Adding Batched DMA support (hipMemcpyBatchAsync), and bmasweep preset

[gilbertlee-amd]

Motivation

This adds a new Executor (B) based on the hipMemcpyBatchAsync call that was introduced in HIP 7.0.
This new Executor supports SubExecutors - namely how many batches a single Transfer is broken into.
This allows comparing performance against standard hipMemcpyAsync.
A new bmasweep preset is also introduced to compare the two versions

Technical Details

This new code also enables more than one destination when using DMA and Batched DMA (BMA) Executor.
When multiple destinations are provided, the copies are performed one after another with DMA, and all as different batches (respecting number of SubExecutors) for Batched DMA.

Test Result

Here are results from bmasweep from a MI355X:

[BMA Sweep Related]
EXE_INDEX            =            0 : Executing on GPU 0
LOCAL_COPY           =            0 : Excluding local copy to GPU 0
GPU_MEM_TYPE         =            0 : Using default GPU (0=default, 1=fine-grained, 2=uncached, 3=managed)
NUM_GPU_DEVICES      =            8 : Using 8 GPUs
NUM_SUB_EXECS        =            4 : 1,2,4,8

Performing 7 simultaneous DMA Transfers from GPU 0 other GPUs
Executing: ...................
┌------------┬--------┬---------------------------------------┐
│      Bytes │    DMA │ BMA (1)   BMA (2)   BMA (4)   BMA (8) │
├------------┼--------┼---------------------------------------┤
│       4096 │   0.68 │    0.67      0.35      0.18      0.09 │
│       8192 │   1.33 │    1.34      0.70      0.36      0.18 │
│      16384 │   2.61 │    2.68      1.41      0.72      0.35 │
│      32768 │   5.42 │    5.36      2.98      1.49      0.76 │
│      65536 │  10.52 │   10.54      5.53      3.01      1.52 │
│     131072 │  20.18 │   19.84     11.18      5.63      3.04 │
│     262144 │  33.90 │   32.19     21.52     11.21      5.67 │
│     524288 │  44.86 │   43.72     34.54     21.98     11.40 │
│    1048576 │  50.01 │   49.64     45.43     35.59     21.95 │
│    2097152 │  47.99 │   48.07     50.96     46.41     35.78 │
│    4194304 │  53.64 │   53.66     48.59     51.27     46.71 │
│    8388608 │  57.09 │   57.07     53.65     48.10     51.54 │
│   16777216 │  59.11 │   59.11     57.22     53.66     48.16 │
│   33554432 │  60.10 │   59.97     59.11     57.10     53.74 │
│   67108864 │  57.39 │   60.65     60.11     59.10     57.18 │
│  134217728 │  60.95 │   60.82     60.67     60.13     59.13 │
│  268435456 │  61.10 │   61.10     60.83     60.64     60.07 │
│  536870912 │  61.19 │   59.33     61.10     60.95     60.54 │
│ 1073741824 │  61.22 │   61.22     61.18     60.41     60.73 │
└------------┴--------┴---------------------------------------┘
Reported numbers are all GB/s, normalized for per Transfer for 7 Transfers

[Copilot]

Pull request overview

This PR introduces a new GPU batched DMA executor (B, backed by hipMemcpyBatchAsync in HIP/ROCm 7.0+) and a new bmasweep preset to compare standard DMA vs batched DMA, while also extending the DMA path to support multiple destination buffers.

Changes:

• Add EXE_GPU_BDMA (“B”) executor support gated by HIP/ROCm 7.0+ (hipMemcpyBatchAsync).
• Allow DMA (and BMA) transfers to specify multiple destinations and execute the copies accordingly.
• Add a bmasweep preset and update docs/changelog to expose the new executor.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 9 comments.

Show a summary per file

File	Description
src/header/TransferBench.hpp	Adds BDMA executor, multi-destination DMA support, BDMA execution path, and related validation/topology updates.
src/client/Utilities.hpp	Adds string mapping for the new executor type.
src/client/Presets/Presets.hpp	Registers the new bmasweep preset.
src/client/Presets/BmaSweep.hpp	New preset to benchmark DMA vs batched DMA across multiple destinations.
examples/example.cfg	Documents the new executor in the example config.
CHANGELOG.md	Notes the new executor/preset and related DMA behavior changes.

Comments suppressed due to low confidence (1)

src/header/TransferBench.hpp:5834

• Wildcard expansion for executor subindices treats EXE_GPU_BDMA like GFX/DMA and iterates over GetNumExecutorSubIndices(). Since BDMA reports 0 subindices, this branch currently generates no transfers when exeSubIndices is -2, instead of recursing once with subindex -1. BDMA should likely be handled like CPU here (set -1 and recurse once).

      case EXE_GPU_GFX: case EXE_GPU_DMA: case EXE_GPU_BDMA:
      {
        // Iterate over all available subindices
        ExeDevice exeDevice = {wc.exe.exeType, wc.exe.exeIndices[0], wc.exe.exeRanks[0], 0};
        int numSubIndices = GetNumExecutorSubIndices(exeDevice);
        for (int x = 0; x < numSubIndices; x++) {
          wc.exe.exeSubIndices = {x};
          result |= RecursiveWildcardTransferExpansion(wc, baseRankIndex, numBytes, numSubExecs, transfers);
        }
        wc.exe.exeSubIndices = {-1};
        return result;

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 6 out of 6 changed files in this pull request and generated 4 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[alex-breslow-amd]

LGTM