TensorFlow build fails with --config=sycl

[ashokbugude]

TensorFlow build fails with --config=sycl

[DuncanMcBain]

Hi @ashokbugude, thanks for your report. Could you give us some more information, please? We'd like to know what operating system you're building for, and what branch you're building from. Ideally you should use this branch, as while we will upstream our modifications to TensorFlow there is sometimes a delay before they are accepted. Therefore, more recent commits can be found there first.
Thanks,
Duncan.

[ashokbugude]

Hi @DuncanMcBain . Thanks for your response. Below are the details of tensorflow I am using

System information

Have I written custom code (as opposed to using a stock example script provided in TensorFlow):No
OS Platform and Distribution (e.g., Linux Ubuntu 16.04): 17.04
TensorFlow installed from (source or binary): source
TensorFlow version (use command below): 1.3
Python version: 2.7
Bazel version (if compiling from source): 0.5.1
GCC/Compiler version (if compiling from source): 6.0.3
CUDA/cuDNN version: N/A
GPU model and memory: N/A

Actual Requirement : To cross-build Tensorflow using --config=sycl for arm 32 bit system

Things Done
1. Able to cross-build Tensorflow without --config=sycl for arm 32 bit system using below command

bazel build --crosstool_top=//armv7l-compiler:toolchain --cpu=armeabi-v7a tensorflow:libtensorflow_cc.so --cxxopt="-std=c++11"

Result : Built Fine without errors

2 cross-build Tensorflow without --config=sycl for arm 32 bit system using below command

bazel build -c opt --copt="-mfpu=neon" --crosstool_top=//armv7l-compiler:toolchain --cpu=armeabi-v7a tensorflow:libtensorflow_cc.so --cxxopt="-std=c++11" --config=sycl
Result
Build Failed at Last moment while trying to link libComputeCpp.so

I have downloaded the sdk from codeplaysoftware website for Linux16.04 and arm64 bit. Can I know if there is 32 bit version .

If not can I please know if I can get source code to build libComputeCpp.so using cross compilation
for arm 32 bit

I basically wanted to improve the speed of TensorFlow on embedded 32 bit arm device by using opencl or sycl. Could you also please suggest any other optimization flags to be used while building.

Thanks

[DuncanMcBain]

I'm sorry, but cross-compiling TensorFlow with SYCL is not currently supported for technical reasons. I expect that this situation will change in the future, but won't work right now (in effect, you are already cross-compiling when using SYCL, and you can't cross twice). While we don't distribute for arm32 right now, there's nothing really stopping us doing that in future. Can you talk about the hardware you're targeting particularly?

[ashokbugude]

Thanks for your response . I am targeting 'arm-v7a' ie 32 bit arm architecture . I was actually looking for source code to build libComputeCpp.so so that I can try cross-build myself as I am in need of it badly

[DuncanMcBain]

I've asked internally, we will likely be releasing an arm32 build of ComputeCpp with our next public version. We're building and testing it internally anyway, so it should be fairly easy.

That said, you will still not be able to cross-compile TensorFlow for the reasons I outlined above. While we have plans to improve this, making modifications to TensorFlow's build scripts is a challenging process, so I can't say when it might be available. We do want to get it done, however, as we expect it will bring a large reduction in build times.

Which OpenCL implementation are you targeting? ComputeCpp requires some non-core feature support to be able to run, in particular it requires either SPIR 1.2 or SPIR-V support to be able to run kernels on-device.

[ashokbugude]

Thanks for detailed response I understood your reasons Sorry ,I do not have much knowledge of SPIR But my embedded device has a Qualcomm cpu with 4 cores with arm-v7a architecture and I am able to cross compile Tensorflow using an sdk similar to arm-gnueabi-v7a as present in https://releases.linaro.org/components/toolchain/gcc-linaro/latest/ I am thinking even though you may release 32 bit arm version , it might not integrate with the device unless it is compiled with the sdk as mentioned in above link, kindly correct me if I am wrong

…

On 4 Dec 2017 23:26, "Duncan McBain" ***@***.***> wrote: I've asked internally, we will likely be releasing an arm32 build of ComputeCpp with our next public version. We're building and testing it internally anyway, so it should be fairly easy. That said, you will still not be able to cross-compile TensorFlow for the reasons I outlined above. While we have plans to improve this, making modifications to TensorFlow's build scripts is a challenging process, so I can't say when it might be available. We do want to get it done, however, as we expect it will bring a large reduction in build times. Which OpenCL implementation are you targeting? ComputeCpp requires some non-core feature support to be able to run, in particular it requires either SPIR 1.2 or SPIR-V support to be able to run kernels on-device. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#68 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AJjYUOG7fzuQuIPILxYN4WdaTtRwAatLks5s9DJYgaJpZM4QznAW> .

[DuncanMcBain]

I don't think there will be any problems linking against code compiled from, say, the linaro toolchain. At least, we've not seen any internally.

SPIR is an intermediate format based on LLVM IR which allows developers not to have to write OpenCL kernels directly in OpenCL C, and is the format that ComputeCpp primarily targets. You can find more information here: https://www.khronos.org/spir/
You can find out if you have SPIR support by running the "clinfo" tool, which you can likely install via your package manager. If you grep the output for the string "cl_khr_spir" and find it, your device will be able to run ComputeCpp.

[DuncanMcBain]

Hi @ashokbugude,
Did you manage to make any progress with this? Did you get any useful output from, say, cl_info and the like?

[ashokbugude]

Hi Although opencl so files are present (without header fules) in the device, it says clinfo command not found. Istalling clinfo tool is not easy as it needs to be cross - compiled for the device.Hence no progress from my side.

…

On 16 Dec 2017 03:13, "Duncan McBain" ***@***.***> wrote: Hi @ashokbugude <https://github.com/ashokbugude>, Did you manage to make any progress with this? Did you get any useful output from, say, cl_info and the like? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#68 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AJjYUKCAMyujsY4BYpfcobCOPqAsQ1YSks5tAugRgaJpZM4QznAW> .