Ciprian Mindru
Ciprian Mindru
**Summary** - Optimize the Insert/Extract/Transpose kernels from the CPU backend by removing the address arithmetic performed at run-time with a simple access pattern based on offsets only, generated at compile-time...
**Summary** - Merge Pad into ChannelwiseQuantizedConvolution: put logic into the existing graph optimization pass `MergePadIntoConvolution` which was merging the Pad only into the Convolution node. - Merge Pad also into...
The graph optimizations should not change the numerical behavior of a graph before/after transformation. Some graph optimizations for quantized operations change the numerical behavior: 1. Merging multiple `RescaleQuantized` nodes together...
Add `stride` parameter in the Slice operator. The `stride` parameter should be allowed to be both positive and negative (slicing from the end). Both ONNX and TensorFlowLite formats define Slice...
The current reduce operators are all either incomplete or inconsistent. 1. The operator names should simply start with `Reduce` and not `BatchedReduce`. I don't understand why in Glow we have...
One such example we can optimize is for this graph: ```Concat(inp1, inp2, ..., inpN)``` For the particular case when the `Concat` node concatenates contiguous slices (e.g. concatenation is done along...
We currently have an optimization to fuse BatchNorm into Conv2D but we need also for FullyConnected (it is similar).
 In this case the RescaleQuantized node should be fused within the ChannelwiseQuantizedConv node output. This fusing type should be refactored in GraphOptz in order to enable it more easily...
Currently the placeholders are allocated in a separate memory buffer called "mutableWeights". For further memory optimization one would choose to allocate the placeholders in the "activations" memory pool such that...
@jfix71 After your change in #4942 some models are not being quantized properly by our backend. Attached here you have everything to reproduce the problem for a MobileNetv1 model. More...