TransferBench v1.71.00

.github/workflows/README_BUILD_PACKAGES.md

@@ -70,9 +70,9 @@ sudo -E BUILD_TYPE=Debug ./build_packages_local.sh
 After the script completes, packages live under `build/`:
 
 ```
-build/amdrocm7-transferbench_1.66.02-<release>_amd64.deb
-build/amdrocm7-transferbench-1.66.02-<release>.x86_64.rpm
-build/amdrocm7-transferbench-1.66.02-Linux.tar.gz
+build/amdrocm7-transferbench_1.71.00-<release>_amd64.deb
+build/amdrocm7-transferbench-1.71.00-<release>.x86_64.rpm
+build/amdrocm7-transferbench-1.71.00-Linux.tar.gz
 ```
 
 ## Installing built packages

CHANGELOG.md

@@ -3,6 +3,34 @@
 Documentation for TransferBench is available at
 [https://rocm.docs.amd.com/projects/TransferBench](https://rocm.docs.amd.com/projects/TransferBench).
 
+## v1.71.00
+### Added
+- Added PingPong operations to measure latency between two GPU Executors. A PingPong joins a Ping and a Pong triplet
+  with `+` and an optional lap count, e.g. `(N->G0->F1)+1000(N->G1->F0)`. Each lap, the Ping Executor writes a
+  1-byte flag to its DST and waits for the Pong Executor to write one back to the Pong DST. Results report the
+  GPU-timed round-trip latency per lap
+  - PingPongs can run in parallel with regular Transfers in the same Test, and do not count toward Executor bandwidth
+  - Each Ping and Pong uses a single threadblock; the number of SubExecutors and bytes to transfer are ignored
+  - Currently supported on GFX Executors only
+  - Ping and Pong may run on different ranks within the same pod (requires pod communication support)
+- Added `PINGPONG_FLAG_BUFFER` and `PINGPONG_STRIDE` to spread PingPong flags across a buffer, moving by the stride each lap
+- Added latency presets built on PingPong operations:
+  - "p2p_latency": latency between pairs of GPU Executors, run serially
+  - "one2all_latency": latency from one GPU Executor to all others, run in parallel
+  - "a2a_latency": latency between all pairs of GPU Executors, run in parallel
+  - Behavior can be modified with `NUM_LAPS`, `GPU_MEM_TYPE`, `USE_REMOTE_READ`, `NUM_GPU_DEVICES` and, for a2a_latency, `A2A_LOCAL`
+- Added gfx1250-strict to the default GPU targets for CMake builds
+- NIC topology output now shows each NIC's IBV max_msg_sz
+### Modified
+- When GFX Executor timing does not use HIP events (multi-stream mode or `USE_HIP_EVENTS=0`), it now spans from the first
+  Transfer launch to the last Transfer completion, excluding executor dispatch overhead
+### Fixed
+- CU IDs reported with `SHOW_ITERATIONS` now match `CU_MASK` indices on gfx90a, gfx942, gfx950 and gfx1250
+- GFX kernel launch failures are now reported instead of being ignored
+- RoCE / GID index fields are now initialized for NICs that are not RoCE or have no active port
+- NIC Transfers are now rejected during validation if a queue pair's work request (the smaller of `NIC_CHUNK_BYTES` and
+  the bytes assigned to that queue pair) exceeds the IBV max_msg_sz of either NIC
+
 ## v1.70.02
 ### Modified
 - rings preset defaults `NUM_SUB_EXEC` to 0, which uses all available subexecutors per Transfer

CMakeLists.txt

@@ -103,8 +103,8 @@ set(ENV{ROCM_PATH} "${ROCM_PATH}")
 # TransferBench project definitions
 #==================================================================================================
 set(TRANSFERBENCH_VERSION_MAJOR 1)
-set(TRANSFERBENCH_VERSION_MINOR 70)
-set(TRANSFERBENCH_VERSION_PATCH_FALLBACK "02")
+set(TRANSFERBENCH_VERSION_MINOR 71)
+set(TRANSFERBENCH_VERSION_PATCH_FALLBACK "00")
 
 # Auto-compute patch from git: count commits since the last v<MAJOR>.<MINOR>.* tag.
 # Falls back to TRANSFERBENCH_VERSION_PATCH_FALLBACK when git is unavailable,
@@ -220,7 +220,8 @@ set(DEFAULT_GPUS
       gfx1151
       gfx1200
       gfx1201
-      gfx1250)
+      gfx1250
+      gfx1250-strict)
 
 if(BUILD_LOCAL_GPU_TARGET_ONLY)
   message(STATUS "Building only for local GPU target")

docs/install/build_from_source.rst

@@ -317,7 +317,7 @@ To modify the CMake behavior, use the following environment variables:
               <td rowspan="3"><b>CMake cache variables</b></td>
               <td><code>GPU_TARGETS</code></td>
               <td>Semicolon-separated GPU architectures. Overridden if <code>BUILD_LOCAL_GPU_TARGET_ONLY</code> is <code>ON</code></td>
-              <td><code style="word-break: break-all;">gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1200;gfx1201;gfx1250</code></td>
+              <td><code style="word-break: break-all;">gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1200;gfx1201;gfx1250;gfx1250-strict</code></td>
               </tr>
               <tr>
                 <td><code>AMD_SMI_EXECUTABLE</code></td>

src/client/Client.cpp

@@ -107,16 +107,24 @@ int main(int argc, char **argv)
     if (transfers.empty()) {
       Print("<none>\n");
     } else {
-      bool isMultiNode = GetNumRanks() > 1;
       for (size_t i = 0; i < transfers.size(); i++) {
         Transfer const& t = transfers[i];
-        Print("Transfer %5lu: (%s->", i, MemDevicesToStr(t.srcs).c_str());
-        if (isMultiNode) Print("R%d", t.exeDevice.exeRank);
-        Print("%c%d", ExeTypeStr[t.exeDevice.exeType], t.exeDevice.exeIndex);
-        if (t.exeDevice.exeSlot) Print("%c", 'A' + t.exeDevice.exeSlot);
-        if (t.exeSubIndex != -1) Print(".%d", t.exeSubIndex);
-        if (t.exeSubSlot != 0) Print("%c", 'A' + t.exeSubSlot);
-        Print("->%s)\n", MemDevicesToStr(t.dsts).c_str());
+        if (t.numLaps > 0) {
+          Print("Transfer %5lu: PingPong +%d: (%s->%s->%s) <+> (%s->%s->%s)\n",
+                i, t.numLaps,
+                MemDeviceToStr(t.srcs[0]).c_str(),
+                ExeDeviceToStr(t.exeDevice, t.exeSubIndex, t.exeSubSlot).c_str(),
+                MemDeviceToStr(t.dsts[0]).c_str(),
+                MemDeviceToStr(t.srcs[1]).c_str(),
+                ExeDeviceToStr(t.exeDevicePong, t.exeSubIndexPong, t.exeSubSlotPong).c_str(),
+                MemDeviceToStr(t.dsts[1]).c_str());
+        } else {
+          Print("Transfer %5lu: (%s->%s->%s)\n",
+                i,
+                MemDevicesToStr(t.srcs).c_str(),
+                ExeDeviceToStr(t.exeDevice, t.exeSubIndex, t.exeSubSlot).c_str(),
+                MemDevicesToStr(t.dsts).c_str());
+        }
       }
     }
     return 0;

src/client/EnvVars.hpp

@@ -44,7 +44,7 @@ THE SOFTWARE.
 #include <random>
 #include <time.h>
 
-#define CLIENT_VERSION "02"
+#define CLIENT_VERSION "00"
 
 #include "TransferBench.hpp"
 using namespace TransferBench;
@@ -74,6 +74,8 @@ class EnvVars
   int numIterations;                 // Number of timed iterations to perform.  If negative, run for -numIterations seconds instead
   int numSubIterations;              // Number of subiterations to perform
   int numWarmups;                    // Number of un-timed warmup iterations to perform
+  int pingpongFlagBuffer;            // Size in bytes of the pingpong flag buffer (must be positive)
+  int pingpongStride;                // Stride in bytes between flag slots for pingpong laps (wraps within the flag buffer)
   int showBorders;                   // Show ASCII box-drawing characaters in tables
   int showIterations;                // Show per-iteration timing info
   int useInteractive;                // Pause for user-input before starting transfer loop
@@ -158,55 +160,57 @@ class EnvVars
     else if (archName == "gfx950")  defaultGfxUnroll = 4;
     else if (archName == "gfx1250") defaultGfxUnroll = 32;
 
-    alwaysValidate    = GetEnvVar("ALWAYS_VALIDATE", 0);
-    blockBytes        = GetEnvVar("BLOCK_BYTES"         , 256);
-    byteOffset        = GetEnvVar("BYTE_OFFSET"         , 0);
-    fillCompress      = GetEnvVarArray("FILL_COMPRESS"  , {});
-    gfxBlockOrder     = GetEnvVar("GFX_BLOCK_ORDER"     , 0);
-    gfxBlockSize      = GetEnvVar("GFX_BLOCK_SIZE"      , 256);
-    gfxKernel         = GetEnvVar("GFX_KERNEL"          , 0);
-    gfxSeType         = GetEnvVar("GFX_SE_TYPE"         , 0);
-    gfxSingleTeam     = GetEnvVar("GFX_SINGLE_TEAM"     , 0);
-    gfxTemporal       = GetEnvVar("GFX_TEMPORAL"        , 0);
-    gfxUnroll         = GetEnvVar("GFX_UNROLL"          , defaultGfxUnroll);
-    gfxWaveOrder      = GetEnvVar("GFX_WAVE_ORDER"      , 0);
-    gfxWordSize       = GetEnvVar("GFX_WORD_SIZE"       , 4);
-    hideEnv           = GetEnvVar("HIDE_ENV"            , 0);
-    minNumVarSubExec  = GetEnvVar("MIN_VAR_SUBEXEC"     , 1);
-    maxNumVarSubExec  = GetEnvVar("MAX_VAR_SUBEXEC"     , 0);
-    numIterations     = GetEnvVar("NUM_ITERATIONS"      , 10);
-    numSubIterations  = GetEnvVar("NUM_SUBITERATIONS"   , 1);
-    numWarmups        = GetEnvVar("NUM_WARMUPS"         , 3);
-    outputToCsv       = GetEnvVar("OUTPUT_TO_CSV"       , 0);
-    samplingFactor    = GetEnvVar("SAMPLING_FACTOR"     , 1);
-    showBorders       = GetEnvVar("SHOW_BORDERS"        , 1);
-    showIterations    = GetEnvVar("SHOW_ITERATIONS"     , 0);
-    showPercentiles   = GetEnvVarArray("SHOW_PERCENTILES", {});
-    sweepMaxPow2      = GetEnvVar("SWEEP_MAX_POW2"      , 29);
-    sweepMinPow2      = GetEnvVar("SWEEP_MIN_POW2"      , 10);
-    sweepMinPow2      = std::clamp(sweepMinPow2, 0, 62);
-    sweepMaxPow2      = std::clamp(sweepMaxPow2, 0, 62);
+    alwaysValidate     = GetEnvVar("ALWAYS_VALIDATE", 0);
+    blockBytes         = GetEnvVar("BLOCK_BYTES"         , 256);
+    byteOffset         = GetEnvVar("BYTE_OFFSET"         , 0);
+    fillCompress       = GetEnvVarArray("FILL_COMPRESS"  , {});
+    gfxBlockOrder      = GetEnvVar("GFX_BLOCK_ORDER"     , 0);
+    gfxBlockSize       = GetEnvVar("GFX_BLOCK_SIZE"      , 256);
+    gfxKernel          = GetEnvVar("GFX_KERNEL"          , 0);
+    gfxSeType          = GetEnvVar("GFX_SE_TYPE"         , 0);
+    gfxSingleTeam      = GetEnvVar("GFX_SINGLE_TEAM"     , 0);
+    gfxTemporal        = GetEnvVar("GFX_TEMPORAL"        , 0);
+    gfxUnroll          = GetEnvVar("GFX_UNROLL"          , defaultGfxUnroll);
+    gfxWaveOrder       = GetEnvVar("GFX_WAVE_ORDER"      , 0);
+    gfxWordSize        = GetEnvVar("GFX_WORD_SIZE"       , 4);
+    hideEnv            = GetEnvVar("HIDE_ENV"            , 0);
+    minNumVarSubExec   = GetEnvVar("MIN_VAR_SUBEXEC"     , 1);
+    maxNumVarSubExec   = GetEnvVar("MAX_VAR_SUBEXEC"     , 0);
+    numIterations      = GetEnvVar("NUM_ITERATIONS"      , 10);
+    numSubIterations   = GetEnvVar("NUM_SUBITERATIONS"   , 1);
+    numWarmups         = GetEnvVar("NUM_WARMUPS"         , 3);
+    pingpongFlagBuffer = GetEnvVar("PINGPONG_FLAG_BUFFER", 1);
+    pingpongStride     = GetEnvVar("PINGPONG_STRIDE"     , 1);
+    outputToCsv        = GetEnvVar("OUTPUT_TO_CSV"       , 0);
+    samplingFactor     = GetEnvVar("SAMPLING_FACTOR"     , 1);
+    showBorders        = GetEnvVar("SHOW_BORDERS"        , 1);
+    showIterations     = GetEnvVar("SHOW_ITERATIONS"     , 0);
+    showPercentiles    = GetEnvVarArray("SHOW_PERCENTILES", {});
+    sweepMaxPow2       = GetEnvVar("SWEEP_MAX_POW2"      , 29);
+    sweepMinPow2       = GetEnvVar("SWEEP_MIN_POW2"      , 10);
+    sweepMinPow2       = std::clamp(sweepMinPow2, 0, 62);
+    sweepMaxPow2       = std::clamp(sweepMaxPow2, 0, 62);
     if (sweepMinPow2 > sweepMaxPow2) std::swap(sweepMinPow2, sweepMaxPow2);
-    tdmBlockOrder     = GetEnvVar("TDM_BLOCK_ORDER"     , 0);
-    tdmBlockSize      = GetEnvVar("TDM_BLOCK_SIZE"      , 256);
-    tdmLdsBytes       = GetEnvVar("TDM_LDS_BYTES"       , 0);
-    useHipEvents      = GetEnvVar("USE_HIP_EVENTS"      , 1);
-    useHsaDma         = GetEnvVar("USE_HSA_DMA"         , 0);
-    useInteractive    = GetEnvVar("USE_INTERACTIVE"     , 0);
-    useSingleStream   = GetEnvVar("USE_SINGLE_STREAM"   , 1);
-    validateDirect    = GetEnvVar("VALIDATE_DIRECT"     , 0);
-    validateOnDevice  = GetEnvVar("VALIDATE_ON_DEVICE"  , 0);
-    validateSource    = GetEnvVar("VALIDATE_SOURCE"     , 0);
-
-    ibGidIndex        = GetEnvVar("IB_GID_INDEX"        ,-1);
-    ibPort            = GetEnvVar("IB_PORT_NUMBER"      , 1);
-    roceVersion       = GetEnvVar("ROCE_VERSION"        , 2);
-    ipAddressFamily   = GetEnvVar("IP_ADDRESS_FAMILY"   , 4);
-    nicChunkBytes     = GetEnvVar("NIC_CHUNK_BYTES"     , 1073741824);
-    nicCqPollBatch    = GetEnvVar("NIC_CQ_POLL_BATCH"   , 4);
-    nicRelaxedOrder   = GetEnvVar("NIC_RELAX_ORDER"     , 1);
-    nicServiceLevel   = GetEnvVar("NIC_SERVICE_LEVEL"   , 0);
-    nicTrafficClass   = GetEnvVar("NIC_TRAFFIC_CLASS"   , 0);
+    tdmBlockOrder      = GetEnvVar("TDM_BLOCK_ORDER"     , 0);
+    tdmBlockSize       = GetEnvVar("TDM_BLOCK_SIZE"      , 256);
+    tdmLdsBytes        = GetEnvVar("TDM_LDS_BYTES"       , 0);
+    useHipEvents       = GetEnvVar("USE_HIP_EVENTS"      , 1);
+    useHsaDma          = GetEnvVar("USE_HSA_DMA"         , 0);
+    useInteractive     = GetEnvVar("USE_INTERACTIVE"     , 0);
+    useSingleStream    = GetEnvVar("USE_SINGLE_STREAM"   , 1);
+    validateDirect     = GetEnvVar("VALIDATE_DIRECT"     , 0);
+    validateOnDevice   = GetEnvVar("VALIDATE_ON_DEVICE"  , 0);
+    validateSource     = GetEnvVar("VALIDATE_SOURCE"     , 0);
+
+    ibGidIndex         = GetEnvVar("IB_GID_INDEX"        ,-1);
+    ibPort             = GetEnvVar("IB_PORT_NUMBER"      , 1);
+    roceVersion        = GetEnvVar("ROCE_VERSION"        , 2);
+    ipAddressFamily    = GetEnvVar("IP_ADDRESS_FAMILY"   , 4);
+    nicChunkBytes      = GetEnvVar("NIC_CHUNK_BYTES"     , 1073741824);
+    nicCqPollBatch     = GetEnvVar("NIC_CQ_POLL_BATCH"   , 4);
+    nicRelaxedOrder    = GetEnvVar("NIC_RELAX_ORDER"     , 1);
+    nicServiceLevel    = GetEnvVar("NIC_SERVICE_LEVEL"   , 0);
+    nicTrafficClass    = GetEnvVar("NIC_TRAFFIC_CLASS"   , 0);
 
     // Check that NIC service level and traffic class are in valid ranges
     if (nicServiceLevel < 0 || nicServiceLevel > 15) {
@@ -404,6 +408,8 @@ class EnvVars
     printf(" NUM_SUBITERATIONS   - # of sub-iterations to run per iteration. Must be non-negative\n");
     printf(" NUM_WARMUPS         - # of untimed warmup iterations per test\n");
     printf(" OUTPUT_TO_CSV       - Outputs to CSV format if set\n");
+    printf(" PINGPONG_FLAG_BUFFER - Size in bytes of the pingpong flag buffer (default 1, must be positive). Each flag slot is 1 byte\n");
+    printf(" PINGPONG_STRIDE     - Stride in bytes between flag slots for pingpong laps (default 1, must be positive unless the flag buffer is 1 byte, wraps within the flag buffer)\n");
     printf(" SAMPLING_FACTOR     - Add this many samples (when possible) between powers of 2 when auto-generating data sizes\n");
     printf(" SHOW_BORDERS        - Show ASCII box-drawing characters in tables\n");
     printf(" SHOW_ITERATIONS     - Show per-iteration timing info\n");
@@ -555,6 +561,10 @@ class EnvVars
           "Running %s subiterations", (numSubIterations == 0 ? "infinite" : std::to_string(numSubIterations)).c_str());
     Print("NUM_WARMUPS", numWarmups,
           "Running %d warmup iteration(s) per Test", numWarmups);
+    Print("PINGPONG_FLAG_BUFFER", pingpongFlagBuffer,
+          "Pingpong flag buffer of %d bytes", pingpongFlagBuffer);
+    Print("PINGPONG_STRIDE", pingpongStride,
+          "Pingpong flag stride %d bytes per lap", pingpongStride);
     Print("SHOW_BORDERS", showBorders, "%s ASCII box-drawing characaters in tables", showBorders ? "Showing" : "Hiding");
     Print("SHOW_ITERATIONS", showIterations,
           "%s per-iteration timing", showIterations ? "Showing" : "Hiding");
@@ -738,6 +748,8 @@ class EnvVars
     cfg.general.numIterations      = numIterations;
     cfg.general.numSubIterations   = numSubIterations;
     cfg.general.numWarmups         = numWarmups;
+    cfg.general.pingpongFlagBuffer = pingpongFlagBuffer;
+    cfg.general.pingpongStride     = pingpongStride;
     cfg.general.recordPerIteration = ((showIterations != 0) || !showPercentiles.empty()) ? 1 : 0;
     cfg.general.useHipEvents       = useHipEvents;
     cfg.general.useInteractive     = useInteractive;