// Copyright 2024 Imagination Technologies, inc.
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

$assert LMUL in [1, 2, 4, 8]
#include <assert.h>

#include <riscv_vector.h>

#include "xnnpack/common.h"
#include "xnnpack/reduce.h"


void xnn_f32_rdsum_ukernel_${ACCUMULATORS}p${ACCUMULATORS}x__rvv_u${LMUL}v(
    size_t rows,
    size_t channels,
    const float* input,
    size_t input_stride,
    const float* zero,
    float* output,
    const struct xnn_f32_scale_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(rows != 0);
  assert(channels != 0);
  assert(input != NULL);
  assert(output != NULL);

  const float scale = params->scalar.scale;

  size_t input_increment = ${ACCUMULATORS} * input_stride;
  for (; channels > 0; ) {
    size_t n = __riscv_vsetvl_e32m${LMUL}(channels); channels -= n;
    const float* i0 = input;
    $for i in range(1, ACCUMULATORS):
      const float* i${i} = (const float*) ((uintptr_t) i${i - 1} + input_stride);
    vfloat32m${LMUL}_t acc_f32v = __riscv_vfmv_v_f_f32m${LMUL}(0.f, n);

    for (int r = rows; r > 0; r -= ${ACCUMULATORS}) {
      $for N in range(1, ACCUMULATORS, 2):
        if XNN_UNPREDICTABLE(r < ${N+1}) {
          i${N} = zero;
        }
        if XNN_UNPREDICTABLE(r <= ${N+1}) {
          i${N+1} = zero;
        }

      $for N in range(0, ACCUMULATORS):
        vfloat32m${LMUL}_t in${N}_f32v = __riscv_vle32_v_f32m${LMUL}(i${N}, n);
        acc_f32v = __riscv_vfadd_vv_f32m${LMUL}(acc_f32v, in${N}_f32v, n);

      $for N in range(0, ACCUMULATORS):
        i${N} = (const float*) ((uintptr_t) i${N} + input_increment);
    }
    acc_f32v = __riscv_vfmul_vf_f32m${LMUL}(acc_f32v, scale, n);
    vfloat32m${LMUL}_t out_f32v = __riscv_vle32_v_f32m${LMUL}(output, n);
    out_f32v = __riscv_vfadd_vv_f32m${LMUL}(out_f32v, acc_f32v, n);
    __riscv_vse32_v_f32m${LMUL}(output, out_f32v, n); output += n;

    input = (const float*) ((uintptr_t) input + n * sizeof(float));
  }
}