Open tt-vjovanovic opened 1 month ago
A good candidate to write a test for.
Not urgent since there is a workaround of splitting the data into chunks up to 1KB.
@tt-vjovanovic can you share a git commit hash of your repo that demonstrates the problem?
Metal has this defined const uint32_t MAX_BLOCK_SIZE = (0x1 << 10); // Max 1024 bytes
@tt-vjovanovic can you share a git commit hash of your repo that demonstrates the problem?
3bff9d422e35b8ff2e28f99d6ec63eb31d00cf1e third_party/umd
I'm not saying this commit caused the problem, but that is where our submodule points to :)
Related, to be fixed after this issue is resolved: https://github.com/tenstorrent/tt-debuda/issues/124
We are using UMD in tt-debuda and having trouble when reading/writing more than 1KB of data from second chp on N300. Either our initialization of UMD is bad or there is a bug in UMD. In function
tt_SiliconDevice::read_from_non_mmio_device
there is conditionuse_dram = size_in_bytes > 1024;
... We are usingtt_SiliconDevice::read_from_device
andtt_SiliconDevice::write_to_device
...