Revert "added better test by rlblaster / ypsu / kbalazs"

This reverts commit cc96941dca32c222cbdb12808e8abb2dec78b5e6.

magyarsort.h

@@ -9,9 +9,9 @@
  * LICENCE: CC3 - look it up, you need to mention me but that is all
  */
 
-#include <vector>
 #include <cstdio>
 #include <cstdint>
+#include <cstring> // memset
 
 namespace MagyarSort {
 
@@ -23,44 +23,67 @@ namespace MagyarSort {
 	// - DIGIT_RANGE and BITS_PER_DIGIT should correspond
 	// - DIGITS should also correspond with the uint32_t
 	// - and DIGIT_RANGE should be 2^n value (16 or 256)
-	static constexpr int DIGITS = 4; // "helyiérték"
-	static constexpr int BITS_PER_DIGIT = 8; // "bit / helyiérték"
-	static constexpr int DIGIT_RANGE = 256; // "helyiérték állapottér"
+	static constexpr int DIGITS = 8; // "helyiérték"
+	static constexpr int BITS_PER_DIGIT = 4; // "bit / helyiérték"
+	static constexpr int DIGIT_RANGE = 16; // "helyiérték állapottér"
 
-	template<int digitChoice>
+	template<int DIGIT_CHOICE>
 	static inline uint32_t getDigit(uint32_t num) noexcept {
-		static constexpr int SHIFT = digitChoice * BITS_PER_DIGIT;
+		static constexpr int SHIFT = DIGIT_CHOICE * BITS_PER_DIGIT;
 
 		uint32_t shifted = num >> SHIFT;
 		return shifted & (DIGIT_RANGE - 1);
 	}
 
-	/** Sort the given array (in-place sorting) with the given size */
-	inline void sort(uint32_t arr[], size_t size) noexcept {
-		// Holds "digit" occurences, prefix sums, whatevers
-		// First "DIGIT_RANGE" elem is for MSB "DIGITS", last is for LSB
-		static thread_local size_t radics[DIGITS * DIGIT_RANGE];
-		for(int i = 0; i < (DIGITS * DIGIT_RANGE); ++i) { radics[i] = 0; }
-
-		// Calculate occurences of digits
-		for(size_t i = 0; i < size; ++i) {
-			auto d0 = getDigit<0>(arr[i]);
-			auto d1 = getDigit<1>(arr[i]);
-			auto d2 = getDigit<2>(arr[i]);
-			auto d3 = getDigit<3>(arr[i]);
-
-			printf("d0:%u, arr[i]: %u\n", d0, arr[i]);
-			printf("d1:%u, arr[i]: %u\n", d1, arr[i]);
-			printf("d2:%u, arr[i]: %u\n", d2, arr[i]);
-			printf("d3:%u, arr[i]: %u\n", d3, arr[i]);
-
-			++radics[d0];
-			++radics[d1 + DIGIT_RANGE * 1];
-			++radics[d2 + DIGIT_RANGE * 2];
-			++radics[d3 + DIGIT_RANGE * 3];
+	/** Recursive Functor: no class should be generated I think (compiler should be smart) */
+	template<int DIGIT>
+	struct OccurenceMagic : public OccurenceMagic<DIGIT - 1> {
+		inline OccurenceMagic(uint32_t arr[], size_t i, size_t *radicsOut) noexcept
+				: OccurenceMagic<DIGIT -1 >(arr, i, radicsOut) {
+			// Parents run first so template recursion runs DIGIT=0 first...
+			++radicsOut[getDigit<DIGIT>(arr[i]) + DIGIT_RANGE * DIGIT];
 		}
+	};
+	/** Ends template recursion */
+	template<>
+	struct OccurenceMagic<-1> {
+		inline OccurenceMagic(uint32_t arr[], size_t i, size_t *radicsOut) noexcept {}
+	};
 
-		// TODO: remove debug stuffz
+	static inline void countOccurences(uint32_t arr[], size_t size, size_t *radicsOut) noexcept {
+		for(size_t i = 0; i < size; ++i) {
+			// Creates no object, struct is empty
+			OccurenceMagic<DIGITS - 1>(arr, i, radicsOut);
+		}
+	}
+
+	/** Recursive Functor: no class should be generated I think (compiler should be smart) */
+	template<int DIGIT>
+	struct PrefixMagic : public PrefixMagic<DIGIT - 1> {
+		inline PrefixMagic(size_t *radics, size_t *prev, int i) noexcept
+				: PrefixMagic<DIGIT - 1>(radics, prev, i) {
+			static constexpr int DSTART = (DIGIT * DIGIT_RANGE);
+			radics[DSTART + i] += prev[DIGIT];
+			prev[DIGIT] = radics[DSTART + i];
+		}
+	};
+	/** Ends template recursion */
+	template<>
+	struct PrefixMagic<-1> {
+		inline PrefixMagic(size_t *radics, size_t *prev, int i) noexcept {}
+	};
+
+	static inline void calcPrefixSums(size_t *radics) noexcept {
+		static thread_local size_t prev[DIGITS];
+		memset(prev, 0, sizeof(prev));
+
+		for(int i = 0; i < DIGIT_RANGE; ++i) {
+			// This is a template-unrolled loop too
+			PrefixMagic<DIGITS - 1>(radics, prev, i);
+		}
+	}
+
+	void debugIt(size_t *radics) {
 		for(size_t j = 0; j < DIGITS; ++j) {
 			printf("d%d: ", j);
 			for(size_t i = 0; i < DIGIT_RANGE; ++i) {
@@ -69,6 +92,24 @@ namespace MagyarSort {
 			printf("\n\n");
 		}
 	}
+
+	/** Sort the given array (in-place sorting) with the given size */
+	inline void sort(uint32_t arr[], size_t size) noexcept {
+		// Holds "digit" occurences, prefix sums, whatevers
+		// First "DIGIT_RANGE" elem is for MSB "DIGITS", last is for LSB
+		static thread_local size_t radics[DIGITS * DIGIT_RANGE];
+		memset(radics, 0, sizeof(radics));
+
+		// Calculate occurences of digits
+		countOccurences(arr, size, radics);
+
+		debugIt(radics);
+
+		// Calculate prefix sums
+		calcPrefixSums(radics);
+
+		debugIt(radics);
+	}
 };
 
 #endif