// 
// sub_1022D4 from 0x1022d4 to 0x1024f0 (540 bytes)
// 
// 1 xrefs:
// > <???> @ 0x11a100
// 
// 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_1022D4(float32x4_t *result, __int64 a2, __int64 a3, __int64 a4, const float *a5)
{
  __int64 v5; // x30
  float32x4_t v6; // q0
  float32x4_t v7; // q1
  __int64 v8; // x8
  long double v9; // q2
  long double v10; // q3
  __int64 v11; // x9
  __int64 v12; // x5
  long double v13; // q0
  int8x16_t v14; // q1
  long double v15; // q2
  int8x16_t v16; // q3
  long double v17; // q4
  int v18; // w8
  unsigned __int64 v19; // d6
  __int64 v20; // x10
  unsigned int v21; // w9
  int16x8_t v22; // q16
  float32x4_t v23; // q17
  float32x4_t v24; // q6
  long double v25; // q5
  __int64 v26; // x9
  int16x8_t v27; // q16
  float32x4_t v28; // q7
  float32x4_t v29; // q6
  long double *v30; // x9
  float32x4_t *v31; // x0
  long double v32; // q5
  float32x4_t v33; // q16
  float32x4_t v34; // q17
  float32x4_t v35; // q6
  float32x4_t v36; // q7
  long double v37; // q4
  long double *v38; // x9
  int64x2_t v39; // q3
  long double v40; // q0
  long double v41; // q1
  long double v42; // q2
  float32x4_t *v43; // x0
  float32x4_t v44; // q5
  float32x4_t v45; // q16
  float32x4_t v46; // q17
  float32x4_t v47; // q6
  float32x4_t v48; // q7
  int v49; // w8
  float32x4_t v50; // q0
  float32x4_t v51; // q1
  __int64 v52; // x9
  long double v53; // q2
  long double v54; // q3
  __int64 v55; // x10
  __int64 v56; // x0
  __int64 v57; // x1
  int v58; // w8

  if ( (_DWORD)a2 && ((unsigned __int8)result & 0xF) != 0 )
  {
    v6 = vld1q_dup_f32(a5);
    v7 = vld1q_dup_f32((const float *)a4);
    v8 = a3 + 16;
    *(_QWORD *)&v9 = 0x3000000030000000LL;
    *((_QWORD *)&v9 + 1) = 0x3000000030000000LL;
    *(_QWORD *)&v10 = 0x3800000038000000LL;
    *((_QWORD *)&v10 + 1) = 0x3800000038000000LL;
    do
    {
      if ( ((unsigned __int8)result & 0xF) == 0 )
        break;
      v11 = 2LL * (unsigned int)(8 * *(_DWORD *)(a4 + 4));
      result = (float32x4_t *)sub_102D18(
                                *(long double *)&v6,
                                COERCE_LONG_DOUBLE(vaddq_s32(v7, v6)),
                                v9,
                                v10,
                                COERCE_LONG_DOUBLE(vmulq_f32(vcvtq_f32_u32(vshrq_n_u32(v7, 1u)), *(float32x4_t *)&v9)),
                                *(long double *)(a3 + v11),
                                *(long double *)(v8 + v11));
      v5 = v12;
      *(_QWORD *)a4 += *(_QWORD *)a5;
    }
    while ( (_DWORD)a2 );
  }
  if ( (int)a2 >> 2 )
  {
    result = (float32x4_t *)sub_102CA4(result, a2, a3, a4, a5, v5);
    if ( v18 )
    {
      do
      {
        v19 = vshl_n_s32(vzip2_s32((int32x2_t)v14.n128_u64[0], (int32x2_t)vextq_s8(v14, v14, 8u).n128_u64[0]), 3u).n64_u64[0];
        v20 = a3 + 2LL * (unsigned int)v19;
        v21 = HIDWORD(v19);
        v22 = *(int16x8_t *)(v20 + 16);
        v23 = vcvtq_f32_s32(vmovl_s16(*(int16x4_t *)v20));
        v24 = vcvtq_f32_s32(vmovl_high_s16(*(int16x8_t *)v20));
        v25 = COERCE_LONG_DOUBLE(vmulq_f32(vcvtq_f32_u32(vshrq_n_u32(vuzp1q_s32(v14, v16), 1u)), *(float32x4_t *)&v13));
        v26 = a3 + 2LL * v21;
        *result = vmulq_f32(
                    vmlaq_n_f32(
                      v23,
                      vsubq_f32(vcvtq_f32_s32(vmovl_s16((int16x4_t)v22.n128_u64[0])), v23),
                      *(float *)&v25),
                    *(float32x4_t *)&v17);
        result[1] = vmulq_f32(
                      vmlaq_n_f32(v24, vsubq_f32(vcvtq_f32_s32(vmovl_high_s16(v22)), v24), *(float *)&v25),
                      *(float32x4_t *)&v17);
        v27 = *(int16x8_t *)(v26 + 16);
        v28 = vcvtq_f32_s32(vmovl_s16(*(int16x4_t *)v26));
        v29 = vcvtq_f32_s32(vmovl_high_s16(*(int16x8_t *)v26));
        v30 = (long double *)(a3
                            + 2LL
                            * vshl_n_s32(
                                vzip2_s32((int32x2_t)v16.n128_u64[0], (int32x2_t)vextq_s8(v16, v16, 8u).n128_u64[0]),
                                3u).n64_u32[0]);
        result[2] = vmulq_f32(
                      vmlaq_lane_f32(
                        v28,
                        vsubq_f32(vcvtq_f32_s32(vmovl_s16((int16x4_t)v27.n128_u64[0])), v28),
                        *(float32x2_t *)&v25,
                        1),
                      *(float32x4_t *)&v17);
        result[3] = vmulq_f32(
                      vmlaq_lane_f32(v29, vsubq_f32(vcvtq_f32_s32(vmovl_high_s16(v27)), v29), *(float32x2_t *)&v25, 1),
                      *(float32x4_t *)&v17);
        v31 = (float32x4_t *)sub_102DE4(
                               v13,
                               COERCE_LONG_DOUBLE(vaddq_s64(v14, *(int64x2_t *)&v15)),
                               v15,
                               *(long double *)&v16,
                               v17,
                               v25,
                               *v30,
                               v30[1]);
        v31[4] = vmulq_f32(vmlaq_laneq_f32(v33, v34, *(float32x4_t *)&v32, 2), *(float32x4_t *)&v37);
        v31[5] = vmulq_f32(vmlaq_laneq_f32(v35, v36, *(float32x4_t *)&v32, 2), *(float32x4_t *)&v37);
        v43 = (float32x4_t *)sub_102DE4(
                               v40,
                               v41,
                               v42,
                               COERCE_LONG_DOUBLE(vaddq_s64(v39, *(int64x2_t *)&v42)),
                               v37,
                               v32,
                               *v38,
                               v38[1]);
        v43[6] = vmulq_f32(vmlaq_laneq_f32(v45, v46, v44, 3), *(float32x4_t *)&v17);
        v43[7] = vmulq_f32(vmlaq_laneq_f32(v47, v48, v44, 3), *(float32x4_t *)&v17);
        result = v43 + 8;
      }
      while ( v49 );
    }
    *(_QWORD *)a4 = v14.n128_u64[0];
  }
  if ( (a2 & 3) != 0 )
  {
    v50 = vld1q_dup_f32(a5);
    v51 = vld1q_dup_f32((const float *)a4);
    v52 = a3 + 16;
    *(_QWORD *)&v53 = 0x3000000030000000LL;
    *((_QWORD *)&v53 + 1) = 0x3000000030000000LL;
    *(_QWORD *)&v54 = 0x3800000038000000LL;
    *((_QWORD *)&v54 + 1) = 0x3800000038000000LL;
    do
    {
      v55 = 2LL * (unsigned int)(8 * *(_DWORD *)(a4 + 4));
      v56 = sub_102D18(
              *(long double *)&v50,
              COERCE_LONG_DOUBLE(vaddq_s32(v51, v50)),
              v53,
              v54,
              COERCE_LONG_DOUBLE(vmulq_f32(vcvtq_f32_u32(vshrq_n_u32(v51, 1u)), *(float32x4_t *)&v53)),
              *(long double *)(a3 + v55),
              *(long double *)(v52 + v55));
      result = (float32x4_t *)sub_102E10(v56, v57);
    }
    while ( v58 );
  }
  return result;
}