Method exec

dnn/src/cuda/local_share/forward/batch_size_aware_chwn.cpp:68–89

}

void LocalShareForwardImpl::AlgoCHWNBatchSizeAware::exec(const ExecArgs& args) const {
    local_share::Param kern_param;
    auto&& param = args.opr->param();
    unpack_local_share_params(
            args.src_layout, args.filter_layout, args.dst_layout, param);
    kern_param.n = n, kern_param.co = co, kern_param.ci = ci, kern_param.hi = hi,
    kern_param.wi = wi, kern_param.ph = ph, kern_param.pw = pw,
    kern_param.grp_ho = ho / sgh, kern_param.grp_wo = wo / sgw, kern_param.sgh = sgh,
    kern_param.sgw = sgw;
    auto&& handle = concrete_handle(args.opr->handle());
    auto&& cublas_hdl = cublas_handle(args.opr->handle());
    auto&& stream = cuda_stream(args.opr->handle());

    auto one = handle->one_device();
    auto zero = handle->zero_device();

    local_share::_do_local_share_convolution_large_batch_size(
            args.src_tensor->ptr<dt_float32>(), args.filter_tensor->ptr<dt_float32>(),
            args.dst_tensor->ptr<dt_float32>(),
            reinterpret_cast<float*>(args.workspace.raw_ptr), fh, fw, sh, sw,
            kern_param, cublas_hdl, stream, one, zero);
}

// vim: syntax=cpp.doxygen