From cdcaa0ccb416d48a0689839cfdf78faaf67bf8a9 Mon Sep 17 00:00:00 2001
From: Sebastian Deorowicz <sebastian.deorowicz@polsl.pl>
Date: Fri, 11 Oct 2024 15:20:14 +0200
Subject: [PATCH] Feature/faster decompression (#6)

---
 src/defs.h          |  4 +--
 src/parser.h        | 18 ++++++++++---
 src/seq_reservoir.h | 66 ++++++++++++++++++++++++++++++++++-----------
 3 files changed, 68 insertions(+), 20 deletions(-)

diff --git a/src/defs.h b/src/defs.h
index c555f52..845a577 100644
--- a/src/defs.h
+++ b/src/defs.h
@@ -15,8 +15,8 @@
 #include <string>
 #include "params.h"
 
-const std::string LZ_ANI_VER = "lz-ani 1.2.0";
-const std::string LZ_ANI_DATE = "2024-10-09";
+const std::string LZ_ANI_VER = "lz-ani 1.2.1";
+const std::string LZ_ANI_DATE = "2024-10-11";
 const std::string LZ_ANI_AUTHORS = "Sebastian Deorowicz, Adam Gudys";
 const std::string LZ_ANI_INFO = LZ_ANI_VER + " (" + LZ_ANI_DATE + ") by " + LZ_ANI_AUTHORS;
 
diff --git a/src/parser.h b/src/parser.h
index 4fea926..0e4f08f 100644
--- a/src/parser.h
+++ b/src/parser.h
@@ -30,6 +30,8 @@ class CParser
 
 	CParams params;
 
+	seq_t seq_working;
+
 	seq_t seq_ref;
 	seq_t seq_data;
 	uint32_t n_ref_seqs = 0; 
@@ -59,21 +61,31 @@ class CParser
 
 	void append(seq_t& seq, const seq_view& sv, const uint8_t allowed_N, const uint8_t forbidden_N)
 	{
-		for (uint32_t i = 0; i < sv.size(); ++i)
+/*		for (uint32_t i = 0; i < sv.size(); ++i)
 		{
 			auto c = sv[i];
 			if (c == forbidden_N)
 				seq.emplace_back(allowed_N);
 			else
 				seq.emplace_back(c);
-		}
+		}*/
+
+		auto prev_size = seq.size();
+		seq.resize(prev_size + sv.size());
+		sv.unpack(seq.data() + prev_size);
+		replace(seq.begin() + prev_size, seq.end(), forbidden_N, allowed_N);
 	}
 
 	void append_rc(seq_t& seq, const seq_view& sv, const uint8_t allowed_N, const uint8_t forbidden_N)
 	{
+		seq.reserve(seq.size() + sv.size());
+
+		seq_working.resize(sv.size());
+		sv.unpack(seq_working.data());
+
 		for (uint32_t i = 0; i < sv.size(); ++i)
 		{
-			auto c = sv[sv.size() - 1 - i];
+			auto c = seq_working[seq_working.size() - 1 - i];
 			if (c == forbidden_N)
 				seq.emplace_back(allowed_N);
 			else if (c < code_N)
diff --git a/src/seq_reservoir.h b/src/seq_reservoir.h
index 2f88c01..c77fe76 100644
--- a/src/seq_reservoir.h
+++ b/src/seq_reservoir.h
@@ -33,6 +33,25 @@ class seq_view
 	const uint32_t len;
 	const internal_packing_t internal_packing;
 
+	static inline uint8_t triples[256][4];
+
+	struct _si
+	{
+		_si()
+		{
+			for(uint8_t i = 0; i < 6; ++i)
+				for(uint8_t j = 0; j < 6; ++j)
+					for (uint8_t k = 0; k < 6; ++k)
+					{
+						auto idx = 36 * i + 6 * j + k;
+						triples[idx][0] = i;
+						triples[idx][1] = j;
+						triples[idx][2] = k;
+						triples[idx][3] = 0;
+					}
+		}
+	} static inline _init;
+
 public:
 	seq_view(const uint8_t *data = 0, const uint32_t len = 0, internal_packing_t internal_packing = internal_packing_t::none) :
 		data(data), 
@@ -90,21 +109,21 @@ class seq_view
 
 		if (len & 1)
 			dest[len / 2] = src[len - 1] << 4;
-	}
+	}*/
 
-	static void unpack2(uint8_t* dest, uint8_t* src, uint32_t len)
+	void unpack2(uint8_t* dest) const
 	{
 		for (uint32_t i = 0; i < len / 2; ++i)
 		{
-			dest[2 * i] = src[i] >> 4;
-			dest[2 * i + 1] = src[i] & 0xf;
+			dest[2 * i] = data[i] >> 4;
+			dest[2 * i + 1] = data[i] & 0xf;
 		}
 
 		if (len & 1)
-			dest[len - 1] = src[len / 2] >> 4;
+			dest[len - 1] = data[len / 2] >> 4;
 	}
 
-	static void pack3(uint8_t* dest, uint8_t* src, uint32_t len)
+/*	static void pack3(uint8_t* dest, uint8_t* src, uint32_t len)
 	{
 		for (uint32_t i = 0; i < len / 3; ++i)
 			dest[i] = 36 * src[3 * i] + 6 * src[3 * i + 1] + src[3 * i + 2];
@@ -121,32 +140,49 @@ class seq_view
 			// Nothing
 			break;
 		}
-	}
+	}*/
 
-	static void unpack3(uint8_t * dest, uint8_t * src, uint32_t len)
+	void unpack3(uint8_t *dest) const
 	{
 		uint32_t len_div_3 = len / 3;
 
 		for (uint32_t i = 0; i < len_div_3; ++i)
 		{
-			dest[3 * i] = src[i] / 36;
-			dest[3 * i + 1] = src[i] / 6 - 6 * dest[3 * i];
-			dest[3 * i + 2] = src[i] - 36 * dest[3 * i] - 6 * dest[3 * i + 1];
+			dest[3 * i] = triples[data[i]][0];
+			dest[3 * i + 1] = triples[data[i]][1];
+			dest[3 * i + 2] = triples[data[i]][2];
 		}
 
 		switch (len % 3)
 		{
 		case 2:
-			dest[len - 2] = src[len_div_3] / 6;
-			dest[len - 1] = src[len_div_3] - 6 * dest[len - 2];
+			dest[len - 2] = triples[data[len_div_3]][0];
+			dest[len - 1] = triples[data[len_div_3]][1];
 			break;
 		case 1:
-			dest[len - 1] = src[len_div_3];
+			dest[len - 1] = triples[data[len_div_3]][0];
+			break;
 		case 0:
 			// Nothing
 			break;
 		}
-	}*/
+	}
+
+	void unpack(uint8_t* dest) const 
+	{
+		switch (internal_packing)
+		{
+		case internal_packing_t::none:
+			copy_n(data, len, dest);
+			break;
+		case internal_packing_t::two_in_byte:
+			unpack2(dest);
+			break;
+		case internal_packing_t::three_in_byte:
+			unpack3(dest);
+			break;
+		}
+	}
 };
 
 class CSeqReservoir