1) With upcoming firmware AMD NPU will be exposed as /dev/accel/accel0.
This device is owned by root:render, similarly to GPU.
When present, tools like rocminfo try to query device capabilities,
breaking sandbox. To fix this issue, this device has now addwrite in
check_amdgpu.
2) There are a bunch of bugs from tinderbox and users who forgot to
enable KFD in kernel. Instead of recommendation to check permissions,
they will see a better message that AMD device is missing.
3) In cases when we just want to addwrite to AMD devices, new function
rocm_add_sandbox (similar to cuda_add_sandbox) was added. No errors are
raised if device is missing.
Part-of: https://github.com/gentoo/gentoo/pull/44355
Bug: https://bugs.gentoo.org/965198
Signed-off-by: Sv. Lockal <[email protected]>
---
eclass/rocm.eclass | 37 ++++++++++++++++++++++++++++++++++---
1 file changed, 34 insertions(+), 3 deletions(-)
diff --git a/eclass/rocm.eclass b/eclass/rocm.eclass
index 0fa99a4178e71..c24666e33e8d6 100644
--- a/eclass/rocm.eclass
+++ b/eclass/rocm.eclass
@@ -248,13 +248,44 @@ get_amdgpu_flags() {
echo $(printf "%s;" ${AMDGPU_TARGETS[@]})
}
+# @FUNCTION: rocm_add_sandbox
+# @USAGE: [-w]
+# @DESCRIPTION:
+# Add AMD GPU/NPU dev nodes to the sandbox predict list.
+# with -w, add to the sandbox write list.
+rocm_add_sandbox() {
+ debug-print-function "${FUNCNAME[0]}" "$@"
+
+ local i
+ for i in /dev/kfd /dev/dri/render* /dev/accel/accel*; do
+ if [[ ! -c $i ]]; then
+ continue
+ elif [[ $1 == '-w' ]]; then
+ addwrite "$i"
+ else
+ addpredict "$i"
+ fi
+ done
+}
+
# @FUNCTION: check_amdgpu
# @USAGE: check_amdgpu
# @DESCRIPTION:
-# grant and check read-write permissions on AMDGPU devices, die if not
available.
+# Grant and check read-write permissions on AMDGPU and AMDNPU devices.
+# Die if no AMDGPU devices are available.
check_amdgpu() {
- for device in /dev/kfd /dev/dri/render*; do
- addwrite ${device}
+ # Common case: no AMDGPU device or the kernel fusion driver is
disabled in the kernel.
+ if [[ ! -c /dev/kfd ]]; then
+ eerror "Device /dev/kfd does not exist!"
+ eerror "To proceed, you need to have an AMD GPU and have
CONFIG_HSA_AMD set in your kernel config."
+ die "/dev/kfd is missing"
+ fi
+
+ local device
+ for device in /dev/kfd /dev/dri/render* /dev/accel/accel*; do
+ [[ ! -c ${device} ]] && continue
+
+ addwrite "${device}"
if [[ ! -r ${device} || ! -w ${device} ]]; then
eerror "Cannot read or write ${device}!"
eerror "Make sure it is present and check the
permission."
