_var->tensor_view; // Make sure the tensor view is taken over by the output.

          old_var->tensor_view = 0;

        }

      }

    }

  }

  ccv_nnc_tensor_t* output_tensors[ccv_max(1, per_output_size)];

  if (parallel_count > 1)

  {

    const int max_device_id_size = per_input_size + per_output_size;

    assert(max_device_id_size > 0);

    int device_ids[max_device_id_size];

    ccv_nnc_stream_context_t* streams[parallel_count];

    ccv_nnc_stream_signal_t* signal;

    if (stream_context)

      signal = ccv_nnc_stream_context_emit_signal_new(stream_context);

    for (i = 0; i < parallel_count; 
i++74
)

    {

      int flag = 0;

      for (j = 0; !flag && 
j < per_input_size78
; 
j++74
)

        if (input_tensors[i * per_input_size + j])

          flag = (CCV_TENSOR_GET_MEMORY(input_tensors[i * per_input_size + j]->info.type) == CCV_TENSOR_GPU_MEMORY);

      for (j = 0; j < per_output_size; 
j++82
)

      {

        output_tensors[j] = outputs[j + i * per_output_size] ? ccv_nnc_tensor_from_variable(graph, outputs[j + i * per_output_size], stream_context) : 
08
;

        if (output_tensors[j] && 
!flag74
)

          flag = (CCV_TENSOR_GET_MEMORY(output_tensors[j]->info.type) == CCV_TENSOR_GPU_MEMORY);

      }

      const int stream_type = flag ? 
CCV_STREAM_CONTEXT_GPU70
 : 
CCV_STREAM_CONTEXT_CPU4
;

      const int tensor_type = flag ? 
CCV_TENSOR_GPU_MEMORY70
 : 
CCV_TENSOR_CPU_MEMORY4
;

      const int device_id_size = ccv_nnc_device_ids_for_io(input_tensors + i * per_input_size, per_input_size, output_tensors, per_output_size, tensor_type, device_ids, max_device_id_size);

      ccv_nnc_stream_context_t* stream_0 = 0;

      for (j = 0; j < device_id_size; 
j++72
)

      {

        int type = stream_type;

        CCV_STREAM_SET_DEVICE_ID(type, device_ids[j]);

        ccv_nnc_stream_context_t* const stream = _ccv_nnc_dynamic_graph_get_stream(graph, type);

        if (!stream_0)

          stream_0 = stream;

      }

      // Wait signal to finish.

      if (stream_context)

      {

        if (stream_0)

          ccv_nnc_stream_context_wait_signal(stream_0, signal);

        else

          ccv_nnc_stream_context_wait(stream_context);

      }

      if (stream_0)

      {

        ccv_nnc_dynamic_graph_neighbor_context_discovery_t discovery = {

          .graph = graph,

          .stream_type = stream_type

        };

        ccv_nnc_stream_context_set_neighbor_discovery(stream_0, _ccv_nnc_dynamic_graph_neighbor_context_discovery, &discovery);

      }

      PRINT(CCV_CLI_INFO, "%s: [%d] -> [%d]\n", ccv_nnc_cmd_name(cmd.cmd), per_input_size, per_output_size);

      int k;

      for (k = 0; k < per_input_size; 
k++130
)

      {

        PRINT(CCV_CLI_INFO, "|-> %d. %p (%p:%d)", 
k + 1, input_tensors[k + i * per_input_size], (input_tensors[k + i * per_input_size] ? input_tensors[k + i * per_input_size]->data.u8 : 0), (input_tensors[k + i * per_input_size] ? 0
CCV_TENSOR_GET_DEVICE_ID0
(input_tensors[k + i * per_input_size]->info.type) : -1));

        if (input_tensors[k + i * per_input_size] && CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_INFO))

          ccv_nnc_print_tensor_info(input_tensors[k + i * per_input_size]);

        PRINT(CCV_CLI_INFO, "\n");

      }

      for (k = 0; k < per_output_size; 
k++82
)

      {

        PRINT(CCV_CLI_INFO, "|<- %d. %p (%p:%d)", 
k + 1, output_tensors[k], (output_tensors[k] ? output_tensors[k]->data.u8 : 0), (output_tensors[k] ? 0
CCV_TENSOR_GET_DEVICE_ID0
(output_tensors[k]->info.type) : -1));

        if (output_tensors[k] && 
CCV_CLI_OUTPUT_LEVEL_IS74
(CCV_CLI_INFO))

          ccv_nnc_print_tensor_shape(output_tensors[k]);

        PRINT(CCV_CLI_INFO, "\n");

      }

      const int status = ccv_nnc_cmd_exec(cmd, hint, flags, input_tensors + i * per_input_size, per_input_size, output_tensors, per_output_size, stream_0);

      if (status != 0)

        PRINT(CCV_CLI_INFO, "Invalid Status: %d\n", status);

      if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE))

      {

        for (k = 0; k < per_output_size; k++)

        {

          PRINT(CCV_CLI_VERBOSE, "POST: |<- %d. %p (%p:%d)", k + 1, output_tensors[k], (output_tensors[k] ? output_tensors[k]->data.u8 : 0), (output_tensors[k] ? CCV_TENSOR_GET_DEVICE_ID(output_tensors[k]->info.type) : -1));

          if (output_tensors[k])

            ccv_nnc_print_tensor_info(output_tensors[k]);

          PRINT(CCV_CLI_VERBOSE, "\n");

        }

      }

      if (stream_context && 
stream_044
)

      {

        ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);

        ccv_nnc_stream_context_wait_signal(stream_context, signal);

      }

      streams[i] = stream_0;

    }

    if (!stream_context)

      
for (i = 0; 9
i < parallel_count; 
i++30
)

        if (streams[i])

          ccv_nnc_stream_context_wait(streams[i]);

  } else {

    for (i = 0; i < per_output_size; 
i++15.5k
)

      output_tensors[i] = outputs[i] ? ccv_nnc_tensor_from_variable(graph, outputs[i], stream_context) : 
0199
;

    PRINT(CCV_CLI_INFO, "%s: [%d] -> [%d]\n", ccv_nnc_cmd_name(cmd.cmd), per_input_size, per_output_size);

    for (i = 0; i < per_input_size; 
i++28.0k
)

    {

      PRINT(CCV_CLI_INFO, "|-> %d. %p (%p:%d)", 
i + 1, input_tensors[i], (input_tensors[i] ? input_tensors[i]->data.u8 : 0), (input_tensors[i] ? 0
CCV_TENSOR_GET_DEVICE_ID0
(input_tensors[i]->info.type) : -1));

      if (input_tensors[i] && 
CCV_CLI_OUTPUT_LEVEL_IS28.0k
(CCV_CLI_INFO))

        ccv_nnc_print_tensor_info(input_tensors[i]);

      PRINT(CCV_CLI_INFO, "\n");

    }

    ccv_nnc_cmd_exec(cmd, hint, flags, input_tensors, per_input_size, output_tensors, per_output_size, stream_context);

    for (i = 0; i < per_output_size; 
i++15.5k
)

    {

      PRINT(CCV_CLI_INFO, "|<- %d. %p (%p:%d)", 
i + 1, output_tensors[i], (output_tensors[i] ? output_tensors[i]->data.u8 : 0), (output_tensors[i] ? 0
CCV_TENSOR_GET_DEVICE_ID0
(output_tensors[i]->info.type) : -1));

      if (output_tensors[i] && 
CCV_CLI_OUTPUT_LEVEL_IS15.3k
(CCV_CLI_INFO))

        ccv_nnc_print_tensor_info(output_tensors[i]);

      PRINT(CCV_CLI_INFO, "\n");

    }

  }

  int inputs_are_constants = 1;

  for (i = 0; inputs_are_constants && 
i < input_size15.3k
; 
i++15.3k
)

    if (inputs[i] && 
inputs[i]->type != CCV_NNC_TENSOR_CONSTANT15.3k
)

      inputs_are_constants = 0;

  if (input_size > 0 && 
!inputs_are_constants15.3k
 && 
!graph->no_grad15.3k
) // No need to record the execution if there is no input or we disabled gradient computation.

  {

    ccv_nnc_tensor_symbol_t output_symbols[ccv_max(1, output_size)];

    for (i = 0; i < output_size; 
i++15.5k
)

      if (outputs[i])

      {

        assert(outputs[i]->type != CCV_NNC_TENSOR_CONSTANT);

        output_symbols[i] = _ccv_nnc_tensor_symbol_from_variable(graph, outputs[i]);

      } else

        output_symbols[i] = NO_TENSOR_SYMBOL;

    int t;

    for (t = 0; t < parallel_count; 
t++15.3k
)

    {

      ccv_nnc_graph_exec_symbol_t graph_exec = ccv_nnc_graph_exec_symbol_new(graph->tape, cmd, input_symbols + t * per_input_size, per_input_size, output_symbols + t * per_output_size, per_output_size, 0);

      if (graph_execs)

        graph_execs[t] = graph_exec;

      // This needs to be done before we set the new sources on the outputs.

      for (i = 0; i < per_input_size; 
i++28.0k
)

      {

        ccv_array_t* const input_source = input_sources[i + t * per_input_size];

        if (input_source)

          
for (j = 0; 14.0k
j < input_source->rnum; 
j++14.0k
)

            ccv_nnc_graph_exec_symbol_concat(graph->tape, (ccv_nnc_graph_exec_symbol_t){

              .d = *(int*)ccv_array_get(input_source, j),

              .graph = graph->tape

            }, graph_exec);

        ccv_array_t* const input_alias_source = input_alias_sources[i + t * per_input_size];

        if (input_alias_source)

          
for (j = 0; 1.01k
j < input_alias_source->rnum; 
j++1.01k
)

            ccv_nnc_graph_exec_symbol_concat(graph->tape, (ccv_nnc_graph_exec_symbol_t){

              .d = *(int*)ccv_array_get(input_alias_source, j),

              .graph = graph->tape

            }, graph_exec);

      }

      for (i = 0; i < per_input_size; 
i++28.0k
)

      {

        ccv_nnc_tensor_variable_t const input = inputs[i + t * per_input_size];

        if (!input || 
input_symbols[i + t * per_input_size].d == CCV_NNC_NO_TENSOR_SYMBOL28.0k
)

          continue;

        // Constant inputs still need lifetime tracking while this exec is alive because

        // backward may read their concrete tensor buffers even though they do not require

        // gradients themselves.

        ccv_nnc_tensor_variable_graph_bind_t* const bind = (ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(graph->binds, input_symbols[i + t * per_input_size].d);

        if (!bind->destinations)

          bind->destinations = ccv_array_new(sizeof(int), 1, 0);

        ccv_array_add_unique_int(bind->destinations, graph_exec.d);

        if (input->alias_index_ref)

        {

            const int alias_index = input->alias_index_ref - 1;

            assert(alias_index >= 0);

            ccv_nnc_tensor_variable_t variable_to = *(ccv_nnc_tensor_variable_t*)ccv_array_get(graph->vars, alias_index);

            ccv_nnc_tensor_variable_graph_bind_t* const root_bind = (ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(graph->binds, variable_to->symbol.d);

            if (!root_bind->destinations)

              root_bind->destinations = ccv_array_new(sizeof(int), 1, 0);

            ccv_array_add_unique_int(root_bind->destinations, graph_exec.d);

        }

      }

      
for (i = 0; 15.3k
i < per_output_size; 
i++15.5k
)

      {

        ccv_nnc_tensor_variable_t const output = outputs[i + t * per_output_size];

        if (!output)

          continue;

        ccv_nnc_tensor_variable_graph_bind_t* const bind = (ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(graph->binds, output_symbols[i + t * per_output_size].d);

        assert(!bind->sources); // This is a new symbol, therefore, no binded sources associated yet.

        bind->sources = ccv_array_new(sizeof(int), 1, 0);

        ccv_array_add_unique_int(bind->sources, graph_exec.d);

        if (output->alias_index_ref)

        {

          const int alias_index = output->alias_index_ref - 1;

          assert(alias_index >= 0);

          ccv_nnc_tensor_variable_t variable_to = *(ccv_nnc_tensor_variable_t*)ccv_array_get(graph->vars, alias_index);

          ccv_nnc_tensor_variable_graph_bind_t* const root_bind = (ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(graph->binds, variable_to->symbol.d);

          if (!root_bind->sources)

            root_bind->sources = ccv_array_new(sizeof(int), 1, 0);

          ccv_array_add_unique_int(root_bind->sources, graph_exec.d);

        }

      }

    }

  }

  // Now, able to free some of the reused outputs.

  
for (i = 0; 15.3k
i < freeable_size; 
i++415
)

    ccv_nnc_tensor_variable_free(graph, freeables[i]);

}

int ccv_nnc_dynamic_graph_exec(ccv_nnc_dynamic_graph_t* const graph, const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, const ccv_nnc_tensor_variable_t* const inputs, const int input_size, ccv_nnc_tensor_variable_t* const outputs, const int output_size, const int parallel, ccv_nnc_stream_context_t* const stream_context)

{

  ccv_nnc_dynamic_graph_exec_ret(graph, cmd, hint, flags, inputs, input_size, outputs, output_size, parallel, stream_context, 0);

  return CCV_NNC_EXEC_SUCCESS;

}

static int _ccv_nnc_tensor_variable_is_only_output(ccv_nnc_dynamic_graph_t* const graph, const ccv_nnc_tensor_variable_graph_bind_t* bind, const int symbol_d)

{

  if (bind->alias_ref)

    bind = (ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(graph->binds, bind->alias_ref - 1);

  if (!bind->sources || bind->sources->rnum == 0)

    return 1;

  int i;

  for (i = 0; i < bind->sources->rnum; 
i++16.4k
)

  {

    const int exec_symbol_d = *(int*)ccv_array_get(bind->sources, i);

    const ccv_nnc_graph_exec_symbol_t exec_symbol = {

      .d = exec_symbol_d,

      .graph = graph->tape

    };

    const int* outputs; int output_size;

    ccv_nnc_graph_exec_symbol_io(graph->tape, exec_symbol, 0, 0, &outputs, &output_size);

    int j;

    for (j = 0; j < output_size; 
j++16.5k
)

      if (outputs[j] >= 0 && 
outputs[j] != symbol_d17.5k
) // If output is me, it is the only output.

      {

        assert(outputs[j] < graph->binds->rnum);

        const ccv_nnc_tensor_variable_graph_bind_t* other_bind = (ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(graph->binds, outputs[j]);

        // This is in use and is it not a constant symbol.

        if (other_bind->index >= 0 && 
other_bind->type != CCV_NNC_TENSOR_CONSTANT25
)

          return 0;

        if (other_bind->alias_ref) // If this is alias, use its original's destinations.

          other_bind = (ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(graph->binds, other_bind->alias_ref - 1);

        // The original is in use and is it not a constant symbol.

        if (other_bind->index >= 0 && 
other_bind->type != CCV_NNC_TENSOR_CONSTANT1
)

          return 0;

        if (other_bind->destinations && other_bind->destinations->rnum > 0)

          return 0;

      }

  }

  return 1;

}

static void _ccv_nnc_update_bind_destinations_when_free(ccv_nnc_dynamic_graph_t* const graph, const int freed_exec_symbol_d, ccv_array_t* const binds, ccv_nnc_tensor_variable_graph_bind_t* const bind, const int tensor_index, ccv_array_t* const ws)

{

  int i;

  if (bind->destinations)

  {

    int flag = 0;

    for (i = 0; !flag && 
i < bind->destinations->rnum25.0k
; 
i++24.8k
)

    {

      const int exec_symbol_d = *(int*)ccv_array_get(bind->destinations, i);

      if (exec_symbol_d == freed_exec_symbol_d)

      {

        if (i < bind->destinations->rnum - 1)

          *(int*)ccv_array_get(bind->destinations, i) = *(int*)ccv_array_get(bind->destinations, bind->destinations->rnum - 1);

        --bind->destinations->rnum;

        flag = 1;

      }

    }

    // This symbol can be freed.

    if (flag && 
bind->index == CCV_NNC_TENSOR_NO_VARIABLE_BUT_USED24.8k
)

    {

      ccv_nnc_tensor_variable_graph_bind_t* root_bind = bind;

      if (bind->alias_ref)

      {

        root_bind = (ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(binds, bind->alias_ref - 1);

        if (root_bind->index == CCV_NNC_TENSOR_NO_VARIABLE)

          root_bind = bind;

      }

      // If the alias_ref is not freed, we cannot free this, unless it is very clear there is no reference to this any more.

      // It is possible because exec will be freed already, thus, it is safe to remove this alias out.

      if (root_bind->index == CCV_NNC_TENSOR_NO_VARIABLE_BUT_USED &&

        
(17.6k
(17.6k
!root_bind->sources17.6k
 || 
root_bind->sources->rnum == 08.64k
) || 
_ccv_nnc_tensor_variable_is_only_output(graph, bind, tensor_index)6.01k
) &&

        
root_bind->destinations->rnum == 017.6k
)

      {

        if (root_bind->sources)

          
for (i = 0; 8.63k
i < root_bind->sources->rnum; 
i++6.00k
)

            ccv_array_add_unique_int(ws, *(int*)ccv_array_get(root_bind->sources, i));

        _ccv_nnc_tensor_variable_graph_bind_free(graph, bind, 1);

        ccv_nnc_tensor_symbol_free(graph->tape, (ccv_nnc_tensor_symbol_t){

          .d = tensor_index,

          .graph = graph->tape

        });

      } else 
if (8
bind->index == CCV_NNC_TENSOR_NO_VARIABLE_BUT_USED8
 && // Handle the case the bind is already freed, and it doesn't have any sources or destinations.

        bind->alias_ref && 
(2
!bind->sources2
 || 
bind->sources->rnum == 00
) && 
(2
!bind->destinations2
 || 
bind->destinations->rnum == 02
)) {

        _ccv_nnc_tensor_variable_graph_bind_free(graph, bind, 1);

        ccv_nnc_tensor_symbol_free(graph->tape, (ccv_nnc_tensor_symbol_t){

          .d = tensor_index,

          .graph = graph->tape

        });

      }

    }

  }

}

static void _ccv_nnc_update_bind_sources_when_free(ccv_nnc_dynamic_graph_t* const graph, const int freed_exec_symbol_d, ccv_array_t* const binds, ccv_nnc_tensor_variable_graph_bind_t* const bind, const int tensor_index, ccv_array_t* const ws)

{

  int i;

  if (bind->sources)

  {

    int flag = 0;

    for (i = 0; !flag && 
i < bind->sources->rnum7.29k
; 
i++7.29k
)

    {

      const int exec_symbol_d = *(int*)ccv_array_get(bind->sources, i);

      if (exec_symbol_d == freed_exec_symbol_d)

      {

        if (i < bind->sources->rnum - 1)

          *(int*)ccv_array_get(bind->sources, i) = *(int*)ccv_array_get(bind->sources, bind->sources->rnum - 1);

        --bind->sources->rnum;

        flag = 1;

      }

    }

    if (flag && !bind->alias_ref && 
bind->index >= 07.28k
 && 
bind->type == CCV_NNC_TENSOR_CONSTANT4.84k
 && // If it is detached (constant but previously has sources). Now can check again.

      
(3
bind->sources->rnum == 03
 || 
_ccv_nnc_tensor_variable_is_only_output(graph, bind, tensor_index)0
) &&

      
(3
!bind->destinations3
 || 
bind->destinations->rnum == 03
))

    {

      // If this is constant, set it to be no symbol again.

      ccv_nnc_tensor_variable_t tv = *(ccv_nnc_tensor_variable_t*)ccv_array_get(graph->vars, bind->index);

      tv->symbol = NO_TENSOR_SYMBOL;

      _ccv_nnc_tensor_variable_graph_bind_free(graph, bind, 1);

      ccv_nnc_tensor_symbol_free(graph->tape, (ccv_nnc_tensor_symbol_t){

        .d = tensor_index,

        .graph = graph->tape

      });

    } else if (flag && bind->index == CCV_NNC_TENSOR_NO_VARIABLE_BUT_USED) {

      // This symbol can be freed.

      ccv_nnc_tensor_variable_graph_bind_t* root_bind = bind;

      if (bind->alias_ref)

      {

        root_bind = (ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(binds, bind->alias_ref - 1);

        if (root_bind->index == CCV_NNC_TENSOR_NO_VARIABLE)

          root_bind = bind;

      }

      // If the alias_ref is not freed, we cannot free this, unless it is very clear there is no reference to this any more.

      // It is possible because exec will be freed already, thus, it is safe to remove this alias out.

      if (root_bind->index == CCV_NNC_TENSOR_NO_VARIABLE_BUT_USED &&

        
(2.43k
root_bind->sources->rnum == 02.43k
 || 
_ccv_nnc_tensor_variable_is_only_output(graph, bind, tensor_index)0
) &&

        
(2.43k
!root_bind->destinations2.43k
 || 
root_bind->destinations->rnum == 02.43k
))

      {

        for (i = 0; i < root_bind->sources->rnum; 
i++0
)

          ccv_array_add_unique_int(ws, *(int*)ccv_array_get(root_bind->sources, i));

        _ccv_nnc_tensor_variable_graph_bind_free(graph, bind, 1);

        ccv_nnc_tensor_symbol_free(graph->tape, (ccv_nnc_tensor_symbol_t){

          .d = tensor_index,

          .graph = graph->tape

        });

      } else if (bind->index == CCV_NNC_TENSOR_NO_VARIABLE_BUT_USED && // Handle the case the bind is already freed, and it doesn't have any sources or destinations.

        bind->alias_ref && 
(3
!bind->sources3
 || 
bind->sources->rnum == 03
) && 
(3
!bind->destinations3
 || 
bind->destinations->rnum == 00
)) {

        _ccv_nnc_tensor_variable_graph_bind_free(graph, bind, 1);

        ccv_nnc_tensor_symbol_free(graph->tape, (ccv_nnc_tensor_symbol_t){

          .d = tensor_index,

          .graph = graph->tape

        });

      }

    }

  }

}

static void _ccv_nnc_update_bind_sources_destinations_when_free(ccv_nnc_dynamic_graph_t* const graph, const int freed_exec_symbol_d, ccv_array_t* const binds, const int* const inputs, const int input_size, const int* const outputs, const int output_size, ccv_array_t* const ws)

{

  int i;

  for (i = 0; i < input_size; 
i++27.9k
)

    if (inputs[i] >= 0 && inputs[i] < binds->rnum)

    {

      ccv_nnc_tensor_variable_graph_bind_t* const bind = (ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(binds, inputs[i]);

      if (bind->index == CCV_NNC_TENSOR_NO_VARIABLE)

        continue;

      if (bind->alias_ref)

      {

        const int alias_to = bind->alias_ref - 1;

        ccv_nnc_tensor_variable_graph_bind_t* const root_bind = (ccv_nnc_tensor_variable_graph_bind_t*)ccv_array_get(binds, alias_to);