X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=decoder.shader;h=616c3e2dfd0896e5cd035bacc42ec0720c5174a1;hb=5e1d27014149311318e97b8e04a6e05ec858e57c;hp=7136e03e5ff0de76095f1fd5974f68fe0bc5a3ba;hpb=bed1185c791ae562ae0b5934a64a95440285729b;p=narabu

diff --git a/decoder.shader b/decoder.shader
index 7136e03..616c3e2 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;
@@ -161,8 +162,8 @@ void idct_1d(inout float y0, inout float y1, inout float y2, inout float y3, ino
 	y0 = p6_0 + p6_7;
 	y1 = p6_1 + p6_6;
 	y2 = p6_2 + p6_5;
-	y3 = p6_3 - p4_4;
-	y4 = p6_3 + p4_4;
+	y3 = p6_3 + p4_4;
+	y4 = p6_3 - p4_4;
 	y5 = p6_2 - p6_5;
 	y6 = p6_1 - p6_6;
 	y7 = p6_0 - p6_7;
@@ -203,6 +204,8 @@ void main()
 	local_timing[0] = start;
 #endif
 
+	const uint blocks_per_row = (imageSize(out_tex).x + 7) / 8;
+
 	const uint local_x = gl_LocalInvocationID.x % 8;
 	const uint local_y = (gl_LocalInvocationID.x / 8) % 8;
 	const uint local_z = gl_LocalInvocationID.x / 64;
@@ -228,7 +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).
@@ -306,8 +309,8 @@ void main()
 		pick_timer(start, local_timing[6]);
 
 		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 block_x = global_block_idx % blocks_per_row;
+		uint block_y = global_block_idx / blocks_per_row;
 
 		uint y = block_y * 8;
 		uint x = block_x * 8 + local_x;