// 
// sub_103050 from 0x103050 to 0x103a14 (2500 bytes)
// 
// 3 xrefs:
// > sub_103A14 @ 0x103a20
// > sub_103A28 @ 0x103a34
// > sub_103A3C @ 0x103a48
// 
// 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_103050(__int64 a1, __int64 a2, __int64 a3, const float *a4, float *a5)
{
  bool v5; // w8
  long double v6; // q0
  long double v7; // q1
  long double v8; // q2
  long double v9; // q3
  long double v10; // q4
  __int64 result; // x0
  __int64 v12; // x1
  int v13; // w2
  __int64 v14; // x3
  unsigned __int32 *v15; // x4
  int v16; // w6
  int v17; // w7
  __int64 v18; // x8
  int v19; // w9
  __int64 v20; // x10
  __int64 v21; // x11
  __int64 v22; // x12
  __int64 v23; // x13
  __int64 v24; // x14
  __int64 v25; // x15
  __int64 v26; // x16
  __int64 v27; // x17
  float32x4_t v28; // q0
  float32x4_t v29; // q1
  float32x4_t v30; // q2
  float32x4_t v31; // q3
  float32x4_t v32; // q4
  float32x4_t v33; // q5
  float32x4_t v34; // q6
  float32x4_t v35; // q7
  float32x4_t v36; // q16
  float32x4_t v37; // q17
  __int64 v38; // x5
  __int64 v39; // x5
  int v40; // w19
  __int64 i; // x20
  _DWORD *v42; // x21
  int v43; // w22
  float32x4_t v44; // q24
  float32x4_t v45; // q26
  float32x4_t v46; // q18
  float32x4_t v47; // q21
  float32x4_t v48; // q19
  float32x4_t v49; // q25
  float32x4_t v50; // q20
  float32x4_t v51; // q23
  __int64 v52; // x20
  float32x4_t v53; // q22
  float32x4_t v54; // q27
  float32x4_t v55; // q28
  float32x4_t v56; // q29
  float32x4_t v57; // q30
  int j; // w21
  float32x4_t v59; // q8
  float32x4_t v60; // q12
  float32x4_t v61; // q14
  float32x4_t v62; // q10
  float32x4_t v63; // q24
  float32x4_t v64; // q31
  float32x4_t v65; // q9
  float32x4_t v66; // q11
  float32x4_t v67; // q21
  float32x4_t v68; // q13
  float32x4_t v69; // q19
  float32x4_t v70; // q20
  float32x4_t v71; // q15
  float32x4_t v72; // q14
  __int64 v73; // x22
  float32x4_t v74; // q26
  float32x4_t v75; // q10
  __int64 k; // x20
  _DWORD *v77; // x22
  __int64 v78; // x1
  int v79; // w2
  __int64 v80; // x3
  int v81; // w6
  int v82; // w7
  __int64 v83; // x8
  int v84; // w9
  __int64 v85; // x10
  __int64 v86; // x11
  __int64 v87; // x12
  __int64 v88; // x13
  __int64 v89; // x14
  __int64 v90; // x15
  __int64 v91; // x16
  __int64 v92; // x17
  float32x4_t v93; // q0
  float32x4_t v94; // q1
  float32x4_t v95; // q2
  float32x4_t v96; // q3
  float32x4_t v97; // q4
  __int64 v98; // x4
  __int64 v99; // x4
  __int64 v100; // x5
  __int64 v101; // x5
  int v102; // w20
  __int64 m; // x21
  _DWORD *v104; // x22
  int v105; // w23
  float32x4_t v106; // q16
  float32x4_t v107; // q20
  float32x4_t v108; // q7
  float32x4_t v109; // q19
  float32x4_t v110; // q5
  float32x4_t v111; // q18
  float32x4_t v112; // q6
  float32x4_t v113; // q17
  __int64 v114; // x21
  __int64 v115; // x21
  int n; // w22
  float32x4_t v117; // q21
  float32x4_t v118; // q16
  float32x4_t v119; // q7
  float32x4_t v120; // q5
  float32x4_t v121; // q6
  float32x4_t v122; // q23
  float32x4_t v123; // q17
  float32x4_t v124; // q23
  float32x4_t v125; // q18
  float32x4_t v126; // q23
  float32x4_t v127; // q22
  __int64 v128; // x21
  int ii; // w22
  float32x4_t v130; // q21
  float32x4_t v131; // q16
  float32x4_t v132; // q7
  float32x4_t v133; // q5
  float32x4_t v134; // q6
  float32x4_t v135; // q23
  float32x4_t v136; // q17
  float32x4_t v137; // q23
  float32x4_t v138; // q18
  float32x4_t v139; // q23
  float32x4_t v140; // q22
  __int64 v141; // x21
  int jj; // w22
  float32x4_t v143; // q21
  __int64 v144; // x23
  float32x4_t v145; // q16
  float32x4_t v146; // q7
  float32x4_t v147; // q5
  float32x4_t v148; // q6
  float32x4_t v149; // q23
  float32x4_t v150; // q17
  float32x4_t v151; // q23
  float32x4_t v152; // q18
  float32x4_t v153; // q23
  float32x4_t v154; // q22
  __int64 v155; // x21
  int kk; // w22
  float32x4_t v157; // q21
  float32x4_t v158; // q16
  float32x4_t v159; // q7
  float32x4_t v160; // q5
  float32x4_t v161; // q6
  float32x4_t v162; // q23
  float32x4_t v163; // q17
  float32x4_t v164; // q23
  float32x4_t v165; // q18
  float32x4_t v166; // q23
  float32x4_t v167; // q22
  __int64 mm; // x21
  _DWORD *v169; // x23

  v5 = *a5 != 0.0 || a5[1] != 0.0 || a5[2] != 0.0 || a5[3] != 0.0 || a5[4] != 0.0;
  *(float32x4_t *)&v6 = vld1q_dup_f32(a4);
  *(float32x4_t *)&v7 = vld1q_dup_f32(a4 + 2);
  *(float32x4_t *)&v8 = vld1q_dup_f32(a4 + 3);
  *(float32x4_t *)&v9 = vld1q_dup_f32(a4 + 1);
  *(float32x4_t *)&v10 = vld1q_dup_f32(a4 + 4);
  if ( v5 )
  {
    result = sub_103A50(a1, a2, a3, a4, a5, 323584, v6, v7, v8, v9, v10);
    v39 = v38 + 2148;
    while ( v16 >= 1 )
    {
      v33.n128_u32[0] = *v15;
      v34.n128_u32[0] = v15[1];
      v35.n128_u32[0] = v15[2];
      v36.n128_u32[0] = v15[3];
      v37.n128_u32[0] = v15[4];
      v40 = v16 - 4;
      if ( v16 >= 4 )
        v16 = v19;
      for ( i = 0; i < v16; ++i )
      {
        v42 = *(_DWORD **)(v20 + 8 * i);
        *(_DWORD *)(v21 + 4 * i) = *v42;
        *(_DWORD *)(v22 + 4 * i) = v42[1];
        *(_DWORD *)(v23 + 4 * i) = v42[2];
        *(_DWORD *)(v24 + 4 * i) = v42[3];
        *(_DWORD *)(v25 + 4 * i) = v42[4];
        *(_DWORD *)(v26 + 4 * i) = v42[5];
        v43 = v42[6];
        LODWORD(v42) = v42[7];
        *(_DWORD *)(v27 + 4 * i) = v43;
        *(_DWORD *)(v14 + 4 * i) = (_DWORD)v42;
      }
      v45 = 0u;
      v44 = 0u;
      v46 = 0u;
      v47 = 0u;
      v49 = 0u;
      v48 = 0u;
      v51 = 0u;
      v50 = 0u;
      v52 = 0;
      v33 = vdupq_lane_s32((int32x2_t)v33.n128_u64[0], 0);
      v34 = vdupq_lane_s32((int32x2_t)v34.n128_u64[0], 0);
      v35 = vdupq_lane_s32((int32x2_t)v35.n128_u64[0], 0);
      v36 = vdupq_lane_s32((int32x2_t)v36.n128_u64[0], 0);
      v37 = vdupq_lane_s32((int32x2_t)v37.n128_u64[0], 0);
      v53 = 0u;
      v54 = 0u;
      v55 = 0u;
      v56 = 0u;
      v57 = 0u;
      for ( j = v13; j; v45 = vmlaq_f32(v75, v65, v59) )
      {
        v59 = *(float32x4_t *)(result + v52);
        v60 = vaddq_f32(v57, v28);
        v61 = vaddq_f32(v53, v32);
        v62 = vaddq_f32(v56, v31);
        v63 = vmlaq_f32(v44, v60, v59);
        v64 = vaddq_f32(v54, v30);
        v65 = vaddq_f32(v55, v29);
        v66 = vmulq_f32(v63, v61);
        if ( v17 >= 1 )
        {
          v67 = vmlaq_f32(v47, v60, v63);
          v68 = vaddq_f32(vmlaq_f32(vnegq_f32(vmulq_f32(v67, v64)), v62, v63), v46);
          v46 = vmlaq_f32(vnegq_f32(vmulq_f32(v67, v61)), v65, v63);
          if ( v17 != 1 )
          {
            v69 = vmlaq_f32(v48, v60, v67);
            v70 = vmlaq_f32(v50, v60, v69);
            v71 = vmulq_f32(v69, v61);
            v72 = vmulq_f32(v70, v61);
            v50 = vmlsq_f32(vmlaq_f32(v51, v62, v69), v64, v70);
            v51 = vmlaq_f32(vnegq_f32(v72), v65, v69);
            v48 = vmlsq_f32(vmlaq_f32(v49, v62, v67), v64, v69);
            v49 = vmlaq_f32(vnegq_f32(v71), v65, v67);
          }
          v47 = v68;
        }
        v73 = (unsigned int)(v16 - 1);
        v74 = vmlaq_f32(v45, v62, v59);
        v75 = vnegq_f32(v66);
        if ( (unsigned int)v73 <= 3 )
          __asm { BR              X23 }
        v57 = vaddq_f32(v57, v33);
        v56 = vaddq_f32(v56, v34);
        v55 = vaddq_f32(v55, v35);
        v54 = vaddq_f32(v54, v36);
        v53 = vaddq_f32(v53, v37);
        --j;
        v52 += v18;
        v44 = vmlsq_f32(v74, v64, v63);
      }
      for ( k = 0; k < v16; ++k )
      {
        v77 = *(_DWORD **)(v20 + 8 * k);
        *v77 = *(_DWORD *)(v21 + 4 * k);
        v77[1] = *(_DWORD *)(v22 + 4 * k);
        v77[2] = *(_DWORD *)(v23 + 4 * k);
        v77[3] = *(_DWORD *)(v24 + 4 * k);
        v77[4] = *(_DWORD *)(v25 + 4 * k);
        v77[5] = *(_DWORD *)(v26 + 4 * k);
        v77[6] = *(_DWORD *)(v27 + 4 * k);
        v77[7] = *(_DWORD *)(v14 + 4 * k);
        *(_QWORD *)(v20 + 8 * k) = v77 + 32;
      }
      v12 += 16;
      result += 16;
      v16 = v40;
    }
  }
  else
  {
    result = sub_103A50(a1, a2, a3, a4, 323584, 323584, v6, v7, v8, v9, v10);
    v99 = v98 + 2136;
    v101 = v100 + 2140;
    while ( v81 >= 1 )
    {
      v102 = v81 - 4;
      if ( v81 >= 4 )
        v81 = v84;
      for ( m = 0; m < v81; ++m )
      {
        v104 = *(_DWORD **)(v85 + 8 * m);
        *(_DWORD *)(v86 + 4 * m) = *v104;
        *(_DWORD *)(v87 + 4 * m) = v104[1];
        *(_DWORD *)(v88 + 4 * m) = v104[2];
        *(_DWORD *)(v89 + 4 * m) = v104[3];
        *(_DWORD *)(v90 + 4 * m) = v104[4];
        *(_DWORD *)(v91 + 4 * m) = v104[5];
        v105 = v104[6];
        LODWORD(v104) = v104[7];
        *(_DWORD *)(v92 + 4 * m) = v105;
        *(_DWORD *)(v80 + 4 * m) = (_DWORD)v104;
      }
      v107 = 0u;
      v106 = 0u;
      v109 = 0u;
      v108 = 0u;
      v111 = 0u;
      v110 = 0u;
      v113 = 0u;
      v112 = 0u;
      v114 = (unsigned int)(v81 - 1);
      if ( v82 == 1 )
      {
        if ( (unsigned int)v114 <= 3 )
          __asm { BR              X22 }
      }
      else
      {
        if ( !v82 )
          __asm { BR              X22 }
        switch ( v81 )
        {
          case 1:
            v115 = 0;
            for ( n = v79; n; v115 += v83 )
            {
              v117 = *(float32x4_t *)(result + v115);
              --n;
              v118 = vmlaq_f32(v106, v93, v117);
              v119 = vmlaq_f32(v108, v93, v118);
              v120 = vmlaq_f32(v110, v93, v119);
              v121 = vmlaq_f32(v112, v93, v120);
              v122 = vmulq_f32(v121, v97);
              *(_DWORD *)(v78 + v115) = v121.n128_u32[0];
              v112 = vmlsq_f32(vmlaq_f32(v113, v96, v120), v95, v121);
              v123 = vnegq_f32(v122);
              v124 = vmulq_f32(v120, v97);
              v113 = vmlaq_f32(v123, v94, v120);
              v110 = vmlsq_f32(vmlaq_f32(v111, v96, v119), v95, v120);
              v125 = vnegq_f32(v124);
              v126 = vmulq_f32(v119, v97);
              v127 = vmulq_f32(v118, v97);
              v111 = vmlaq_f32(v125, v94, v119);
              v108 = vmlsq_f32(vmlaq_f32(v109, v96, v118), v95, v119);
              v109 = vmlaq_f32(vnegq_f32(v126), v94, v118);
              v106 = vmlsq_f32(vmlaq_f32(v107, v96, v117), v95, v118);
              v107 = vmlaq_f32(vnegq_f32(v127), v94, v117);
            }
            break;
          case 2:
            v128 = 0;
            for ( ii = v79; ii; v128 += v83 )
            {
              v130 = *(float32x4_t *)(result + v128);
              --ii;
              v131 = vmlaq_f32(v106, v93, v130);
              v132 = vmlaq_f32(v108, v93, v131);
              v133 = vmlaq_f32(v110, v93, v132);
              v134 = vmlaq_f32(v112, v93, v133);
              v135 = vmulq_f32(v134, v97);
              *(_QWORD *)(v78 + v128) = v134.n128_u64[0];
              v112 = vmlsq_f32(vmlaq_f32(v113, v96, v133), v95, v134);
              v136 = vnegq_f32(v135);
              v137 = vmulq_f32(v133, v97);
              v113 = vmlaq_f32(v136, v94, v133);
              v110 = vmlsq_f32(vmlaq_f32(v111, v96, v132), v95, v133);
              v138 = vnegq_f32(v137);
              v139 = vmulq_f32(v132, v97);
              v140 = vmulq_f32(v131, v97);
              v111 = vmlaq_f32(v138, v94, v132);
              v108 = vmlsq_f32(vmlaq_f32(v109, v96, v131), v95, v132);
              v109 = vmlaq_f32(vnegq_f32(v139), v94, v131);
              v106 = vmlsq_f32(vmlaq_f32(v107, v96, v130), v95, v131);
              v107 = vmlaq_f32(vnegq_f32(v140), v94, v130);
            }
            break;
          case 3:
            v141 = 0;
            for ( jj = v79; jj; v107 = vmlaq_f32(vnegq_f32(v154), v94, v143) )
            {
              v143 = *(float32x4_t *)(result + v141);
              v144 = v78 + v141;
              --jj;
              v141 += v83;
              v145 = vmlaq_f32(v106, v93, v143);
              v146 = vmlaq_f32(v108, v93, v145);
              v147 = vmlaq_f32(v110, v93, v146);
              v148 = vmlaq_f32(v112, v93, v147);
              *(_QWORD *)v144 = v148.n128_u64[0];
              v149 = vmulq_f32(v148, v97);
              *(_DWORD *)(v144 + 8) = v148.n128_u32[2];
              v112 = vmlsq_f32(vmlaq_f32(v113, v96, v147), v95, v148);
              v150 = vnegq_f32(v149);
              v151 = vmulq_f32(v147, v97);
              v113 = vmlaq_f32(v150, v94, v147);
              v110 = vmlsq_f32(vmlaq_f32(v111, v96, v146), v95, v147);
              v152 = vnegq_f32(v151);
              v153 = vmulq_f32(v146, v97);
              v154 = vmulq_f32(v145, v97);
              v111 = vmlaq_f32(v152, v94, v146);
              v108 = vmlsq_f32(vmlaq_f32(v109, v96, v145), v95, v146);
              v109 = vmlaq_f32(vnegq_f32(v153), v94, v145);
              v106 = vmlsq_f32(vmlaq_f32(v107, v96, v143), v95, v145);
            }
            break;
          case 4:
            v155 = 0;
            for ( kk = v79; kk; v155 += v83 )
            {
              v157 = *(float32x4_t *)(result + v155);
              --kk;
              v158 = vmlaq_f32(v106, v93, v157);
              v159 = vmlaq_f32(v108, v93, v158);
              v160 = vmlaq_f32(v110, v93, v159);
              v161 = vmlaq_f32(v112, v93, v160);
              v162 = vmulq_f32(v161, v97);
              *(float32x4_t *)(v78 + v155) = v161;
              v112 = vmlsq_f32(vmlaq_f32(v113, v96, v160), v95, v161);
              v163 = vnegq_f32(v162);
              v164 = vmulq_f32(v160, v97);
              v113 = vmlaq_f32(v163, v94, v160);
              v110 = vmlsq_f32(vmlaq_f32(v111, v96, v159), v95, v160);
              v165 = vnegq_f32(v164);
              v166 = vmulq_f32(v159, v97);
              v167 = vmulq_f32(v158, v97);
              v111 = vmlaq_f32(v165, v94, v159);
              v108 = vmlsq_f32(vmlaq_f32(v109, v96, v158), v95, v159);
              v109 = vmlaq_f32(vnegq_f32(v166), v94, v158);
              v106 = vmlsq_f32(vmlaq_f32(v107, v96, v157), v95, v158);
              v107 = vmlaq_f32(vnegq_f32(v167), v94, v157);
            }
            break;
          default:
            break;
        }
      }
      for ( mm = 0; mm < v81; ++mm )
      {
        v169 = *(_DWORD **)(v85 + 8 * mm);
        *v169 = *(_DWORD *)(v86 + 4 * mm);
        v169[1] = *(_DWORD *)(v87 + 4 * mm);
        v169[2] = *(_DWORD *)(v88 + 4 * mm);
        v169[3] = *(_DWORD *)(v89 + 4 * mm);
        v169[4] = *(_DWORD *)(v90 + 4 * mm);
        v169[5] = *(_DWORD *)(v91 + 4 * mm);
        v169[6] = *(_DWORD *)(v92 + 4 * mm);
        v169[7] = *(_DWORD *)(v80 + 4 * mm);
        *(_QWORD *)(v85 + 8 * mm) = v169 + 32;
      }
      v78 += 16;
      result += 16;
      v81 = v102;
    }
  }
  return result;
}