/*
 * 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_2x2(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 = 0 };

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

  // 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;

  // Block info to determine if we stop here.
  Single_cta<Kernel_traits::VERSION> const binfo(params, bidb, bidh, 0, tidx);
  if (binfo.stop_early()) {
    return;
  }

  // The structure to hold the mask.
  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);
  // 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 O.
  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_[0], tidx);

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

  // 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);

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

  // Load the fragments for Q.
  typename Smem_tile_q::Fragment frag_q[2][Mma_tile_p::MMAS_M];
  smem_q.load(frag_q[0], 0);

  // Load the fragments for K.
  typename Smem_tile_k::Fragment frag_k[2][Mma_tile_p::MMAS_N];
  smem_k.load(frag_k[0], 0);

  // Declare the accumulators for the 1st gemm.
  fmha::Fragment_accumulator<Traits_p> acc_p[Mma_tile_p::MMAS_M][Mma_tile_p::MMAS_N];
  fmha::Clear_accumulator<typename Traits_p::Accumulator_type, 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) {
    // Trigger the load from shared memory for the next series of Q/K values.
    smem_q.load(frag_q[ki & 1], ki);
    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) & 1], frag_k[(ki - 1) & 1]);
    }
  }

  // Do the final stage of math.
  if (Mma_tile_p::MMAS_K <= Mma_tile_p::VALID_MMAS_K) {
    int ki = Mma_tile_p::MMAS_K;
    fmha::gemm(acc_p, frag_q[(ki - 1) & 1], frag_k[(ki - 1) & 1]);
  }

  // Store the P matrix.
#if defined(STORE_P)
  enum { BITS_PER_ELT_P = sizeof(typename Traits_p::Accumulator_type) * 8 };

  using Gmem_tile_p = fmha::Gmem_tile_ps<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);
  gmem_p.store(acc_p);
#endif

  // Make sure the shared memory was consumed.
  __syncthreads();

  // Allocate the global memory tile loader for V.
  Gmem_tile_v gmem_v(params, 2, binfo, tidx);
  // Allocate the shared memory tile loader for V. We use the same as K so be careful!!!
  Smem_tile_v smem_v(smem_, tidx);
  // Trigger the loads for V.
  gmem_v.load(smem_v);

  // Load the mask.
  mask.load(0);

  // Create the object to do the softmax.
  using Softmax = fmha::Softmax<Traits_p, Cta_tile_p, Kernel_traits>;
  Softmax softmax(params, &smem_[0], bidb, tidx);

  // 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);
  }

  // Enable our trick to use the max for INT8 to scale.
  if (params.use_int8_scale_max) {
    // 16129 == 127 ^ 2.
    float p_max = reinterpret_cast<float const&>(params.scale_bmm1) * 16129.f;
    softmax.apply_exp(p_max);
  } else {
    float p_max[Softmax::ROWS_PER_THREAD];
    softmax.template reduce<fmha::Max_>(p_max);

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

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

  // Commit the data to shared memory for V. It must happen after the last "reduce".
  gmem_v.commit(smem_v);

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

  // Finalize softmax on the accumulators of P^T.
  softmax.scale(p_sum);

  // Store the P matrix.
#if defined(STORE_S)
  enum { BITS_PER_ELT_S = sizeof(typename Traits_p::A_type) * 8 };

  using Gmem_tile_s = fmha::Gmem_tile_ps<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);
  softmax.store(gmem_s);
#endif

  // Repack the transformed P elements to fragments for the next GEMM.
  fmha::Fragment_a<Traits_p, fmha::Row> frag_p[Mma_tile_o::MMAS_K][Mma_tile_o::MMAS_M];
  softmax.pack(frag_p);

  // 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];
  fmha::Clear_accumulator<typename Traits_o::Accumulator_type, Cta_tile_o::WARPS_K>::apply(acc_o);

  // Load the fragments for K. We keep the data in registers during the entire kernel.
  typename Smem_tile_v::Fragment frag_v[2][Mma_tile_o::VALID_MMAS_N];
  smem_v.load(frag_v[0], 0);

// Do this part of O = P^T * V^T.
#pragma unroll
  for (int ki = 1; ki < Mma_tile_o::MMAS_K; ++ki) {
    // Trigger the load from shared memory for the next series of Q/K values.
    smem_v.load(frag_v[ki & 1], ki);
    // Do the math.
    fmha::gemm(acc_o, frag_p[(ki - 1)], frag_v[(ki - 1) & 1]);
  }

  // Do the final stage of math.
  {
    int ki = Mma_tile_o::MMAS_K;
    fmha::gemm(acc_o, frag_p[(ki - 1)], frag_v[(ki - 1) & 1]);
  }

// // DEBUG.
// printf("tidx=%3d acc_o[0][0]=0x%08x\n", tidx, acc_o[0][0].reg(0));
// printf("tidx=%3d acc_o[1][0]=0x%08x\n", tidx, acc_o[1][0].reg(0));
// printf("tidx=%3d acc_o[2][0]=0x%08x\n", tidx, acc_o[2][0].reg(0));
// printf("tidx=%3d acc_o[3][0]=0x%08x\n", tidx, acc_o[3][0].reg(0));
// // END OF DEBUG.

// Loop over MMAS_M.
#pragma unroll
  for (int ii = 0; ii < Gmem_tile_o::LOOPS; ++ii) {
    // Make sure the data was read from shared memory.
    __syncthreads();

    // 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.
    static_assert(Gmem_tile_o::STGS_PER_LOOP == Smem_tile_o::LDS_PER_LOOP, "");
    uint4 out[Gmem_tile_o::STGS_PER_LOOP];
    smem_o.load(out);

    // // DEBUG.
    // #pragma unroll
    // for( int jj = 0; jj < Gmem_tile_o::STGS_PER_LOOP; ++jj ) {
    //     printf("tidx=%3d loop=%d out[%d]=0x%08x\n", tidx, ii, jj, out[jj].x);
    // }
    // // END OF DEBUG.

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

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

}  // namespace fused_multihead_attention