// 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 % 16 == 0
$assert KERNEL_TILE >= 2
$assert ACCUMULATORS >= 1
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <immintrin.h>

#include <xnnpack/dwconv.h>
#include <xnnpack/intrinsics-polyfill.h>


void xnn_f32_dwconv_minmax_ukernel_${KERNEL_TILE}p${CHANNEL_TILE}c__avx512f${"" 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,
    const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(channels != 0);
  assert(output_width != 0);

  const __m512 vmin = _mm512_set1_ps(params->scalar.min);
  const __m512 vmax = _mm512_set1_ps(params->scalar.max);
  do {
    $for K in range(KERNEL_TILE):
      const float* i${K} = input[${K}];
      assert(i${K} != NULL);
      if XNN_UNPREDICTABLE(i${K} != zero) {
        i${K} = (const float*) ((uintptr_t) i${K} + input_offset);
      }
    input = (const float**) ((uintptr_t) input + input_stride);

    size_t c = channels;
    const float* w = weights;
    for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
      __m512 vacc${ABC[0:16]}p0 = _mm512_load_ps(w);
      $for C in range(16, CHANNEL_TILE, 16):
        __m512 vacc${ABC[C:C+16]}p0 = _mm512_load_ps(w + ${C});

      $for K in range(KERNEL_TILE):

        const __m512 vi${K}x${ABC[0:16]} = _mm512_loadu_ps(i${K});
        $for C in range(16, CHANNEL_TILE, 16):
          const __m512 vi${K}x${ABC[C:C+16]} = _mm512_loadu_ps(i${K} + ${C});
        i${K} += ${CHANNEL_TILE};

        $for C in range(0, CHANNEL_TILE, 16):
          const __m512 vk${K}x${ABC[C:C+16]} = _mm512_load_ps(w + ${(K + 1) * CHANNEL_TILE + C});
        $for C in range(0, CHANNEL_TILE, 16):
          $if 1 <= K < ACCUMULATORS:
            __m512 vacc${ABC[C:C+16]}p${K} = _mm512_mul_ps(vi${K}x${ABC[C:C+16]}, vk${K}x${ABC[C:C+16]});
          $else:
            vacc${ABC[C:C+16]}p${K % ACCUMULATORS} = _mm512_fmadd_ps(vi${K}x${ABC[C:C+16]}, vk${K}x${ABC[C:C+16]}, vacc${ABC[C:C+16]}p${K % ACCUMULATORS});

      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:
              $for C in range(0, CHANNEL_TILE, 16):
                vacc${ABC[C:C+16]}p${A} = _mm512_add_ps(vacc${ABC[C:C+16]}p${A}, vacc${ABC[C:C+16]}p${A + ACC_SLICE});
          $ACC_SLICE *= 2

      $for C in range(0, CHANNEL_TILE, 16):
        __m512 vacc${ABC[C:C+16]} = _mm512_max_ps(vmin, vacc${ABC[C:C+16]}p0);
      $for C in range(0, CHANNEL_TILE, 16):
        vacc${ABC[C:C+16]} = _mm512_min_ps(vmax, vacc${ABC[C:C+16]});

      _mm512_storeu_ps(output, vacc${ABC[0:16]});
      $for C in range(16, CHANNEL_TILE, 16):
        _mm512_storeu_ps(output + ${C}, vacc${ABC[C:C+16]});
      output += ${CHANNEL_TILE};
    }
    $if CHANNEL_TILE > 16:
      for (; c >= 16; c -= 16) {
        __m512 vacc${ABC[0:16]}p0 = _mm512_load_ps(w);
        $for K in range(KERNEL_TILE):

          const __m512 vi${K}x${ABC[0:16]} = _mm512_loadu_ps(i${K});
          i${K} += 16;

          const __m512 vk${K}x${ABC[0:16]} = _mm512_load_ps(w + ${(K + 1) * CHANNEL_TILE});
          $if 1 <= K < ACCUMULATORS:
            __m512 vacc${ABC[0:16]}p${K} = _mm512_mul_ps(vi${K}x${ABC[0:16]}, vk${K}x${ABC[0:16]});
          $else:
            vacc${ABC[0:16]}p${K % ACCUMULATORS} = _mm512_fmadd_ps(vi${K}x${ABC[0:16]}, vk${K}x${ABC[0:16]}, vacc${ABC[0:16]}p${K % ACCUMULATORS});

        w += 16;

        $if ACCUMULATORS > 1:
          // Add up all accumulators to vacc${ABC[0:16]}p0
          $ACC_SLICE = 1
          $while ACC_SLICE < ACCUMULATORS:
            $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
              $if A + ACC_SLICE < ACCUMULATORS:
                vacc${ABC[0:16]}p${A} = _mm512_add_ps(vacc${ABC[0:16]}p${A}, vacc${ABC[0:16]}p${A + ACC_SLICE});
            $ACC_SLICE *= 2

        __m512 vacc${ABC[0:16]} = _mm512_max_ps(vmin, vacc${ABC[0:16]}p0);
        vacc${ABC[0:16]} = _mm512_min_ps(vmax, vacc${ABC[0:16]});

        _mm512_storeu_ps(output, vacc${ABC[0:16]});
        output += 16;
      }
    if XNN_UNLIKELY(c != 0) {
      assert(c >= 1);
      assert(c <= 16);
      // Prepare mask for valid 32-bit elements (depends on nc).
      const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << c) - UINT32_C(1)));

      __m512 vacc${ABC[0:16]}p0 = _mm512_maskz_loadu_ps(vmask, w);
      $for K in range(KERNEL_TILE):

        const __m512 vi${K}x${ABC[0:16]} = _mm512_maskz_loadu_ps(vmask, i${K});
        const __m512 vk${K}x${ABC[0:16]} = _mm512_maskz_loadu_ps(vmask, w + ${(K + 1) * CHANNEL_TILE});
        $if 1 <= K < ACCUMULATORS:
          __m512 vacc${ABC[0:16]}p${K} = _mm512_mul_ps(vi${K}x${ABC[0:16]}, vk${K}x${ABC[0:16]});
        $else:
          vacc${ABC[0:16]}p${K % ACCUMULATORS} = _mm512_fmadd_ps(vi${K}x${ABC[0:16]}, vk${K}x${ABC[0:16]}, vacc${ABC[0:16]}p${K % ACCUMULATORS});

      $if ACCUMULATORS > 1:
        // Add up all accumulators to vacc${ABC[0:16]}p0
        $ACC_SLICE = 1
        $while ACC_SLICE < ACCUMULATORS:
          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
            $if A + ACC_SLICE < ACCUMULATORS:
              vacc${ABC[0:16]}p${A} = _mm512_add_ps(vacc${ABC[0:16]}p${A}, vacc${ABC[0:16]}p${A + ACC_SLICE});
          $ACC_SLICE *= 2

      __m512 vacc${ABC[0:16]} = _mm512_max_ps(vmin, vacc${ABC[0:16]}p0);
      vacc${ABC[0:16]} = _mm512_min_ps(vmax, vacc${ABC[0:16]});

      _mm512_mask_storeu_ps(output, vmask, vacc${ABC[0:16]});
      output += c;
    }

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