diff --git a/runbench b/runbench
index 76651fc..7330f4c 100755
--- a/runbench
+++ b/runbench
@@ -20,8 +20,7 @@ bins = []
 {
   "base"         => "-O3", 
   "march-native" => "-O3 -march=native",
-# Fast math seems to return broken results and not be a benefit
-#  "ffast-math"   => "-O3 -march=native -ffast-math",
+  "ffast-math"   => "-O3 -march=native -ffast-math",
 }.each do |name, opts|
   bench = "#{BINDIR}/clangbench-#{name}"
   if !File.exist?(bench)
diff --git a/src/main.rs b/src/main.rs
index e9ee24d..1c765af 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -47,9 +47,9 @@ fn main() {
     pixout[2] = r * matrix[2][0] + g * matrix[2][1] + b * matrix[2][2] + e * matrix[2][3];
   }
   let to_time = time::precise_time_ns();
-  let mut sum = 0f32;
+  let mut sum = 0f64;
   for v in out {
-    sum += v;
+    sum += v as f64;
   }
   println!("{:.2} ms/megapixel (sum is {})",
          ((to_time - from_time) as f32)/((num_pixels as f32)),
@@ -70,9 +70,9 @@ fn main() {
     pixout[2] = z_comps.extract(0) + z_comps.extract(1) + z_comps.extract(2);
   }
   let to_time = time::precise_time_ns();
-  let mut sum = 0f32;
+  let mut sum = 0f64;
   for v in out {
-    sum += v;
+    sum += v as f64;
   }
   println!("{:.2} ms/megapixel (sum is {}) (explicit simd)",
          ((to_time - from_time) as f32)/((num_pixels as f32)),
diff --git a/test.c b/test.c
index d5789ba..ee4ab5b 100644
--- a/test.c
+++ b/test.c
@@ -56,7 +56,7 @@ int main(void) {
   int64_t to_time = time_in_micros();
 
   // Calculate the pixel average
-  float sum = 0.0f;
+  double sum = 0.0f;
   for(int i=0; i<NUM_PIXELS*3; i++) {
     sum += out[i];
   }