)]}' { "commit": "68cc71ce5d7ce58751343ccf9cb5aec6dcd60f8e", "tree": "ed178a0ccc72eab5c1fc8ed108057b52671fa3c4", "parents": [ "f27082a17463803e4deed9eea28075883e58a31c" ], "author": { "name": "yangfan", "email": "yang0773@gmail.com", "time": "Mon Mar 12 10:57:52 2018 +0800" }, "committer": { "name": "Sameer Agarwal", "email": "sameeragarwal@google.com", "time": "Sun Apr 08 14:23:10 2018 +0000" }, "message": "Optimization for custom small blas multiplication with dynamic\ntemplate parameters in C level.\n\n- unroll for loops\n- matrix access more cache coherent\n- platform independant\n\nBriefly, this commit brings 1~50% performance improvments for\nmost cases in small_blas_gem(m/v)_benchmark, but a small drop\nfor corner cases with small dimensions especially 1,2,3. Here\nwe list the results partially, which show decrease percentage\nof executing time, compared to unoptimized version.\n\nPlatform: desktop PC (i7-7700 CPU MP8@3.60GHz + ubuntu 17.10)\n(Lenovo Research Device+ Lab, \u003cyangfan34@lenovo.com\u003e)\n\nBenchmark Time CPU\n-----------------------------------------------------------\nBM_MatrixMatrixMultiplyDynamic/2/2/2 -0.1082 -0.1083\nBM_MatrixMatrixMultiplyDynamic/2/2/15 -0.1270 -0.1270\nBM_MatrixMatrixMultiplyDynamic/2/4/2 -0.1433 -0.1433\nBM_MatrixMatrixMultiplyDynamic/2/4/15 -0.2069 -0.2068\nBM_MatrixMatrixMultiplyDynamic/2/6/2 -0.1446 -0.1446\nBM_MatrixMatrixMultiplyDynamic/2/6/15 -0.2156 -0.2156\nBM_MatrixMatrixMultiplyDynamic/2/8/2 -0.1788 -0.1788\nBM_MatrixMatrixMultiplyDynamic/2/8/15 -0.3316 -0.3316\nBM_MatrixMatrixMultiplyDynamic/2/10/2 -0.2025 -0.2025\nBM_MatrixMatrixMultiplyDynamic/2/10/15 -0.3444 -0.3444\nBM_MatrixMatrixMultiplyDynamic/2/12/2 -0.0515 -0.0515\nBM_MatrixMatrixMultiplyDynamic/2/12/15 -0.3733 -0.3733\nBM_MatrixMatrixMultiplyDynamic/2/15/2 -0.2784 -0.2784\nBM_MatrixMatrixMultiplyDynamic/2/15/15 -0.3704 -0.3704\nBM_MatrixMatrixMultiplyDynamic/4/2/2 -0.1839 -0.1839\nBM_MatrixMatrixMultiplyDynamic/4/2/15 -0.1922 -0.1922\nBM_MatrixMatrixMultiplyDynamic/4/4/2 -0.2248 -0.2248\nBM_MatrixMatrixMultiplyDynamic/4/4/15 -0.3132 -0.3132\nBM_MatrixMatrixMultiplyDynamic/4/6/2 -0.2311 -0.2311\nBM_MatrixMatrixMultiplyDynamic/4/6/15 -0.3239 -0.3239\nBM_MatrixMatrixMultiplyDynamic/4/8/2 -0.0574 -0.0574\nBM_MatrixMatrixMultiplyDynamic/4/8/15 -0.4173 -0.4173\nBM_MatrixMatrixMultiplyDynamic/4/10/2 -0.2861 -0.2861\nBM_MatrixMatrixMultiplyDynamic/4/10/15 -0.4065 -0.4064\nBM_MatrixMatrixMultiplyDynamic/4/12/2 -0.2976 -0.2975\nBM_MatrixMatrixMultiplyDynamic/4/12/15 -0.4218 -0.4218\nBM_MatrixMatrixMultiplyDynamic/4/15/2 -0.3116 -0.3116\nBM_MatrixMatrixMultiplyDynamic/4/15/15 -0.4242 -0.4241\nBM_MatrixMatrixMultiplyDynamic/8/12/2 -0.3675 -0.3674\nBM_MatrixMatrixMultiplyDynamic/8/12/4 -0.5055 -0.5055\nBM_MatrixMatrixMultiplyDynamic/8/12/6 -0.4302 -0.4302\nBM_MatrixMatrixMultiplyDynamic/8/12/8 -0.4854 -0.4854\nBM_MatrixMatrixMultiplyDynamic/8/12/10 -0.4882 -0.4882\nBM_MatrixMatrixMultiplyDynamic/8/12/12 -0.5209 -0.5209\nBM_MatrixMatrixMultiplyDynamic/8/12/15 -0.4558 -0.4558\nBM_MatrixMatrixMultiplyDynamic/8/15/2 -0.2319 -0.2319\nBM_MatrixMatrixMultiplyDynamic/8/15/4 -0.5105 -0.5105\nBM_MatrixMatrixMultiplyDynamic/8/15/6 -0.4477 -0.4477\nBM_MatrixMatrixMultiplyDynamic/8/15/8 -0.5479 -0.5479\nBM_MatrixMatrixMultiplyDynamic/8/15/10 -0.4843 -0.4843\nBM_MatrixMatrixMultiplyDynamic/8/15/12 -0.5212 -0.5212\nBM_MatrixMatrixMultiplyDynamic/8/15/15 -0.4459 -0.4459\n\nBM_MatrixVectorMultiply/1/1 +0.0978 +0.0978\nBM_MatrixVectorMultiply/1/2 +0.0551 +0.0551\nBM_MatrixVectorMultiply/1/3 -0.0019 -0.0020\nBM_MatrixVectorMultiply/1/4 +0.0563 +0.0562\nBM_MatrixVectorMultiply/1/6 +0.1379 +0.1379\nBM_MatrixVectorMultiply/1/7 +0.1090 +0.1090\nBM_MatrixVectorMultiply/1/12 +0.0901 +0.0901\nBM_MatrixVectorMultiply/1/16 +0.0493 +0.0493\nBM_MatrixVectorMultiply/1/20 +0.2255 +0.2255\nBM_MatrixVectorMultiply/2/1 +0.1261 +0.1261\nBM_MatrixVectorMultiply/2/2 +0.2328 +0.2328\nBM_MatrixVectorMultiply/2/3 +0.1404 +0.1403\nBM_MatrixVectorMultiply/2/4 +0.0257 +0.0256\nBM_MatrixVectorMultiply/2/6 -0.1691 -0.1691\nBM_MatrixVectorMultiply/2/7 -0.2619 -0.2619\nBM_MatrixVectorMultiply/2/12 -0.4261 -0.4261\nBM_MatrixVectorMultiply/2/16 -0.5387 -0.5387\nBM_MatrixVectorMultiply/2/20 -0.6171 -0.6171\nBM_MatrixVectorMultiply/3/1 +0.1664 +0.1664\nBM_MatrixVectorMultiply/3/2 +0.0848 +0.0848\nBM_MatrixVectorMultiply/3/3 -0.0044 -0.0044\nBM_MatrixVectorMultiply/3/4 -0.0683 -0.0684\nBM_MatrixVectorMultiply/3/6 -0.1652 -0.1652\nBM_MatrixVectorMultiply/3/7 -0.1633 -0.1633\nBM_MatrixVectorMultiply/3/12 -0.1921 -0.1921\nBM_MatrixVectorMultiply/3/16 -0.3659 -0.3659\nBM_MatrixVectorMultiply/3/20 -0.4137 -0.4137\nBM_MatrixVectorMultiply/4/1 -0.0577 -0.0577\nBM_MatrixVectorMultiply/4/2 -0.1337 -0.1338\nBM_MatrixVectorMultiply/4/3 -0.1443 -0.1443\nBM_MatrixVectorMultiply/4/4 +0.0013 +0.0013\nBM_MatrixVectorMultiply/4/6 -0.1071 -0.1071\nBM_MatrixVectorMultiply/4/7 -0.1396 -0.1397\nBM_MatrixVectorMultiply/4/12 -0.2792 -0.2792\nBM_MatrixVectorMultiply/4/16 -0.4485 -0.4486\nBM_MatrixVectorMultiply/4/20 -0.3588 -0.3588\n\nChange-Id: I64a8cf11391e3d06341a2b8764cd1b4f1b8a23f1\n", "tree_diff": [ { "type": "modify", "old_id": "264ac53047ddaaf8a45c339f4759b281efcf910c", "old_mode": 33188, "old_path": "internal/ceres/small_blas.h", "new_id": "34b4ec7adcba61f6842a44c7fcd6de0d0f56ad57", "new_mode": 33188, "new_path": "internal/ceres/small_blas.h" }, { "type": "add", "old_id": "0000000000000000000000000000000000000000", "old_mode": 0, "old_path": "/dev/null", "new_id": "978c5d56a8427978c625c7684a0869890aa93359", "new_mode": 33188, "new_path": "internal/ceres/small_blas_generic.h" } ] }