From: Steinar H. Gunderson <sgunderson@bigfoot.com>
Date: Fri, 6 Oct 2017 18:08:46 +0000 (+0200)
Subject: Add support for repeating blocks. About 2% size reduction.
X-Git-Url: https://git.sesse.net/?p=narabu;a=commitdiff_plain;h=056ba403e3ee3b9a362ac85196d4bcce6a73d98c

Add support for repeating blocks. About 2% size reduction.
---

diff --git a/coded.dat b/coded.dat
index b249f4c..a811389 100644
Binary files a/coded.dat and b/coded.dat differ
diff --git a/narabu.cpp b/narabu.cpp
index bb1209c..0ac4f71 100644
--- a/narabu.cpp
+++ b/narabu.cpp
@@ -223,6 +223,7 @@ int main(int argc, char **argv)
 	printf("%d err=0x%x\n", __LINE__, glGetError());
 
 	// Decode all luma blocks.
+	size_t last_src_offset = 0, last_src_len = 0;
 	for (unsigned y = 0; y < 8; ++y) {
                 for (unsigned x = 0; x < 8; ++x) {
 			unsigned coeff_num = y * 8 + x;
@@ -235,8 +236,16 @@ int main(int argc, char **argv)
 				}
 
 				CoeffStream *stream = &streams[coeff_num * num_blocks + block_idx / BLOCKS_PER_STREAM];
-				stream->src_offset = ptr - coded.data();
-				stream->src_len = *num_rans_bytes;
+				if (*num_rans_bytes == 0) {
+					// Repeat last stream.
+					stream->src_offset = last_src_offset;
+					stream->src_len = last_src_len;
+				} else {
+					stream->src_offset = ptr - coded.data();
+					stream->src_len = *num_rans_bytes;
+					last_src_offset = stream->src_offset;
+					last_src_len = last_src_len;
+				}
 
 				// TODO: check len
 				ptr += *num_rans_bytes;
diff --git a/qdc.cpp b/qdc.cpp
index 41ef04d..0339e12 100644
--- a/qdc.cpp
+++ b/qdc.cpp
@@ -235,22 +235,14 @@ public:
 		//printf("post-flush = %08x\n", rans);
 
 		uint32_t num_rans_bytes = out_end - ptr;
-#if 0
-		if (num_rans_bytes == 4) {
-			uint32_t block;
-			memcpy(&block, ptr, 4);
-
-			if (block == last_block) {
-				write_varint(0, codedfp);
-				clear();
-				return 1;
-			}
-
-			last_block = block;
+		if (num_rans_bytes == last_block.size() &&
+		    memcmp(last_block.data(), ptr, last_block.size()) == 0) {
+			write_varint(0, codedfp);
+			clear();
+			return 1;
 		} else {
-			last_block = 0;
+			last_block = string((const char *)ptr, num_rans_bytes);
 		}
-#endif
 
 		write_varint(num_rans_bytes, codedfp);
 		//fwrite(&num_rans_bytes, 1, 4, codedfp);
@@ -294,7 +286,7 @@ private:
 	RansEncSymbol esyms[NUM_SYMS];
 	uint32_t sign_bias;
 
-	uint32_t last_block = 0;  // Not a valid 4-byte rANS block (?)
+	std::string last_block;
 };
 
 static constexpr int dc_scalefac = 8;  // Matches the FDCT's gain.