// 
// sub_101BE4 from 0x101be4 to 0x101d30 (332 bytes)
// 
// 1 xrefs:
// > <???> @ 0x11a0e8
// 
// 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_101BE4(float32x4_t *result, __int64 a2, __int64 a3, __int64 a4, __int64 a5)
{
  __int64 i; // x30
  int v6; // w8
  float32x4_t v7; // q0
  int32x4_t v8; // q1
  int64x2_t v9; // q2
  int32x4_t v10; // q3
  float32x4_t v11; // q4
  __int64 v12; // x5
  int64x2_t v13; // q6
  int64x2_t v14; // q7
  int16x8_t v15; // q16
  int16x8_t v16; // q6
  int32x4_t v17; // q5
  float32x4_t v18; // q7
  float32x4_t v19; // q16
  float v20; // w9
  float v21; // s0
  float32x2_t v22; // d1
  unsigned int v23; // w10
  __int64 v24; // x1
  int v25; // w8

  if ( (int)a2 >> 2 )
  {
    result = (float32x4_t *)sub_102CA4(result, a2, a3, a4, a5, i);
    for ( i = v12; v6; result += 2 )
    {
      v13 = vzip1q_s32(*(int32x4_t *)(a3 + 4LL * v8.n128_u32[1]), *(int32x4_t *)(a3 + 4LL * v8.n128_u32[3]));
      v14 = vzip1q_s32(*(int32x4_t *)(a3 + 4LL * v10.n128_u32[1]), *(int32x4_t *)(a3 + 4LL * v10.n128_u32[3]));
      v15 = vzip1q_s64(v13, v14);
      v16 = vzip2q_s64(v13, v14);
      v17 = vmulq_f32(vcvtq_f32_u32(vshrq_n_u32(vuzp1q_s32(v8, v10), 1u)), v7);
      v18 = vcvtq_f32_s32(vmovl_s16((int16x4_t)v15.n128_u64[0]));
      v19 = vcvtq_f32_s32(vmovl_high_s16(v15));
      v8 = vaddq_s64(v8, v9);
      v10 = vaddq_s64(v10, v9);
      --v6;
      *result = vmulq_f32(
                  vmlaq_f32(
                    v18,
                    vzip1q_s32(v17, v17),
                    vsubq_f32(vcvtq_f32_s32(vmovl_s16((int16x4_t)v16.n128_u64[0])), v18)),
                  v11);
      result[1] = vmulq_f32(
                    vmlaq_f32(v19, vzip2q_s32(v17, v17), vsubq_f32(vcvtq_f32_s32(vmovl_high_s16(v16)), v19)),
                    v11);
    }
    *(_QWORD *)a4 = v8.n128_u64[0];
  }
  v20 = 2.3283e-10;
  v21 = 1.0;
  v22.n64_u64[0] = 0x3800000038000000LL;
  if ( (a2 & 3) != 0 )
  {
    do
    {
      v23 = 2 * *(_DWORD *)(a4 + 4);
      result->n128_u64[0] = vmul_f32(
                              vmla_n_f32(
                                vmul_n_f32(
                                  vcvt_f32_s32((int32x2_t)__PAIR64__(
                                                            *(__int16 *)(a3 + 2LL * v23 + 2),
                                                            *(__int16 *)(a3 + 2LL * v23))),
                                  v21 - (float)((float)*(unsigned int *)a4 * v20)),
                                vcvt_f32_s32((int32x2_t)__PAIR64__(
                                                          *(__int16 *)(a3 + 2LL * (v23 + 2) + 2),
                                                          *(__int16 *)(a3 + 2LL * (v23 + 2)))),
                                (float)*(unsigned int *)a4 * v20),
                              v22).n64_u64[0];
      result = (float32x4_t *)sub_102E10(&result->n128_i8[8], i);
      i = v24;
    }
    while ( v25 );
  }
  return result;
}