v1.46 Fixing GFX_UNROLL for gfx906, enabling GFX_SINGLE_TEAM by default (#80)

CHANGELOG.md

@@ -3,6 +3,15 @@
 Documentation for TransferBench is available at
 [https://rocm.docs.amd.com/projects/TransferBench](https://rocm.docs.amd.com/projects/TransferBench).
 
+## v1.46
+
+### Fixes
+* Fixing GFX_UNROLL set to 13 (past 8) on gfx906 cards
+
+### Modifications
+* GFX_SINGLE_TEAM=1 by default
+* Adding field showing summation of individual Transfer bandwidths for Executors
+
 ## v1.45
 
 ### Additions

src/TransferBench.cpp

@@ -597,18 +597,24 @@ void ExecuteTransfers(EnvVars const& ev,
       double exeBandwidthGbs = (exeInfo.totalBytes / 1.0E9) / exeDurationMsec * 1000.0f;
       maxGpuTime = std::max(maxGpuTime, exeDurationMsec);
 
+      double sumBandwidthGbs = 0.0;
+      for (auto& transfer: exeInfo.transfers)
+      {
+        transfer->transferTime /= (1.0 * numTimedIterations);
+        transfer->transferBandwidth = (transfer->numBytesActual / 1.0E9) / transfer->transferTime * 1000.0f;
+        transfer->executorBandwidth = exeBandwidthGbs;
+        sumBandwidthGbs += transfer->transferBandwidth;
+      }
+
       if (verbose && !ev.outputToCsv)
       {
-        printf(" Executor: %3s %02d | %7.3f GB/s | %8.3f ms | %12lu bytes\n",
-               ExeTypeName[exeType], exeIndex, exeBandwidthGbs, exeDurationMsec, exeInfo.totalBytes);
+        printf(" Executor: %3s %02d | %7.3f GB/s | %8.3f ms | %12lu bytes | %-7.3f GB/s (sum)\n",
+               ExeTypeName[exeType], exeIndex, exeBandwidthGbs, exeDurationMsec, exeInfo.totalBytes, sumBandwidthGbs);
       }
 
       int totalCUs = 0;
       for (auto const& transfer : exeInfo.transfers)
       {
-        transfer->transferTime /= (1.0 * numTimedIterations);
-        transfer->transferBandwidth = (transfer->numBytesActual / 1.0E9) / transfer->transferTime * 1000.0f;
-        transfer->executorBandwidth = exeBandwidthGbs;
         totalCUs += transfer->numSubExecs;
 
         char exeSubIndexStr[32] = "";

src/include/EnvVars.hpp

@@ -29,7 +29,7 @@ THE SOFTWARE.
 #include "Compatibility.hpp"
 #include "Kernels.hpp"
 
-#define TB_VERSION "1.45"
+#define TB_VERSION "1.46"
 
 extern char const MemTypeStr[];
 extern char const ExeTypeStr[];
@@ -171,8 +171,8 @@ public:
     // Different hardware pick different GPU kernels
     // This performance difference is generally only noticable when executing fewer CUs
     int defaultGfxUnroll = 4;
-    if      (archName == "gfx906") defaultGfxUnroll = 13;
-    else if (archName == "gfx90a") defaultGfxUnroll = 9;
+    if      (archName == "gfx906") defaultGfxUnroll = 8;
+    else if (archName == "gfx90a") defaultGfxUnroll = 8;
     else if (archName == "gfx940") defaultGfxUnroll = 6;
     else if (archName == "gfx941") defaultGfxUnroll = 6;
     else if (archName == "gfx942") defaultGfxUnroll = 4;
@@ -183,7 +183,7 @@ public:
     byteOffset        = GetEnvVar("BYTE_OFFSET"         , 0);
     continueOnError   = GetEnvVar("CONTINUE_ON_ERROR"   , 0);
     gfxBlockSize      = GetEnvVar("GFX_BLOCK_SIZE"      , 256);
-    gfxSingleTeam     = GetEnvVar("GFX_SINGLE_TEAM"     , 0);
+    gfxSingleTeam     = GetEnvVar("GFX_SINGLE_TEAM"     , 1);
     gfxUnroll         = GetEnvVar("GFX_UNROLL"          , defaultGfxUnroll);
     gfxWaveOrder      = GetEnvVar("GFX_WAVE_ORDER"      , 0);
     hideEnv           = GetEnvVar("HIDE_ENV"            , 0);
@@ -509,7 +509,7 @@ public:
 
     if (gfxUnroll < 1 || gfxUnroll > MAX_UNROLL)
     {
-      printf("[ERROR] GFX kernel unroll factor must be between 1 and %d\n", MAX_UNROLL);
+      printf("[ERROR] GFX kernel unroll factor must be between 1 and %d (Not %d)\n", MAX_UNROLL, gfxUnroll);
       exit(1);
     }