Skip to content

Add amdgpu tests to benchmark SDMA, read wavefront states and Trap registers. - #1

Open
samikhawaja wants to merge 6 commits into
AMD-CloudGPU:2019_Q3_RC6_Releasefrom
samikhawaja:2019_Q3_RC6_Release
Open

Add amdgpu tests to benchmark SDMA, read wavefront states and Trap registers.#1
samikhawaja wants to merge 6 commits into
AMD-CloudGPU:2019_Q3_RC6_Releasefrom
samikhawaja:2019_Q3_RC6_Release

Conversation

@samikhawaja

Copy link
Copy Markdown

Add following new amdgpu tests,

  • Benchmarking SDMA memcpy from System RAM to VRAM.
  • Benchmarking SDMA memcpy from VRAM to System RAM.
  • Read wavefront states.
  • Read TBA/TMA registers visible from the shader code.

Use the 'SDMA_OPCODE_COPY' to copy a 512MB buffer from System RAM to a
buffer of same size in VRAM. Measure time it takes to complete this
operation and calculate the transfer speed.
Carve out a 'SDMA copy' helper function from 'System RAM to VRAM copy
test' and use it to add a test to benchmark the memory copy from VRAM to
System RAM using SDMA.
Use the amdgpu kernel debugfs to read the wave states. Only support of
AI Family AMD gpus is available.
Use the 'gpr' file in amdgpu kernel debugfs to read the sgpr and vgpr
registers when a compute unit is active.
Dispatch a shader program with a long loop and read wavefront states.
Read the TBA and TMA registers in a shader and write the values into a
buffer. Verify that the values are as set in the kernel mode driver.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant