how to save the generated opencl code about tvm-mali

[merrymercy]

After you build the function, you can get the opencl source code.

s = tvm.create_schedule(C.op)
func = tvm.build(s, [A, B, C], target='opencl')
src = func.imported_modules[0].get_source()

Variable src is a string of opencl code. Then you can write it to a file.
See the tutorial (http://docs.tvmlang.org/tutorials/get_started.html) for more explanation on this example.

For the convolution example in my blog post, I get the source code by uncommenting line 135

tvm-mali/layer-test/test_conv2d.py

Lines 131 to 135 in 8b0ba39

	with target:
	B = topi.nn.conv2d(A, W, stride, padding)
	s = topi.generic.schedule_conv2d_nchw([B])
	func = tvm.build(s, [A, W, B], target_host=target_host)
	#print(func.imported_modules[0].get_source())

[janboeye]

@merrymercy
Thanks
How could I get opencl kernel after nnvm.compiler.build?

[merrymercy]

After this block

tvm-mali/mali_imagenet_bench.py

Lines 32 to 35 in 90395b5

	with nnvm.compiler.build_config(opt_level=opt_level):
	graph, lib, params = nnvm.compiler.build(
	net, tvm.target.mali(), shape={"data": data_shape}, params=params,
	dtype=dtype, target_host=args.target_host)

insert this line

print(lib.imported_modules[0].get_source())

[janboeye]

Thanks a lot!
What mean is lib.imported_modules?
And how to get host runtime source code too?

[merrymercy]

lib is the host module (llvm ir)
lib.imported_modules is the device module (cuda, opencl)

lib.get_source()  # get host llvm ir
lib.imported_modules[0].get_source() # get opencl/cuda source

[janboeye]

There is no way to get host c code, only LLVM IR, right?
export_library does not generate opencl runtime code, right?

[janboeye]

Is there any document to explain how this host library and opencl are executed on target platform?

[merrymercy]

1. TVM translate its Halide IR to LLVM IR directly, so there is no host c code.
2. nnvm.compiler.build generates both host code and opencl code. export_library only saves them to files

Since the original question is answered, we can close this issue for now and open new issue on tvm repo.