diff --git a/benchmarks/macbook-pro-2014-small_blas_gemv_benchmark.json b/benchmarks/macbook-pro-2014-small_blas_gemv_benchmark.json
new file mode 100644
index 0000000..3e483a2
--- /dev/null
+++ b/benchmarks/macbook-pro-2014-small_blas_gemv_benchmark.json
@@ -0,0 +1,515 @@
+{
+  "context": {
+    "date": "2018-02-25 13:20:04",
+    "num_cpus": 8,
+    "mhz_per_cpu": 2200,
+    "cpu_scaling_enabled": false,
+    "library_build_type": "release"
+  },
+  "benchmarks": [
+    {
+      "name": "BM_MatrixVectorMultiply/1/1",
+      "iterations": 69298697,
+      "real_time": 1.0097105894512250e+01,
+      "cpu_time": 1.0040275360444367e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/1/2",
+      "iterations": 65913992,
+      "real_time": 1.0302522610563768e+01,
+      "cpu_time": 1.0245351244998181e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/1/3",
+      "iterations": 73595895,
+      "real_time": 9.6440729347083387e+00,
+      "cpu_time": 9.5853172245544389e+00,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/1/4",
+      "iterations": 69574897,
+      "real_time": 1.0073530211782117e+01,
+      "cpu_time": 9.9997560901886722e+00,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/1/6",
+      "iterations": 58273603,
+      "real_time": 1.1879482448448192e+01,
+      "cpu_time": 1.1746073089045133e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/1/7",
+      "iterations": 54426846,
+      "real_time": 1.2970374490544540e+01,
+      "cpu_time": 1.2881400476522181e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/1/12",
+      "iterations": 40754541,
+      "real_time": 1.7086407941937370e+01,
+      "cpu_time": 1.6969201051730625e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/1/16",
+      "iterations": 36813042,
+      "real_time": 2.0977509004292560e+01,
+      "cpu_time": 2.0338009556504460e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/1/20",
+      "iterations": 32061264,
+      "real_time": 2.1783493596539344e+01,
+      "cpu_time": 2.1625753744456222e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/2/1",
+      "iterations": 71987577,
+      "real_time": 9.9393425620120528e+00,
+      "cpu_time": 9.8528944792793975e+00,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/2/2",
+      "iterations": 63771443,
+      "real_time": 1.1381383373240160e+01,
+      "cpu_time": 1.1294600939169571e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/2/3",
+      "iterations": 52469043,
+      "real_time": 1.5624869525055580e+01,
+      "cpu_time": 1.5158995752981452e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/2/4",
+      "iterations": 45512470,
+      "real_time": 1.5249180278513519e+01,
+      "cpu_time": 1.4811259419671170e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/2/6",
+      "iterations": 40479275,
+      "real_time": 1.7419527499445600e+01,
+      "cpu_time": 1.7267799386229168e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/2/7",
+      "iterations": 35497677,
+      "real_time": 1.9656466957520109e+01,
+      "cpu_time": 1.9009525609239173e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/2/12",
+      "iterations": 27042793,
+      "real_time": 2.4197042293219681e+01,
+      "cpu_time": 2.3855672008434897e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/2/16",
+      "iterations": 24077820,
+      "real_time": 2.8851556123593411e+01,
+      "cpu_time": 2.8603793865059156e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/2/20",
+      "iterations": 20977240,
+      "real_time": 3.2981163630488950e+01,
+      "cpu_time": 3.2740865814568508e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/3/1",
+      "iterations": 61687596,
+      "real_time": 1.1764908379779497e+01,
+      "cpu_time": 1.1687698123298562e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/3/2",
+      "iterations": 43755469,
+      "real_time": 1.6423369476207299e+01,
+      "cpu_time": 1.6309206970218952e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/3/3",
+      "iterations": 37693381,
+      "real_time": 1.8031485450223030e+01,
+      "cpu_time": 1.7915983710774054e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/3/4",
+      "iterations": 36812849,
+      "real_time": 1.9044860858008274e+01,
+      "cpu_time": 1.8915433575923469e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/3/6",
+      "iterations": 31380003,
+      "real_time": 2.2716422938798129e+01,
+      "cpu_time": 2.2573452271499175e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/3/7",
+      "iterations": 29979614,
+      "real_time": 2.5305456738143899e+01,
+      "cpu_time": 2.4368525892294780e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/3/12",
+      "iterations": 22574601,
+      "real_time": 3.2935694633936649e+01,
+      "cpu_time": 3.1806276443158403e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/3/16",
+      "iterations": 18534847,
+      "real_time": 3.7167092181413921e+01,
+      "cpu_time": 3.6952179858835549e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/3/20",
+      "iterations": 14379061,
+      "real_time": 4.9144608746192276e+01,
+      "cpu_time": 4.8366649254773812e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/4/1",
+      "iterations": 48585469,
+      "real_time": 1.4383808851738836e+01,
+      "cpu_time": 1.4028535980582999e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/4/2",
+      "iterations": 36885397,
+      "real_time": 1.9225010159014570e+01,
+      "cpu_time": 1.9118216349955460e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/4/3",
+      "iterations": 31222959,
+      "real_time": 2.2854289371385235e+01,
+      "cpu_time": 2.2695991113462323e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/4/4",
+      "iterations": 31518186,
+      "real_time": 2.2779098836662452e+01,
+      "cpu_time": 2.2609264378349625e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/4/6",
+      "iterations": 27870791,
+      "real_time": 2.6312030431589900e+01,
+      "cpu_time": 2.6164704116219813e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/4/7",
+      "iterations": 26635110,
+      "real_time": 2.8015773282089135e+01,
+      "cpu_time": 2.7610248277555474e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/4/12",
+      "iterations": 17797575,
+      "real_time": 3.9333237811242320e+01,
+      "cpu_time": 3.8665042849938892e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/4/16",
+      "iterations": 14592636,
+      "real_time": 4.8254791185539325e+01,
+      "cpu_time": 4.7936644208764051e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixVectorMultiply/4/20",
+      "iterations": 11582884,
+      "real_time": 6.1053182527591900e+01,
+      "cpu_time": 6.0339031280982766e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/1/1",
+      "iterations": 76783011,
+      "real_time": 9.3956197029191628e+00,
+      "cpu_time": 9.3391232078669315e+00,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/1/2",
+      "iterations": 70815082,
+      "real_time": 9.6104494527203634e+00,
+      "cpu_time": 9.5503808072974952e+00,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/1/3",
+      "iterations": 66582329,
+      "real_time": 1.0727912012247707e+01,
+      "cpu_time": 1.0640345728969692e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/1/4",
+      "iterations": 53793602,
+      "real_time": 1.3614292048217301e+01,
+      "cpu_time": 1.3500583210620450e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/1/6",
+      "iterations": 40321652,
+      "real_time": 1.7766203749914158e+01,
+      "cpu_time": 1.7570882264446855e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/1/7",
+      "iterations": 37201986,
+      "real_time": 1.9978757341790001e+01,
+      "cpu_time": 1.8992776353391505e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/1/12",
+      "iterations": 27434090,
+      "real_time": 2.6018652924849363e+01,
+      "cpu_time": 2.5589622254647360e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/1/16",
+      "iterations": 18985007,
+      "real_time": 3.8675490929263397e+01,
+      "cpu_time": 3.8356793863705086e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/1/20",
+      "iterations": 16250235,
+      "real_time": 4.2684055338167965e+01,
+      "cpu_time": 4.2342464585896622e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/2/1",
+      "iterations": 72217809,
+      "real_time": 9.8704761588567429e+00,
+      "cpu_time": 9.7949108370208933e+00,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/2/2",
+      "iterations": 58405854,
+      "real_time": 1.2449916972022011e+01,
+      "cpu_time": 1.2345851496324279e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/2/3",
+      "iterations": 44559025,
+      "real_time": 1.6004563317294398e+01,
+      "cpu_time": 1.5868367855894519e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/2/4",
+      "iterations": 35586284,
+      "real_time": 1.8873825462252668e+01,
+      "cpu_time": 1.8742417724761761e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/2/6",
+      "iterations": 30603233,
+      "real_time": 2.3746124174457755e+01,
+      "cpu_time": 2.3614759917685756e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/2/7",
+      "iterations": 25503512,
+      "real_time": 2.7041517616016900e+01,
+      "cpu_time": 2.6823050880208051e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/2/12",
+      "iterations": 17142745,
+      "real_time": 4.0841419326141207e+01,
+      "cpu_time": 4.0593790551046581e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/2/16",
+      "iterations": 13289037,
+      "real_time": 5.3266691859451711e+01,
+      "cpu_time": 5.2837914440301496e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/2/20",
+      "iterations": 10179301,
+      "real_time": 7.1242744951979475e+01,
+      "cpu_time": 6.9943702421217523e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/3/1",
+      "iterations": 69897251,
+      "real_time": 1.0455159873353184e+01,
+      "cpu_time": 1.0183404780826111e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/3/2",
+      "iterations": 48687523,
+      "real_time": 1.4473279406174216e+01,
+      "cpu_time": 1.4368362917127623e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/3/3",
+      "iterations": 36199282,
+      "real_time": 1.8768082416320095e+01,
+      "cpu_time": 1.8594457204980039e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/3/4",
+      "iterations": 32392711,
+      "real_time": 2.1793111421538484e+01,
+      "cpu_time": 2.1554787433506387e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/3/6",
+      "iterations": 25449918,
+      "real_time": 2.7307808579096385e+01,
+      "cpu_time": 2.7092071573668875e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/3/7",
+      "iterations": 22970703,
+      "real_time": 3.0684889528257191e+01,
+      "cpu_time": 3.0455576392242047e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/3/12",
+      "iterations": 15189357,
+      "real_time": 4.8938380408733977e+01,
+      "cpu_time": 4.7137676729830986e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/3/16",
+      "iterations": 12187690,
+      "real_time": 6.4202213625161093e+01,
+      "cpu_time": 6.1960141749584857e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/3/20",
+      "iterations": 8626958,
+      "real_time": 7.7660120865570065e+01,
+      "cpu_time": 7.6987740058547374e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/4/1",
+      "iterations": 68685362,
+      "real_time": 1.0764260905433320e+01,
+      "cpu_time": 1.0544197175520464e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/4/2",
+      "iterations": 41396113,
+      "real_time": 1.6881927031100783e+01,
+      "cpu_time": 1.6629145833088174e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/4/3",
+      "iterations": 35592798,
+      "real_time": 2.0520861720163161e+01,
+      "cpu_time": 2.0318998242284707e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/4/4",
+      "iterations": 29245508,
+      "real_time": 2.4356611964112989e+01,
+      "cpu_time": 2.4129585986333215e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/4/6",
+      "iterations": 20036065,
+      "real_time": 3.4106466514085753e+01,
+      "cpu_time": 3.3619725230478345e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/4/7",
+      "iterations": 18768417,
+      "real_time": 3.6661212077921938e+01,
+      "cpu_time": 3.6314090847405801e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/4/12",
+      "iterations": 11789871,
+      "real_time": 6.1986798330774114e+01,
+      "cpu_time": 6.1171322400389727e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/4/16",
+      "iterations": 7869236,
+      "real_time": 8.5527959635890710e+01,
+      "cpu_time": 8.2749964545479799e+01,
+      "time_unit": "ns"
+    },
+    {
+      "name": "BM_MatrixTransposeVectorMultiply/4/20",
+      "iterations": 7417847,
+      "real_time": 9.9073482499474821e+01,
+      "cpu_time": 9.7701125407411183e+01,
+      "time_unit": "ns"
+    }
+  ]
+}
diff --git a/internal/ceres/CMakeLists.txt b/internal/ceres/CMakeLists.txt
index 4080ea7..618bd72 100644
--- a/internal/ceres/CMakeLists.txt
+++ b/internal/ceres/CMakeLists.txt
@@ -383,7 +383,6 @@
   add_executable(autodiff_cost_function_benchmark autodiff_cost_function_benchmark.cc)
   target_link_libraries(autodiff_cost_function_benchmark ceres benchmark::benchmark)
 
-  add_executable(small_blas_benchmark small_blas_benchmark.cc)
-  target_link_libraries(small_blas_benchmark ceres benchmark::benchmark)
-
+  add_executable(small_blas_gemv_benchmark small_blas_gemv_benchmark.cc)
+  target_link_libraries(small_blas_gemv_benchmark ceres benchmark::benchmark)
 endif (BUILD_BENCHMARKS)
diff --git a/internal/ceres/small_blas_benchmark.cc b/internal/ceres/small_blas_benchmark.cc
deleted file mode 100644
index c78a69d..0000000
--- a/internal/ceres/small_blas_benchmark.cc
+++ /dev/null
@@ -1,296 +0,0 @@
-// Ceres Solver - A fast non-linear least squares minimizer
-// Copyright 2018 Google Inc. All rights reserved.
-// http://ceres-solver.org/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are met:
-//
-// * Redistributions of source code must retain the above copyright notice,
-//   this list of conditions and the following disclaimer.
-// * Redistributions in binary form must reproduce the above copyright notice,
-//   this list of conditions and the following disclaimer in the documentation
-//   and/or other materials provided with the distribution.
-// * Neither the name of Google Inc. nor the names of its contributors may be
-//   used to endorse or promote products derived from this software without
-//   specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
-// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
-// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
-// ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
-// LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
-// CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
-// SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
-// INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
-// CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
-// ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
-// POSSIBILITY OF SUCH DAMAGE.
-//
-// Authors: sameeragarwal@google.com (Sameer Agarwal)
-
-#include <iostream>
-#include "Eigen/Dense"
-#include "benchmark/benchmark.h"
-#include "ceres/small_blas.h"
-
-namespace ceres {
-
-// Benchmarking matrix-vector multiply routines and optimizing memory
-// access requires that we make sure that they are not just sitting in
-// the cache. So, as the benchmarking routine iterates, we need to
-// multiply new/different matrice and vectors. Allocating/creating
-// these objects in the benchmarking loop is too heavy duty, so we
-// create them before hand and cycle through them in the
-// benchmark. This class, given the size of the matrix creates such
-// matrix and vector objects for use in the benchmark.
-class MatrixVectorMultiplyData {
- public:
-  MatrixVectorMultiplyData(int rows, int cols) {
-    num_elements_ = 1000;
-    // A single memory buffer for all the matrices & vectors.
-
-    size_t buffer_size = num_elements_ * (200);
-    data_.resize(buffer_size, 1.00000000000001);
-
-    // Each element is three points, corresponding to the three
-    // elements of the expression c = A * b.
-    ptrs_.resize(3 * num_elements_, NULL);
-    double* p = &data_[0];
-    for (int i = 0; i < num_elements_; ++i) {
-      // Matrix X.
-      ptrs_[3 * i] = p;
-      p += rows * cols;
-      // Vector b.
-      ptrs_[3 * i + 1] = p;
-      p += cols;
-      // Vector c.
-      ptrs_[3 * i + 2] = p;
-      p += rows;
-    }
-  }
-
-  int num_elements() const { return num_elements_; }
-  double* data() { return &data_[0]; }
-  const std::vector<double*>& ptrs() const { return ptrs_; }
-
- private:
-  int num_elements_;
-  std::vector<double> data_;
-  std::vector<double*> ptrs_;
-};
-
-// Run on (8 X 2200 MHz CPU s)
-// 2018-02-06 21:23:59
-// ---------------------------------------------------------------------------
-// Benchmark                                    Time           CPU Iterations
-// ---------------------------------------------------------------------------
-// BM_MatrixVectorMultiplyDynamic/1/1           4 ns          4 ns  165611093
-// BM_MatrixVectorMultiplyDynamic/1/2           5 ns          5 ns  140648672
-// BM_MatrixVectorMultiplyDynamic/1/3           5 ns          5 ns  139414459
-// BM_MatrixVectorMultiplyDynamic/1/4           5 ns          5 ns  144247512
-// BM_MatrixVectorMultiplyDynamic/1/6           6 ns          6 ns  106639042
-// BM_MatrixVectorMultiplyDynamic/1/8           7 ns          7 ns  102367617
-// BM_MatrixVectorMultiplyDynamic/1/10          9 ns          9 ns   82419847
-// BM_MatrixVectorMultiplyDynamic/1/12         10 ns         10 ns   65129002
-// BM_MatrixVectorMultiplyDynamic/1/16         12 ns         12 ns   53500867
-// BM_MatrixVectorMultiplyDynamic/1/20         16 ns         16 ns   46067179
-// BM_MatrixVectorMultiplyDynamic/2/1           5 ns          5 ns  128880215
-// BM_MatrixVectorMultiplyDynamic/2/2           8 ns          8 ns   81938429
-// BM_MatrixVectorMultiplyDynamic/2/3          10 ns         10 ns   68807565
-// BM_MatrixVectorMultiplyDynamic/2/4           8 ns          8 ns   91833388
-// BM_MatrixVectorMultiplyDynamic/2/6          10 ns         10 ns   64031028
-// BM_MatrixVectorMultiplyDynamic/2/8          12 ns         12 ns   59788179
-// BM_MatrixVectorMultiplyDynamic/2/10         15 ns         15 ns   44737868
-// BM_MatrixVectorMultiplyDynamic/2/12         17 ns         17 ns   37423949
-// BM_MatrixVectorMultiplyDynamic/2/16         22 ns         22 ns   33470723
-// BM_MatrixVectorMultiplyDynamic/2/20         26 ns         26 ns   27076057
-// BM_MatrixVectorMultiplyDynamic/3/1           6 ns          6 ns  100932908
-// BM_MatrixVectorMultiplyDynamic/3/2          12 ns         12 ns   65591589
-// BM_MatrixVectorMultiplyDynamic/3/3          14 ns         14 ns   48182819
-// BM_MatrixVectorMultiplyDynamic/3/4          11 ns         11 ns   61770338
-// BM_MatrixVectorMultiplyDynamic/3/6          15 ns         15 ns   44712435
-// BM_MatrixVectorMultiplyDynamic/3/8          18 ns         18 ns   35177294
-// BM_MatrixVectorMultiplyDynamic/3/10         21 ns         21 ns   32164683
-// BM_MatrixVectorMultiplyDynamic/3/12         24 ns         24 ns   28222279
-// BM_MatrixVectorMultiplyDynamic/3/16         30 ns         30 ns   23050731
-// BM_MatrixVectorMultiplyDynamic/3/20         38 ns         38 ns   17832714
-// BM_MatrixVectorMultiplyDynamic/4/1           8 ns          8 ns   85763293
-// BM_MatrixVectorMultiplyDynamic/4/2          16 ns         16 ns   41959886
-// BM_MatrixVectorMultiplyDynamic/4/3          19 ns         19 ns   36674176
-// BM_MatrixVectorMultiplyDynamic/4/4          15 ns         15 ns   43561867
-// BM_MatrixVectorMultiplyDynamic/4/6          21 ns         21 ns   34278607
-// BM_MatrixVectorMultiplyDynamic/4/8          22 ns         22 ns   31484163
-// BM_MatrixVectorMultiplyDynamic/4/10         26 ns         26 ns   25605197
-// BM_MatrixVectorMultiplyDynamic/4/12         31 ns         31 ns   23380172
-// BM_MatrixVectorMultiplyDynamic/4/16         38 ns         38 ns   18054638
-// BM_MatrixVectorMultiplyDynamic/4/20         49 ns         49 ns   14771703
-void BM_MatrixVectorMultiplyDynamic(benchmark::State& state) {
-  const int rows = state.range(0);
-  const int cols = state.range(1);
-  MatrixVectorMultiplyData data(rows, cols);
-  const std::vector<double*> ptrs = data.ptrs();
-  const int num_elements = data.num_elements();
-
-  int i = 0;
-  for (auto _ : state) {
-    double* a_ptr = ptrs[3 * i];
-    double* b_ptr = ptrs[3 * i + 1];
-    double* c_ptr = ptrs[3 * i + 2];
-    internal::MatrixVectorMultiply<Eigen::Dynamic, Eigen::Dynamic, 1>(
-        a_ptr, rows, cols, b_ptr, c_ptr);
-    i = (i + 1) % num_elements;
-  }
-}
-
-// Each ArgPair specifies a row and column size of the matrix.
-BENCHMARK(BM_MatrixVectorMultiplyDynamic)
-->ArgPair(1, 1)
-->ArgPair(1, 2)
-->ArgPair(1, 3)
-->ArgPair(1, 4)
-->ArgPair(1, 6)
-->ArgPair(1, 8)
-->ArgPair(1, 10)
-->ArgPair(1, 12)
-->ArgPair(1, 16)
-->ArgPair(1, 20)
-->ArgPair(2, 1)
-->ArgPair(2, 2)
-->ArgPair(2, 3)
-->ArgPair(2, 4)
-->ArgPair(2, 6)
-->ArgPair(2, 8)
-->ArgPair(2, 10)
-->ArgPair(2, 12)
-->ArgPair(2, 16)
-->ArgPair(2, 20)
-->ArgPair(3, 1)
-->ArgPair(3, 2)
-->ArgPair(3, 3)
-->ArgPair(3, 4)
-->ArgPair(3, 6)
-->ArgPair(3, 8)
-->ArgPair(3, 10)
-->ArgPair(3, 12)
-->ArgPair(3, 16)
-->ArgPair(3, 20)
-->ArgPair(4, 1)
-->ArgPair(4, 2)
-->ArgPair(4, 3)
-->ArgPair(4, 4)
-->ArgPair(4, 6)
-->ArgPair(4, 8)
-->ArgPair(4, 10)
-->ArgPair(4, 12)
-->ArgPair(4, 16)
-->ArgPair(4, 20);
-
-// Run on (8 X 2200 MHz CPU s)
-// 2018-02-06 21:18:17
-// ------------------------------------------------------------------------------------
-// Benchmark                                             Time           CPU Iterations
-// ------------------------------------------------------------------------------------
-// BM_MatrixTransposeVectorMultiplyDynamic/1/1           5 ns          5 ns  139356174
-// BM_MatrixTransposeVectorMultiplyDynamic/1/2           6 ns          6 ns  120800041
-// BM_MatrixTransposeVectorMultiplyDynamic/1/3           7 ns          7 ns  100267858
-// BM_MatrixTransposeVectorMultiplyDynamic/1/4           9 ns          9 ns   70778564
-// BM_MatrixTransposeVectorMultiplyDynamic/1/6          14 ns         14 ns   47748651
-// BM_MatrixTransposeVectorMultiplyDynamic/1/8          16 ns         16 ns   43903663
-// BM_MatrixTransposeVectorMultiplyDynamic/1/10         18 ns         18 ns   34838177
-// BM_MatrixTransposeVectorMultiplyDynamic/1/12         20 ns         20 ns   36138731
-// BM_MatrixTransposeVectorMultiplyDynamic/1/16         23 ns         23 ns   27063704
-// BM_MatrixTransposeVectorMultiplyDynamic/1/20         29 ns         29 ns   23400336
-// BM_MatrixTransposeVectorMultiplyDynamic/2/1           6 ns          6 ns  121572101
-// BM_MatrixTransposeVectorMultiplyDynamic/2/2           8 ns          8 ns   82896155
-// BM_MatrixTransposeVectorMultiplyDynamic/2/3          12 ns         12 ns   56705415
-// BM_MatrixTransposeVectorMultiplyDynamic/2/4          14 ns         14 ns   51241509
-// BM_MatrixTransposeVectorMultiplyDynamic/2/6          18 ns         18 ns   38377403
-// BM_MatrixTransposeVectorMultiplyDynamic/2/8          25 ns         25 ns   28560121
-// BM_MatrixTransposeVectorMultiplyDynamic/2/10         29 ns         29 ns   23608052
-// BM_MatrixTransposeVectorMultiplyDynamic/2/12         33 ns         33 ns   20668478
-// BM_MatrixTransposeVectorMultiplyDynamic/2/16         44 ns         44 ns   16335446
-// BM_MatrixTransposeVectorMultiplyDynamic/2/20         53 ns         53 ns   13462315
-// BM_MatrixTransposeVectorMultiplyDynamic/3/1           6 ns          6 ns  117031415
-// BM_MatrixTransposeVectorMultiplyDynamic/3/2          10 ns         10 ns   71040747
-// BM_MatrixTransposeVectorMultiplyDynamic/3/3          14 ns         14 ns   49453538
-// BM_MatrixTransposeVectorMultiplyDynamic/3/4          17 ns         17 ns   39161935
-// BM_MatrixTransposeVectorMultiplyDynamic/3/6          22 ns         22 ns   32118490
-// BM_MatrixTransposeVectorMultiplyDynamic/3/8          28 ns         28 ns   25295689
-// BM_MatrixTransposeVectorMultiplyDynamic/3/10         34 ns         34 ns   20900389
-// BM_MatrixTransposeVectorMultiplyDynamic/3/12         39 ns         39 ns   17934922
-// BM_MatrixTransposeVectorMultiplyDynamic/3/16         51 ns         51 ns   10000000
-// BM_MatrixTransposeVectorMultiplyDynamic/3/20         64 ns         64 ns   10594824
-// BM_MatrixTransposeVectorMultiplyDynamic/4/1           7 ns          7 ns   98903583
-// BM_MatrixTransposeVectorMultiplyDynamic/4/2          13 ns         13 ns   57301899
-// BM_MatrixTransposeVectorMultiplyDynamic/4/3          16 ns         16 ns   44622083
-// BM_MatrixTransposeVectorMultiplyDynamic/4/4          18 ns         18 ns   39645007
-// BM_MatrixTransposeVectorMultiplyDynamic/4/6          26 ns         26 ns   27239262
-// BM_MatrixTransposeVectorMultiplyDynamic/4/8          33 ns         33 ns   20869171
-// BM_MatrixTransposeVectorMultiplyDynamic/4/10         39 ns         39 ns   17169614
-// BM_MatrixTransposeVectorMultiplyDynamic/4/12         47 ns         47 ns   15045286
-// BM_MatrixTransposeVectorMultiplyDynamic/4/16         62 ns         62 ns   11437535
-// BM_MatrixTransposeVectorMultiplyDynamic/4/20         77 ns         77 ns    8351428
-void BM_MatrixTransposeVectorMultiplyDynamic(benchmark::State& state) {
-  const int rows = state.range(0);
-  const int cols = state.range(1);
-  MatrixVectorMultiplyData data(rows, cols);
-  const std::vector<double*> ptrs = data.ptrs();
-  const int num_elements = data.num_elements();
-
-  int i = 0;
-  for (auto _ : state) {
-    double* a_ptr = ptrs[3 * i];
-    double* b_ptr = ptrs[3 * i + 1];
-    double* c_ptr = ptrs[3 * i + 2];
-    internal::MatrixTransposeVectorMultiply<Eigen::Dynamic, Eigen::Dynamic, 1>(
-        a_ptr, rows, cols, c_ptr, b_ptr);
-    i = (i + 1) % num_elements;
-  }
-}
-
-// Each ArgPair specifies a row and column size of the matrix.
-BENCHMARK(BM_MatrixTransposeVectorMultiplyDynamic)
-->ArgPair(1, 1)
-->ArgPair(1, 2)
-->ArgPair(1, 3)
-->ArgPair(1, 4)
-->ArgPair(1, 6)
-->ArgPair(1, 8)
-->ArgPair(1, 10)
-->ArgPair(1, 12)
-->ArgPair(1, 16)
-->ArgPair(1, 20)
-->ArgPair(2, 1)
-->ArgPair(2, 2)
-->ArgPair(2, 3)
-->ArgPair(2, 4)
-->ArgPair(2, 6)
-->ArgPair(2, 8)
-->ArgPair(2, 10)
-->ArgPair(2, 12)
-->ArgPair(2, 16)
-->ArgPair(2, 20)
-->ArgPair(3, 1)
-->ArgPair(3, 2)
-->ArgPair(3, 3)
-->ArgPair(3, 4)
-->ArgPair(3, 6)
-->ArgPair(3, 8)
-->ArgPair(3, 10)
-->ArgPair(3, 12)
-->ArgPair(3, 16)
-->ArgPair(3, 20)
-->ArgPair(4, 1)
-->ArgPair(4, 2)
-->ArgPair(4, 3)
-->ArgPair(4, 4)
-->ArgPair(4, 6)
-->ArgPair(4, 8)
-->ArgPair(4, 10)
-->ArgPair(4, 12)
-->ArgPair(4, 16)
-->ArgPair(4, 20);
-
-}  // namespace ceres
-
-BENCHMARK_MAIN();
diff --git a/internal/ceres/small_blas_gemv_benchmark.cc b/internal/ceres/small_blas_gemv_benchmark.cc
new file mode 100644
index 0000000..513ff12
--- /dev/null
+++ b/internal/ceres/small_blas_gemv_benchmark.cc
@@ -0,0 +1,116 @@
+// Ceres Solver - A fast non-linear least squares minimizer
+// Copyright 2018 Google Inc. All rights reserved.
+// http://ceres-solver.org/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are met:
+//
+// * Redistributions of source code must retain the above copyright notice,
+//   this list of conditions and the following disclaimer.
+// * Redistributions in binary form must reproduce the above copyright notice,
+//   this list of conditions and the following disclaimer in the documentation
+//   and/or other materials provided with the distribution.
+// * Neither the name of Google Inc. nor the names of its contributors may be
+//   used to endorse or promote products derived from this software without
+//   specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+// ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+// LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+// CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+// SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+// INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+// CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+// ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+// POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: sameeragarwal@google.com (Sameer Agarwal)
+
+#include "Eigen/Dense"
+#include "benchmark/benchmark.h"
+#include "ceres/small_blas.h"
+
+namespace ceres {
+
+// Benchmarking matrix-vector multiply routines and optimizing memory
+// access requires that we make sure that they are not just sitting in
+// the cache. So, as the benchmarking routine iterates, we need to
+// multiply new/different matrice and vectors. Allocating/creating
+// these objects in the benchmarking loop is too heavy duty, so we
+// create them before hand and cycle through them in the
+// benchmark. This class, given the size of the matrix creates such
+// matrix and vector objects for use in the benchmark.
+class MatrixVectorMultiplyData {
+ public:
+  MatrixVectorMultiplyData(int rows, int cols) {
+    rows_ = rows;
+    cols_ = cols;
+
+    num_elements_ = 1000;
+    a_.resize(num_elements_ * rows, 1.00001);
+    b_.resize(num_elements_ * rows * cols, 1.00002);
+    c_.resize(num_elements_ * cols, 1.00003);
+  }
+
+  int num_elements() const { return num_elements_; }
+  double* GetA(int i) { return &a_[i * rows_]; };
+  double* GetB(int i) { return &b_[i * rows_ * cols_]; };
+  double* GetC(int i) { return &c_[i * cols_]; };
+
+ private:
+  int num_elements_;
+  int rows_;
+  int cols_;
+  std::vector<double> a_;
+  std::vector<double> b_;
+  std::vector<double> c_;
+};
+
+// Helper function to generate the various matrix sizes for which we
+// run the benchmark.
+static void MatrixSizeArguments(benchmark::internal::Benchmark* benchmark) {
+  std::vector<int> rows = {1, 2, 3, 4};
+  std::vector<int> cols = {1, 2, 3, 4, 6, 7, 12, 16, 20};
+  for (int r : rows) {
+    for (int c : cols) {
+      benchmark->Args({r, c});
+    }
+  }
+}
+
+void BM_MatrixVectorMultiply(benchmark::State& state) {
+  const int rows = state.range(0);
+  const int cols = state.range(1);
+  MatrixVectorMultiplyData data(rows, cols);
+  const int num_elements = data.num_elements();
+  int i = 0;
+  for (auto _ : state) {
+    // A += B * C;
+    internal::MatrixVectorMultiply<Eigen::Dynamic, Eigen::Dynamic, 1>(
+        data.GetB(i), rows, cols, data.GetC(i), data.GetA(i));
+    i = (i + 1) % num_elements;
+  }
+}
+
+BENCHMARK(BM_MatrixVectorMultiply)->Apply(MatrixSizeArguments);
+
+void BM_MatrixTransposeVectorMultiply(benchmark::State& state) {
+  const int rows = state.range(0);
+  const int cols = state.range(1);
+  MatrixVectorMultiplyData data(cols, rows);
+  const int num_elements = data.num_elements();
+  int i = 0;
+  for (auto _ : state) {
+    internal::MatrixTransposeVectorMultiply<Eigen::Dynamic, Eigen::Dynamic, 1>(
+        data.GetB(i), rows, cols, data.GetC(i), data.GetA(i));
+    i = (i + 1) % num_elements;
+  }
+}
+
+BENCHMARK(BM_MatrixTransposeVectorMultiply)->Apply(MatrixSizeArguments);
+
+}  // namespace ceres
+
+BENCHMARK_MAIN();