// 
// sub_103E84 from 0x103e84 to 0x10430c (1160 bytes)
// 
// 1 xrefs:
// > <???> @ 0x11f738
// 
// 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_103E84(__int64 a1, __int64 a2, __int64 a3, int a4)
{
  int v4; // w8
  unsigned __int64 v5; // x22
  __int64 v7; // x20
  unsigned int v9; // w23
  int v10; // w23
  long double v11; // q0
  float32x4_t v12; // q0
  __int64 v13; // x0
  __int64 v14; // x1
  int32x4_t v15; // q0
  int32x4_t v16; // q1
  int32x4_t v17; // q16
  uint32x4_t v18; // q3
  uint32x4_t v19; // q1
  int32x4_t v20; // q2
  uint32x4_t v21; // q16
  float32x4_t v22; // q4
  float32x4_t v23; // q5
  float32x4_t v24; // q7
  __int64 v25; // x23
  __int64 v26; // x22
  __int64 v27; // x0
  __int64 v28; // x1
  long double v29; // q0
  __int128 v30; // q1
  long double v31; // q0
  __int128 v32; // q1
  __int64 v33; // x8
  int v34; // w23
  int8x16_t v35; // q0
  int32x4_t v36; // q2
  int32x4_t v37; // q0
  __int64 v38; // x23
  __int64 v39; // x22
  int8x16_t v40; // q0
  float32x4_t v41; // q2
  int v42; // w22
  __int64 j; // x0
  int32x4_t v44; // q0
  int32x4_t v45; // q1
  int32x4_t v46; // q19
  uint32x4_t v47; // q3
  uint32x4_t v48; // q1
  int8x16_t v49; // q5
  int8x16_t v50; // q4
  int32x4_t v51; // q19
  int8x16_t v52; // q7
  int8x16_t v53; // q6
  int32x4_t v54; // q5
  uint32x4_t v55; // q7
  uint32x4_t v56; // q5
  int32x4_t v57; // q2
  uint32x4_t v58; // q19
  uint32x4_t v59; // q0
  uint32x4_t v60; // q2
  uint32x4_t v61; // q4
  float32x4_t v62; // q20
  float32x4_t v63; // q18
  float32x4_t v64; // q6
  float32x4_t v65; // q3
  float32x4_t v66; // q7
  float32x4_t v67; // q16
  float32x4_t v68; // q17
  float32x4_t v69; // q1
  float32x4_t v70; // q5
  float32x4_t v71; // q2
  float32x4_t v72; // q19
  float32x4_t v73; // q4
  float32x4_t v74; // q0
  float32x4_t v75; // q2
  float32x4_t v76; // q3
  float32x4_t v77; // q6
  float32x4_t v78; // q7
  int k; // w22
  float32x4_t v80; // q0
  int16x8_t i; // [xsp+40h] [xbp-B0h]
  float32x4_t v83; // [xsp+50h] [xbp-A0h]
  long double v84; // [xsp+50h] [xbp-A0h]
  int8x16_t v85; // [xsp+50h] [xbp-A0h]
  float32x4_t v86; // [xsp+60h] [xbp-90h]
  float32x4_t v87; // [xsp+60h] [xbp-90h]
  float32x4_t v88; // [xsp+60h] [xbp-90h]
  float32x4_t v89; // [xsp+70h] [xbp-80h]
  float32x4_t v90; // [xsp+80h] [xbp-70h]
  long double v91; // [xsp+90h] [xbp-60h]
  int8x16_t v92; // [xsp+A0h] [xbp-50h]
  int8x16_t v93; // [xsp+B0h] [xbp-40h]

  v4 = *(unsigned __int8 *)(a1 + 102);
  v91 = *(long double *)a1;
  v5 = (unsigned int)a4;
  v7 = a1;
  v90 = *(float32x4_t *)(a1 + 16);
  v9 = a4 >> 2;
  v89 = *(float32x4_t *)(a1 + 32);
  switch ( *(_BYTE *)(a1 + 100) )
  {
    case 0:
    case 1:
      while ( v9 )
      {
        v31 = sub_104768(a1);
        --v9;
        a1 = sub_10479C(v31, *(long double *)a2, v91);
        *(_OWORD *)a2 = v32;
        a2 = v33;
      }
      break;
    case 2:
    case 3:
      v34 = a4 >> 3;
      for ( i = vdupq_n_s16(-v4); v34; a2 += 32 )
      {
        v84 = *(long double *)a2;
        v87 = *(float32x4_t *)(a2 + 16);
        v35.n128_u64[0] = sub_103C94(v7, a3);
        --v34;
        v36 = vmlaq_f32(v90, vcvtq_f32_u32(vandq_s8(v35, *(int8x16_t *)&v91)), v89);
        v37 = vmlaq_f32(v90, vcvtq_f32_u32(vandq_s8(*(int8x16_t *)&v91, vshlq_s16(v35, i))), v89);
        *(float32x4_t *)a2 = vaddq_f32(vzip1q_s32(v36, v37), *(float32x4_t *)&v84);
        *(float32x4_t *)(a2 + 16) = vaddq_f32(vzip2q_s32(v36, v37), v87);
      }
      v38 = 0;
      v39 = (v5 >> 2) & 1;
      while ( (_DWORD)v39 != (_DWORD)v38 )
      {
        v88 = COERCE_FLOAT32X4_T(sub_104774(*(long double *)(a2 + 16 * v38)));
        v85 = COERCE_INT8X16_T(((long double (*)(void))sub_10430C)());
        *(long double *)&v40 = sub_10430C(v7, a3);
        v41.n128_u64[0] = vmlaq_f32(v90, vcvtq_f32_u32(vandq_s8(v85, *(int8x16_t *)&v91)), v89).n128_u64[0];
        v41.n128_u64[1] = vmlaq_f32(v90, vcvtq_f32_u32(vandq_s8(v40, *(int8x16_t *)&v91)), v89).n128_u64[0];
        *(float32x4_t *)(a2 + 16 * v38++) = vaddq_f32(v41, v88);
      }
      break;
    case 4:
    case 5:
    case 6:
    case 7:
      sub_104780();
      v10 = (int)v5 >> 4;
      sub_1047CC(v11);
      if ( (int)v5 >> 4 )
      {
        do
        {
          sub_104774(*(long double *)a2);
          v83 = *(float32x4_t *)(a2 + 32);
          v86 = *(float32x4_t *)(a2 + 48);
          sub_103C94(v13, v14);
          *(long double *)&v15 = sub_1047F0();
          v18 = vuzp1q_s32(v17, v16);
          v19 = vuzp2q_s32(v17, v16);
          v21 = vuzp1q_s32(v15, v20);
          v12 = vcvtq_f32_u32(vuzp2q_s32(v15, v20));
          *(float32x4_t *)(a2 + 32) = vmlaq_f32(vaddq_f32(v83, v24), vcvtq_f32_u32(v18), v89);
          *(float32x4_t *)(a2 + 48) = vmlaq_f32(vaddq_f32(v86, v24), vcvtq_f32_u32(v19), v89);
          *(float32x4_t *)a2 = vmlaq_f32(vaddq_f32(v22, v24), vcvtq_f32_u32(v21), v89);
          *(float32x4_t *)(a2 + 16) = vmlaq_f32(vaddq_f32(v23, v24), v12, v89);
          a2 += 64;
          --v10;
        }
        while ( v10 );
      }
      v25 = 0;
      v26 = (v5 >> 2) & 3;
      while ( (_DWORD)v26 != (_DWORD)v25 )
      {
        sub_104774(*(long double *)&v12);
        v29 = sub_10430C(v27, v28);
        sub_10479C(v29, *(long double *)(a2 + 16 * v25), v91);
        *(_OWORD *)(a2 + 16 * v25++) = v30;
      }
      break;
    case 8:
      v42 = a4 >> 5;
      for ( j = sub_1047CC(COERCE_LONG_DOUBLE(vdupq_n_s16(-6 * v4))); v42; a2 += 128 )
      {
        sub_104768(j);
        *(long double *)&v44 = sub_1047F0();
        v47 = vuzp1q_s32(v46, v45);
        v48 = vuzp2q_s32(v46, v45);
        v51 = vextq_s8(v50, vextq_s8(v49, v50, 8u), 8u);
        v50.n128_u64[1] = v49.n128_u64[0];
        v54 = vextq_s8(v53, vextq_s8(v52, v53, 8u), 8u);
        v53.n128_u64[1] = v52.n128_u64[0];
        v55 = vuzp1q_s32(v51, v54);
        v56 = vuzp2q_s32(v51, v54);
        v58 = vuzp1q_s32(v44, v57);
        v59 = vuzp2q_s32(v44, v57);
        v60 = vuzp1q_s32(v50, v53);
        v61 = vuzp2q_s32(v50, v53);
        v63 = vmlaq_f32(v62, vcvtq_f32_u32(v47), v89);
        v64 = vmlaq_f32(v62, vcvtq_f32_u32(v55), v89);
        v65 = vmlaq_f32(v62, vcvtq_f32_u32(v48), v89);
        v66 = vmlaq_f32(v62, vcvtq_f32_u32(v56), v89);
        v67 = vmlaq_f32(v62, vcvtq_f32_u32(v59), v89);
        v68 = vmlaq_f32(v62, vcvtq_f32_u32(v61), v89);
        v69 = vmlaq_f32(vaddq_f32(*(float32x4_t *)a2, v62), vcvtq_f32_u32(v58), v89);
        v70 = vmlaq_f32(vaddq_f32(*(float32x4_t *)(a2 + 16), v62), vcvtq_f32_u32(v60), v89);
        v71 = *(float32x4_t *)(a2 + 96);
        v72 = *(float32x4_t *)(a2 + 112);
        v73 = vaddq_f32(v64, *(float32x4_t *)(a2 + 80));
        v74 = vaddq_f32(v63, *(float32x4_t *)(a2 + 64));
        *(float32x4_t *)a2 = v69;
        *(float32x4_t *)(a2 + 16) = v70;
        v75 = vaddq_f32(v65, v71);
        v76 = vaddq_f32(v66, v72);
        v77 = *(float32x4_t *)(a2 + 32);
        v78 = *(float32x4_t *)(a2 + 48);
        *(float32x4_t *)(a2 + 64) = v74;
        *(float32x4_t *)(a2 + 80) = v73;
        *(float32x4_t *)(a2 + 96) = v75;
        *(float32x4_t *)(a2 + 112) = v76;
        --v42;
        *(float32x4_t *)(a2 + 32) = vaddq_f32(v67, v77);
        *(float32x4_t *)(a2 + 48) = vaddq_f32(v68, v78);
      }
      for ( k = (v9 >> 1) & 3; k; a2 += 32 )
      {
        j = sub_1047B8(j);
        --k;
        v92 = vandq_s8(v92, *(int8x16_t *)&v91);
        v93 = vandq_s8(v93, *(int8x16_t *)&v91);
        v80 = vaddq_f32(vmlaq_f32(v90, vcvtq_f32_u32(v92), v89), *(float32x4_t *)(a2 + 16));
        *(float32x4_t *)a2 = vaddq_f32(vmlaq_f32(v90, vcvtq_f32_u32(v93), v89), *(float32x4_t *)a2);
        *(float32x4_t *)(a2 + 16) = v80;
      }
      break;
    default:
      return 0;
  }
  return 0;
}