X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=rans.shader;h=925543ef7dfa284d0fddbce411a61444438dbc9f;hb=dc3e3a2a88b2b1f0e177f613a8b355116e6020ff;hp=b77d2407adab28496e783a1953d76f043f312b51;hpb=b877c1eae5d301a3ce36e0ba9b0b830a95e06ec8;p=narabu

diff --git a/rans.shader b/rans.shader
index b77d240..925543e 100644
--- a/rans.shader
+++ b/rans.shader
@@ -3,7 +3,7 @@
 
 layout(local_size_x = 1) in;
 
-const uint prob_bits = 13;  // Note!
+const uint prob_bits = 12;
 const uint prob_scale = 1 << prob_bits;
 const uint RANS_BYTE_L = (1u << 23);
 const uint BLOCKS_PER_STREAM = 320;
@@ -97,13 +97,15 @@ void encode_coeff(int signed_k, uint sign_bias, inout RansEncoder enc)
 	uint k = abs(signed_k);
 
 	if (k >= ESCAPE_LIMIT) {
-		// ... boring stuff here
-		RansEncPut(enc.rans, enc.rans_offset, k, 1, prob_bits - 1);
+		// Put the coefficient as a 1/(2^12) symbol _before_
+		// the 255 coefficient, since the decoder will read the
+		// 255 coefficient first.
+		RansEncPut(enc.rans, enc.rans_offset, k, 1, prob_bits);
 		k = ESCAPE_LIMIT;
 	}
 
-	uvec2 sym = ransdist[enc.lut_base + (k + NUM_SYMS - 1) & (NUM_SYMS - 1)];
-	RansEncPut(enc.rans, enc.rans_offset, sym.x, sym.y, prob_bits);
+	uvec2 sym = ransdist[enc.lut_base + ((k - 1) & (NUM_SYMS - 1))];
+	RansEncPut(enc.rans, enc.rans_offset, sym.x, sym.y, prob_bits + 1);
 	
 	if (signed_k < 0) {
 		enc.rans += sign_bias;
@@ -125,7 +127,7 @@ void encode_9_7(uint streamgroup_num, uint coeff_row, layout(r16ui) restrict rea
 	uint sign_bias1 = ransdist[enc1.lut_base + 255].x + ransdist[enc1.lut_base + 255].y;
 	uint sign_bias2 = ransdist[enc2.lut_base + 255].x + ransdist[enc2.lut_base + 255].y;
 
-	for (uint subblock_idx = BLOCKS_PER_STREAM; subblock_idx --> 0; ) {
+	for (uint subblock_idx = 0; subblock_idx < BLOCKS_PER_STREAM; ++subblock_idx) {
 		// TODO: Use SSBOs instead of a texture?
 		uint x = (streamgroup_num * BLOCKS_PER_STREAM + subblock_idx) % 160;
 		uint y = (streamgroup_num * BLOCKS_PER_STREAM + subblock_idx) / 160;