[GitHub] [incubator-tvm] FrozenGene commented on pull request #5754: [RFC] Improve quantized convolution performance for armv8 architectures

2020-06-22 Thread GitBox
FrozenGene commented on pull request #5754: URL: https://github.com/apache/incubator-tvm/pull/5754#issuecomment-647914319 Thanks @giuseros @anijain2305 MERGED NOW. This is an automated message from the Apache Git Service. To

[GitHub] [incubator-tvm] FrozenGene commented on pull request #5754: [RFC] Improve quantized convolution performance for armv8 architectures

2020-06-22 Thread GitBox
FrozenGene commented on pull request #5754: URL: https://github.com/apache/incubator-tvm/pull/5754#issuecomment-647889871 @anijain2305 could you have a look another round? This is an automated message from the Apache Git

[GitHub] [incubator-tvm] FrozenGene commented on pull request #5754: [RFC] Improve quantized convolution performance for armv8 architectures

2020-06-22 Thread GitBox
FrozenGene commented on pull request #5754: URL: https://github.com/apache/incubator-tvm/pull/5754#issuecomment-647615333 > Hi @FrozenGene , @anijain2305 , > Any update on this review? > Also, is there a way to retrigger the tests? Or should I contact someone in particular? >

[GitHub] [incubator-tvm] FrozenGene commented on pull request #5754: [RFC] Improve quantized convolution performance for armv8 architectures

2020-06-17 Thread GitBox
FrozenGene commented on pull request #5754: URL: https://github.com/apache/incubator-tvm/pull/5754#issuecomment-645406262 > @FrozenGene Can you please review when you get time? Yep. I could review it tomorrow. This is

[GitHub] [incubator-tvm] FrozenGene commented on pull request #5754: [RFC] Improve quantized convolution performance for armv8 architectures

2020-06-12 Thread GitBox
FrozenGene commented on pull request #5754: URL: https://github.com/apache/incubator-tvm/pull/5754#issuecomment-643286046 > Hi @FrozenGene , > I gave it another go, but switching legalization on the strategy seems very hard (since we would need the auto-tuner to pick the best data-type

[GitHub] [incubator-tvm] FrozenGene commented on pull request #5754: [RFC] Improve quantized convolution performance for armv8 architectures

2020-06-11 Thread GitBox
FrozenGene commented on pull request #5754: URL: https://github.com/apache/incubator-tvm/pull/5754#issuecomment-642690898 > Hi @FrozenGene , > I agree that different strategies should be available to the auto-tuner. See if the solution proposed is good enough for you (at least as a

[GitHub] [incubator-tvm] FrozenGene commented on pull request #5754: [RFC] Improve quantized convolution performance for armv8 architectures

2020-06-11 Thread GitBox
FrozenGene commented on pull request #5754: URL: https://github.com/apache/incubator-tvm/pull/5754#issuecomment-642671388 @giuseros I suddenly think of auto scheduler will have one environment value. So the change of legalization won't affect auto scheduler. We could check the value of

[GitHub] [incubator-tvm] FrozenGene commented on pull request #5754: [RFC] Improve quantized convolution performance for armv8 architectures

2020-06-11 Thread GitBox
FrozenGene commented on pull request #5754: URL: https://github.com/apache/incubator-tvm/pull/5754#issuecomment-642651252 > So I mean to add a `convert_data_type` pass that is similar to `alter_op_layout` but converts datatype (and we can do something like `if topi_impl == 'spatial_nhwc'

[GitHub] [incubator-tvm] FrozenGene commented on pull request #5754: [RFC] Improve quantized convolution performance for armv8 architectures

2020-06-11 Thread GitBox
FrozenGene commented on pull request #5754: URL: https://github.com/apache/incubator-tvm/pull/5754#issuecomment-642601817 > Hi @FrozenGene , > > The idea of adding the algorithm name to the attributes would work if the legalization step was run after we pick the strategy. It is

[GitHub] [incubator-tvm] FrozenGene commented on pull request #5754: [RFC] Improve quantized convolution performance for armv8 architectures

2020-06-11 Thread GitBox
FrozenGene commented on pull request #5754: URL: https://github.com/apache/incubator-tvm/pull/5754#issuecomment-642577581 > 1. It will be hard to do this. The point is that the legalization is done in Relay before picking the strategy (thus, it is unaware of the strategy picked). To keep

[GitHub] [incubator-tvm] FrozenGene commented on pull request #5754: [RFC] Improve quantized convolution performance for armv8 architectures

2020-06-11 Thread GitBox
FrozenGene commented on pull request #5754: URL: https://github.com/apache/incubator-tvm/pull/5754#issuecomment-642541198 Glad to see we have the same thought we should let autotvm select the best. Autoscheduler reley on the legalization pass to generate smlal inst(After auto

[GitHub] [incubator-tvm] FrozenGene commented on pull request #5754: [RFC] Improve quantized convolution performance for armv8 architectures

2020-06-10 Thread GitBox
FrozenGene commented on pull request #5754: URL: https://github.com/apache/incubator-tvm/pull/5754#issuecomment-642375802 cc @ajtulloch This is an automated message from the Apache Git Service. To respond to the message,

[GitHub] [incubator-tvm] FrozenGene commented on pull request #5754: [RFC] Improve quantized convolution performance for armv8 architectures

2020-06-10 Thread GitBox
FrozenGene commented on pull request #5754: URL: https://github.com/apache/incubator-tvm/pull/5754#issuecomment-642374967 Thanks for the great work! I have some quick question: 1. Have you tested various models arm cpu? (like A53, A72, A55, A75 and so on). According to fb qnnpack blog,