// 
// sub_101D30 from 0x101d30 to 0x101ef4 (452 bytes)
// 
// 1 xrefs:
// > <???> @ 0x11a0f0
// 
// This function is hookable on this platform!
// 
// Using BromaIDA 8.0.0 @ https://github.com/Stazzical/BromaIDA
// Using bindings at commit ba9f177b at Thu Jul 16 22:52:06 2026 from https://github.com/geode-sdk/bindings
// 

float32x4_t *__fastcall sub_101D30(float32x4_t *result, __int64 a2, __int64 a3, __int64 a4, _QWORD *a5)
{
  __int64 v5; // x30
  float v6; // w8
  long double v7; // q0
  __int64 v8; // x5
  int v9; // w8
  float32x4_t v10; // q0
  int8x16_t v11; // q1
  int64x2_t v12; // q2
  int8x16_t v13; // q3
  float32x4_t v14; // q4
  unsigned __int64 v15; // d5
  __int64 v16; // x9
  int16x8_t v17; // q5
  float32x4_t v18; // q16
  float32x4_t v19; // q6
  int16x8_t v20; // q5
  unsigned __int64 v21; // d7
  float32x4_t v22; // q16
  int16x8_t v23; // q5
  __int64 v24; // x9
  float32x4_t v25; // q7
  int16x8_t v26; // q5
  float32x4_t v27; // q7
  float v28; // w9
  long double v29; // q0
  __int64 v30; // x0
  __int64 v31; // x1
  int v32; // w8

  v6 = 4.6566e-10;
  *(_QWORD *)&v7 = 0x3800000038000000LL;
  for ( *((_QWORD *)&v7 + 1) = 0x3800000038000000LL; (_DWORD)a2; *(_QWORD *)a4 += *a5 )
  {
    if ( ((unsigned __int8)result & 0xF) == 0 )
      break;
    result = (float32x4_t *)sub_102DB8(
                              v7,
                              v6,
                              (float)(*(_DWORD *)a4 >> 1),
                              *(double *)(a3 + 2LL * (unsigned int)(4 * *(_DWORD *)(a4 + 4))),
                              *(double *)(a3 + 2LL * (unsigned int)(4 * *(_DWORD *)(a4 + 4) + 4)));
    v5 = v8;
  }
  if ( (int)a2 >> 2 )
  {
    for ( result = (float32x4_t *)sub_102CA4(result, a2, a3, a4, a5, v5); v9; result += 4 )
    {
      v15 = vshl_n_s32(vzip2_s32((int32x2_t)v11.n128_u64[0], (int32x2_t)vextq_s8(v11, v11, 8u).n128_u64[0]), 2u).n64_u64[0];
      v16 = HIDWORD(v15);
      v17 = *(int16x8_t *)(a3 + 2LL * (unsigned int)v15);
      v18 = vcvtq_f32_s32(vmovl_s16((int16x4_t)v17.n128_u64[0]));
      v19 = vmulq_f32(vcvtq_f32_u32(vshrq_n_u32(vuzp1q_s32(v11, v13), 1u)), v10);
      *result = vmulq_f32(vmlaq_n_f32(v18, vsubq_f32(vcvtq_f32_s32(vmovl_high_s16(v17)), v18), v19.n128_f32[0]), v14);
      v20 = *(int16x8_t *)(a3 + 2 * v16);
      v21 = vshl_n_s32(vzip2_s32((int32x2_t)v13.n128_u64[0], (int32x2_t)vextq_s8(v13, v13, 8u).n128_u64[0]), 2u).n64_u64[0];
      v22 = vcvtq_f32_s32(vmovl_s16((int16x4_t)v20.n128_u64[0]));
      result[1] = vmulq_f32(
                    vmlaq_lane_f32(
                      v22,
                      vsubq_f32(vcvtq_f32_s32(vmovl_high_s16(v20)), v22),
                      (float32x2_t)v19.n128_u64[0],
                      1),
                    v14);
      v23 = *(int16x8_t *)(a3 + 2LL * (unsigned int)v21);
      v24 = 2LL * HIDWORD(v21);
      --v9;
      v25 = vcvtq_f32_s32(vmovl_s16((int16x4_t)v23.n128_u64[0]));
      result[2] = vmulq_f32(vmlaq_laneq_f32(v25, vsubq_f32(vcvtq_f32_s32(vmovl_high_s16(v23)), v25), v19, 2), v14);
      v26 = *(int16x8_t *)(a3 + v24);
      v11 = vaddq_s64(v11, v12);
      v13 = vaddq_s64(v13, v12);
      v27 = vcvtq_f32_s32(vmovl_s16((int16x4_t)v26.n128_u64[0]));
      result[3] = vmulq_f32(vmlaq_laneq_f32(v27, vsubq_f32(vcvtq_f32_s32(vmovl_high_s16(v26)), v27), v19, 3), v14);
    }
    *(_QWORD *)a4 = v11.n128_u64[0];
  }
  v28 = 4.6566e-10;
  *(_QWORD *)&v29 = 0x3800000038000000LL;
  *((_QWORD *)&v29 + 1) = 0x3800000038000000LL;
  if ( (a2 & 3) != 0 )
  {
    do
    {
      v30 = sub_102DB8(
              v29,
              v28,
              (float)(*(_DWORD *)a4 >> 1),
              *(double *)(a3 + 2LL * (unsigned int)(4 * *(_DWORD *)(a4 + 4))),
              *(double *)(a3 + 2LL * (unsigned int)(4 * *(_DWORD *)(a4 + 4) + 4)));
      result = (float32x4_t *)sub_102E10(v30, v31);
    }
    while ( v32 );
  }
  return result;
}