Merge branch 'master' into master

Author: preda

src/TrigBufCache.cpp

@@ -334,7 +334,7 @@ TrigBufCache::~TrigBufCache() = default;
 TrigPtr TrigBufCache::smallTrig(u32 W, u32 nW) {
   lock_guard lock{mut};
   auto& m = small;
-  decay_t<decltype(m)>::key_type key{W, nW};
+  decay_t<decltype(m)>::key_type key{W, nW, 0, 0};
 
   TrigPtr p{};
   auto it = m.find(key);
@@ -347,21 +347,22 @@ TrigPtr TrigBufCache::smallTrig(u32 W, u32 nW) {
 }
 
 TrigPtr TrigBufCache::smallTrigCombo(u32 width, u32 middle, u32 W, u32 nW) {
-  lock_guard lock{mut};
-  auto& m = small;
 #if PREFER_DP_TO_MEM == 2             // No pre-computed trig values
-  decay_t<decltype(m)>::key_type key{W, nW};
-#else
-  // Hack so that width 512 and height 512 don't share the same buffer.  Width could share the height buffer since it is a subset of the combo height buffer.
-  // Also add in middle so that 512:15:512 does not share buffer with 512:14:512.
-  decay_t<decltype(m)>::key_type key{W, nW+middle};
+  return smallTrig(W, nW);
 #endif
 
+  lock_guard lock{mut};
+  auto& m = small;
+  decay_t<decltype(m)>::key_type key1{W, nW, width, middle};
+  // We write the "combo" under two keys, so it can also be retrieved as non-combo by smallTrig()
+  decay_t<decltype(m)>::key_type key2{W, nW, 0, 0};
+
   TrigPtr p{};
-  auto it = m.find(key);
+  auto it = m.find(key1);
   if (it == m.end() || !(p = it->second.lock())) {
     p = make_shared<TrigBuf>(context, genSmallTrigCombo(width, middle, W, nW));
-    m[key] = p;
+    m[key1] = p;
+    m[key2] = p;
     smallCache.add(p);
   }
   return p;

src/TrigBufCache.h

@@ -27,13 +27,13 @@ class TrigBufCache {
   const Context* context;
   std::mutex mut;
 
-  std::map<tuple<u32, u32>, TrigPtr::weak_type> small;
+  std::map<tuple<u32, u32, u32, u32>, TrigPtr::weak_type> small;
   std::map<tuple<u32, u32, u32>, TrigPtr::weak_type> middle;
 
   // The shared-pointers below keep the most recent set of buffers alive even without any Gpu instance
   // referencing them. This allows a single worker to delete & re-create the Gpu instance and still reuse the buffers.
-  StrongCache smallCache{6};
-  StrongCache middleCache{6};
+  StrongCache smallCache{4};
+  StrongCache middleCache{4};
 
 public:
   TrigBufCache(const Context* context) :

tools/primenet.py

@@ -74,10 +74,15 @@ def fetch(what):
     # print(res)
 
     BEGIN_MARK = "<!--BEGIN_ASSIGNMENTS_BLOCK-->"
-    begin = res.find("<!--BEGIN_ASSIGNMENTS_BLOCK-->")
-    if begin == -1: raise(AssertionError("assignment no BEGIN mark"))
-    begin += len(BEGIN_MARK)
-    end   = res.find("<!--END_ASSIGNMENTS_BLOCK-->", begin)
+    # begin = res.find(BEGIN_MARK)
+    begin = res.find(">PRP=")
+    if begin == -1: begin = res.find(">LL=")
+    if begin == -1:
+        print(res)
+        raise(AssertionError("assignment no BEGIN mark"))
+    begin += 1
+    # begin += len(BEGIN_MARK)
+    end   = res.find("</", begin)
     if end == -1: raise(AssertionError("assignemnt no END mark"))
     line = res[begin:end].strip().strip('\n')
     print(datetime.now(), " New assignment: ", line)