X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=narabu.cpp;h=0ac4f71d75ab58794da432e64c22ea0811cdb437;hb=daf421e32981645e551621551c6b82697ad078de;hp=a0e58717e6b32da507d8621d4b73e908769e4806;hpb=de545fc4a6740687509e9ffd871e808a48281fb8;p=narabu

diff --git a/narabu.cpp b/narabu.cpp
index a0e5871..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;
@@ -271,7 +280,8 @@ int main(int argc, char **argv)
 
 #define PARALLEL_SLICES 1
 	steady_clock::time_point start = steady_clock::now();
-	for (int i = 0; i < 1000; ++i) {
+	unsigned num_iterations = 1000;
+	for (unsigned i = 0; i < num_iterations; ++i) {
 		unsigned num_slices = (WIDTH/8)*(HEIGHT/8)/BLOCKS_PER_STREAM;
 		glDispatchCompute(1, (num_slices+PARALLEL_SLICES-1)/PARALLEL_SLICES, 1);
 	}
@@ -373,5 +383,5 @@ int main(int argc, char **argv)
 	glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0); // unbind
 	
 	printf("foo = 0x%x\n", glGetError());
-	printf("Each iteration took %.3f ms.\n", 1e3 * duration<double>(now - start).count() / 1000);
+	printf("Each iteration took %.3f ms.\n", 1e3 * duration<double>(now - start).count() / num_iterations);
 }