Method exec

dnn/src/cuda/conv_bias/group_conv.cpp:132–188

}

void ConvBiasForwardImpl::AlgoGroupConvGeneral::exec(const ExecArgs& args) const {
    auto bundle = get_workspace_bundle(args.workspace.raw_ptr, args);
    TensorND conv_dst_tensor = *args.dst_tensor;
    if (args.dst_layout->dtype.enumv() != args.bias_layout->dtype.enumv()) {
        conv_dst_tensor = TensorND{
                bundle.get(bundle.nr_workspace() - 1), args.dst_tensor->layout};
        conv_dst_tensor.layout.dtype = DType();
        args.opr->check_or_deduce_dtype_fwd(
                args.src_layout->dtype, args.filter_layout->dtype,
                conv_dst_tensor.layout.dtype);
    }
    {
        auto sub_args = args;
        sub_args.dst_tensor = &conv_dst_tensor;
        sub_args.dst_layout = &conv_dst_tensor.layout;

        auto config = prepare_sub_opr(sub_args);
        TensorND tsrc{args.src_tensor->raw_ptr(), config.first[0]};
        TensorND tfilter{args.filter_tensor->raw_ptr(), config.first[1]};
        TensorND tbias{args.bias_tensor->raw_ptr(), config.first[2]};
        TensorND tz{args.z_tensor->raw_ptr(), config.first[3]};
        TensorND tdst{conv_dst_tensor.raw_ptr(), config.first[4]};

        size_t c_pos;
        if (args.filter_meta.format == Param::Format::NCHW ||
            args.filter_meta.format == Param::Format::NCHW4) {
            c_pos = 1;
        } else {
            megdnn_assert(
                    args.filter_meta.format == Param::Format::NHWC,
                    "invalid conv format");
            c_pos = 3;
        }

        auto grp = args.filter_meta.group;

        auto&& fm = args.filter_meta;
        auto strd_src = tsrc.layout.stride[c_pos] * fm.icpg * tsrc.layout.dtype.size(),
             strd_dst = tdst.layout.stride[c_pos] * fm.ocpg * tdst.layout.dtype.size(),
             strd_flt = fm.icpg * fm.ocpg * fm.spatial[0] * fm.spatial[1] *
                        tfilter.layout.dtype.size();
        if (args.filter_meta.format == Param::Format::NCHW4) {
            strd_src >>= 2;
            strd_dst >>= 2;
        }
        for (uint32_t g = 0; g < grp; ++g) {
            config.second->exec(
                    tsrc, tfilter, tbias, tz, tdst, nullptr, bundle.get_workspace(0));
            incr_refp(tsrc.get_ref_ptr(), strd_src);
            incr_refp(tdst.get_ref_ptr(), strd_dst);
            incr_refp(tfilter.get_ref_ptr(), strd_flt);
        }
    }
    handle_bias_and_nonlinear(
            args.handle, args.nonlinear_mode, &conv_dst_tensor, args.dst_tensor,
            args.bias_tensor);
}