Method exec_internal

dnn/src/cuda/convolution/backward_data/matmul.cpp:97–146

template <typename T>
void ConvolutionBackwardDataImpl::AlgoMatmul::exec_internal(const ExecArgs& args) {
    auto&& fm = args.filter_meta;
    size_t N = args.grad_layout->shape[0], IC = fm.icpg,
           IH = args.grad_layout->shape[2], IW = args.grad_layout->shape[3],
           OC = fm.ocpg, OH = args.diff_layout->shape[2],
           OW = args.diff_layout->shape[3], FH = fm.spatial[0], FW = fm.spatial[1],
           PH = fm.padding[0], PW = fm.padding[1], SH = fm.stride[0], SW = fm.stride[1],
           DH = fm.dilation[0], DW = fm.dilation[1];
    auto stream = cuda_stream(args.handle);

    auto config = prepare_sub_opr(args);

    auto&& sizes = matmul_get_workspace_bundle(args.as_fwd_args());
    sizes.push_back(config.second->get_workspace_in_bytes(
            config.first[0], config.first[1], config.first[2]));
    auto wbundle = WorkspaceBundle(args.workspace.raw_ptr, sizes);

    T* diff_t = static_cast<T*>(wbundle.get(0));
    T* col = static_cast<T*>(wbundle.get(1));
    {
        // transpose diff
        TensorLayout froml({N, OC * OH * OW}, typename DTypeTrait<T>::dtype()),
                tol(froml);
        froml.stride[0] = args.diff_layout->stride[0];
        tol.stride[0] = 1;
        tol.stride[1] = N;
        TensorND from(args.diff_tensor->ptr<T>(), froml), to(diff_t, tol);
        args.handle->relayout_opr()->exec(from, to);
    }
    {
        // take gemm grad
        TensorLayout Al({OC, IC * FH * FW}, typename DTypeTrait<T>::dtype()),
                Bl({IC * FH * FW, OH * OW * N}, typename DTypeTrait<T>::dtype()),
                Cl({OC, OH * OW * N}, typename DTypeTrait<T>::dtype());
        TensorND A(args.filter_tensor->ptr<T>(), Al), B(col, Bl), C(diff_t, Cl);
        if (fm.should_flip) {
            convolution::flip_filter(
                    args.as_fwd_args(), wbundle.get_workspace(2), A.get_ref_ptr());
            config.second->exec(A, C, B, wbundle.get_workspace(3));
        } else {
            config.second->exec(A, C, B, wbundle.get_workspace(2));
        }
    }
    {
        // col2im
        convolution::col2im<T>(
                col, args.grad_tensor->ptr<T>(), N, args.grad_layout->stride[0], IC, IH,
                IW, FH, FW, OH, OW, PH, PW, SH, SW, DH, DW, stream);
    }
}

// vim: syntax=cpp.doxygen