Merge df10ccbc8c into 36fdfa2694

Optimized autoconfig for AMD CPUs with < 2 MB L3 cache per thread

src/backend/cpu/CpuWorker.cpp

@@ -36,7 +36,6 @@
 #include "crypto/rx/RxCache.h"
 #include "crypto/rx/RxDataset.h"
 #include "crypto/rx/RxVm.h"
-#include "crypto/ghostrider/ghostrider.h"
 #include "net/JobResults.h"
 
 
@@ -45,6 +44,11 @@
 #endif
 
 
+#ifdef XMRIG_ALGO_GHOSTRIDER
+#include "crypto/ghostrider/ghostrider.h"
+#endif
+
+
 #ifdef XMRIG_FEATURE_BENCHMARK
 #   include "backend/common/benchmark/BenchState.h"
 #endif

src/backend/cpu/platform/HwlocCpuInfo.cpp

@@ -320,11 +320,16 @@ void xmrig::HwlocCpuInfo::processTopLevelCache(hwloc_obj_t cache, const Algorith
             L2 += l2->attr->cache.size;
             L2_associativity = l2->attr->cache.associativity;
 
-            if (L3_exclusive && l2->attr->cache.size >= scratchpad) {
+            if (L3_exclusive) {
+                if (vendor() == VENDOR_AMD) {
+                    extra += std::min<size_t>(l2->attr->cache.size, scratchpad);
+                }
+                else if (l2->attr->cache.size >= scratchpad) {
                     extra += scratchpad;
                 }
             }
         }
+    }
 
     // This code is supposed to run only on Intel CPUs
     if ((vendor() == VENDOR_INTEL) && (scratchpad == 2 * oneMiB)) {