// 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]
$WASM_X8X16_NARROW_I16X8 = {"QS8": "wasm_i8x16_narrow_i16x8", "QU8": "wasm_u8x16_narrow_i16x8"}[DATATYPE]
void xnn_f32_${DATATYPE.lower()}_vcvt_ukernel__wasmsimd_cvt_u${BATCH_TILE}(
    size_t batch,
    const float* input,
    ${XINT8_T}* output,
    const struct xnn_f32_${DATATYPE.lower()}_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const v128_t vscale = wasm_v128_load32_splat(&params->scalar.scale);
  const v128_t voutput_zero_point = wasm_v128_load16_splat(&params->scalar.output_zero_point);
  XNN_FORCE_REALIZATION(vscale);
  XNN_FORCE_REALIZATION(voutput_zero_point);
  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      v128_t vx${ABC[0:4]} = wasm_v128_load(input);
      $for N in range(4, BATCH_TILE, 4):
        v128_t vx${ABC[N:N+4]} = wasm_v128_load(input + ${N});
      input += ${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]});

      wasm_v128_store(output, vy${ABC[0:16]});
      $for N in range(16, BATCH_TILE, 16):
        $if N + 8 < BATCH_TILE:
          wasm_v128_store(output + ${N}, vy${ABC[N:N+16]});
        $else:
          wasm_v128_store64_lane(output + ${N}, vy${ABC[N:N+8]}, 0);
      output += ${BATCH_TILE};
    }
  for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
    v128_t vx_lo = wasm_v128_load(input);
    v128_t vx_hi = wasm_v128_load(input + 4);
    input += 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);

    wasm_v128_store64_lane(output, vy, 0);
    output += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch >= 1 * sizeof(float));
    assert(batch <= 7 * sizeof(float));
    v128_t vx_lo = wasm_v128_load(input);
    const float* x_hi = (const float*) ((uintptr_t) input + (batch & (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);

    if (batch & (4 * sizeof(float))) {
      wasm_v128_store32_lane(output, vy, 0);
      vy = wasm_u64x2_shr(vy, 32);
      output += 4;
    }
    if (batch & (2 * sizeof(float))) {
      wasm_v128_store16_lane(output, vy, 0);
      vy = wasm_u32x4_shr(vy, 16);
      output += 2;
    }
    if (batch & (1 * sizeof(float))) {
      wasm_v128_store8_lane(output, vy, 0);
    }
  }
}