2 #extension GL_NV_gpu_shader5 : enable
4 layout(local_size_x = 1) in;
6 const uint prob_bits = 12;
7 const uint prob_scale = 1 << prob_bits;
8 const uint RANS_BYTE_L = (1u << 23);
9 const uint BLOCKS_PER_STREAM = 320;
10 const uint STREAM_BUF_SIZE = 1024; // 1 kB per stream ought to be enough for everyone :-)
11 const uint NUM_SYMS = 256;
12 const uint ESCAPE_LIMIT = NUM_SYMS - 1;
14 #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))
16 const uint luma_mapping[8] = {
17 MAPPING(0, 0, 1, 1, 2, 2, 3, 3),
18 MAPPING(0, 0, 1, 2, 2, 2, 3, 3),
19 MAPPING(1, 1, 2, 2, 2, 3, 3, 3),
20 MAPPING(1, 1, 2, 2, 2, 3, 3, 3),
21 MAPPING(1, 2, 2, 2, 2, 3, 3, 3),
22 MAPPING(2, 2, 2, 2, 3, 3, 3, 3),
23 MAPPING(2, 2, 3, 3, 3, 3, 3, 3),
24 MAPPING(3, 3, 3, 3, 3, 3, 3, 3),
27 layout(std430, binding = 10) buffer outputBuf
29 uint8_t rans_output[];
32 layout(std430, binding = 11) buffer outputBuf2
34 uint rans_start_offset[];
37 layout(std140, binding = 13) uniform DistBlock
39 uvec4 ransdist[4 * 256];
44 uint stream_num; // const
45 uint lut_base; // const
46 uint sign_bias; // const
51 layout(r16ui) uniform restrict readonly uimage2D dc_ac7_tex;
52 layout(r16ui) uniform restrict readonly uimage2D ac1_ac6_tex;
53 layout(r16ui) uniform restrict readonly uimage2D ac2_ac5_tex;
54 layout(r8i) uniform restrict readonly iimage2D ac3_tex;
55 layout(r8i) uniform restrict readonly iimage2D ac4_tex;
57 void RansEncInit(uint streamgroup_num, uint coeff_row, uint coeff_col, uint dist_num, out RansEncoder enc)
59 enc.stream_num = streamgroup_num * 64 + coeff_row * 8 + coeff_col;
60 enc.lut_base = dist_num * 256;
61 enc.sign_bias = sign_biases[dist_num];
62 enc.rans_offset = enc.stream_num * STREAM_BUF_SIZE + STREAM_BUF_SIZE; // Starts at the end.
63 enc.rans = RANS_BYTE_L;
66 void RansEncRenorm(inout uint rans, inout uint rans_offset, uint freq, uint prob_bits)
68 uint x_max = ((RANS_BYTE_L >> prob_bits) << 8) * freq; // this turns into a shift.
69 while (rans >= x_max) {
70 rans_output[--rans_offset] = uint8_t(rans & 0xffu);
75 void RansEncPut(inout uint rans, inout uint rans_offset, uint start, uint freq, uint prob_bits)
77 RansEncRenorm(rans, rans_offset, freq, prob_bits);
78 rans = ((rans / freq) << prob_bits) + (rans % freq) + start;
81 void RansEncPutSymbol(inout uint rans, inout uint rans_offset, uvec4 sym)
84 uint rcp_freq = sym.y;
86 uint rcp_shift = (sym.w & 0xffffu);
87 uint cmpl_freq = (sym.w >> 16);
92 rans_output[--rans_offset] = uint8_t(rans & 0xffu);
94 } while (rans >= x_max);
98 umulExtended(rans, rcp_freq, q, unused);
99 rans += bias + (q >> rcp_shift) * cmpl_freq;
102 void RansEncFlush(uint rans, inout uint rans_offset)
105 rans_output[rans_offset + 0] = uint8_t(rans >> 0);
106 rans_output[rans_offset + 1] = uint8_t(rans >> 8);
107 rans_output[rans_offset + 2] = uint8_t(rans >> 16);
108 rans_output[rans_offset + 3] = uint8_t(rans >> 24);
111 int sign_extend(uint coeff, uint bits)
113 return int(coeff << (32 - bits)) >> (32 - bits);
116 void encode_coeff(int signed_k, inout RansEncoder enc)
118 uint k = abs(signed_k);
120 if (k >= ESCAPE_LIMIT) {
121 // Put the coefficient as a 1/(2^12) symbol _before_
122 // the 255 coefficient, since the decoder will read the
123 // 255 coefficient first.
124 RansEncPut(enc.rans, enc.rans_offset, k, 1, prob_bits);
128 uvec4 sym = ransdist[enc.lut_base + ((k - 1) & (NUM_SYMS - 1))];
129 RansEncPutSymbol(enc.rans, enc.rans_offset, sym);
132 enc.rans += enc.sign_bias;
136 void encode_end(inout RansEncoder enc)
138 RansEncFlush(enc.rans, enc.rans_offset);
139 rans_start_offset[enc.stream_num] = enc.rans_offset;
142 void encode_9_7(uint streamgroup_num, uint coeff_row, layout(r16ui) restrict readonly uimage2D tex, uint col1, uint col2, uint dist1, uint dist2)
144 RansEncoder enc1, enc2;
145 RansEncInit(streamgroup_num, coeff_row, col1, dist1, enc1);
146 RansEncInit(streamgroup_num, coeff_row, col2, dist2, enc2);
148 for (uint subblock_idx = 0; subblock_idx < BLOCKS_PER_STREAM; ++subblock_idx) {
149 // TODO: Use SSBOs instead of a texture?
150 uint x = (streamgroup_num * BLOCKS_PER_STREAM + subblock_idx) % 160;
151 uint y = (streamgroup_num * BLOCKS_PER_STREAM + subblock_idx) / 160;
152 uint f = imageLoad(tex, ivec2(x, y * 8 + coeff_row)).x;
154 encode_coeff(sign_extend(f & 0x1ffu, 9), enc1);
155 encode_coeff(sign_extend(f >> 9, 7), enc2);
162 void encode_8(uint streamgroup_num, uint coeff_row, layout(r8i) restrict readonly iimage2D tex, uint col, uint dist)
165 RansEncInit(streamgroup_num, coeff_row, col, dist, enc);
167 for (uint subblock_idx = 0; subblock_idx < BLOCKS_PER_STREAM; ++subblock_idx) {
168 // TODO: Use SSBOs instead of a texture?
169 uint x = (streamgroup_num * BLOCKS_PER_STREAM + subblock_idx) % 160;
170 uint y = (streamgroup_num * BLOCKS_PER_STREAM + subblock_idx) / 160;
171 int f = imageLoad(tex, ivec2(x, y * 8 + coeff_row)).x;
173 encode_coeff(f, enc);
181 uint streamgroup_num = gl_WorkGroupID.x;
182 uint coeff_row = gl_WorkGroupID.y; // 0..7
183 uint coeff_colset = gl_WorkGroupID.z; // 0 = dc+ac7, 1 = ac1+ac6, 2 = ac2+ac5, 3 = ac3, 4 = ac5
184 uint m = luma_mapping[coeff_row];
186 // TODO: DC coeff pred
188 if (coeff_colset == 0) {
189 uint dist_dc = bitfieldExtract(m, 0, 2);
190 uint dist_ac7 = bitfieldExtract(m, 14, 2);
191 encode_9_7(streamgroup_num, coeff_row, dc_ac7_tex, 0, 7, dist_dc, dist_ac7);
192 } else if (coeff_colset == 1) {
193 uint dist_ac1 = bitfieldExtract(m, 2, 2);
194 uint dist_ac6 = bitfieldExtract(m, 12, 2);
195 encode_9_7(streamgroup_num, coeff_row, ac1_ac6_tex, 1, 6, dist_ac1, dist_ac6);
196 } else if (coeff_colset == 2) {
197 uint dist_ac2 = bitfieldExtract(m, 4, 2);
198 uint dist_ac5 = bitfieldExtract(m, 10, 2);
199 encode_9_7(streamgroup_num, coeff_row, ac2_ac5_tex, 2, 5, dist_ac2, dist_ac5);
200 } else if (coeff_colset == 3) {
201 uint dist_ac3 = bitfieldExtract(m, 6, 2);
202 encode_8(streamgroup_num, coeff_row, ac3_tex, 3, dist_ac3);
204 uint dist_ac4 = bitfieldExtract(m, 8, 2);
205 encode_8(streamgroup_num, coeff_row, ac4_tex, 4, dist_ac4);