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


$UNROLL = CHANNELS_BATCH >> 3
void xnn_f16_f32acc_rdsum_ukernel_${ACCUMULATORS}p${ACCUMULATORS}x__f16c_c${CHANNELS_BATCH}(
    size_t rows,
    size_t channels,
    const xnn_float16* input,
    size_t input_stride,
    const xnn_float16* zero,
    float* output,
    const struct xnn_f16_f32acc_scale_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  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_BATCH}; channels -= ${CHANNELS_BATCH}) {
    const uint16_t* i0 = (const uint16_t*) input;
    $for ACC in range(1, ACCUMULATORS):
      const uint16_t* i${ACC} = (const uint16_t*) ((uintptr_t) input + ${ACC} * input_stride);

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

    for (int r = rows; r > 0; r -= ${ACCUMULATORS}) {
      $for ACC in range(1, ACCUMULATORS, 2):
        if XNN_UNPREDICTABLE(r < ${ACC+1}) {
          i${ACC} = (const uint16_t*) zero;
        }
        if XNN_UNPREDICTABLE(r <= ${ACC+1}) {
          i${ACC+1} = (const uint16_t*) zero;
        }
      $for c in range(UNROLL):
        __m256 vin${c};
      $for j in range(ACCUMULATORS):
        $for c in range(UNROLL):
          vin${c} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (&i${j}[${c*8}])));
        $for c in range(UNROLL):
          vacc${c} = _mm256_add_ps(vin${c}, vacc${c});
      $for ACC in range(0, ACCUMULATORS):
        i${ACC} = (const uint16_t*) ((uintptr_t) i${ACC} + 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 = (const void*) ((uintptr_t) o + 8 * sizeof(float));
    $for i in range(0, UNROLL):
      vacc${i} = _mm256_add_ps(vo${i}, vacc${i});
    $for i in range(0, UNROLL):
      _mm256_storeu_ps(output, vacc${i}); output = (void*) ((uintptr_t) output + 8 * sizeof(float));

    input = (const xnn_float16*) ((uintptr_t) input + ${CHANNELS_BATCH} * sizeof(xnn_float16));
  }
  if (channels != 0) {
    input_increment = ${ACCUMULATORS} * input_stride;
    const uint16_t* i0 = (const uint16_t*) input;
    $for ACC in range(1, ACCUMULATORS):
      const uint16_t* i${ACC} = (const uint16_t*) ((uintptr_t) input + ${ACC} * input_stride);
    __m256 vacc[${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 ACC in range(1, ACCUMULATORS, 2):
        if XNN_UNPREDICTABLE(r < ${ACC+1}) {
          i${ACC} = (const uint16_t*) zero;
        }
        if XNN_UNPREDICTABLE(r <= ${ACC+1}) {
          i${ACC+1} = (const uint16_t*) zero;
        }
      for (int i = 0; i < num_full_chunks; ++i) {
        $for c in range(ACCUMULATORS):
          vacc[i] = _mm256_add_ps(_mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) &i${c}[i*8])), vacc[i]);
      }

      if (remainder) {
        $for c in range(ACCUMULATORS):
          vacc[num_full_chunks] = _mm256_add_ps(vacc[num_full_chunks],  _mm256_cvtph_ps(_mm_loadu_si128((const  __m128i*) &i${c}[num_full_chunks*8])));
      }
      $for ACC in range(ACCUMULATORS):
        i${ACC} = (const uint16_t*) ((uintptr_t) i${ACC} + 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 < num_full_chunks; ++i) {
      vo[i] = _mm256_loadu_ps(o); o = (const void*) ((uintptr_t) o + 8 * sizeof(float));
    }
    for (int i = 0; i < num_full_chunks; ++i) {
      vacc[i] = _mm256_add_ps(vo[i], vacc[i]);
    }
    for (int i = 0; i < num_full_chunks; ++i) {
      _mm256_storeu_ps(output, vacc[i]); output = (void*) ((uintptr_t) output + 8 * sizeof(float));
    }
    if (remainder) {
      __m256 vout = vacc[num_full_chunks];
      __m128 vout_low = _mm256_castps256_ps128(vout);
      if (channels & 4) {
        __m128 vo =  _mm_loadu_ps(output);
        vo = _mm_add_ps(vout_low, vo);
        _mm_storeu_ps(output, vo);
        vout_low  = _mm256_castps256_ps128(_mm256_permute2f128_ps(vout, vout, 1));
        output = (void*) ((uintptr_t) output + 4 * sizeof(float));
      }
      if (channels & 2) {
        __m128 vo =  _mm_castsi128_ps(_mm_loadl_epi64((__m128i*) output));
        vo = _mm_add_ps(vout_low, vo);
        _mm_storel_pi((__m64*) output, vo);
        vout_low = _mm_movehl_ps(vout_low, vout_low);
        output = (void*) ((uintptr_t) output + 2 * sizeof(float));
      }
      if (channels & 1) {
        __m128 vo = _mm_castsi128_ps(_mm_cvtsi32_si128(unaligned_load_s32(output)));
        vo = _mm_add_ps(vout_low, vo);
        _mm_store_ss(output, vo);
      }
    }
  }
}