From: Steinar H. Gunderson <sgunderson@bigfoot.com>
Date: Sun, 24 Sep 2017 17:45:40 +0000 (+0200)
Subject: Make the GPU decoder (finally) work with any resolution.
X-Git-Url: https://git.sesse.net/?p=narabu;a=commitdiff_plain;h=de545fc4a6740687509e9ffd871e808a48281fb8

Make the GPU decoder (finally) work with any resolution.
---

diff --git a/narabu.cpp b/narabu.cpp
index 898f7ef..a0e5871 100644
--- a/narabu.cpp
+++ b/narabu.cpp
@@ -19,6 +19,11 @@ using namespace std::chrono;
 
 #define WIDTH 1280
 #define HEIGHT 720
+#define WIDTH_BLOCKS (WIDTH/8)
+#define WIDTH_BLOCKS_CHROMA (WIDTH/16)
+#define HEIGHT_BLOCKS (HEIGHT/8)
+#define NUM_BLOCKS (WIDTH_BLOCKS * HEIGHT_BLOCKS)
+#define NUM_BLOCKS_CHROMA (WIDTH_BLOCKS_CHROMA * HEIGHT_BLOCKS)
 
 const unsigned prob_bits = 12;
 const unsigned prob_scale = 1 << prob_bits;
@@ -54,10 +59,12 @@ optional<uint32_t> read_varint(const char **ptr, const char *end)
 	return nullopt;  // Error: EOF.
 }
 
+const unsigned num_blocks = ((NUM_BLOCKS + BLOCKS_PER_STREAM - 1) / BLOCKS_PER_STREAM);
+
 struct CoeffStream {
         uint src_offset, src_len;
 };
-CoeffStream streams[45 * 64];  // HACK
+CoeffStream streams[num_blocks * 64];
 
 int main(int argc, char **argv)
 {
@@ -202,8 +209,6 @@ int main(int argc, char **argv)
 	GLint num_blocks_pos = glGetUniformLocation(glsl_program_num, "num_blocks");
 	printf("%d err=0x%x pos=%d,%d,%d,%d\n", __LINE__, glGetError(), cum2sym_tex_pos, dsyms_tex_pos, out_tex_pos, sign_bias_pos);
 
-	unsigned num_blocks = (HEIGHT / 16);
-
 	// Bind the textures.
 	glUniform1i(cum2sym_tex_pos, 0);
 	glUniform1i(dsyms_tex_pos, 1);
@@ -222,14 +227,14 @@ int main(int argc, char **argv)
                 for (unsigned x = 0; x < 8; ++x) {
 			unsigned coeff_num = y * 8 + x;
 
-			for (unsigned yb = 0; yb < HEIGHT; yb += 16) {
+			for (unsigned block_idx = 0; block_idx < NUM_BLOCKS; block_idx += BLOCKS_PER_STREAM) {
 				optional<uint32_t> num_rans_bytes = read_varint(&ptr, end);
 				if (!num_rans_bytes) {
-					fprintf(stderr, "Error parsing varint for block %d rANS bytes\n", yb);
+					fprintf(stderr, "Error parsing varint for block %d rANS bytes\n", block_idx);
 					exit(1);
 				}
 
-				CoeffStream *stream = &streams[coeff_num * num_blocks + (yb/16)];
+				CoeffStream *stream = &streams[coeff_num * num_blocks + block_idx / BLOCKS_PER_STREAM];
 				stream->src_offset = ptr - coded.data();
 				stream->src_len = *num_rans_bytes;