diff --git a/benchmarks/dot/dot.f90 b/benchmarks/dot/dot.f90
index d95aea2..4342ca6 100644
--- a/benchmarks/dot/dot.f90
+++ b/benchmarks/dot/dot.f90
@@ -1,7 +1,7 @@
 program benchmark_dot
 
    use kinds
-   use fordot
+   use fordot, only: fdot_product => dot_product ! to avoid overloading
    use fast_math, only: fprod, fprod_kahan
    use forbenchmark
 
@@ -19,7 +19,7 @@ program benchmark_dot
    allocate(seed_array(seed_size))
    seed_array = 123456789
 
-   call bench%init(7,'Benchmark dot_product','benchmarks/dot/results/dot', 10000)
+   call bench%init(8,'Benchmark dot_product','benchmarks/dot/results/dot', 10000)
 
    num_elements = [1000_ik, 10000_ik, 100000_ik, 1000000_ik]
 
@@ -48,9 +48,9 @@ program benchmark_dot
 
 
       !===============================================================================
-      call bench%start_benchmark(2,'m1', "a = dot_product(u,v,'m1')",[p])
+      call bench%start_benchmark(2,'blas', "a = dot_product(u,v,'m2')",[p])
       do nl = 1,bench%nloops
-         a = dot_product(u,v,'m1')
+         a = fdot_product(u,v,'m2')
          call prevent_optimization(a,nl) ! loop-invariant
       end do
       call bench%stop_benchmark(cmp_gflops)
@@ -58,9 +58,9 @@ program benchmark_dot
 
 
       !===============================================================================
-      call bench%start_benchmark(3,'m2', "a = dot_product(u,v,'m2')",[p])
+      call bench%start_benchmark(3,'m1_b16', "a = f(u,v,'m1',16)",[p])
       do nl = 1,bench%nloops
-         a = dot_product(u,v,'m2')
+         a = fdot_product(u,v,'m1',16)
          call prevent_optimization(a,nl) ! loop-invariant
       end do
       call bench%stop_benchmark(cmp_gflops)
@@ -68,9 +68,9 @@ program benchmark_dot
 
 
       !===============================================================================
-      call bench%start_benchmark(4,'m3', "a = dot_product(u,v,'m3')",[p])
+      call bench%start_benchmark(4,'blas_b16', "a = dot_product(u,v,'m2',16)",[p])
       do nl = 1,bench%nloops
-         a = dot_product(u,v,'m3')
+         a = fdot_product(u,v,'m2',16)
          call prevent_optimization(a,nl) ! loop-invariant
       end do
       call bench%stop_benchmark(cmp_gflops)
@@ -78,9 +78,9 @@ program benchmark_dot
 
 
       !===============================================================================
-      call bench%start_benchmark(5,'m4', "a = dot_product(u,v,'m4')",[p])
+      call bench%start_benchmark(5,'m3_b16', "a = dot_product(u,v,'m3',16)",[p])
       do nl = 1,bench%nloops
-         a = dot_product(u,v,'m4')
+         a = fdot_product(u,v,'m3',16)
          call prevent_optimization(a,nl) ! loop-invariant
       end do
       call bench%stop_benchmark(cmp_gflops)
@@ -88,7 +88,17 @@ program benchmark_dot
 
 
       !===============================================================================
-      call bench%start_benchmark(6,'chunks', "a = fprod(u,v)",[p])
+      call bench%start_benchmark(6,'m4_b16', "a = dot_product(u,v,'m4',16)",[p])
+      do nl = 1,bench%nloops
+         a = fdot_product(u,v,'m4',16)
+         call prevent_optimization(a,nl) ! loop-invariant
+      end do
+      call bench%stop_benchmark(cmp_gflops)
+      !===============================================================================
+
+
+      !===============================================================================
+      call bench%start_benchmark(7,'chunks', "a = fprod(u,v)",[p])
       do nl = 1,bench%nloops
          a = fprod(u,v)
          call prevent_optimization(a,nl) ! loop-invariant
@@ -98,7 +108,7 @@ program benchmark_dot
 
 
       !===============================================================================
-      call bench%start_benchmark(7,'kahan', "a = fprod_kahan(u,v)",[p])
+      call bench%start_benchmark(8,'kahan', "a = fprod_kahan(u,v)",[p])
       do nl = 1,bench%nloops
          a = fprod_kahan(u,v)
          call prevent_optimization(a,nl) ! loop-invariant
@@ -128,7 +138,7 @@ end function cmp_gflops
    subroutine prevent_optimization(a, nl)
       real(rk), intent(in) :: a
       integer, intent(in)  :: nl
-      if (a == 0.0_rk) print*, nl, 'a = 0.0'
+      if (abs(a)<tiny(0.0_rk)) print*, nl, 'a = 0.0'
    end subroutine prevent_optimization
    !===============================================================================
 
diff --git a/benchmarks/dot/dot_co.f90 b/benchmarks/dot/dot_co.f90
index 1be39da..ade0c19 100644
--- a/benchmarks/dot/dot_co.f90
+++ b/benchmarks/dot/dot_co.f90
@@ -3,7 +3,7 @@ program benchmark_dot_coarray
 #if defined(USE_COARRAY)
 
    use kinds
-   use fordot
+   use fordot, only: fdot_product => dot_product ! to avoid overloading
    use forbenchmark
 
    implicit none
@@ -21,7 +21,7 @@ program benchmark_dot_coarray
    allocate(seed_array(seed_size))
    seed_array = 123456789
 
-   call bench%init(4, 'Fordot_product','benchmarks/dot/results/dot', 1000)
+   call bench%init(6, 'Fordot_product','benchmarks/dot/results/dot', 1000)
 
    num_elements = [1000_ik, 10000_ik, 100000_ik, 1000000_ik]
 
@@ -52,7 +52,7 @@ program benchmark_dot_coarray
       !===============================================================================
       call bench%start_benchmark(2,'m1_co', "a = dot_product(u,v,'m1',coarray=.true.)",[p])
       do nl = 1,bench%nloops
-         a = dot_product(u,v,'m1',coarray=.true.)
+         a = fdot_product(u,v,'m1',coarray=.true.)
          call prevent_optimization(a,nl) ! loop-invariant
       end do
       call bench%stop_benchmark(cmp_gflops)
@@ -62,7 +62,7 @@ program benchmark_dot_coarray
       !===============================================================================
       call bench%start_benchmark(3,'m2_co', "a = dot_product(u,v,'m2',coarray=.true.)",[p])
       do nl = 1,bench%nloops
-         a = dot_product(u,v,'m2',coarray=.true.)
+         a = fdot_product(u,v,'m2',coarray=.true.)
          call prevent_optimization(a,nl) ! loop-invariant
       end do
       call bench%stop_benchmark(cmp_gflops)
@@ -72,7 +72,7 @@ program benchmark_dot_coarray
       !===============================================================================
       call bench%start_benchmark(4,'m3_co', "a = dot_product(u,v,'m3',coarray=.true.)",[p])
       do nl = 1,bench%nloops
-         a = dot_product(u,v,'m3',coarray=.true.)
+         a = fdot_product(u,v,'m3',coarray=.true.)
          call prevent_optimization(a,nl) ! loop-invariant
       end do
       call bench%stop_benchmark(cmp_gflops)
diff --git a/benchmarks/dot/results/dot_gfortran.html b/benchmarks/dot/results/dot_gfortran.html
index c0154b6..e9f3fdb 100644
--- a/benchmarks/dot/results/dot_gfortran.html
+++ b/benchmarks/dot/results/dot_gfortran.html
@@ -40,48 +40,58 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>2.400000e-06</td>
-      <td>0.416667</td>
+      <td>1.600000e-06</td>
+      <td>0.625000</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
     </tr>
     <tr>
-      <td>m1</td>
-      <td>1.043478</td>
-      <td>2.300000e-06</td>
-      <td>0.434783</td>
+      <td>blas</td>
+      <td>5.333333</td>
+      <td>3.000000e-07</td>
+      <td>3.333333</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
     </tr>
     <tr>
-      <td>m2</td>
-      <td>12.000000</td>
-      <td>2.000000e-07</td>
-      <td>5.000000</td>
+      <td>m1_b16</td>
+      <td>0.571429</td>
+      <td>2.800000e-06</td>
+      <td>0.357143</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
     </tr>
     <tr>
-      <td>m3</td>
-      <td>1.142857</td>
-      <td>2.100000e-06</td>
-      <td>0.476190</td>
+      <td>blas_b16</td>
+      <td>0.842105</td>
+      <td>1.900000e-06</td>
+      <td>0.526316</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
     </tr>
     <tr>
-      <td>m4</td>
-      <td>1.043478</td>
-      <td>2.300000e-06</td>
-      <td>0.434783</td>
+      <td>m3_b16</td>
+      <td>0.571429</td>
+      <td>2.800000e-06</td>
+      <td>0.357143</td>
+      <td>10000</td>
+      <td>1000</td>
+      <td>1000</td>
+      <td>1000</td>
+    </tr>
+    <tr>
+      <td>m4_b16</td>
+      <td>0.592593</td>
+      <td>2.700000e-06</td>
+      <td>0.370370</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
@@ -90,8 +100,8 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>chunks</td>
       <td>8.000000</td>
-      <td>3.000000e-07</td>
-      <td>3.333333</td>
+      <td>2.000000e-07</td>
+      <td>5.000000</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
@@ -99,7 +109,7 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>kahan</td>
-      <td>6.000000</td>
+      <td>4.000000</td>
       <td>4.000000e-07</td>
       <td>2.500000</td>
       <td>10000</td>
@@ -110,48 +120,58 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>1.750000e-05</td>
-      <td>0.571429</td>
+      <td>1.670000e-05</td>
+      <td>0.598802</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
     </tr>
     <tr>
-      <td>m1</td>
-      <td>1.023392</td>
-      <td>1.710000e-05</td>
-      <td>0.584795</td>
+      <td>blas</td>
+      <td>7.260870</td>
+      <td>2.300000e-06</td>
+      <td>4.347826</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
     </tr>
     <tr>
-      <td>m2</td>
-      <td>5.645161</td>
-      <td>3.100000e-06</td>
-      <td>3.225806</td>
+      <td>m1_b16</td>
+      <td>0.970930</td>
+      <td>1.720000e-05</td>
+      <td>0.581395</td>
+      <td>10000</td>
+      <td>10000</td>
+      <td>10000</td>
+      <td>10000</td>
+    </tr>
+    <tr>
+      <td>blas_b16</td>
+      <td>5.218750</td>
+      <td>3.200000e-06</td>
+      <td>3.125000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
     </tr>
     <tr>
-      <td>m3</td>
-      <td>1.035503</td>
-      <td>1.690000e-05</td>
-      <td>0.591716</td>
+      <td>m3_b16</td>
+      <td>0.976608</td>
+      <td>1.710000e-05</td>
+      <td>0.584795</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
     </tr>
     <tr>
-      <td>m4</td>
-      <td>1.041667</td>
-      <td>1.680000e-05</td>
-      <td>0.595238</td>
+      <td>m4_b16</td>
+      <td>0.959770</td>
+      <td>1.740000e-05</td>
+      <td>0.574713</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
@@ -159,9 +179,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>chunks</td>
-      <td>5.468750</td>
-      <td>3.200000e-06</td>
-      <td>3.125000</td>
+      <td>6.680000</td>
+      <td>2.500000e-06</td>
+      <td>4.000000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
@@ -169,9 +189,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>kahan</td>
-      <td>4.605263</td>
-      <td>3.800000e-06</td>
-      <td>2.631579</td>
+      <td>5.387097</td>
+      <td>3.100000e-06</td>
+      <td>3.225806</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
@@ -180,48 +200,58 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>1.680000e-04</td>
-      <td>0.595238</td>
+      <td>1.674000e-04</td>
+      <td>0.597372</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
       <td>100000</td>
     </tr>
     <tr>
-      <td>m1</td>
-      <td>0.998217</td>
-      <td>1.683000e-04</td>
-      <td>0.594177</td>
+      <td>blas</td>
+      <td>9.732558</td>
+      <td>1.720000e-05</td>
+      <td>5.813953</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
       <td>100000</td>
     </tr>
     <tr>
-      <td>m2</td>
-      <td>12.824427</td>
-      <td>1.310000e-05</td>
-      <td>7.633588</td>
+      <td>m1_b16</td>
+      <td>0.989362</td>
+      <td>1.692000e-04</td>
+      <td>0.591017</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
       <td>100000</td>
     </tr>
     <tr>
-      <td>m3</td>
-      <td>0.993495</td>
-      <td>1.691000e-04</td>
-      <td>0.591366</td>
+      <td>blas_b16</td>
+      <td>3.647059</td>
+      <td>4.590000e-05</td>
+      <td>2.178649</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
       <td>100000</td>
     </tr>
     <tr>
-      <td>m4</td>
-      <td>0.987654</td>
-      <td>1.701000e-04</td>
-      <td>0.587889</td>
+      <td>m3_b16</td>
+      <td>0.998807</td>
+      <td>1.676000e-04</td>
+      <td>0.596659</td>
+      <td>10000</td>
+      <td>100000</td>
+      <td>100000</td>
+      <td>100000</td>
+    </tr>
+    <tr>
+      <td>m4_b16</td>
+      <td>0.997616</td>
+      <td>1.678000e-04</td>
+      <td>0.595948</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
@@ -229,9 +259,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>chunks</td>
-      <td>3.193916</td>
-      <td>5.260000e-05</td>
-      <td>1.901141</td>
+      <td>3.607759</td>
+      <td>4.640000e-05</td>
+      <td>2.155172</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
@@ -239,9 +269,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>kahan</td>
-      <td>2.957746</td>
-      <td>5.680000e-05</td>
-      <td>1.760563</td>
+      <td>3.348000</td>
+      <td>5.000000e-05</td>
+      <td>2.000000</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
@@ -250,48 +280,58 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>1.716600e-03</td>
-      <td>0.582547</td>
+      <td>1.723200e-03</td>
+      <td>0.580316</td>
+      <td>10000</td>
+      <td>1000000</td>
+      <td>1000000</td>
+      <td>1000000</td>
+    </tr>
+    <tr>
+      <td>blas</td>
+      <td>8.082552</td>
+      <td>2.132000e-04</td>
+      <td>4.690432</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
       <td>1000000</td>
     </tr>
     <tr>
-      <td>m1</td>
-      <td>1.000117</td>
-      <td>1.716400e-03</td>
-      <td>0.582615</td>
+      <td>m1_b16</td>
+      <td>0.998783</td>
+      <td>1.725300e-03</td>
+      <td>0.579609</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
       <td>1000000</td>
     </tr>
     <tr>
-      <td>m2</td>
-      <td>10.038596</td>
-      <td>1.710000e-04</td>
-      <td>5.847953</td>
+      <td>blas_b16</td>
+      <td>6.216450</td>
+      <td>2.772000e-04</td>
+      <td>3.607504</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
       <td>1000000</td>
     </tr>
     <tr>
-      <td>m3</td>
-      <td>0.999011</td>
-      <td>1.718300e-03</td>
-      <td>0.581971</td>
+      <td>m3_b16</td>
+      <td>0.998378</td>
+      <td>1.726000e-03</td>
+      <td>0.579374</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
       <td>1000000</td>
     </tr>
     <tr>
-      <td>m4</td>
-      <td>1.003566</td>
-      <td>1.710500e-03</td>
-      <td>0.584624</td>
+      <td>m4_b16</td>
+      <td>0.999246</td>
+      <td>1.724500e-03</td>
+      <td>0.579878</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
@@ -299,9 +339,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>chunks</td>
-      <td>2.615572</td>
-      <td>6.563000e-04</td>
-      <td>1.523693</td>
+      <td>2.691238</td>
+      <td>6.403000e-04</td>
+      <td>1.561768</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
@@ -309,9 +349,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>kahan</td>
-      <td>2.418087</td>
-      <td>7.099000e-04</td>
-      <td>1.408649</td>
+      <td>2.487657</td>
+      <td>6.927000e-04</td>
+      <td>1.443626</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
diff --git a/benchmarks/dot/results/dot_gfortran_perf.png b/benchmarks/dot/results/dot_gfortran_perf.png
index d3832c8..bf334be 100644
Binary files a/benchmarks/dot/results/dot_gfortran_perf.png and b/benchmarks/dot/results/dot_gfortran_perf.png differ
diff --git a/benchmarks/dot/results/dot_gfortran_speedup.png b/benchmarks/dot/results/dot_gfortran_speedup.png
index edb05a0..a6be2b9 100644
Binary files a/benchmarks/dot/results/dot_gfortran_speedup.png and b/benchmarks/dot/results/dot_gfortran_speedup.png differ
diff --git a/benchmarks/dot/results/dot_gfortran_speedup_avg.png b/benchmarks/dot/results/dot_gfortran_speedup_avg.png
index 546d07d..6df0599 100644
Binary files a/benchmarks/dot/results/dot_gfortran_speedup_avg.png and b/benchmarks/dot/results/dot_gfortran_speedup_avg.png differ
diff --git a/benchmarks/dot/results/dot_gfortran_time.png b/benchmarks/dot/results/dot_gfortran_time.png
index 7915ecb..f35431d 100644
Binary files a/benchmarks/dot/results/dot_gfortran_time.png and b/benchmarks/dot/results/dot_gfortran_time.png differ
diff --git a/benchmarks/dot/results/dot_ifort.html b/benchmarks/dot/results/dot_ifort.html
index 2be7264..58c5f66 100644
--- a/benchmarks/dot/results/dot_ifort.html
+++ b/benchmarks/dot/results/dot_ifort.html
@@ -48,40 +48,50 @@ <h2>dot_product Benchmark</h2>
       <td>1000</td>
     </tr>
     <tr>
-      <td>m1</td>
-      <td>0.614545</td>
-      <td>2.750000e-06</td>
-      <td>0.363636</td>
+      <td>blas</td>
+      <td>6.259259</td>
+      <td>2.700000e-07</td>
+      <td>3.703704</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
     </tr>
     <tr>
-      <td>m2</td>
-      <td>6.500000</td>
-      <td>2.600000e-07</td>
-      <td>3.846154</td>
+      <td>m1_b16</td>
+      <td>0.710084</td>
+      <td>2.380000e-06</td>
+      <td>0.420168</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
     </tr>
     <tr>
-      <td>m3</td>
-      <td>0.612319</td>
-      <td>2.760000e-06</td>
-      <td>0.362319</td>
+      <td>blas_b16</td>
+      <td>0.880208</td>
+      <td>1.920000e-06</td>
+      <td>0.520833</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
     </tr>
     <tr>
-      <td>m4</td>
-      <td>0.866667</td>
-      <td>1.950000e-06</td>
-      <td>0.512821</td>
+      <td>m3_b16</td>
+      <td>0.590909</td>
+      <td>2.860000e-06</td>
+      <td>0.349650</td>
+      <td>10000</td>
+      <td>1000</td>
+      <td>1000</td>
+      <td>1000</td>
+    </tr>
+    <tr>
+      <td>m4_b16</td>
+      <td>0.637736</td>
+      <td>2.650000e-06</td>
+      <td>0.377358</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
@@ -89,9 +99,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>chunks</td>
-      <td>9.388889</td>
-      <td>1.800000e-07</td>
-      <td>5.555556</td>
+      <td>9.941176</td>
+      <td>1.700000e-07</td>
+      <td>5.882353</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
@@ -99,9 +109,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>kahan</td>
-      <td>3.673913</td>
-      <td>4.600000e-07</td>
-      <td>2.173913</td>
+      <td>3.595745</td>
+      <td>4.700000e-07</td>
+      <td>2.127660</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
@@ -110,48 +120,58 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>1.677000e-05</td>
-      <td>0.596303</td>
+      <td>1.674000e-05</td>
+      <td>0.597372</td>
+      <td>10000</td>
+      <td>10000</td>
+      <td>10000</td>
+      <td>10000</td>
+    </tr>
+    <tr>
+      <td>blas</td>
+      <td>5.978571</td>
+      <td>2.800000e-06</td>
+      <td>3.571429</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
     </tr>
     <tr>
-      <td>m1</td>
-      <td>0.997620</td>
-      <td>1.681000e-05</td>
-      <td>0.594884</td>
+      <td>m1_b16</td>
+      <td>0.996429</td>
+      <td>1.680000e-05</td>
+      <td>0.595238</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
     </tr>
     <tr>
-      <td>m2</td>
-      <td>6.142857</td>
-      <td>2.730000e-06</td>
-      <td>3.663004</td>
+      <td>blas_b16</td>
+      <td>4.755682</td>
+      <td>3.520000e-06</td>
+      <td>2.840909</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
     </tr>
     <tr>
-      <td>m3</td>
-      <td>0.995252</td>
-      <td>1.685000e-05</td>
-      <td>0.593472</td>
+      <td>m3_b16</td>
+      <td>0.959862</td>
+      <td>1.744000e-05</td>
+      <td>0.573394</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
     </tr>
     <tr>
-      <td>m4</td>
-      <td>0.994072</td>
-      <td>1.687000e-05</td>
-      <td>0.592768</td>
+      <td>m4_b16</td>
+      <td>0.958214</td>
+      <td>1.747000e-05</td>
+      <td>0.572410</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
@@ -159,9 +179,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>chunks</td>
-      <td>6.576471</td>
-      <td>2.550000e-06</td>
-      <td>3.921569</td>
+      <td>6.389313</td>
+      <td>2.620000e-06</td>
+      <td>3.816794</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
@@ -169,9 +189,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>kahan</td>
-      <td>3.701987</td>
-      <td>4.530000e-06</td>
-      <td>2.207506</td>
+      <td>3.623377</td>
+      <td>4.620000e-06</td>
+      <td>2.164502</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
@@ -180,48 +200,58 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>1.673300e-04</td>
-      <td>0.597621</td>
+      <td>1.674000e-04</td>
+      <td>0.597372</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
       <td>100000</td>
     </tr>
     <tr>
-      <td>m1</td>
-      <td>0.999164</td>
-      <td>1.674700e-04</td>
-      <td>0.597122</td>
+      <td>blas</td>
+      <td>12.381657</td>
+      <td>1.352000e-05</td>
+      <td>7.396450</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
       <td>100000</td>
     </tr>
     <tr>
-      <td>m2</td>
-      <td>14.639545</td>
-      <td>1.143000e-05</td>
-      <td>8.748906</td>
+      <td>m1_b16</td>
+      <td>0.993295</td>
+      <td>1.685300e-04</td>
+      <td>0.593366</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
       <td>100000</td>
     </tr>
     <tr>
-      <td>m3</td>
-      <td>0.998032</td>
-      <td>1.676600e-04</td>
-      <td>0.596445</td>
+      <td>blas_b16</td>
+      <td>3.674276</td>
+      <td>4.556000e-05</td>
+      <td>2.194908</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
       <td>100000</td>
     </tr>
     <tr>
-      <td>m4</td>
-      <td>0.998985</td>
-      <td>1.675000e-04</td>
-      <td>0.597015</td>
+      <td>m3_b16</td>
+      <td>0.991002</td>
+      <td>1.689200e-04</td>
+      <td>0.591996</td>
+      <td>10000</td>
+      <td>100000</td>
+      <td>100000</td>
+      <td>100000</td>
+    </tr>
+    <tr>
+      <td>m4_b16</td>
+      <td>0.992353</td>
+      <td>1.686900e-04</td>
+      <td>0.592803</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
@@ -229,9 +259,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>chunks</td>
-      <td>3.329288</td>
-      <td>5.026000e-05</td>
-      <td>1.989654</td>
+      <td>3.688037</td>
+      <td>4.539000e-05</td>
+      <td>2.203128</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
@@ -239,9 +269,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>kahan</td>
-      <td>2.520789</td>
-      <td>6.638000e-05</td>
-      <td>1.506478</td>
+      <td>2.799331</td>
+      <td>5.980000e-05</td>
+      <td>1.672241</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
@@ -250,48 +280,58 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>1.730490e-03</td>
-      <td>0.577871</td>
+      <td>1.747730e-03</td>
+      <td>0.572171</td>
+      <td>10000</td>
+      <td>1000000</td>
+      <td>1000000</td>
+      <td>1000000</td>
+    </tr>
+    <tr>
+      <td>blas</td>
+      <td>8.625228</td>
+      <td>2.026300e-04</td>
+      <td>4.935103</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
       <td>1000000</td>
     </tr>
     <tr>
-      <td>m1</td>
-      <td>1.000844</td>
-      <td>1.729030e-03</td>
-      <td>0.578359</td>
+      <td>m1_b16</td>
+      <td>0.987262</td>
+      <td>1.770280e-03</td>
+      <td>0.564882</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
       <td>1000000</td>
     </tr>
     <tr>
-      <td>m2</td>
-      <td>9.674027</td>
-      <td>1.788800e-04</td>
-      <td>5.590340</td>
+      <td>blas_b16</td>
+      <td>5.937390</td>
+      <td>2.943600e-04</td>
+      <td>3.397201</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
       <td>1000000</td>
     </tr>
     <tr>
-      <td>m3</td>
-      <td>0.999059</td>
-      <td>1.732120e-03</td>
-      <td>0.577327</td>
+      <td>m3_b16</td>
+      <td>0.981904</td>
+      <td>1.779940e-03</td>
+      <td>0.561817</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
       <td>1000000</td>
     </tr>
     <tr>
-      <td>m4</td>
-      <td>0.996160</td>
-      <td>1.737160e-03</td>
-      <td>0.575652</td>
+      <td>m4_b16</td>
+      <td>0.986593</td>
+      <td>1.771480e-03</td>
+      <td>0.564500</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
@@ -299,9 +339,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>chunks</td>
-      <td>2.802685</td>
-      <td>6.174400e-04</td>
-      <td>1.619591</td>
+      <td>2.796815</td>
+      <td>6.249000e-04</td>
+      <td>1.600256</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
@@ -309,9 +349,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>kahan</td>
-      <td>2.338216</td>
-      <td>7.400900e-04</td>
-      <td>1.351187</td>
+      <td>2.260561</td>
+      <td>7.731400e-04</td>
+      <td>1.293427</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
diff --git a/benchmarks/dot/results/dot_ifort_co.html b/benchmarks/dot/results/dot_ifort_co.html
index 5335b37..b3d6fe2 100644
--- a/benchmarks/dot/results/dot_ifort_co.html
+++ b/benchmarks/dot/results/dot_ifort_co.html
@@ -42,10 +42,10 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>0.000003</td>
       <td>0.000002</td>
       <td>0.000002</td>
-      <td>8.906157</td>
+      <td>0.000002</td>
+      <td>9.349845</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
@@ -53,11 +53,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m1_co</td>
-      <td>0.000324</td>
-      <td>0.008962</td>
-      <td>0.008962</td>
-      <td>0.008962</td>
-      <td>0.001785</td>
+      <td>0.000329</td>
+      <td>0.005775</td>
+      <td>0.005775</td>
+      <td>0.005775</td>
+      <td>0.002770</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
@@ -65,11 +65,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m2_co</td>
-      <td>0.000346</td>
-      <td>0.008380</td>
-      <td>0.008380</td>
-      <td>0.008380</td>
-      <td>0.001909</td>
+      <td>0.000273</td>
+      <td>0.006964</td>
+      <td>0.006964</td>
+      <td>0.006964</td>
+      <td>0.002297</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
@@ -77,11 +77,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m3_co</td>
-      <td>0.000327</td>
-      <td>0.008864</td>
-      <td>0.008864</td>
-      <td>0.008864</td>
-      <td>0.001805</td>
+      <td>0.000223</td>
+      <td>0.008538</td>
+      <td>0.008538</td>
+      <td>0.008538</td>
+      <td>0.001874</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
@@ -90,10 +90,10 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>0.000030</td>
+      <td>0.000035</td>
       <td>0.000017</td>
-      <td>0.000018</td>
-      <td>8.917131</td>
+      <td>0.000020</td>
+      <td>8.565535</td>
       <td>1000</td>
       <td>10000</td>
       <td>10000</td>
@@ -101,11 +101,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m1_co</td>
-      <td>0.003719</td>
-      <td>0.008121</td>
-      <td>0.008120</td>
-      <td>0.008121</td>
-      <td>0.019703</td>
+      <td>0.005151</td>
+      <td>0.006833</td>
+      <td>0.006833</td>
+      <td>0.006833</td>
+      <td>0.023416</td>
       <td>1000</td>
       <td>10000</td>
       <td>10000</td>
@@ -113,11 +113,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m2_co</td>
-      <td>0.003765</td>
-      <td>0.008021</td>
-      <td>0.008021</td>
-      <td>0.008021</td>
-      <td>0.019947</td>
+      <td>0.005102</td>
+      <td>0.006899</td>
+      <td>0.006895</td>
+      <td>0.006897</td>
+      <td>0.023197</td>
       <td>1000</td>
       <td>10000</td>
       <td>10000</td>
@@ -125,11 +125,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m3_co</td>
-      <td>0.003371</td>
-      <td>0.008959</td>
-      <td>0.008959</td>
-      <td>0.008959</td>
-      <td>0.017860</td>
+      <td>0.005644</td>
+      <td>0.006236</td>
+      <td>0.006236</td>
+      <td>0.006236</td>
+      <td>0.025656</td>
       <td>1000</td>
       <td>10000</td>
       <td>10000</td>
@@ -138,10 +138,10 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>0.000467</td>
-      <td>0.000225</td>
-      <td>0.000359</td>
-      <td>4.655775</td>
+      <td>0.000447</td>
+      <td>0.000222</td>
+      <td>0.000335</td>
+      <td>4.950822</td>
       <td>1000</td>
       <td>100000</td>
       <td>100000</td>
@@ -149,11 +149,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m1_co</td>
-      <td>0.056323</td>
-      <td>0.008292</td>
-      <td>0.008291</td>
-      <td>0.008291</td>
-      <td>0.192969</td>
+      <td>0.071489</td>
+      <td>0.006248</td>
+      <td>0.006248</td>
+      <td>0.006248</td>
+      <td>0.256062</td>
       <td>1000</td>
       <td>100000</td>
       <td>100000</td>
@@ -161,11 +161,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m2_co</td>
-      <td>0.050779</td>
-      <td>0.009197</td>
-      <td>0.009197</td>
-      <td>0.009197</td>
-      <td>0.173975</td>
+      <td>0.044525</td>
+      <td>0.010033</td>
+      <td>0.010032</td>
+      <td>0.010033</td>
+      <td>0.159481</td>
       <td>1000</td>
       <td>100000</td>
       <td>100000</td>
@@ -173,11 +173,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m3_co</td>
-      <td>0.055709</td>
-      <td>0.008383</td>
-      <td>0.008383</td>
-      <td>0.008383</td>
-      <td>0.190867</td>
+      <td>0.066330</td>
+      <td>0.006735</td>
+      <td>0.006734</td>
+      <td>0.006734</td>
+      <td>0.237583</td>
       <td>1000</td>
       <td>100000</td>
       <td>100000</td>
@@ -186,10 +186,10 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>0.009344</td>
-      <td>0.009065</td>
-      <td>0.009204</td>
-      <td>1.738573</td>
+      <td>0.008509</td>
+      <td>0.008337</td>
+      <td>0.008413</td>
+      <td>1.901813</td>
       <td>1000</td>
       <td>1000000</td>
       <td>1000000</td>
@@ -197,11 +197,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m1_co</td>
-      <td>0.678207</td>
-      <td>0.013777</td>
-      <td>0.013777</td>
-      <td>0.013777</td>
-      <td>1.161314</td>
+      <td>0.827831</td>
+      <td>0.010278</td>
+      <td>0.010278</td>
+      <td>0.010278</td>
+      <td>1.556690</td>
       <td>1000</td>
       <td>1000000</td>
       <td>1000000</td>
@@ -209,11 +209,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m2_co</td>
-      <td>0.668718</td>
-      <td>0.013973</td>
-      <td>0.013973</td>
-      <td>0.013973</td>
-      <td>1.145068</td>
+      <td>0.543472</td>
+      <td>0.015656</td>
+      <td>0.015656</td>
+      <td>0.015656</td>
+      <td>1.021959</td>
       <td>1000</td>
       <td>1000000</td>
       <td>1000000</td>
@@ -221,11 +221,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m3_co</td>
-      <td>0.706626</td>
-      <td>0.013223</td>
-      <td>0.013223</td>
-      <td>0.013223</td>
-      <td>1.209977</td>
+      <td>0.607708</td>
+      <td>0.014001</td>
+      <td>0.014001</td>
+      <td>0.014001</td>
+      <td>1.142752</td>
       <td>1000</td>
       <td>1000000</td>
       <td>1000000</td>
diff --git a/benchmarks/dot/results/dot_ifort_co_perf_tot.png b/benchmarks/dot/results/dot_ifort_co_perf_tot.png
index 10aec84..55329fe 100644
Binary files a/benchmarks/dot/results/dot_ifort_co_perf_tot.png and b/benchmarks/dot/results/dot_ifort_co_perf_tot.png differ
diff --git a/benchmarks/dot/results/dot_ifort_co_speedup.png b/benchmarks/dot/results/dot_ifort_co_speedup.png
index 0b721ee..edea9b8 100644
Binary files a/benchmarks/dot/results/dot_ifort_co_speedup.png and b/benchmarks/dot/results/dot_ifort_co_speedup.png differ
diff --git a/benchmarks/dot/results/dot_ifort_co_speedup_max_avg.png b/benchmarks/dot/results/dot_ifort_co_speedup_max_avg.png
index 5094cc0..f2457d2 100644
Binary files a/benchmarks/dot/results/dot_ifort_co_speedup_max_avg.png and b/benchmarks/dot/results/dot_ifort_co_speedup_max_avg.png differ
diff --git a/benchmarks/dot/results/dot_ifort_co_time_avg.png b/benchmarks/dot/results/dot_ifort_co_time_avg.png
index b15b6d3..355915a 100644
Binary files a/benchmarks/dot/results/dot_ifort_co_time_avg.png and b/benchmarks/dot/results/dot_ifort_co_time_avg.png differ
diff --git a/benchmarks/dot/results/dot_ifort_co_time_max.png b/benchmarks/dot/results/dot_ifort_co_time_max.png
index 19ff1bf..f4830f8 100644
Binary files a/benchmarks/dot/results/dot_ifort_co_time_max.png and b/benchmarks/dot/results/dot_ifort_co_time_max.png differ
diff --git a/benchmarks/dot/results/dot_ifort_perf.png b/benchmarks/dot/results/dot_ifort_perf.png
index 4e512ce..eb0cf7e 100644
Binary files a/benchmarks/dot/results/dot_ifort_perf.png and b/benchmarks/dot/results/dot_ifort_perf.png differ
diff --git a/benchmarks/dot/results/dot_ifort_speedup.png b/benchmarks/dot/results/dot_ifort_speedup.png
index 37a7400..9075eaa 100644
Binary files a/benchmarks/dot/results/dot_ifort_speedup.png and b/benchmarks/dot/results/dot_ifort_speedup.png differ
diff --git a/benchmarks/dot/results/dot_ifort_speedup_avg.png b/benchmarks/dot/results/dot_ifort_speedup_avg.png
index 2041e4a..be1d17d 100644
Binary files a/benchmarks/dot/results/dot_ifort_speedup_avg.png and b/benchmarks/dot/results/dot_ifort_speedup_avg.png differ
diff --git a/benchmarks/dot/results/dot_ifort_time.png b/benchmarks/dot/results/dot_ifort_time.png
index 9f2442c..63b3cc9 100644
Binary files a/benchmarks/dot/results/dot_ifort_time.png and b/benchmarks/dot/results/dot_ifort_time.png differ
diff --git a/benchmarks/dot/results/dot_ifx.html b/benchmarks/dot/results/dot_ifx.html
index 674fd41..5e0d5f9 100644
--- a/benchmarks/dot/results/dot_ifx.html
+++ b/benchmarks/dot/results/dot_ifx.html
@@ -40,48 +40,58 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>1.680000e-06</td>
-      <td>0.595238</td>
+      <td>1.690000e-06</td>
+      <td>0.591716</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
     </tr>
     <tr>
-      <td>m1</td>
-      <td>0.641221</td>
-      <td>2.620000e-06</td>
-      <td>0.381679</td>
+      <td>blas</td>
+      <td>6.500000</td>
+      <td>2.600000e-07</td>
+      <td>3.846154</td>
+      <td>10000</td>
+      <td>1000</td>
+      <td>1000</td>
+      <td>1000</td>
+    </tr>
+    <tr>
+      <td>m1_b16</td>
+      <td>0.051368</td>
+      <td>3.290000e-05</td>
+      <td>0.030395</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
     </tr>
     <tr>
-      <td>m2</td>
-      <td>6.222222</td>
-      <td>2.700000e-07</td>
-      <td>3.703704</td>
+      <td>blas_b16</td>
+      <td>0.200000</td>
+      <td>8.450000e-06</td>
+      <td>0.118343</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
     </tr>
     <tr>
-      <td>m3</td>
-      <td>0.501493</td>
-      <td>3.350000e-06</td>
-      <td>0.298507</td>
+      <td>m3_b16</td>
+      <td>0.207872</td>
+      <td>8.130000e-06</td>
+      <td>0.123001</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
     </tr>
     <tr>
-      <td>m4</td>
-      <td>0.807692</td>
-      <td>2.080000e-06</td>
-      <td>0.480769</td>
+      <td>m4_b16</td>
+      <td>0.113499</td>
+      <td>1.489000e-05</td>
+      <td>0.067159</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
@@ -89,9 +99,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>chunks</td>
-      <td>1.826087</td>
-      <td>9.200000e-07</td>
-      <td>1.086957</td>
+      <td>1.362903</td>
+      <td>1.240000e-06</td>
+      <td>0.806452</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
@@ -99,9 +109,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>kahan</td>
-      <td>1.166667</td>
-      <td>1.440000e-06</td>
-      <td>0.694444</td>
+      <td>0.734783</td>
+      <td>2.300000e-06</td>
+      <td>0.434783</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
@@ -110,48 +120,58 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>1.729000e-05</td>
-      <td>0.578369</td>
+      <td>1.717000e-05</td>
+      <td>0.582411</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
     </tr>
     <tr>
-      <td>m1</td>
-      <td>1.029780</td>
-      <td>1.679000e-05</td>
-      <td>0.595593</td>
+      <td>blas</td>
+      <td>6.553435</td>
+      <td>2.620000e-06</td>
+      <td>3.816794</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
     </tr>
     <tr>
-      <td>m2</td>
-      <td>5.523962</td>
-      <td>3.130000e-06</td>
-      <td>3.194888</td>
+      <td>m1_b16</td>
+      <td>1.558076</td>
+      <td>1.102000e-05</td>
+      <td>0.907441</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
     </tr>
     <tr>
-      <td>m3</td>
-      <td>1.024289</td>
-      <td>1.688000e-05</td>
-      <td>0.592417</td>
+      <td>blas_b16</td>
+      <td>1.980392</td>
+      <td>8.670000e-06</td>
+      <td>1.153403</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
     </tr>
     <tr>
-      <td>m4</td>
-      <td>1.029167</td>
-      <td>1.680000e-05</td>
-      <td>0.595238</td>
+      <td>m3_b16</td>
+      <td>1.385795</td>
+      <td>1.239000e-05</td>
+      <td>0.807103</td>
+      <td>10000</td>
+      <td>10000</td>
+      <td>10000</td>
+      <td>10000</td>
+    </tr>
+    <tr>
+      <td>m4_b16</td>
+      <td>1.630579</td>
+      <td>1.053000e-05</td>
+      <td>0.949668</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
@@ -159,9 +179,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>chunks</td>
-      <td>1.996536</td>
-      <td>8.660000e-06</td>
-      <td>1.154734</td>
+      <td>1.430833</td>
+      <td>1.200000e-05</td>
+      <td>0.833333</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
@@ -169,9 +189,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>kahan</td>
-      <td>1.262044</td>
-      <td>1.370000e-05</td>
-      <td>0.729927</td>
+      <td>1.036836</td>
+      <td>1.656000e-05</td>
+      <td>0.603865</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
@@ -180,48 +200,58 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>1.672400e-04</td>
-      <td>0.597943</td>
+      <td>1.673900e-04</td>
+      <td>0.597407</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
       <td>100000</td>
     </tr>
     <tr>
-      <td>m1</td>
-      <td>0.999522</td>
-      <td>1.673200e-04</td>
-      <td>0.597657</td>
+      <td>blas</td>
+      <td>15.121048</td>
+      <td>1.107000e-05</td>
+      <td>9.033424</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
       <td>100000</td>
     </tr>
     <tr>
-      <td>m2</td>
-      <td>14.892253</td>
-      <td>1.123000e-05</td>
-      <td>8.904720</td>
+      <td>m1_b16</td>
+      <td>5.320725</td>
+      <td>3.146000e-05</td>
+      <td>3.178640</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
       <td>100000</td>
     </tr>
     <tr>
-      <td>m3</td>
-      <td>0.998031</td>
-      <td>1.675700e-04</td>
-      <td>0.596766</td>
+      <td>blas_b16</td>
+      <td>13.016330</td>
+      <td>1.286000e-05</td>
+      <td>7.776050</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
       <td>100000</td>
     </tr>
     <tr>
-      <td>m4</td>
-      <td>0.999104</td>
-      <td>1.673900e-04</td>
-      <td>0.597407</td>
+      <td>m3_b16</td>
+      <td>8.908462</td>
+      <td>1.879000e-05</td>
+      <td>5.321980</td>
+      <td>10000</td>
+      <td>100000</td>
+      <td>100000</td>
+      <td>100000</td>
+    </tr>
+    <tr>
+      <td>m4_b16</td>
+      <td>8.814639</td>
+      <td>1.899000e-05</td>
+      <td>5.265929</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
@@ -229,9 +259,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>chunks</td>
-      <td>1.830560</td>
-      <td>9.136000e-05</td>
-      <td>1.094571</td>
+      <td>1.756638</td>
+      <td>9.529000e-05</td>
+      <td>1.049428</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
@@ -239,9 +269,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>kahan</td>
-      <td>1.217088</td>
-      <td>1.374100e-04</td>
-      <td>0.727749</td>
+      <td>1.213499</td>
+      <td>1.379400e-04</td>
+      <td>0.724953</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
@@ -250,48 +280,58 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>1.733590e-03</td>
-      <td>0.576838</td>
+      <td>1.729100e-03</td>
+      <td>0.578336</td>
+      <td>10000</td>
+      <td>1000000</td>
+      <td>1000000</td>
+      <td>1000000</td>
+    </tr>
+    <tr>
+      <td>blas</td>
+      <td>8.900957</td>
+      <td>1.942600e-04</td>
+      <td>5.147740</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
       <td>1000000</td>
     </tr>
     <tr>
-      <td>m1</td>
-      <td>1.001826</td>
-      <td>1.730430e-03</td>
-      <td>0.577891</td>
+      <td>m1_b16</td>
+      <td>11.585260</td>
+      <td>1.492500e-04</td>
+      <td>6.700168</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
       <td>1000000</td>
     </tr>
     <tr>
-      <td>m2</td>
-      <td>9.999366</td>
-      <td>1.733700e-04</td>
-      <td>5.768011</td>
+      <td>blas_b16</td>
+      <td>3.498998</td>
+      <td>4.941700e-04</td>
+      <td>2.023595</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
       <td>1000000</td>
     </tr>
     <tr>
-      <td>m3</td>
-      <td>1.000543</td>
-      <td>1.732650e-03</td>
-      <td>0.577151</td>
+      <td>m3_b16</td>
+      <td>11.294663</td>
+      <td>1.530900e-04</td>
+      <td>6.532105</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
       <td>1000000</td>
     </tr>
     <tr>
-      <td>m4</td>
-      <td>0.999130</td>
-      <td>1.735100e-03</td>
-      <td>0.576336</td>
+      <td>m4_b16</td>
+      <td>11.542724</td>
+      <td>1.498000e-04</td>
+      <td>6.675567</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
@@ -299,9 +339,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>chunks</td>
-      <td>1.751349</td>
-      <td>9.898600e-04</td>
-      <td>1.010244</td>
+      <td>1.716876</td>
+      <td>1.007120e-03</td>
+      <td>0.992930</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
@@ -309,9 +349,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>kahan</td>
-      <td>1.167409</td>
-      <td>1.484990e-03</td>
-      <td>0.673405</td>
+      <td>1.165672</td>
+      <td>1.483350e-03</td>
+      <td>0.674150</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
diff --git a/benchmarks/dot/results/dot_ifx_co.html b/benchmarks/dot/results/dot_ifx_co.html
index c069e00..c970e90 100644
--- a/benchmarks/dot/results/dot_ifx_co.html
+++ b/benchmarks/dot/results/dot_ifx_co.html
@@ -45,7 +45,7 @@ <h2>dot_product Benchmark</h2>
       <td>0.000002</td>
       <td>0.000002</td>
       <td>0.000002</td>
-      <td>9.180036</td>
+      <td>9.284486</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
@@ -53,11 +53,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m1_co</td>
-      <td>0.000241</td>
-      <td>0.009144</td>
-      <td>0.009144</td>
-      <td>0.009144</td>
-      <td>0.001750</td>
+      <td>0.000325</td>
+      <td>0.005838</td>
+      <td>0.005838</td>
+      <td>0.005838</td>
+      <td>0.002741</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
@@ -65,11 +65,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m2_co</td>
-      <td>0.000259</td>
-      <td>0.008494</td>
-      <td>0.008493</td>
-      <td>0.008494</td>
-      <td>0.001884</td>
+      <td>0.000297</td>
+      <td>0.006401</td>
+      <td>0.006401</td>
+      <td>0.006401</td>
+      <td>0.002500</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
@@ -77,11 +77,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m3_co</td>
-      <td>0.000233</td>
-      <td>0.009444</td>
-      <td>0.009444</td>
-      <td>0.009444</td>
-      <td>0.001694</td>
+      <td>0.000273</td>
+      <td>0.006963</td>
+      <td>0.006963</td>
+      <td>0.006963</td>
+      <td>0.002298</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
@@ -90,10 +90,10 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>0.000029</td>
+      <td>0.000033</td>
       <td>0.000017</td>
       <td>0.000018</td>
-      <td>8.917157</td>
+      <td>8.891551</td>
       <td>1000</td>
       <td>10000</td>
       <td>10000</td>
@@ -101,11 +101,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m1_co</td>
-      <td>0.003570</td>
-      <td>0.008207</td>
-      <td>0.008207</td>
-      <td>0.008207</td>
-      <td>0.019495</td>
+      <td>0.004790</td>
+      <td>0.006911</td>
+      <td>0.006911</td>
+      <td>0.006911</td>
+      <td>0.023153</td>
       <td>1000</td>
       <td>10000</td>
       <td>10000</td>
@@ -113,11 +113,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m2_co</td>
-      <td>0.003295</td>
-      <td>0.008891</td>
-      <td>0.008891</td>
-      <td>0.008891</td>
-      <td>0.017996</td>
+      <td>0.003659</td>
+      <td>0.009046</td>
+      <td>0.009046</td>
+      <td>0.009046</td>
+      <td>0.017688</td>
       <td>1000</td>
       <td>10000</td>
       <td>10000</td>
@@ -125,11 +125,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m3_co</td>
-      <td>0.003533</td>
-      <td>0.008292</td>
-      <td>0.008292</td>
-      <td>0.008292</td>
-      <td>0.019295</td>
+      <td>0.004636</td>
+      <td>0.007140</td>
+      <td>0.007140</td>
+      <td>0.007140</td>
+      <td>0.022410</td>
       <td>1000</td>
       <td>10000</td>
       <td>10000</td>
@@ -138,10 +138,10 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>0.000448</td>
-      <td>0.000226</td>
-      <td>0.000368</td>
-      <td>4.584413</td>
+      <td>0.000437</td>
+      <td>0.000250</td>
+      <td>0.000361</td>
+      <td>4.530980</td>
       <td>1000</td>
       <td>100000</td>
       <td>100000</td>
@@ -149,11 +149,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m1_co</td>
-      <td>0.054181</td>
-      <td>0.008270</td>
-      <td>0.008270</td>
-      <td>0.008270</td>
-      <td>0.193461</td>
+      <td>0.070626</td>
+      <td>0.006183</td>
+      <td>0.006183</td>
+      <td>0.006183</td>
+      <td>0.258762</td>
       <td>1000</td>
       <td>100000</td>
       <td>100000</td>
@@ -161,11 +161,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m2_co</td>
-      <td>0.047967</td>
-      <td>0.009342</td>
-      <td>0.009342</td>
-      <td>0.009342</td>
-      <td>0.171273</td>
+      <td>0.070399</td>
+      <td>0.006203</td>
+      <td>0.006203</td>
+      <td>0.006203</td>
+      <td>0.257932</td>
       <td>1000</td>
       <td>100000</td>
       <td>100000</td>
@@ -173,11 +173,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m3_co</td>
-      <td>0.052354</td>
-      <td>0.008559</td>
-      <td>0.008559</td>
-      <td>0.008559</td>
-      <td>0.186936</td>
+      <td>0.062081</td>
+      <td>0.007034</td>
+      <td>0.007034</td>
+      <td>0.007034</td>
+      <td>0.227456</td>
       <td>1000</td>
       <td>100000</td>
       <td>100000</td>
@@ -186,10 +186,10 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>0.009598</td>
-      <td>0.009041</td>
-      <td>0.009318</td>
-      <td>1.717448</td>
+      <td>0.008663</td>
+      <td>0.008174</td>
+      <td>0.008402</td>
+      <td>1.904804</td>
       <td>1000</td>
       <td>1000000</td>
       <td>1000000</td>
@@ -197,11 +197,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m1_co</td>
-      <td>0.701147</td>
-      <td>0.013689</td>
-      <td>0.013689</td>
-      <td>0.013689</td>
-      <td>1.168822</td>
+      <td>0.711372</td>
+      <td>0.012178</td>
+      <td>0.012178</td>
+      <td>0.012178</td>
+      <td>1.313888</td>
       <td>1000</td>
       <td>1000000</td>
       <td>1000000</td>
@@ -209,11 +209,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m2_co</td>
-      <td>0.651640</td>
-      <td>0.014729</td>
-      <td>0.014729</td>
-      <td>0.014729</td>
-      <td>1.086292</td>
+      <td>0.796191</td>
+      <td>0.010880</td>
+      <td>0.010876</td>
+      <td>0.010878</td>
+      <td>1.470868</td>
       <td>1000</td>
       <td>1000000</td>
       <td>1000000</td>
@@ -221,11 +221,11 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>m3_co</td>
-      <td>0.650201</td>
-      <td>0.014762</td>
-      <td>0.014761</td>
-      <td>0.014762</td>
-      <td>1.083896</td>
+      <td>0.577058</td>
+      <td>0.015012</td>
+      <td>0.015012</td>
+      <td>0.015012</td>
+      <td>1.065814</td>
       <td>1000</td>
       <td>1000000</td>
       <td>1000000</td>
diff --git a/benchmarks/dot/results/dot_ifx_co_perf_tot.png b/benchmarks/dot/results/dot_ifx_co_perf_tot.png
index 5c2b995..dc5f724 100644
Binary files a/benchmarks/dot/results/dot_ifx_co_perf_tot.png and b/benchmarks/dot/results/dot_ifx_co_perf_tot.png differ
diff --git a/benchmarks/dot/results/dot_ifx_co_speedup.png b/benchmarks/dot/results/dot_ifx_co_speedup.png
index 07f23bf..a054ed6 100644
Binary files a/benchmarks/dot/results/dot_ifx_co_speedup.png and b/benchmarks/dot/results/dot_ifx_co_speedup.png differ
diff --git a/benchmarks/dot/results/dot_ifx_co_speedup_max_avg.png b/benchmarks/dot/results/dot_ifx_co_speedup_max_avg.png
index 7eb7a43..19fc7fc 100644
Binary files a/benchmarks/dot/results/dot_ifx_co_speedup_max_avg.png and b/benchmarks/dot/results/dot_ifx_co_speedup_max_avg.png differ
diff --git a/benchmarks/dot/results/dot_ifx_co_time_avg.png b/benchmarks/dot/results/dot_ifx_co_time_avg.png
index 798b906..815b7e4 100644
Binary files a/benchmarks/dot/results/dot_ifx_co_time_avg.png and b/benchmarks/dot/results/dot_ifx_co_time_avg.png differ
diff --git a/benchmarks/dot/results/dot_ifx_co_time_max.png b/benchmarks/dot/results/dot_ifx_co_time_max.png
index 48278f6..3a2bda8 100644
Binary files a/benchmarks/dot/results/dot_ifx_co_time_max.png and b/benchmarks/dot/results/dot_ifx_co_time_max.png differ
diff --git a/benchmarks/dot/results/dot_ifx_perf.png b/benchmarks/dot/results/dot_ifx_perf.png
index 52fadf6..392abfe 100644
Binary files a/benchmarks/dot/results/dot_ifx_perf.png and b/benchmarks/dot/results/dot_ifx_perf.png differ
diff --git a/benchmarks/dot/results/dot_ifx_speedup.png b/benchmarks/dot/results/dot_ifx_speedup.png
index 5c068e2..8c18cd8 100644
Binary files a/benchmarks/dot/results/dot_ifx_speedup.png and b/benchmarks/dot/results/dot_ifx_speedup.png differ
diff --git a/benchmarks/dot/results/dot_ifx_speedup_avg.png b/benchmarks/dot/results/dot_ifx_speedup_avg.png
index 2b8790e..5137fdd 100644
Binary files a/benchmarks/dot/results/dot_ifx_speedup_avg.png and b/benchmarks/dot/results/dot_ifx_speedup_avg.png differ
diff --git a/benchmarks/dot/results/dot_ifx_time.png b/benchmarks/dot/results/dot_ifx_time.png
index 9855f9e..db3982c 100644
Binary files a/benchmarks/dot/results/dot_ifx_time.png and b/benchmarks/dot/results/dot_ifx_time.png differ
diff --git a/benchmarks/dot/results/dot_nvfortran.html b/benchmarks/dot/results/dot_nvfortran.html
index b278183..ccd3690 100644
--- a/benchmarks/dot/results/dot_nvfortran.html
+++ b/benchmarks/dot/results/dot_nvfortran.html
@@ -40,48 +40,58 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>2.136000e-07</td>
-      <td>4.681648</td>
+      <td>2.133000e-07</td>
+      <td>4.688233</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
     </tr>
     <tr>
-      <td>m1</td>
-      <td>0.962596</td>
-      <td>2.219000e-07</td>
-      <td>4.506534</td>
+      <td>blas</td>
+      <td>0.854225</td>
+      <td>2.497000e-07</td>
+      <td>4.004806</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
     </tr>
     <tr>
-      <td>m2</td>
-      <td>0.844603</td>
-      <td>2.529000e-07</td>
-      <td>3.954132</td>
+      <td>m1_b16</td>
+      <td>0.161677</td>
+      <td>1.319300e-06</td>
+      <td>0.757978</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
     </tr>
     <tr>
-      <td>m3</td>
-      <td>0.900126</td>
-      <td>2.373000e-07</td>
-      <td>4.214075</td>
+      <td>blas_b16</td>
+      <td>0.120454</td>
+      <td>1.770800e-06</td>
+      <td>0.564717</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
       <td>1000</td>
     </tr>
     <tr>
-      <td>m4</td>
-      <td>0.881552</td>
-      <td>2.423000e-07</td>
-      <td>4.127115</td>
+      <td>m3_b16</td>
+      <td>0.131895</td>
+      <td>1.617200e-06</td>
+      <td>0.618353</td>
+      <td>10000</td>
+      <td>1000</td>
+      <td>1000</td>
+      <td>1000</td>
+    </tr>
+    <tr>
+      <td>m4_b16</td>
+      <td>0.125618</td>
+      <td>1.698000e-06</td>
+      <td>0.588928</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
@@ -89,9 +99,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>chunks</td>
-      <td>0.752643</td>
-      <td>2.838000e-07</td>
-      <td>3.523608</td>
+      <td>0.755847</td>
+      <td>2.822000e-07</td>
+      <td>3.543586</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
@@ -99,9 +109,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>kahan</td>
-      <td>0.384796</td>
-      <td>5.551000e-07</td>
-      <td>1.801477</td>
+      <td>0.390159</td>
+      <td>5.467000e-07</td>
+      <td>1.829157</td>
       <td>10000</td>
       <td>1000</td>
       <td>1000</td>
@@ -110,48 +120,58 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>2.449800e-06</td>
-      <td>4.081966</td>
+      <td>2.623500e-06</td>
+      <td>3.811702</td>
+      <td>10000</td>
+      <td>10000</td>
+      <td>10000</td>
+      <td>10000</td>
+    </tr>
+    <tr>
+      <td>blas</td>
+      <td>0.980198</td>
+      <td>2.676500e-06</td>
+      <td>3.736223</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
     </tr>
     <tr>
-      <td>m1</td>
-      <td>0.991380</td>
-      <td>2.471100e-06</td>
-      <td>4.046781</td>
+      <td>m1_b16</td>
+      <td>0.632779</td>
+      <td>4.146000e-06</td>
+      <td>2.411963</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
     </tr>
     <tr>
-      <td>m2</td>
-      <td>0.980783</td>
-      <td>2.497800e-06</td>
-      <td>4.003523</td>
+      <td>blas_b16</td>
+      <td>0.578220</td>
+      <td>4.537200e-06</td>
+      <td>2.204002</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
     </tr>
     <tr>
-      <td>m3</td>
-      <td>0.977613</td>
-      <td>2.505900e-06</td>
-      <td>3.990582</td>
+      <td>m3_b16</td>
+      <td>0.601072</td>
+      <td>4.364700e-06</td>
+      <td>2.291108</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
     </tr>
     <tr>
-      <td>m4</td>
-      <td>0.973727</td>
-      <td>2.515900e-06</td>
-      <td>3.974721</td>
+      <td>m4_b16</td>
+      <td>0.590958</td>
+      <td>4.439400e-06</td>
+      <td>2.252557</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
@@ -159,9 +179,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>chunks</td>
-      <td>0.557457</td>
-      <td>4.394600e-06</td>
-      <td>2.275520</td>
+      <td>0.574774</td>
+      <td>4.564400e-06</td>
+      <td>2.190868</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
@@ -169,9 +189,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>kahan</td>
-      <td>0.441971</td>
-      <td>5.542900e-06</td>
-      <td>1.804110</td>
+      <td>0.455952</td>
+      <td>5.753900e-06</td>
+      <td>1.737952</td>
       <td>10000</td>
       <td>10000</td>
       <td>10000</td>
@@ -180,48 +200,58 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>4.245560e-05</td>
-      <td>2.355402</td>
+      <td>4.319610e-05</td>
+      <td>2.315024</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
       <td>100000</td>
     </tr>
     <tr>
-      <td>m1</td>
-      <td>1.006646</td>
-      <td>4.217530e-05</td>
-      <td>2.371056</td>
+      <td>blas</td>
+      <td>1.378156</td>
+      <td>3.134340e-05</td>
+      <td>3.190464</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
       <td>100000</td>
     </tr>
     <tr>
-      <td>m2</td>
-      <td>3.816268</td>
-      <td>1.112490e-05</td>
-      <td>8.988845</td>
+      <td>m1_b16</td>
+      <td>0.982489</td>
+      <td>4.396600e-05</td>
+      <td>2.274485</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
       <td>100000</td>
     </tr>
     <tr>
-      <td>m3</td>
-      <td>1.005133</td>
-      <td>4.223880e-05</td>
-      <td>2.367492</td>
+      <td>blas_b16</td>
+      <td>0.972519</td>
+      <td>4.441670e-05</td>
+      <td>2.251405</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
       <td>100000</td>
     </tr>
     <tr>
-      <td>m4</td>
-      <td>1.000754</td>
-      <td>4.242360e-05</td>
-      <td>2.357179</td>
+      <td>m3_b16</td>
+      <td>0.978518</td>
+      <td>4.414440e-05</td>
+      <td>2.265293</td>
+      <td>10000</td>
+      <td>100000</td>
+      <td>100000</td>
+      <td>100000</td>
+    </tr>
+    <tr>
+      <td>m4_b16</td>
+      <td>0.975412</td>
+      <td>4.428500e-05</td>
+      <td>2.258101</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
@@ -229,9 +259,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>chunks</td>
-      <td>0.898362</td>
-      <td>4.725890e-05</td>
-      <td>2.116004</td>
+      <td>0.926702</td>
+      <td>4.661270e-05</td>
+      <td>2.145338</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
@@ -239,9 +269,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>kahan</td>
-      <td>0.657020</td>
-      <td>6.461840e-05</td>
-      <td>1.547547</td>
+      <td>0.661962</td>
+      <td>6.525470e-05</td>
+      <td>1.532457</td>
       <td>10000</td>
       <td>100000</td>
       <td>100000</td>
@@ -250,48 +280,58 @@ <h2>dot_product Benchmark</h2>
     <tr>
       <td>dot_product</td>
       <td>1.000000</td>
-      <td>5.856337e-04</td>
-      <td>1.707552</td>
+      <td>5.770094e-04</td>
+      <td>1.733074</td>
+      <td>10000</td>
+      <td>1000000</td>
+      <td>1000000</td>
+      <td>1000000</td>
+    </tr>
+    <tr>
+      <td>blas</td>
+      <td>1.568794</td>
+      <td>3.678045e-04</td>
+      <td>2.718836</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
       <td>1000000</td>
     </tr>
     <tr>
-      <td>m1</td>
-      <td>1.024743</td>
-      <td>5.714932e-04</td>
-      <td>1.749802</td>
+      <td>m1_b16</td>
+      <td>0.962880</td>
+      <td>5.992540e-04</td>
+      <td>1.668741</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
       <td>1000000</td>
     </tr>
     <tr>
-      <td>m2</td>
-      <td>3.353743</td>
-      <td>1.746209e-04</td>
-      <td>5.726691</td>
+      <td>blas_b16</td>
+      <td>1.843979</td>
+      <td>3.129154e-04</td>
+      <td>3.195752</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
       <td>1000000</td>
     </tr>
     <tr>
-      <td>m3</td>
-      <td>1.022575</td>
-      <td>5.727051e-04</td>
-      <td>1.746099</td>
+      <td>m3_b16</td>
+      <td>0.950563</td>
+      <td>6.070184e-04</td>
+      <td>1.647397</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
       <td>1000000</td>
     </tr>
     <tr>
-      <td>m4</td>
-      <td>1.019593</td>
-      <td>5.743798e-04</td>
-      <td>1.741008</td>
+      <td>m4_b16</td>
+      <td>0.920440</td>
+      <td>6.268845e-04</td>
+      <td>1.595190</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
@@ -299,9 +339,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>chunks</td>
-      <td>0.903414</td>
-      <td>6.482454e-04</td>
-      <td>1.542626</td>
+      <td>0.830726</td>
+      <td>6.945842e-04</td>
+      <td>1.439710</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
@@ -309,9 +349,9 @@ <h2>dot_product Benchmark</h2>
     </tr>
     <tr>
       <td>kahan</td>
-      <td>0.762443</td>
-      <td>7.681020e-04</td>
-      <td>1.301910</td>
+      <td>0.663607</td>
+      <td>8.695051e-04</td>
+      <td>1.150080</td>
       <td>10000</td>
       <td>1000000</td>
       <td>1000000</td>
diff --git a/benchmarks/dot/results/dot_nvfortran_perf.png b/benchmarks/dot/results/dot_nvfortran_perf.png
index 880945f..cccbf91 100644
Binary files a/benchmarks/dot/results/dot_nvfortran_perf.png and b/benchmarks/dot/results/dot_nvfortran_perf.png differ
diff --git a/benchmarks/dot/results/dot_nvfortran_speedup.png b/benchmarks/dot/results/dot_nvfortran_speedup.png
index 1f125f6..1579fa4 100644
Binary files a/benchmarks/dot/results/dot_nvfortran_speedup.png and b/benchmarks/dot/results/dot_nvfortran_speedup.png differ
diff --git a/benchmarks/dot/results/dot_nvfortran_speedup_avg.png b/benchmarks/dot/results/dot_nvfortran_speedup_avg.png
index fad0e90..a2d6403 100644
Binary files a/benchmarks/dot/results/dot_nvfortran_speedup_avg.png and b/benchmarks/dot/results/dot_nvfortran_speedup_avg.png differ
diff --git a/benchmarks/dot/results/dot_nvfortran_time.png b/benchmarks/dot/results/dot_nvfortran_time.png
index 055d4b7..61a38e1 100644
Binary files a/benchmarks/dot/results/dot_nvfortran_time.png and b/benchmarks/dot/results/dot_nvfortran_time.png differ
diff --git a/benchmarks/dot/results/export.py b/benchmarks/dot/results/export.py
index 0ec24e7..1c628d0 100644
--- a/benchmarks/dot/results/export.py
+++ b/benchmarks/dot/results/export.py
@@ -36,7 +36,7 @@
                                    title='dot_product Benchmark - Speedup',
                                    xlabel='Number of Elements',
                                    ylabel='Speedup [-]',
-                                   bar_width=0.12)
+                                   bar_width=0.10)
 
 
 fpd.plot_speedup_avg(file_path, benchmark_data, x_data, 
diff --git a/benchmarks/dot/test/test_dot.f90 b/benchmarks/dot/test/test_dot.f90
index 20a412e..9b7a61e 100644
--- a/benchmarks/dot/test/test_dot.f90
+++ b/benchmarks/dot/test/test_dot.f90
@@ -1,7 +1,7 @@
 program test_dot
 
    use kinds
-   use fordot
+   use fordot, only: fdot_product => dot_product
    use fast_math, only: fprod, fprod_kahan
    use forunittest
 
@@ -12,22 +12,24 @@ program test_dot
    integer               :: m
    type(unit_test)       :: ut
 
+   m = 300
+
    allocate(u(m),v(m))
    call random_number(u)
    call random_number(v)
 
    a_ref = dot_product(u,v)
 
-   a = dot_product(u,v, option='m1')
+   a = fdot_product(u,v, option='m1')
    call ut%check(a, a_ref, tol=1e-12_rk, msg="dot_product m1")
 
-   a = dot_product(u,v, option='m2')
+   a = fdot_product(u,v, option='m2')
    call ut%check(a, a_ref, tol=1e-12_rk, msg="dot_product m2")
 
-   a = dot_product(u,v, option='m3')
+   a = fdot_product(u,v, option='m3')
    call ut%check(a, a_ref, tol=1e-12_rk, msg="dot_product m3")
 
-   a = dot_product(u,v, option='m4')
+   a = fdot_product(u,v, option='m4')
    call ut%check(a, a_ref, tol=1e-12_rk, msg="dot_product m4")
 
    a = fprod(u,v)
@@ -36,4 +38,16 @@ program test_dot
    a = fprod_kahan(u,v)
    call ut%check(a, a_ref, tol=1e-12_rk, msg="fprod_kahan")
 
+   a = fdot_product(u,v, option='m1',nblock=16)
+   call ut%check(a, a_ref, tol=1e-12_rk, msg="dot_product m1b")
+
+   a = fdot_product(u,v, option='m2',nblock=16)
+   call ut%check(a, a_ref, tol=1e-12_rk, msg="dot_product m2b")
+
+   a = fdot_product(u,v, option='m3',nblock=16)
+   call ut%check(a, a_ref, tol=1e-12_rk, msg="dot_product m3b")
+
+   a = fdot_product(u,v, option='m4',nblock=16)
+   call ut%check(a, a_ref, tol=1e-12_rk, msg="dot_product m4b")
+
 end program test_dot
\ No newline at end of file
diff --git a/fpm.rsp b/fpm.rsp
index cd566ff..2c57110 100644
--- a/fpm.rsp
+++ b/fpm.rsp
@@ -3,7 +3,7 @@ options run --example
 options --target dot
 options --profile release
 options --compiler ifx
-options --flag "-O3 -mtune=native -xHost -lblas -qopenmp -flto=full -DINT64"
+options --flag "-O3 -mtune=native -xHost -lblas -qopenmp -flto=full -DINT64 -DUSE_DO_CONCURRENT"
 
 @benchmark-dot-ifort
 options run --example
@@ -31,7 +31,7 @@ options run --example
 options --target dot_co
 options --profile release
 options --compiler ifx
-options --flag "-O3 -mtune=native -xHost -lblas -qopenmp -DINT64 -flto=full -coarray -DUSE_COARRAY"
+options --flag "-O3 -mtune=native -xHost -lblas -qopenmp -DINT64 -flto=full -coarray -DUSE_COARRAY -DUSE_DO_CONCURRENT"
 
 @benchmark-dot-ifort-coarray
 options run --example
@@ -48,7 +48,7 @@ options test
 options --target test_dot
 options --profile release
 options --compiler ifx
-options --flag "-O3 -mtune=native -xHost -lblas -qopenmp -flto=full -DINT64"
+options --flag "-O3 -mtune=native -xHost -lblas -qopenmp -flto=full -DINT64 -DUSE_DO_CONCURRENT"
 
 @benchmark-dot-test-ifort
 options test