X-Git-Url: https://git.sesse.net/?p=narabu;a=blobdiff_plain;f=narabu-encoder.cpp;h=fe270641634652debd37edba61470d585da5fbf8;hp=36804d616f70681f411af412961a23ac87a542b4;hb=0f84eb2a5fa77dbc20f7a073f83bdb674ebb7b75;hpb=fedc6308264085999789854152c9577ddfe1bc6e

diff --git a/narabu-encoder.cpp b/narabu-encoder.cpp
index 36804d6..fe27064 100644
--- a/narabu-encoder.cpp
+++ b/narabu-encoder.cpp
@@ -227,10 +227,10 @@ int main(int argc, char **argv)
 	glNamedBufferStorage(output_ssbo, 45 * 64 * 1024, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 	check_error();
 
-	GLuint output_offset_ssbo;
-	glGenBuffers(1, &output_offset_ssbo);
-	glBindBuffer(GL_SHADER_STORAGE_BUFFER, output_offset_ssbo);
-	glNamedBufferStorage(output_offset_ssbo, 45 * 64 * sizeof(uint32_t), nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
+	GLuint bytes_written_ssbo;
+	glGenBuffers(1, &bytes_written_ssbo);
+	glBindBuffer(GL_SHADER_STORAGE_BUFFER, bytes_written_ssbo);
+	glNamedBufferStorage(bytes_written_ssbo, 45 * 64 * sizeof(uint32_t), nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 	check_error();
 
 	// Bind SSBOs.
@@ -243,7 +243,7 @@ int main(int argc, char **argv)
 
 	glUseProgram(glsl_rans_program_num);
 	glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 10, output_ssbo);
-	glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 11, output_offset_ssbo);
+	glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 11, bytes_written_ssbo);
 	glBindBufferBase(GL_UNIFORM_BUFFER, 13, dist_ubo);
 
 	glUseProgram(glsl_program_num);
@@ -383,9 +383,8 @@ int main(int argc, char **argv)
 	}
 
 	// Write out the actual data.
-	// TODO: Do the deduplication.
 
-	const uint32_t *offsets = (const uint32_t *)glMapNamedBufferRange(output_offset_ssbo, 0, 45 * 64 * sizeof(uint32_t), GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
+	const uint32_t *bytes_written = (const uint32_t *)glMapNamedBufferRange(bytes_written_ssbo, 0, 45 * 64 * sizeof(uint32_t), GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 #if 0
 	for (int i = 0; i < 45*64; ++i) {
 		printf("%d,%d,%d: %u\n", i / 64, (i / 8) % 8, i % 8, 1024 * (i + 1) - offsets[i]);
@@ -399,8 +398,8 @@ int main(int argc, char **argv)
 		for (unsigned x = 0; x < 8; ++x) {
 			for (unsigned int stream_idx = 0; stream_idx < 45; ++stream_idx) {
 				const uint8_t *out_end = data + (stream_idx * 64 + y * 8 + x + 1) * 1024;
-				const uint8_t *ptr = data + offsets[stream_idx * 64 + y * 8 + x];
-				uint32_t num_rans_bytes = out_end - ptr;
+				uint32_t num_rans_bytes = bytes_written[stream_idx * 64 + y * 8 + x];
+				const uint8_t *ptr = out_end - num_rans_bytes;
 				assert(num_rans_bytes <= 1024);
 
 				if (num_rans_bytes == last_block.size() &&