From: Steinar H. Gunderson <sgunderson@bigfoot.com>
Date: Sun, 24 Sep 2017 17:39:25 +0000 (+0200)
Subject: Make blocks per stream a named constant.
X-Git-Url: https://git.sesse.net/?p=narabu;a=commitdiff_plain;h=69e36d8b488476d9d0e23f5b54eb18a4f9b5fbdd

Make blocks per stream a named constant.
---

diff --git a/decoder.shader b/decoder.shader
index 724a2f5..3b72126 100644
--- a/decoder.shader
+++ b/decoder.shader
@@ -16,6 +16,7 @@ const uint prob_bits = 12;
 const uint prob_scale = 1 << prob_bits;
 const uint NUM_SYMS = 256;
 const uint ESCAPE_LIMIT = NUM_SYMS - 1;
+const uint BLOCKS_PER_STREAM = 320;
 
 // These need to be folded into quant_matrix.
 const float dc_scalefac = 8.0;
@@ -230,7 +231,7 @@ void main()
 
 	pick_timer(start, local_timing[0]);
 
-	for (uint block_idx = 40; block_idx --> 0; ) {
+	for (uint block_idx = BLOCKS_PER_STREAM / 8; block_idx --> 0; ) {
 		pick_timer(start, local_timing[1]);
 
 		// rANS decode one coefficient across eight blocks (so 64x8 coefficients).
diff --git a/narabu.cpp b/narabu.cpp
index 11117f4..898f7ef 100644
--- a/narabu.cpp
+++ b/narabu.cpp
@@ -24,6 +24,7 @@ const unsigned prob_bits = 12;
 const unsigned prob_scale = 1 << prob_bits;
 const unsigned NUM_SYMS = 256;
 const unsigned NUM_TABLES = 8;
+const unsigned BLOCKS_PER_STREAM = 320;
 
 struct RansDecSymbol {
         unsigned sym_start;
@@ -266,7 +267,7 @@ 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_slices = (WIDTH/8)*(HEIGHT/8)/320;
+		unsigned num_slices = (WIDTH/8)*(HEIGHT/8)/BLOCKS_PER_STREAM;
 		glDispatchCompute(1, (num_slices+PARALLEL_SLICES-1)/PARALLEL_SLICES, 1);
 	}
 	check_error();
diff --git a/qdc.cpp b/qdc.cpp
index 5a4a9d5..a0919bd 100644
--- a/qdc.cpp
+++ b/qdc.cpp
@@ -23,8 +23,10 @@
 #define HEIGHT_BLOCKS (HEIGHT/8)
 #define NUM_BLOCKS (WIDTH_BLOCKS * HEIGHT_BLOCKS)
 #define NUM_BLOCKS_CHROMA (WIDTH_BLOCKS_CHROMA * HEIGHT_BLOCKS)
+
 #define NUM_SYMS 256
 #define ESCAPE_LIMIT (NUM_SYMS - 1)
+#define BLOCKS_PER_STREAM 320
 
 // If you set this to 1, the program will try to optimize the placement
 // of coefficients to rANS probability distributions. This is randomized,
@@ -720,10 +722,10 @@ int main(int argc, char **argv)
 #endif
 
 	// DC coefficient pred from the right to left (within each slice)
-	for (unsigned block_idx = 0; block_idx < NUM_BLOCKS; block_idx += 320) {
+	for (unsigned block_idx = 0; block_idx < NUM_BLOCKS; block_idx += BLOCKS_PER_STREAM) {
 		int prev_k = 128;
 
-		for (unsigned subblock_idx = 320; subblock_idx --> 0; ) {
+		for (unsigned subblock_idx = BLOCKS_PER_STREAM; subblock_idx --> 0; ) {
 			unsigned yb = (block_idx + subblock_idx) / WIDTH_BLOCKS;
 			unsigned xb = (block_idx + subblock_idx) % WIDTH_BLOCKS;
 			int k = coeff_y[(yb * 8) * WIDTH + (xb * 8)];
@@ -733,11 +735,11 @@ int main(int argc, char **argv)
 			prev_k = k;
 		}
 	}
-	for (unsigned block_idx = 0; block_idx < NUM_BLOCKS_CHROMA; block_idx += 320) {
+	for (unsigned block_idx = 0; block_idx < NUM_BLOCKS_CHROMA; block_idx += BLOCKS_PER_STREAM) {
 		int prev_k_cb = 0;
 		int prev_k_cr = 0;
 
-		for (unsigned subblock_idx = 320; subblock_idx --> 0; ) {
+		for (unsigned subblock_idx = BLOCKS_PER_STREAM; subblock_idx --> 0; ) {
 			unsigned yb = (block_idx + subblock_idx) / WIDTH_BLOCKS_CHROMA;
 			unsigned xb = (block_idx + subblock_idx) % WIDTH_BLOCKS_CHROMA;
 			int k_cb = coeff_cb[(yb * 8) * WIDTH/2 + (xb * 8)];
@@ -880,7 +882,7 @@ int main(int argc, char **argv)
 				//printf("encoding coeff %d xb,yb=%d,%d: %d\n", y*8+x, xb, yb, k);
 				rans_encoder.encode_coeff(k);
 
-				if (block_idx % 320 == 319 || block_idx == NUM_BLOCKS - 1) {
+				if (block_idx % BLOCKS_PER_STREAM == (BLOCKS_PER_STREAM - 1) || block_idx == NUM_BLOCKS - 1) {
 					int l = rans_encoder.save_block(codedfp);
 					num_bytes += l;
 					lens.push_back(l);
@@ -920,7 +922,7 @@ int main(int argc, char **argv)
 				//printf("encoding coeff %d xb,yb=%d,%d: %d\n", y*8+x, xb, yb, k);
 				rans_encoder.encode_coeff(k);
 
-				if (block_idx % 320 == 319 || block_idx == NUM_BLOCKS - 1) {
+				if (block_idx % BLOCKS_PER_STREAM == (BLOCKS_PER_STREAM - 1) || block_idx == NUM_BLOCKS - 1) {
 					num_bytes += rans_encoder.save_block(codedfp);
 				}
 			}
@@ -945,7 +947,7 @@ int main(int argc, char **argv)
 				//printf("encoding coeff %d xb,yb=%d,%d: %d\n", y*8+x, xb, yb, k);
 				rans_encoder.encode_coeff(k);
 
-				if (block_idx % 320 == 319 || block_idx == NUM_BLOCKS - 1) {
+				if (block_idx % BLOCKS_PER_STREAM == (BLOCKS_PER_STREAM - 1) || block_idx == NUM_BLOCKS - 1) {
 					num_bytes += rans_encoder.save_block(codedfp);
 				}
 			}