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

$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <immintrin.h>

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


$UNROLL = CHANNELS >> 3
void xnn_f32_rdsum_ukernel_${ACCUMULATORS}p${ACCUMULATORS}x__avx_c${CHANNELS}(
    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)])
{
  static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};

  assert(rows != 0);
  assert(channels != 0);
  assert(input != NULL);
  assert(output != NULL);

  const __m256 vscale = _mm256_set1_ps(params->scalar.scale);

  size_t input_increment = ${ACCUMULATORS} * input_stride;
  for (; channels >= ${CHANNELS}; channels -= ${CHANNELS}) {
    const float* i0 = input;
    $for i in range(1, ACCUMULATORS):
      const float* i${i} = (const float*) ((uintptr_t) input + ${i} * input_stride);

    $for i in range(UNROLL):
      __m256 vacc${i} = _mm256_setzero_ps();

    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;
        }
      $for c in range(UNROLL):
        __m256 vin${c};
      $for j in range(ACCUMULATORS):
        $for c in range(UNROLL):
          vin${c} = _mm256_loadu_ps(&i${j}[${c*8}]);
        $for c in range(UNROLL):
          vacc${c} = _mm256_add_ps(vin${c}, vacc${c});
      $for N in range(0, ACCUMULATORS):
        i${N} = (const float*) ((uintptr_t) i${N} + input_increment);
    }
    $for i in range(UNROLL):
      vacc${i} = _mm256_mul_ps(vacc${i}, vscale);

    const float* o = output;
    $for i in range(0, UNROLL):
      __m256 vo${i} = _mm256_loadu_ps(o); o += 8;
    $for i in range(0, UNROLL):
      vacc${i} = _mm256_add_ps(vo${i}, vacc${i});
    $for i in range(0, UNROLL):
      _mm256_storeu_ps(output, vacc${i}); output += 8;

    input = (const float*) ((uintptr_t) input + ${CHANNELS} * sizeof(float));
  }
  __m256i vmask;
  if (channels != 0) {
    input_increment = ${ACCUMULATORS} * input_stride;
    const float* i0 = input;
    $for i in range(1, ACCUMULATORS):
      const float* i${i} = (const float*) ((uintptr_t) input + ${i} * input_stride);
    __m256 vacc[${UNROLL}];
    $for i in range(UNROLL):
      vacc[${i}] = _mm256_setzero_ps();

    const size_t num_full_chunks = channels >> 3;
    const size_t num_chunks = round_up_po2(channels, 8) >> 3;
    const size_t remainder = channels & 0x7;
    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 (int i = 0; i < num_full_chunks; ++i) {
        $for c in range(ACCUMULATORS):
          vacc[i] = _mm256_add_ps(_mm256_loadu_ps(&i${c}[i*8]), vacc[i]);
      }

      if (remainder) {
        vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - (channels & 0x7) * sizeof(float)));
        $for c in range(ACCUMULATORS):
          vacc[num_full_chunks] = _mm256_add_ps(_mm256_maskload_ps(&i${c}[num_full_chunks*8], vmask), vacc[num_full_chunks]);
      }
      $for N in range(ACCUMULATORS):
        i${N} = (const float*) ((uintptr_t) i${N} + input_increment);
    }
    for (size_t i = 0; i < num_chunks; ++i) {
      vacc[i] = _mm256_mul_ps(vacc[i], vscale);
    }

    __m256 vo[${UNROLL}];
    const float* o = output;
    for (int i = 0; i < channels >> 3; ++i) {
      vo[i] = _mm256_loadu_ps(o); o += 8;
    }
    for (int i = 0; i < channels >> 3; ++i) {
      vacc[i] = _mm256_add_ps(vo[i], vacc[i]);
    }
    for (int i = 0; i < channels >> 3; ++i) {
      _mm256_storeu_ps(output, vacc[i]); output += 8;
    }
    if (remainder) {
      const size_t pos = num_full_chunks;
      __m256 vout = vacc[pos];
      const __m256 vdata = _mm256_maskload_ps(output, vmask);
      vout = _mm256_add_ps(vout, vdata);
      __m128 vout_lo = _mm256_castps256_ps128(vout);
      if (channels & 4) {
        _mm_storeu_ps(output, vout_lo);
        vout_lo = _mm256_extractf128_ps(vout, 1);
        output += 4;
      }
      if (channels & 2) {
        _mm_storel_pi((__m64*) output, vout_lo);
        vout_lo = _mm_movehl_ps(vout_lo, vout_lo);
        output += 2;
      }
      if (channels & 1) {
        _mm_store_ss(output, vout_lo);
      }
    }
  }
}