galkinvv
commented
1 year ago In short: thanks for reporting. This is known behavior and is mostly harmless. (I've noted the crash on Ctrl+C while piped as a separate issue #17).
Detailed:
The reason of a problem is a some incompatibility between RADV driver and memtest_vulkan. memtest_vulkan insists on allocating contiguous buffer (this is a technical design choice), but RADV is failing to allocate such buffer contiguously.
However, the good news: being able to test only 15.0GB out of 24.0 GB is not a problem for nearly all usages, since
- if you are using memtest_vulkan to check the GPU for errors or overclocking stability - in practice, 99% of VRAM errors actually found by memtest_vulkan are not bound to a specific memory location, but to a data transmission mechanisms. So, they are equally found while testing ~24GB and 16GB (and even 3-4GB).
- if you are using memtest_vulkan for cooling/performance check - 16GB and 24GB are the same too.
So the for 99% cases - this is just fine. The Failed to allocate a buffer messages above are generated by a driver, and from the memtest_vualkn usage scenario can be just ignored.
I can't make the driver less verbose since it is driver's unconditional stderr output
Actually some other drivers sometimes fails allocations of large contiguous buffers too, memtest_vulkan silently auto-selects a bit smaller, it's just ok.
As a half fix - I plan to detect RADV driver and apply some minor tunes:
- add a notice the to the memtest_vulkan that the errors from driver are harmless
- tune the memory size auto selection coefficient to faster decreasing memory size - so the count of errors won't be so huge.
Also, there is a chance that the 16GB limit is specified in some limits exposed by the driver, but I'm not sure.
Can you please install the package containing the vulkaninfo utility (something like vulkan-tools) and attach the output of VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/radeon_icd.x86_64.json vulkaninfo ?
A side note: AMD GPUs on Linux has 2 different opensource vulkan drivers that can be installed simultaneously without severe conflicts. So you can additionally install the AMDVLK driver, and specify the driver to use by a VK_ICD_FILENAMES environment variable: amd_icd64 is AMDVLK, and radeon_icd.x86_64 is the RADV you are using now.
(for your vulkan loader's libvulkan1 version. Newer libvulkan1 loader version renamed that var to VK_DRIVER_FILES).
So, with AMDVLK driver additionally installed you can run
[user@host]$ VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/amd_icd64.json ./memtest_vulkan
[user@host]$ VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/radeon_icd.x86_64.json ./memtest_vulkanI have no 16+GB AMD GPU, but from my experience with 12 GB RX6700 - AMDVLK allows testing 10.5GB (1.5GB is auto-skipped by memtest_vulkan to avoid desktop lockup), while RADV limits it to 4GB. So I hope that for RX 7900 the AMDVLK can allow going above 16GB.
The same environment var applies to all other vulkan apps, including vulkaninfo
jhumlick
I have a RX 7900 XTX, and it looks like not all of my memory is being tested. When I launch normally, I see:
I have resizable BAR turned on in my bios.
I will attach the output I see running with the file renamed to memtest_vulkan_verbose.
It also appears that the tool crashes if I write to a log file with tee using a pipe. (i.e.
./memtest_vulkan_verbose| teememtest_vulkan_verbose.txtwill crash when ctrl+c is pressed) memtest_vulkan_verbose.txt