X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=rans.shader;h=925543ef7dfa284d0fddbce411a61444438dbc9f;hb=dc3e3a2a88b2b1f0e177f613a8b355116e6020ff;hp=a7c6f8cce21abee5bca5cfbff55739ff0e323ef3;hpb=5e1d27014149311318e97b8e04a6e05ec858e57c;p=narabu

diff --git a/rans.shader b/rans.shader
index a7c6f8c..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;
@@ -87,23 +87,29 @@ void RansEncFlush(uint rans, inout uint rans_offset)
 	rans_output[rans_offset + 3] = uint8_t(rans >> 24);
 }
 
-void encode_coeff(uint coeff, uint bits, inout RansEncoder enc)
+int sign_extend(uint coeff, uint bits)
+{
+	return int(coeff << (32 - bits)) >> (32 - bits);
+}
+
+void encode_coeff(int signed_k, uint sign_bias, inout RansEncoder enc)
 {
-	// Sign-extend to recover the coefficient.
-	// FIXME: not needed for the bits == 8 case!
-	int signed_k = int(coeff << (32 - bits)) >> (32 - bits);
 	uint k = abs(signed_k);
 
 	if (k >= ESCAPE_LIMIT) {
-		// ... boring stuff here
+		// 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 - 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);
 	
-	// fix some bias stuff here
+	if (signed_k < 0) {
+		enc.rans += sign_bias;
+	}
 }
 
 void encode_end(inout RansEncoder enc)
@@ -118,14 +124,17 @@ void encode_9_7(uint streamgroup_num, uint coeff_row, layout(r16ui) restrict rea
 	RansEncInit(streamgroup_num, coeff_row, col1, dist1, enc1);
 	RansEncInit(streamgroup_num, coeff_row, col2, dist2, enc2);
 
-	for (uint subblock_idx = BLOCKS_PER_STREAM; subblock_idx --> 0; ) {
+	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 = 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;
 		uint f = imageLoad(tex, ivec2(x, y * 8 + coeff_row)).x;
 
-		encode_coeff(f & 0x1ffu, 9, enc1);
-		encode_coeff(f >> 9, 7, enc2);
+		encode_coeff(sign_extend(f & 0x1ffu, 9), sign_bias1, enc1);
+		encode_coeff(sign_extend(f >> 9, 7), sign_bias2, enc2);
 	}
 
 	encode_end(enc1);
@@ -137,13 +146,15 @@ void encode_8(uint streamgroup_num, uint coeff_row, layout(r8i) restrict readonl
 	RansEncoder enc;
 	RansEncInit(streamgroup_num, coeff_row, col, dist, enc);
 
+	uint sign_bias = ransdist[enc.lut_base + 255].x + ransdist[enc.lut_base + 255].y;
+
 	for (uint subblock_idx = BLOCKS_PER_STREAM; subblock_idx --> 0; ) {
 		// 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;
 		int f = imageLoad(tex, ivec2(x, y * 8 + coeff_row)).x;
 
-		encode_coeff(f, 8, enc);
+		encode_coeff(f, sign_bias, enc);
 	}
 
 	encode_end(enc);