// 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.
$assert CHANNEL_TILE % 32 == 0
$assert CHANNEL_TILE >= 32
$SIMD_TILE = CHANNEL_TILE // 32
$assert ACCUMULATORS <= SIMD_TILE
$assert AVX in [2, 10]

#include <assert.h>

#include <immintrin.h>

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

$ACC_SUFFIX = "" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS
$ISA = "avx2" if AVX == 2 else "avx256skx"
void xnn_qs8_rsum_ukernel__${ISA}_u${CHANNEL_TILE}${ACC_SUFFIX}(
    size_t batch,
    const int8_t* input,
    int32_t* output,
    const struct xnn_qs8_rsum_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(input != NULL);
  assert(output != NULL);
  assert(params != NULL);

  XNN_ALIGN(32) static const int8_t onemask_table[64] = {
    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
  };

  const __m256i vone = _mm256_load_si256((const __m256i*) &onemask_table[0]);
  const __m256i vone_16 = _mm256_srli_epi16(vone, 8);
  $for ACC in range(ACCUMULATORS):
    __m256i vacc${ACC} = _mm256_setzero_si256();

  // 256 int8s may be summed into an int16 before overflowing.
  // Each register has 32 lanes and there are ${ACCUMULATORS} accumulators so batch size is ${ACCUMULATORS*256*32}
  for (; batch >= ${ACCUMULATORS*256*32}; batch -= ${ACCUMULATORS*256*32}) {
    $for ACC in range(ACCUMULATORS):
      __m256i vacc16_${ACC} = _mm256_setzero_si256();
    for (size_t current_batch = ${ACCUMULATORS*256*32}; current_batch > 0; current_batch -= ${CHANNEL_TILE}) {
      $for N in range(SIMD_TILE):
        const __m256i vt${N} = _mm256_maddubs_epi16(vone, _mm256_loadu_si256((const __m256i*) input)); input += 32;
      $for N in range(SIMD_TILE):
        vacc16_${N % ACCUMULATORS} = _mm256_add_epi16(vacc16_${N % ACCUMULATORS}, vt${N});
    }
    $for ACC in range(ACCUMULATORS):
      vacc${ACC} = _mm256_add_epi32(vacc${ACC}, _mm256_madd_epi16(vone_16, vacc16_${ACC}));
  }

  $if CHANNEL_TILE > 32:
    if (XNN_LIKELY(batch >= ${CHANNEL_TILE})) {
      assert(batch >= 1 && batch < ${ACCUMULATORS*256*32});
      $for ACC in range(ACCUMULATORS):
        __m256i vacc16_${ACC} = _mm256_setzero_si256();
      for (; batch >= ${CHANNEL_TILE}; batch -= ${CHANNEL_TILE}) {
        $for N in range(SIMD_TILE):
          const __m256i vt${N} = _mm256_maddubs_epi16(vone, _mm256_loadu_si256((const __m256i*) input)); input += 32;
        $for N in range(SIMD_TILE):
          vacc16_${N % ACCUMULATORS} = _mm256_add_epi16(vacc16_${N % ACCUMULATORS}, vt${N});
      }
      $for ACC in range(ACCUMULATORS):
        vacc${ACC} = _mm256_add_epi32(vacc${ACC}, _mm256_madd_epi16(vone_16, vacc16_${ACC}));
    }
  if (XNN_UNLIKELY(batch != 0)) {
    assert(batch >= 1 && batch < ${256*32});
    __m256i vacc16 = _mm256_setzero_si256();
    for (; batch >= 32; batch -= 32) {
      const __m256i vt = _mm256_maddubs_epi16(vone, _mm256_loadu_si256((const __m256i*) input)); input += 32;
      vacc16 = _mm256_add_epi16(vacc16, vt);
    }
    $if AVX == 2:
      // Remainder is between 17 and 31 bytes, so process 32 bytes (overread of up to 15)
      if (XNN_UNLIKELY(batch >= 17)) {
        assert(batch >= 17 && batch <= 31);
        const __m256i vonemask = _mm256_loadu_si256((const __m256i*) &onemask_table[32 - batch]);
        const __m256i vt = _mm256_maddubs_epi16(vonemask, _mm256_loadu_si256((const __m256i*) input));
        vacc16 = _mm256_add_epi16(vacc16, vt);
      // Remainder is between 1 and 16 bytes, so process 16 bytes (overread of up to 15)
      } else if (XNN_UNLIKELY(batch != 0)) {
        assert(batch >= 1 && batch <= 16);
        const __m256i vonemask = _mm256_loadu_si256((const __m256i*) &onemask_table[32 - batch]);
        const __m256i vt = _mm256_maddubs_epi16(vonemask, _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*) input)));
        vacc16 = _mm256_add_epi16(vacc16, vt);
      }
    $else:
      if (XNN_UNLIKELY(batch != 0)) {
        assert(batch >= 1 && batch <= 31);
        // Prepare mask for valid 8-bit elements (depends on batch).
        const __mmask32 vmask = _cvtu32_mask32((UINT32_C(1) << batch) - 1);
        const __m256i vt = _mm256_maddubs_epi16(vone, _mm256_maskz_loadu_epi8(vmask, input));
        vacc16 = _mm256_add_epi16(vacc16, vt);
      }
    vacc0 = _mm256_add_epi32(vacc0, _mm256_madd_epi16(vone_16, vacc16));
  }
  $if ACCUMULATORS > 1:
    $ACC_SLICE = 1
    $while ACC_SLICE < ACCUMULATORS:
      $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
        $if A + ACC_SLICE < ACCUMULATORS:
          vacc${A} = _mm256_add_epi32(vacc${A}, vacc${A + ACC_SLICE});
      $ACC_SLICE *= 2

  __m128i vacc_lo = _mm_add_epi32(_mm256_castsi256_si128(vacc0), _mm256_extractf128_si256(vacc0, 1));
  vacc_lo = _mm_hadd_epi32(vacc_lo, vacc_lo);
  vacc_lo = _mm_hadd_epi32(vacc_lo, vacc_lo);

  *output += _mm_cvtsi128_si32(vacc_lo);
}