// 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_int32_x${BATCH_TILE}(
    size_t batch,
    const void* input,
    float* output,
    const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const uint32x4_t vsign_mask = vmovq_n_u32(0x80000000);
  const uint32x4_t vexp_offset = vmovq_n_u32(0x70000000);
  const float32x4_t vexp_scale = vld1q_dup_f32(&params->neon.exp_scale);
  const uint32x4_t vmagic_bias = vmovq_n_u32(0x3F000000);
  const uint32x4_t vdenorm_cutoff = vmovq_n_u32(0x04000000);

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

      $for N in range(SIMD_TILE):
        const uint32x4_t vw${2*N} = vshll_n_u16(vget_low_u16(vh${N}), 16);
        const uint32x4_t vw${2*N+1} = vshll_n_u16(vget_high_u16(vh${N}), 16);

      $for N in range(2*SIMD_TILE):
        const uint32x4_t vsign${N} = vandq_u32(vw${N}, vsign_mask);

      $for N in range(2*SIMD_TILE):
        const uint32x4_t vnonsign${N} = veorq_u32(vw${N}, vsign${N});

      $for N in range(2*SIMD_TILE):
        const float32x4_t vnorm${N} = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign${N}, 3)), vexp_scale);

      $for N in range(2*SIMD_TILE):
        const float32x4_t vdenorm${N} = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign${N}, 16)), vreinterpretq_f32_u32(vmagic_bias));

      $for N in range(2*SIMD_TILE):
        const uint32x4_t vxmask${N} = vcgtq_u32(vnonsign${N}, vdenorm_cutoff);

      $for N in range(2*SIMD_TILE):
        const uint32x4_t vf${N} = vorrq_u32(vsign${N}, vreinterpretq_u32_f32(vbslq_f32(vxmask${N}, vnorm${N}, vdenorm${N})));

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

    const uint32x4_t vw_lo = vshll_n_u16(vget_low_u16(vh), 16);
    const uint32x4_t vw_hi = vshll_n_u16(vget_high_u16(vh), 16);

    const uint32x4_t vsign_lo = vandq_u32(vw_lo, vsign_mask);
    const uint32x4_t vsign_hi = vandq_u32(vw_hi, vsign_mask);

    const uint32x4_t vnonsign_lo = veorq_u32(vw_lo, vsign_lo);
    const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi);

    const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_lo, 3)), vexp_scale);
    const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scale);

    const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_lo, 16)), vreinterpretq_f32_u32(vmagic_bias));
    const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinterpretq_f32_u32(vmagic_bias));

    const uint32x4_t vxmask_lo = vcgtq_u32(vnonsign_lo, vdenorm_cutoff);
    const uint32x4_t vf_lo = vorrq_u32(vsign_lo, vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));

    const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff);
    const uint32x4_t vf_hi = vorrq_u32(vsign_hi, 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(batch != 0) {
    const uint16x8_t vh = vld1q_u16(i); i += 8;

    const uint32x4_t vw_lo = vshll_n_u16(vget_low_u16(vh), 16);
    const uint32x4_t vw_hi = vshll_n_u16(vget_high_u16(vh), 16);

    const uint32x4_t vsign_lo = vandq_u32(vw_lo, vsign_mask);
    const uint32x4_t vsign_hi = vandq_u32(vw_hi, vsign_mask);

    const uint32x4_t vnonsign_lo = veorq_u32(vw_lo, vsign_lo);
    const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi);

    const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_lo, 3)), vexp_scale);
    const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scale);

    const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_lo, 16)), vreinterpretq_f32_u32(vmagic_bias));
    const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinterpretq_f32_u32(vmagic_bias));

    const uint32x4_t vxmask_lo = vcgtq_u32(vnonsign_lo, vdenorm_cutoff);
    uint32x4_t vf = vorrq_u32(vsign_lo, vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));

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

      const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff);
      vf = vorrq_u32(vsign_hi, vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi)));
    }
    uint32x2_t vf_lo = vget_low_u32(vf);
    if (batch & (2 * sizeof(uint16_t))) {
      vst1_f32(output, vreinterpret_f32_u32(vf_lo)); output += 2;
      vf_lo = vget_high_u32(vf);
    }
    if (batch & (1 * sizeof(uint16_t))) {
      vst1_lane_f32(output, vreinterpret_f32_u32(vf_lo), 0);
    }
  }
}