NVIDIA / cutlass

CUDA Templates for Linear Algebra Subroutines
Other
5.52k stars 941 forks source link

[QST] does cutlass depthwise conv support 1x1 filter #1598

Open tengdecheng opened 3 months ago

tengdecheng commented 3 months ago

hi, i have a question about depthwise conv with params like 1x1 filter, stride=1, pad=0, dilation=1, i have a compile error raise by checking kWarpGemmIterations in cutlass/conv/threadblock/depthwise_mma_base.h:

template <
    /// Size of the Gemm problem - concept: gemm::GemmShape<>
    typename Shape_,
    /// Policy describing tuning details (concept: MmaPolicy)
    typename Policy_,
    /// Number of stages,
    int Stages,
    /// Used for partial specialization
    typename Enable = bool>
class DepthwiseDirectConvMmaBase {
    using Policy = Policy_;

    /// Warp-level Mma
    using Operator = typename Policy::Operator;

    /// Shape describing the overall GEMM computed from shared memory
    /// by each warp.
    using WarpGemm = typename Policy::Operator::Shape;
    static int const kWarpGemmIterations = (WarpGemm::kK / Operator::Policy::MmaShape::kK);

    static_assert(kWarpGemmIterations > 1,
                    "The pipelined structure requires at least two warp-level "
                    "GEMM operations.");
};

so i want know how could i do to fix it, thanks !

github-actions[bot] commented 2 months ago

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.