#include "ccv.h"

#include "nnc/ccv_nnc.h"

#include "nnc/ccv_nnc_internal.h"

static int _ccv_nnc_scaled_dot_product_attention_forw_bitmask(const ccv_nnc_cmd_param_t cmd, const int input_size, const int output_size, const uint64_t* const input_bitmasks, const int input_bitmask_size, const uint64_t* const output_bitmasks, const int output_bitmask_size)

{

  // 6 inputs (query, key, value, [attn_mask], [unify head weight], [unify head bias])

  // 8 inputs with varlen (query, key, value, 0, 0, 0, q_seq_offsets, kv_seq_offsets)

  // 3 outputs (y, softmax_lse, [qkv])

  if (cmd.scaled_dot_product_attention.is_varlen && 
input_size == 83
 && 
(input_bitmasks[0] & 255u) == ((1u << 0) | (1u << 1) | (1u << 2) | (1u << 6) | (1u << 7))3
 && 
(output_bitmasks[0] & 1u) == 1u3
)

    return 1;

  if (input_size == 6 && (input_bitmasks[0] & 55u) == 55u && 
(output_bitmasks[0] & 7u) == 7u6
)

    return 1;

  if (input_size == 5 && 
(input_bitmasks[0] & 23u) == 23u0
 && 
(output_bitmasks[0] & 7u) == 7u0
)

    return 1;

  if ((input_bitmasks[0] & 55u) == 7u && 
(output_bitmasks[0] & 3u) == 3u8
)

    return 1;

  return 0;

}

static int _ccv_nnc_allow_query_inplace(const ccv_nnc_cmd_param_t cmd, const int input_idx, const int input_size, const int output_idx, const int output_size)

{

  if (input_idx == 0 && 
output_idx == 012
)

    return 1;

  return 0;

}

static int _ccv_nnc_scaled_dot_product_attention_back_bitmask(const ccv_nnc_cmd_param_t cmd, const int input_size, const int output_size, const uint64_t* const input_bitmasks, const int input_bitmask_size, const uint64_t* const output_bitmasks, const int output_bitmask_size)

{

  // 1, 0, 0, 8, 16, 32, 64?, 128?, 256?, 512, 1024, 2048?

  // 1, 2, 4, 8, 16, 32

  // Inputs (gradient, 0, 0, q, k, v, [attn_mask], [head weight], [bias], y, saved softmax_lse, qkv)

  // Output (dquery, dkey, dvalue, [attn mask], dweight, dbias) [cannot diff against attn_mask]

  if ((input_bitmasks[0] & 4025u) == 4025u && 
(output_bitmasks[0] & 63u) == 55u0
)

    return 1;

  if ((input_bitmasks[0] & 3769u) == 3769u && 
(output_bitmasks[0] & 31u) == 23u0
)

    return 1;

  if ((input_bitmasks[0] & 1593u) == 1593u && 
(output_bitmasks[0] & 7u) == 7u0
)

    return 1;

  return 0;

}

static void _ccv_nnc_scaled_dot_product_attention_tensor_auto_forw(const ccv_nnc_cmd_param_t cmd, const ccv_nnc_tensor_param_t* const inputs, const int input_size, const ccv_nnc_hint_t hint, ccv_nnc_tensor_param_t* const outputs, const int output_size)

{

  assert(input_size >= 3);

  assert(output_size >= 1);

  const int q_nd = ccv_nnc_tensor_nd(inputs[0].dim);

  assert(q_nd == 3 || q_nd == 4);

  const int k_nd = ccv_nnc_tensor_nd(inputs[1].dim);

  assert(k_nd == 3 || k_nd == 4);

  const int v_nd = ccv_nnc_tensor_nd(inputs[2].dim);

  assert(v_nd == 3 || v_nd == 4);

  assert(q_nd == k_nd && k_nd == v_nd);

  if (!cmd.scaled_dot_product_attention.is_varlen && 
input_size > 419
)

  {

    assert(output_size >= 3);

    outputs[0] = inputs[0];

    outputs[0].dim[1] = inputs[0].dim[1]; // sequence length matches query, embedding size matches value * num_head.

    outputs[0].dim[2] = inputs[2].dim[v_nd - 1] * (q_nd == 4 ? inputs[0].dim[2] : 
10
);

    outputs[0].dim[3] = 0;

    // This is saved softmax_lse, which would be in 32F if exists.

    outputs[1] = inputs[0];

    outputs[1].dim[q_nd - 3] = inputs[0].dim[q_nd - 2];

    outputs[1].dim[q_nd - 2] = inputs[0].dim[q_nd - 3];

    outputs[1].dim[q_nd - 1] = 0;

    outputs[1].datatype = CCV_32F;

    outputs[2] = inputs[0];

    outputs[2].dim[q_nd - 1] = inputs[2].dim[v_nd - 1]; // sequence length matches query, embedding size matches value.

  } else {

    outputs[0] = inputs[0];

    outputs[0].dim[q_nd - 1] = inputs[2].dim[v_nd - 1]; // sequence length matches query, embedding size matches value.

    if (output_size == 1)

      return;

    assert
(output_size > 1)8
;

    // This is saved softmax_lse, which would be in 32F if exists.

    outputs[1] = inputs[0];

    if (cmd.scaled_dot_product_attention.is_varlen)

    {

      assert(input_size >= 8);

      assert(q_nd == 4);

      assert(inputs[6].dim[0] > 0);

      assert(cmd.scaled_dot_product_attention.max_seqlen_q > 0);

      outputs[1].dim[0] = inputs[6].dim[0] - 1;

      outputs[1].dim[1] = inputs[0].dim[2];

      outputs[1].dim[2] = cmd.scaled_dot_product_attention.max_seqlen_q;

    } else {

      outputs[1].dim[q_nd - 3] = inputs[0].dim[q_nd - 2];

      outputs[1].dim[q_nd - 2] = inputs[0].dim[q_nd - 3];

    }

    outputs[1].dim[q_nd - 1] = 0;

    outputs[1].datatype = CCV_32F;

  }

}

static void _ccv_nnc_scaled_dot_product_attention_tensor_auto_back(const ccv_nnc_cmd_param_t cmd, const ccv_nnc_tensor_param_t* const inputs, const int input_size, const ccv_nnc_hint_t hint, ccv_nnc_tensor_param_t* const outputs, const int output_size)

{

  assert(input_size >= 6);

  assert(output_size >= 3);

  int i;

  for (i = 0; i < output_size; 
i++3
)

    outputs[i] = inputs[3 + i];

}

REGISTER_COMMAND(CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_FORWARD)(ccv_nnc_cmd_registry_t* const registry)

  FIND_BACKEND(ccv_nnc_scaled_dot_product_attention_cpu_ref.c, mps/ccv_nnc_scaled_dot_product_attention_mps.m, gpu/ccv_nnc_scaled_dot_product_attention_flash_attn.cu)

{

  registry->bitmask = _ccv_nnc_scaled_dot_product_attention_forw_bitmask;

  registry->tensor_auto = _ccv_nnc_scaled_dot_product_attention_tensor_auto_forw;

  registry->allow_inplace = _ccv_nnc_allow_query_inplace;

}

REGISTER_COMMAND(CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_BACKWARD)(ccv_nnc_cmd_registry_t* const registry)

  FIND_BACKEND(ccv_nnc_scaled_dot_product_attention_cpu_ref.c, mps/ccv_nnc_scaled_dot_product_attention_mps.m, gpu/ccv_nnc_scaled_dot_product_attention_flash_attn.cu)

{

  registry->bitmask = _ccv_nnc_scaled_dot_product_attention_back_bitmask;

  registry->tensor_auto = _ccv_nnc_scaled_dot_product_attention_tensor_auto_back;

}

//@REGISTER_EASY_COMMAND_MACRO(CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_FORWARD)

#define CMD_SCALED_DOT_PRODUCT_ATTENTION_FORWARD(_scale, _is_causal) ccv_nnc_cmd(CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_FORWARD, 0, ((ccv_nnc_cmd_param_t){.size={.dim={1,1,1}},.scaled_dot_product_attention={.scale=_scale,.is_causal=_is_causal}}), 0)

//@REGISTER_EASY_COMMAND_MACRO(CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_BACKWARD)

#define CMD_SCALED_DOT_PRODUCT_ATTENTION_BACKWARD(_scale, _is_causal) ccv_nnc_cmd(CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_BACKWARD, 0, ((ccv_nnc_cmd_param_t){.size={.dim={1,1,1}},.scaled_dot_product_attention={.scale=_scale,.is_causal=_is_causal}}), 0)