csullivan commented on a change in pull request #8823:
URL: https://github.com/apache/tvm/pull/8823#discussion_r696091607
##########
File path: python/tvm/target/target.py
##########
@@ -413,79 +413,118 @@ def bifrost(model="unknown", options=None):
return Target(" ".join(["opencl"] + opts))
-def hexagon(cpu_ver="v66", sim_args=None, llvm_args=None, hvx=128):
+def hexagon(cpu_ver="v66", **kwargs):
"""Returns a Hexagon target.
Parameters
----------
- cpu_ver : str
+ cpu_ver : str (default: "v66")
Review comment:
You are correct for the way Vulkan currently retrieves device attributes.
However the proposal would be to elevate the device attribute querying to
the DeviceAPI. In this case the RPC support that already exists with the
DeviceAPI would be used to query the remote DeviceAPI. The intention is that
the device attributes would be queried remotely and then sent back to the local
machine used for cross compiling. It could then used to build a device specific
target used to inform lowering and codegen. In short, the approach should not
require the device to be local to the machine doing the compilation, as indeed
there are many cases when this is not practical. Do you still see technical
feasibility issues with this approach?
The goal would be to improve performance and accessibility by requiring less
information provided by the user of TVM in order to achieve optimal device
specific code.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
