// 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.

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

#include <wasm_simd128.h>

#include <xnnpack/vadd.h>


void xnn_qs8_vadd_minmax_ukernel__wasmsimd_x${BATCH_TILE}(
    size_t n,
    const int8_t* input_x,
    const int8_t* input_y,
    int8_t* output,
    const union xnn_qs8_add_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  const v128_t vzero_point_product = wasm_v128_load(params->wasmsimd.zero_point_product);
  const v128_t vx_multiplier = wasm_v128_load(params->wasmsimd.x_multiplier);
  const v128_t vy_multiplier = wasm_v128_load(params->wasmsimd.y_multiplier);
  const v128_t vremainder_mask = wasm_v128_load(params->wasmsimd.remainder_mask);
  const v128_t vremainder_threshold = wasm_v128_load(params->wasmsimd.remainder_threshold);
  const int32_t vshift = params->wasmsimd.shift;
  const v128_t voutput_zero_point = wasm_v128_load(params->wasmsimd.output_zero_point);
  const v128_t voutput_min = wasm_v128_load(params->wasmsimd.output_min);
  const v128_t voutput_max = wasm_v128_load(params->wasmsimd.output_max);

  for (; n >= ${BATCH_TILE} * sizeof(int8_t); n -= ${BATCH_TILE} * sizeof(int8_t)) {
    const v128_t vx${ABC[0:8]} = wasm_i16x8_load_8x8(input_x);
    const v128_t vy${ABC[0:8]} = wasm_i16x8_load_8x8(input_y);
    $for N in range(8, BATCH_TILE, 8):
      const v128_t vx${ABC[N:N+8]} = wasm_i16x8_load_8x8(input_x + ${N});
      const v128_t vy${ABC[N:N+8]} = wasm_i16x8_load_8x8(input_y + ${N});
    input_x += ${BATCH_TILE};
    input_y += ${BATCH_TILE};

    $for N in range(0, BATCH_TILE, 8):
      v128_t vacc${ABC[N:N+4]} = wasm_i32x4_add(vzero_point_product, wasm_i32x4_mul(wasm_i32x4_widen_low_i16x8(vx${ABC[N:N+8]}), vx_multiplier));
      v128_t vacc${ABC[N+4:N+8]} = wasm_i32x4_add(vzero_point_product, wasm_i32x4_mul(wasm_i32x4_widen_high_i16x8(vx${ABC[N:N+8]}), vx_multiplier));

    $for N in range(0, BATCH_TILE, 8):
      vacc${ABC[N:N+4]} = wasm_i32x4_add(vacc${ABC[N:N+4]}, wasm_i32x4_mul(wasm_i32x4_widen_low_i16x8(vy${ABC[N:N+8]}), vy_multiplier));
      vacc${ABC[N+4:N+8]} = wasm_i32x4_add(vacc${ABC[N+4:N+8]}, wasm_i32x4_mul(wasm_i32x4_widen_high_i16x8(vy${ABC[N:N+8]}), vy_multiplier));

    $for N in range(0, BATCH_TILE, 4):
      const v128_t vrem${ABC[N:N+4]} = wasm_i32x4_add(wasm_v128_and(vacc${ABC[N:N+4]}, vremainder_mask), wasm_i32x4_shr(vacc${ABC[N:N+4]}, 31));

    $for N in range(0, BATCH_TILE, 4):
      vacc${ABC[N:N+4]} = wasm_i32x4_sub(wasm_i32x4_shr(vacc${ABC[N:N+4]}, vshift), wasm_i32x4_gt(vrem${ABC[N:N+4]}, vremainder_threshold));

    $for N in range(0, BATCH_TILE, 8):
      v128_t vout${ABC[N:N+8]} = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc${ABC[N:N+4]}, vacc${ABC[N+4:N+8]}), voutput_zero_point);

    $for N in range(0, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        v128_t vout${ABC[N:N+16]} = wasm_i8x16_narrow_i16x8(vout${ABC[N:N+8]}, vout${ABC[N+8:N+16]});
      $else:
        v128_t vout${ABC[N:N+8]}${ABC[N:N+8]} = wasm_i8x16_narrow_i16x8(vout${ABC[N:N+8]}, vout${ABC[N:N+8]});

    $for N in range(0, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        vout${ABC[N:N+16]} = wasm_i8x16_max(vout${ABC[N:N+16]}, voutput_min);
      $else:
        vout${ABC[N:N+8]}${ABC[N:N+8]} = wasm_i8x16_max(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_min);

    $for N in range(0, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        vout${ABC[N:N+16]} = wasm_i8x16_min(vout${ABC[N:N+16]}, voutput_max);
      $else:
        vout${ABC[N:N+8]}${ABC[N:N+8]} = wasm_i8x16_min(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_max);

    $if BATCH_TILE >= 16:
      wasm_v128_store(output, vout${ABC[0:16]});
    $else:
      *((double*) output) = wasm_f64x2_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
    $for N in range(16, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        wasm_v128_store(output + ${N}, vout${ABC[N:N+16]});
      $else:
        *((double*) (output + ${N})) = wasm_f64x2_extract_lane(vout${ABC[N:N+8]}${ABC[N:N+8]}, 0);
    output += ${BATCH_TILE};
  }
  if XNN_UNLIKELY(n != 0) {
    ${"do " if BATCH_TILE > 8 else ""}{
      const v128_t vx${ABC[0:8]} = wasm_i16x8_load_8x8(input_x);
      const v128_t vy${ABC[0:8]} = wasm_i16x8_load_8x8(input_y);
      $if BATCH_TILE > 8:
        input_x += 8;
        input_y += 8;

      v128_t vacc${ABC[0:4]} = wasm_i32x4_add(vzero_point_product, wasm_i32x4_mul(wasm_i32x4_widen_low_i16x8(vx${ABC[0:8]}), vx_multiplier));
      v128_t vacc${ABC[4:8]} = wasm_i32x4_add(vzero_point_product, wasm_i32x4_mul(wasm_i32x4_widen_high_i16x8(vx${ABC[0:8]}), vx_multiplier));

      vacc${ABC[0:4]} = wasm_i32x4_add(vacc${ABC[0:4]}, wasm_i32x4_mul(wasm_i32x4_widen_low_i16x8(vy${ABC[0:8]}), vy_multiplier));
      vacc${ABC[4:8]} = wasm_i32x4_add(vacc${ABC[4:8]}, wasm_i32x4_mul(wasm_i32x4_widen_high_i16x8(vy${ABC[0:8]}), vy_multiplier));

      const v128_t vrem${ABC[0:4]} = wasm_i32x4_add(wasm_v128_and(vacc${ABC[0:4]}, vremainder_mask), wasm_i32x4_shr(vacc${ABC[0:4]}, 31));
      const v128_t vrem${ABC[4:8]} = wasm_i32x4_add(wasm_v128_and(vacc${ABC[4:8]}, vremainder_mask), wasm_i32x4_shr(vacc${ABC[4:8]}, 31));

      vacc${ABC[0:4]} = wasm_i32x4_sub(wasm_i32x4_shr(vacc${ABC[0:4]}, vshift), wasm_i32x4_gt(vrem${ABC[0:4]}, vremainder_threshold));
      vacc${ABC[4:8]} = wasm_i32x4_sub(wasm_i32x4_shr(vacc${ABC[4:8]}, vshift), wasm_i32x4_gt(vrem${ABC[4:8]}, vremainder_threshold));

      v128_t vout${ABC[0:8]} = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc${ABC[0:4]}, vacc${ABC[4:8]}), voutput_zero_point);

      v128_t vout${ABC[0:8]}${ABC[0:8]} = wasm_i8x16_narrow_i16x8(vout${ABC[0:8]}, vout${ABC[0:8]});
      vout${ABC[0:8]}${ABC[0:8]} = wasm_i8x16_max(vout${ABC[0:8]}${ABC[0:8]}, voutput_min);
      vout${ABC[0:8]}${ABC[0:8]} = wasm_i8x16_min(vout${ABC[0:8]}${ABC[0:8]}, voutput_max);

      $if BATCH_TILE > 8:
        if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
          *((double*) output) = wasm_f64x2_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
          output += 8;
          n -= 8 * sizeof(int8_t);
        } else {
          if (n & (4 * sizeof(int8_t))) {
            *((uint32_t*) output) = (uint32_t) wasm_i32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
            vout${ABC[0:8]}${ABC[0:8]} = wasm_u64x2_shr(vout${ABC[0:8]}${ABC[0:8]}, 32);
            output += 4;
          }
          if (n & (2 * sizeof(int8_t))) {
            *((uint16_t*) output) = (uint16_t) wasm_i16x8_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
            vout${ABC[0:8]}${ABC[0:8]} = wasm_u32x4_shr(vout${ABC[0:8]}${ABC[0:8]}, 16);
            output += 2;
          }
          if (n & (1 * sizeof(int8_t))) {
            *output = wasm_i8x16_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
          }
          n = 0;
        }
      $else:
        if (n & (4 * sizeof(int8_t))) {
          *((uint32_t*) output) = (uint32_t) wasm_i32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
          vout${ABC[0:8]}${ABC[0:8]} = wasm_u64x2_shr(vout${ABC[0:8]}${ABC[0:8]}, 32);
          output += 4;
        }
        if (n & (2 * sizeof(int8_t))) {
          *((uint16_t*) output) = (uint16_t) wasm_i16x8_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
          vout${ABC[0:8]}${ABC[0:8]} = wasm_u32x4_shr(vout${ABC[0:8]}${ABC[0:8]}, 16);
          output += 2;
        }
        if (n & (1 * sizeof(int8_t))) {
          *output = wasm_i8x16_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
        }
    }${" while (n != 0);" if BATCH_TILE > 8 else ""}
  }
}