rocm-7.1.1

ROCm release v7.1.1

TransferBench v1.65.00

v1.65.00

Added

• Added warp-level dispatch support via GFX_SE_TYPE environment variable
  • GFX_SE_TYPE=0 (default): Threadblock-level dispatch, each subexecutor is a threadblock
  • GFX_SE_TYPE=1: Warp-level dispatch, each subexecutor is a single warp

rocm-7.1.0

ROCm release v7.1.0

rocm-7.0.2

ROCm release v7.0.2

rocm-6.4.4

ROCm release v6.4.4

rocm-7.0.1

ROCm release v7.0.1

rocm-7.0.0

ROCm release v7.0.0

TransferBench v1.64.00

v1.64.00

Added

• Added BLOCKSIZES to a2asweep preset to allow also sweeping over threadblock sizes
• Added FILL_COMPRESS to allow more control over input data pattern
  • FILL_COMPRESS takes in a comma-separated list of integer percentages (that must add up to 100)
    that sets the percentages of 64B lines to be filled by random/1B0/2B0/4B0/32B0 data patterns
    • Bins:
      • 0 - random
      • 1 - 1B0 upper 1 byte of each aligned 2 bytes is 0
      • 2 - 2B0 upper 2 bytes of each aligned 4 bytes is 0
      • 3 - 4B0 upper 4 bytes of each aligned 8 bytes is 0
      • 4 - 32B0 upper 32 bytes of each aligned 64-byte line are 0
  • FILL_PATTERN will be ignored if FILL_COMPRESS is specified
• Additional details about data patterns generated will be printed if the debug env var DUMP_LINES is
  set to a non-zero value, which also corresponds to how many 64 byte lines will be printed

Modified

• Increased GFX_BLOCKSIZE limit from 512 to 1024 (still requires multiple of 64)

Fixed

• Fixed bug when using BYTE_OFFSET

TransferBench v1.63.00

v1.63.00

Added

• Added gfx950, gfx1150, and gfx1151 to default GPU targets list in CMake builds

Modified

• Removing self-GPU check for DMA engine copies
• Switched to amdclang++ as primary compiler
• healthcheck preset adds HBM testing and support for more MI3XX variants

Fixed

• Fixed issue when using "P" memory type and specific DMA subengines
• Fixed issue with subiteration timing reports

rocm-6.4.3

ROCm release v6.4.3