X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=decoder.shader;h=09027915633af93b69a883039836bdecc16f8df2;hb=34138d5b1a1302a7a1050fd46d2fb95c0186140a;hp=6d54e4dc5671173f540a1574666e51870a14b1c2;hpb=28409aed1a0cbf8d2e8d9d157d08c3f6d9a3f51a;p=narabu

diff --git a/decoder.shader b/decoder.shader
index 6d54e4d..0902791 100644
--- a/decoder.shader
+++ b/decoder.shader
@@ -1,12 +1,15 @@
-#version 430
+#version 440
 #extension GL_ARB_shader_clock : enable
 
+#define PARALLEL_SLICES 1
+
 #define ENABLE_TIMING 0
 
-layout(local_size_x = 8, local_size_y = 8) in;
+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;
 
 const uint prob_bits = 12;
 const uint prob_scale = 1 << prob_bits;
@@ -37,6 +40,16 @@ const uint ff_zigzag_direct[64] = {
     58, 59, 52, 45, 38, 31, 39, 46,
     53, 60, 61, 54, 47, 55, 62, 63
 };
+const uint stream_mapping[64] = {
+	0, 0, 1, 1, 2, 2, 3, 3,
+	0, 0, 1, 2, 2, 2, 3, 3,
+	1, 1, 2, 2, 2, 3, 3, 3,
+	1, 1, 2, 2, 2, 3, 3, 3,
+	1, 2, 2, 2, 2, 3, 3, 3,
+	2, 2, 2, 2, 3, 3, 3, 3,
+	2, 2, 3, 3, 3, 3, 3, 3,
+	3, 3, 3, 3, 3, 3, 3, 3,
+};
 
 layout(std430, binding = 9) buffer layoutName
 {
@@ -48,32 +61,23 @@ layout(std430, binding = 10) buffer layoutName2
 };
 
 struct CoeffStream {
-	uint src_offset, src_len, sign_offset, sign_len, extra_bits;
+	uint src_offset, src_len;
 };
 layout(std430, binding = 0) buffer whatever3
 {
 	CoeffStream streams[];
 };
-
-uniform uint src_offset, src_len, sign_offset, sign_len, extra_bits;
+uniform uint sign_bias_per_model[16];
 
 const uint RANS_BYTE_L = (1u << 23);  // lower bound of our normalization interval
 
-uint last_offset = -1, ransbuf;
-
 uint get_rans_byte(uint offset)
 {
-	if (last_offset != (offset >> 2)) {
-		last_offset = offset >> 2;
-		ransbuf = data_SSBO[offset >> 2];
-	}
-	return bitfieldExtract(ransbuf, 8 * int(offset & 3u), 8);
-
 	// We assume little endian.
-//	return bitfieldExtract(data_SSBO[offset >> 2], 8 * int(offset & 3u), 8);
+	return bitfieldExtract(data_SSBO[offset >> 2], 8 * int(offset & 3u), 8);
 }
 
-void RansDecInit(out uint r, inout uint offset)
+uint RansDecInit(inout uint offset)
 {
 	uint x;
 
@@ -83,7 +87,7 @@ void RansDecInit(out uint r, inout uint offset)
 	x |= get_rans_byte(offset + 3) << 24;
 	offset += 4;
 
-	r = x;
+	return x;
 }
 
 uint RansDecGet(uint r, uint scale_bits)
@@ -163,7 +167,7 @@ void idct_1d(inout float y0, inout float y1, inout float y2, inout float y3, ino
 	y7 = p6_0 - p6_7;
 }
 
-shared float temp[64 * 8];
+shared float temp[64 * 8 * PARALLEL_SLICES];
 
 void pick_timer(inout uvec2 start, inout uvec2 t)
 {
@@ -194,28 +198,28 @@ void main()
 	}
 	uvec2 start = clock2x32ARB();
 #else
-	uvec2 start;
+	uvec2 start = uvec2(0, 0);
+	local_timing[0] = start;
 #endif
 
+	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;
+
 	const uint num_blocks = 720 / 16;  // FIXME: make a uniform
-	const uint thread_num = gl_LocalInvocationID.y * 8 + gl_LocalInvocationID.x;
+	const uint slice_num = local_z;
+	const uint thread_num = local_y * 8 + local_x;
 
-	const uint block_row = gl_WorkGroupID.y;
+	const uint block_row = gl_WorkGroupID.y * PARALLEL_SLICES + slice_num;
 	//const uint coeff_num = ff_zigzag_direct[thread_num];
 	const uint coeff_num = thread_num;
 	const uint stream_num = coeff_num * num_blocks + block_row;
-	//const uint stream_num = block_row * num_blocks + coeff_num;  // HACK
-	const uint model_num = min((coeff_num % 8) + (coeff_num / 8), 7);
+	const uint model_num = stream_mapping[coeff_num];
+	const uint sign_bias = sign_bias_per_model[model_num];
 
 	// Initialize rANS decoder.
 	uint offset = streams[stream_num].src_offset;
-	uint rans;
-	RansDecInit(rans, offset);
-
-	// Initialize sign bit decoder. TODO: this ought to be 32-bit-aligned instead!
-	uint soffset = streams[stream_num].sign_offset;
-	uint sign_buf = get_rans_byte(soffset++) >> streams[stream_num].extra_bits;
-	uint sign_bits_left = 8 - streams[stream_num].extra_bits;
+	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);
@@ -234,22 +238,23 @@ void main()
 		// rANS decode one coefficient across eight blocks (so 64x8 coefficients).
 		for (uint subblock_idx = 8; subblock_idx --> 0; ) {
 			// Read a symbol.
-			int k = int(cum2sym(RansDecGet(rans, prob_bits), model_num));
+			uint bottom_bits = RansDecGet(rans, prob_bits + 1);
+			bool sign = false;
+			if (bottom_bits >= sign_bias) {
+				bottom_bits -= sign_bias;
+				rans -= sign_bias;
+				sign = true;
+			}
+			int k = int(cum2sym(bottom_bits, model_num));  // Can go out-of-bounds; that will return zero.
 			uvec2 sym = get_dsym(k, model_num);
-			RansDecAdvance(rans, offset, sym.x, sym.y, prob_bits);
+			RansDecAdvance(rans, offset, sym.x, sym.y, prob_bits + 1);
 
 			if (k == ESCAPE_LIMIT) {
 				k = int(RansDecGet(rans, prob_bits));
 				RansDecAdvance(rans, offset, k, 1, prob_bits);
 			}
-			if (k != 0) {
-				if (sign_bits_left == 0) {
-					sign_buf = get_rans_byte(soffset++);
-					sign_bits_left = 8;
-				}
-				if ((sign_buf & 1u) == 1u) k = -k;
-				--sign_bits_left;
-				sign_buf >>= 1;
+			if (sign) {
+				k = -k;
 			}
 
 			if (coeff_num == 0) {
@@ -257,7 +262,13 @@ void main()
 				last_k = k;
 			}
 
-			temp[subblock_idx * 64 + coeff_num] = k * q;
+#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
 		}
 
@@ -269,14 +280,14 @@ void main()
 		pick_timer(start, local_timing[3]);
 
 		// Horizontal DCT one row (so 64 rows).
-		idct_1d(temp[thread_num * 8 + 0],
-		        temp[thread_num * 8 + 1],
-		        temp[thread_num * 8 + 2],
-		        temp[thread_num * 8 + 3],
-		        temp[thread_num * 8 + 4],
-		        temp[thread_num * 8 + 5],
-		        temp[thread_num * 8 + 6],
-		        temp[thread_num * 8 + 7]);
+		idct_1d(temp[slice_num * 64 * 8 + thread_num * 8 + 0],
+		        temp[slice_num * 64 * 8 + thread_num * 8 + 1],
+		        temp[slice_num * 64 * 8 + thread_num * 8 + 2],
+		        temp[slice_num * 64 * 8 + thread_num * 8 + 3],
+		        temp[slice_num * 64 * 8 + thread_num * 8 + 4],
+		        temp[slice_num * 64 * 8 + thread_num * 8 + 5],
+		        temp[slice_num * 64 * 8 + thread_num * 8 + 6],
+		        temp[slice_num * 64 * 8 + thread_num * 8 + 7]);
 
 		pick_timer(start, local_timing[4]);
 
@@ -286,7 +297,7 @@ void main()
 		pick_timer(start, local_timing[5]);
 
 		// Vertical DCT one row (so 64 columns).
-		uint row_offset = gl_LocalInvocationID.y * 64 + gl_LocalInvocationID.x;
+		uint row_offset = local_z * 64 * 8 + local_y * 64 + local_x;
 		idct_1d(temp[row_offset + 0 * 8],
 		        temp[row_offset + 1 * 8],
 		        temp[row_offset + 2 * 8],
@@ -299,7 +310,7 @@ void main()
 		pick_timer(start, local_timing[6]);
 
 		uint y = block_row * 16 + block_y * 8;
-		uint x = block_x * 64 + gl_LocalInvocationID.y * 8 + gl_LocalInvocationID.x;
+		uint x = block_x * 64 + local_y * 8 + local_x;
 		for (uint yl = 0; yl < 8; ++yl) {
 			imageStore(out_tex, ivec2(x, yl + y), vec4(temp[row_offset + yl * 8], 0.0, 0.0, 1.0));
 		}