tune VEGA auto suggestion

- allow to use up to 2024 threads on VEGA gpus

fix #5

1 files changed, 15 insertions, 3 deletions

diff --git a/xmrstak/backend/amd/autoAdjust.hpp b/xmrstak/backend/amd/autoAdjust.hpp
index 8f9aa50..41dd11a 100644
--- a/xmrstak/backend/amd/autoAdjust.hpp
+++ b/xmrstak/backend/amd/autoAdjust.hpp
@@ -97,13 +97,25 @@ private:
         int i = 0;
         for(auto& ctx : devVec)
         {
+			/* 1000 is a magic selected limit, the reason is that more than 2GiB memory
+			 * sowing down the memory performance because of TLB cache misses
+			 */
+			size_t maxThreads = 1000u;
+			if(ctx.name.compare("gfx901") == 0)
+			{
+				/* Increase the number of threads for AMD VEGA gpus.
+				 * Limit the number of threads based on the issue: https://github.com/fireice-uk/xmr-stak/issues/5#issuecomment-339425089
+				 * to avoid out of memory errors
+				 */
+				maxThreads = 2024u;
+			}
+
 			// keep 128MiB memory free (value is randomly chosen)
 			size_t availableMem = ctx.freeMem - (128u * byteToMiB);
 			// 224byte extra memory is used per thread for meta data
 			size_t perThread = hashMemSize + 224u;
-			size_t max_intensity = availableMem / perThread;
-			// 1000 is a magic selected limit \todo select max intensity depending of the gpu type
-			size_t possibleIntensity = std::min( size_t(1000u) , max_intensity );
+			size_t maxIntensity = availableMem / perThread;
+			size_t possibleIntensity = std::min( maxThreads , maxIntensity );
 			// map intensity to a multiple of the compute unit count, 8 is the number of threads per work group
 			size_t intensity = (possibleIntensity / (8 * ctx.computeUnits)) * ctx.computeUnits * 8;
 			conf += std::string("  // gpu: ") + ctx.name + " memory:" + std::to_string(availableMem / byteToMiB) + "\n";