// Copyright 2020 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.

$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$assert CHANNEL_TILE % 8 == 0
$assert CHANNEL_TILE >= 8
$assert KERNEL_TILE >= 2
#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/dwconv.h>


void xnn_qs8_dwconv_minmax_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__neon_mul16(
    size_t channels,
    size_t output_width,
    const int8_t** input,
    const void* weights,
    int8_t* output,
    size_t input_stride,
    size_t output_increment,
    size_t input_offset,
    const int8_t* zero,
    const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  assert(channels != 0);
  assert(output_width != 0);

  const int32x4_t vmultiplier = vld1q_dup_s32(&params->neon.multiplier);
  const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
  const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
  const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
  $if CHANNEL_TILE == 8:
    const int8x8_t voutput_min = vld1_dup_s8(&params->neon.output_min);
    const int8x8_t voutput_max = vld1_dup_s8(&params->neon.output_max);
  $else:
    const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
    const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);
  do {
    $for K in range(KERNEL_TILE):
      const int8_t* i${K} = input[${K}];
      assert(i${K} != NULL);
      if XNN_UNPREDICTABLE(i${K} != zero) {
        i${K} = (const int8_t*) ((uintptr_t) i${K} + input_offset);
      }
    input = (const int8_t**) ((uintptr_t) input + input_stride);

    size_t c = channels;
    const void* w = weights;
    for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
      $for C in range(0, CHANNEL_TILE, 4):
        int32x4_t vacc${ABC[C:C+4]} = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));

      $for K in range(KERNEL_TILE):

        $for C in range(0, CHANNEL_TILE, 8):
          const int16x8_t vi${K}x${ABC[C:C+8]} = vmovl_s8(vld1_s8(i${K})); i${K} += 8;
          const int16x8_t vk${K}x${ABC[C:C+8]} = vmovl_s8(vld1_s8(w)); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));

        $for C in range(0, CHANNEL_TILE, 8):
          vacc${ABC[C:C+4]} = vmlal_s16(vacc${ABC[C:C+4]}, vget_low_s16(vi${K}x${ABC[C:C+8]}), vget_low_s16(vk${K}x${ABC[C:C+8]}));
          vacc${ABC[C+4:C+8]} = vmlal_s16(vacc${ABC[C+4:C+8]}, vget_high_s16(vi${K}x${ABC[C:C+8]}), vget_high_s16(vk${K}x${ABC[C:C+8]}));

      $for C in range(0, CHANNEL_TILE, 4):
        vacc${ABC[C:C+4]} = vqrdmulhq_s32(vacc${ABC[C:C+4]}, vmultiplier);

      $for C in range(0, CHANNEL_TILE, 4):
        vacc${ABC[C:C+4]} = vsraq_n_s32(vacc${ABC[C:C+4]}, vbicq_s32(vacc${ABC[C:C+4]}, vzero_shift_mask), 31);

      $for C in range(0, CHANNEL_TILE, 4):
        vacc${ABC[C:C+4]} = vrshlq_s32(vacc${ABC[C:C+4]}, vright_shift);

#if XNN_ARCH_ARM64
      $for C in range(0, CHANNEL_TILE, 8):
        const int16x8_t vacc${ABC[C:C+8]} = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc${ABC[C:C+4]}), vacc${ABC[C+4:C+8]}), voutput_zero_point);

      $for C in range(0, CHANNEL_TILE, 16):
        $if C + 8 < CHANNEL_TILE:
          int8x16_t vout${ABC[C:C+16]} = vqmovn_high_s16(vqmovn_s16(vacc${ABC[C:C+8]}), vacc${ABC[C+8:C+16]});
        $else:
          int8x8_t vout${ABC[C:C+8]} = vqmovn_s16(vacc${ABC[C:C+8]});
#else
      $for C in range(0, CHANNEL_TILE, 8):
        const int16x8_t vacc${ABC[C:C+8]} = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc${ABC[C:C+4]}), vqmovn_s32(vacc${ABC[C+4:C+8]})), voutput_zero_point);

      $for C in range(0, CHANNEL_TILE, 16):
        $if C + 8 < CHANNEL_TILE:
          int8x16_t vout${ABC[C:C+16]} = vcombine_s8(vqmovn_s16(vacc${ABC[C:C+8]}), vqmovn_s16(vacc${ABC[C+8:C+16]}));
        $else:
          int8x8_t vout${ABC[C:C+8]} = vqmovn_s16(vacc${ABC[C:C+8]});
#endif

      $for C in range(0, CHANNEL_TILE, 16):
        $if C + 8 < CHANNEL_TILE:
          vout${ABC[C:C+16]} = vmaxq_s8(vout${ABC[C:C+16]}, voutput_min);
        $else:
          $if CHANNEL_TILE == 8:
            vout${ABC[C:C+8]} = vmax_s8(vout${ABC[C:C+8]}, voutput_min);
          $else:
            vout${ABC[C:C+8]} = vmax_s8(vout${ABC[C:C+8]}, vget_low_s8(voutput_min));

      $for C in range(0, CHANNEL_TILE, 16):
        $if C + 8 < CHANNEL_TILE:
          vout${ABC[C:C+16]} = vminq_s8(vout${ABC[C:C+16]}, voutput_max);
        $else:
          $if CHANNEL_TILE == 8:
            vout${ABC[C:C+8]} = vmin_s8(vout${ABC[C:C+8]}, voutput_max);
          $else:
            vout${ABC[C:C+8]} = vmin_s8(vout${ABC[C:C+8]}, vget_low_s8(voutput_max));

      $for C in range(0, CHANNEL_TILE, 16):
        $if C + 8 < CHANNEL_TILE:
          vst1q_s8(output, vout${ABC[C:C+16]}); output += 16;
        $else:
          vst1_s8(output, vout${ABC[C:C+8]}); output += 8;
    }
    if XNN_UNLIKELY(c != 0) {
      $if CHANNEL_TILE > 8:
        const int8_t* k = (const int8_t*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t));
      ${"do " if CHANNEL_TILE > 8 else ""}{
        int32x4_t vacc${ABC[0:4]} = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
        int32x4_t vacc${ABC[4:8]} = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));

        $for K in range(KERNEL_TILE):
          $if CHANNEL_TILE > 8:
            const int16x8_t vi${K}x${ABC[0:8]} = vmovl_s8(vld1_s8(i${K})); i${K} += 8;
          $else:
            const int16x8_t vi${K}x${ABC[0:8]} = vmovl_s8(vld1_s8(i${K}));
          $if CHANNEL_TILE > 8:
            $if K == 0:
              const int16x8_t vk${K}x${ABC[0:8]} = vmovl_s8(vld1_s8(k)); k += 8;
            $else:
              const int16x8_t vk${K}x${ABC[0:8]} = vmovl_s8(vld1_s8((const void*) (k + ${K * CHANNEL_TILE - 8})));
          $else:
            $if K == 0:
              const int16x8_t vk${K}x${ABC[0:8]} = vmovl_s8(vld1_s8(w));
            $else:
              const int16x8_t vk${K}x${ABC[0:8]} = vmovl_s8(vld1_s8((const void*) ((uintptr_t) w + ${K * CHANNEL_TILE} * sizeof(int8_t))));

          vacc${ABC[0:4]} = vmlal_s16(vacc${ABC[0:4]}, vget_low_s16(vi${K}x${ABC[0:8]}), vget_low_s16(vk${K}x${ABC[0:8]}));
          vacc${ABC[4:8]} = vmlal_s16(vacc${ABC[4:8]}, vget_high_s16(vi${K}x${ABC[0:8]}), vget_high_s16(vk${K}x${ABC[0:8]}));

        vacc${ABC[0:4]} = vqrdmulhq_s32(vacc${ABC[0:4]}, vmultiplier);
        vacc${ABC[4:8]} = vqrdmulhq_s32(vacc${ABC[4:8]}, vmultiplier);

        vacc${ABC[0:4]} = vsraq_n_s32(vacc${ABC[0:4]}, vbicq_s32(vacc${ABC[0:4]}, vzero_shift_mask), 31);
        vacc${ABC[4:8]} = vsraq_n_s32(vacc${ABC[4:8]}, vbicq_s32(vacc${ABC[4:8]}, vzero_shift_mask), 31);

        vacc${ABC[0:4]} = vrshlq_s32(vacc${ABC[0:4]}, vright_shift);
        vacc${ABC[4:8]} = vrshlq_s32(vacc${ABC[4:8]}, vright_shift);

#if XNN_ARCH_ARM64
        const int16x8_t vacc${ABC[0:8]} = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc${ABC[0:4]}), vacc${ABC[4:8]}), voutput_zero_point);

        int8x8_t vout${ABC[0:8]} = vqmovn_s16(vacc${ABC[0:8]});
#else
        const int16x8_t vacc${ABC[0:8]} = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc${ABC[0:4]}), vqmovn_s32(vacc${ABC[4:8]})), voutput_zero_point);

        int8x8_t vout${ABC[0:8]} = vqmovn_s16(vacc${ABC[0:8]});
#endif

        $if CHANNEL_TILE == 8:
          vout${ABC[0:8]} = vmax_s8(vout${ABC[0:8]}, voutput_min);
          vout${ABC[0:8]} = vmin_s8(vout${ABC[0:8]}, voutput_max);
        $else:
          vout${ABC[0:8]} = vmax_s8(vout${ABC[0:8]}, vget_low_s8(voutput_min));
          vout${ABC[0:8]} = vmin_s8(vout${ABC[0:8]}, vget_low_s8(voutput_max));

        $if CHANNEL_TILE > 8:
          if XNN_LIKELY(c >= 8) {
            vst1_s8(output, vout${ABC[0:8]}); output += 8;
            c -= 8;
          } else {
            if (c & 4) {
              vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_s8(vout${ABC[0:8]}), 0); output += 4;
              vout${ABC[0:8]} = vext_s8(vout${ABC[0:8]}, vout${ABC[0:8]}, 4);
            }
            if (c & 2) {
              vst1_lane_u16(__builtin_assume_aligned(output, 1), vreinterpret_u16_s8(vout${ABC[0:8]}), 0); output += 2;
              vout${ABC[0:8]} = vext_s8(vout${ABC[0:8]}, vout${ABC[0:8]}, 2);
            }
            if (c & 1) {
              vst1_lane_s8(output, vout${ABC[0:8]}, 0); output += 1;
            }
            c = 0;
          }
        $else:
          if (c & 4) {
            vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_s8(vout${ABC[0:8]}), 0); output += 4;
            vout${ABC[0:8]} = vext_s8(vout${ABC[0:8]}, vout${ABC[0:8]}, 4);
          }
          if (c & 2) {
            vst1_lane_u16(__builtin_assume_aligned(output, 1), vreinterpret_u16_s8(vout${ABC[0:8]}), 0); output += 2;
            vout${ABC[0:8]} = vext_s8(vout${ABC[0:8]}, vout${ABC[0:8]}, 2);
          }
          if (c & 1) {
            vst1_lane_s8(output, vout${ABC[0:8]}, 0); output += 1;
          }
      }${" while (c != 0);" if CHANNEL_TILE > 8 else ""}
    }

    output = (int8_t*) ((uintptr_t) output + output_increment);
  } while (--output_width != 0);
}