Summary: 121 instances, 109 unique

Text	Count
// TODO Add support for other input types, e.g., tensor[]	1
// TODO: think about whether this is better off computed inside a Node.	2
// TODO: any kernel size check in respect to input ? In contrast to Conv,	1
// TODO we visited many redundent nodes during this process,	1
// TODO: We could come up with a mechanism to lazy compute that	1
// TODO: intend to find a way to reuse the following function later	1
# TODO: ensure that the compiler supports these options before adding	1
// TODO: add checks for number of inputs and argument values	1
// TODO: Use a more specific ONNXIFI error code here to denote what about	1
// TODO: replace users of ReLU input after ReLU with ReLU output.	1
// TODO - not quantized support yet in libjit.	1
// TODO Currently we only support symmetric quantization.	1
// TODO: Use a proper type based on the JIT's output type.	1
// TODO: Remove ->getLayout() enum and take a string like transpose. Refactor	2
// TODO: use a rw mutex here for efficiency.	1
// TODO: consider refactoring boilerplate code to new trait: DottyPrintable<ADP>	1
// TODO: Uncomment this once #5729 gets fixed.	2
// TODO: fix this so that a HostManager is deleted when all backends	1
// TODO: support any dimensional transposes.	1
// TODO Add support for other output types, e.g., tensor[]	1
// TODO: implement 3rd argument rounding_mode option	1
// TODO - registered once to avoid error:	1
// TODO: Use a more specific ONNXIFI error code here to denote what	1
// TODO: Only add to externalPlaceholders_ of PH is external placeholder	1
// TODO: This backend does not have a 64-bit type, but Glow uses	1
// FIXME: Right now, the TensorQuantizationParams only tracks one	1
// TODO: Remove the quantization step to potentially improve performance.	1
// Initialize bias (TODO take out to a separate function when quant is in).	1
// TODO: Change Visitors to return whether they modified the Function they	1
// FIXME: Remove InOut!	1
// TODO: Erase N during CSE? If we don't do it here,	1
// FIXME: looking for symbols external to libjit in the process is	2
// TODO: verify with garret we don't need to lock here - i.e. host manager	1
// TODO: This limitation can be lifted, but that is for simplicity.	1
// TODO: these may need to be tweaked depending on specific CPU.	1
// FIXME: this is a temporary solution for the case when NonZero returns	1
/// TODO: Currently, for all supported types wider mantissa also means wider	1
// TODO: Move this into the GraphOptimizer once Glow supports reduce	1
// TODO: We can also support Splat here if needed.	1
// FIXME: This can starve inactive topos.	1
// TODO: should we check is_multiparam?	1
// TODO: OpenCL fast convolution kernel itself has some issue with group >	1
// TODO: May be disallow usage of dest interval for src?	1
// TODO: allow correct type mapping from double to float	1
// TODO: Only if enabled.	1
// TODO: Need to add different flavours of dumping.	1
// TODO: these may need to be tweaked depending on interpreter overheads.	1
// TODO: add AVX implementation.	1
// TODO: Fix bad assumption? See issue 3499, for now workaround it.	1
usedMemoryBytes_ += functionCost_; // TODO:: static moduleSize.	1
/// FIXME: This is a workaround, because defining the hash_code	2
// TODO: We can also support quantized constants if there is a need in it.	1
// TODO: for intermediate placeholders in DRT/P2P cases, we don't need	1
// TODO: extend this to allow non-constant scalars.	3
// TODO: need to remove the Clip ops once the FC inputs can be put on SRAM	1
// TODO: Unload functions that were loaded successfully.	2
// TODO: Try to produce semantically meaningful parameter names, e.g. by	1
usedMemoryBytes_ -= functionCost_; // TODO: static moduleSize.	1
// FIXME: looking for symbols external to libjit in the process is	1
} // FIXME: else make sure networkName does not have any sequence of	1
// TODO: Do we need support for different quant params of copy?	1
// TODO: Handle other dimensions.	2
// TODO reverse map of jit-glow node should resolve this problem.	1
// TODO: add a TRACE_EVENT entry	1
// TODO: Verify model integrity using flatbuffers::Verifier class.	1
/// TODO: can use one mutex per set if performance becomes an issue.	1
/// TODO: Remove after modifying InstrGen to use MemberTypeInfo as well?	1
// TODO: move all initialization out of constructor.	1
/// TODO: check Dtype is float (optional value).	1
// TODO: add axis/dim support	1
// TODO: Rename "BatchDims" member to "Axis". This was attempted in #5565 but	2
// TODO: If it would introduce a last write into an observable, do not	1
// TODO: add description	1
// TODO: Add here any backend specific logic.	1
// TODO: remove the shuffle and replace it with layout.	1
// TODO: this function is largely duplicated with allowsPartialInput()	1
// TODO: Support quantized constants if needed.	2
// TODO: synchronize clocks better, this can be off the thread was yielded	1
// TODO: delete this once this is fixed by	1
/// TODO: distinguish between data types with different peak flops.	1
// TODO - support other types.	1
// TODO: Loader should provide function to register callbacks.	1
// TODO:	1
// TODO: this should be done only on Glow subgraphs to avoid modifying parts	1
// TODO: consider adding this functionality to the main operator set.	1
// TODO: support more info type values. Here is the minimal required	1
// TODO investigate if removing "noalias" can be used to create bigger	1
// TODO: look through the nodes in FXIR to determine if a custom DSP op is	1
// TODO: Add only constants used by F to the compiled function. This should	1
// TODO: ONNX spec disallows using "pads" and "auto_pad" together. However,	1
/// literature).  TODO: Generalize these parameters for other cache sizes.	1
// TODO: Check that ptr is an allocated address.	1
// TODO: When Glow supports reduce operators with multiple axes remove this!	1
// TODO: check maybe we can support more dimensions to be reduced	1
// TODO: Handle this case with FloorDiv	1
// TODO: Use synGetMemInfo once implemented.	1
// TODO Add support for other input types, e.g., tensor list	1
// TODO: Add support for strides different than 1 (positive or negative) once	1
/// TODO: Default tensor loader could be extended to support various data types	1
// TODO: the logic here need to be improved.	1
// TODO: broadcast inputs if input is not a node.	1
// TODO : will improve the algorithm for different memory size.	1
// TODO: Only run the appropriate passes as needed.	1
/// TODO: Handle this case with FloorDiv	2
// TODO: fix UINT_MAX	1
// TODO: Change Glow Type to use sdim_t to be consistent	1
/// TODO: Multi-dimension slicing will be supported later.	1
// TODO: @hwwang T80910607 Only support None dtype (4th argument)	1
// TODO: pass dim as a second tensor	1