/*
 * SPDX-FileCopyrightText: Copyright (c) 2011-2024 NVIDIA CORPORATION & AFFILIATES. All rights
 * reserved. SPDX-License-Identifier: NVIDIA TensorRT Source Code License Agreement
 *
 * NVIDIA CORPORATION, its affiliates and licensors retain all intellectual
 * property and proprietary rights in and to this material, related
 * documentation and any modifications thereto. Any use, reproduction,
 * disclosure or distribution of this material and related documentation
 * without an express license agreement from NVIDIA CORPORATION or
 * its affiliates is strictly prohibited.
 */

#pragma once

#include <fmha/gemm.h>
#include <fmha/kernel_traits.h>
#include <fused_multihead_attention_kernel.h>

namespace fused_multihead_attention {

////////////////////////////////////////////////////////////////////////////////////////////////////

template <typename Kernel_traits, typename Params>
inline __device__ void device_flash_attention(Params const& params) {
  // The instruction traits.
  using Traits_p = typename Kernel_traits::Traits_p;
  using Traits_o = typename Kernel_traits::Traits_o;

  // The description of the CTA tile for the 1st batched GEMM.
  using Cta_tile_p = typename Kernel_traits::Cta_tile_p;
  // The description of the CTA tile for the 2nd batched GEMM.
  using Cta_tile_o = typename Kernel_traits::Cta_tile_o;

  // The MMA tile for the 1st GEMM.
  using Mma_tile_p = typename Traits_p::template Mma_tile<Cta_tile_p>;
  // The MMA tile for the 2nd GEMM.
  using Mma_tile_o = typename Traits_o::template Mma_tile<Cta_tile_o>;

  // The global memory tile to load Q.
  using Gmem_tile_q = typename Kernel_traits::Gmem_tile_q;
  // The shared memory tile to swizzle Q.
  using Smem_tile_q = typename Kernel_traits::Smem_tile_q;

  // The global memory tile to load K.
  using Gmem_tile_k = typename Kernel_traits::Gmem_tile_k;
  // The shared memory tile to swizzle K.
  using Smem_tile_k = typename Kernel_traits::Smem_tile_k;

  // The global memory tile to load V.
  using Gmem_tile_v = typename Kernel_traits::Gmem_tile_v;
  // The shared memory tile to swizzle V.
  using Smem_tile_v = typename Kernel_traits::Smem_tile_v;

  // The global memory tile to store O.
  using Gmem_tile_o = typename Kernel_traits::Gmem_tile_o;
  // The shared memory tile to swizzle O.
  using Smem_tile_o = typename Kernel_traits::Smem_tile_o;

  // Do we use LDGSTS for Q, K or V?
  enum { USE_LDGSTS_Q = Kernel_traits::USE_LDGSTS_Q };

  enum { USE_LDGSTS_K = Kernel_traits::USE_LDGSTS_K };

  enum { USE_LDGSTS_V = Kernel_traits::USE_LDGSTS_V };

  enum { USE_LDGSTS_KV = Kernel_traits::USE_LDGSTS_K || Kernel_traits::USE_LDGSTS_V };

  // Do we use LDGSTS for any of the 3 input matrices.
  enum { USE_LDGSTS = USE_LDGSTS_Q || USE_LDGSTS_K || USE_LDGSTS_V };

  // If either K or V uses LDGSTS, they cannot share a buffer.
  static_assert(!(USE_LDGSTS_K || USE_LDGSTS_V) || !Kernel_traits::SHARE_SMEM_FOR_K_AND_V, "");

  // Fragment double buffer (reduce register pressure)
  enum {
    FRAGMENT_K_SIZE_IN_K_DIM = (Kernel_traits::LIMIT_QK_FRAGMENTS) * 2 +
                               !(Kernel_traits::LIMIT_QK_FRAGMENTS)*Mma_tile_p::MMAS_K
  };

  static_assert(!(Kernel_traits::LIMIT_QK_FRAGMENTS && USE_LDGSTS_K), "");

  // Shared memory.
  extern __shared__ char smem_[];

  // The block index for the batch.
  int const bidb = blockIdx.y;
  // The block index for the head.
  int const bidh = blockIdx.x;
  // The thread index.
  int const tidx = threadIdx.x;

  // The block info.
  Single_cta<Kernel_traits::VERSION> const binfo(params, bidb, bidh, 0, tidx);
  if (binfo.stop_early()) {
    return;
  }

  // Create the object to control the masks.
  fmha::Mask<Traits_p, Cta_tile_p, Kernel_traits::MASK_VERSION> mask(params, binfo, tidx);

  // Allocate the global memory tile loader for Q.
  // Gmem_tile_q gmem_q(params, 0, binfo, tidx);
  Gmem_tile_q gmem_q(params, 0, binfo, tidx);
  // Allocate the shared memory tile loader for Q.
  Smem_tile_q smem_q(&smem_[0], tidx);

  // Allocate the global memory tile loader for K.
  Gmem_tile_k gmem_k(params, 1, binfo, tidx);
  // Allocate the shared memory tile loader for K.
  Smem_tile_k smem_k(&smem_[Smem_tile_q::BYTES_PER_TILE], tidx);

  // Allocate the global memory tile loader for V.
  Gmem_tile_v gmem_v(params, 2, binfo, tidx);

  // The base pointer of smem_v;
  char* smem_v_ = nullptr;
  if (Kernel_traits::SHARE_SMEM_FOR_K_AND_V) {
    smem_v_ = &smem_[Smem_tile_q::BYTES_PER_TILE];
  } else {
    smem_v_ = &smem_[Smem_tile_q::BYTES_PER_TILE + Smem_tile_k::BYTES_PER_TILE];
  }
  // Allocate the shared memory tile loader for V. We use the same as K so be careful!!!
  Smem_tile_v smem_v(smem_v_, tidx);

  // Allocate the global memory tile loader for O.
  // Gmem_tile_o gmem_o(params, binfo, tidx);
  Gmem_tile_o gmem_o(params, binfo, tidx);
  // Allocate the shared memory tile loader for O. We use the same as K so be careful!!!
  Smem_tile_o smem_o(&smem_[Smem_tile_q::BYTES_PER_TILE], tidx);

  // Trigger the loads for Q.
  gmem_q.load(smem_q);
  // Trigger the loads for K.
  gmem_k.load(smem_k);
  // Trigger the loads for K.
  gmem_v.load(smem_v);

  // Push the LDGDEPBAR instruction after the loads for Q, K and V.
  fmha::ldgdepbar<USE_LDGSTS>();

  // Commit the data for Q and K to shared memory.
  gmem_q.commit(smem_q);
  gmem_k.commit(smem_k);

  // Commit the data for V to shared memory.
  if (!Kernel_traits::SHARE_SMEM_FOR_K_AND_V) {
    gmem_v.commit(smem_v);
  }

  // Make sure the data is in shared memory.
  fmha::depbar<USE_LDGSTS, 1>();
  __syncthreads();

  // Load the fragments for Q.
  // NOTE: lds full frag_q as it will be used by all kv loops
  typename Smem_tile_q::Fragment frag_q[Mma_tile_p::MMAS_K][Mma_tile_p::MMAS_M];
#pragma unroll
  for (int ki = 0; ki < Mma_tile_p::MMAS_K; ++ki) {
    smem_q.load(frag_q[ki], ki);
  }

  // Load the fragments for K. We keep the data in registers during the entire kernel.
  typename Smem_tile_k::Fragment frag_k[FRAGMENT_K_SIZE_IN_K_DIM][Mma_tile_p::MMAS_N];
  if (Kernel_traits::LIMIT_QK_FRAGMENTS) {
    smem_k.load(frag_k[0], 0);
  } else {
#pragma unroll
    for (int ki = 0; ki < Mma_tile_p::MMAS_K; ++ki) {
      smem_k.load(frag_k[ki], ki);
    }
  }

  // Commit the data for V to shared memory if it has not been done already.
  if (Kernel_traits::SHARE_SMEM_FOR_K_AND_V) {
    // Make sure we are done loading the fragments for K.
    __syncthreads();

    // Commit the data to shared memory for V.
    gmem_v.commit(smem_v);

    // Make sure the data is in shared memory.
    __syncthreads();
  }

  // Load the fragments for V. We keep the data in registers during the entire kernel.
  typename Smem_tile_v::Fragment frag_v[Mma_tile_o::MMAS_K][Mma_tile_o::VALID_MMAS_N];
#pragma unroll
  for (int ki = 0; ki < Mma_tile_o::MMAS_K; ++ki) {
    smem_v.load(frag_v[ki], ki);
  }

  // Store/load P to/from memory (for debugging).
#if defined(STORE_P)
  enum { BITS_PER_ELT_P = sizeof(typename Traits_p::Accumulator_type) * 8 };

  using Gmem_tile_p = fmha::Gmem_tile_p<Traits_p, Cta_tile_p, BITS_PER_ELT_P>;
  Gmem_tile_p gmem_p(params.p_ptr, params.p_stride_in_bytes, params.scale_bmm1, tidx);
#endif

  // Store S to memory (for debugging). NOTE: We use A_type as C_type is int32 for IMMA???
#if defined(STORE_S)
  enum { BITS_PER_ELT_S = sizeof(typename Traits_p::A_type) * 8 };

  using Gmem_tile_s = fmha::Gmem_tile_s<Traits_p, Cta_tile_p, BITS_PER_ELT_S>;
  Gmem_tile_s gmem_s(params.s_ptr, params.s_stride_in_bytes, params.scale_softmax, tidx);
#endif

  // Create the object to do the softmax.
  using Softmax = fmha::Softmax<Traits_p, Cta_tile_p, Kernel_traits>;
  Softmax softmax(params, &smem_[Smem_tile_q::BYTES_PER_TILE], bidb, tidx);

  // Prefetch next kv buffer to share memory
  enum {
    PREFETCH_K_BUFFER_TO_SMEM = !Kernel_traits::LIMIT_QK_FRAGMENTS && !Softmax::USE_SHARED_MEMORY
  };

  enum { PREFETCH_V_BUFFER_TO_SMEM = !Kernel_traits::SHARE_SMEM_FOR_K_AND_V };

  enum { PREFETCH_KV_BUFFER_TO_SMEM = PREFETCH_K_BUFFER_TO_SMEM && PREFETCH_V_BUFFER_TO_SMEM };

  // The number of threads per row.
  // enum { THREADS_PER_ROW = Cta_tile_p::WARPS_N * 8 };
  // DEBUG.
  // static_assert(THREADS_PER_ROW == 32, "");
  // END OF DEBUG.
  enum { THREADS_PER_ROW = 32 };

  int const q_loop_bound =
      ((binfo.actual_seqlen + Cta_tile_p::M - 1) / Cta_tile_p::M) * Cta_tile_p::M;

  int const kv_loop_bound =
      ((binfo.actual_seqlen + Cta_tile_p::N - 1) / Cta_tile_p::N) * Cta_tile_p::N;
  // Load over the entire sequence length.
  for (int q_loop = 0, outer = 0; q_loop < q_loop_bound; q_loop += Cta_tile_p::M, outer++) {
    // Reset the mask col.
    mask.reset();
    // Load the mask row.
    mask.load(outer);

    // Declare the accumulators for the 1st gemm.
    fmha::Fragment_accumulator<Traits_o> acc_o[Mma_tile_o::MMAS_M][Mma_tile_o::VALID_MMAS_N];
    using Acc_type_o = typename Traits_o::Accumulator_type;
    fmha::Clear_accumulator<Acc_type_o, Cta_tile_o::WARPS_K>::apply(acc_o);

    // Flash attention updater
    fmha::Fragment_updater<Traits_o, Cta_tile_o> acc_o_updater;

    for (int kv_loop = 0; kv_loop < kv_loop_bound; kv_loop += Cta_tile_p::N) {
      // NOTE: causal mask
      if (Kernel_traits::MASK_VERSION == 3 && q_loop + Cta_tile_p::M <= kv_loop) {
        break;
      }

      // Trigger the load for the next K/V values (smem_k double buffer).
      if (kv_loop + Cta_tile_p::N < kv_loop_bound && PREFETCH_KV_BUFFER_TO_SMEM) {
        // Make sure we are done reading the data (smem).
        if (!(Smem_tile_k::BUFFERS_PER_TILE > 1 && Smem_tile_v::BUFFERS_PER_TILE > 1)) {
          __syncthreads();
        }

        gmem_k.move();
        smem_k.move_to_next_write_buffer();
        gmem_k.load(smem_k);

        gmem_v.move();
        smem_v.move_to_next_write_buffer();
        gmem_v.load(smem_v);

        // Push the LDGDEPBAR instruction after the loads for K.
        fmha::ldgdepbar<USE_LDGSTS>();

        gmem_k.commit(smem_k);
        if (!Kernel_traits::SHARE_SMEM_FOR_K_AND_V) {
          gmem_v.commit(smem_v);
        }
      }

      // Declare the accumulators for the 1st gemm.
      fmha::Fragment_accumulator<Traits_p> acc_p[Mma_tile_p::MMAS_M][Mma_tile_p::MMAS_N];
      using Acc_type_p = typename Traits_p::Accumulator_type;
      fmha::Clear_accumulator<Acc_type_p, Cta_tile_p::WARPS_K>::apply(acc_p);

// Do this part of P^T = (Q * K^T)^T.
#pragma unroll
      for (int ki = 1; ki < Mma_tile_p::MMAS_K; ++ki) {
        int k_ki = (ki - 1);
        if (Kernel_traits::LIMIT_QK_FRAGMENTS) {
          k_ki = (ki - 1) & 1;
          smem_k.load(frag_k[(ki & 1)], ki);
        }

        // Do the math for the values already in registers.
        if (ki <= Mma_tile_p::VALID_MMAS_K) {
          fmha::gemm(acc_p, frag_q[(ki - 1)], frag_k[k_ki]);
        }
      }

      if (Mma_tile_p::MMAS_K <= Mma_tile_p::VALID_MMAS_K) {
        int ki = Mma_tile_p::MMAS_K;
        int k_ki = (ki - 1);
        if (Kernel_traits::LIMIT_QK_FRAGMENTS) {
          k_ki = (ki - 1) & 1;
        }
        fmha::gemm(acc_p, frag_q[(ki - 1)], frag_k[k_ki]);
      }

      // Store the P matrix.
#if defined(STORE_P)
      gmem_p.store(acc_p);
#endif

      // Convert from the accumulator type to FP32 for Softmax.
      softmax.unpack(acc_p);

      // Apply the mask.
      if (params.has_alibi) {
        softmax.apply_mask_alibi(mask, bidh, params.alibi_params);
      } else {
        softmax.apply_mask(mask);
      }

      // Move mask to next kv tile
      mask.move();

      // Make sure we are done reading the data (smem_v).
      if ((Kernel_traits::SHARE_SMEM_FOR_K_AND_V || Kernel_traits::LIMIT_QK_FRAGMENTS) &&
          Softmax::USE_SHARED_MEMORY) {
        __syncthreads();
      }

      // Enable our trick to use the max for INT8 to scale.
      // float p_max[Softmax::ROWS_PER_THREAD];
      if (Kernel_traits::USE_SCALE_MAX) {
        // 16129 == 127 ^ 2.
        float p_max = reinterpret_cast<float const&>(params.scale_bmm1) * 16129.f;
        softmax.apply_exp(p_max);
      } else {
        // Compute the max.
        softmax.template reduce<fmha::Max_>(acc_o_updater.prev_max_);

        // Update acc_max scale of flash attention
        acc_o_updater.update_acc_max();

        if (Softmax::USE_SHARED_MEMORY) {
          // Make sure we are done reading shared memory.
          __syncthreads();
        }

        // Compute the exponential value.
        softmax.apply_exp(acc_o_updater.curr_max_);
      }

      // Compute the sum.
      // float p_sum[Softmax::ROWS_PER_THREAD];
      softmax.template reduce<fmha::Sum_>(acc_o_updater.prev_sum_);

      // Trigger the load for the next K/V values.
      if (kv_loop + Cta_tile_p::N < kv_loop_bound && !PREFETCH_KV_BUFFER_TO_SMEM) {
        gmem_k.move();
        gmem_v.move();
        smem_k.move_to_next_write_buffer();
        smem_v.move_to_next_write_buffer();
        gmem_k.load(smem_k);
        gmem_v.load(smem_v);

        // Push the LDGDEPBAR instruction after the loads for Q, K and V.
        fmha::ldgdepbar<USE_LDGSTS>();

        gmem_k.commit(smem_k);
        if (!Kernel_traits::SHARE_SMEM_FOR_K_AND_V) {
          gmem_v.commit(smem_v);
        }
      }

      // Update acc_sum scale of flash attention
      acc_o_updater.update_acc_sum();

      // Finalize softmax on the accumulators of P^T.
      // NOTE: don't scale here (scale finally)
      // softmax.scale(p_sum);

      // Store the P matrix.
#if defined(STORE_S)
      softmax.store(gmem_s);
#endif

#if defined(STORE_P)
      gmem_p.move();
#endif

#if defined(STORE_S)
      gmem_s.move();
#endif

      if (kv_loop + Cta_tile_p::N < kv_loop_bound) {
        fmha::depbar<USE_LDGSTS, 1>();
        __syncthreads();

        smem_k.move_to_next_read_buffer();
        if (Kernel_traits::LIMIT_QK_FRAGMENTS) {
          smem_k.load(frag_k[0], 0);
        } else {
#pragma unroll
          for (int ki = 0; ki < Mma_tile_p::MMAS_K; ++ki) {
            smem_k.load(frag_k[ki], ki);
          }
        }
      }

      // Repack for the next BMM.
      fmha::Fragment_a<Traits_p, fmha::Row> frag_p[Mma_tile_o::MMAS_K][Mma_tile_o::MMAS_M];
      softmax.pack(frag_p);

      fmha::Fragment_accumulator<Traits_o> local_acc_o[Mma_tile_o::MMAS_M]
                                                      [Mma_tile_o::VALID_MMAS_N];
      using Acc_type_o = typename Traits_o::Accumulator_type;
      fmha::Clear_accumulator<Acc_type_o, Cta_tile_o::WARPS_K>::apply(local_acc_o);

// Do this part of O = P^T * V^T.
#pragma unroll
      for (int ki = 0; ki < Mma_tile_o::MMAS_K; ++ki) {
        fmha::gemm(local_acc_o, frag_p[ki], frag_v[ki]);
      }

      // Update acc_o of flash attention
      acc_o_updater.update_o(acc_o, local_acc_o);

      if (kv_loop + Cta_tile_p::N < kv_loop_bound) {
        // Commit the data for V to shared memory if it has not been done already.
        if (Kernel_traits::SHARE_SMEM_FOR_K_AND_V) {
          // Make sure we are done loading the fragments for K.
          __syncthreads();

          // Commit the data to shared memory for V.
          gmem_v.commit(smem_v);

          // Make sure the data is in shared memory.
          __syncthreads();
        }

        smem_v.move_next_read_buffer();

#pragma unroll
        for (int ki = 0; ki < Mma_tile_o::MMAS_K; ++ki) {
          smem_v.load(frag_v[ki], ki);
        }
      }
    }  // Inner loop over the key/value sequence length.

    // Trigger the load for the next Q, K, V values.
    if (q_loop + Cta_tile_p::M < q_loop_bound) {
      smem_q.move_to_next_write_buffer();
      gmem_q.move();
      gmem_q.load(smem_q);
      smem_k.move_to_next_write_buffer();
      smem_v.move_to_next_write_buffer();
      gmem_k.reset();
      gmem_v.reset();
      if (!USE_LDGSTS_KV) {
        gmem_k.load(smem_k);
        gmem_v.load(smem_v);
      }
    }

    // Make sure we have the LDGDEPBAR in place.
    fmha::ldgdepbar<USE_LDGSTS>();

// Loop over MMAS_M.
// NOTE: o and might shared the same buffer with kv
#pragma unroll
    for (int ii = 0; ii < Gmem_tile_o::LOOPS; ++ii) {
      // Swizzle the elements and do the final reduction.
      smem_o.store(acc_o, ii);

      // Make sure the data is in shared memory.
      __syncthreads();

      // Load from shared memory.
      uint4 out[Gmem_tile_o::STGS_PER_LOOP];
      smem_o.load(out);

      // Make sure the data was read from shared memory.
      __syncthreads();

      // Output the values.
      gmem_o.store(out, ii);
    }

    // Move to the next part of the output.
    gmem_o.move();

    // Trigger the load for the next Q, K, V values.
    if (q_loop + Cta_tile_p::M < q_loop_bound) {
      if (USE_LDGSTS_KV) {
        gmem_k.load(smem_k);
        gmem_v.load(smem_v);
      }

      // Make sure we have the LDGDEPBAR in place.
      fmha::ldgdepbar<USE_LDGSTS>();

      // Commit the data for Q, K, V to shared memory.
      gmem_q.commit(smem_q);

      gmem_k.commit(smem_k);

      if (!Kernel_traits::SHARE_SMEM_FOR_K_AND_V) {
        gmem_v.commit(smem_v);
      }
    }

    // Make sure the data is in shared memory.
    fmha::depbar<USE_LDGSTS, 1>();
    __syncthreads();

    // Make sure we are reading from the correct buffer.
    if (USE_LDGSTS_Q) {
      smem_q.move_to_next_read_buffer();
    }

// Trigger the loads for the values of Q for the next iteration.
#pragma unroll
    for (int ki = 0; ki < Mma_tile_p::MMAS_K; ++ki) {
      smem_q.load(frag_q[ki], ki);
    }

    if (USE_LDGSTS_K) {
      smem_k.move_to_next_read_buffer();
    }

    // Trigger the loads for the values of K, V for the next iteration.
    if (Kernel_traits::LIMIT_QK_FRAGMENTS) {
      smem_k.load(frag_k[0], 0);
    } else {
#pragma unroll
      for (int ki = 0; ki < Mma_tile_p::MMAS_K; ++ki) {
        smem_k.load(frag_k[ki], ki);
      }
    }

    // Commit the data for V to shared memory if it has not been done already.
    if (Kernel_traits::SHARE_SMEM_FOR_K_AND_V) {
      // Make sure we are done loading the fragments for K.
      __syncthreads();

      // Commit the data to shared memory for V.
      gmem_v.commit(smem_v);

      // Make sure the data is in shared memory.
      __syncthreads();
    }

    if (USE_LDGSTS_V) {
      smem_v.move_to_next_read_buffer();
    }

#pragma unroll
    for (int ki = 0; ki < Mma_tile_o::MMAS_K; ++ki) {
      smem_v.load(frag_v[ki], ki);
    }

  }  // Outer loop over the query sequence length.

}  // device_flash_attention_1xN

////////////////////////////////////////////////////////////////////////////////////////////////////

}  // namespace fused_multihead_attention