@@ -21,6 +21,7 @@
 #include <ireduce.hpp>
 #include <logic.hpp>
 #include <reduce.hpp>
+#include <scan.hpp>
 #include <sobel.hpp>
 #include <tile.hpp>
 #include <transpose.hpp>
@@ -76,19 +77,20 @@ Array<float> otsuThreshold(const Array<float>& supEdges,
     Array<uint> hist = histogram<float>(supEdges, NUM_BINS, 0, maxVal, false);
 
     const dim4& hDims = hist.dims();
+    auto fhist        = cast<float, uint>(hist);
 
-    // reduce along histogram dimension i.e. 0th dimension
-    auto totals = reduce<af_add_t, uint, float>(hist, 0);
-
-    // tile histogram total along 0th dimension
-    auto ttotals = tile(totals, dim4(hDims[0]));
+    auto totals = createValueArray<float>(1.f, supEdges.elements());
 
     // pixel frequency probabilities
-    auto probability =
-        arithOp<float, af_div_t>(cast<float, uint>(hist), ttotals, hDims);
+    auto probability = arithOp<float, af_div_t>(fhist, totals, hDims);
+    auto fscan       = detail::scan<af_add_t, float, float>(probability, 0);
+
+    vector<float> hfscan(fscan.elements());
+    detail::copyData<float>(hfscan.data(), fscan);
 
     vector<af_seq> seqBegin(4, af_span);
     vector<af_seq> seqRest(4, af_span);
+    vector<af_seq> sliceIndex(4, af_span);
 
     seqBegin[0] = af_make_seq(0, static_cast<double>(hDims[0] - 1), 1);
     seqRest[0]  = af_make_seq(0, static_cast<double>(hDims[0] - 1), 1);
@@ -104,8 +106,8 @@ Array<float> otsuThreshold(const Array<float>& supEdges,
         auto frontPartition = createSubArray(probability, seqBegin, false);
         auto endPartition   = createSubArray(probability, seqRest, false);
 
-        auto qL = reduce<af_add_t, float, float>(frontPartition, 0);
-        auto qH = reduce<af_add_t, float, float>(endPartition, 0);
+        auto qL = createValueArray<float>(1, hfscan[b]);
+        auto qH = createValueArray<float>(1, 1.f - hfscan[b]);
 
         const dim4 fdims(b + 1, hDims[1], hDims[2], hDims[3]);
         const dim4 edims(NUM_BINS - 1 - b, hDims[1], hDims[2], hDims[3]);
@@ -129,11 +131,8 @@ Array<float> otsuThreshold(const Array<float>& supEdges,
         auto op2   = arithOp<float, af_mul_t>(qL, qH, tdims);
         auto sigma = arithOp<float, af_mul_t>(sqrd, op2, tdims);
 
-        vector<af_seq> sliceIndex(4, af_span);
         sliceIndex[0] = {double(b), double(b), 1};
-
-        auto binRes = createSubArray<float>(sigmas, sliceIndex, false);
-
+        auto binRes   = createSubArray<float>(sigmas, sliceIndex, false);
         copyArray(binRes, sigma);
     }