Kun Lu
Kun Lu
The following Bitcast related test cases would fail on s390x (BE machines): - `//tensorflow/python/ops/ragged:ragged_bitcast_op_test` - `//tensorflow/compiler/tests:cast_ops_test_cpu` - `//tensorflow/compiler/tests:cast_ops_test_cpu_mlir_bridge_test` I found the root cause is that the basic Bitcast operation in...
Test case `//tensorflow/compiler/mlir/quantization/tensorflow/python:quantize_model_test` would fail on s390x (BE machines) due to endianness issue in `static range quantization` process. This PR is to swap the tensor content in `graph_def` after the...
Test case `//tensorflow/compiler/xla/tests:bitcast_convert_test_cpu` would fail on s390x (BE machines). This test case applies two XLA backends (HOST/CPU and Interpreter) to execute the `BitcastConvertHloTest` and compare the results. However, the `Interpreter`...
Test case `//tensorflow/compiler/mlir/lite/quantization/lite:quantize_weights_test` would fail on s390x (BE machines). The cause is that the un-quantized output `quant_model` built with `FlatBufferBuilder` has endianness issue in tensor contents. This PR is to...
**Description of problem:** Hi, We noticed that recent CI builds are failing on the s390x builder ([https://build.gluster.org/job/centos8-s390-regression/6/consoleFull](https://build.gluster.org/job/centos8-s390-regression/6/consoleFull)). It took more than 7 hours to build and run the regression test...
This PR aims to achieve a holistic/generic solution for TFLite FlatBuffers endianness issue on s390x(big-endian machines). It follows the below guidelines on BE machines: 1. Provide the endianness conversion feature...
Test case `//tensorflow/python/framework:tensor_util_test` failed on s390x (big-endian arch) because in `testHalf()` function the `tensor_content` data which is used for comparison with the tensor proto string is hard-coded in little-endian format....
Sub-test `tensorflow.python.ops.io_ops.serialize_tensor` in test case `//tensorflow/tools/docs:tf_doctest` failed on s390x (big-endian arch), because the serialized binary strings which represent the tensor data are different between little-endian and big-endian platforms. This PR...
Hi, Since the `luajit2` code change has been merged into the latest version of `OpenResty`, now we can build `OpenResty` from source on s390x with similar steps to amd64 arch....