2 #extension GL_ARB_shader_clock : enable
4 #define PARALLEL_SLICES 1
6 #define ENABLE_TIMING 0
8 layout(local_size_x = 64*PARALLEL_SLICES) in;
9 layout(r8ui) uniform restrict readonly uimage2D cum2sym_tex;
10 layout(rg16ui) uniform restrict readonly uimage2D dsyms_tex;
11 layout(r8) uniform restrict writeonly image2D out_tex;
12 layout(r16i) uniform restrict writeonly iimage2D coeff_tex;
14 const uint prob_bits = 12;
15 const uint prob_scale = 1 << prob_bits;
16 const uint NUM_SYMS = 256;
17 const uint ESCAPE_LIMIT = NUM_SYMS - 1;
19 // These need to be folded into quant_matrix.
20 const float dc_scalefac = 8.0;
21 const float quant_scalefac = 4.0;
23 const float quant_matrix[64] = {
24 8, 16, 19, 22, 26, 27, 29, 34,
25 16, 16, 22, 24, 27, 29, 34, 37,
26 19, 22, 26, 27, 29, 34, 34, 38,
27 22, 22, 26, 27, 29, 34, 37, 40,
28 22, 26, 27, 29, 32, 35, 40, 48,
29 26, 27, 29, 32, 35, 40, 48, 58,
30 26, 27, 29, 34, 38, 46, 56, 69,
31 27, 29, 35, 38, 46, 56, 69, 83
33 const uint ff_zigzag_direct[64] = {
34 0, 1, 8, 16, 9, 2, 3, 10,
35 17, 24, 32, 25, 18, 11, 4, 5,
36 12, 19, 26, 33, 40, 48, 41, 34,
37 27, 20, 13, 6, 7, 14, 21, 28,
38 35, 42, 49, 56, 57, 50, 43, 36,
39 29, 22, 15, 23, 30, 37, 44, 51,
40 58, 59, 52, 45, 38, 31, 39, 46,
41 53, 60, 61, 54, 47, 55, 62, 63
44 layout(std430, binding = 9) buffer layoutName
48 layout(std430, binding = 10) buffer layoutName2
50 uvec2 timing[10 * 64];
54 uint src_offset, src_len;
56 layout(std430, binding = 0) buffer whatever3
58 CoeffStream streams[];
60 uniform uint sign_bias_per_model[16];
62 const uint RANS_BYTE_L = (1u << 23); // lower bound of our normalization interval
64 uint get_rans_byte(uint offset)
66 // We assume little endian.
67 return bitfieldExtract(data_SSBO[offset >> 2], 8 * int(offset & 3u), 8);
70 uint RansDecInit(inout uint offset)
74 x = get_rans_byte(offset);
75 x |= get_rans_byte(offset + 1) << 8;
76 x |= get_rans_byte(offset + 2) << 16;
77 x |= get_rans_byte(offset + 3) << 24;
83 uint RansDecGet(uint r, uint scale_bits)
85 return r & ((1u << scale_bits) - 1);
88 void RansDecAdvance(inout uint rans, inout uint offset, const uint start, const uint freq, uint prob_bits)
90 const uint mask = (1u << prob_bits) - 1;
91 rans = freq * (rans >> prob_bits) + (rans & mask) - start;
94 while (rans < RANS_BYTE_L) {
95 rans = (rans << 8) | get_rans_byte(offset++);
99 uint cum2sym(uint bits, uint table)
101 return imageLoad(cum2sym_tex, ivec2(bits, table)).x;
104 uvec2 get_dsym(uint k, uint table)
106 return imageLoad(dsyms_tex, ivec2(k, table)).xy;
109 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)
111 const float a1 = 0.7071067811865474; // sqrt(2)
112 const float a2 = 0.5411961001461971; // cos(3/8 pi) * sqrt(2)
113 const float a4 = 1.3065629648763766; // cos(pi/8) * sqrt(2)
114 // static const float a5 = 0.5 * (a4 - a2);
115 const float a5 = 0.3826834323650897;
117 // phase 2 (phase 1 is just moving around)
118 const float p2_4 = y5 - y3;
119 const float p2_5 = y1 + y7;
120 const float p2_6 = y1 - y7;
121 const float p2_7 = y5 + y3;
124 const float p3_2 = y2 - y6;
125 const float p3_3 = y2 + y6;
126 const float p3_5 = p2_5 - p2_7;
127 const float p3_7 = p2_5 + p2_7;
130 const float p4_2 = a1 * p3_2;
131 const float p4_4 = p2_4 * a2 + (p2_4 + p2_6) * a5; // Inverted.
132 const float p4_5 = a1 * p3_5;
133 const float p4_6 = p2_6 * a4 - (p2_4 + p2_6) * a5;
136 const float p5_0 = y0 + y4;
137 const float p5_1 = y0 - y4;
138 const float p5_3 = p4_2 + p3_3;
141 const float p6_0 = p5_0 + p5_3;
142 const float p6_1 = p5_1 + p4_2;
143 const float p6_2 = p5_1 - p4_2;
144 const float p6_3 = p5_0 - p5_3;
145 const float p6_5 = p4_5 + p4_4;
146 const float p6_6 = p4_5 + p4_6;
147 const float p6_7 = p4_6 + p3_7;
160 shared float temp[64 * 8 * PARALLEL_SLICES];
162 void pick_timer(inout uvec2 start, inout uvec2 t)
165 uvec2 now = clock2x32ARB();
167 uvec2 delta = now - start;
168 if (now.x < start.x) {
172 uvec2 new_t = t + delta;
178 start = clock2x32ARB();
184 uvec2 local_timing[10];
186 for (int timer_idx = 0; timer_idx < 10; ++timer_idx) {
187 local_timing[timer_idx] = uvec2(0, 0);
189 uvec2 start = clock2x32ARB();
191 uvec2 start = uvec2(0, 0);
192 local_timing[0] = start;
195 const uint local_x = gl_LocalInvocationID.x % 8;
196 const uint local_y = (gl_LocalInvocationID.x / 8) % 8;
197 const uint local_z = gl_LocalInvocationID.x / 64;
199 const uint num_blocks = 720 / 16; // FIXME: make a uniform
200 const uint slice_num = local_z;
201 const uint thread_num = local_y * 8 + local_x;
203 const uint block_row = gl_WorkGroupID.y * PARALLEL_SLICES + slice_num;
204 //const uint coeff_num = ff_zigzag_direct[thread_num];
205 const uint coeff_num = thread_num;
206 const uint stream_num = coeff_num * num_blocks + block_row;
207 const uint model_num = min((coeff_num % 8) + (coeff_num / 8), 7);
208 const uint sign_bias = sign_bias_per_model[model_num];
210 // Initialize rANS decoder.
211 uint offset = streams[stream_num].src_offset;
212 uint rans = RansDecInit(offset);
214 float q = (coeff_num == 0) ? 1.0 : (quant_matrix[coeff_num] * quant_scalefac / 128.0 / sqrt(2.0)); // FIXME: fold
216 //int w = (coeff_num == 0) ? 32 : int(quant_matrix[coeff_num]);
219 pick_timer(start, local_timing[0]);
221 for (uint block_idx = 40; block_idx --> 0; ) {
222 uint block_x = block_idx % 20;
223 uint block_y = block_idx / 20;
224 if (block_x == 19) last_k = 0;
226 pick_timer(start, local_timing[1]);
228 // rANS decode one coefficient across eight blocks (so 64x8 coefficients).
229 for (uint subblock_idx = 8; subblock_idx --> 0; ) {
231 uint bottom_bits = RansDecGet(rans, prob_bits + 1);
233 if (bottom_bits >= sign_bias) {
234 bottom_bits -= sign_bias;
238 int k = int(cum2sym(bottom_bits, model_num)); // Can go out-of-bounds; that will return zero.
239 uvec2 sym = get_dsym(k, model_num);
240 RansDecAdvance(rans, offset, sym.x, sym.y, prob_bits + 1);
242 if (k == ESCAPE_LIMIT) {
243 k = int(RansDecGet(rans, prob_bits));
244 RansDecAdvance(rans, offset, k, 1, prob_bits);
250 if (coeff_num == 0) {
256 uint y = block_row * 16 + block_y * 8 + local_y;
257 uint x = block_x * 64 + subblock_idx * 8 + local_x;
258 imageStore(coeff_tex, ivec2(x, y), ivec4(k, 0,0,0));
261 temp[slice_num * 64 * 8 + subblock_idx * 64 + coeff_num] = k * q;
262 //temp[subblock_idx * 64 + 8 * y + x] = (2 * k * w * 4) / 32; // 100% matching unquant
265 pick_timer(start, local_timing[2]);
267 memoryBarrierShared();
270 pick_timer(start, local_timing[3]);
272 // Horizontal DCT one row (so 64 rows).
273 idct_1d(temp[slice_num * 64 * 8 + thread_num * 8 + 0],
274 temp[slice_num * 64 * 8 + thread_num * 8 + 1],
275 temp[slice_num * 64 * 8 + thread_num * 8 + 2],
276 temp[slice_num * 64 * 8 + thread_num * 8 + 3],
277 temp[slice_num * 64 * 8 + thread_num * 8 + 4],
278 temp[slice_num * 64 * 8 + thread_num * 8 + 5],
279 temp[slice_num * 64 * 8 + thread_num * 8 + 6],
280 temp[slice_num * 64 * 8 + thread_num * 8 + 7]);
282 pick_timer(start, local_timing[4]);
284 memoryBarrierShared();
287 pick_timer(start, local_timing[5]);
289 // Vertical DCT one row (so 64 columns).
290 uint row_offset = local_z * 64 * 8 + local_y * 64 + local_x;
291 idct_1d(temp[row_offset + 0 * 8],
292 temp[row_offset + 1 * 8],
293 temp[row_offset + 2 * 8],
294 temp[row_offset + 3 * 8],
295 temp[row_offset + 4 * 8],
296 temp[row_offset + 5 * 8],
297 temp[row_offset + 6 * 8],
298 temp[row_offset + 7 * 8]);
300 pick_timer(start, local_timing[6]);
302 uint y = block_row * 16 + block_y * 8;
303 uint x = block_x * 64 + local_y * 8 + local_x;
304 for (uint yl = 0; yl < 8; ++yl) {
305 imageStore(out_tex, ivec2(x, yl + y), vec4(temp[row_offset + yl * 8], 0.0, 0.0, 1.0));
308 pick_timer(start, local_timing[7]);
310 memoryBarrierShared(); // is this needed?
313 pick_timer(start, local_timing[8]);
314 pick_timer(start, local_timing[9]); // should be nearly nothing
318 for (int timer_idx = 0; timer_idx < 10; ++timer_idx) {
319 uint global_idx = thread_num * 10 + timer_idx;
321 uint old_val = atomicAdd(timing[global_idx].x, local_timing[timer_idx].x);
322 if (old_val + local_timing[timer_idx].x < old_val) {
323 ++local_timing[timer_idx].y;
325 atomicAdd(timing[global_idx].y, local_timing[timer_idx].y);