diff --git a/src/BC1/BC1Encoder.cpp b/src/BC1/BC1Encoder.cpp
index f8f09a6..52d854a 100644
--- a/src/BC1/BC1Encoder.cpp
+++ b/src/BC1/BC1Encoder.cpp
@@ -19,23 +19,22 @@
 
 #include "BC1Encoder.h"
 
+#include <algorithm>
 #include <array>
 #include <cassert>
-#include <climits>
 #include <cmath>
-#include <cstddef>
 #include <cstdint>
 #include <memory>
 #include <type_traits>
 
 #include "../BlockView.h"
 #include "../Color.h"
-#include "../Interpolator.h"
 #include "../Matrix4x4.h"
 #include "../Vector4.h"
 #include "../Vector4Int.h"
 #include "../bitwiseEnums.h"
 #include "../util.h"
+#include "Histogram.h"
 #include "OrderTable.h"
 #include "SingleColorTable.h"
 
@@ -288,6 +287,7 @@ void BC1Encoder::FindEndpoints(Color4x4 pixels, EncodeResults &block, const Bloc
             uint8_t hr6 = scale8To6(metrics.max.r);
 
             block.low = Color(lr5, lr6, lr5);
+            block.high = Color(hr5, hr6, hr5);
         }
     } else if (endpoint_mode == EndpointMode::LeastSquares) {
         //  2D Least Squares approach from Humus's example, with added inset and optimal rounding.
@@ -297,7 +297,6 @@ void BC1Encoder::FindEndpoints(Color4x4 pixels, EncodeResults &block, const Bloc
 
         auto &sums = metrics.sums;
         auto &min = metrics.min;
-        auto &max = metrics.max;
 
         unsigned chan0 = (unsigned)diff.MaxChannelRGB();  // primary axis of the bounding box
         l[chan0] = (float)min[chan0];
@@ -312,8 +311,8 @@ void BC1Encoder::FindEndpoints(Color4x4 pixels, EncodeResults &block, const Bloc
             for (unsigned c = 0; c < 3; c++) { sums_xy[c] += val[chan0] * val[c]; }
         }
 
-        const auto &sum_x = sums[chan0];
-        const auto &sum_xx = sums_xy[chan0];
+        const unsigned sum_x = (unsigned)sums[chan0];
+        const unsigned sum_xx = sums_xy[chan0];
 
         float denominator = (float)(16 * sum_xx) - (float)(sum_x * sum_x);
 
@@ -326,9 +325,9 @@ void BC1Encoder::FindEndpoints(Color4x4 pixels, EncodeResults &block, const Bloc
                  *  a = (m∑xy - ∑x∑y) / m∑x² - (∑x)²
                  *  b = (∑x²∑y - ∑xy∑x) / m∑x² - (∑x)²
                  * see Giordano/Weir pg.103 */
-                const auto chan = (chan0 + i) % 3;
-                const auto &sum_y = sums[chan];
-                const auto &sum_xy = sums_xy[chan];
+                const unsigned chan = (chan0 + i) % 3;
+                const unsigned sum_y = (unsigned)sums[chan];
+                const unsigned sum_xy = sums_xy[chan];
 
                 float a = (float)((16 * sum_xy) - (sum_x * sum_y)) / denominator;
                 float b = (float)((sum_xx * sum_y) - (sum_xy * sum_x)) / denominator;
@@ -709,9 +708,8 @@ void BC1Encoder::RefineBlockCF(Color4x4 &pixels, EncodeResults &block, BlockMetr
         sums[i + 1] = sums[i] + color_vectors[p];
     }
 
-    const Hash q_total = ((_flags & Flags::Exhaustive) != Flags::None) ? OrderTable::OrderCount : orderings;
+    const unsigned q_total = ((_flags & Flags::Exhaustive) != Flags::None) ? OrderTable::OrderCount : orderings;
     for (Hash q = 0; q < q_total; q++) {
-
         Hash trial_hash = ((_flags & Flags::Exhaustive) != Flags::None) ? q : OrderTable::BestOrders[start_hash][q];
         Vector4 trial_matrix = OrderTable::GetFactors(trial_hash);
 
diff --git a/src/BC1/BC1Encoder.h b/src/BC1/BC1Encoder.h
index 01d563e..40b5f70 100644
--- a/src/BC1/BC1Encoder.h
+++ b/src/BC1/BC1Encoder.h
@@ -20,21 +20,19 @@
 #pragma once
 
 #include <array>
-#include <cassert>
+#include <climits>
 #include <cstdint>
-#include <mutex>
+#include <memory>
 
 #include "../BlockEncoder.h"
 #include "../BlockView.h"
-#include "../Interpolator.h"
-#include "../bitwiseEnums.h"
-#include "../ndebug.h"
+#include "../Color.h"
 #include "BC1Block.h"
-#include "OrderTable.h"
 #include "SingleColorTable.h"
-#include "Tables.h"
 
 namespace rgbcx {
+class Interpolator;
+class Vector4;
 
 class BC1Encoder final : public BlockEncoder<BC1Block, 4, 4> {
    public:
@@ -163,6 +161,6 @@ class BC1Encoder final : public BlockEncoder<BC1Block, 4, 4> {
 
     template <ColorMode M> void RefineBlockLS(Color4x4 &pixels, EncodeResults &block, BlockMetrics &metrics, ErrorMode error_mode, unsigned passes) const;
 
-    template<ColorMode M> void RefineBlockCF(Color4x4 &pixels, EncodeResults &block, BlockMetrics &metrics, ErrorMode error_mode, unsigned orderings) const;
+    template <ColorMode M> void RefineBlockCF(Color4x4 &pixels, EncodeResults &block, BlockMetrics &metrics, ErrorMode error_mode, unsigned orderings) const;
 };
 }  // namespace rgbcx
diff --git a/src/BC1/Histogram.h b/src/BC1/Histogram.h
index 812ae70..94d586e 100644
--- a/src/BC1/Histogram.h
+++ b/src/BC1/Histogram.h
@@ -72,7 +72,7 @@ template <size_t N> class Histogram {
 
     unsigned GetPacked() const {
         unsigned packed = 0;
-        for (unsigned i = 0; i < (N - 1); i++) { packed |= (_bins[i] << (4 * i)); }
+        for (unsigned i = 0; i < (N - 1); i++) { packed |= (_bins[i] << (4U * i)); }
 
         //        assert(packed < HashCount);
 
diff --git a/src/BC1/OrderTable.cpp b/src/BC1/OrderTable.cpp
index 2924972..1388836 100644
--- a/src/BC1/OrderTable.cpp
+++ b/src/BC1/OrderTable.cpp
@@ -22,7 +22,6 @@
 #include <array>
 
 #include "../Vector4.h"
-#include "Histogram.h"
 
 namespace rgbcx {
 using Hash = uint16_t;
diff --git a/src/BC1/OrderTable.h b/src/BC1/OrderTable.h
index e4410ad..0a3a901 100644
--- a/src/BC1/OrderTable.h
+++ b/src/BC1/OrderTable.h
@@ -21,14 +21,12 @@
 
 #include <algorithm>
 #include <array>
+#include <atomic>
+#include <cassert>
+#include <cmath>
 #include <cstdint>
-#include <mutex>
-#include <numeric>
 
 #include "../Vector4.h"
-#include "../util.h"
-#include "Histogram.h"
-#include "Tables.h"
 
 namespace rgbcx {
 template <size_t N> class Histogram;
@@ -106,7 +104,7 @@ template <size_t N> class OrderTable {
     }
 
     static Vector4 GetFactors(Hash hash) {
-        assert(generated);
+        assert(generated.load());
         assert(factors != nullptr);
 
         return factors->at(hash);
diff --git a/src/BlockEncoder.h b/src/BlockEncoder.h
index 9568a3e..bf78fab 100644
--- a/src/BlockEncoder.h
+++ b/src/BlockEncoder.h
@@ -19,8 +19,8 @@
 
 #pragma once
 
-#include <cstdint>
 #include <climits>
+#include <cstdint>
 
 #include "BlockView.h"
 
@@ -59,11 +59,7 @@ template <class B, size_t M, size_t N> class BlockEncoder {
                 assert(pixel_x + N <= image_width);
 
                 unsigned top_left = pixel_x + (pixel_y * image_width);
-                auto src = DecodedBlock(&decoded[top_left], image_width);
-
-                if (pixel_x == 272 && pixel_y == 748) {
-                    int foo = 3;
-                }
+                auto src = DecodedBlock(&decoded[top_left], (int)image_width);
 
                 EncodeBlock(src, &blocks[x + block_width * y]);
             }
diff --git a/src/BlockView.h b/src/BlockView.h
index d064f89..feffc86 100644
--- a/src/BlockView.h
+++ b/src/BlockView.h
@@ -150,7 +150,7 @@ template <size_t M, size_t N> class ColorBlockView : public BlockView<Color, M,
             metrics.has_black |= (val.r | val.g | val.b < black_threshold);
         }
 
-        for (unsigned c = 0; c < 3; c++) { metrics.avg[c] = (uint8_t)(metrics.sums[c] / (M * N)); }
+        for (unsigned c = 0; c < 3; c++) { metrics.avg[c] = (uint8_t)((unsigned)metrics.sums[c] / (M * N)); }
 
         return metrics;
     }
diff --git a/src/Vector4Int.h b/src/Vector4Int.h
index dfbf783..186bebe 100644
--- a/src/Vector4Int.h
+++ b/src/Vector4Int.h
@@ -64,7 +64,7 @@ class Vector4Int {
         return _c[index];
     }
 
-    operator Vector4() const { return Vector4(_c[0], _c[1], _c[2], _c[3]); }
+    operator Vector4() const { return Vector4((float)_c[0], (float)_c[1], (float)_c[2], (float)_c[3]); }
 
     friend Vector4Int operator+(const Vector4Int &lhs, const Vector4Int &rhs) { return DoOp(lhs, rhs, std::plus()); }
     friend Vector4Int operator-(const Vector4Int &lhs, const Vector4Int &rhs) { return DoOp(lhs, rhs, std::minus()); }