// 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
$SIMD_TILE = BATCH_TILE // 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/common.h>
#include <xnnpack/vcvt.h>


void xnn_f16_f32_vcvt_ukernel__neon_int16_x${BATCH_TILE}(
    size_t n,
    const void* input,
    float* output,
    const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const uint16x8_t vsign_mask = vmovq_n_u16(0x8000);
  const uint16x8_t vexp_offset = vmovq_n_u16(0x7000);
  const float32x4_t vexp_scale = vld1q_dup_f32(&params->neon.exp_scale);
  const uint32x4_t vmagic_bias = vmovq_n_u32(0x3F000000);
  const uint16x8_t vdenorm_cutoff = vmovq_n_u16(0x0400);

  const uint16_t* i = (const uint16_t*) input;
  $if BATCH_TILE > 8:
    for (; n >= ${BATCH_TILE} * sizeof(uint16_t); n -= ${BATCH_TILE} * sizeof(uint16_t)) {
      $for N in range(SIMD_TILE):
        const uint16x8_t vh${N} = vld1q_u16(i); i += 8;

      $for N in range(SIMD_TILE):
        const uint16x8_t vsign${N} = vandq_u16(vh${N}, vsign_mask);

      $for N in range(SIMD_TILE):
        const uint16x8_t vnonsign${N} = veorq_u16(vh${N}, vsign${N});

      $for N in range(SIMD_TILE):
        const uint16x8x2_t vprenorm${N} = vzipq_u16(vshlq_n_u16(vnonsign${N}, 13), vsraq_n_u16(vexp_offset, vnonsign${N}, 3));

      $for N in range(SIMD_TILE):
        const float32x4_t vnorm${2*N} = vmulq_f32(vreinterpretq_f32_u16(vprenorm${N}.val[0]), vexp_scale);
        const float32x4_t vnorm${2*N+1} = vmulq_f32(vreinterpretq_f32_u16(vprenorm${N}.val[1]), vexp_scale);

      $for N in range(SIMD_TILE):
        const float32x4_t vdenorm${2*N} = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(vnonsign${N}))), vreinterpretq_f32_u32(vmagic_bias));
        const float32x4_t vdenorm${2*N+1} = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_high_u16(vnonsign${N}))), vreinterpretq_f32_u32(vmagic_bias));

      $for N in range(SIMD_TILE):
        const uint16x8_t vmask${N} = vcgtq_u16(vnonsign${N}, vdenorm_cutoff);

      $for N in range(SIMD_TILE):
        const uint32x4_t vxmask${2*N} = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_low_u16(vmask${N}))));
        const uint32x4_t vf${2*N} = vorrq_u32(vshll_n_u16(vget_low_u16(vsign${N}), 16),
          vreinterpretq_u32_f32(vbslq_f32(vxmask${2*N}, vnorm${2*N}, vdenorm${2*N})));

      $for N in range(SIMD_TILE):
        const uint32x4_t vxmask${2*N+1} = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_high_u16(vmask${N}))));
        const uint32x4_t vf${2*N+1} = vorrq_u32(vshll_n_u16(vget_high_u16(vsign${N}), 16),
          vreinterpretq_u32_f32(vbslq_f32(vxmask${2*N+1}, vnorm${2*N+1}, vdenorm${2*N+1})));

      $for N in range(2*SIMD_TILE):
        vst1q_f32(output, vreinterpretq_f32_u32(vf${N})); output += 4;
    }
  for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
    const uint16x8_t vh = vld1q_u16(i); i += 8;

    const uint16x8_t vsign = vandq_u16(vh, vsign_mask);

    const uint16x8_t vnonsign = veorq_u16(vh, vsign);

    const uint16x8x2_t vprenorm = vzipq_u16(vshlq_n_u16(vnonsign, 13), vsraq_n_u16(vexp_offset, vnonsign, 3));
    const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[0]), vexp_scale);
    const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[1]), vexp_scale);

    const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias));
    const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_high_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias));

    const uint16x8_t vmask = vcgtq_u16(vnonsign, vdenorm_cutoff);

    const uint32x4_t vxmask_lo = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_low_u16(vmask))));
    const uint32x4_t vf_lo = vorrq_u32(vshll_n_u16(vget_low_u16(vsign), 16),
      vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));

    const uint32x4_t vxmask_hi = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_high_u16(vmask))));
    const uint32x4_t vf_hi = vorrq_u32(vshll_n_u16(vget_high_u16(vsign), 16),
      vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi)));

    vst1q_f32(output, vreinterpretq_f32_u32(vf_lo)); output += 4;
    vst1q_f32(output, vreinterpretq_f32_u32(vf_hi)); output += 4;
  }
  if XNN_UNPREDICTABLE(n != 0) {
    const uint16x8_t vh = vld1q_u16(i); i += 8;

    const uint16x8_t vsign = vandq_u16(vh, vsign_mask);

    const uint16x8_t vnonsign = veorq_u16(vh, vsign);

    const uint16x8x2_t vprenorm = vzipq_u16(vshlq_n_u16(vnonsign, 13), vsraq_n_u16(vexp_offset, vnonsign, 3));
    const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[0]), vexp_scale);
    const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[1]), vexp_scale);

    const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias));
    const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_high_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias));

    const uint16x8_t vmask = vcgtq_u16(vnonsign, vdenorm_cutoff);

    const uint32x4_t vxmask_lo = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_low_u16(vmask))));
    uint32x4_t vf = vorrq_u32(vshll_n_u16(vget_low_u16(vsign), 16),
      vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));

    if (n & (4 * sizeof(uint16_t))) {
      vst1q_f32(output, vreinterpretq_f32_u32(vf)); output += 4;

      const uint32x4_t vxmask_hi = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_high_u16(vmask))));
      vf = vorrq_u32(vshll_n_u16(vget_high_u16(vsign), 16),
        vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi)));
    }
    uint32x2_t vf_lo = vget_low_u32(vf);
    if (n & (2 * sizeof(uint16_t))) {
      vst1_f32(output, vreinterpret_f32_u32(vf_lo)); output += 2;
      vf_lo = vget_high_u32(vf);
    }
    if (n & (1 * sizeof(uint16_t))) {
      vst1_lane_f32(output, vreinterpret_f32_u32(vf_lo), 0);
    }
  }
}