// 
// sub_103A80 from 0x103a80 to 0x103c94 (532 bytes)
// 
// 1 xrefs:
// > <???> @ 0x11f740
// 
// 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
// 

__int64 __fastcall sub_103A80(int8x16_t *a1, float32x4_t *a2, __int64 a3, int a4)
{
  unsigned int v4; // w9
  int v5; // w8
  int v6; // w9
  int v8; // w23
  int v9; // w9
  __int64 v10; // x23
  int v11; // w22
  __int64 v12; // x24
  __int64 v13; // x25
  int8x16_t v14; // q0
  int v15; // w22
  __int64 v16; // x23
  long double v17; // q0
  int16x8_t v18; // q0
  float32x4_t v19; // q1
  int8x16_t v20; // q3
  float32x4_t v21; // q4
  float32x4_t v22; // q5
  int16x8_t i; // [xsp+0h] [xbp-A0h]
  int16x8_t v25; // [xsp+10h] [xbp-90h]
  int16x8_t v26; // [xsp+20h] [xbp-80h]
  int16x8_t j; // [xsp+20h] [xbp-80h]
  float32x4_t v28; // [xsp+30h] [xbp-70h]
  float32x4_t v29; // [xsp+40h] [xbp-60h]
  int8x16_t v30; // [xsp+50h] [xbp-50h]

  v4 = a1[6].n128_u8[4];
  if ( v4 <= 7 )
  {
    v5 = a1[6].n128_u8[6];
    v6 = 1 << v4;
    v30 = *a1;
    v29 = a1[1];
    v8 = a4 >> 2;
    v28 = a1[2];
    if ( (v6 & 0xF0) != 0 )
    {
      v9 = v8 / 4;
      v26 = vdupq_n_s16(-v5);
      v25 = vdupq_n_s16(-2 * v5);
      v10 = v8 / 4;
      v11 = a4 >> 4;
      v12 = 2 * v9;
      v13 = 3 * v9;
      for ( i = vdupq_n_s16(-3 * v5); v11; ++a2 )
      {
        *(long double *)&v14 = sub_104768();
        --v11;
        *a2 = vaddq_f32(vmlaq_f32(v29, vcvtq_f32_u32(vandq_s8(v14, v30)), v28), *a2);
        a2[v10] = vaddq_f32(vmlaq_f32(v29, vcvtq_f32_u32(vandq_s8(v30, vshlq_s16(v14, v26))), v28), a2[v10]);
        a2[v12] = vaddq_f32(vmlaq_f32(v29, vcvtq_f32_u32(vandq_s8(v30, vshlq_s16(v14, v25))), v28), a2[v12]);
        a2[v13] = vaddq_f32(vmlaq_f32(v29, vcvtq_f32_u32(vandq_s8(v30, vshlq_s16(v14, i))), v28), a2[v13]);
      }
    }
    else if ( (v6 & 3) != 0 )
    {
      for ( ; v8; ++a2 )
      {
        --v8;
        *a2 = vmlaq_f32(vaddq_f32(*a2, v29), vcvtq_f32_u32(vandq_s8(COERCE_INT8X16_T(sub_104768()), v30)), v28);
      }
    }
    else
    {
      v15 = a4 >> 3;
      v16 = v8 / 2;
      for ( j = vdupq_n_s16(-v5); v15; ++a2 )
      {
        v17 = sub_104768();
        --v15;
        *(long double *)&v18 = sub_104804(v17, *(long double *)a2);
        *a2 = v19;
        a2[v16] = vmlaq_f32(vaddq_f32(a2[v16], v21), vcvtq_f32_u32(vandq_s8(v20, vshlq_s16(v18, j))), v22);
      }
    }
  }
  return 0;
}