From: Steinar H. Gunderson <sgunderson@bigfoot.com>
Date: Mon, 16 Oct 2017 20:16:41 +0000 (+0200)
Subject: Implement sign_bias in the rANS GPU encoder. Still not working properly.
X-Git-Url: https://git.sesse.net/?p=narabu;a=commitdiff_plain;h=86c0bf043378f95d1431990f378814af9fcfe6fd

Implement sign_bias in the rANS GPU encoder. Still not working properly.
---

diff --git a/coded.dat b/coded.dat
index a811389..3980407 100644
Binary files a/coded.dat and b/coded.dat differ
diff --git a/rans.shader b/rans.shader
index a7c6f8c..36feb65 100644
--- a/rans.shader
+++ b/rans.shader
@@ -87,7 +87,7 @@ 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)
+void encode_coeff(uint coeff, uint bits, uint sign_bias, inout RansEncoder enc)
 {
 	// Sign-extend to recover the coefficient.
 	// FIXME: not needed for the bits == 8 case!
@@ -100,10 +100,12 @@ void encode_coeff(uint coeff, uint bits, inout RansEncoder enc)
 		k = ESCAPE_LIMIT;
 	}
 
-	uvec2 sym = ransdist[enc.lut_base + (k - 1) & (NUM_SYMS - 1)];
+	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);
 	
-	// fix some bias stuff here
+	if (signed_k < 0) {
+		enc.rans += sign_bias;
+	}
 }
 
 void encode_end(inout RansEncoder enc)
@@ -118,14 +120,17 @@ 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);
 
+	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; ) {
 		// 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(f & 0x1ffu, 9, sign_bias1, enc1);
+		encode_coeff(f >> 9, 7, sign_bias2, enc2);
 	}
 
 	encode_end(enc1);
@@ -137,13 +142,15 @@ 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, 8, sign_bias, enc);
 	}
 
 	encode_end(enc);