[paklui]

The ROCm memory backend reported can_init = true for buffers allocated with hipMalloc(). When can_init = true is set, then in initialize_buffer_content(), it would fill the GPU buffer with host CPU stores.

The other accelerator memory backends use can_init = false for device memory in their memory allocate buffer routine and leave the buffer content uninitialized, so we want to also be consistent and leave the buffer content uninitialized. To be consistent, this change sets can_init = false in rocm_memory_allocate_buffer(), matching the other accelerator backends so to have the same behavior.

This 1-liner change is limited to the ROCm memory backend. Verified with --enable-rocm --enable-rocm_dmabuf in build, and ran various perftests, like ib_write_bw and ib_write_lat with --use_rocm and --use_rocm_dmabuf between 2 nodes with AMD Instinct GPUs, runs complete, and bandwidth reaches line rate and latency unchanged.

Please review and thanks in advance.

[vijaysampath]

LGTM

[paklui]

hi @sshaulnv would you be able to review this PR for this 1-line change in the ROCm backend so it matches the other backends, thanks in advance!

[sshaulnv]

@paklui , thanks for the contribution, merged!