)]}'
{
  "commit": "842466f5f94e7f2a3d8f5118d9b5db984b163baa",
  "tree": "71dbee615c17013b344e28ff53976a68e14191eb",
  "parents": [
    "2e124f462b87996c6d6ddb1fab843c04bdaa14a6"
  ],
  "author": {
    "name": "Gerda Zsejke More",
    "email": "gerdazsejke.more@arm.com",
    "time": "Sat Oct 18 18:43:54 2025 +0200"
  },
  "committer": {
    "name": "Jonathan Wright",
    "email": "jonathan.wright@arm.com",
    "time": "Mon Nov 17 10:05:15 2025 -0800"
  },
  "message": "Optimize Neon I8MM implementation of av1_warp_affine\n\nIn case of beta \u003d\u003d 0 and alpha \u003d\u003d 0 we know filter values before\nprocessing loops so add new logic using USMMLA instruction for\nhorizontal filtering.\n\nBy permuting the input samples and the 6-tap filter we can use the\nArmv8.6 I8MM USMMLA matrix multiply instructions to accelerate\nhorizontal 6-tap convolutions. The 2x8 by 8x2 matrix multiply\ninstruction does twice the work of a USDOT dot product instruction.\n\nIn case of 8-tap filter we can replace the USDOT instruction with\nUSMMLA, to apply a 7-tap filter, and an extra multiplication.\n\nChange-Id: I5caf759db5a047335d9b12b6d4b77d8df605f23f\n",
  "tree_diff": [
    {
      "type": "modify",
      "old_id": "15ac0043cac7be6e8b55114d021b1e6776ac35c2",
      "old_mode": 33188,
      "old_path": "av1/common/arm/warp_plane_neon_i8mm.c",
      "new_id": "2d0297452739fe0f1fa4e1e21894846a4166a45b",
      "new_mode": 33188,
      "new_path": "av1/common/arm/warp_plane_neon_i8mm.c"
    }
  ]
}