Fixing specific DMA engine transfers, enabling GFX_SINGLE_TEAM=1 by default (#166)

CHANGELOG.md

@@ -3,10 +3,17 @@
 Documentation for TransferBench is available at
 [https://rocm.docs.amd.com/projects/TransferBench](https://rocm.docs.amd.com/projects/TransferBench).
 
+## v1.60.00
+### Modified
+- Reverted GFX_SINGLE_TEAM default back to 1
+
+### Fixed
+- Fixed bug where peer memory access was not enabled for DMA transfers, which would break specific DMA engine transfers
+
 ## v1.59.01
 ### Added
 - The a2a preset A2A_MODE variable has been enhanced to allow for customizing the number of srcs/dsts to use
-  This is specified by setting A2A_MODE to numSrcs:numDsts.  Extra destinations past 1 will be "local" writes (i.e. if one sets A2A_MODE=1:3, then transfers will follow this pattern: Fx Gx FyFxFx) 
+  This is specified by setting A2A_MODE to numSrcs:numDsts.  Extra destinations past 1 will be "local" writes (i.e. if one sets A2A_MODE=1:3, then transfers will follow this pattern: Fx Gx FyFxFx)
   to simulate similar conditions normally used during collective algorithms such as ring-based AllReduce
 
 ## v1.59.00

src/client/Client.hpp

@@ -23,7 +23,7 @@ THE SOFTWARE.
 #pragma once
 
 // TransferBench client version
-#define CLIENT_VERSION "01"
+#define CLIENT_VERSION "00"
 
 #include "TransferBench.hpp"
 #include "EnvVars.hpp"

src/client/EnvVars.hpp

@@ -136,7 +136,7 @@ public:
     blockBytes        = GetEnvVar("BLOCK_BYTES"         , 256);
     byteOffset        = GetEnvVar("BYTE_OFFSET"         , 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);

src/header/TransferBench.hpp

@@ -64,7 +64,7 @@ namespace TransferBench
   using std::set;
   using std::vector;
 
-  constexpr char VERSION[] = "1.59";
+  constexpr char VERSION[] = "1.60";
 
   /**
    * Enumeration of supported Executor types
@@ -2297,7 +2297,7 @@ namespace {
         MemDevice const& srcMemDevice = t.srcs[iSrc];
 
         // Ensure executing GPU can access source memory
-        if (exeDevice.exeType == EXE_GPU_GFX && IsGpuMemType(srcMemDevice.memType) &&
+        if (IsGpuExeType(exeDevice.exeType) && IsGpuMemType(srcMemDevice.memType) &&
             srcMemDevice.memIndex != exeDevice.exeIndex) {
           ERR_CHECK(EnablePeerAccess(exeDevice.exeIndex, srcMemDevice.memIndex));
         }
@@ -2310,7 +2310,7 @@ namespace {
         MemDevice const& dstMemDevice = t.dsts[iDst];
 
         // Ensure executing GPU can access destination memory
-        if (exeDevice.exeType == EXE_GPU_GFX && IsGpuMemType(dstMemDevice.memType) &&
+        if (IsGpuExeType(exeDevice.exeType) && IsGpuMemType(dstMemDevice.memType) &&
             dstMemDevice.memIndex != exeDevice.exeIndex) {
           ERR_CHECK(EnablePeerAccess(exeDevice.exeIndex, dstMemDevice.memIndex));
         }