)]}'
{
  "commit": "a3ad84f95cdf5eb02f59c65dbe584d41c98a7c5a",
  "tree": "800bf09b7705bab28165cff3486aadf02d325ef1",
  "parents": [
    "eee6faf6a1f3107315167618159ff6a85484c01f"
  ],
  "author": {
    "name": "George Steed",
    "email": "george.steed@arm.com",
    "time": "Thu Nov 30 07:39:43 2023 +0000"
  },
  "committer": {
    "name": "James Zern",
    "email": "jzern@google.com",
    "time": "Fri Dec 01 20:38:23 2023 +0000"
  },
  "message": "av1_fwd_txfm2d_neon.c: Use switch for small square problem sizes\n\nFor the smallest problem sizes we have a significant overhead from\nneeding to load and store the transform intermediate vectors. Avoiding\nthe kernel lookup and calling the kernels directly significantly\nimproves performance in these cases.\n\nBenchmarking on a Neoverse N2 machine with Clang 16 and GCC 12, the\nspeed tests report a geomean ~22% reduction in times reported for 4x4\nand a ~4.6% reduction for 8x8.\n\nChange-Id: I5907fab09e40b1cea57c446fd7e604ae911ceae8\n",
  "tree_diff": [
    {
      "type": "modify",
      "old_id": "a17a41ad13c2729b22aa10c1b3e507d8090a0f79",
      "old_mode": 33188,
      "old_path": "av1/encoder/arm/neon/av1_fwd_txfm2d_neon.c",
      "new_id": "d70f5a5a398f48a94fd900fa39ad9ecd58e49506",
      "new_mode": 33188,
      "new_path": "av1/encoder/arm/neon/av1_fwd_txfm2d_neon.c"
    }
  ]
}