// Copyright 2021 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 <arm_neon.h>

#include <xnnpack/common.h>
#include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/vcvt.h>


$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
$WASM_X8X16_NARROW_I16X8 = {"QS8": "wasm_i8x16_narrow_i16x8", "QU8": "wasm_u8x16_narrow_i16x8"}[DATATYPE]
$WASM_X8X16_MIN = {"QS8": "wasm_i8x16_min", "QU8": "wasm_u8x16_min"}[DATATYPE]
$WASM_X8X16_MAX = {"QS8": "wasm_i8x16_max", "QU8": "wasm_u8x16_max"}[DATATYPE]
void xnn_f32_${DATATYPE.lower()}_vcvt_ukernel__wasmsimd_cvt_x${BATCH_TILE}(
    size_t n,
    const float* x,
    ${XINT8_T}* y,
    const union xnn_f32_${DATATYPE.lower()}_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);
  assert(x != NULL);
  assert(y != NULL);

  const v128_t vscale = wasm_v128_load64_splat(params->wasmsimd_cvt.scale);
  const v128_t voutput_zero_point = wasm_v128_load64_splat(params->wasmsimd_cvt.output_zero_point);
  const v128_t voutput_min = wasm_v128_load64_splat(params->wasmsimd_cvt.output_min);
  const v128_t voutput_max = wasm_v128_load64_splat(params->wasmsimd_cvt.output_max);
  $if BATCH_TILE > 8:
    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
      v128_t vx${ABC[0:4]} = wasm_v128_load(x);
      $for N in range(4, BATCH_TILE, 4):
        v128_t vx${ABC[N:N+4]} = wasm_v128_load(x + ${N});
      x += ${BATCH_TILE};

      $for N in range(0, BATCH_TILE, 4):
        vx${ABC[N:N+4]} = wasm_f32x4_mul(vx${ABC[N:N+4]}, vscale);

      $for N in range(0, BATCH_TILE, 4):
        vx${ABC[N:N+4]} = wasm_f32x4_nearest(vx${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        v128_t vacc${ABC[N:N+4]} = wasm_i32x4_trunc_sat_f32x4(vx${ABC[N:N+4]});

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

      $for N in range(0, BATCH_TILE, 8):
        vacc${ABC[N:N+8]} = wasm_i16x8_add_sat(vacc${ABC[N:N+8]}, voutput_zero_point);

      $for N in range(0, BATCH_TILE, 16):
        $if N + 8 < BATCH_TILE:
          v128_t vy${ABC[N:N+16]} = ${WASM_X8X16_NARROW_I16X8}(vacc${ABC[N:N+8]}, vacc${ABC[N+8:N+16]});
        $else:
          v128_t vy${ABC[N:N+8]} = ${WASM_X8X16_NARROW_I16X8}(vacc${ABC[N:N+8]}, vacc${ABC[N:N+8]});

      $for N in range(0, BATCH_TILE, 16):
        $if N + 8 < BATCH_TILE:
          vy${ABC[N:N+16]} = ${WASM_X8X16_MAX}(vy${ABC[N:N+16]}, voutput_min);
        $else:
          vy${ABC[N:N+8]} = ${WASM_X8X16_MAX}(vy${ABC[N:N+8]}, voutput_min);

      $for N in range(0, BATCH_TILE, 16):
        $if N + 8 < BATCH_TILE:
          vy${ABC[N:N+16]} = ${WASM_X8X16_MIN}(vy${ABC[N:N+16]}, voutput_max);
        $else:
          vy${ABC[N:N+8]} = ${WASM_X8X16_MIN}(vy${ABC[N:N+8]}, voutput_max);

      wasm_v128_store(y, vy${ABC[0:16]});
      $for N in range(16, BATCH_TILE, 16):
        $if N + 8 < BATCH_TILE:
          wasm_v128_store(y + ${N}, vy${ABC[N:N+16]});
        $else:
          *((double*) (y + ${N})) = wasm_f64x2_extract_lane(vy${ABC[N:N+8]}, 0);
      y += ${BATCH_TILE};
    }
  for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
    v128_t vx_lo = wasm_v128_load(x);
    v128_t vx_hi = wasm_v128_load(x + 4);
    x += 8;

    vx_lo = wasm_f32x4_mul(vx_lo, vscale);
    vx_hi = wasm_f32x4_mul(vx_hi, vscale);

    vx_lo = wasm_f32x4_nearest(vx_lo);
    vx_hi = wasm_f32x4_nearest(vx_hi);

    v128_t vacc_lo = wasm_i32x4_trunc_sat_f32x4(vx_lo);
    v128_t vacc_hi = wasm_i32x4_trunc_sat_f32x4(vx_hi);

    v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);
    vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point);

    v128_t vy = ${WASM_X8X16_NARROW_I16X8}(vacc, vacc);
    vy = ${WASM_X8X16_MAX}(vy, voutput_min);
    vy = ${WASM_X8X16_MIN}(vy, voutput_max);

    *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
    y += 8;
  }
  if XNN_UNLIKELY(n != 0) {
    assert(n >= 1 * sizeof(float));
    assert(n <= 7 * sizeof(float));
    v128_t vx_lo = wasm_v128_load(x);
    const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
    v128_t vx_hi = wasm_v128_load(x_hi);

    vx_lo = wasm_f32x4_mul(vx_lo, vscale);
    vx_hi = wasm_f32x4_mul(vx_hi, vscale);

    vx_lo = wasm_f32x4_nearest(vx_lo);
    vx_hi = wasm_f32x4_nearest(vx_hi);

    v128_t vacc_lo = wasm_i32x4_trunc_sat_f32x4(vx_lo);
    v128_t vacc_hi = wasm_i32x4_trunc_sat_f32x4(vx_hi);

    v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);
    vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point);

    v128_t vy = ${WASM_X8X16_NARROW_I16X8}(vacc, vacc);
    vy = ${WASM_X8X16_MAX}(vy, voutput_min);
    vy = ${WASM_X8X16_MIN}(vy, voutput_max);

    if (n & (4 * sizeof(float))) {
      *((float*) y) = wasm_f32x4_extract_lane(vy, 0);
      y += 4;
      vy = wasm_u64x2_shr(vy, 32);
    }
    uint32_t vy_lo = (uint32_t) wasm_i32x4_extract_lane(vy, 0);
    if (n & (2 * sizeof(float))) {
      *((uint16_t*) y) = (uint16_t) vy_lo;
      y += 2;
      vy_lo >>= 16;
    }
    if (n & (1 * sizeof(float))) {
      *y = (${XINT8_T}) vy_lo;
    }
  }
}