X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=decoder.shader;h=616c3e2dfd0896e5cd035bacc42ec0720c5174a1;hb=5e1d27014149311318e97b8e04a6e05ec858e57c;hp=012cbe3bbac0ec13bb40b0cf42bd9ecc4bdcadf0;hpb=3fb87c6b953be3382cd216c74ff6aa025c8eaa2a;p=narabu

diff --git a/decoder.shader b/decoder.shader
index 012cbe3..616c3e2 100644
--- a/decoder.shader
+++ b/decoder.shader
@@ -9,8 +9,7 @@ layout(local_size_x = 64*PARALLEL_SLICES) in;
 layout(r8ui) uniform restrict readonly uimage2D cum2sym_tex;
 layout(rg16ui) uniform restrict readonly uimage2D dsyms_tex;
 layout(r8) uniform restrict writeonly image2D out_tex;
-layout(r32i) uniform restrict writeonly iimage2D coeff_tex;
-layout(r32i) uniform restrict writeonly iimage2D coeff2_tex;
+layout(r16i) uniform restrict writeonly iimage2D coeff_tex;
 uniform int num_blocks;
 
 const uint prob_bits = 12;
@@ -72,50 +71,40 @@ layout(std430, binding = 0) buffer whatever3
 };
 uniform uint sign_bias_per_model[16];
 
-struct myuint64 {
-	uint high, low;
-};
+const uint RANS_BYTE_L = (1u << 23);  // lower bound of our normalization interval
 
-const uint RANS64_L = (1u << 31);  // lower bound of our normalization interval
+uint get_rans_byte(uint offset)
+{
+	// We assume little endian.
+	return bitfieldExtract(data_SSBO[offset >> 2], 8 * int(offset & 3u), 8);
+}
 
-myuint64 RansDecInit(inout uint offset)
+uint RansDecInit(inout uint offset)
 {
-	myuint64 x;
-	x.low  = data_SSBO[offset++];
-	x.high = data_SSBO[offset++];
+	uint x;
+
+	x  = get_rans_byte(offset);
+	x |= get_rans_byte(offset + 1) << 8;
+	x |= get_rans_byte(offset + 2) << 16;
+	x |= get_rans_byte(offset + 3) << 24;
+	offset += 4;
+
 	return x;
 }
 
-uint RansDecGet(myuint64 r, uint scale_bits)
+uint RansDecGet(uint r, uint scale_bits)
 {
-	return r.low & ((1u << scale_bits) - 1);
+	return r & ((1u << scale_bits) - 1);
 }
 
-void RansDecAdvance(inout myuint64 rans, inout uint offset, const uint start, const uint freq, uint prob_bits)
+void RansDecAdvance(inout uint rans, inout uint offset, const uint start, const uint freq, uint prob_bits)
 {
 	const uint mask = (1u << prob_bits) - 1;
-	const uint recovered_lowbits = (rans.low & mask) - start;
-
-	// rans >>= prob_bits;
-	rans.low = (rans.low >> prob_bits) | ((rans.high & mask) << (32 - prob_bits));
-	rans.high >>= prob_bits;
-
-	// rans *= freq;
-	uint h1, l1, h2, l2;
-	umulExtended(rans.low, freq, h1, l1);
-	umulExtended(rans.high, freq, h2, l2);
-	rans.low = l1;
-	rans.high = l2 + h1;
-
-	// rans += recovered_lowbits;
-	uint carry;
-	rans.low = uaddCarry(rans.low, recovered_lowbits, carry);
-	rans.high += carry;
-
+	rans = freq * (rans >> prob_bits) + (rans & mask) - start;
+	
 	// renormalize
-	if (rans.high == 0 && rans.low < RANS64_L) {
-		rans.high = rans.low;
-		rans.low = data_SSBO[offset++];
+	while (rans < RANS_BYTE_L) {
+		rans = (rans << 8) | get_rans_byte(offset++);
 	}
 }
 
@@ -173,8 +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;
@@ -232,8 +221,8 @@ void main()
 	const uint sign_bias = sign_bias_per_model[model_num];
 
 	// Initialize rANS decoder.
-	uint offset = streams[stream_num].src_offset >> 2;
-	myuint64 rans = RansDecInit(offset);
+	uint offset = streams[stream_num].src_offset;
+	uint rans = RansDecInit(offset);
 
 	float q = (coeff_num == 0) ? 1.0 : (quant_matrix[coeff_num] * quant_scalefac / 128.0 / sqrt(2.0));  // FIXME: fold
 	q *= (1.0 / 255.0);
@@ -252,7 +241,7 @@ void main()
 			bool sign = false;
 			if (bottom_bits >= sign_bias) {
 				bottom_bits -= sign_bias;
-				rans.low -= sign_bias;
+				rans -= sign_bias;
 				sign = true;
 			}
 			int k = int(cum2sym(bottom_bits, model_num));  // Can go out-of-bounds; that will return zero.
@@ -266,19 +255,17 @@ void main()
 			if (sign) {
 				k = -k;
 			}
-#if 0
-			if (coeff_num == 0) {
-				//imageStore(coeff_tex, ivec2((block_row * 40 + block_idx) * 8 + subblock_idx, 0), ivec4(k, 0,0,0));
-				imageStore(coeff_tex, ivec2((block_row * 40 + block_idx) * 8 + subblock_idx, 0), ivec4(rans.low, 0,0,0));
-				imageStore(coeff2_tex, ivec2((block_row * 40 + block_idx) * 8 + subblock_idx, 0), ivec4(rans.high, 0,0,0));
-			}
-#endif
 
 			if (coeff_num == 0) {
 				k += last_k;
 				last_k = k;
 			}
 
+#if 0
+			uint y = block_row * 16 + block_y * 8 + local_y;
+			uint x = block_x * 64 + subblock_idx * 8 + local_x;
+			imageStore(coeff_tex, ivec2(x, y), ivec4(k, 0,0,0));
+#endif
 
 			temp[slice_num * 64 * 8 + subblock_idx * 64 + coeff_num] = k * q;
 			//temp[subblock_idx * 64 + 8 * y + x] = (2 * k * w * 4) / 32;  // 100% matching unquant