NEON intrinsics blurb.

gistfile1.txt

// NEON ASM:

  // 8x8 transpose
  vswp      d1,d8
  vswp      d3,d10
  vswp      d5,d12
  vswp      d7,d14

  vtrn.32   q0,q2
  vtrn.32   q1,q3
  vtrn.32   q4,q6
  vtrn.32   q5,q7

  vtrn.16   q0,q1
  vtrn.16   q2,q3
  vtrn.16   q4,q5
  vtrn.16   q6,q7

// NEON intrinsics to accomplish the exact same thing:

  int32x4_t a0 = vreinterpretq_s32_s16(vcombine_s16(vget_low_s16(i0), vget_low_s16(i4)));
  int32x4_t a1 = vreinterpretq_s32_s16(vcombine_s16(vget_low_s16(i1), vget_low_s16(i5)));
  int32x4_t a2 = vreinterpretq_s32_s16(vcombine_s16(vget_low_s16(i2), vget_low_s16(i6)));
  int32x4_t a3 = vreinterpretq_s32_s16(vcombine_s16(vget_low_s16(i3), vget_low_s16(i7)));
  int32x4_t a4 = vreinterpretq_s32_s16(vcombine_s16(vget_high_s16(i0), vget_high_s16(i4)));
  int32x4_t a5 = vreinterpretq_s32_s16(vcombine_s16(vget_high_s16(i1), vget_high_s16(i5)));
  int32x4_t a6 = vreinterpretq_s32_s16(vcombine_s16(vget_high_s16(i2), vget_high_s16(i6)));
  int32x4_t a7 = vreinterpretq_s32_s16(vcombine_s16(vget_high_s16(i3), vget_high_s16(i7)));

  int16x8_t b0 = vreinterpretq_s16_s32(vtrnq_s32(a0, a2).val[0]);
  int16x8_t b1 = vreinterpretq_s16_s32(vtrnq_s32(a1, a3).val[0]);
  int16x8_t b2 = vreinterpretq_s16_s32(vtrnq_s32(a0, a2).val[1]);
  int16x8_t b3 = vreinterpretq_s16_s32(vtrnq_s32(a1, a3).val[1]);
  int16x8_t b4 = vreinterpretq_s16_s32(vtrnq_s32(a4, a6).val[0]);
  int16x8_t b5 = vreinterpretq_s16_s32(vtrnq_s32(a5, a7).val[0]);
  int16x8_t b6 = vreinterpretq_s16_s32(vtrnq_s32(a4, a6).val[1]);
  int16x8_t b7 = vreinterpretq_s16_s32(vtrnq_s32(a5, a7).val[1]);

  int16x8_t o0 = vtrnq_s16(b0, b1).val[0];
  int16x8_t o1 = vtrnq_s16(b0, b1).val[1];
  int16x8_t o2 = vtrnq_s16(b2, b3).val[0];
  int16x8_t o3 = vtrnq_s16(b2, b3).val[1];
  int16x8_t o4 = vtrnq_s16(b4, b5).val[0];
  int16x8_t o5 = vtrnq_s16(b4, b5).val[1];
  int16x8_t o6 = vtrnq_s16(b6, b7).val[0];
  int16x8_t o7 = vtrnq_s16(b6, b7).val[1];