#include "cuda_utils.cuh"
#include <cmath>
#include <stdint.h>

const int BLOCK_SIZE = 1024;

// TODO: Maybe add some fast_sum_f16_f32 variant that not only accumulate in f32
// but also expect a f32 output so that this can be used for normalization e.g.
// in softmax.

// Fast reduce sum kernel, this assumes that the dimensions to loop over are at
// the end, each block is responsible for populating one value in the output
// array. There are at most 1024 threads per block.
template <typename T>
__device__ void
fast_sum(const size_t src_numel, const size_t el_to_sum_per_block,
         const size_t num_dims, const size_t *info, const T *src, T *dst) {
  const size_t *dims = info;
  const size_t *strides = info + num_dims;

  __shared__ T shr[BLOCK_SIZE];
  size_t tid = threadIdx.x;
  size_t dst_id = blockIdx.x;

  shr[tid] = 0;
  // Elements summed in this block range from dst_id * el_to_sum_per_block
  // to (dst_id + 1) * el_to_sum_per_block.
  size_t start_idx = dst_id * el_to_sum_per_block;
  size_t stop_idx = min(start_idx + el_to_sum_per_block, src_numel);
  size_t idx = start_idx + tid;

  while (idx < stop_idx) {
    // TODO: Fast version for the contiguous case.
    size_t strided_i = get_strided_index(idx, num_dims, dims, strides);
    shr[tid] += src[strided_i];
    idx += blockDim.x;
  }

  // Parallel reduction, see the slides:
  // https://www.olcf.ornl.gov/wp-content/uploads/2019/12/05_Atomics_Reductions_Warp_Shuffle.pdf
  // https://stackoverflow.com/questions/66078814/is-cuda-atomicadd-operation-faster-than-launch-another-kernel-when-we-do-reduce
  for (int s = blockDim.x / 2; s > 0; s >>= 1) {
    __syncthreads();
    if (tid < s)
      shr[tid] += shr[tid + s];
  }

  if (tid == 0)
    dst[dst_id] = shr[0];
}

template <typename T>
__device__ void
fast_max(const size_t src_numel, const size_t el_to_sum_per_block,
         const size_t num_dims, const size_t *info, const T *src, T *dst) {
  const size_t *dims = info;
  const size_t *strides = info + num_dims;

  __shared__ T shr[BLOCK_SIZE];
  size_t tid = threadIdx.x;
  size_t dst_id = blockIdx.x;

  shr[tid] = -INFINITY;
  // Elements summed in this block range from dst_id * el_to_sum_per_block
  // to (dst_id + 1) * el_to_sum_per_block.
  size_t start_idx = dst_id * el_to_sum_per_block;
  size_t stop_idx = min(start_idx + el_to_sum_per_block, src_numel);
  size_t idx = start_idx + tid;

  while (idx < stop_idx) {
    // TODO: Fast version for the contiguous case.
    size_t strided_i = get_strided_index(idx, num_dims, dims, strides);
    shr[tid] = maxg(shr[tid], src[strided_i]);
    idx += blockDim.x;
  }

  // Parallel reduction, see the slides:
  // https://www.olcf.ornl.gov/wp-content/uploads/2019/12/05_Atomics_Reductions_Warp_Shuffle.pdf
  // https://stackoverflow.com/questions/66078814/is-cuda-atomicadd-operation-faster-than-launch-another-kernel-when-we-do-reduce
  for (int s = blockDim.x / 2; s > 0; s >>= 1) {
    __syncthreads();
    if (tid < s)
      shr[tid] = maxg(shr[tid], shr[tid + s]);
  }

  if (tid == 0)
    dst[dst_id] = shr[0];
}

template <typename T>
__device__ void
fast_min(const size_t src_numel, const size_t el_to_sum_per_block,
         const size_t num_dims, const size_t *info, const T *src, T *dst) {
  const size_t *dims = info;
  const size_t *strides = info + num_dims;

  __shared__ T shr[BLOCK_SIZE];
  size_t tid = threadIdx.x;
  size_t dst_id = blockIdx.x;

  shr[tid] = INFINITY;
  // Elements summed in this block range from dst_id * el_to_sum_per_block
  // to (dst_id + 1) * el_to_sum_per_block.
  size_t start_idx = dst_id * el_to_sum_per_block;
  size_t stop_idx = min(start_idx + el_to_sum_per_block, src_numel);
  size_t idx = start_idx + tid;

  while (idx < stop_idx) {
    // TODO: Fast version for the contiguous case.
    size_t strided_i = get_strided_index(idx, num_dims, dims, strides);
    shr[tid] = ming(shr[tid], src[strided_i]);
    idx += blockDim.x;
  }

  // Parallel reduction, see the slides:
  // https://www.olcf.ornl.gov/wp-content/uploads/2019/12/05_Atomics_Reductions_Warp_Shuffle.pdf
  // https://stackoverflow.com/questions/66078814/is-cuda-atomicadd-operation-faster-than-launch-another-kernel-when-we-do-reduce
  for (int s = blockDim.x / 2; s > 0; s >>= 1) {
    __syncthreads();
    if (tid < s)
      shr[tid] = ming(shr[tid], shr[tid + s]);
  }

  if (tid == 0)
    dst[dst_id] = shr[0];
}

template <typename T>
__device__ void
fast_argmin(const size_t src_numel, const size_t el_to_sum_per_block,
         const size_t num_dims, const size_t *info, const T *src, uint32_t *dst) {
  const size_t *dims = info;
  const size_t *strides = info + num_dims;

  __shared__ T shr[BLOCK_SIZE];
  __shared__ uint32_t shr_index[BLOCK_SIZE];
  size_t tid = threadIdx.x;
  size_t dst_id = blockIdx.x;

  // Not sure how that works on uint32_t and uint8_t but it seems to do ok.
  shr[tid] = INFINITY;
  shr_index[tid] = 0xFFFFFFFF;
  bool not_set = true;
  // Elements summed in this block range from dst_id * el_to_sum_per_block
  // to (dst_id + 1) * el_to_sum_per_block.
  size_t start_idx = dst_id * el_to_sum_per_block;
  size_t stop_idx = min(start_idx + el_to_sum_per_block, src_numel);
  size_t idx = start_idx + tid;

  while (idx < stop_idx) {
    // TODO: Fast version for the contiguous case.
    size_t strided_i = get_strided_index(idx, num_dims, dims, strides);
    if (not_set || src[strided_i] < shr[tid]) {
      shr[tid] = src[strided_i];
      // Assume that the reduction takes place over the last dimension which is contiguous.
      shr_index[tid] = idx % dims[num_dims - 1];
      not_set = false;
    }
    idx += blockDim.x;
  }

  // Parallel reduction, see the slides:
  // https://www.olcf.ornl.gov/wp-content/uploads/2019/12/05_Atomics_Reductions_Warp_Shuffle.pdf
  // https://stackoverflow.com/questions/66078814/is-cuda-atomicadd-operation-faster-than-launch-another-kernel-when-we-do-reduce
  for (int s = blockDim.x / 2; s > 0; s >>= 1) {
    __syncthreads();
    if (tid < s && shr[tid + s] < shr[tid]) {
      shr[tid] = shr[tid + s];
      shr_index[tid] = shr_index[tid + s];
    }
  }

  if (tid == 0)
    dst[dst_id] = shr_index[0];
}

template <typename T>
__device__ void
fast_argmax(const size_t src_numel, const size_t el_to_sum_per_block,
         const size_t num_dims, const size_t *info, const T *src, uint32_t *dst) {
  const size_t *dims = info;
  const size_t *strides = info + num_dims;

  __shared__ T shr[BLOCK_SIZE];
  __shared__ uint32_t shr_index[BLOCK_SIZE];
  size_t tid = threadIdx.x;
  size_t dst_id = blockIdx.x;

  shr[tid] = -INFINITY;
  shr_index[tid] = 0xFFFFFFFF;
  bool not_set = true;
  // Elements summed in this block range from dst_id * el_to_sum_per_block
  // to (dst_id + 1) * el_to_sum_per_block.
  size_t start_idx = dst_id * el_to_sum_per_block;
  size_t stop_idx = min(start_idx + el_to_sum_per_block, src_numel);
  size_t idx = start_idx + tid;

  while (idx < stop_idx) {
    // TODO: Fast version for the contiguous case.
    size_t strided_i = get_strided_index(idx, num_dims, dims, strides);
    if (not_set || src[strided_i] > shr[tid]) {
      shr[tid] = src[strided_i];
      // Assume that the reduction takes place over the last dimension which is contiguous.
      shr_index[tid] = idx % dims[num_dims - 1];
      not_set = false;
    }
    idx += blockDim.x;
  }

  // Parallel reduction, see the slides:
  // https://www.olcf.ornl.gov/wp-content/uploads/2019/12/05_Atomics_Reductions_Warp_Shuffle.pdf
  // https://stackoverflow.com/questions/66078814/is-cuda-atomicadd-operation-faster-than-launch-another-kernel-when-we-do-reduce
  for (int s = blockDim.x / 2; s > 0; s >>= 1) {
    __syncthreads();
    if (tid < s && shr[tid + s] > shr[tid]) {
      shr[tid] = shr[tid + s];
      shr_index[tid] = shr_index[tid + s];
    }
  }

  if (tid == 0)
    dst[dst_id] = shr_index[0];
}

#define FAST_OP(TYPENAME, MIN_NAME, MAX_NAME, ARGMIN_NAME, ARGMAX_NAME, SUM_NAME) \
  extern "C" __global__ void ARGMIN_NAME(                                      \
      const size_t src_numel, const size_t el_to_sum_per_block,                \
      const size_t num_dims, const size_t *info, const TYPENAME *src,          \
      uint32_t *dst) {                                                         \
    fast_argmin(src_numel, el_to_sum_per_block, num_dims, info, src, dst);     \
  }                                                                            \
  extern "C" __global__ void ARGMAX_NAME(                                     \
      const size_t src_numel, const size_t el_to_sum_per_block,                \
      const size_t num_dims, const size_t *info, const TYPENAME *src,          \
      uint32_t *dst) {                                                         \
    fast_argmax(src_numel, el_to_sum_per_block, num_dims, info, src, dst);     \
  }                                                                            \
  extern "C" __global__ void MIN_NAME(                                         \
      const size_t src_numel, const size_t el_to_sum_per_block,                \
      const size_t num_dims, const size_t *info, const TYPENAME *src,          \
      TYPENAME *dst) {                                                         \
    fast_min(src_numel, el_to_sum_per_block, num_dims, info, src, dst);        \
  }                                                                            \
  extern "C" __global__ void MAX_NAME(                                         \
      const size_t src_numel, const size_t el_to_sum_per_block,                \
      const size_t num_dims, const size_t *info, const TYPENAME *src,          \
      TYPENAME *dst) {                                                         \
    fast_max(src_numel, el_to_sum_per_block, num_dims, info, src, dst);        \
  }                                                                            \
  extern "C" __global__ void SUM_NAME(                                         \
      const size_t src_numel, const size_t el_to_sum_per_block,                \
      const size_t num_dims, const size_t *info, const TYPENAME *src,          \
      TYPENAME *dst) {                                                         \
    fast_sum(src_numel, el_to_sum_per_block, num_dims, info, src, dst);        \
  }

#define SUM_OP(TYPENAME, FN_NAME)                                              \
  extern "C" __global__ void FN_NAME(                                          \
      const size_t numel, const size_t num_dims, const size_t num_sum_dims,    \
      const size_t *info, const TYPENAME *inp, TYPENAME *out) {                \
    const size_t *dims = info;                                                 \
    const size_t *strides = info + num_dims;                                   \
    const size_t *sum_dims_l = info + 2 * num_dims;                            \
    const size_t *sum_dims_s = info + 2 * num_dims + num_sum_dims;             \
    if (is_contiguous(num_dims, dims, strides)) {                              \
      for (unsigned int i = blockIdx.x * blockDim.x + threadIdx.x; i < numel;  \
           i += blockDim.x * gridDim.x) {                                      \
        size_t dst_index = i;                                                  \
        for (unsigned int nd = 0; nd < num_sum_dims; ++nd) {                   \
          size_t stride = sum_dims_s[nd];                                      \
          size_t pre = dst_index / stride;                                     \
          size_t post = dst_index % stride;                                    \
          dst_index = (pre / sum_dims_l[nd]) * stride + post;                  \
        }                                                                      \
        atomicAdd(out + dst_index, inp[i]);                                    \
      }                                                                        \
    } else {                                                                   \
      for (unsigned int i = blockIdx.x * blockDim.x + threadIdx.x; i < numel;  \
           i += blockDim.x * gridDim.x) {                                      \
        unsigned strided_i = get_strided_index(i, num_dims, dims, strides);    \
        size_t dst_index = i;                                                  \
        for (unsigned int nd = 0; nd < num_sum_dims; ++nd) {                   \
          size_t stride = sum_dims_s[nd];                                      \
          size_t pre = dst_index / stride;                                     \
          size_t post = dst_index % stride;                                    \
          dst_index = (pre / sum_dims_l[nd]) * stride + post;                  \
        }                                                                      \
        atomicAdd(out + dst_index, inp[strided_i]);                            \
      }                                                                        \
    }                                                                          \
  }

#if __CUDA_ARCH__ >= 800
SUM_OP(__nv_bfloat16, sum_bf16)
FAST_OP(__nv_bfloat16, fast_min_bf16, fast_max_bf16, fast_argmin_bf16, fast_argmax_bf16, fast_sum_bf16)
#endif

#if __CUDA_ARCH__ >= 530
SUM_OP(__half, sum_f16)
FAST_OP(__half, fast_min_f16, fast_max_f16, fast_argmin_f16, fast_argmax_f16, fast_sum_f16)
#endif

SUM_OP(float, sum_f32)
SUM_OP(double, sum_f64)
SUM_OP(uint32_t, sum_u32)

FAST_OP(float, fast_min_f32, fast_max_f32, fast_argmin_f32, fast_argmax_f32, fast_sum_f32)
FAST_OP(double, fast_min_f64, fast_max_f64, fast_argmin_f64, fast_argmax_f64, fast_sum_f64)
FAST_OP(uint32_t, fast_min_u32, fast_max_u32, fast_argmin_u32, fast_argmax_u32, fast_sum_u32)
FAST_OP(int64_t, fast_min_i64, fast_max_i64, fast_argmin_i64, fast_argmax_i64, fast_sum_i64)
FAST_OP(uint8_t, fast_min_u8, fast_max_u8, fast_argmin_u8, fast_argmax_u8, fast_sum_u8)