Comments (4)
Hi Digant,
Dynamic input shapes are also interest us. Dynamic shapes are not supported today. The entire graph must be reinitialised if the shape of an input tensor changes. This is expensive as all weights must be re-packed and indirection buffers recalculated.
Support for dynamic input shapes are on our roadmap. This is very important to us too. However, unbounded dynamism will not be supported. A maximum value for each input dimension must be given and memory allocated using this. Some preliminary work has already be done. See here.
Using the lower level APIs to do this will be very slow so I strongly recommend not doing this. I am happy to speak with you to see if our plans will be useful to you.
from xnnpack.
Thanks for getting back to me. Glad to know that you are also interested in this feature.
Support for dynamic input shapes are on our roadmap. This is very important to us too. However, unbounded dynamism will not be supported. A maximum value for each input dimension must be given and memory allocated using this.
Upper-bounded, rank-specialized^ dynamism is something we are interested in at the moment as well. Do you have any timelines you can share for the roadmap?
Using the lower level APIs to do this will be very slow so I strongly recommend not doing this. I am happy to speak with you to see if our plans will be useful to you.
Agree, as I said earlier using lower level APIs is not the most desirable path for us as well. And I would be happy to sync with you on this and for the dynamic shape feature in general.
^ - Tensor rank is not dynamic.
from xnnpack.
Support for dynamic input shapes are on our roadmap.
Hi Alan, that's great news. We have a strong dependency to XNNPACK and this is one the main painpoints we have. I know a lot can chancge, but what's the rough timeline that you are thinking of for this?
from xnnpack.
I am quite hesitant to give timelines publicly as we are extremely short staffed now. This should change soon so I will keep you updated as to our plans
from xnnpack.
Related Issues (20)
- will you Plan to support int8 perchannel quantize for linear op? HOT 2
- building failed on Raspberry Pi 4 HOT 1
- Regarding the issue with f32-gemm-bench. HOT 2
- filtering out -mcpu=native when building with Bazel on Arm 64-bit (aarch64) HOT 2
- Help Wanted: How to use SIMD to accelerate Exponential function on CPU.
- Can RVV Kernels be enable by default? HOT 1
- Build error when including XNNPACK using FetchContent HOT 4
- Help needed: any doc available.
- experiments-config.h is hiding the xnnpack.h header
- GELU support in XNNPACK HOT 4
- Need help: bench the 'sdpa' operator HOT 1
- Concatenate and Split don't support input that has 0 size dimension HOT 4
- Vector extension errors while building on RISC-V HOT 2
- Avoid undefined behavior in memcpy call in `xnn_define_static_reshape`
- Clamp on empty ranges should be valid HOT 2
- f32-raddstoreexpminusmax-rvv-rr2-p6-u4v.c error: no member named 'rvv_rr2_p6' in 'union xnn_f32_expminus_params' HOT 1
- Request for Legacy CPU Support or Improved Error Handling HOT 2
- [QD8_F32_QC4W] Issue with odd number of input_channels HOT 3
- XNNPACK make error scc1: error: invalid feature modifier 'i8mm' in '-march=armv8.2-a+i8mm+fp16' HOT 2
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from xnnpack.