TVM to OpenCL flow

Firstly, I am pretty new to TVM. I understand that TVM can lower deep learning models and run them on OpenCL devices. Is there a document which elaborates this flow? I am interested in understanding the compilation flags for selecting the OpenCL device and also the lowering of models to OpenCL Kernels.

Any references to the documentation is highly appriciated.

Thanks, Harish

I found this link useful, yet I’m not sure how can i select a CPU device in OpenCL.

use ‘’ or ‘’ to generate inference code. OpenCL is ‘target’ . ‘target_host’ is platform your program run on

Here is example for ‘armv7l-linux-gnueabihf’ device with OpenCL support.

This script is original for Android device with opencl. To show ‘target’ and ‘target_host’ meaning in corsscompile replace the ‘target_host’ with arm-linux

import tvm.relay.testing
from tvm import relay
from tvm.contrib.util import tempdir

target = 'opencl'
target_host = 'llvm -device=arm_cpu -model=bcm2835 -mtriple=armv7l-linux-gnueabihf -mattr=+neon'
toolchain = '/root/cross-pi-gcc-8.3.0-1/bin/arm-linux-gnueabihf-gcc'
options = ['-lm']

from import get_model

dshape = (1, 3, 224, 224)
block = get_model("mobilenet0.25", pretrained=True)
shape_dict = {"data": dshape}
mod, params = relay.frontend.from_mxnet(block, shape_dict)
func = mod["main"]
func = relay.Function(
    func.params, relay.nn.softmax(func.body), None, func.type_params, func.attrs

with tvm.transform.PassContext(opt_level=3):
    lib =, target=target, params=params, target_host=target_host)

tmp = tempdir()
filename = ""
lib.export_library(tmp.relpath(filename), cc=toolchain, options=options)

1 Like