X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=rans.shader;h=47d0c8b7f9ad2a5c4cbac0ee959b0efd7adbe5ca;hb=c6928acb2405c30fdaf1a5d9efeb2902d7c544aa;hp=965c6744d8e1151fc0f05eae9ec1dd5af660f60c;hpb=1982179eccafd16269939ae5a5d4a467f68b26fd;p=narabu

diff --git a/rans.shader b/rans.shader
index 965c674..47d0c8b 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;
 
@@ -7,9 +6,10 @@ 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;
+const uint WIDTH_BLOCKS = 160;  // TODO: send in as a uniform.
 
 #define MAPPING(s0, s1, s2, s3, s4, s5, s6, s7) ((s0) | (s1 << 2) | (s2 << 4) | (s3 << 6) | (s4 << 8) | (s5 << 10) | (s6 << 12) | (s7 << 14))
 
@@ -26,12 +26,12 @@ const uint luma_mapping[8] = {
 
 layout(std430, binding = 10) buffer outputBuf
 {
-	uint8_t rans_output[];
+	uint rans_output[];
 };
 
 layout(std430, binding = 11) buffer outputBuf2
 {
-	uint rans_start_offset[];
+	uint rans_bytes_written[];
 };
 
 layout(std140, binding = 13) uniform DistBlock
@@ -41,11 +41,14 @@ layout(std140, binding = 13) uniform DistBlock
 };
 
 struct RansEncoder {
-	uint stream_num;   // const
-	uint lut_base;     // const
-	uint sign_bias;    // 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;
@@ -54,31 +57,43 @@ 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.
-	while (rans >= x_max) {
-		rans_output[--rans_offset] = uint8_t(rans & 0xffu);
-		rans >>= 8;
+	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 RansEncPutSymbol(inout uint rans, inout uint rans_offset, uvec4 sym)
+void RansEncPutSymbol(inout RansEncoder enc, uvec4 sym)
 {
 	uint x_max = sym.x;
 	uint rcp_freq = sym.y;
@@ -87,23 +102,31 @@ void RansEncPutSymbol(inout uint rans, inout uint rans_offset, uvec4 sym)
 	uint cmpl_freq = (sym.w >> 16);
 
 	// renormalize
-	while (rans >= x_max) {
-		rans_output[--rans_offset] = uint8_t(rans & 0xffu);
-		rans >>= 8;
+	while (enc.rans >= x_max) {
+		RansPutByte(enc.rans & 0xffu, enc);
+		enc.rans >>= 8;
 	}
 
 	uint q, unused;
-	umulExtended(rans, rcp_freq, q, unused);
-	rans += bias + (q >> rcp_shift) * cmpl_freq;
+	umulExtended(enc.rans, rcp_freq, q, unused);
+	enc.rans += bias + (q >> rcp_shift) * cmpl_freq;
 }
 
-void RansEncFlush(uint rans, inout uint rans_offset)
+uint RansEncFlush(inout RansEncoder enc)
 {
-	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);
+	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)
@@ -119,12 +142,12 @@ void encode_coeff(int signed_k, 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;
 	}
 
 	uvec4 sym = ransdist[enc.lut_base + ((k - 1) & (NUM_SYMS - 1))];
-	RansEncPutSymbol(enc.rans, enc.rans_offset, sym);
+	RansEncPutSymbol(enc, sym);
 	
 	if (signed_k < 0) {
 		enc.rans += enc.sign_bias;
@@ -133,8 +156,8 @@ void encode_coeff(int signed_k, inout RansEncoder enc)
 
 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)
@@ -145,8 +168,8 @@ void encode_9_7(uint streamgroup_num, uint coeff_row, layout(r16ui) restrict rea
 
 	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 x = (streamgroup_num * BLOCKS_PER_STREAM + subblock_idx) % WIDTH_BLOCKS;
+		uint y = (streamgroup_num * BLOCKS_PER_STREAM + subblock_idx) / WIDTH_BLOCKS;
 		uint f = imageLoad(tex, ivec2(x, y * 8 + coeff_row)).x;
 
 		encode_coeff(sign_extend(f & 0x1ffu, 9), enc1);
@@ -164,8 +187,8 @@ void encode_8(uint streamgroup_num, uint coeff_row, layout(r8i) restrict readonl
 
 	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 x = (streamgroup_num * BLOCKS_PER_STREAM + subblock_idx) % WIDTH_BLOCKS;
+		uint y = (streamgroup_num * BLOCKS_PER_STREAM + subblock_idx) / WIDTH_BLOCKS;
 		int f = imageLoad(tex, ivec2(x, y * 8 + coeff_row)).x;
 
 		encode_coeff(f, enc);