libslic3r: split the system preset load across threads reliably

The parallel_for over the vendors used an auto_partitioner, which splits its
range only while a worker is asking for work. At six vendors that often did not
happen at all. In two runs out of four on this machine every vendor in that step
ran on the calling thread, which the log shows as one thread id for all of them.

A grain of one splits the range up front. The vendors also start slowest first,
so the vendor that decides when the step ends is not the one left until last. A
vendor with no cache is parsed from its JSONs, which costs far more than any
cache load, so those go before cached ones, and within each group the bigger
file goes first, the cache or the profile index that names the vendor's presets.

src/libslic3r/PresetBundle.cpp

@@ -19,6 +19,7 @@
 
 #include <algorithm>
 #include <cstdlib>
+#include <numeric>
 #include <set>
 #include <fstream>
 #include <unordered_set>
@@ -37,6 +38,7 @@
 #include <miniz/miniz.h>
 #include <tbb/blocked_range.h>
 #include <tbb/parallel_for.h>
+#include <tbb/partitioner.h>
 
 // Mark string for localization and translate.
 #define L(s) Slic3r::I18N::translate(s)
@@ -2552,9 +2554,32 @@ std::pair<PresetsConfigSubstitutions, std::string> PresetBundle::load_system_pre
     std::vector<PresetsConfigSubstitutions>         parallel_substitutions(other_vendors.size());
     std::vector<std::string>                        parallel_errors(other_vendors.size());
 
-    tbb::parallel_for(tbb::blocked_range<size_t>(0, other_vendors.size()),
+    // Started slowest first, since this step ends when the slowest vendor does.
+    std::vector<size_t> by_cost(other_vendors.size());
+    std::iota(by_cost.begin(), by_cost.end(), size_t(0));
+    // A vendor with no cache is parsed from its JSONs, which costs far more than
+    // any cache load, so those come first, and within each group the bigger file
+    // goes first, the cache or the <vendor>.json that names every preset to parse.
+    std::vector<std::pair<bool, uintmax_t>> vendor_costs(other_vendors.size());
+    for (size_t i = 0; i < other_vendors.size(); ++i) {
+        boost::system::error_code ec;
+        const uintmax_t cache_size = boost::filesystem::file_size(dir / (other_vendors[i] + ".opc"), ec);
+        if (! ec) {
+            vendor_costs[i] = { false, cache_size };
+            continue;
+        }
+        const uintmax_t index_size = boost::filesystem::file_size(dir / (other_vendors[i] + ".json"), ec);
+        vendor_costs[i] = { true, ec ? 0 : index_size };
+    }
+    std::stable_sort(by_cost.begin(), by_cost.end(),
+        [&](size_t a, size_t b) { return vendor_costs[a] > vendor_costs[b]; });
+
+    // An auto_partitioner splits the range only while a worker is asking for work,
+    // which at this size left every vendor on the calling thread.
+    tbb::parallel_for(tbb::blocked_range<size_t>(0, other_vendors.size(), 1),
         [&](const tbb::blocked_range<size_t>& range) {
-            for (size_t i = range.begin(); i < range.end(); ++i) {
+            for (size_t k = range.begin(); k < range.end(); ++k) {
+                const size_t i = by_cost[k];
                 auto bundle = std::make_unique<PresetBundle>();
                 bundle->set_is_validation_mode(validation_mode);
                 bundle->set_generate_vendor_caches(m_generate_vendor_caches);
@@ -2567,7 +2592,7 @@ std::pair<PresetsConfigSubstitutions, std::string> PresetBundle::load_system_pre
                     parallel_errors[i] = err.what();
                 }
             }
-        });
+        }, tbb::simple_partitioner());
 
     // Step 3: Sequentially merge the parallel-loaded bundles into `this`.
     // The merge order is the original vendor order so any duplicate-warning