Method exec

dnn/src/cuda/roi_align/opr_impl.cpp:10–51

namespace cuda {

void ROIAlignForwardImpl::exec(
        _megdnn_tensor_in src, _megdnn_tensor_in rois, _megdnn_tensor_out dst,
        _megdnn_tensor_out index, _megdnn_workspace workspace) {
    check_exec(src.layout, rois.layout, dst.layout, index.layout, workspace.size);
    auto stream = cuda_stream(handle());
    int nthreads = dst.layout.total_nr_elems();
    float spatial_scale = param().spatial_scale;
    float offset = param().offset;
    int sample_height = param().sample_height;
    int sample_width = param().sample_width;
    int channels = src.layout.shape[1];
    int height = src.layout.shape[2];
    int width = src.layout.shape[3];
    int pooled_height = dst.layout.shape[2];
    int pooled_width = dst.layout.shape[3];
    using namespace ::megdnn::roi_align;
    using namespace ::megdnn::cuda::roi_align;
#define cb(DType)                                                                  \
    if (src.layout.dtype == DType()) {                                             \
        using T = typename DTypeTrait<DType>::ctype;                               \
        switch (param().mode) {                                                    \
            case param::ROIAlign::Mode::MAX:                                       \
                forward_proxy<T, MaxPooler<T>>(                                    \
                        nthreads, src.ptr<T>(), spatial_scale, offset, channels,   \
                        height, width, pooled_height, pooled_width, sample_height, \
                        sample_width, rois.ptr<T>(), dst.ptr<T>(),                 \
                        index.ptr<dt_int32>(), stream);                            \
                break;                                                             \
            case param::ROIAlign::Mode::AVERAGE:                                   \
                forward_proxy<T, AveragePooler<T>>(                                \
                        nthreads, src.ptr<T>(), spatial_scale, offset, channels,   \
                        height, width, pooled_height, pooled_width, sample_height, \
                        sample_width, rois.ptr<T>(), dst.ptr<T>(),                 \
                        index.ptr<dt_int32>(), stream);                            \
                break;                                                             \
            default:                                                               \
                megdnn_assert_internal(false);                                     \
        }                                                                          \
    }
    MEGDNN_FOREACH_COMPUTING_DTYPE_FLOAT(cb)
#undef cb
}

void ROIAlignBackwardImpl::exec(
        _megdnn_tensor_in diff, _megdnn_tensor_in rois, _megdnn_tensor_in index,