)]}' { "commit": "6c835f81a7c4315518c0ed766b1eef511335bf0b", "tree": "c7398b8f2bd27ae334eaeddbce3caa6901d009a0", "parents": [ "72cb66cb8f9ffa8d482d00a5dfe4784d7978fe2e" ], "author": { "name": "Sameer Agarwal", "email": "sameeragarwal@google.com", "time": "Sun Feb 25 13:36:33 2018 -0800" }, "committer": { "name": "Sameer Agarwal", "email": "sameeragarwal@google.com", "time": "Sun Feb 25 13:44:22 2018 -0800" }, "message": "Improve the performance of MatrixTransposeVector Multiply.\n\nThis is done by making the matrix access more cache coherent.\nFor small matrices there isn\u0027t much difference, but for\nlarger matrices like 4x20, this leads to ~50% performance\nimprovement.\n\nBenchmark Time CPU Time Old Time New CPU Old CPU New\n-----------------------------------------------------------------------------------------------------------------------------\nBM_MatrixVectorMultiply/1/1 +0.0138 +0.0046 10 10 10 10\nBM_MatrixVectorMultiply/1/2 +0.0215 +0.0146 10 11 10 10\nBM_MatrixVectorMultiply/1/3 +0.0469 +0.0422 10 10 10 10\nBM_MatrixVectorMultiply/1/4 +0.0696 +0.0671 10 11 10 11\nBM_MatrixVectorMultiply/1/6 +0.0827 +0.0795 12 13 12 13\nBM_MatrixVectorMultiply/1/7 +0.1408 +0.1152 13 15 13 14\nBM_MatrixVectorMultiply/1/12 +0.0317 +0.0232 17 18 17 17\nBM_MatrixVectorMultiply/1/16 -0.0362 -0.0168 21 20 20 20\nBM_MatrixVectorMultiply/1/20 +0.0304 +0.0290 22 22 22 22\nBM_MatrixVectorMultiply/2/1 +0.0019 +0.0009 10 10 10 10\nBM_MatrixVectorMultiply/2/2 +0.0285 +0.0291 11 12 11 12\nBM_MatrixVectorMultiply/2/3 -0.1178 -0.0977 16 14 15 14\nBM_MatrixVectorMultiply/2/4 +0.0938 +0.1144 15 17 15 17\nBM_MatrixVectorMultiply/2/6 +0.0617 +0.0629 17 18 17 18\nBM_MatrixVectorMultiply/2/7 +0.0956 +0.0876 20 22 19 21\nBM_MatrixVectorMultiply/2/12 +0.0043 +0.0119 24 24 24 24\nBM_MatrixVectorMultiply/2/16 +0.0361 +0.0337 29 30 29 30\nBM_MatrixVectorMultiply/2/20 +0.0463 +0.0365 33 35 33 34\nBM_MatrixVectorMultiply/3/1 -0.0201 -0.0210 12 12 12 11\nBM_MatrixVectorMultiply/3/2 -0.0741 -0.0766 16 15 16 15\nBM_MatrixVectorMultiply/3/3 -0.0076 -0.0118 18 18 18 18\nBM_MatrixVectorMultiply/3/4 +0.1071 +0.0963 19 21 19 21\nBM_MatrixVectorMultiply/3/6 +0.0449 +0.0390 23 24 23 23\nBM_MatrixVectorMultiply/3/7 +0.1099 +0.1018 25 28 24 27\nBM_MatrixVectorMultiply/3/12 +0.1512 +0.0992 33 38 32 35\nBM_MatrixVectorMultiply/3/16 +0.2281 +0.2005 37 46 37 44\nBM_MatrixVectorMultiply/3/20 +0.2387 +0.1799 49 61 48 57\nBM_MatrixVectorMultiply/4/1 +0.4444 +0.2574 14 21 14 18\nBM_MatrixVectorMultiply/4/2 +0.0313 +0.0230 19 20 19 20\nBM_MatrixVectorMultiply/4/3 +0.0626 +0.0596 23 24 23 24\nBM_MatrixVectorMultiply/4/4 +0.2322 +0.1440 23 28 23 26\nBM_MatrixVectorMultiply/4/6 +0.0936 +0.0768 26 29 26 28\nBM_MatrixVectorMultiply/4/7 +0.0848 +0.0835 28 30 28 30\nBM_MatrixVectorMultiply/4/12 +0.1607 +0.1101 39 46 39 43\nBM_MatrixVectorMultiply/4/16 +0.0752 +0.0687 48 52 48 51\nBM_MatrixVectorMultiply/4/20 +0.1782 +0.1463 61 72 60 69\nBM_MatrixTransposeVectorMultiply/1/1 +0.3609 +0.2857 9 13 9 12\nBM_MatrixTransposeVectorMultiply/1/2 +0.3106 +0.2970 10 13 10 12\nBM_MatrixTransposeVectorMultiply/1/3 +0.3018 +0.2383 11 14 11 13\nBM_MatrixTransposeVectorMultiply/1/4 -0.0795 -0.0819 14 13 14 12\nBM_MatrixTransposeVectorMultiply/1/6 -0.0108 -0.0629 18 18 18 16\nBM_MatrixTransposeVectorMultiply/1/7 -0.1073 -0.0879 20 18 19 17\nBM_MatrixTransposeVectorMultiply/1/12 -0.3035 -0.3016 26 18 26 18\nBM_MatrixTransposeVectorMultiply/1/16 -0.4909 -0.4951 39 20 38 19\nBM_MatrixTransposeVectorMultiply/1/20 -0.4619 -0.4985 43 23 42 21\nBM_MatrixTransposeVectorMultiply/2/1 +0.3471 +0.2906 10 13 10 13\nBM_MatrixTransposeVectorMultiply/2/2 +0.2323 +0.2266 12 15 12 15\nBM_MatrixTransposeVectorMultiply/2/3 +0.0802 +0.0779 16 17 16 17\nBM_MatrixTransposeVectorMultiply/2/4 -0.0951 -0.0983 19 17 19 17\nBM_MatrixTransposeVectorMultiply/2/6 -0.0974 -0.1064 24 21 24 21\nBM_MatrixTransposeVectorMultiply/2/7 +0.0612 -0.0457 27 29 27 26\nBM_MatrixTransposeVectorMultiply/2/12 -0.3777 -0.3838 41 25 41 25\nBM_MatrixTransposeVectorMultiply/2/16 -0.4783 -0.4843 53 28 53 27\nBM_MatrixTransposeVectorMultiply/2/20 -0.5567 -0.5566 71 32 70 31\nBM_MatrixTransposeVectorMultiply/3/1 +0.4607 +0.4753 10 15 10 15\nBM_MatrixTransposeVectorMultiply/3/2 +0.2870 +0.2754 14 19 14 18\nBM_MatrixTransposeVectorMultiply/3/3 +0.1270 +0.1245 19 21 19 21\nBM_MatrixTransposeVectorMultiply/3/4 +0.0160 +0.0076 22 22 22 22\nBM_MatrixTransposeVectorMultiply/3/6 -0.0612 -0.0635 27 26 27 25\nBM_MatrixTransposeVectorMultiply/3/7 -0.0531 -0.0695 31 29 30 28\nBM_MatrixTransposeVectorMultiply/3/12 -0.4009 -0.3938 49 29 47 29\nBM_MatrixTransposeVectorMultiply/3/16 -0.4584 -0.4537 64 35 62 34\nBM_MatrixTransposeVectorMultiply/3/20 -0.5087 -0.5098 78 38 77 38\nBM_MatrixTransposeVectorMultiply/4/1 +0.6696 +0.6837 11 18 11 18\nBM_MatrixTransposeVectorMultiply/4/2 +0.3085 +0.3085 17 22 17 22\nBM_MatrixTransposeVectorMultiply/4/3 +0.2908 +0.2821 21 26 20 26\nBM_MatrixTransposeVectorMultiply/4/4 +0.0076 -0.0031 24 25 24 24\nBM_MatrixTransposeVectorMultiply/4/6 -0.0884 -0.0841 34 31 34 31\nBM_MatrixTransposeVectorMultiply/4/7 -0.0834 -0.0825 37 34 36 33\nBM_MatrixTransposeVectorMultiply/4/12 -0.4477 -0.4453 62 34 61 34\nBM_MatrixTransposeVectorMultiply/4/16 -0.5324 -0.5203 86 40 83 40\nBM_MatrixTransposeVectorMultiply/4/20 -0.4905 -0.4933 99 50 98 50\n\nChange-Id: I7f2a1c986e4a345bb67cb9eb0235234573024889\n", "tree_diff": [ { "type": "delete", "old_id": "3e483a2392212f5855f42d442efb68f964226bf3", "old_mode": 33188, "old_path": "benchmarks/macbook-pro-2014-small_blas_gemv_benchmark.json", "new_id": "0000000000000000000000000000000000000000", "new_mode": 0, "new_path": "/dev/null" }, { "type": "add", "old_id": "0000000000000000000000000000000000000000", "old_mode": 0, "old_path": "/dev/null", "new_id": "1798ccbd1f6e5703a5ec143526554172e024e522", "new_mode": 33188, "new_path": "internal/ceres/benchmarks/macbook-pro-2014-small_blas_gemv_benchmark.json" }, { "type": "modify", "old_id": "264ac53047ddaaf8a45c339f4759b281efcf910c", "old_mode": 33188, "old_path": "internal/ceres/small_blas.h", "new_id": "a603437174e4c3666d4f7b9f3b90171cd00ae385", "new_mode": 33188, "new_path": "internal/ceres/small_blas.h" } ] }