TransferBench v1.57.00

v1.57.00

Modified

• Removing use of default starship operator / C++20 requirement to enable compilation of more OSs
• Changing how version is reported. Client version is now just last two digits, and increments only if
  no changes are made to the backend header-only library file, and resets to 0 when header is updated
• GFX_SINGLE_TEAM=0 is set by default

TransferBench v1.56

v1.56

Fixed

• Fixed bug when using interactive mode. Interactive mode now starts prior to all warmup iterations

TransferBench v1.55

v1.55

Fixed

• Fixed missing header error when compiling on CentOS
• Fixed issues when using multi-stream mode for GFX executor

TransferBench v1.54

v1.54

Modified

• Refactored TransferBench into a header-only library combined with a thin client to facilitate the
  use of TransferBench as the backend for other applications
• Optimized how data validation is handled - this should speed up Tests with many parallel transfers as data is only
  generated once
• Preset benchmarks now no longer take in any extra command line arguments. Preset settings are only accessed via
  environment variables. Details for each preset are printed
• The a2a preset benchmark now defaults to using fine-grained memory and GFX unroll of 2
• Refactored how Transfers are launched in parallel which has reduced some CPU-side overheads
• CPU and DMA executor timing now use CPU wall clock timing instead of slowest Transfer time

Added

• New one2all preset which sweeps over all subests of parallel transfers from one GPU to others
• Adding new warnings for DMA execution relating to how HIP will default to using agents from the source memory

Removed

• CU scaling preset has been removed. Similar functionality already exists in the schmoo preset benchmark
• Preparation of source data via GFX kernel has been removed (USE_PREP_KERNEL)
• Removed GFX block-reordering (BLOCK_ORDER)
• Removed NUM_CPU_DEVICES and NUM_GPU_DEVICES from common env vars and only into the presets they apply to.
• Removed SHARED_MEM_BYTES option for GFX executor
• Removed USE_PCIE_INDEX, and SHARED_MEM_BYTES

Fixed

• Fixed a potential timing reporting issue when DMA executed Transfers end up getting serialized.

TransferBench v1.53

v1.53

Added

• Added ability to specify NULL for sweep preset as source or destination memory type

TransferBench v1.52

Added

• Added USE_HSA_DMA env var to switch to using hsa_amd_memory_async_copy instead of hipMemcpyAsync for DMA execution
• Added ability to set USE_GPU_DMA env var for a2a benchmark
• Adding check for large BAR enablement for GPU devices during topology check

Fixed

• Potential memory leak if HSA reports 0 hops between GPUs and CPUs

TransferBench v1.51

v1.51

Modified

• CSV output has been modified slightly to match normal terminal output
• Output for non single stream mode has been changed to match single stream mode (results per Executor)

Added

• Support for sub-iterations via NUM_SUBITERATIONS. This allows for additional looping during an iteration
  If set to 0, this should infinitely loop (which may be useful for some debug purposes)
• Support for variable number of subexecutors (currently for GPU-GFX executor only). Setting subExecutors to
  0 will run over a range of CUs to use, and report only the results of the best one found. This can be tuned
  for performance by setting the MIN_VAR_SUBEXEC and MAX_VAR_SUBEXEC environment variables to narrow the
  search space. The number of CUs used will be identical for all variable subExecutor transfers
• Experimental new "healthcheck" preset config which currently only supports MI300 series. This preset runs
  through CPU to GPU bandwidth tests and all-to-all XGMI bandwidth tests and compares against expected values
  Pass criteria limits can be modified (due to platform differences) via the environment variables
  LIMIT_UDIR (undirectional), LIMIT_BDIR (bidirectional), and LIMIT_A2A (Per GPU-GPU link bandwidth)

Fixed

• Fixed out-of-bounds memory access during topology detection that can happen if the number of
  CPUs is less than the number of NUMA domains
• Fixed CU masking functionality on multi-XCD architectures (e.g. MI300)

TransferBench v1.50

Added

• Adding new parallel copy preset benchmark (pcopy)
  • Usage: ./TransferBench pcopy <numBytes=64M> <#CUs=8> <srcGpu=0> <minGpus=1> <maxGpus=#GPU-1>

Fixed

• Removed non-copies DMA Transfers (this had previously been using hipMemset)
• Fixed CPU executor when operating on null destination

TransferBench v1.49

Fixes

• Enumerating previously missed DMA engines used only for CPU traffic in topology display

TransferBench v1.48

v1.48

Fixes

• Various fixes for TransferBenchCuda

Additions

• Support for targeting specific DMA engines via executor subindex (e.g. D0.1)
• Printing warnings when exeuctors are overcommited

Modifications

• USE_REMOTE_READ supported for rwrite preset benchmark