Implement streaming encoder mode with constant memory usage. (libjxl#…

…3017) * Implement streaming encoder mode with constant memory usage. All the internal images are now limited to the 2048x2048 DC group size, and the maximum memory usage is around 500 MB. Benchmark results: ``` Encoding kPixels Bytes BPP E MP/s D MP/s Max norm SSIMULACRA2 PSNR pnorm BPP*pnorm QABPP Bugs --------------------------------------------------------------------------------------------------------------------------------------------- jxl:d0.1:1 156870 125499004 6.4001333 2.883 17.712 0.33339094 93.97017505 55.04 0.11910771 0.762305231867 6.400 0 jxl:d0.1:3 156870 118196370 6.0277173 1.342 14.221 0.33339094 93.97017505 55.04 0.11910771 0.717947620036 6.028 0 jxl:d0.1:4 156870 117892031 6.0121968 1.501 14.289 0.33382890 94.00462704 55.04 0.11907577 0.715906976850 6.012 0 jxl:d0.1:5 156870 139407423 7.1094276 0.777 14.252 0.34058568 93.91823707 57.02 0.10181720 0.723862019767 7.109 0 jxl:d0.1:6 156870 140721309 7.1764325 0.568 13.979 0.34023632 93.89339858 57.11 0.10187573 0.731104323064 7.176 0 jxl:d0.1:7 156870 141506387 7.2164695 0.560 14.325 0.33378362 93.90823930 57.19 0.10142266 0.731913544176 7.216 0 jxl:d0.1:1:buf3 156870 124186341 6.3331908 4.689 16.793 0.33339094 93.97017505 55.04 0.11910771 0.754331862831 6.333 0 jxl:d0.1:3:buf3 156870 118645467 6.0506202 3.936 15.506 0.33339094 93.97017505 55.04 0.11910771 0.720675522105 6.051 0 jxl:d0.1:4:buf3 156870 119051450 6.0713243 4.091 15.353 0.33382890 94.00462704 55.04 0.11907577 0.722947623654 6.071 0 jxl:d0.1:5:buf3 156870 124960819 6.3726872 1.788 13.883 0.42522492 93.86107006 55.38 0.11737111 0.747969378745 6.373 0 jxl:d0.1:6:buf3 156870 125038021 6.3766243 1.194 14.652 0.42531900 93.88796957 55.40 0.11745910 0.748992546623 6.377 0 jxl:d0.1:7:buf3 156870 125896374 6.4203982 1.156 14.181 0.43496868 93.90811739 55.47 0.11689811 0.750532425816 6.420 0 jxl:d1:1 156870 30373676 1.5489810 3.437 24.764 1.43588506 86.14940783 43.59 0.58841531 0.911444150987 2.257 0 jxl:d1:3 156870 27841570 1.4198500 2.069 26.847 1.43588506 86.14940783 43.59 0.58841531 0.835461474297 2.068 0 jxl:d1:4 156870 28110273 1.4335532 2.860 24.650 1.43953718 86.43275684 43.61 0.58675396 0.841142996644 2.092 0 jxl:d1:5 156870 23502237 1.1985549 0.913 23.353 1.49287327 84.49628039 43.63 0.58180371 0.697323707375 1.816 0 jxl:d1:6 156870 23686906 1.2079726 0.852 24.001 1.48966351 84.58298559 43.65 0.58294516 0.704181767609 1.829 0 jxl:d1:7 156870 23680833 1.2076629 0.662 24.676 1.49379655 84.59488978 43.66 0.58316869 0.704271179346 1.829 0 jxl:d1:1:buf3 156870 30144399 1.5372885 6.648 28.425 1.43588506 86.14940783 43.59 0.58841531 0.904564075601 2.240 0 jxl:d1:3:buf3 156870 28077983 1.4319065 6.575 26.224 1.43588506 86.14940783 43.59 0.58841531 0.842555684628 2.085 0 jxl:d1:4:buf3 156870 28586708 1.4578501 6.951 28.615 1.43953718 86.43275684 43.61 0.58675396 0.855399349245 2.127 0 jxl:d1:5:buf3 156870 23664692 1.2068397 2.236 25.647 1.79721683 84.31625183 43.00 0.62510227 0.754398246796 2.203 0 jxl:d1:6:buf3 156870 23700533 1.2086675 1.586 25.146 1.79445658 84.33377882 43.02 0.62494042 0.755345188024 2.201 0 jxl:d1:7:buf3 156870 23699715 1.2086258 1.441 24.907 1.79768241 84.34806803 43.03 0.62501552 0.755409881140 2.202 0 jxl:d10:1 156870 4634629 0.2363544 6.664 10.740 6.71570493 38.19963812 35.71 2.25288348 0.532478951288 1.600 0 jxl:d10:3 156870 3819655 0.1947928 5.277 12.029 6.71570493 38.19963812 35.71 2.25288348 0.438845458543 1.320 0 jxl:d10:4 156870 4137919 0.2110235 5.224 11.493 6.71726114 40.72074578 35.94 2.19382440 0.462948398734 1.427 0 jxl:d10:5 156870 3064736 0.1562938 1.756 12.541 9.14517281 30.37032353 35.23 2.69294022 0.420889917659 1.429 0 jxl:d10:6 156870 3123745 0.1593031 1.146 13.002 9.35053938 31.50052643 35.32 2.68063191 0.427033058530 1.496 0 jxl:d10:7 156870 3129105 0.1595765 0.913 13.258 9.23112087 31.60086736 35.33 2.67094507 0.426220007652 1.486 0 jxl:d10:1:buf3 156870 4788892 0.2442214 9.607 12.413 6.71570493 38.19963812 35.71 2.25288348 0.550202441230 1.652 0 jxl:d10:3:buf3 156870 4017087 0.2048613 9.304 11.483 6.71570493 38.19963812 35.71 2.25288348 0.461528694744 1.388 0 jxl:d10:4:buf3 156870 4396053 0.2241876 8.637 11.715 6.71726114 40.72074578 35.94 2.19382440 0.491828307200 1.516 0 jxl:d10:5:buf3 156870 3738546 0.1906564 2.497 11.388 7.36731798 37.57962463 35.82 2.26285872 0.431428576133 1.419 0 jxl:d10:6:buf3 156870 3790243 0.1932929 1.851 11.352 7.39668835 38.22880038 35.86 2.25680857 0.436224963822 1.442 0 jxl:d10:7:buf3 156870 3791016 0.1933323 1.679 11.463 7.40258595 38.25278127 35.86 2.25490305 0.435945531927 1.444 0 Aggregate: 156870 23301908 1.1883387 2.288 16.536 1.59684548 66.47757179 44.19 0.54200030 0.644079906108 2.696 0 ``` * Address review comments. * Rename update_global_state to initialize_global_state
libsdl-org · Dec 13, 2023 · 68c57e7 · 68c57e7
1 parent e986e5c
commit 68c57e7
Show file tree

Hide file tree

Showing 26 changed files with 932 additions and 400 deletions.
diff --git a/lib/jxl/ans_test.cc b/lib/jxl/ans_test.cc
@@ -39,7 +39,7 @@ void RoundtripTestcase(int n_histograms, int alphabet_size,
 
   BuildAndEncodeHistograms(HistogramParams(), n_histograms, input_values_vec,
                            &codes, &context_map, &writer, 0, nullptr);
-  WriteTokens(input_values_vec[0], codes, context_map, &writer, 0, nullptr);
+  WriteTokens(input_values_vec[0], codes, context_map, 0, &writer, 0, nullptr);
 
   // Magic bytes + padding
   BitWriter::Allotment allotment_magic2(&writer, 24);
@@ -211,7 +211,8 @@ void TestCheckpointing(bool ans, bool lz77) {
     auto input_values_copy = input_values;
     BuildAndEncodeHistograms(params, 1, input_values_copy, &codes, &context_map,
                              &writer, 0, nullptr);
-    WriteTokens(input_values_copy[0], codes, context_map, &writer, 0, nullptr);
+    WriteTokens(input_values_copy[0], codes, context_map, 0, &writer, 0,
+                nullptr);
     writer.ZeroPadToByte();
   }
 

diff --git a/lib/jxl/dec_ans.cc b/lib/jxl/dec_ans.cc
@@ -341,6 +341,9 @@ Status DecodeHistograms(BitReader* br, size_t num_contexts, ANSCode* code,
   if (num_contexts > 1) {
     JXL_RETURN_IF_ERROR(DecodeContextMap(context_map, &num_histograms, br));
   }
+  JXL_DEBUG_V(
+      4, "Decoded context map of size %" PRIuS " and %" PRIuS " histograms",
+      num_contexts, num_histograms);
   code->lz77.nonserialized_distance_context = context_map->back();
   code->use_prefix_code = br->ReadFixedBits<1>();
   if (code->use_prefix_code) {
@@ -354,17 +357,6 @@ Status DecodeHistograms(BitReader* br, size_t num_contexts, ANSCode* code,
   const size_t max_alphabet_size = 1 << code->log_alpha_size;
   JXL_RETURN_IF_ERROR(
       DecodeANSCodes(num_histograms, max_alphabet_size, br, code));
-  // When using LZ77, flat codes might result in valid codestreams with
-  // histograms that potentially allow very large bit counts.
-  // TODO(veluca): in principle, a valid codestream might contain a histogram
-  // that could allow very large numbers of bits that is never used during ANS
-  // decoding. There's no benefit to doing that, though.
-  if (!code->lz77.enabled && code->max_num_bits > 32) {
-    // Just emit a warning as there are many opportunities for false positives.
-    JXL_WARNING("Histogram can represent numbers that are too large: %" PRIuS
-                "\n",
-                code->max_num_bits);
-  }
   return true;
 }