kice edited a comment on issue #4523: Optimization for subpixel layer on Tensor core URL: https://github.com/apache/incubator-tvm/issues/4523#issuecomment-569522065 After testing it, I am happy to let you know, we have no significant difference at all. And I even found a bug. xD TVM build on Win10 MSVC, CUDA 10.1, Test with RTX 2060 Super ``` v0.0.4 fn (%data: Tensor[(1, 3, 720, 1280), float16], %head.0.weight: Tensor[(64, 3, 3, 3), float16], %head.0.bias: Tensor[(64), float16], %body.0.body.0.weight: Tensor[(64, 64, 3, 3), float16], %body.0.body.0.bias: Tensor[(64), float16], %body.0.body.2.weight: Tensor[(64, 64, 3, 3), float16], %body.0.body.2.bias: Tensor[(64), float16], %body.1.body.0.weight: Tensor[(64, 64, 3, 3), float16], %body.1.body.0.bias: Tensor[(64), float16], %body.1.body.2.weight: Tensor[(64, 64, 3, 3), float16], %body.1.body.2.bias: Tensor[(64), float16], %body.2.weight: Tensor[(64, 64, 3, 3), float16], %body.2.bias: Tensor[(64), float16], %tail.0.0.weight: Tensor[(128, 64, 3, 3), float16], %tail.0.0.bias: Tensor[(128), float16], %tail.0.2.weight: Tensor[(64, 32, 3, 3), float16], %tail.0.2.bias: Tensor[(64), float16], %tail.1.weight: Tensor[(3, 16, 3, 3), float16], %tail.1.bias: Tensor[(3), float16]) -> Tensor[(1, 3, 2880, 5120), float16] { %0 = nn.conv2d(%data, %head.0.weight, padding=[1, 1], kernel_size=[3, 3]) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %1 = nn.bias_add(%0, %head.0.bias) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %2 = nn.conv2d(%1, %body.0.body.0.weight, padding=[1, 1], kernel_size=[3, 3]) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %3 = nn.bias_add(%2, %body.0.body.0.bias) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %4 = nn.leaky_relu(%3, alpha=0.1f) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %5 = nn.conv2d(%4, %body.0.body.2.weight, padding=[1, 1], kernel_size=[3, 3]) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %6 = nn.bias_add(%5, %body.0.body.2.bias) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %7 = add(%6, %1) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %8 = nn.conv2d(%7, %body.1.body.0.weight, padding=[1, 1], kernel_size=[3, 3]) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %9 = nn.bias_add(%8, %body.1.body.0.bias) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %10 = nn.leaky_relu(%9, alpha=0.1f) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %11 = nn.conv2d(%10, %body.1.body.2.weight, padding=[1, 1], kernel_size=[3, 3]) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %12 = nn.bias_add(%11, %body.1.body.2.bias) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %13 = add(%12, %7) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %14 = nn.conv2d(%13, %body.2.weight, padding=[1, 1], kernel_size=[3, 3]) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %15 = nn.bias_add(%14, %body.2.bias) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %16 = add(%15, %1) /* ty=Tensor[(1, 64, 720, 1280), float16] */; %17 = nn.conv2d(%16, %tail.0.0.weight, padding=[1, 1], kernel_size=[3, 3]) /* ty=Tensor[(1, 128, 720, 1280), float16] */; %18 = nn.bias_add(%17, %tail.0.0.bias) /* ty=Tensor[(1, 128, 720, 1280), float16] */; %19 = nn.depth_to_space(%18, block_size=2, mode="CRD") /* ty=Tensor[(1, 32, 1440, 2560), float16] */; %20 = nn.conv2d(%19, %tail.0.2.weight, padding=[1, 1], kernel_size=[3, 3]) /* ty=Tensor[(1, 64, 1440, 2560), float16] */; %21 = nn.bias_add(%20, %tail.0.2.bias) /* ty=Tensor[(1, 64, 1440, 2560), float16] */; %22 = nn.depth_to_space(%21, block_size=2, mode="CRD") /* ty=Tensor[(1, 16, 2880, 5120), float16] */; %23 = nn.conv2d(%22, %tail.1.weight, padding=[1, 1], kernel_size=[3, 3]) /* ty=Tensor[(1, 3, 2880, 5120), float16] */; nn.bias_add(%23, %tail.1.bias) /* ty=Tensor[(1, 3, 2880, 5120), float16] */ } ```
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
