Example of how to use XLA AOT via tfcompile to build a Keras model into a shared library.

tfcompile.ipynb

did you solve the error ?

no. are you experiencing the same?

I am trying to use XLA compiler from Tensorflow following your jupyter example

During execution of bazel build I always end up on the following build error:

error: Could not find include file 'tensorflow/compiler/mlir/xla/ir/hlo_ops_base.td' include "tensorflow/compiler/mlir/xla/ir/hlo_ops_base.td" ^ external/org_tensorflow/tensorflow/compiler/mlir/xla/ir/hlo_ops.td:22:9: error: Unexpected input at top level include "tensorflow/compiler/mlir/xla/ir/hlo_ops_base.td"

> ERROR: /home/ubuntu/.cache/bazel/_bazel_ubuntu/e5cce820cc082410b4fcc604db349066/external/org_tensorflow/tensorflow/compiler/mlir/xla/BUILD:465:1: Executing genrule @org_tensorflow//tensorflow/compiler/mlir/xla:operator_writer_inc failed (Exit 1)
[6,144 / 7,191] 3 actions running
    @org_tensorflow//tensorflow/compiler/xla/client:global_data; 4s local
    @org_tensorflow//tensorflow/core/kernels/tensor_forest:resources; 1s local
    ...//tensorflow/core/kernels:eigen_contraction_kernel_with_mkl; 1s local
external/org_tensorflow/tensorflow/compiler/mlir/xla/ir/hlo_ops.td:22:9: error: Could not find include file 'tensorflow/compiler/mlir/xla/ir/hlo_ops_base.td'
include "tensorflow/compiler/mlir/xla/ir/hlo_ops_base.td"
        ^
external/org_tensorflow/tensorflow/compiler/mlir/xla/ir/hlo_ops.td:22:9: error: Unexpected input at top level
include "tensorflow/compiler/mlir/xla/ir/hlo_ops_base.td"
        ^
[6,144 / 7,191] 3 actions running
    @org_tensorflow//tensorflow/compiler/xla/client:global_data; 4s local
    @org_tensorflow//tensorflow/core/kernels/tensor_forest:resources; 1s local
    ...//tensorflow/core/kernels:eigen_contraction_kernel_with_mkl; 1s local
Target @org_tensorflow//:graph failed to build
[6,147 / 7,191] checking cached actions
Use --verbose_failures to see the command lines of failed build steps.
[6,147 / 7,191] checking cached actions
INFO: Elapsed time: 7903.567s, Critical Path: 204.12s
[6,147 / 7,191] checking cached actions
INFO: 5961 processes: 5961 local.
[6,147 / 7,191] checking cached actions
FAILED: Build did NOT complete successfully
FAILED: Build did NOT complete successfully

So, it does not find the hlo_ops_base.td file, which of course is present in the path (I checked it)

The first time I have tried this, it worked like a charm.

Afterwards I have executed it again on different machines (also perfect clean VMs on different platforms), but always had the same issue.

I am using:

• bazel 1.1.0,
• tensorflow 1.14 (cpu),
• protobuf 3.0.0,
• python 2.7

Does anyone have any clue on how to solve this? I have tried to search it online and it seems no one else is having this issue...

Thanks, Matteo

same

Model => 150 ms ± 199 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
XLA binary => 191 ms ± 604 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

Why XLA compiled binary is slower than the model itself?