X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=decoder.shader;h=012cbe3bbac0ec13bb40b0cf42bd9ecc4bdcadf0;hb=3fb87c6b953be3382cd216c74ff6aa025c8eaa2a;hp=cab6f8cbecbb508421b01740a208b431a8387136;hpb=2161c821c9c65ebae24eff9c2e4809c11c67cd02;p=narabu

diff --git a/decoder.shader b/decoder.shader
index cab6f8c..012cbe3 100644
--- a/decoder.shader
+++ b/decoder.shader
@@ -9,13 +9,15 @@ 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(r16i) uniform restrict writeonly iimage2D coeff_tex;
+layout(r32i) uniform restrict writeonly iimage2D coeff_tex;
+layout(r32i) uniform restrict writeonly iimage2D coeff2_tex;
 uniform int num_blocks;
 
 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;
@@ -70,40 +72,50 @@ layout(std430, binding = 0) buffer whatever3
 };
 uniform uint sign_bias_per_model[16];
 
-const uint RANS_BYTE_L = (1u << 23);  // lower bound of our normalization interval
+struct myuint64 {
+	uint high, low;
+};
 
-uint get_rans_byte(uint offset)
-{
-	// We assume little endian.
-	return bitfieldExtract(data_SSBO[offset >> 2], 8 * int(offset & 3u), 8);
-}
+const uint RANS64_L = (1u << 31);  // lower bound of our normalization interval
 
-uint RansDecInit(inout uint offset)
+myuint64 RansDecInit(inout uint 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;
-
+	myuint64 x;
+	x.low  = data_SSBO[offset++];
+	x.high = data_SSBO[offset++];
 	return x;
 }
 
-uint RansDecGet(uint r, uint scale_bits)
+uint RansDecGet(myuint64 r, uint scale_bits)
 {
-	return r & ((1u << scale_bits) - 1);
+	return r.low & ((1u << scale_bits) - 1);
 }
 
-void RansDecAdvance(inout uint rans, inout uint offset, const uint start, const uint freq, uint prob_bits)
+void RansDecAdvance(inout myuint64 rans, inout uint offset, const uint start, const uint freq, uint prob_bits)
 {
 	const uint mask = (1u << prob_bits) - 1;
-	rans = freq * (rans >> prob_bits) + (rans & mask) - start;
-	
+	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;
+
 	// renormalize
-	while (rans < RANS_BYTE_L) {
-		rans = (rans << 8) | get_rans_byte(offset++);
+	if (rans.high == 0 && rans.low < RANS64_L) {
+		rans.high = rans.low;
+		rans.low = data_SSBO[offset++];
 	}
 }
 
@@ -203,6 +215,8 @@ 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;
@@ -218,17 +232,17 @@ void main()
 	const uint sign_bias = sign_bias_per_model[model_num];
 
 	// Initialize rANS decoder.
-	uint offset = streams[stream_num].src_offset;
-	uint rans = RansDecInit(offset);
+	uint offset = streams[stream_num].src_offset >> 2;
+	myuint64 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);
 	//int w = (coeff_num == 0) ? 32 : int(quant_matrix[coeff_num]);
-	int last_k = 0;
+	int last_k = 128;
 
 	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).
@@ -238,7 +252,7 @@ void main()
 			bool sign = false;
 			if (bottom_bits >= sign_bias) {
 				bottom_bits -= sign_bias;
-				rans -= sign_bias;
+				rans.low -= sign_bias;
 				sign = true;
 			}
 			int k = int(cum2sym(bottom_bits, model_num));  // Can go out-of-bounds; that will return zero.
@@ -252,17 +266,19 @@ 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
@@ -306,8 +322,8 @@ 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;