Christoph Berganski
Christoph Berganski
Currently, they key matrix is transposed first before being quantized. To simplify the streamlining and detection of the operator pattern in FINN, I would like to have the transpose directly...
This was already implemented, it was just not called from the `step_synthesize_bitfile`. According to the last synthesis I ran using Vitis/Vivado 2022.2, the DSP48 utilization seems to be reported in...
I am facing the `AssertionError: Thresholds can't be expressed with type INT8` (the assertion is [here](https://github.com/Xilinx/finn/blob/f2424e7fa9bb834f192bdec632bbdebb84b64fb1/src/finn/custom_op/fpgadataflow/thresholding_batch.py#L350) in the weight file generation of `Thresholding_Batch`), which I cannot really explain, at least...
Trying ideas and bug fixes for streamlining the scaled dot-product attention operator. Related to issue/discussion #878 - [x] Refactor `MoveScalarMulPastMatMul` for two-input join-node matmuls - [x] Validate that these changes...
More of a personal preference, but I assume I am not alone. I like to specify my model or "experiment" configurations as well as the metrics collected in YAML format...
When running some automated finn builds by some experiment manager doing parameter sweeps, the jobs failed with some TTY issues from within docker, which I have tracked down to the...
## Quick summary This is not really a bug report, it is more a collection of missing support and minor annoyances I am encountering while working towards streamlining and eventually...
See https://github.com/Xilinx/finn/issues/978 for some explanation of what it wrong with this transformation. See https://github.com/Xilinx/finn/issues/875 for more context. - [x] Rework the `RoundAndClipThresholds` transformation to correctly respect the range of the...
This includes a set of `HWCustomOp` and `HLSBackend` operator templates which can be specialized in just a few lines of code to implement arbitrary elementwise binary operations, like `Add`, `Mul`,...
This is required for implementing floating-point support for generic element-wise binary operations in FINN, proposed here: https://github.com/Xilinx/finn/pull/1040 This does not change the behavior of `Slice` for already supported types, it...