CMakeLists.txt

# NOTE we leave the add_kernel not protected by LITE_WITH_LIGHT_WEIGHT_FRAMEWORK so that all the kernels will be registered
# to the model_optimize_tool.
if((NOT LITE_ON_MODEL_OPTIMIZE_TOOL) AND (NOT LITE_WITH_PYTHON) AND (NOT (LITE_WITH_LIGHT_WEIGHT_FRAMEWORK AND LITE_WITH_ARM)))
    return()
endif()

message(STATUS "compile with lite ARM kernels")

# 1. basic kernels for basic models
# for conv op
add_kernel(conv_depthwise ARM basic SRCS conv_depthwise.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(conv_direct ARM basic SRCS conv_direct.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(conv_gemmlike ARM basic SRCS conv_gemmlike.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(conv_winograd ARM basic SRCS conv_winograd.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(conv_compute_arm ARM basic SRCS conv_compute.cc DEPS ${lite_kernel_deps}
        conv_depthwise conv_direct conv_gemmlike conv_winograd)

add_kernel(fc_compute_arm ARM basic SRCS fc_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(activation_compute_arm ARM basic SRCS activation_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(mul_compute_arm ARM basic SRCS mul_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(matmul_compute_arm ARM basic SRCS matmul_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(scale_compute_arm ARM basic SRCS scale_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(softmax_compute_arm ARM basic SRCS softmax_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(batch_norm_compute_arm ARM basic SRCS batch_norm_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(elementwise_compute_arm ARM basic SRCS elementwise_compute.cc DEPS ${lite_kernel_deps} math_arm)

add_kernel(pool_compute_arm ARM basic SRCS pool_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(split_compute_arm ARM basic SRCS split_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(concat_compute_arm ARM basic SRCS concat_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(pad2d_compute_arm ARM basic SRCS pad2d_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(prior_box_compute_arm ARM basic SRCS prior_box_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(calib_compute_arm ARM basic SRCS calib_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(transpose_compute_arm ARM basic SRCS transpose_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(yolo_box_compute_arm ARM basic SRCS yolo_box_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(shuffle_channel_compute_arm ARM basic SRCS shuffle_channel_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(argmax_compute_arm ARM basic SRCS argmax_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(conv_transpose_compute_arm ARM basic SRCS conv_transpose_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(interpolate_compute_arm ARM basic SRCS interpolate_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(box_coder_compute_arm ARM basic SRCS box_coder_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(slice_compute_arm ARM basic SRCS slice_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(cast_compute_arm ARM basic SRCS cast_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(squeeze_compute_arm ARM basic SRCS squeeze_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(reduce_mean_compute_arm ARM basic SRCS reduce_mean_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(stack_compute_arm ARM basic SRCS stack_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(affine_channel_compute_arm ARM basic SRCS affine_channel_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(range_compute_arm ARM basic SRCS range_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(dropout_compute_arm ARM basic SRCS dropout_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(layout_compute_arm ARM basic SRCS layout_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(instance_norm_compute_arm ARM basic SRCS instance_norm_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(grid_sampler_compute_arm ARM basic SRCS grid_sampler_compute.cc DEPS ${lite_kernel_deps} math_arm)

## 2.other basic kernels: basic kernels that not used in basic models
add_kernel(negative_compute_arm ARM extra SRCS negative_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(crop_compute_arm ARM extra SRCS crop_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(power_compute_arm ARM extra SRCS power_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(norm_compute_arm ARM extra SRCS norm_compute.cc DEPS ${lite_kernel_deps} math_arm)

## 3. extra kernels
add_kernel(lrn_compute_arm ARM extra SRCS lrn_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(decode_bboxes_compute_arm ARM extra SRCS decode_bboxes_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(density_prior_box_compute_arm ARM basic SRCS density_prior_box_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(axpy_compute_arm ARM extra SRCS axpy_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(reduce_max_compute_arm ARM extra SRCS reduce_max_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(sequence_expand_compute_arm ARM extra SRCS sequence_expand_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(im2sequence_compute_arm ARM extra SRCS im2sequence_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(sequence_pool_compute_arm ARM extra SRCS sequence_pool_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(sequence_conv_compute_arm ARM extra SRCS sequence_conv_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(layer_norm_compute_arm ARM extra SRCS layer_norm_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(gather_compute_arm ARM extra SRCS gather_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(reduce_prod_compute_arm ARM extra SRCS reduce_prod_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(split_lod_tensor_compute_arm ARM extra SRCS split_lod_tensor_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(merge_lod_tensor_compute_arm ARM extra SRCS merge_lod_tensor_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(anchor_generator_compute_arm ARM extra SRCS anchor_generator_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(generate_proposals_compute_arm ARM extra SRCS generate_proposals_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(roi_align_compute_arm ARM extra SRCS roi_align_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(box_clip_compute_arm ARM extra SRCS box_clip_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(assign_value_compute_arm ARM basic SRCS assign_value_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(conditional_block_compute_arm ARM extra SRCS conditional_block_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(collect_fpn_proposals_compute_arm ARM extra SRCS collect_fpn_proposals_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(distribute_fpn_proposals_compute_arm ARM extra SRCS distribute_fpn_proposals_compute.cc DEPS ${lite_kernel_deps} math_arm)

# for OCR specific
add_kernel(gru_unit_compute_arm ARM extra SRCS gru_unit_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(gru_compute_arm ARM extra SRCS gru_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(beam_search_decode_compute_arm ARM extra SRCS beam_search_decode_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(lookup_table_compute_arm ARM extra SRCS lookup_table_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(lookup_table_dequant_compute_arm ARM extra SRCS lookup_table_dequant_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(sequence_softmax_compute_arm ARM extra SRCS sequence_softmax_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(while_compute_arm ARM extra SRCS while_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(topk_compute_arm ARM extra SRCS topk_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(increment_compute_arm ARM extra SRCS increment_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(beam_search_compute_arm ARM extra SRCS beam_search_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(fill_constant_compute_arm ARM basic SRCS fill_constant_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(fill_constant_batch_size_like_compute_arm ARM basic SRCS fill_constant_batch_size_like_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(lod_reset_compute_arm ARM extra SRCS lod_reset_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(lstm_arm ARM extra SRCS lstm_compute.cc DEPS ${lite_kernel_deps} math_arm)

# 4. training kernels
add_kernel(mean_compute_arm ARM extra SRCS mean_compute.cc DEPS ${lite_kernel_deps} math_arm)

add_kernel(mean_grad_compute_arm ARM train SRCS mean_grad_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(activation_grad_compute_arm ARM train SRCS activation_grad_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(elementwise_grad_compute_arm ARM train SRCS elementwise_grad_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(mul_grad_compute_arm ARM train SRCS mul_grad_compute.cc DEPS ${lite_kernel_deps} math_arm)
add_kernel(sgd_compute_arm ARM train SRCS sgd_compute.cc DEPS ${lite_kernel_deps} math_arm)

lite_cc_test(test_scale_compute_arm SRCS scale_compute_test.cc DEPS scale_compute_arm)
lite_cc_test(test_softmax_compute_arm SRCS softmax_compute_test.cc DEPS softmax_compute_arm)
lite_cc_test(test_batch_norm_compute_arm SRCS batch_norm_compute_test.cc DEPS batch_norm_compute_arm)
lite_cc_test(test_elementwise_compute_arm SRCS elementwise_compute_test.cc DEPS elementwise_compute_arm)
lite_cc_test(test_pool_compute_arm SRCS pool_compute_test.cc DEPS pool_compute_arm)
lite_cc_test(test_mul_compute_arm SRCS mul_compute_test.cc DEPS mul_compute_arm)
lite_cc_test(test_split_compute_arm SRCS split_compute_test.cc DEPS split_compute_arm)
lite_cc_test(test_concat_compute_arm SRCS concat_compute_test.cc DEPS concat_compute_arm)
lite_cc_test(test_transpose_compute_arm SRCS transpose_compute_test.cc DEPS transpose_compute_arm COMPILE_LEVEL extra)
lite_cc_test(test_argmax_compute_arm SRCS argmax_compute_test.cc DEPS argmax_compute_arm)
lite_cc_test(test_dropout_compute_arm SRCS dropout_compute_test.cc DEPS dropout_compute_arm)
if(LITE_BUILD_EXTRA)
    lite_cc_test(test_split_lod_tensor_compute_arm SRCS split_lod_tensor_compute_test.cc DEPS split_lod_tensor_compute_arm)
    lite_cc_test(test_merge_lod_tensor_compute_arm SRCS merge_lod_tensor_compute_test.cc DEPS merge_lod_tensor_compute_arm)
    lite_cc_test(test_lrn_compute_arm SRCS lrn_compute_test.cc DEPS lrn_compute_arm)
    lite_cc_test(test_decode_bboxes_compute_arm SRCS decode_bboxes_compute_test.cc DEPS decode_bboxes_compute_arm)
    lite_cc_test(test_axpy_compute_arm SRCS axpy_compute_test.cc DEPS axpy_compute_arm)
    lite_cc_test(test_layer_norm_compute_arm SRCS layer_norm_compute_test.cc DEPS layer_norm_compute_arm)
endif()