piano_123

using ThreadblockShape0 = cutlass::gemm::GemmShape; using WarpShape0 = cutlass::gemm::GemmShape; using InstructionShape = cutlass::gemm::GemmShape; I have a doubt how to determine the values of these three variables（ThreadblockShape，WarpShape， InstructionShape） and what are they...

@hwu36 it works，thanks. And I have a doubt: for input Activation=f16:nhwc --Filter=f16:nhwc --Output=f16 --accumulator-type=f16, the input and output channel requirements are multiples of 8, and for channels that are not...

For the situation:input Activation=f16:nhwc --Filter=f16:nhwc --Output=f16 --accumulator-type=f16 , Does setting accumulator-type=f16 cause overflow?

Hi @hwu36 ,I still have a doubt about Silu. In Example 17 ,under the same configuration except EpilogueOp, I just replaced EpilogueOp from LinearCombination to LinearCombinationSilu, and the run time...

@hwu36 Hi，does cutlass support dethpthwise convlution? I haven't seen the introduction of dethwise in the documentation, sample and header files. I set group=inchannel=outchannel in the conv problem size in example...

@hwu36 “You can set the channel number as 1 just like what you did” means set group=inchannel=outchannel ? Do you plan to do efficient depthwise conv in the future?

@quic-ssiddego How can I run successfully the script in the video and where I can find the script?

@exiawsh Hi ， would you like to explain more detail? In PETRV2, the loaded data camera=12, after backbone processing, is equivalent to the 2d feature map concat of the t-1...

@exiawsh When i run test.py,https://github.com/megvii-research/PETR/blob/e48faec8aa24bdd14f95692428ddc4982f2f71cb/projects/mmdet3d_plugin/datasets/pipelines/loading.py#L114 choices=14,and this is not key frame , so I wondering if it can directly use the backbone results of the previous key frame, saving half...

@exiawsh Thanks a lot！