#include "ccv.h"

#include "ccv_internal.h"

#include "nnc/ccv_nnc.h"

#include "nnc/ccv_nnc_easy.h"

#include "nnc/ccv_nnc_internal.h"

#ifdef USE_OPENMP

#include <omp.h>

#endif

#ifdef USE_DISPATCH

#include <dispatch/dispatch.h>

#endif

#include "../_ccv_nnc_cpu_ref.h"

void _ccv_nnc_ewsum_forw_cpu_ref_f32(ccv_nnc_tensor_view_t* const* const inputs, const int input_size, ccv_nnc_tensor_view_t* const* const outputs, const int output_size)

{

  if (input_size == 1 && 
output_size == 10
)

  {

    _ccv_nnc_tensor_transfer_cpu_ref_f32(inputs[0], outputs[0]);

    return;

  }

  // Assuming this is float 32.

  int dim[CCV_NNC_MAX_DIM_ALLOC];

  int astride[CCV_NNC_MAX_DIM_ALLOC];

  int bstride[CCV_NNC_MAX_DIM_ALLOC];

  int cstride[CCV_NNC_MAX_DIM_ALLOC];

  int x, z;

  int k = 0;

  // Bad, I promised this can be inplace operation. Need to first find out if there are share the same pointer first.

  for (z = 1; z < input_size; 
z++36.0k
)

  {

    ccv_nnc_tensor_view_t* c = outputs[0];

    ccv_nnc_tensor_view_t* a = inputs[z];

    if (c->data.f32 == a->data.f32)

    {

      k = z;

      break;

    }

  }

  for (z = 0; z < input_size - 1; 
z++36.0k
)

  {

    ccv_nnc_tensor_view_t* c = outputs[0];

    ccv_nnc_tensor_view_t* a = z > 0 ? 
c28
 : 
inputs[k]36.0k
;

    ccv_nnc_tensor_view_t* b = z >= k ? 
inputs[z + 1]36.0k
 : 
inputs[z]10
;

    assert(ccv_nnc_tensor_nd(a->info.dim) <= CCV_NNC_MAX_DIM + 2);

    assert(ccv_nnc_tensor_nd(b->info.dim) <= CCV_NNC_MAX_DIM + 2);

    assert(ccv_nnc_tensor_nd(c->info.dim) <= CCV_NNC_MAX_DIM + 2);

    ccv_nnc_tensor_view_get_dim(a, dim);

    assert(ccv_nnc_tensor_view_check_dim(b, dim));

    assert(ccv_nnc_tensor_view_check_dim(c, dim));

    if (!CCV_IS_TENSOR_VIEW(a) && 
!36.0k
CCV_IS_TENSOR_VIEW36.0k
(b) && 
!36.0k
CCV_IS_TENSOR_VIEW36.0k
(c))

    {

      // Super optimal case, just do one for-loop for sum.

      const int tensor_count = ccv_nnc_tensor_count(a->info);

      for (x = 0; x < tensor_count; 
x++15.4M
)

        c->data.f32[x] = a->data.f32[x] + b->data.f32[x];

      continue;

    }

    assert
(CCV_NNC_MAX_DIM == 2)3
; // Need to change this logic for CCV_NNC_MAX_DIM == other number.

    ccv_nnc_tensor_view_get_stride(a, astride);

    ccv_nnc_tensor_view_get_stride(b, bstride);

    ccv_nnc_tensor_view_get_stride(c, cstride);

    int i[CCV_NNC_MAX_DIM + 2];

    float* const ap = a->data.f32;

    float* const bp = b->data.f32;

    float* const cp = c->data.f32;

    const int count = dim[2] * dim[3];

    if (astride[2] == dim[3] && bstride[2] == dim[3] && cstride[2] == dim[3] && astride[3] == 1 && bstride[3] == 1 && cstride[3] == 1)

    {

      // Special casing if the ainc[3] is the same as dim[3] (do memcpy for the last two dim)

      for (i[0] = 0; i[0] < dim[0]; 
i[0]++3
)

      {

        float* ap0 = ap + i[0] * astride[0];

        float* bp0 = bp + i[0] * bstride[0];

        float* cp0 = cp + i[0] * cstride[0];

        for (i[1] = 0; i[1] < dim[1]; 
i[1]++3
)

        {

          for (x = 0; x < count; 
x++3
)

            cp0[x] = ap0[x] + bp0[x];

          ap0 += astride[1];

          bp0 += bstride[1];

          cp0 += cstride[1];

        }

      }

      continue;

    }

    // Non-optimal case, need to do skip copy.

    for (i[0] = 0; i[0] < dim[0]; i[0]++)

    {

      float* const ap0 = ap + i[0] * astride[0];

      float* const bp0 = bp + i[0] * bstride[0];

      float* const cp0 = cp + i[0] * cstride[0];

      for (i[1] = 0; i[1] < dim[1]; i[1]++)

      {

        float* ap1 = ap0 + i[1] * astride[1];

        float* bp1 = bp0 + i[1] * bstride[1];

        float* cp1 = cp0 + i[1] * cstride[1];

        for (i[2] = 0; i[2] < dim[2]; i[2]++)

        {

          for (x = 0; x < dim[3]; x++)

            cp1[x * cstride[3]] = ap1[x * astride[3]] + bp1[x * bstride[3]];

          ap1 += astride[2];

          bp1 += bstride[2];

          cp1 += cstride[2];

        }

      }

    }

  }

}

void _ccv_nnc_ewsum_forw_cpu_ref_i32(ccv_nnc_tensor_view_t* const* const inputs, const int input_size, ccv_nnc_tensor_view_t* const* const outputs, const int output_size)

{

  if (input_size == 1 && output_size == 1)

  {

    _ccv_nnc_tensor_transfer_cpu_ref_f32(inputs[0], outputs[0]);

    return;

  }

  // Assuming this is float 32.

  int dim[CCV_NNC_MAX_DIM_ALLOC];

  int astride[CCV_NNC_MAX_DIM_ALLOC];

  int bstride[CCV_NNC_MAX_DIM_ALLOC];

  int cstride[CCV_NNC_MAX_DIM_ALLOC];

  int x, z;

  int k = 0;

  // Bad, I promised this can be inplace operation. Need to first find out if there are share the same pointer first.

  for (z = 1; z < input_size; z++)

  {

    ccv_nnc_tensor_view_t* c = outputs[0];

    ccv_nnc_tensor_view_t* a = inputs[z];

    if (c->data.f32 == a->data.f32)

    {

      k = z;

      break;

    }

  }

  for (z = 0; z < input_size - 1; z++)

  {

    ccv_nnc_tensor_view_t* c = outputs[0];

    ccv_nnc_tensor_view_t* a = z > 0 ? c : inputs[k];

    ccv_nnc_tensor_view_t* b = z >= k ? inputs[z + 1] : inputs[z];

    assert(ccv_nnc_tensor_nd(a->info.dim) <= CCV_NNC_MAX_DIM + 2);

    assert(ccv_nnc_tensor_nd(b->info.dim) <= CCV_NNC_MAX_DIM + 2);

    assert(ccv_nnc_tensor_nd(c->info.dim) <= CCV_NNC_MAX_DIM + 2);

    ccv_nnc_tensor_view_get_dim(a, dim);

    assert(ccv_nnc_tensor_view_check_dim(b, dim));

    assert(ccv_nnc_tensor_view_check_dim(c, dim));

    if (!CCV_IS_TENSOR_VIEW(a) && !CCV_IS_TENSOR_VIEW(b) && !CCV_IS_TENSOR_VIEW(c))

    {

      // Super optimal case, just do one for-loop for sum.

      const int tensor_count = ccv_nnc_tensor_count(a->info);

      for (x = 0; x < tensor_count; x++)

        c->data.f32[x] = a->data.f32[x] + b->data.f32[x];

      continue;

    }

    assert(CCV_NNC_MAX_DIM == 2); // Need to change this logic for CCV_NNC_MAX_DIM == other number.

    ccv_nnc_tensor_view_get_stride(a, astride);

    ccv_nnc_tensor_view_get_stride(b, bstride);

    ccv_nnc_tensor_view_get_stride(c, cstride);

    int i[CCV_NNC_MAX_DIM + 2];

    int* const ap = a->data.i32;

    int* const bp = b->data.i32;

    int* const cp = c->data.i32;

    const int count = dim[2] * dim[3];

    if (astride[2] == dim[3] && bstride[2] == dim[3] && cstride[2] == dim[3] && astride[3] == 1 && bstride[3] == 1 && cstride[3] == 1)

    {

      // Special casing if the ainc[3] is the same as dim[3] (do memcpy for the last two dim)

      for (i[0] = 0; i[0] < dim[0]; i[0]++)

      {

        int* ap0 = ap + i[0] * astride[0];

        int* bp0 = bp + i[0] * bstride[0];

        int* cp0 = cp + i[0] * cstride[0];

        for (i[1] = 0; i[1] < dim[1]; i[1]++)

        {

          for (x = 0; x < count; x++)

            cp0[x] = ap0[x] + bp0[x];

          ap0 += astride[1];

          bp0 += bstride[1];

          cp0 += cstride[1];

        }

      }

      continue;

    }

    // Non-optimal case, need to do skip copy.

    for (i[0] = 0; i[0] < dim[0]; i[0]++)

    {

      int* const ap0 = ap + i[0] * astride[0];

      int* const bp0 = bp + i[0] * bstride[0];

      int* const cp0 = cp + i[0] * cstride[0];

      for (i[1] = 0; i[1] < dim[1]; i[1]++)

      {

        int* ap1 = ap0 + i[1] * astride[1];

        int* bp1 = bp0 + i[1] * bstride[1];

        int* cp1 = cp0 + i[1] * cstride[1];

        for (i[2] = 0; i[2] < dim[2]; i[2]++)

        {

          for (x = 0; x < dim[3]; x++)

            cp1[x * cstride[3]] = ap1[x * astride[3]] + bp1[x * bstride[3]];

          ap1 += astride[2];

          bp1 += bstride[2];

          cp1 += cstride[2];

        }

      }

    }

  }

}

static int _ccv_nnc_ewsum_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)

{

  if (outputs[0]->info.datatype == CCV_32S)

    _ccv_nnc_ewsum_forw_cpu_ref_i32((ccv_nnc_tensor_view_t**)inputs, input_size, (ccv_nnc_tensor_view_t**)outputs, output_size);

  else

    _ccv_nnc_ewsum_forw_cpu_ref_f32((ccv_nnc_tensor_view_t**)inputs, input_size, (ccv_nnc_tensor_view_t**)outputs, output_size);

  return CCV_NNC_EXEC_SUCCESS;

}

static int _ccv_nnc_ewsum_back(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)

{

  // D[x + y + z, x] = 1

  int i;

  if (inputs[0] == 0)

  {

    // Set them to 1.

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

      if (outputs[i])

        _ccv_nnc_tensor_set_cpu_ref_f32((ccv_nnc_tensor_view_t*)outputs[i], 1);

  } else {

    // Copy over the gradient (If they are not pointing to the same tensor already).

    for (i = 0; i < output_size; 
i++15.8k
)

      if (outputs[i] && inputs[0]->data.f32 != outputs[i]->data.f32)

        _ccv_nnc_tensor_transfer_cpu_ref_f32((ccv_nnc_tensor_view_t*)inputs[0], (ccv_nnc_tensor_view_t*)outputs[i]);

  }

  return CCV_NNC_EXEC_SUCCESS;

}

void _ccv_nnc_ewprod_forw_cpu_ref(ccv_nnc_tensor_view_t* const* const inputs, const int input_size, ccv_nnc_tensor_view_t* const* const outputs, const int output_size)

{

  if (input_size == 1 && 
output_size == 10
)

  {

    _ccv_nnc_tensor_transfer_cpu_ref_f32(inputs[0], outputs[0]);

    return;

  }

  // Assuming this is float 32.

  int dim[CCV_NNC_MAX_DIM_ALLOC];

  int astride[CCV_NNC_MAX_DIM_ALLOC];

  int bstride[CCV_NNC_MAX_DIM_ALLOC];

  int cstride[CCV_NNC_MAX_DIM_ALLOC];

  int x, z;

  int k = 0;

  // Bad, I promised this can be inplace operation. Need to first find out if there are share the same pointer first.

  for (z = 1; z < input_size; 
z++30.6k
)

  {

    ccv_nnc_tensor_view_t* c = outputs[0];

    ccv_nnc_tensor_view_t* a = inputs[z];

    if (c->data.f32 == a->data.f32)

    {

      k = z;

      break;

    }

  }

  for (z = 0; z < input_size - 1; 
z++30.7k
)

  {

    ccv_nnc_tensor_view_t* c = outputs[0];

    ccv_nnc_tensor_view_t* a = z > 0 ? 
c0
 : inputs[k];

    ccv_nnc_tensor_view_t* b = z >= k ? 
inputs[z + 1]30.6k
 : 
inputs[z]12
;

    assert(ccv_nnc_tensor_nd(a->info.dim) <= CCV_NNC_MAX_DIM + 2);

    assert(ccv_nnc_tensor_nd(b->info.dim) <= CCV_NNC_MAX_DIM + 2);

    assert(ccv_nnc_tensor_nd(c->info.dim) <= CCV_NNC_MAX_DIM + 2);

    ccv_nnc_tensor_view_get_dim(a, dim);

    assert(ccv_nnc_tensor_view_check_dim(b, dim));

    assert(ccv_nnc_tensor_view_check_dim(c, dim));

    if (!CCV_IS_TENSOR_VIEW(a) && 
!30.7k
CCV_IS_TENSOR_VIEW30.7k
(b) && 
!30.7k
CCV_IS_TENSOR_VIEW30.7k
(c))

    {

      // Super optimal case, just do one for-loop for sum.

      const int tensor_count = ccv_nnc_tensor_count(a->info);

      for (x = 0; x < tensor_count; 
x++73.5k
)

        c->data.f32[x] = a->data.f32[x] * b->data.f32[x];

      continue;

    }

    assert
(CCV_NNC_MAX_DIM == 2)3
; // Need to change this logic for CCV_NNC_MAX_DIM == other number.

    ccv_nnc_tensor_view_get_stride(a, astride);

    ccv_nnc_tensor_view_get_stride(b, bstride);

    ccv_nnc_tensor_view_get_stride(c, cstride);

    int i[CCV_NNC_MAX_DIM + 2];

    float* const ap = a->data.f32;

    float* const bp = b->data.f32;

    float* const cp = c->data.f32;

    const int count = dim[2] * dim[3];

    if (astride[2] == dim[3] && bstride[2] == dim[3] && cstride[2] == dim[3])

    {

      // Special casing if the ainc[3] is the same as dim[3]

      for (i[0] = 0; i[0] < dim[0]; 
i[0]++3
)

      {

        float* ap0 = ap + i[0] * astride[0];

        float* bp0 = bp + i[0] * bstride[0];

        float* cp0 = cp + i[0] * cstride[0];

        for (i[1] = 0; i[1] < dim[1]; 
i[1]++3
)

        {

          for (x = 0; x < count; 
x++3
)

            cp0[x] = ap0[x] * bp0[x];

          ap0 += astride[1];

          bp0 += bstride[1];

          cp0 += cstride[1];

        }

      }

      continue;

    }

    // Non-optimal case, need to do skip copy.

    for (i[0] = 0; i[0] < dim[0]; i[0]++)

    {

      float* const ap0 = ap + i[0] * astride[0];

      float* const bp0 = bp + i[0] * bstride[0];

      float* const cp0 = cp + i[0] * cstride[0];

      for (i[1] = 0; i[1] < dim[1]; i[1]++)

      {

        float* ap1 = ap0 + i[1] * astride[1];

        float* bp1 = bp0 + i[1] * bstride[1];

        float* cp1 = cp0 + i[1] * cstride[1];

        for (i[2] = 0; i[2] < dim[2]; i[2]++)

        {

          for (x = 0; x < dim[3]; x++)

            cp1[x] = ap1[x] * bp1[x];

          ap1 += astride[2];

          bp1 += bstride[2];

          cp1 += cstride[2];

        }

      }

    }

  }

}

static int _ccv_nnc_ewprod_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)

{

  _ccv_nnc_ewprod_forw_cpu_ref((ccv_nnc_tensor_view_t**)inputs, input_size, (ccv_nnc_tensor_view_t**)outputs, output_size);

  return CCV_NNC_EXEC_SUCCESS;

}

static int _ccv_nnc_ewprod_back(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)

{

  // D[x * y * z, x] = y * z

  // Assuming this is float 32.

  int dim[CCV_NNC_MAX_DIM_ALLOC];

  int gstride[CCV_NNC_MAX_DIM_ALLOC];

  int astride[CCV_NNC_MAX_DIM_ALLOC];

  int bstride[CCV_NNC_MAX_DIM_ALLOC];

  int hstride[CCV_NNC_MAX_DIM_ALLOC];

  int x, z;

  ccv_nnc_tensor_view_t* g = (ccv_nnc_tensor_view_t*)inputs[0];

  ccv_nnc_tensor_view_t* b = (ccv_nnc_tensor_view_t*)inputs[output_size + 1];

  if (g == 0)

  {

    assert(ccv_nnc_tensor_nd(b->info.dim) <= CCV_NNC_MAX_DIM + 2);

    ccv_nnc_tensor_view_get_dim(b, dim);

    ccv_nnc_tensor_view_get_stride(b, bstride);

    for (z = 0; z < output_size; z++)

    {

      ccv_nnc_tensor_view_t* a = (ccv_nnc_tensor_view_t*)inputs[z + 1];

      ccv_nnc_tensor_view_t* h = (ccv_nnc_tensor_view_t*)outputs[z];

      assert(ccv_nnc_tensor_nd(a->info.dim) <= CCV_NNC_MAX_DIM + 2);

      assert(ccv_nnc_tensor_nd(h->info.dim) <= CCV_NNC_MAX_DIM + 2);

      assert(ccv_nnc_tensor_view_check_dim(a, dim));

      assert(ccv_nnc_tensor_view_check_dim(h, dim));

      ccv_nnc_tensor_view_get_stride(a, astride);

      ccv_nnc_tensor_view_get_stride(h, hstride);

      if (!CCV_IS_TENSOR_VIEW(a) && !CCV_IS_TENSOR_VIEW(b) && !CCV_IS_TENSOR_VIEW(h))

      {

        // Super optimal case, just do one for-loop for sum.

        const int tensor_count = ccv_nnc_tensor_count(b->info);

        for (x = 0; x < tensor_count; x++)

          h->data.f32[x] = b->data.f32[x] / a->data.f32[x];

        continue;

      }

      assert(CCV_NNC_MAX_DIM == 2); // Need to change this logic for CCV_NNC_MAX_DIM == other number.

      int i[CCV_NNC_MAX_DIM + 2];

      float* const ap = a->data.f32;

      float* const bp = b->data.f32;

      float* const hp = h->data.f32;

      const int count = dim[2] * dim[3];

      if (astride[2] == dim[3] && bstride[2] == dim[3] && hstride[2] == dim[3])

      {

        // Special casing if the ainc[3] is the same as dim[3]

        for (i[0] = 0; i[0] < dim[0]; i[0]++)

        {

          float* ap0 = ap + i[0] * astride[0];

          float* bp0 = bp + i[0] * bstride[0];

          float* hp0 = hp + i[0] * hstride[0];

          for (i[1] = 0; i[1] < dim[1]; i[1]++)

          {

            for (x = 0; x < count; x++)

              hp0[x] = bp0[x] / ap0[x];

            ap0 += astride[1];

            bp0 += bstride[1];

            hp0 += hstride[1];

          }

        }

        continue;

      }

      // Non-optimal case, need to do skip copy.

      for (i[0] = 0; i[0] < dim[0]; i[0]++)

      {

        float* const ap0 = ap + i[0] * astride[0];

        float* const bp0 = bp + i[0] * bstride[0];

        float* const hp0 = hp + i[0] * hstride[0];

        for (i[1] = 0; i[1] < dim[1]; i[1]++)

        {

          float* ap1 = ap0 + i[1] * astride[1];

          float* bp1 = bp0 + i[1] * bstride[1];

          float* hp1 = hp0 + i[1] * hstride[1];

          for (i[2] = 0; i[2] < dim[2]; i[2]++)

          {

            for (x = 0; x < dim[3]; x++)

              hp1[x] = bp1[x] / ap1[x];

            ap1 += astride[2];

            bp1 += bstride[2];

            hp1 += hstride[2];

          }

        }

      }

    }

  } else {

    assert(ccv_nnc_tensor_nd(g->info.dim) <= CCV_NNC_MAX_DIM + 2);

    assert(ccv_nnc_tensor_nd(b->info.dim) <= CCV_NNC_MAX_DIM + 2);

    ccv_nnc_tensor_view_get_dim(b, dim);

    assert(ccv_nnc_tensor_view_check_dim(g, dim));

    ccv_nnc_tensor_view_get_stride(b, bstride);

    ccv_nnc_tensor_view_get_stride(g, gstride);

    for (z = 0; z < output_size; 
z++40.1k
)

    {

      ccv_nnc_tensor_view_t* a = (ccv_nnc_tensor_view_t*)inputs[z + 1];

      ccv_nnc_tensor_view_t* h = (ccv_nnc_tensor_view_t*)outputs[z];

      assert(ccv_nnc_tensor_nd(a->info.dim) <= CCV_NNC_MAX_DIM + 2);

      assert(ccv_nnc_tensor_nd(h->info.dim) <= CCV_NNC_MAX_DIM + 2);

      assert(ccv_nnc_tensor_view_check_dim(a, dim));

      assert(ccv_nnc_tensor_view_check_dim(h, dim));

      ccv_nnc_tensor_view_get_stride(a, astride);

      ccv_nnc_tensor_view_get_stride(h, hstride);

      if (!CCV_IS_TENSOR_VIEW(g) && 
!40.1k
CCV_IS_TENSOR_VIEW40.1k
(a) && 
!40.1k
CCV_IS_TENSOR_VIEW40.1k
(b) && 
!40.1k
CCV_IS_TENSOR_VIEW40.1k
(h))

      {

        // Super optimal case, just do one for-loop for sum.

        const int tensor_count = ccv_nnc_tensor_count(g->info);

        for (x = 0; x < tensor_count; 
x++91.8k
)

          h->data.f32[x] = g->data.f32[x] * b->data.f32[x] / a->data.f32[x];

        continue;

      }

      assert
(CCV_NNC_MAX_DIM == 2)3
; // Need to change this logic for CCV_NNC_MAX_DIM == other number.

      int i[CCV_NNC_MAX_DIM + 2];

      float* const gp = g->data.f32;

      float* const ap = a->data.f32;

      float* const bp = b->data.f32;

      float* const hp = h->data.f32;

      const int count = dim[2] * dim[3];

      if (gstride[2] == dim[3] && astride[2] == dim[3] && bstride[2] == dim[3] && hstride[2] == dim[3])

      {

        // Special casing if the ainc[3] is the same as dim[3]

        for (i[0] = 0; i[0] < dim[0]; 
i[0]++3
)

        {

          float* gp0 = gp + i[0] * gstride[0];

          float* ap0 = ap + i[0] * astride[0];

          float* bp0 = bp + i[0] * bstride[0];

          float* hp0 = hp + i[0] * hstride[0];

          for (i[1] = 0; i[1] < dim[1]; 
i[1]++3
)

          {

            for (x = 0; x < count; 
x++3
)

              hp0[x] = gp0[x] * bp0[x] / ap0[x];

            gp0 += gstride[1];

            ap0 += astride[1];

            bp0 += bstride[1];

            hp0 += hstride[1];

          }

        }

        continue;

      }

      // Non-optimal case, need to do skip copy.

      for (i[0] = 0; i[0] < dim[0]; i[0]++)

      {

        float* const gp0 = gp + i[0] * gstride[0];

        float* const ap0 = ap + i[0] * astride[0];

        float* const bp0 = bp + i[0] * bstride[0];

        float* const hp0 = hp + i[0] * hstride[0];

        for (i[1] = 0; i[1] < dim[1]; i[1]++)

        {

          float* gp1 = gp0 + i[1] * gstride[1];

          float* ap1 = ap0 + i[1] * astride[1];

          float* bp1 = bp0 + i[1] * bstride[1];

          float* hp1 = hp0 + i[1] * hstride[1];

          for (i[2] = 0; i[2] < dim[2]; i[2]++)

          {

            for (x = 0; x < dim[3]; x++)

              hp1[x] = gp1[x] * bp1[x] / ap1[x];

            gp1 += gstride[2];

            ap1 += astride[2];

            bp1 += bstride[2];

            hp1 += hstride[2];

          }

        }

      }

    }

  }

  return CCV_NNC_EXEC_SUCCESS;

}

static void _ccv_nnc_ewdiv_forw_cpu_ref(const float p, ccv_nnc_tensor_view_t* const a, ccv_nnc_tensor_view_t* const b, ccv_nnc_tensor_view_t* const c)

{

  // Assuming this is float 32.

  int dim[CCV_NNC_MAX_DIM_ALLOC];

  int astride[CCV_NNC_MAX_DIM_ALLOC];

  int bstride[CCV_NNC_MAX_DIM_ALLOC];

  int cstride[CCV_NNC_MAX_DIM_ALLOC];

  if (a == 0) // Take 0 as all ones tensor.

  {

    assert(ccv_nnc_tensor_nd(b->info.dim) <= CCV_NNC_MAX_DIM + 2);

    assert(ccv_nnc_tensor_nd(c->info.dim) <= CCV_NNC_MAX_DIM + 2);

    ccv_nnc_tensor_view_get_dim(b, dim);

    assert(ccv_nnc_tensor_view_check_dim(c, dim));

    int x;

    if (!CCV_IS_TENSOR_VIEW(b) && !CCV_IS_TENSOR_VIEW(c))

    {

      // Super optimal case, just do one for-loop for sum.

      const int tensor_count = ccv_nnc_tensor_count(b->info);

      for (x = 0; x < tensor_count; 
x++1.98k
)

        c->data.f32[x] = p / b->data.f32[x];

      return;

    }

    assert
(CCV_NNC_MAX_DIM == 2)0
; // Need to change this logic for CCV_NNC_MAX_DIM == other number.

    ccv_nnc_tensor_view_get_stride(b, bstride);

    ccv_nnc_tensor_view_get_stride(c, cstride);

    int i[CCV_NNC_MAX_DIM + 2];

    float* const bp = b->data.f32;

    float* const cp = c->data.f32;

    const int count = dim[2] * dim[3];

    if (bstride[2] == dim[3] && cstride[2] == dim[3])

    {

      // Special casing if the ainc[3] is the same as dim[3]

      for (i[0] = 0; i[0] < dim[0]; i[0]++)

      {

        float* bp0 = bp + i[0] * bstride[0];

        float* cp0 = cp + i[0] * cstride[0];

        for (i[1] = 0; i[1] < dim[1]; i[1]++)

        {

          for (x = 0; x < count; x++)

            cp0[x] = p / bp0[x];

          bp0 += bstride[1];

          cp0 += cstride[1];

        }

      }

      return;

    }

    // Non-optimal case, need to do skip copy.

    for (i[0] = 0; i[0] < dim[0]; i[0]++)

    {

      float* const bp0 = bp + i[0] * bstride[0];

      float* const cp0 = cp + i[0] * cstride[0];

      for (i[1] = 0; i[1] < dim[1]; i[1]++)

      {

        float* bp1 = bp0 + i[1] * bstride[1];

        float* cp1 = cp0 + i[1] * cstride[1];

        for (i[2] = 0; i[2] < dim[2]; i[2]++)

        {

          for (x = 0; x < dim[3]; x++)

            cp1[x] = p / bp1[x];

          bp1 += bstride[2];

          cp1 += cstride[2];

        }

      }

    }

  } else {

    assert(ccv_nnc_tensor_nd(a->info.dim) <= CCV_NNC_MAX_DIM + 2);

    assert(ccv_nnc_tensor_nd(b->info.dim) <= CCV_NNC_MAX_DIM + 2);

    assert(ccv_nnc_tensor_nd(c->info.dim) <= CCV_NNC_MAX_DIM + 2);

    ccv_nnc_tensor_view_get_dim(a, dim);

    assert(ccv_nnc_tensor_view_check_dim(b, dim));

    assert(ccv_nnc_tensor_view_check_dim(c, dim));

    int x;

    if (!CCV_IS_TENSOR_VIEW(a) && !CCV_IS_TENSOR_VIEW(b) && !CCV_IS_TENSOR_VIEW(c))

    {

      // Super optimal case, just do one for-loop for sum.

      const int tensor_count = ccv_nnc_tensor_count(a->info);

      for (x = 0; x < tensor_count; 
x++6.81k
)

        c->data.f32[x] = p * a->data.f32[x] / b->data.f32[x];

      return;

    }

    assert
(CCV_NNC_MAX_DIM == 2)0
; // Need to change this logic for CCV_NNC_MAX_DIM == other number.

    ccv_nnc_tensor_view_get_stride(a, astride);

    ccv_nnc_tensor_view_get_stride(b, bstride);

    ccv_nnc_tensor_view_get_stride(c, cstride);

    int i[CCV_NNC_MAX_DIM + 2];

    float* const ap = a->data.f32;

    float* const bp = b->data.f32;

    float* const cp = c->data.f32;

    const int count = dim[2] * dim[3];

    if (astride[2] == dim[3] && bstride[2] == dim[3] && cstride[2] == dim[3])

    {

      // Special casing if the ainc[3] is the same as dim[3]

      for (i[0] = 0; i[0] < dim[0]; i[0]++)

      {

        float* ap0 = ap + i[0] * astride[0];

        float* bp0 = bp + i[0] * bstride[0];

        float* cp0 = cp + i[0] * cstride[0];

        for (i[1] = 0; i[1] < dim[1]; i[1]++)

        {

          for (x = 0; x < count; x++)

            cp0[x] = p * ap0[x] / bp0[x];

          ap0 += astride[1];

          bp0 += bstride[1];

          cp0 += cstride[1];

        }

      }

      return;

    }

    // Non-optimal case, need to do skip copy.

    for (i[0] = 0; i[0] < dim[0]; i[0]++)

    {

      float* const ap0 = ap + i[0] * astride[0];

      float* const bp0 = bp + i[0] * bstride[0];

      float* const cp0 = cp + i[0] * cstride[0];

      for (i[1] = 0; i[1] < dim[1]; i[1]++)

      {

        float* ap1 = ap0 + i[1] * astride[1];

        float* bp1 = bp0 + i[1] * bstride[1];

        float* cp1 = cp0 + i[1] * cstride[1];

        for (i[2] = 0; i[2] < dim[2]; i[2]++)

        {

          for (x = 0; x < dim[3]; x++)

            cp1[x] = p * ap1[x] / bp1[x];

          ap1 += astride[2];

          bp1 += bstride[2];

          cp1 += cstride[2];

        }

      }

    }

  }

}

static int _ccv_nnc_ewdiv_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)

{

  _ccv_nnc_ewdiv_forw_cpu_ref(1, (ccv_nnc_tensor_view_t*)inputs[0], (ccv_nnc_tensor_view_t*)inputs[1], (ccv_nnc_tensor_view_t*)outputs[0]);

  return CCV_NNC_EXEC_SUCCESS;

}

static int _ccv_nnc_ewdiv_back(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)

{

  // D[x / y, x] = 1 / y, D[x / y, y] = -x / y^2

  if (output_size == 1 || 
outputs[1] == 016
)

  {

    // When we only need D[x / y, x]

    _ccv_nnc_ewdiv_forw_cpu_ref(1, (ccv_nnc_tensor_view_t*)inputs[0], (ccv_nnc_tensor_view_t*)inputs[2], (ccv_nnc_tensor_view_t*)outputs[0]);

    return CCV_NNC_EXEC_SUCCESS;

  }

  int dim[CCV_NNC_MAX_DIM_ALLOC];

  int gstride[CCV_NNC_MAX_DIM_ALLOC];

  int bstride[CCV_NNC_MAX_DIM_ALLOC];

  int cstride[CCV_NNC_MAX_DIM_ALLOC];

  int hastride[CCV_NNC_MAX_DIM_ALLOC];

  int hbstride[CCV_NNC_MAX_DIM_ALLOC];

  ccv_nnc_tensor_view_t* g = (ccv_nnc_tensor_view_t*)inputs[0];

  ccv_nnc_tensor_view_t* b = (ccv_nnc_tensor_view_t*)inputs[2];

  ccv_nnc_tensor_view_t* c = (ccv_nnc_tensor_view_t*)inputs[3];

  ccv_nnc_tensor_view_t* ha = (ccv_nnc_tensor_view_t*)outputs[0];

  ccv_nnc_tensor_view_t* hb = (ccv_nnc_tensor_view_t*)outputs[1];

  if (g == 0)

  {

    assert(ccv_nnc_tensor_nd(b->info.dim) <= CCV_NNC_MAX_DIM + 2);

    assert(ccv_nnc_tensor_nd(c->info.dim) <= CCV_NNC_MAX_DIM + 2);

    assert(ccv_nnc_tensor_nd(hb->info.dim) <= CCV_NNC_MAX_DIM + 2);

    ccv_nnc_tensor_view_get_dim(b, dim);

    assert(ccv_nnc_tensor_view_check_dim(c, dim));

    assert(ccv_nnc_tensor_view_check_dim(hb, dim));

    if (ha)

    {

      assert(ccv_nnc_tensor_nd(ha->info.dim) <= CCV_NNC_MAX_DIM + 2);

      assert(ccv_nnc_tensor_view_check_dim(ha, dim));

    }

    int x;

    if (!CCV_IS_TENSOR_VIEW(b) && !CCV_IS_TENSOR_VIEW(c) && (ha == 0 || !CCV_IS_TENSOR_VIEW(ha)) && !CCV_IS_TENSOR_VIEW(hb))

    {

      // Super optimal case, just do one for-loop for sum.

      const int tensor_count = ccv_nnc_tensor_count(b->info);

      if (ha == 0)

      {

        for (x = 0; x < tensor_count; x++)

        {

          const float v = 1 / b->data.f32[x];

          hb->data.f32[x] = -c->data.f32[x] * v;

        }

      } else {

        for (x = 0; x < tensor_count; x++)

        {

          const float v = 1 / b->data.f32[x];

          ha->data.f32[x] = v;

          hb->data.f32[x] = -c->data.f32[x] * v;

        }

      }

      return CCV_NNC_EXEC_SUCCESS;

    }

    assert(CCV_NNC_MAX_DIM == 2); // Need to change this logic for CCV_NNC_MAX_DIM == other number.

    ccv_nnc_tensor_view_get_stride(b, bstride);

    ccv_nnc_tensor_view_get_stride(c, cstride);

    ccv_nnc_tensor_view_get_stride(hb, hbstride);

    int i[CCV_NNC_MAX_DIM + 2];

    float* const bp = b->data.f32;

    float* const cp = c->data.f32;

    float* const hbp = hb->data.f32;

    const int count = dim[2] * dim[3];

    if (ha == 0)

    {

      if (bstride[2] == dim[3] && cstride[2] == dim[3] && hbstride[2] == dim[3])

      {

        // Special casing if the ainc[3] is the same as dim[3]

        for (i[0] = 0; i[0] < dim[0]; i[0]++)

        {

          float* bp0 = bp + i[0] * bstride[0];

          float* cp0 = cp + i[0] * cstride[0];

          float* hbp0 = hbp + i[0] * hbstride[0];

          for (i[1] = 0; i[1] < dim[1]; i[1]++)

          {

            for (x = 0; x < count; x++)

            {

              const float v = 1 / bp0[x];

              hbp0[x] = -cp0[x] * v;

            }

            bp0 += bstride[1];

            cp0 += cstride[1];

            hbp0 += hbstride[1];

          }

        }

        return CCV_NNC_EXEC_SUCCESS;

      }

      // Non-optimal case, need to do skip copy.

      for (i[0] = 0; i[0] < dim[0]; i[0]++)

      {

        float* const bp0 = bp + i[0] * bstride[0];

        float* const cp0 = cp + i[0] * cstride[0];

        float* const hbp0 = hbp + i[0] * hbstride[0];

        for (i[1] = 0; i[1] < dim[1]; i[1]++)

        {

          float* bp1 = bp0 + i[1] * bstride[1];

          float* cp1 = cp0 + i[1] * cstride[1];

          float* hbp1 = hbp0 + i[1] * hbstride[1];

          for (i[2] = 0; i[2] < dim[2]; i[2]++)

          {

            for (x = 0; x < dim[3]; x++)

            {

              const float v = 1 / bp1[x];

              hbp1[x] = -cp1[x] * v;

            }

            bp1 += bstride[2];

            cp1 += cstride[2];

            hbp1 += hbstride[2];

          }

        }

      }

    } else {

      float* const hap = ha->data.f32;

      ccv_nnc_tensor_view_get_stride(ha, hastride);

      if (bstride[2] == dim[3] && cstride[2] == dim[3] && hastride[2] == dim[3] && hbstride[2] == dim[3])

      {

        // Special casing if the ainc[3] is the same as dim[3]

        for (i[0] = 0; i[0] < dim[0]; i[0]++)

        {

          float* bp0 = bp + i[0] * bstride[0];

          float* cp0 = cp + i[0] * cstride[0];

          float* hap0 = hap + i[0] * hastride[0];

          float* hbp0 = hbp + i[0] * hbstride[0];

          for (i[1] = 0; i[1] < dim[1]; i[1]++)

          {

            for (x = 0; x < count; x++)

            {

              const float v = 1 / bp0[x];

              hap0[x] = v;

              hbp0[x] = -cp0[x] * v;

            }

            bp0 += bstride[1];

            cp0 += cstride[1];

            hap0 += hastride[1];

            hbp0 += hbstride[1];

          }

        }

        return CCV_NNC_EXEC_SUCCESS;

      }

      // Non-optimal case, need to do skip copy.

      for (i[0] = 0; i[0] < dim[0]; i[0]++)

      {

        float* const bp0 = bp + i[0] * bstride[0];

        float* const cp0 = cp + i[0] * cstride[0];

        float* const hap0 = hap + i[0] * hastride[0];

        float* const hbp0 = hbp + i[0] * hbstride[0];

        for (i[1] = 0; i[1] < dim[1]; i[1]++)

        {

          float* bp1 = bp0 + i[1] * bstride[1];

          float* cp1 = cp0 + i[1] * cstride[1];

          float* hap1 = hap0 + i[1] * hastride[1];

          float* hbp1 = hbp0 + i[1] * hbstride[1];

          for (i[2] = 0; i[2] < dim[2]; i[2]++)

          {

            for (x = 0; x < dim[3]; x++)

            {

              const float v = 1 / bp1[x];

              hap1[x] = v;

              hbp1[x] = -cp1[x] * v;

            }

            bp1 += bstride[2];

            cp1 += cstride[2];

            hap1 += hastride[2];

            hbp1 += hbstride[2];

          }

        }

      }

    }

  } else {

    assert(ccv_nnc_tensor_nd(g->info.dim) <= CCV_NNC_MAX_DIM + 2);

    assert(ccv_nnc_tensor_nd(b->info.dim) <= CCV_NNC_MAX_DIM + 2);

    assert(ccv_nnc_tensor_nd(c->info.dim) <= CCV_NNC_MAX_DIM + 2);

    assert(ccv_nnc_tensor_nd(hb->info.dim) <= CCV_NNC_MAX_DIM + 2);

    ccv_nnc_tensor_view_get_dim(b, dim);

    assert(ccv_nnc_tensor_view_check_dim(g, dim));

    assert(ccv_nnc_tensor_view_check_dim(c, dim));

    assert(ccv_nnc_tensor_view_check_dim(hb, dim));

    if (ha)

    {

      assert(ccv_nnc_tensor_nd(ha->info.dim) <= CCV_NNC_MAX_DIM + 2);

      assert(ccv_nnc_tensor_view_check_dim(ha, dim));

    }

    int x;

    if (!CCV_IS_TENSOR_VIEW(g) && !CCV_IS_TENSOR_VIEW(b) && !CCV_IS_TENSOR_VIEW(c) && (ha == 0 || 
!1
CCV_IS_TENSOR_VIEW1
(ha)) && !CCV_IS_TENSOR_VIEW(hb))

    {

      // Super optimal case, just do one for-loop for sum.

      const int tensor_count = ccv_nnc_tensor_count(g->info);

      if (ha == 0)

      {

        for (x = 0; x < tensor_count; 
x++1.49k
)

        {

          const float v = g->data.f32[x] / b->data.f32[x];

          hb->data.f32[x] = -c->data.f32[x] * v;

        }

      } else {

        for (x = 0; x < tensor_count; 
x++1
)

        {

          const float v = g->data.f32[x] / b->data.f32[x];

          ha->data.f32[x] = v;

          hb->data.f32[x] = -c->data.f32[x] * v;

        }

      }

      return CCV_NNC_EXEC_SUCCESS;

    }

    assert
(CCV_NNC_MAX_DIM == 2)0
; // Need to change this logic for CCV_NNC_MAX_DIM == other number.

    ccv_nnc_tensor_view_get_stride(g, gstride);

    ccv_nnc_tensor_view_get_stride(b, bstride);

    ccv_nnc_tensor_view_get_stride(c, cstride);

    ccv_nnc_tensor_view_get_stride(hb, hbstride);

    int i[CCV_NNC_MAX_DIM + 2];

    float* const gp = g->data.f32;

    float* const bp = b->data.f32;

    float* const cp = c->data.f32;

    float* const hbp = hb->data.f32;

    const int count = dim[2] * dim[3];

    if (ha == 0)

    {

      if (gstride[2] == dim[3] && bstride[2] == dim[3] && cstride[2] == dim[3] && hbstride[2] == dim[3])

      {

        // Special casing if the ainc[3] is the same as dim[3]

        for (i[0] = 0; i[0] < dim[0]; i[0]++)

        {

          float* gp0 = gp + i[0] * gstride[0];

          float* bp0 = bp + i[0] * bstride[0];

          float* cp0 = cp + i[0] * cstride[0];

          float* hbp0 = hbp + i[0] * hbstride[0];

          for (i[1] = 0; i[1] < dim[1]; i[1]++)