// 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 = 4
$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"
$VMULADDQ_F32 = "vfmaq_f32" if FMA else "vmlaq_f32"
#include <assert.h>
#include <stddef.h>
#include <stdint.h>

#include <arm_neon.h>

#include <xnnpack/dwconv.h>
#include <xnnpack/math.h>


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__${"neonfma" if FMA else "neon"}${"" 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});

  const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
  const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
  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.
      $if CHANNEL_TILE == 4:
        size_t c = 0;
        for (; c < channels; c += 4) {
          float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;

          $for K in range(FIRST_PASS_TILE):

            const float32x4_t vi${K}x0123 = vld1q_f32(i${K}); i${K} += 4;

            const float32x4_t vk${K}x0123 = vld1q_f32(w); w += 4;
            $if 1 <= K < ACCUMULATORS:
              float32x4_t vacc0123p${K} = vmulq_f32(vi${K}x0123, vk${K}x0123);
            $else:
              vacc0123p${K % ACCUMULATORS} = ${VMULADDQ_F32}(vacc0123p${K % ACCUMULATORS}, vi${K}x0123, vk${K}x0123);

          $if ACCUMULATORS > 1:
            // Add up all accumulators to vacc0123p0
            $ACC_SLICE = 1
            $while ACC_SLICE < ACCUMULATORS:
              $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
                $if A + ACC_SLICE < ACCUMULATORS:
                  vacc0123p${A} = vaddq_f32(vacc0123p${A}, vacc0123p${A + ACC_SLICE});
              $ACC_SLICE *= 2

          vst1q_f32(b, vacc0123p0); b += 4;
        }
      $else:
        size_t c = round_up_po2(channels, ${CHANNEL_ROUND});
        for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
          $for C in range(0, CHANNEL_TILE, 4):
            float32x4_t vacc${ABC[C:C+4]}p0 = vld1q_f32(w); w += 4;

          $for K in range(FIRST_PASS_TILE):

            $for C in range(0, CHANNEL_TILE, 4):
              const float32x4_t vi${K}x${ABC[C:C+4]} = vld1q_f32(i${K}); i${K} += 4;

            $for C in range(0, CHANNEL_TILE, 4):
              const float32x4_t vk${K}x${ABC[C:C+4]} = vld1q_f32(w); w += 4;
            $for C in range(0, CHANNEL_TILE, 4):
              $if 1 <= K < ACCUMULATORS:
                float32x4_t vacc${ABC[C:C+4]}p${K} = vmulq_f32(vi${K}x${ABC[C:C+4]}, vk${K}x${ABC[C:C+4]});
              $else:
                vacc${ABC[C:C+4]}p${K % ACCUMULATORS} = ${VMULADDQ_F32}(vacc${ABC[C:C+4]}p${K % ACCUMULATORS}, vi${K}x${ABC[C:C+4]}, vk${K}x${ABC[C:C+4]});

          $if ACCUMULATORS > 1:
            // Add up all accumulators to vacc${ABC[0:4]}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${A} = vaddq_f32(vacc${ABC[C:C+4]}p${A}, vacc${ABC[C:C+4]}p${A + ACC_SLICE});
              $ACC_SLICE *= 2

          $for C in range(0, CHANNEL_TILE, 4):
            vst1q_f32(b, vacc${ABC[C:C+4]}p0); b += 4;
        }

        $if CHANNEL_TILE == 8:
          if (c != 0) {
            float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;

            $for K in range(FIRST_PASS_TILE):

              const float32x4_t vi${K}x0123 = vld1q_f32(i${K}); i${K} += 4;

              const float32x4_t vk${K}x0123 = vld1q_f32(w); w += 4;
              $if 1 <= K < ACCUMULATORS:
                float32x4_t vacc0123p${K} = vmulq_f32(vi${K}x0123, vk${K}x0123);
              $else:
                vacc0123p${K % ACCUMULATORS} = ${VMULADDQ_F32}(vacc0123p${K % ACCUMULATORS}, vi${K}x0123, vk${K}x0123);

            $if ACCUMULATORS > 1:
              // Add up all accumulators to vacc0123p0
              $ACC_SLICE = 1
              $while ACC_SLICE < ACCUMULATORS:
                $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
                  $if A + ACC_SLICE < ACCUMULATORS:
                    vacc0123p${A} = vaddq_f32(vacc0123p${A}, vacc0123p${A + ACC_SLICE});
                $ACC_SLICE *= 2

            vst1q_f32(b, vacc0123p0); b += 4;
          }
        $else:
          for (; c != 0; c -= 4) {
            float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;

            $for K in range(FIRST_PASS_TILE):

              const float32x4_t vi${K}x0123 = vld1q_f32(i${K}); i${K} += 4;

              const float32x4_t vk${K}x0123 = vld1q_f32(w); w += 4;
              $if 1 <= K < ACCUMULATORS:
                float32x4_t vacc0123p${K} = vmulq_f32(vi${K}x0123, vk${K}x0123);
              $else:
                vacc0123p${K % ACCUMULATORS} = ${VMULADDQ_F32}(vacc0123p${K % ACCUMULATORS}, vi${K}x0123, vk${K}x0123);

            $if ACCUMULATORS > 1:
              // Add up all accumulators to vacc0123p0
              $ACC_SLICE = 1
              $while ACC_SLICE < ACCUMULATORS:
                $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
                  $if A + ACC_SLICE < ACCUMULATORS:
                    vacc0123p${A} = vaddq_f32(vacc0123p${A}, vacc0123p${A + ACC_SLICE});
                $ACC_SLICE *= 2

            vst1q_f32(b, vacc0123p0); b += 4;
          }
    }

    // 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};

      $if CHANNEL_TILE == 4:
        size_t c = 0;
        for (; c < channels; c += 4) {
          float32x4_t vacc0123p0 = vld1q_f32(b);

          $for K in range(MIDDLE_PASS_TILE):

            const float32x4_t vi${K}x0123 = vld1q_f32(i${K}); i${K} += 4;

            const float32x4_t vk${K}x0123 = vld1q_f32(w); w += 4;
            $if 1 <= K < ACCUMULATORS:
              float32x4_t vacc0123p${K} = vmulq_f32(vi${K}x0123, vk${K}x0123);
            $else:
              vacc0123p${K % ACCUMULATORS} = ${VMULADDQ_F32}(vacc0123p${K % ACCUMULATORS}, vi${K}x0123, vk${K}x0123);

          $if ACCUMULATORS > 1:
            // Add up all accumulators to vacc0123p0
            $ACC_SLICE = 1
            $while ACC_SLICE < ACCUMULATORS:
              $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
                $if A + ACC_SLICE < ACCUMULATORS:
                  vacc0123p${A} = vaddq_f32(vacc0123p${A}, vacc0123p${A + ACC_SLICE});
              $ACC_SLICE *= 2

          vst1q_f32(b, vacc0123p0); b += 4;
        }
      $else:
        size_t c = round_up_po2(channels, ${CHANNEL_ROUND});
        for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
          float32x4_t vacc0123p0 = vld1q_f32(b);
          $for C in range(4, CHANNEL_TILE, 4):
            float32x4_t vacc${ABC[C:C+4]}p0 = vld1q_f32(b + ${C});

          $for K in range(MIDDLE_PASS_TILE):

            $for C in range(0, CHANNEL_TILE, 4):
              const float32x4_t vi${K}x${ABC[C:C+4]} = vld1q_f32(i${K}); i${K} += 4;

            $for C in range(0, CHANNEL_TILE, 4):
              const float32x4_t vk${K}x${ABC[C:C+4]} = vld1q_f32(w); w += 4;
            $for C in range(0, CHANNEL_TILE, 4):
              $if 1 <= K < ACCUMULATORS:
                float32x4_t vacc${ABC[C:C+4]}p${K} = vmulq_f32(vi${K}x${ABC[C:C+4]}, vk${K}x${ABC[C:C+4]});
              $else:
                vacc${ABC[C:C+4]}p${K % ACCUMULATORS} = ${VMULADDQ_F32}(vacc${ABC[C:C+4]}p${K % ACCUMULATORS}, vi${K}x${ABC[C:C+4]}, vk${K}x${ABC[C:C+4]});

          $if ACCUMULATORS > 1:
            // Add up all accumulators to vacc${ABC[0:4]}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${A} = vaddq_f32(vacc${ABC[C:C+4]}p${A}, vacc${ABC[C:C+4]}p${A + ACC_SLICE});
              $ACC_SLICE *= 2

          $for C in range(0, CHANNEL_TILE, 4):
            vst1q_f32(b, vacc${ABC[C:C+4]}p0); b += 4;
        }

        $if CHANNEL_TILE == 8:
          if (c != 0) {
            float32x4_t vacc0123p0 = vld1q_f32(b);

            $for K in range(MIDDLE_PASS_TILE):

              const float32x4_t vi${K}x0123 = vld1q_f32(i${K}); i${K} += 4;

              const float32x4_t vk${K}x0123 = vld1q_f32(w); w += 4;
              $if 1 <= K < ACCUMULATORS:
                float32x4_t vacc0123p${K} = vmulq_f32(vi${K}x0123, vk${K}x0123);
              $else:
                vacc0123p${K % ACCUMULATORS} = ${VMULADDQ_F32}(vacc0123p${K % ACCUMULATORS}, vi${K}x0123, vk${K}x0123);

            $if ACCUMULATORS > 1:
              // Add up all accumulators to vacc0123p0
              $ACC_SLICE = 1
              $while ACC_SLICE < ACCUMULATORS:
                $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
                  $if A + ACC_SLICE < ACCUMULATORS:
                    vacc0123p${A} = vaddq_f32(vacc0123p${A}, vacc0123p${A + ACC_SLICE});
                $ACC_SLICE *= 2

            vst1q_f32(b, vacc0123p0); b += 4;
          }
        $else:
          for (; c != 0; c -= 4) {
            float32x4_t vacc0123p0 = vld1q_f32(b);

            $for K in range(MIDDLE_PASS_TILE):

              const float32x4_t vi${K}x0123 = vld1q_f32(i${K}); i${K} += 4;

              const float32x4_t vk${K}x0123 = vld1q_f32(w); w += 4;
              $if 1 <= K < ACCUMULATORS:
                float32x4_t vacc0123p${K} = vmulq_f32(vi${K}x0123, vk${K}x0123);
              $else:
                vacc0123p${K % ACCUMULATORS} = ${VMULADDQ_F32}(vacc0123p${K % ACCUMULATORS}, vi${K}x0123, vk${K}x0123);

            $if ACCUMULATORS > 1:
              // Add up all accumulators to vacc0123p0
              $ACC_SLICE = 1
              $while ACC_SLICE < ACCUMULATORS:
                $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
                  $if A + ACC_SLICE < ACCUMULATORS:
                    vacc0123p${A} = vaddq_f32(vacc0123p${A}, vacc0123p${A + ACC_SLICE});
                $ACC_SLICE *= 2

            vst1q_f32(b, vacc0123p0);
            b += 4;
          }
    }

    // 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 > 4:
        for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
          $for C in range(0, CHANNEL_TILE, 4):
            float32x4_t vacc${ABC[C:C+4]}p0 = vld1q_f32(b); b += 4;

          $for K in range(LAST_PASS_TILE):

            $for C in range(0, CHANNEL_TILE, 4):
              const float32x4_t vi${K}x${ABC[C:C+4]} = vld1q_f32(i${K}); i${K} += 4;

            $for C in range(0, CHANNEL_TILE, 4):
              float32x4_t vk${K}x${ABC[C:C+4]} = vld1q_f32(w); w += 4;

            $for C in range(0, CHANNEL_TILE, 4):
              $if 1 <= K < ACCUMULATORS:
                float32x4_t vacc${ABC[C:C+4]}p${K} = vmulq_f32(vi${K}x${ABC[C:C+4]}, vk${K}x${ABC[C:C+4]});
              $else:
                vacc${ABC[C:C+4]}p${K % ACCUMULATORS} = ${VMULADDQ_F32}(vacc${ABC[C:C+4]}p${K % ACCUMULATORS}, vi${K}x${ABC[C:C+4]}, vk${K}x${ABC[C:C+4]});

          $if ACCUMULATORS > 1:
            // Add up all accumulators to vacc${ABC[0:4]}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${A} = vaddq_f32(vacc${ABC[C:C+4]}p${A}, vacc${ABC[C:C+4]}p${A + ACC_SLICE});
              $ACC_SLICE *= 2

          $for C in range(0, CHANNEL_TILE, 4):
            float32x4_t vacc${ABC[C:C+4]} = vmaxq_f32(vacc${ABC[C:C+4]}p0, vmin);

          $for C in range(0, CHANNEL_TILE, 4):
            vacc${ABC[C:C+4]} = vminq_f32(vacc${ABC[C:C+4]}, vmax);

          $for C in range(0, CHANNEL_TILE, 4):
            vst1q_f32(output, vacc${ABC[C:C+4]}); output += 4;
        }


      for (; c >= 4; c -= 4) {
        float32x4_t vacc0123p0 = vld1q_f32(b); b += 4;

        $for K in range(LAST_PASS_TILE):

          const float32x4_t vi${K}x0123 = vld1q_f32(i${K}); i${K} += 4;

          float32x4_t vk${K}x0123 = vld1q_f32(w); w += 4;

          $if 1 <= K < ACCUMULATORS:
            float32x4_t vacc0123p${K} = vmulq_f32(vi${K}x0123, vk${K}x0123);
          $else:
            vacc0123p${K % ACCUMULATORS} = ${VMULADDQ_F32}(vacc0123p${K % ACCUMULATORS}, vi${K}x0123, vk${K}x0123);


        $if ACCUMULATORS > 1:
          // Add up all accumulators to vacc0123p0
          $ACC_SLICE = 1
          $while ACC_SLICE < ACCUMULATORS:
            $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
              $if A + ACC_SLICE < ACCUMULATORS:
                vacc0123p${A} = vaddq_f32(vacc0123p${A}, vacc0123p${A + ACC_SLICE});
            $ACC_SLICE *= 2

        float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);

        vacc0123 = vminq_f32(vacc0123, vmax);

        vst1q_f32(output, vacc0123); output += 4;
      }

      if XNN_UNLIKELY(c != 0) {
        float32x4_t vacc0123p0 = vld1q_f32(b);
        $for K in range(LAST_PASS_TILE):

          const float32x4_t vi${K}x0123 = vld1q_f32(i${K});
          float32x4_t vk${K}x0123 = vld1q_f32(w); w += 4;
          $if 1 <= K < ACCUMULATORS:
            float32x4_t vacc0123p${K} = vmulq_f32(vi${K}x0123, vk${K}x0123);
          $else:
            vacc0123p${K % ACCUMULATORS} = ${VMULADDQ_F32}(vacc0123p${K % ACCUMULATORS}, vi${K}x0123, vk${K}x0123);

        $if ACCUMULATORS > 1:
          // Add up all accumulators to vacc${ABC[0:4]}p0
          $ACC_SLICE = 1
          $while ACC_SLICE < ACCUMULATORS:
            $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
              $if A + ACC_SLICE < ACCUMULATORS:
                vacc0123p${A} = vaddq_f32(vacc0123p${A}, vacc0123p${A + ACC_SLICE});
            $ACC_SLICE *= 2

        float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
        vacc0123 = vminq_f32(vacc0123, vmax);

        float32x2_t vacc01 = vget_low_f32(vacc0123);
        if (c & 2) {
          vst1_f32(output, vacc01); output += 2;
          vacc01 = vget_high_f32(vacc0123);
        }
        if (c & 1) {
          vst1_lane_f32(output, vacc01, 0); output += 1;
        }
      }

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