Error when building tvm wirh hexagon V66 dsp

#1

when building tvm with backend qualcomm hexagon V66 architeture met some errors:

error: use of undeclared identifier ‘Q6_Vhf_equals_Vqf16’ error: use of undeclared identifier ‘Q6_Vqf16_vadd_VhfVhf’

error: invalid instruction asm volatile(" %0 = dmpause" : “=r”(dm0));

error: invalid instruction asm volatile(" %0 = dmpause" : “=r”(dm0));

#2

@sanirudh @srkreddy1238 @tqchen @kparzysz

anyone can give some ideas?

#3

Hi @Panda6, The error shows that fp16 HVX instructions and user dma instructions are not supported in v66 as they were introduced starting from v68 architecture.

#4

@Panda6 I’ve raised the PR #17169 to fix this issue. Please try it out and let me know.

#5

Well done.

Beside this, I’d like to knwn more about tvm integrated with hexagon dsp:

  1. Can TVM outperform snpe and qnn when used with hexagon backend in terms of performance?
  2. Can we use TVM with hexagon backend to deploy a model end-to-end?
  3. How does the quantization process work? Can we directly use the quantization encoding files exported from AIMET tools in TVM?
  4. With the introduction of V75 architecture and its HMX computing unit, does TVM already support it?

Thanks a lot.

#6

I want to build a model end-2-end like:

import numpy as np

import tvm from tvm import relay, auto_scheduler import tvm.relay.testing from tvm.contrib import graph_executor import os import tvm.contrib.hexagon as hexagon import onnx from tvm.contrib import utils, ndk

onnx_model = onnx.load(“mobilenetv2.onnx”)

path_lib = “./mobilenetv2.so” os.environ[“TVM_NDK_CC”] = “tools/android-ndk-r25c/toolchains/llvm/prebuilt/linux-x86_64/bin/aarch64-linux-android28-clang++”

mod, params = relay.frontend.from_onnx(onnx_model, {“x”: (1, 3, 224, 224)}) target = tvm.target.hexagon(“v66”, hvx=0) with tvm.transform.PassContext(opt_level=3): lib = relay.build(mod, target, params=params) lib.export_library(path_lib, ndk.create_shared)

But failed: clang-14: error: linker command failed with exit code 1 (use -v to see invocation)

@sanirudh @srkreddy1238 @tqchen @kparzysz