X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=rans.shader;h=3e19c83eae83e4f31f6c0c47bd4f0bc4263a7cc5;hb=0f84eb2a5fa77dbc20f7a073f83bdb674ebb7b75;hp=925543ef7dfa284d0fddbce411a61444438dbc9f;hpb=dc3e3a2a88b2b1f0e177f613a8b355116e6020ff;p=narabu

diff --git a/rans.shader b/rans.shader
index 925543e..3e19c83 100644
--- a/rans.shader
+++ b/rans.shader
@@ -1,5 +1,4 @@
 #version 440
-#extension GL_NV_gpu_shader5 : enable
 
 layout(local_size_x = 1) in;
 
@@ -7,7 +6,7 @@ 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;
-const uint STREAM_BUF_SIZE = 1024;  // 1 kB per stream ought to be enough for everyone :-)
+const uint STREAM_BUF_SIZE = 256;  // In uint32s. 1 kB per stream ought to be enough for everyone :-)
 const uint NUM_SYMS = 256;
 const uint ESCAPE_LIMIT = NUM_SYMS - 1;
 
@@ -24,27 +23,31 @@ const uint luma_mapping[8] = {
 	MAPPING(3, 3, 3, 3, 3, 3, 3, 3),
 };
 
-layout(std430, binding = 9) buffer layoutName
+layout(std430, binding = 10) buffer outputBuf
 {
-	uint dist[4 * 256];
-	uvec2 ransdist[4 * 256];
+	uint rans_output[];
 };
 
-layout(std430, binding = 10) buffer outputBuf
+layout(std430, binding = 11) buffer outputBuf2
 {
-	uint8_t rans_output[];
+	uint rans_bytes_written[];
 };
 
-layout(std430, binding = 11) buffer outputBuf2
+layout(std140, binding = 13) uniform DistBlock
 {
-	uint rans_start_offset[];
+	uvec4 ransdist[4 * 256];
+	uint sign_biases[4];
 };
 
 struct RansEncoder {
-	uint stream_num;   // const
-	uint lut_base;     // const
+	uint stream_num;         // const
+	uint lut_base;           // const
+	uint sign_bias;          // const
+	uint rans_start_offset;  // const
 	uint rans_offset;
 	uint rans;
+	uint bit_buffer;
+	uint bytes_in_buffer;
 };
 
 layout(r16ui) uniform restrict readonly uimage2D dc_ac7_tex;
@@ -53,38 +56,76 @@ layout(r16ui) uniform restrict readonly uimage2D ac2_ac5_tex;
 layout(r8i) uniform restrict readonly iimage2D ac3_tex;
 layout(r8i) uniform restrict readonly iimage2D ac4_tex;
 
+void RansPutByte(uint x, inout RansEncoder enc)
+{
+	enc.bit_buffer = (enc.bit_buffer << 8) | x;
+	if (++enc.bytes_in_buffer == 4) {
+		rans_output[--enc.rans_offset] = enc.bit_buffer;
+		enc.bytes_in_buffer = 0;
+	}
+}
+
 void RansEncInit(uint streamgroup_num, uint coeff_row, uint coeff_col, uint dist_num, out RansEncoder enc)
 {
 	enc.stream_num = streamgroup_num * 64 + coeff_row * 8 + coeff_col;
 	enc.lut_base = dist_num * 256;
+	enc.sign_bias = sign_biases[dist_num];
 	enc.rans_offset = enc.stream_num * STREAM_BUF_SIZE + STREAM_BUF_SIZE;  // Starts at the end.
+	enc.rans_start_offset = enc.rans_offset;
 	enc.rans = RANS_BYTE_L;
+	enc.bit_buffer = 0;
+	enc.bytes_in_buffer = 0;
 }
 
-void RansEncRenorm(inout uint rans, inout uint rans_offset, uint freq, uint prob_bits)
+void RansEncRenorm(inout RansEncoder enc, uint freq, uint prob_bits)
 {
 	uint x_max = ((RANS_BYTE_L >> prob_bits) << 8) * freq; // this turns into a shift.
-	if (rans >= x_max) {
-		do {
-			rans_output[--rans_offset] = uint8_t(rans & 0xff);
-			rans >>= 8;
-		} while (rans >= x_max);
+	while (enc.rans >= x_max) {
+		RansPutByte(enc.rans & 0xffu, enc);
+		enc.rans >>= 8;
 	}
 }
 
-void RansEncPut(inout uint rans, inout uint rans_offset, uint start, uint freq, uint prob_bits)
+void RansEncPut(inout RansEncoder enc, uint start, uint freq, uint prob_bits)
 {
-	RansEncRenorm(rans, rans_offset, freq, prob_bits);
-	rans = ((rans / freq) << prob_bits) + (rans % freq) + start;
+	RansEncRenorm(enc, freq, prob_bits);
+	enc.rans = ((enc.rans / freq) << prob_bits) + (enc.rans % freq) + start;
 }
 
-void RansEncFlush(uint rans, inout uint rans_offset)
+void RansEncPutSymbol(inout RansEncoder enc, uvec4 sym)
 {
-	rans_offset -= 4;
-	rans_output[rans_offset + 0] = uint8_t(rans >> 0);
-	rans_output[rans_offset + 1] = uint8_t(rans >> 8);
-	rans_output[rans_offset + 2] = uint8_t(rans >> 16);
-	rans_output[rans_offset + 3] = uint8_t(rans >> 24);
+	uint x_max = sym.x;
+	uint rcp_freq = sym.y;
+	uint bias = sym.z;
+	uint rcp_shift = (sym.w & 0xffffu);
+	uint cmpl_freq = (sym.w >> 16);
+
+	// renormalize
+	while (enc.rans >= x_max) {
+		RansPutByte(enc.rans & 0xffu, enc);
+		enc.rans >>= 8;
+	}
+
+	uint q, unused;
+	umulExtended(enc.rans, rcp_freq, q, unused);
+	enc.rans += bias + (q >> rcp_shift) * cmpl_freq;
+}
+
+uint RansEncFlush(inout RansEncoder enc)
+{
+	RansPutByte(enc.rans >> 24, enc);
+	RansPutByte(enc.rans >> 16, enc);
+	RansPutByte(enc.rans >> 8, enc);
+	RansPutByte(enc.rans >> 0, enc);
+
+	uint num_bytes_written = (enc.rans_start_offset - enc.rans_offset) * 4 + enc.bytes_in_buffer;
+
+	// Make sure there's nothing left in the buffer.
+	RansPutByte(0, enc);
+	RansPutByte(0, enc);
+	RansPutByte(0, enc);
+
+	return num_bytes_written;
 }
 
 int sign_extend(uint coeff, uint bits)
@@ -92,7 +133,7 @@ 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)
+void encode_coeff(int signed_k, inout RansEncoder enc)
 {
 	uint k = abs(signed_k);
 
@@ -100,22 +141,22 @@ void encode_coeff(int signed_k, uint sign_bias, inout RansEncoder enc)
 		// 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);
+		RansEncPut(enc, 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 + 1);
+	uvec4 sym = ransdist[enc.lut_base + ((k - 1) & (NUM_SYMS - 1))];
+	RansEncPutSymbol(enc, sym);
 	
 	if (signed_k < 0) {
-		enc.rans += sign_bias;
+		enc.rans += enc.sign_bias;
 	}
 }
 
 void encode_end(inout RansEncoder enc)
 {
-	RansEncFlush(enc.rans, enc.rans_offset);
-	rans_start_offset[enc.stream_num] = enc.rans_offset;
+	uint bytes_written = RansEncFlush(enc);
+	rans_bytes_written[enc.stream_num] = bytes_written;
 }
 
 void encode_9_7(uint streamgroup_num, uint coeff_row, layout(r16ui) restrict readonly uimage2D tex, uint col1, uint col2, uint dist1, uint dist2)
@@ -124,17 +165,14 @@ 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 = 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(sign_extend(f & 0x1ffu, 9), sign_bias1, enc1);
-		encode_coeff(sign_extend(f >> 9, 7), sign_bias2, enc2);
+		encode_coeff(sign_extend(f & 0x1ffu, 9), enc1);
+		encode_coeff(sign_extend(f >> 9, 7), enc2);
 	}
 
 	encode_end(enc1);
@@ -146,15 +184,13 @@ 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; ) {
+	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;
 		int f = imageLoad(tex, ivec2(x, y * 8 + coeff_row)).x;
 
-		encode_coeff(f, sign_bias, enc);
+		encode_coeff(f, enc);
 	}
 
 	encode_end(enc);