From: Steinar H. Gunderson <sgunderson@bigfoot.com>
Date: Sun, 24 Sep 2017 13:29:44 +0000 (+0200)
Subject: Predict DC across the entire slice instead of resetting each row. Opens up for slices... 
X-Git-Url: https://git.sesse.net/?a=commitdiff_plain;ds=sidebyside;h=2161c821c9c65ebae24eff9c2e4809c11c67cd02;p=narabu

Predict DC across the entire slice instead of resetting each row. Opens up for slices crossing rows easier.
---

diff --git a/coded.dat b/coded.dat
index 50795ae..ea88547 100644
Binary files a/coded.dat and b/coded.dat differ
diff --git a/decoder.shader b/decoder.shader
index e7b83e0..cab6f8c 100644
--- a/decoder.shader
+++ b/decoder.shader
@@ -229,10 +229,6 @@ void main()
 	pick_timer(start, local_timing[0]);
 
 	for (uint block_idx = 40; block_idx --> 0; ) {
-		uint block_x = block_idx % 20;
-		uint block_y = block_idx / 20;
-		if (block_x == 19) last_k = 0;
-
 		pick_timer(start, local_timing[1]);
 
 		// rANS decode one coefficient across eight blocks (so 64x8 coefficients).
@@ -309,8 +305,12 @@ void main()
 
 		pick_timer(start, local_timing[6]);
 
-		uint y = block_row * 16 + block_y * 8;
-		uint x = block_x * 64 + local_y * 8 + local_x;
+		uint global_block_idx = (block_row * 40 + block_idx) * 8 + local_y;
+		uint block_x = global_block_idx % 160;
+		uint block_y = global_block_idx / 160;
+
+		uint y = block_y * 8;
+		uint x = block_x * 8 + local_x;
 		for (uint yl = 0; yl < 8; ++yl) {
 			imageStore(out_tex, ivec2(x, yl + y), vec4(temp[row_offset + yl * 8], 0.0, 0.0, 1.0));
 		}
diff --git a/qdc.cpp b/qdc.cpp
index 556a8a5..f1ff254 100644
--- a/qdc.cpp
+++ b/qdc.cpp
@@ -719,16 +719,35 @@ int main(int argc, char **argv)
 		chroma_energy / (WIDTH * HEIGHT), chroma_energy_pred / (WIDTH * HEIGHT));
 #endif
 
-	// DC coefficient pred from the right to left
-	for (unsigned yb = 0; yb < HEIGHT; yb += 8) {
-		for (unsigned xb = 0; xb < WIDTH - 8; xb += 8) {
-			coeff_y[yb * WIDTH + xb] -= coeff_y[yb * WIDTH + (xb + 8)];
+	// DC coefficient pred from the right to left (within each slice)
+	for (unsigned block_idx = 0; block_idx < NUM_BLOCKS; block_idx += 320) {
+		int prev_k = 0;
+
+		for (unsigned subblock_idx = 320; 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)];
+
+			coeff_y[(yb * 8) * WIDTH + (xb * 8)] = k - prev_k;
+
+			prev_k = k;
 		}
 	}
-	for (unsigned yb = 0; yb < HEIGHT; yb += 8) {
-		for (unsigned xb = 0; xb < WIDTH/2 - 8; xb += 8) {
-			coeff_cb[yb * WIDTH/2 + xb] -= coeff_cb[yb * WIDTH/2 + (xb + 8)];
-			coeff_cr[yb * WIDTH/2 + xb] -= coeff_cr[yb * WIDTH/2 + (xb + 8)];
+	for (unsigned block_idx = 0; block_idx < NUM_BLOCKS_CHROMA; block_idx += 320) {
+		int prev_k_cb = 0;
+		int prev_k_cr = 0;
+
+		for (unsigned subblock_idx = 320; 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)];
+			int k_cr = coeff_cr[(yb * 8) * WIDTH/2 + (xb * 8)];
+
+			coeff_cb[(yb * 8) * WIDTH/2 + (xb * 8)] = k_cb - prev_k_cb;
+			coeff_cr[(yb * 8) * WIDTH/2 + (xb * 8)] = k_cr - prev_k_cr;
+
+			prev_k_cb = k_cb;
+			prev_k_cr = k_cr;
 		}
 	}