2 #extension GL_ARB_shader_clock : enable
4 #define ENABLE_TIMING 0
6 layout(local_size_x = 8, local_size_y = 8) in;
7 layout(r8ui) uniform restrict readonly uimage2D cum2sym_tex;
8 layout(rg16ui) uniform restrict readonly uimage2D dsyms_tex;
9 layout(r8) uniform restrict writeonly image2D out_tex;
11 const uint prob_bits = 12;
12 const uint prob_scale = 1 << prob_bits;
13 const uint NUM_SYMS = 256;
14 const uint ESCAPE_LIMIT = NUM_SYMS - 1;
16 // These need to be folded into quant_matrix.
17 const float dc_scalefac = 8.0;
18 const float quant_scalefac = 4.0;
20 const float quant_matrix[64] = {
21 8, 16, 19, 22, 26, 27, 29, 34,
22 16, 16, 22, 24, 27, 29, 34, 37,
23 19, 22, 26, 27, 29, 34, 34, 38,
24 22, 22, 26, 27, 29, 34, 37, 40,
25 22, 26, 27, 29, 32, 35, 40, 48,
26 26, 27, 29, 32, 35, 40, 48, 58,
27 26, 27, 29, 34, 38, 46, 56, 69,
28 27, 29, 35, 38, 46, 56, 69, 83
30 const uint ff_zigzag_direct[64] = {
31 0, 1, 8, 16, 9, 2, 3, 10,
32 17, 24, 32, 25, 18, 11, 4, 5,
33 12, 19, 26, 33, 40, 48, 41, 34,
34 27, 20, 13, 6, 7, 14, 21, 28,
35 35, 42, 49, 56, 57, 50, 43, 36,
36 29, 22, 15, 23, 30, 37, 44, 51,
37 58, 59, 52, 45, 38, 31, 39, 46,
38 53, 60, 61, 54, 47, 55, 62, 63
41 layout(std430, binding = 9) buffer layoutName
45 layout(std430, binding = 10) buffer layoutName2
47 uvec2 timing[10 * 64];
51 uint src_offset, src_len;
53 layout(std430, binding = 0) buffer whatever3
55 CoeffStream streams[];
57 uniform uint sign_bias_per_model[16];
59 const uint RANS_BYTE_L = (1u << 23); // lower bound of our normalization interval
61 uint get_rans_byte(uint offset)
63 // We assume little endian.
64 return bitfieldExtract(data_SSBO[offset >> 2], 8 * int(offset & 3u), 8);
67 uint RansDecInit(inout uint offset)
71 x = get_rans_byte(offset);
72 x |= get_rans_byte(offset + 1) << 8;
73 x |= get_rans_byte(offset + 2) << 16;
74 x |= get_rans_byte(offset + 3) << 24;
80 uint RansDecGet(uint r, uint scale_bits)
82 return r & ((1u << scale_bits) - 1);
85 void RansDecAdvance(inout uint rans, inout uint offset, const uint start, const uint freq, uint prob_bits)
87 const uint mask = (1u << prob_bits) - 1;
88 rans = freq * (rans >> prob_bits) + (rans & mask) - start;
91 while (rans < RANS_BYTE_L) {
92 rans = (rans << 8) | get_rans_byte(offset++);
96 uint cum2sym(uint bits, uint table)
98 return imageLoad(cum2sym_tex, ivec2(bits, table)).x;
101 uvec2 get_dsym(uint k, uint table)
103 return imageLoad(dsyms_tex, ivec2(k, table)).xy;
106 void idct_1d(inout float y0, inout float y1, inout float y2, inout float y3, inout float y4, inout float y5, inout float y6, inout float y7)
108 const float a1 = 0.7071067811865474; // sqrt(2)
109 const float a2 = 0.5411961001461971; // cos(3/8 pi) * sqrt(2)
110 const float a4 = 1.3065629648763766; // cos(pi/8) * sqrt(2)
111 // static const float a5 = 0.5 * (a4 - a2);
112 const float a5 = 0.3826834323650897;
114 // phase 2 (phase 1 is just moving around)
115 const float p2_4 = y5 - y3;
116 const float p2_5 = y1 + y7;
117 const float p2_6 = y1 - y7;
118 const float p2_7 = y5 + y3;
121 const float p3_2 = y2 - y6;
122 const float p3_3 = y2 + y6;
123 const float p3_5 = p2_5 - p2_7;
124 const float p3_7 = p2_5 + p2_7;
127 const float p4_2 = a1 * p3_2;
128 const float p4_4 = p2_4 * a2 + (p2_4 + p2_6) * a5; // Inverted.
129 const float p4_5 = a1 * p3_5;
130 const float p4_6 = p2_6 * a4 - (p2_4 + p2_6) * a5;
133 const float p5_0 = y0 + y4;
134 const float p5_1 = y0 - y4;
135 const float p5_3 = p4_2 + p3_3;
138 const float p6_0 = p5_0 + p5_3;
139 const float p6_1 = p5_1 + p4_2;
140 const float p6_2 = p5_1 - p4_2;
141 const float p6_3 = p5_0 - p5_3;
142 const float p6_5 = p4_5 + p4_4;
143 const float p6_6 = p4_5 + p4_6;
144 const float p6_7 = p4_6 + p3_7;
157 shared float temp[64 * 8];
159 void pick_timer(inout uvec2 start, inout uvec2 t)
162 uvec2 now = clock2x32ARB();
164 uvec2 delta = now - start;
165 if (now.x < start.x) {
169 uvec2 new_t = t + delta;
175 start = clock2x32ARB();
181 uvec2 local_timing[10];
183 for (int timer_idx = 0; timer_idx < 10; ++timer_idx) {
184 local_timing[timer_idx] = uvec2(0, 0);
186 uvec2 start = clock2x32ARB();
191 const uint num_blocks = 720 / 16; // FIXME: make a uniform
192 const uint thread_num = gl_LocalInvocationID.y * 8 + gl_LocalInvocationID.x;
194 const uint block_row = gl_WorkGroupID.y;
195 //const uint coeff_num = ff_zigzag_direct[thread_num];
196 const uint coeff_num = thread_num;
197 const uint stream_num = coeff_num * num_blocks + block_row;
198 //const uint stream_num = block_row * num_blocks + coeff_num; // HACK
199 const uint model_num = min((coeff_num % 8) + (coeff_num / 8), 7);
200 const uint sign_bias = sign_bias_per_model[model_num];
202 // Initialize rANS decoder.
203 uint offset = streams[stream_num].src_offset;
204 uint rans = RansDecInit(offset);
206 float q = (coeff_num == 0) ? 1.0 : (quant_matrix[coeff_num] * quant_scalefac / 128.0 / sqrt(2.0)); // FIXME: fold
208 //int w = (coeff_num == 0) ? 32 : int(quant_matrix[coeff_num]);
211 pick_timer(start, local_timing[0]);
213 for (uint block_idx = 40; block_idx --> 0; ) {
214 uint block_x = block_idx % 20;
215 uint block_y = block_idx / 20;
216 if (block_x == 19) last_k = 0;
218 pick_timer(start, local_timing[1]);
220 // rANS decode one coefficient across eight blocks (so 64x8 coefficients).
221 for (uint subblock_idx = 8; subblock_idx --> 0; ) {
223 uint bottom_bits = RansDecGet(rans, prob_bits + 1);
225 if (bottom_bits >= sign_bias) {
226 bottom_bits -= sign_bias;
230 int k = int(cum2sym(bottom_bits, model_num)); // Can go out-of-bounds; that will return zero.
231 uvec2 sym = get_dsym(k, model_num);
232 RansDecAdvance(rans, offset, sym.x, sym.y, prob_bits + 1);
234 if (k == ESCAPE_LIMIT) {
235 k = int(RansDecGet(rans, prob_bits));
236 RansDecAdvance(rans, offset, k, 1, prob_bits);
242 if (coeff_num == 0) {
247 temp[subblock_idx * 64 + coeff_num] = k * q;
248 //temp[subblock_idx * 64 + 8 * y + x] = (2 * k * w * 4) / 32; // 100% matching unquant
251 pick_timer(start, local_timing[2]);
253 memoryBarrierShared();
256 pick_timer(start, local_timing[3]);
258 // Horizontal DCT one row (so 64 rows).
259 idct_1d(temp[thread_num * 8 + 0],
260 temp[thread_num * 8 + 1],
261 temp[thread_num * 8 + 2],
262 temp[thread_num * 8 + 3],
263 temp[thread_num * 8 + 4],
264 temp[thread_num * 8 + 5],
265 temp[thread_num * 8 + 6],
266 temp[thread_num * 8 + 7]);
268 pick_timer(start, local_timing[4]);
270 memoryBarrierShared();
273 pick_timer(start, local_timing[5]);
275 // Vertical DCT one row (so 64 columns).
276 uint row_offset = gl_LocalInvocationID.y * 64 + gl_LocalInvocationID.x;
277 idct_1d(temp[row_offset + 0 * 8],
278 temp[row_offset + 1 * 8],
279 temp[row_offset + 2 * 8],
280 temp[row_offset + 3 * 8],
281 temp[row_offset + 4 * 8],
282 temp[row_offset + 5 * 8],
283 temp[row_offset + 6 * 8],
284 temp[row_offset + 7 * 8]);
286 pick_timer(start, local_timing[6]);
288 uint y = block_row * 16 + block_y * 8;
289 uint x = block_x * 64 + gl_LocalInvocationID.y * 8 + gl_LocalInvocationID.x;
290 for (uint yl = 0; yl < 8; ++yl) {
291 imageStore(out_tex, ivec2(x, yl + y), vec4(temp[row_offset + yl * 8], 0.0, 0.0, 1.0));
294 pick_timer(start, local_timing[7]);
296 memoryBarrierShared(); // is this needed?
299 pick_timer(start, local_timing[8]);
300 pick_timer(start, local_timing[9]); // should be nearly nothing
304 for (int timer_idx = 0; timer_idx < 10; ++timer_idx) {
305 uint global_idx = thread_num * 10 + timer_idx;
307 uint old_val = atomicAdd(timing[global_idx].x, local_timing[timer_idx].x);
308 if (old_val + local_timing[timer_idx].x < old_val) {
309 ++local_timing[timer_idx].y;
311 atomicAdd(timing[global_idx].y, local_timing[timer_idx].y);