In the current sw repository, the runtime layer hardcodes the number of NVDLA devices to 1.
As a result, nvdla_runtime and any applications using the IRuntime API cannot access or enumerate multiple NVDLA hardware instances present in the system.
This effectively disables multi-instance support even when:
-The device tree defines multiple NVDLA cores (nvdla@..., nvdla1@...).
-The kernel driver successfully registers multiple nvdla instance.
Load path always targets instance 0.
This means:
-The runtime cannot enumerate multiple hardware cores.
-Applications have no way to select a specific device instance
Impact
-This limitation prevents full hardware utilization on multi-core NVDLA designs (e.g., SoCs with nvdla0 and nvdla1).
-Allowing enumeration and instance selection would enable concurrent inference workloads across multiple DLAs.
In the current sw repository, the runtime layer hardcodes the number of NVDLA devices to 1.
As a result, nvdla_runtime and any applications using the IRuntime API cannot access or enumerate multiple NVDLA hardware instances present in the system.
This effectively disables multi-instance support even when:
-The device tree defines multiple NVDLA cores (nvdla@..., nvdla1@...).
-The kernel driver successfully registers multiple nvdla instance.
Load path always targets instance 0.
This means:
-The runtime cannot enumerate multiple hardware cores.
-Applications have no way to select a specific device instance
Impact
-This limitation prevents full hardware utilization on multi-core NVDLA designs (e.g., SoCs with nvdla0 and nvdla1).
-Allowing enumeration and instance selection would enable concurrent inference workloads across multiple DLAs.