Fix normalization when the result overflows (#701)

Using uint for max values now because uint Max32 can't be represented in IEEE754 single precision float. In calculations it will be promoted to float.

Changed normalization function for single channel images to return max of the data type assuming (x, 0, 0, 0) vectors. And changed the test values accordingly

---------

Co-authored-by: Wasim Abbas <abbas.wasim@gmail.com>

tests/unittests/image_unittests.cc

@@ -106,7 +106,7 @@ void test_by_channel( const uint32_t min_res[], const uint32_t max_res[],
 }
 
 // Hand tested vectors and their normalised results in GeoGebra
-static constexpr uint32_t min_res[3]={54, 13849, 907633408}; // These are -0.577350 float in 8bit-UNORM, 16bit-UNORM and 32bit-UNORM values
+static constexpr uint32_t min_res[3]={54, 13849, 907633408};     // These are -0.577350 float in 8bit-UNORM, 16bit-UNORM and 32bit-UNORM values
 static constexpr uint32_t max_res[3]={201, 51686, 3387333888};
 
 static constexpr uint32_t some_res8[3]={30, 192, 179};
@@ -120,10 +120,10 @@ static constexpr uint32_t some_res16_2[3]={9541, 9654, 32768};
 static constexpr uint32_t some_res32_2[3]={628983424, 628983424, 2147483648};
 
 static constexpr uint32_t min_res1[3]={0, 0, 0};
-static constexpr uint32_t max_res1[3]={255, 65535, 0}; // Last is not 4294967295 because of overflow
-static constexpr uint32_t some_res8_1[3]={0, 198, 128};
-static constexpr uint32_t some_res16_1[3]={0, 32768, 32768};
-static constexpr uint32_t some_res32_1[3]={0, 2147483648, 2147483648};
+static constexpr uint32_t max_res1[3]={255, 65535, 4294967295};
+static constexpr uint32_t some_res8_1[3]={255, 0, 0};            // Second and third values are unused because its single channel image result
+static constexpr uint32_t some_res16_1[3]={65535, 0, 0};         // Same as above
+static constexpr uint32_t some_res32_1[3]={4294967295, 0, 0};    // Same as above
 
 TEST(NormaliseColorTest, multi_channel_test) {
     test_by_channel<4>(min_res, max_res, some_res8, some_res16, some_res32);

tools/toktx/image.hpp

@@ -185,7 +185,7 @@ struct vec3_base {
 };
 
 static constexpr float gc_m[5]={0.0f, 128.0f, 32768.0f, 0.0f, 2147483648.0f};
-static constexpr float gc_s[5]={0.0f, 255.0f, 65535.0f, 0.0f, 4294967295.0f};
+static constexpr uint32_t gc_s[5]={0, 255, 65535, 0, 4294967295};
 
 template <typename componentType>
 struct vec3 : public vec3_base {
@@ -199,15 +199,15 @@ struct vec3 : public vec3_base {
         if (gc_m[i] == r && gc_m[i] == g && gc_m[i] == b) {
             return;
         } else {
-            r = r / gc_s[i] * 2.0f - 1.0f;
-            g = g / gc_s[i] * 2.0f - 1.0f;
-            b = b / gc_s[i] * 2.0f - 1.0f;
+            r = (float)(r / (double)gc_s[i]) * 2.0f - 1.0f;
+            g = (float)(g / (double)gc_s[i]) * 2.0f - 1.0f;
+            b = (float)(b / (double)gc_s[i]) * 2.0f - 1.0f;
             clamp(-1.0f, 1.0f);
             base_normalize();
-            r = (std::floor((r + 1.0f) * gc_s[i] * 0.5f + 0.5f));
-            g = (std::floor((g + 1.0f) * gc_s[i] * 0.5f + 0.5f));
-            b = (std::floor((b + 1.0f) * gc_s[i] * 0.5f + 0.5f));
-            clamp(0, gc_s[i]);
+            r = (std::floor((r + 1.0f) * (float)gc_s[i] * 0.5f + 0.5f));
+            g = (std::floor((g + 1.0f) * (float)gc_s[i] * 0.5f + 0.5f));
+            b = (std::floor((b + 1.0f) * (float)gc_s[i] * 0.5f + 0.5f));
+            clamp(0.0f, (float)gc_s[i]);
         }
     }
 };
@@ -342,7 +342,7 @@ class color<componentType, 2> : public color_base<componentType, 2> {
     }
     void normalize() {
         vec3<componentType> v((float)r, (float)g,
-                              gc_s[sizeof(componentType)] * 0.5f);
+                              (float)gc_s[sizeof(componentType)] * 0.5f);
         v.normalize();
         r = (componentType)v.r;
         g = (componentType)v.g;
@@ -381,10 +381,10 @@ class color<componentType, 1> : public color_base<componentType, 1> {
         return 1;
     }
     void normalize() {
-        vec3<componentType> v((float)r, gc_s[sizeof(componentType)] * 0.5f,
-                              gc_s[sizeof(componentType)] * 0.5f);
-        v.normalize();
-        r = (componentType)v.r;
+		// Normalizing single channel image doesn't make much sense
+		// Here I assume single channel color is (X, 0, 0, 0)
+		if (r != 0)
+			r = (componentType)gc_s[sizeof(componentType)];
     }
 };