--- /dev/null
+#version 440
+#extension GL_NV_gpu_shader5 : enable
+
+layout(local_size_x = 1) in;
+
+const uint prob_bits = 13; // Note!
+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 NUM_SYMS = 256;
+const uint ESCAPE_LIMIT = NUM_SYMS - 1;
+
+#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))
+
+const uint luma_mapping[8] = {
+ MAPPING(0, 0, 1, 1, 2, 2, 3, 3),
+ MAPPING(0, 0, 1, 2, 2, 2, 3, 3),
+ MAPPING(1, 1, 2, 2, 2, 3, 3, 3),
+ MAPPING(1, 1, 2, 2, 2, 3, 3, 3),
+ MAPPING(1, 2, 2, 2, 2, 3, 3, 3),
+ MAPPING(2, 2, 2, 2, 3, 3, 3, 3),
+ MAPPING(2, 2, 3, 3, 3, 3, 3, 3),
+ MAPPING(3, 3, 3, 3, 3, 3, 3, 3),
+};
+
+layout(std430, binding = 9) buffer layoutName
+{
+ uint dist[4 * 256];
+ uvec2 ransdist[4 * 256];
+};
+
+layout(std430, binding = 10) buffer outputBuf
+{
+ uint8_t rans_output[];
+};
+
+layout(std430, binding = 11) buffer outputBuf2
+{
+ uint rans_start_offset[];
+};
+
+struct RansEncoder {
+ uint stream_num; // const
+ uint lut_base; // const
+ uint rans_offset;
+ uint rans;
+};
+
+layout(r16ui) uniform restrict readonly uimage2D dc_ac7_tex;
+layout(r16ui) uniform restrict readonly uimage2D ac1_ac6_tex;
+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 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.rans_offset = enc.stream_num * STREAM_BUF_SIZE + STREAM_BUF_SIZE; // Starts at the end.
+ enc.rans = RANS_BYTE_L;
+}
+
+void RansEncRenorm(inout uint rans, inout uint rans_offset, 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);
+ }
+}
+
+void RansEncPut(inout uint rans, inout uint rans_offset, uint start, uint freq, uint prob_bits)
+{
+ RansEncRenorm(rans, rans_offset, freq, prob_bits);
+ rans = ((rans / freq) << prob_bits) + (rans % freq) + start;
+}
+
+void RansEncFlush(uint rans, inout uint rans_offset)
+{
+ 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);
+}
+
+void encode_coeff(uint coeff, uint bits, inout RansEncoder enc)
+{
+ // Sign-extend to recover the coefficient.
+ // FIXME: not needed for the bits == 8 case!
+ int signed_k = int(coeff << (32 - bits)) >> (32 - bits);
+ uint k = abs(signed_k);
+
+ if (k >= ESCAPE_LIMIT) {
+ // ... boring stuff here
+ RansEncPut(enc.rans, enc.rans_offset, 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);
+
+ // fix some bias stuff here
+}
+
+void encode_end(inout RansEncoder enc)
+{
+ RansEncFlush(enc.rans, enc.rans_offset);
+ rans_start_offset[enc.stream_num] = enc.rans_offset;
+}
+
+void encode_9_7(uint streamgroup_num, uint coeff_row, layout(r16ui) restrict readonly uimage2D tex, uint col1, uint col2, uint dist1, uint dist2)
+{
+ RansEncoder enc1, enc2;
+ RansEncInit(streamgroup_num, coeff_row, col1, dist1, enc1);
+ RansEncInit(streamgroup_num, coeff_row, col2, dist2, enc2);
+
+ for (uint subblock_idx = BLOCKS_PER_STREAM; subblock_idx --> 0; ) {
+ // 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(f & 0x1ffu, 9, enc1);
+ encode_coeff(f >> 9, 7, enc2);
+ }
+
+ encode_end(enc1);
+ encode_end(enc2);
+}
+
+void encode_8(uint streamgroup_num, uint coeff_row, layout(r8i) restrict readonly iimage2D tex, uint col, uint dist)
+{
+ RansEncoder enc;
+ RansEncInit(streamgroup_num, coeff_row, col, dist, enc);
+
+ for (uint subblock_idx = BLOCKS_PER_STREAM; subblock_idx --> 0; ) {
+ // 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, 8, enc);
+ }
+
+ encode_end(enc);
+}
+
+void main()
+{
+ uint streamgroup_num = gl_WorkGroupID.x;
+ uint coeff_row = gl_WorkGroupID.y; // 0..7
+ uint coeff_colset = gl_WorkGroupID.z; // 0 = dc+ac7, 1 = ac1+ac6, 2 = ac2+ac5, 3 = ac3, 4 = ac5
+ uint m = luma_mapping[coeff_row];
+
+ // TODO: DC coeff pred
+
+ if (coeff_colset == 0) {
+ uint dist_dc = bitfieldExtract(m, 0, 2);
+ uint dist_ac7 = bitfieldExtract(m, 14, 2);
+ encode_9_7(streamgroup_num, coeff_row, dc_ac7_tex, 0, 7, dist_dc, dist_ac7);
+ } else if (coeff_colset == 1) {
+ uint dist_ac1 = bitfieldExtract(m, 2, 2);
+ uint dist_ac6 = bitfieldExtract(m, 12, 2);
+ encode_9_7(streamgroup_num, coeff_row, ac1_ac6_tex, 1, 6, dist_ac1, dist_ac6);
+ } else if (coeff_colset == 2) {
+ uint dist_ac2 = bitfieldExtract(m, 4, 2);
+ uint dist_ac5 = bitfieldExtract(m, 10, 2);
+ encode_9_7(streamgroup_num, coeff_row, ac2_ac5_tex, 2, 5, dist_ac2, dist_ac5);
+ } else if (coeff_colset == 3) {
+ uint dist_ac3 = bitfieldExtract(m, 6, 2);
+ encode_8(streamgroup_num, coeff_row, ac3_tex, 3, dist_ac3);
+ } else {
+ uint dist_ac4 = bitfieldExtract(m, 8, 2);
+ encode_8(streamgroup_num, coeff_row, ac4_tex, 4, dist_ac4);
+ }
+}