// Copyright 2019 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 % 4 == 0
$assert KERNEL_TILE >= 2
$assert ACCUMULATORS >= 1
$ABC = "0123456789ABCDEF"
#include <assert.h>

#include <xmmintrin.h>

#include <xnnpack/dwconv.h>


void xnn_f32_dwconv_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__sse${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
    size_t channels,
    size_t output_width,
    const float** input,
    const float* weights,
    float* output,
    size_t input_stride,
    size_t output_increment,
    const union xnn_f32_output_params params[restrict static 1])
{
  assert(channels != 0);
  assert(output_width != 0);

  const __m128 vmax = _mm_load_ps(params->sse.max);
  const __m128 vmin = _mm_load_ps(params->sse.min);
  do {
    $for K in range(KERNEL_TILE):
      const float* i${K} = input[${K}];
      assert(i${K} != NULL);
    input = (const float**) ((uintptr_t) input + input_stride);

    size_t c = channels;
    const float* w = weights;
    for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
      __m128 vacc${ABC[0:4]}p0 = _mm_load_ps(w);
      $for C in range(4, CHANNEL_TILE, 4):
        __m128 vacc${ABC[C:C+4]}p0 = _mm_load_ps(w + ${C});

      $for K in range(KERNEL_TILE):

        const __m128 vi${K}x${ABC[0:4]} = _mm_loadu_ps(i${K});
        $for C in range(4, CHANNEL_TILE, 4):
          const __m128 vi${K}x${ABC[C:C+4]} = _mm_loadu_ps(i${K} + ${C});
        i${K} += ${CHANNEL_TILE};

        $for C in range(0, CHANNEL_TILE, 4):
          const __m128 vk${K}x${ABC[C:C+4]} = _mm_load_ps(w + ${(K + 1) * CHANNEL_TILE + C});
        $for C in range(0, CHANNEL_TILE, 4):
          $if 1 <= K < ACCUMULATORS:
            __m128 vacc${ABC[C:C+4]}p${K} = _mm_mul_ps(vi${K}x${ABC[C:C+4]}, vk${K}x${ABC[C:C+4]});
          $else:
            vacc${ABC[C:C+4]}p${K % ACCUMULATORS} = _mm_add_ps(vacc${ABC[C:C+4]}p${K % ACCUMULATORS}, _mm_mul_ps(vi${K}x${ABC[C:C+4]}, vk${K}x${ABC[C:C+4]}));

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

      $if ACCUMULATORS > 1:
        // Add up all accumulators to vacc${ABC[0:CHANNEL_TILE]}p0
        $ACC_SLICE = 1
        $while ACC_SLICE < ACCUMULATORS:
          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
            $if A + ACC_SLICE < ACCUMULATORS:
              $for C in range(0, CHANNEL_TILE, 4):
                vacc${ABC[C:C+4]}p${A} = _mm_add_ps(vacc${ABC[C:C+4]}p${A}, vacc${ABC[C:C+4]}p${A + ACC_SLICE});
          $ACC_SLICE *= 2

      $for C in range(0, CHANNEL_TILE, 4):
        __m128 vacc${ABC[C:C+4]} = _mm_max_ps(vacc${ABC[C:C+4]}p0, vmin);
      $for C in range(0, CHANNEL_TILE, 4):
        vacc${ABC[C:C+4]} = _mm_min_ps(vacc${ABC[C:C+4]}, vmax);

      _mm_storeu_ps(output, vacc${ABC[0:4]});
      $for C in range(4, CHANNEL_TILE, 4):
        _mm_storeu_ps(output + ${C}, vacc${ABC[C:C+4]});
      output += ${CHANNEL_TILE};
    }
    $if CHANNEL_TILE > 4:
      for (; c >= 4; c -= 4) {
        __m128 vacc0123p0 = _mm_load_ps(w);
        $for K in range(KERNEL_TILE):

          const __m128 vi${K}x0123 = _mm_loadu_ps(i${K});
          i${K} += 4;

          const __m128 vk${K}x0123 = _mm_load_ps(w + ${(K + 1) * CHANNEL_TILE});
          $if 1 <= K < ACCUMULATORS:
            __m128 vacc0123p${K} = _mm_mul_ps(vi${K}x0123, vk${K}x0123);
          $else:
            vacc0123p${K % ACCUMULATORS} = _mm_add_ps(vacc0123p${K % ACCUMULATORS}, _mm_mul_ps(vi${K}x0123, vk${K}x0123));

        w += 4;

        $if ACCUMULATORS > 1:
          // Add up all accumulators to vacc${ABC[0:CHANNEL_TILE]}p0
          $ACC_SLICE = 1
          $while ACC_SLICE < ACCUMULATORS:
            $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
              $if A + ACC_SLICE < ACCUMULATORS:
                vacc0123p${A} = _mm_add_ps(vacc0123p${A}, vacc0123p${A + ACC_SLICE});
            $ACC_SLICE *= 2

        __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
        vacc0123 = _mm_min_ps(vacc0123, vmax);

        _mm_storeu_ps(output, vacc0123);
        output += 4;
      }
    if XNN_UNLIKELY(c != 0) {
      __m128 vacc0123p0 = _mm_load_ps(w);
      $for K in range(KERNEL_TILE):

        const __m128 vi${K}x0123 = _mm_loadu_ps(i${K});
        const __m128 vk${K}x0123 = _mm_load_ps(w + ${(K + 1) * CHANNEL_TILE});
        $if 1 <= K < ACCUMULATORS:
          __m128 vacc0123p${K} = _mm_mul_ps(vi${K}x0123, vk${K}x0123);
        $else:
          vacc0123p${K % ACCUMULATORS} = _mm_add_ps(vacc0123p${K % ACCUMULATORS}, _mm_mul_ps(vi${K}x0123, vk${K}x0123));

      $if ACCUMULATORS > 1:
        // Add up all accumulators to vacc${ABC[0:CHANNEL_TILE]}p0
        $ACC_SLICE = 1
        $while ACC_SLICE < ACCUMULATORS:
          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
            $if A + ACC_SLICE < ACCUMULATORS:
              vacc0123p${A} = _mm_add_ps(vacc0123p${A}, vacc0123p${A + ACC_SLICE});
          $ACC_SLICE *= 2

      __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
      vacc0123 = _mm_min_ps(vacc0123, vmax);

      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;
      }
    }

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