Backend/API: First IntelGPU backend based on clDNN with empty functions (#1116)

* First IntelGPU backend based on clDNN with empty functions

* Backend/API:Conflicts resolved and comments addressed

CMakeLists.txt

@@ -84,6 +84,7 @@ include(var_functions)
 option(NGRAPH_UNIT_TEST_ENABLE "Control the building of unit tests" TRUE)
 option(NGRAPH_TOOLS_ENABLE "Control the building of tool" TRUE)
 option(NGRAPH_CPU_ENABLE "Control the building of the CPU backend" TRUE)
+option(NGRAPH_INTELGPU_ENABLE "Control the building of the Intel GPU backend with clDNN" FALSE)
 option(NGRAPH_GPU_ENABLE "Control the building of the GPU backend" FALSE)
 option(NGRAPH_INTERPRETER_ENABLE "Control the building of the INTERPRETER backend" TRUE)
 option(NGRAPH_DISTRIBUTED_ENABLE "Add distributed mode to the CPU backend" FALSE)
@@ -175,6 +176,10 @@ if (NGRAPH_CPU_ENABLE)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DNGRAPH_CPU_ENABLE")
 endif()
 
+if (NGRAPH_INTELGPU_ENABLE)
+    unit_test_control(BACKEND INTELGPU MANIFEST src/ngraph/runtime/intelgpu/unit_test.manifest)
+endif()
+
 if (NGRAPH_GPU_ENABLE)
     unit_test_control(BACKEND GPU MANIFEST src/ngraph/runtime/gpu/unit_test.manifest)
 endif()

src/ngraph/runtime/CMakeLists.txt

@@ -20,6 +20,10 @@ if (NGRAPH_CPU_ENABLE)
     add_subdirectory(cpu)
 endif()
 
+if (NGRAPH_INTELGPU_ENABLE)
+    add_subdirectory(intelgpu)
+endif()
+
 if (NGRAPH_GPU_ENABLE)
     add_subdirectory(gpu)
 endif()

src/ngraph/runtime/intelgpu/CMakeLists.txt

+# ******************************************************************************
+# Copyright 2017-2018 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ******************************************************************************
+
+set(SRC
+    intelgpu_backend.cpp
+)
+
+if (NGRAPH_INTELGPU_ENABLE)
+    add_library(intelgpu_backend SHARED ${SRC})
+    set_target_properties(intelgpu_backend PROPERTIES VERSION ${NGRAPH_VERSION} SOVERSION ${NGRAPH_API_VERSION})
+    target_link_libraries(intelgpu_backend PUBLIC ngraph)
+
+    install(TARGETS intelgpu_backend LIBRARY DESTINATION ${NGRAPH_INSTALL_LIB})
+endif()

src/ngraph/runtime/intelgpu/intelgpu_backend.cpp

+/*******************************************************************************
+* Copyright 2017-2018 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#include "ngraph/runtime/intelgpu/intelgpu_backend.hpp"
+
+using namespace std;
+using namespace ngraph;
+
+extern "C" void create_backend(void)
+{
+    runtime::Backend::register_backend("INTELGPU",
+                                       make_shared<runtime::intelgpu::IntelGPUBackend>());
+};
+
+shared_ptr<runtime::TensorView>
+    runtime::intelgpu::IntelGPUBackend::create_tensor(const element::Type& element_type,
+                                                      const Shape& shape)
+{
+    throw runtime_error("IntelGPUBackend::create_tensor: Not implemented yet");
+}
+
+shared_ptr<runtime::TensorView> runtime::intelgpu::IntelGPUBackend::create_tensor(
+    const element::Type& element_type, const Shape& shape, void* memory_pointer)
+{
+    throw runtime_error("IntelGPUBackend::create_tensor: Not implemented yet");
+}
+
+bool runtime::intelgpu::IntelGPUBackend::compile(shared_ptr<Function> func)
+{
+    throw runtime_error("IntelGPUBackend::compile: Not implemented yet");
+}
+
+bool runtime::intelgpu::IntelGPUBackend::call(
+    shared_ptr<Function> func,
+    const vector<shared_ptr<runtime::TensorView>>& outputs,
+    const vector<shared_ptr<runtime::TensorView>>& inputs)
+{
+    throw runtime_error("IntelGPUBackend::call: Not implemented yet");
+}

src/ngraph/runtime/intelgpu/intelgpu_backend.hpp

+/*******************************************************************************
+* Copyright 2017-2018 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#pragma once
+
+#include <map>
+
+#include "ngraph/runtime/backend.hpp"
+
+namespace ngraph
+{
+    namespace runtime
+    {
+        namespace intelgpu
+        {
+            class IntelGPUBackend;
+        }
+    }
+}
+
+class ngraph::runtime::intelgpu::IntelGPUBackend : public runtime::Backend
+{
+public:
+    std::shared_ptr<ngraph::runtime::TensorView>
+        create_tensor(const ngraph::element::Type& element_type,
+                      const Shape& shape,
+                      void* memory_pointer) override;
+
+    std::shared_ptr<ngraph::runtime::TensorView>
+        create_tensor(const ngraph::element::Type& element_type, const Shape& shape) override;
+
+    bool compile(std::shared_ptr<Function> func) override;
+
+    bool call(std::shared_ptr<Function> func,
+              const std::vector<std::shared_ptr<runtime::TensorView>>& outputs,
+              const std::vector<std::shared_ptr<runtime::TensorView>>& inputs) override;
+
+private:
+};

src/ngraph/runtime/intelgpu/unit_test.manifest

+ab
+abc
+abc_int64
+abs
+acos
+aliased_output
+asin
+atan
+avg_pool_1d_1channel_1image
+avg_pool_1d_1channel_2image
+avg_pool_1d_2channel_2image
+avg_pool_2d_1channel_1image_padded
+avg_pool_2d_1channel_1image_strided
+avg_pool_2d_2channel_2image
+avg_pool_2d_2channel_2image_padded
+avg_pool_2d_2channel_2image_padded_3x3
+avg_pool_2d_2channel_2image_padded_3x3_strided
+avg_pool_2d_2channel_2image_padded_3x3_strided_uneven
+avg_pool_2d_2channel_2image_padded_only_above
+avg_pool_2d_2channel_2image_padded_only_below
+avg_pool_3d
+backwards_abc
+backwards_abs
+backwards_acos
+backwards_add
+backwards_add_nested
+backwards_asin
+backwards_atan
+backwards_avgpool_n1_c1_hw2x2
+backwards_avgpool_n1_c1_hw4x4
+backwards_avgpool_n2_c2_hw2x2_win_2x2_str_1x1_padding_numeric
+backwards_avgpool_n2_c2_hw4x4
+backwards_avgpool_n2_c2_hw4x4_numeric
+backwards_avgpool_n2_c2_hw4x4_win_2x2_str_1x1_numeric
+backwards_broadcast0
+backwards_broadcast1
+backwards_ceiling
+backwards_concat_axis_0
+backwards_concat_axis_1
+backwards_concat_vector
+backwards_cos
+backwards_cosh
+backwards_divide
+backwards_dot_scalar_scalar
+backwards_dot_scalar_tensor
+backwards_dot_tensor2_tensor2
+backwards_dot_tensor3_tensor3
+backwards_dot_tensor_scalar
+backwards_dot_tensor_vector
+backwards_dot_vector_vector
+backwards_exp
+backwards_floor
+backwards_log
+backwards_maximum
+backwards_maxpool_n2_c1_hw5_3x3_str2_max
+backwards_maxpool_n2c1h5w5_kh3kw3_sh2sw2
+backwards_maxpool_n4_c1_hw4_2x2_max
+backwards_maxpool_n4c1h4w4_kh2kw2_sh1sw1
+backwards_minimum
+backwards_multiply
+backwards_negative
+backwards_parameter
+backwards_power
+backwards_relu
+backwards_replace_slice
+backwards_reshape
+backwards_reverse_3d_02
+backwards_reverse_sequence_n3_c2_h3
+backwards_reverse_sequence_n4d2c3h2w2
+backwards_select
+backwards_select_nested
+backwards_sign
+backwards_sin
+backwards_sinh
+backwards_slice
+backwards_softmax_3d
+backwards_softmax_all
+backwards_softmax_axis
+backwards_softmax_underflow
+backwards_subtract
+backwards_sum_m2s
+backwards_sum_m2v_0
+backwards_sum_m2v_1
+backwards_sum_v2s
+backwards_tan
+backwards_tanh
+batch_norm_one_output
+batch_norm_three_outputs
+batchnorm_bprop_n4c3h2w2
+batchnorm_fprop_b1c2h2w2
+batchnorm_fprop_b2c2h2w1
+batchnorm_fprop_globalstats_b2c2w2h1
+batchnorm_fprop_inference_b2c2h2w1
+broadcast_matrix_0
+broadcast_matrix_1
+broadcast_matrix_2
+broadcast_scalar_matrix
+broadcast_scalar_tensor
+broadcast_scalar_vector
+broadcast_trivial
+broadcast_vector_colwise
+broadcast_vector_rowwise
+broadcast_vector_rowwise_int64
+broadcast_vector_rowwise_reversed
+ceiling
+computation_reuse
+concat_2d_tensor
+concat_4d_tensor
+concat_5d
+concat_matrix_colwise
+concat_matrix_int64
+concat_matrix_rowwise
+concat_vector
+constant_broadcast
+constant_equality_bool
+constant_multi_use
+convert_float32_bool
+convert_int32_bool
+convert_int32_float32
+convolution_2d_1item
+convolution_2d_1item_1o1i_data_dilated
+convolution_2d_1item_2o1i_data_dilated
+convolution_2d_1item_2o2i_data_dilated
+convolution_2d_1item_5o3i_data_dilated
+convolution_2d_1item_padded_1_1x1_1
+convolution_2d_1item_padded_2_3x4_5
+convolution_2d_2item_5o3i_data_dilated
+convolution_2d_2items
+convolution_2d_2items_dilated
+convolution_2d_2items_dilated_padded
+convolution_2d_2items_strided
+convolution_2d_2items_strided_padded
+convolution_2d_2items_strided_padded_same
+convolution_2d_8item_large_5o3i_data_dilated
+convolution_2d_8item_large_5o3i_uneven_filter_data_dilated
+convolution_2d_8item_large_5o3i_uneven_filter_uneven_data_dilation_data_dilated
+convolution_3d_1item_large_5o3i_padded_uneven_filter_uneven_data_dilation_data_dilated
+convolution_3d_2item_large_5o3i_padded_strided_uneven_filter_uneven_data_dilation_data_dilated
+convolution_3d_2item_large_5o3i_padded_strided_uneven_filter_uneven_data_dilation_filter_dilated_data_dilated
+convolution_3d_2item_large_5o3i_uneven_filter_uneven_data_dilation_data_dilated
+convolution_3d_2items
+convolution_4d_2items
+convolution_4d_4items
+convolution_4d_4items_dilated
+convolution_4d_4items_padded_neg
+convolution_4d_4items_strided
+convolution_4d_4items_strided_dilated
+convolution_4d_4items_strided_dilated_padded
+convolution_4d_4items_strided_dilated_padded_neg
+convolution_4d_4items_strided_dilated_padded_same
+convolution_outlining
+cos
+cosh
+divide
+divide_adjoint_stability
+divide_by_zero_float32
+divide_by_zero_int32
+dot1d
+dot2d
+dot3d_2d
+dot3d_3d
+dot_0_0
+dot_2x0_0
+dot_matrix_0x2_2x0
+dot_matrix_2x0_0x2
+dot_matrix_3x2_2x0
+dot_matrix_vector
+dot_matrix_vector_4_3
+dot_matrix_vector_int64
+dot_scalar_0x2
+dot_scalar_scalar
+dot_scalar_tensor_arg0
+dot_scalar_tensor_arg1
+equal
+exp
+floor
+function_call
+function_name
+fuse_max_with_constant_zero_input_as_relu
+greater
+greatereq
+kahan_sum_3d_to_vector
+kahan_sum_to_scalar
+less
+lesseq
+lesseq_bool
+log
+logical_and
+logical_or
+max_3d_eliminate_zero_dim
+max_3d_to_matrix_least_sig
+max_3d_to_matrix_most_sig
+max_3d_to_scalar
+max_3d_to_vector
+max_matrix_cols_zero
+max_matrix_columns
+max_matrix_rows
+max_matrix_rows_zero
+max_matrix_to_scalar_zero_by_zero
+max_pool_1d_1channel_1image
+max_pool_1d_1channel_2image
+max_pool_1d_2channel_2image
+max_pool_2d_1channel_1image_overpadded
+max_pool_2d_1channel_1image_padded
+max_pool_2d_1channel_1image_padded_negative_values
+max_pool_2d_1channel_1image_strided
+max_pool_2d_2channel_2image
+max_pool_3d
+max_to_scalar
+max_trivial
+max_trivial_5d
+max_vector_zero
+maximum
+min_3d_eliminate_zero_dim
+min_3d_to_matrix_least_sig
+min_3d_to_matrix_most_sig
+min_3d_to_scalar
+min_3d_to_vector
+min_matrix_cols_zero
+min_matrix_columns
+min_matrix_rows
+min_matrix_rows_zero
+min_matrix_to_scalar_zero_by_zero
+min_to_scalar
+min_trivial
+min_trivial_5d
+min_vector_zero
+minimum
+mkldnn_layouts
+multiple_backends
+multiple_result
+negative
+node_name
+not
+notequal
+numeric_double_inf
+numeric_double_nan
+numeric_float_inf
+numeric_float_nan
+one_hot_matrix_0
+one_hot_scalar_0_in_3
+one_hot_scalar_1_in_3
+one_hot_scalar_2_in_3
+one_hot_scalar_fp_nonint_in_3
+one_hot_scalar_oob_in_3
+one_hot_vector_0
+one_hot_vector_1
+one_hot_vector_1_barely_oob
+one_hot_vector_1_far_oob
+one_hot_vector_1_fp
+one_hot_vector_1_fp_nonint
+pad_exterior_1d
+pad_exterior_2d_0x0
+pad_exterior_2d_0x3
+pad_exterior_2d_3x0
+pad_exterior_4d_1x2x2x2
+pad_interior_1d
+pad_interior_exterior_1d
+pad_interior_exterior_2d
+pad_interior_exterior_4d_2x0x3x2
+parameter_as_output
+power
+product_3d_eliminate_zero_dim
+product_3d_to_matrix_least_sig
+product_3d_to_matrix_most_sig
+product_3d_to_scalar
+product_3d_to_vector
+product_matrix_cols_zero
+product_matrix_columns
+product_matrix_rows
+product_matrix_rows_zero
+product_matrix_to_scalar_zero_by_zero
+product_to_scalar
+product_trivial
+product_trivial_5d
+product_vector_zero
+reduce_3d_to_vector
+reduce_matrix_cols_zero
+reduce_matrix_columns
+reduce_matrix_rows
+reduce_matrix_rows_zero
+reduce_matrix_to_scalar_zero_by_zero
+reduce_to_scalar
+reduce_trivial
+reduce_vector_zero
+reduce_window_emulating_max_pool_1d_1channel_1image
+reduce_window_emulating_max_pool_1d_1channel_2image
+reduce_window_emulating_max_pool_1d_2channel_2image
+reduce_window_emulating_max_pool_2d_1channel_1image_strided
+reduce_window_emulating_max_pool_2d_2channel_2image
+relu_2Dbackprop
+relu_2Dfprop
+relu_4Dbackprop
+relu_4Dfprop
+replace_slice_3d
+replace_slice_3d_strided
+replace_slice_3d_strided_different_strides
+replace_slice_matrix
+replace_slice_scalar
+replace_slice_vector
+reshape_6d
+reshape_m2m_dim_change_transpose
+reshape_m2m_same
+reshape_m2m_transpose
+reshape_s2t
+reshape_t2s_012
+reshape_t2s_120
+reshape_t2v_012
+reshape_v2m_col
+reshape_v2m_row
+reshape_v2t_middle
+reverse_0d
+reverse_1d_0
+reverse_1d_nochange
+reverse_2d_0
+reverse_2d_01
+reverse_2d_1
+reverse_2d_nochange
+reverse_3d_0
+reverse_3d_01
+reverse_3d_012
+reverse_3d_02
+reverse_3d_1
+reverse_3d_12
+reverse_3d_2
+reverse_3d_nochange
+reverse_sequence_n2c3h4w2
+reverse_sequence_n4c3h2w2
+reverse_sequence_n4d2c3h2w2
+scalar_constant_float32
+scalar_constant_int64
+select
+select_and_scatter_3d_without_overlap
+select_and_scatter_with_overlap
+select_and_scatter_without_overlap
+sign
+sin
+sinh
+slice_3d
+slice_3d_strided
+slice_3d_strided_different_strides
+slice_matrix
+slice_matrix_strided
+slice_scalar
+slice_vector
+softmax_all
+softmax_axis
+softmax_axis_2
+softmax_axis_3d
+softmax_axis_3d_trivial
+softmax_underflow
+sqrt
+subtract
+sum_3d_eliminate_zero_dim
+sum_3d_to_matrix_least_sig
+sum_3d_to_matrix_most_sig
+sum_3d_to_scalar
+sum_3d_to_vector
+sum_5d_to_scalar
+sum_matrix_cols_zero
+sum_matrix_columns
+sum_matrix_rows
+sum_matrix_rows_zero
+sum_matrix_to_scalar_zero_by_zero
+sum_to_scalar
+sum_trivial
+sum_trivial_5d
+sum_vector_zero
+tan
+tanh
+tensor_constant
+tensor_constant_float32
+tensor_constant_int64
+tensor_constant_with_op
+tensorview_custom_mem
+validate_call_input_count
+validate_call_input_shape
+validate_call_input_type
+validate_call_output_count
+validate_call_output_shape
+validate_call_output_type
+zero_sized_abs
+zero_sized_acos
+zero_sized_add
+zero_sized_asin
+zero_sized_atan
+zero_sized_ceiling
+zero_sized_cos
+zero_sized_cosh
+zero_sized_divide
+zero_sized_eq
+zero_sized_floor
+zero_sized_greater
+zero_sized_greatereq
+zero_sized_less
+zero_sized_lesseq
+zero_sized_log
+zero_sized_maximum
+zero_sized_minimum
+zero_sized_multiply
+zero_sized_negative
+zero_sized_not
+zero_sized_not_equal
+zero_sized_power
+zero_sized_sign
+zero_sized_sin
+zero_sized_sinh
+zero_sized_sqrt
+zero_sized_subtract
+zero_sized_tan
+zero_sized_tanh

src/tools/nbench/CMakeLists.txt

@@ -26,6 +26,9 @@ target_link_libraries(nbench ngraph)
 if (NGRAPH_CPU_ENABLE)
     target_link_libraries(nbench cpu_backend)
 endif()
+if (NGRAPH_INTELGPU_ENABLE)
+    target_link_libraries(nbench intelgpu_backend)
+endif()
 if (NGRAPH_GPU_ENABLE)
     target_link_libraries(nbench gpu_backend)
 endif()