mi300a-xnack

Test HSA_XNACK behavior on AMD MI300A

XNACK refers to the GPU's ability to retry memory accesses that failed due a page fault (which normally would lead to a memory access error), and instead retrieve the missing page (source)

Quick Start

./run.sh

ROCm 6.2.1 Results

Device Name

HSA_XNACK=0 (or unset)	HSA_XNACK=1
amdgcn-amd-amdhsa--gfx942:sramecc+:xnack-	amdgcn-amd-amdhsa--gfx942:sramecc+:xnack+

• xnack+: XNACK is available and enabled
• xnack-: XNACK is available and disabled (source)

Compiled with --offload-arch=gfx942 or no --offload-arch provided

GPU kernels will run regardless of whether XNACK is enabled

Allocator	Access	HSA_XNACK=0 (or unset)	HSA_XNACK=1
hipMalloc	GPU Kernel	yes	yes
hipMalloc	Host Loop	yes	yes
hipMalloc	hipMemset	yes	yes
malloc	GPU Kernel	segfault	yes
malloc	Host Loop	yes	yes
malloc	hipMemset	HIP runtime error	HIP runtime error
hipMallocManaged	GPU Kernel	yes	yes
hipMallocManaged	Host Loop	yes	yes
hipMallocManaged	hipMemset	yes	yes

Compiled with --offload-arch=gfx942:xnack-

GPU kernels will run only if XNACK is disabled

Allocator	Access	HSA_XNACK=0 (or unset)	HSA_XNACK=1
hipMalloc	GPU Kernel	yes	N/A
hipMalloc	Host Loop	yes	yes
hipMalloc	hipMemset	yes	yes
malloc	GPU Kernel	segfault	N/A
malloc	Host Loop	yes	yes
malloc	hipMemset	HIP runtime error	HIP runtime error
hipMallocManaged	GPU Kernel	yes	N/A
hipMallocManaged	Host Loop	yes	yes
hipMallocManaged	hipMemset	yes	yes

Compiled with --offload-arch=gfx942:xnack+

GPU kernels will run only if XNACK is enabled

Allocator	Access	HSA_XNACK=0 (or unset)	HSA_XNACK=1
hipMalloc	GPU Kernel	N/A	yes
hipMalloc	Host Loop	yes	yes
hipMalloc	hipMemset	yes	yes
malloc	GPU Kernel	N/A	yes
malloc	Host Loop	yes	yes
malloc	hipMemset	HIP runtime error	HIP runtime error
hipMallocManaged	GPU Kernel	N/A	yes
hipMallocManaged	Host Loop	yes	yes
hipMallocManaged	hipMemset	yes	yes