// Copyright 2022 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.

$CHANNEL_SUBTILE = 8
$assert CHANNEL_TILE % CHANNEL_SUBTILE == 0
$CHANNEL_ROUND = 4
$assert MIDDLE_PASS_TILE <= LAST_PASS_TILE
$assert FIRST_PASS_TILE >= 1
$assert MIDDLE_PASS_TILE >= 1
$assert LAST_PASS_TILE >= 1
$assert ACCUMULATORS >= 1
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>
#include <stddef.h>
#include <stdint.h>

#include <immintrin.h>

#include "xnnpack/dwconv.h"
#include "xnnpack/math.h"


$ISA = {0: "avx", 3: "fma3"}[FMA]
void xnn_f32_dwconv_minmax_ukernel_${FIRST_PASS_TILE}f${MIDDLE_PASS_TILE}m${LAST_PASS_TILE}l${CHANNEL_TILE}c${CHANNEL_SUBTILE}s${CHANNEL_ROUND}r__${ISA}${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
    size_t channels,
    size_t output_width,
    const float** input,
    const float* weights,
    float* output,
    intptr_t input_stride,
    size_t output_increment,
    size_t input_offset,
    const float* zero,
    size_t kernel_size,
    float* buffer,
    const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(channels != 0);
  assert(output_width != 0);
  assert(kernel_size > ${FIRST_PASS_TILE});

  static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};

  const __m256 vmin = _mm256_set1_ps(params->scalar.min);
  const __m256 vmax = _mm256_set1_ps(params->scalar.max);
  XNN_FORCE_REALIZATION(vmin);
  XNN_FORCE_REALIZATION(vmax);
  do {
    const float* w = weights;

    // First pass to process ${FIRST_PASS_TILE} inputs.
    {
      float* b = buffer;
      $for K in range(FIRST_PASS_TILE):
        const float* i${K} = input[${K}];
        assert(i${K} != NULL);
        if XNN_UNPREDICTABLE(i${K} != zero) {
          i${K} = (const float*) ((uintptr_t) i${K} + input_offset);
        }
      input += ${FIRST_PASS_TILE};

      // Process c channels and write to buffer.
      size_t c = round_up_po2(channels, ${CHANNEL_ROUND});
      $if CHANNEL_TILE > 8:
        for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
          $for C in range(0, CHANNEL_TILE, 8):
            $if C == 0:
              __m256 vacc${ABC[C:C+8]}p0 = _mm256_load_ps(w);
            $else:
              __m256 vacc${ABC[C:C+8]}p0 = _mm256_load_ps(w + ${C});

          $for K in range(FIRST_PASS_TILE):

            $for C in range(0, CHANNEL_TILE, 8):
              $if C == 0:
                const __m256 vi${K}x${ABC[C:C+8]} = _mm256_loadu_ps(i${K});
              $else:
                const __m256 vi${K}x${ABC[C:C+8]} = _mm256_loadu_ps(i${K} + ${C});
            i${K} += ${CHANNEL_TILE};

            $for C in range(0, CHANNEL_TILE, 8):
              const __m256 vk${K}x${ABC[C:C+8]} = _mm256_load_ps(w + ${(K + 1) * CHANNEL_TILE + C});
            $for C in range(0, CHANNEL_TILE, 8):
              $if 1 <= K < ACCUMULATORS:
                __m256 vacc${ABC[C:C+8]}p${K} = _mm256_mul_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]});
              $elif FMA == 3:
                vacc${ABC[C:C+8]}p${K % ACCUMULATORS} = _mm256_fmadd_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]}, vacc${ABC[C:C+8]}p${K % ACCUMULATORS});
              $else:
                vacc${ABC[C:C+8]}p${K % ACCUMULATORS} = _mm256_add_ps(vacc${ABC[C:C+8]}p${K % ACCUMULATORS}, _mm256_mul_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]}));

          w += ${(FIRST_PASS_TILE + 1) * CHANNEL_TILE};

          $if ACCUMULATORS > 1:
            // Add up all accumulators to vacc${ABC[0:8]}p0
            $ACC_SLICE = 1
            $while ACC_SLICE < ACCUMULATORS:
              $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
                $if A + ACC_SLICE < ACCUMULATORS:
                  $for C in range(0, CHANNEL_TILE, 8):
                    vacc${ABC[C:C+8]}p${A} = _mm256_add_ps(vacc${ABC[C:C+8]}p${A}, vacc${ABC[C:C+8]}p${A + ACC_SLICE});
              $ACC_SLICE *= 2

          $for C in range(0, CHANNEL_TILE, 8):
            $if C == 0:
              _mm256_store_ps(b, vacc${ABC[C:C+8]}p0);
            $else:
              _mm256_store_ps(b + ${C}, vacc${ABC[C:C+8]}p0);
          b += ${CHANNEL_TILE};
        }

      for (; c >= ${CHANNEL_SUBTILE}; c -= ${CHANNEL_SUBTILE}) {
        __m256 vacc01234567p0 = _mm256_load_ps(w);

        $for K in range(FIRST_PASS_TILE):

          const __m256 vi${K}x01234567 = _mm256_loadu_ps(i${K});
          i${K} += ${CHANNEL_SUBTILE};

          const __m256 vk${K}x01234567 = _mm256_load_ps(w + ${(K + 1) * CHANNEL_SUBTILE});
          $if 1 <= K < ACCUMULATORS:
            __m256 vacc01234567p${K} = _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567);
          $elif FMA == 3:
            vacc01234567p${K % ACCUMULATORS} = _mm256_fmadd_ps(vi${K}x01234567, vk${K}x01234567, vacc01234567p${K % ACCUMULATORS});
          $else:
            vacc01234567p${K % ACCUMULATORS} = _mm256_add_ps(vacc01234567p${K % ACCUMULATORS}, _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567));

        w += ${(FIRST_PASS_TILE + 1) * CHANNEL_SUBTILE};

        $if ACCUMULATORS > 1:
          // Add up all accumulators to vacc01234567p0
          $ACC_SLICE = 1
          $while ACC_SLICE < ACCUMULATORS:
            $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
              $if A + ACC_SLICE < ACCUMULATORS:
                vacc01234567p${A} = _mm256_add_ps(vacc01234567p${A}, vacc01234567p${A + ACC_SLICE});
            $ACC_SLICE *= 2

        _mm256_store_ps(b, vacc01234567p0);
        b += ${CHANNEL_SUBTILE};
      }

      if (c != 0) {
        assert(c >= 1);
        assert(c <= ${CHANNEL_SUBTILE-1});
        const __m256i vmask = _mm256_loadu_si256((const __m256i*) &mask_table[7 - c]);
        __m256 vacc01234567p0 = _mm256_load_ps(w);

        $for K in range(FIRST_PASS_TILE):

          const __m256 vi${K}x01234567 = _mm256_maskload_ps(i${K}, vmask);

          const __m256 vk${K}x01234567 = _mm256_load_ps(w + ${(K + 1) * CHANNEL_SUBTILE});
          $if 1 <= K < ACCUMULATORS:
            __m256 vacc01234567p${K} = _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567);
          $elif FMA == 3:
            vacc01234567p${K % ACCUMULATORS} = _mm256_fmadd_ps(vi${K}x01234567, vk${K}x01234567, vacc01234567p${K % ACCUMULATORS});
          $else:
            vacc01234567p${K % ACCUMULATORS} = _mm256_add_ps(vacc01234567p${K % ACCUMULATORS}, _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567));

        w += ${(FIRST_PASS_TILE + 1) * CHANNEL_SUBTILE};

        $if ACCUMULATORS > 1:
          // Add up all accumulators to vacc01234567p0
          $ACC_SLICE = 1
          $while ACC_SLICE < ACCUMULATORS:
            $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
              $if A + ACC_SLICE < ACCUMULATORS:
                vacc01234567p${A} = _mm256_add_ps(vacc01234567p${A}, vacc01234567p${A + ACC_SLICE});
            $ACC_SLICE *= 2

        _mm256_store_ps(b, vacc01234567p0);
      }
    }

    // Middle pass to process ${MIDDLE_PASS_TILE} inputs in each iteration.
    for (size_t ks = kernel_size - ${FIRST_PASS_TILE}; ks > ${LAST_PASS_TILE}; ks -= ${MIDDLE_PASS_TILE}) {
      float* b = buffer;
      $for K in range(MIDDLE_PASS_TILE):
        const float* i${K} = input[${K}];
        assert(i${K} != NULL);
        if XNN_UNPREDICTABLE(i${K} != zero) {
          i${K} = (const float*) ((uintptr_t) i${K} + input_offset);
        }
      input += ${MIDDLE_PASS_TILE};

      size_t c = round_up_po2(channels, ${CHANNEL_ROUND});
      $if CHANNEL_TILE > 8:
        for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
          $for C in range(0, CHANNEL_TILE, 8):
            $if C == 0:
              __m256 vacc${ABC[C:C+8]}p0 = _mm256_load_ps(b);
            $else:
              __m256 vacc${ABC[C:C+8]}p0 = _mm256_load_ps(b + ${C});

          $for K in range(MIDDLE_PASS_TILE):

            $for C in range(0, CHANNEL_TILE, 8):
              $if C == 0:
                const __m256 vi${K}x${ABC[C:C+8]} = _mm256_loadu_ps(i${K});
              $else:
                const __m256 vi${K}x${ABC[C:C+8]} = _mm256_loadu_ps(i${K} + ${C});
            i${K} += ${CHANNEL_TILE};

            $for C in range(0, CHANNEL_TILE, 8):
              $if K == 0 and C == 0:
                const __m256 vk${K}x${ABC[C:C+8]} = _mm256_load_ps(w);
              $else:
                const __m256 vk${K}x${ABC[C:C+8]} = _mm256_load_ps(w + ${K * CHANNEL_TILE + C});
            $for C in range(0, CHANNEL_TILE, 8):
              $if 1 <= K < ACCUMULATORS:
                __m256 vacc${ABC[C:C+8]}p${K} = _mm256_mul_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]});
              $elif FMA == 3:
                vacc${ABC[C:C+8]}p${K % ACCUMULATORS} = _mm256_fmadd_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]}, vacc${ABC[C:C+8]}p${K % ACCUMULATORS});
              $else:
                vacc${ABC[C:C+8]}p${K % ACCUMULATORS} = _mm256_add_ps(vacc${ABC[C:C+8]}p${K % ACCUMULATORS}, _mm256_mul_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]}));

          w += ${MIDDLE_PASS_TILE * CHANNEL_TILE};

          $if ACCUMULATORS > 1:
            // Add up all accumulators to vacc${ABC[0:8]}p0
            $ACC_SLICE = 1
            $while ACC_SLICE < ACCUMULATORS:
              $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
                $if A + ACC_SLICE < ACCUMULATORS:
                  $for C in range(0, CHANNEL_TILE, 8):
                    vacc${ABC[C:C+8]}p${A} = _mm256_add_ps(vacc${ABC[C:C+8]}p${A}, vacc${ABC[C:C+8]}p${A + ACC_SLICE});
              $ACC_SLICE *= 2

          $for C in range(0, CHANNEL_TILE, 8):
            $if C == 0:
              _mm256_store_ps(b, vacc${ABC[C:C+8]}p0);
            $else:
              _mm256_store_ps(b + ${C}, vacc${ABC[C:C+8]}p0);
          b += ${CHANNEL_TILE};
        }

      for (; c >= 8; c -= 8) {
        __m256 vacc01234567p0 = _mm256_load_ps(b);

        $for K in range(MIDDLE_PASS_TILE):

          const __m256 vi${K}x01234567 = _mm256_loadu_ps(i${K});
          i${K} += ${CHANNEL_SUBTILE};

          $if K == 0:
            const __m256 vk${K}x01234567 = _mm256_load_ps(w);
          $else:
            const __m256 vk${K}x01234567 = _mm256_load_ps(w + ${K * CHANNEL_SUBTILE});
          $if 1 <= K < ACCUMULATORS:
            __m256 vacc01234567p${K} = _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567);
          $elif FMA == 3:
            vacc01234567p${K % ACCUMULATORS} = _mm256_fmadd_ps(vi${K}x01234567, vk${K}x01234567, vacc01234567p${K % ACCUMULATORS});
          $else:
            vacc01234567p${K % ACCUMULATORS} = _mm256_add_ps(vacc01234567p${K % ACCUMULATORS}, _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567));

        w += ${MIDDLE_PASS_TILE * CHANNEL_SUBTILE};

        $if ACCUMULATORS > 1:
          // Add up all accumulators to vacc01234567p0
          $ACC_SLICE = 1
          $while ACC_SLICE < ACCUMULATORS:
            $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
              $if A + ACC_SLICE < ACCUMULATORS:
                vacc01234567p${A} = _mm256_add_ps(vacc01234567p${A}, vacc01234567p${A + ACC_SLICE});
            $ACC_SLICE *= 2

        _mm256_store_ps(b, vacc01234567p0);
        b += ${CHANNEL_SUBTILE};
      }

      if (c != 0) {
        assert(c >= 1);
        assert(c <= ${CHANNEL_SUBTILE-1});
        const __m256i vmask = _mm256_loadu_si256((const __m256i*) &mask_table[7 - c]);
        __m256 vacc01234567p0 = _mm256_load_ps(b);

        $for K in range(MIDDLE_PASS_TILE):

          const __m256 vi${K}x01234567 = _mm256_maskload_ps(i${K}, vmask);

          $if K == 0:
            const __m256 vk${K}x01234567 = _mm256_load_ps(w);
          $else:
            const __m256 vk${K}x01234567 = _mm256_load_ps(w + ${(K) * CHANNEL_SUBTILE});
          $if 1 <= K < ACCUMULATORS:
            __m256 vacc01234567p${K} = _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567);
          $elif FMA == 3:
            vacc01234567p${K % ACCUMULATORS} = _mm256_fmadd_ps(vi${K}x01234567, vk${K}x01234567, vacc01234567p${K % ACCUMULATORS});
          $else:
            vacc01234567p${K % ACCUMULATORS} = _mm256_add_ps(vacc01234567p${K % ACCUMULATORS}, _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567));

        w += ${(MIDDLE_PASS_TILE) * CHANNEL_SUBTILE};

        $if ACCUMULATORS > 1:
          // Add up all accumulators to vacc01234567p0
          $ACC_SLICE = 1
          $while ACC_SLICE < ACCUMULATORS:
            $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
              $if A + ACC_SLICE < ACCUMULATORS:
                vacc01234567p${A} = _mm256_add_ps(vacc01234567p${A}, vacc01234567p${A + ACC_SLICE});
            $ACC_SLICE *= 2

        _mm256_store_ps(b, vacc01234567p0);
      }
    }

    // Last pass to process up to ${LAST_PASS_TILE} inputs.
    {
      float* b = buffer;
      $for K in range(0, LAST_PASS_TILE):
        const float* i${K} = input[${K}];
        assert(i${K} != NULL);
        if XNN_UNPREDICTABLE(i${K} != zero) {
          i${K} = (const float*) ((uintptr_t) i${K} + input_offset);
        }

      size_t c = channels;
      $if CHANNEL_TILE > 8:
        for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
          $for C in range(0, CHANNEL_TILE, 8):
            $if C == 0:
              __m256 vacc${ABC[C:C+8]}p0 = _mm256_load_ps(b);
            $else:
              __m256 vacc${ABC[C:C+8]}p0 = _mm256_load_ps(b + ${C});
          b += ${CHANNEL_TILE};

          $for K in range(LAST_PASS_TILE):

            $for C in range(0, CHANNEL_TILE, 8):
              $if C == 0:
                const __m256 vi${K}x${ABC[C:C+8]} = _mm256_loadu_ps(i${K});
              $else:
                const __m256 vi${K}x${ABC[C:C+8]} = _mm256_loadu_ps(i${K} + ${C});
            i${K} += ${CHANNEL_TILE};

            $for C in range(0, CHANNEL_TILE, 8):
              $if K == 0 and C == 0:
                __m256 vk${K}x${ABC[C:C+8]} = _mm256_load_ps(w);
              $else:
                __m256 vk${K}x${ABC[C:C+8]} = _mm256_load_ps(w + ${K * CHANNEL_TILE + C});

            $for C in range(0, CHANNEL_TILE, 8):
              $if 1 <= K < ACCUMULATORS:
                __m256 vacc${ABC[C:C+8]}p${K} = _mm256_mul_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]});
              $elif FMA == 3:
                vacc${ABC[C:C+8]}p${K % ACCUMULATORS} = _mm256_fmadd_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]}, vacc${ABC[C:C+8]}p${K % ACCUMULATORS});
              $else:
                vacc${ABC[C:C+8]}p${K % ACCUMULATORS} = _mm256_add_ps(vacc${ABC[C:C+8]}p${K % ACCUMULATORS}, _mm256_mul_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]}));

          w += ${LAST_PASS_TILE * CHANNEL_TILE};

          $if ACCUMULATORS > 1:
            // Add up all accumulators to vacc${ABC[0:8]}p0
            $ACC_SLICE = 1
            $while ACC_SLICE < ACCUMULATORS:
              $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
                $if A + ACC_SLICE < ACCUMULATORS:
                  $for C in range(0, CHANNEL_TILE, 8):
                    vacc${ABC[C:C+8]}p${A} = _mm256_add_ps(vacc${ABC[C:C+8]}p${A}, vacc${ABC[C:C+8]}p${A + ACC_SLICE});
              $ACC_SLICE *= 2

          $for C in range(0, CHANNEL_TILE, 8):
            __m256 vacc${ABC[C:C+8]} = _mm256_max_ps(vmin, vacc${ABC[C:C+8]}p0);

          $for C in range(0, CHANNEL_TILE, 8):
            vacc${ABC[C:C+8]} = _mm256_min_ps(vmax, vacc${ABC[C:C+8]});

          $for C in range(0, CHANNEL_TILE, 8):
            $if C == 0:
              _mm256_storeu_ps(output, vacc${ABC[C:C+8]});
            $else:
              _mm256_storeu_ps(output + ${C}, vacc${ABC[C:C+8]});
          output += ${CHANNEL_TILE};
        }


      for (; c >= 8; c -= 8) {
        __m256 vacc01234567p0 = _mm256_load_ps(b);
        b += 8;

        $for K in range(LAST_PASS_TILE):

          const __m256 vi${K}x01234567 = _mm256_loadu_ps(i${K});
          i${K} += 8;

          $if K == 0:
            __m256 vk${K}x01234567 = _mm256_load_ps(w);
          $else:
            __m256 vk${K}x01234567 = _mm256_load_ps(w + ${K * 8});

          $if 1 <= K < ACCUMULATORS:
            __m256 vacc01234567p${K} = _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567);
          $elif FMA == 3:
            vacc01234567p${K % ACCUMULATORS} = _mm256_fmadd_ps(vi${K}x01234567, vk${K}x01234567, vacc01234567p${K % ACCUMULATORS});
          $else:
            vacc01234567p${K % ACCUMULATORS} = _mm256_add_ps(vacc01234567p${K % ACCUMULATORS}, _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567));

        $if CHANNEL_TILE > 8:
          w += ${LAST_PASS_TILE * 8};
        $else:
          w += ${LAST_PASS_TILE * CHANNEL_TILE};


        $if ACCUMULATORS > 1:
          // Add up all accumulators to vacc01234567p0
          $ACC_SLICE = 1
          $while ACC_SLICE < ACCUMULATORS:
            $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
              $if A + ACC_SLICE < ACCUMULATORS:
                vacc01234567p${A} = _mm256_add_ps(vacc01234567p${A}, vacc01234567p${A + ACC_SLICE});
            $ACC_SLICE *= 2

        __m256 vacc01234567 = _mm256_max_ps(vmin, vacc01234567p0);

        vacc01234567 = _mm256_min_ps(vmax, vacc01234567);

        _mm256_storeu_ps(output, vacc01234567);
        output += 8;
      }

      if XNN_UNLIKELY(c != 0) {
        assert(c >= 1);
        assert(c <= ${CHANNEL_SUBTILE-1});
        __m256 vacc01234567p0 = _mm256_load_ps(b);
        const __m256i vmask = _mm256_loadu_si256((const __m256i*) &mask_table[7 - c]);
        $for K in range(LAST_PASS_TILE):

          const __m256 vi${K}x01234567 = _mm256_maskload_ps(i${K}, vmask);
          $if K == 0:
            __m256 vk${K}x01234567 = _mm256_load_ps(w);
          $else:
            __m256 vk${K}x01234567 = _mm256_load_ps(w + ${K * 8});
          $if 1 <= K < ACCUMULATORS:
            __m256 vacc01234567p${K} = _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567);
          $elif FMA == 3:
            vacc01234567p${K % ACCUMULATORS} = _mm256_fmadd_ps(vi${K}x01234567, vk${K}x01234567, vacc01234567p${K % ACCUMULATORS});
          $else:
            vacc01234567p${K % ACCUMULATORS} = _mm256_add_ps(vacc01234567p${K % ACCUMULATORS}, _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567));

        $if ACCUMULATORS > 1:
          // Add up all accumulators to vacc${ABC[0:8]}p0
          $ACC_SLICE = 1
          $while ACC_SLICE < ACCUMULATORS:
            $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
              $if A + ACC_SLICE < ACCUMULATORS:
                vacc01234567p${A} = _mm256_add_ps(vacc01234567p${A}, vacc01234567p${A + ACC_SLICE});
            $ACC_SLICE *= 2

        __m256 vacc01234567 = _mm256_max_ps(vmin, vacc01234567p0);
        vacc01234567 = _mm256_min_ps(vmax, vacc01234567);

        __m128 vacc0123 = _mm256_castps256_ps128(vacc01234567);
        if (c & 4) {
          _mm_storeu_ps(output, vacc0123);
          vacc0123 = _mm256_extractf128_ps(vacc01234567, 1);
          output += 4;
        }
        if (c & 2) {
          _mm_storel_pi((__m64*) output, vacc0123);
          vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
          output += 2;
        }
        if (c & 1) {
          _mm_store_ss(output, vacc0123);
          output += 1;
        }
      }

    }
    input = (const float**) ((uintptr_t) input + input_stride);
    output = (float*) ((uintptr_t) output + output_increment);
  } while (--output_width != 0);
}