MLIR bridge doesn't support tf.NonMaxSuppressionV4 legalization that is
conditionally generated by non_max_suppression_padded function.
PiperOrigin-RevId: 320197235
Change-Id: If7242133254680b366771ced50de074ed6180563
LHLO uses output-parameter, but TupleSelectOp outputs into a tuple on the device. The current type constraints are wrong, and there is not enough expressiveness in LHLO to define a device-memory representation that can be passed in between kernel launches.
PiperOrigin-RevId: 317749542
Change-Id: I6d3350cbf9decf006f239a7208f6bbef0175ac61
tf.BesselI0e and tf.BesselI1e are moved to tf_ops.td as they are now python generated ops, so they do not have summaries and descriptions in the op registry. int8 and int16 support has been added to tf.Acos, tf.Atan, tf.Inv, tf.Reciprocal, tf.Round, and tf.Tan.
PiperOrigin-RevId: 317748438
Change-Id: Icb86560d5118c38c69819fd02daa3a3841e113b2
DeviceIndex op: given a list of device names, this operation returns the index of the device this op runs. In the case of XLA, we are not executing on any device, we return the length of the list.
PiperOrigin-RevId: 317740778
Change-Id: I0679aa0adc5508b83502eee0d2044584577ed5b4
Previously, `BuildXlaCompilerArguments` required taking a snapshot of all
resource variables in order to start compiling.
In this CL it can operate with a span of pointers to resource variables instead
(we CHECK at runtime that the lock is held).
That refactoring allows to launch the XLA compilation without creating an extra
reference to the underlying Tensor of the passed resource variables.
PiperOrigin-RevId: 317706126
Change-Id: I37a97601a08f165b23b4745e1b032bf91c21c313
CUDA 11 changed the libdevice implementation of expf from previously two to one ex2.approx.ftz. Change the CHECK directives to handle both cases.
PiperOrigin-RevId: 317698875
Change-Id: Idd7b28c77427f299b80fa1b7f4b9be8c7881f963
By this order, the default quantization parameter is only applied on the
activations and the weight quantization parameter will use the parameters from
the weight content.
PiperOrigin-RevId: 317672365
Change-Id: Ib7b02ae19105db124721242ea51a5ccc1d5aa68e
MLIR recently introduced a new idiom for constructing loop nests. Use it to
make the legalization of LHLO to affine loops more concise and readable.
PiperOrigin-RevId: 317649515
Change-Id: Idfab27b4655d6df90d940fb7b064ea9941d8a700
Don't use inline variables, our MSVC builds do not like them.
Instead, simulate them using inline functions + function-scope static variables.
PiperOrigin-RevId: 317428053
Change-Id: Icf8838c159ab9f132ad32360633046f4c2224a79
begin_mask and end_mask needs to be adjusted while simplifying StridedSlice op by removing ellipsis_mask. new_axis_mask is already removed before this and we don't yet support shrink_axis_mask.
PiperOrigin-RevId: 317419960
Change-Id: Ie4a5f404f95f5b909065311a54cbbed64d4ccf4b
It is possible for more than one replicated device to be populated (e.g. replicated host).
PiperOrigin-RevId: 317419924
Change-Id: I1d8fa95654f324557d78adf4570adc6c3cfdabb4
TF saved model dialect.
-Mark the init function referenced by SessionInitializerOp as an exported
function with the reserved name "__tf_saved_model_session_initializer".
-Remove variable init logic in init function if it is already imported as a global tensor.
-Add a canonicalizer to SessionInitializerOp to remove empty init functions.
-Return error on SessionInitializerOp in IREE compiler pipeline.
PiperOrigin-RevId: 317395165
Change-Id: Idb7e54cac08add9fc8f2ccfbaf341135fdf59e3b
This is important for the upcoming DeviceIndex op which can be used to select
one of many implementations depending on the device, and some of them may not be
compilable by tf2xla.
PiperOrigin-RevId: 317376420
Change-Id: I6428df6f4da238e5d2bc3618d51c579e34454945
The default value of the flag is True. When the flag value is false, the
bridge will report an error when the native segment of a TRTEngineOp is
executed.
Add test cases.
PiperOrigin-RevId: 317340632
Change-Id: Iacded09b38e63442bbd93076a079d385fb8a77e6
Add LOG_WARNING_WITH_PREFIX to common/utils.h. Replace the use of LOG(WARNING)
with this new macro.
PiperOrigin-RevId: 317330336
Change-Id: Ife0aa0347dd72f6eb0f8805af4d46a7d4cb099ea
Basically rolledback for cl/305641881, the pattern could hurt performance
because the operation can't be fully tiled in Linalg transformation. In this
context, not everyone wants this pattern, so remove it from canonicalize
patterns.
PiperOrigin-RevId: 317302072
Change-Id: I19aa64e14eecccfd738ad3f775f3670974bc68f9
Also, enable mlir bridge for image ops compilers test. ResizeBilinear op
lowering usese CustomCall in case of TPU lowerings.
PiperOrigin-RevId: 317272443
Change-Id: I134c828cdc76552a0cbfdeb7c65532aa986314e2
Added this util upstream in D81435, so now using that instead and deleting the unneeded code here.
PiperOrigin-RevId: 317209256
Change-Id: Id2d8a1fca34ca85e59a05a85bf7f6f59b425c7c1
