Implement FP32/16 Depthwise Conv2d operator in dynamic fusion

This patch adds Depthwise Conv2d operator into dynamic fusion interface and adds the associated tests.

Resolves: COMPMID-5517
Change-Id: I385c94dff7fd40c72b8337ef797e508df4499a82
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8678
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 395 insertions, 6 deletions

diff --git a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDepthwiseConv2d.cpp b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDepthwiseConv2d.cpp
new file mode 100644
index 0000000000..5626093079
--- /dev/null
+++ b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDepthwiseConv2d.cpp
@@ -0,0 +1,220 @@
+/*
+ * Copyright (c) 2022 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "ClComponentDepthwiseConv2d.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/dynamic_fusion/sketch/attributes/DepthwiseConv2dAttributes.h"
+#include "src/core/CL/CLValidate.h"
+#include "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDepthwiseConv2d.h"
+
+namespace arm_compute
+{
+namespace experimental
+{
+namespace dynamic_fusion
+{
+using Settings = ClComponentDepthwiseConv2dSettings;
+
+Settings &Settings::export_input_to_cl_image(bool cl_image)
+{
+    _export_input_to_cl_image = cl_image;
+    return *this;
+}
+
+bool Settings::export_input_to_cl_image() const
+{
+    return _export_input_to_cl_image;
+}
+
+Settings &Settings::export_weights_to_cl_image(bool cl_image)
+{
+    _export_weights_to_cl_image = cl_image;
+    return *this;
+}
+
+bool Settings::export_weights_to_cl_image() const
+{
+    return _export_weights_to_cl_image;
+}
+
+Settings &Settings::fast_relaxed_math(bool fast_relaxed_math)
+{
+    _fast_relaxed_math = fast_relaxed_math;
+    return *this;
+}
+
+bool Settings::fast_relaxed_math() const
+{
+    return _fast_relaxed_math;
+}
+
+Settings &Settings::is_fma_available(bool is_fma_available)
+{
+    _is_fma_available = is_fma_available;
+    return *this;
+}
+
+bool Settings::is_fma_available() const
+{
+    return _is_fma_available;
+}
+
+Settings &Settings::n0(unsigned int n0)
+{
+    _n0 = n0;
+    return *this;
+}
+
+unsigned int Settings::n0() const
+{
+    return _n0;
+}
+
+Settings &Settings::m0(unsigned int m0)
+{
+    _m0 = m0;
+    return *this;
+}
+
+unsigned int Settings::m0() const
+{
+    return _m0;
+}
+
+Status ClComponentDepthwiseConv2d::validate(
+    const Properties                &properties,
+    const ArgumentPack<ITensorInfo> &tensors,
+    const Attributes                &attributes,
+    const Settings                  &settings)
+{
+    ARM_COMPUTE_UNUSED(properties, settings);
+    const auto src = tensors.get_const_tensor(TensorType::ACL_SRC_0);
+    const auto wei = tensors.get_const_tensor(TensorType::ACL_SRC_1);
+    const auto bia = tensors.get_const_tensor(TensorType::ACL_SRC_2);
+    const auto dst = tensors.get_const_tensor(TensorType::ACL_DST_0);
+
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, wei, dst);
+
+    // 1. Check validity
+    // Matching data type
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, wei);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
+    if(bia != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bia);
+    }
+
+    // Matching data layout
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, wei);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, dst);
+    if(bia != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, bia);
+    }
+
+    // All tensor infos are initialized
+    ARM_COMPUTE_RETURN_ERROR_ON(src->tensor_shape().total_size() == 0);
+    ARM_COMPUTE_RETURN_ERROR_ON(wei->tensor_shape().total_size() == 0);
+    ARM_COMPUTE_RETURN_ERROR_ON(dst->tensor_shape().total_size() == 0);
+    if(bia != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(bia->tensor_shape().total_size() == 0);
+    }
+    // Device requirements are met
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
+    // wei shape is correct
+    const DataLayout data_layout = src->data_layout();
+    const size_t     channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+
+    ARM_COMPUTE_RETURN_ERROR_ON(wei->dimension(channel_idx) != (src->dimension(channel_idx) * attributes.depth_multiplier()));
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(wei->num_dimensions() > 3, "Weights can be at most 3 dimensional");
+
+    // dst shape is correct
+    const PadStrideInfo pad_stride_info = PadStrideInfo(attributes.stride().x(), attributes.stride().y(),
+                                                        attributes.pad().left, attributes.pad().right,
+                                                        attributes.pad().top, attributes.pad().bottom,
+                                                        attributes.dimension_rounding_type());
+    const ConvolutionInfo conv_info{ pad_stride_info, attributes.depth_multiplier(), ActivationLayerInfo(), attributes.dilation() };
+    const TensorShape     output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*src, *wei, conv_info);
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(), output_shape);
+
+    // Check strides and dilation
+    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_stride_info.stride().first < 1);
+    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_stride_info.stride().second < 1);
+    ARM_COMPUTE_RETURN_ERROR_ON((conv_info.dilation.x() < 1) || (conv_info.dilation.y() < 1));
+    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_stride_info.stride().first > 1 && settings.m0() != 1);
+    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.dilation.x() > 1 && settings.m0() != 1);
+
+    if(conv_info.depth_multiplier > 1 && settings.n0() > 1)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON((conv_info.depth_multiplier % settings.n0()) != 0);
+    }
+
+    // Check export weights to cl image
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((settings.export_weights_to_cl_image() == true) && (export_to_cl_image(wei) == false), "Weights cannot be exported to cl_image!");
+    ARM_COMPUTE_RETURN_ERROR_ON((settings.export_weights_to_cl_image() == true) && ((settings.n0() % 4) != 0));
+
+    ARM_COMPUTE_RETURN_ERROR_ON(wei->dimension(channel_idx) != (src->dimension(channel_idx) * conv_info.depth_multiplier));
+
+    // bia shape is correct
+    if(bia != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(bia->dimension(0) != output_shape[channel_idx],
+                                        "Biases size and number of dst feature maps should match");
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(bia->num_dimensions() > 1, "Biases should be one dimensional");
+    }
+
+    // 2. Check support level
+    // Data type
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32);
+    // Data layout
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(src, DataLayout::NHWC);
+    // Texture in the input tensor
+    ARM_COMPUTE_RETURN_ERROR_ON((settings.export_input_to_cl_image() == true));
+
+    return Status{};
+}
+
+ClComponentDepthwiseConv2d::ClComponentDepthwiseConv2d(
+    ComponentId                      id,
+    const Properties                &properties,
+    const ArgumentPack<ITensorInfo> &tensors,
+    const Attributes                &attributes,
+    const Settings                  &settings)
+    : IGpuKernelComponent{ id, properties, tensors },
+      _component_writer{ std::make_unique<ClTemplateDepthwiseConv2d>(id, tensors, attributes, settings) }
+{
+}
+ClComponentDepthwiseConv2d::~ClComponentDepthwiseConv2d()
+{
+}
+const IGpuTemplateComponentWriter *ClComponentDepthwiseConv2d::template_writer() const
+{
+    return _component_writer.get();
+}
+} // namespace dynamic_fusion
+} // namespace experimental
+} // namespace arm_compute
diff --git a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDepthwiseConv2d.h b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDepthwiseConv2d.h
new file mode 100644
index 0000000000..0e2b5f14cb
--- /dev/null
+++ b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDepthwiseConv2d.h
@@ -0,0 +1,171 @@
+/*
+ * Copyright (c) 2022 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef SRC_DYNAMIC_FUSION_SKETCH_GPU_COMPONENTS_CL_CLCOMPONENTDEPTHWISECONV2D
+#define SRC_DYNAMIC_FUSION_SKETCH_GPU_COMPONENTS_CL_CLCOMPONENTDEPTHWISECONV2D
+
+#include "arm_compute/core/Error.h"
+#include "src/dynamic_fusion/sketch/gpu/components/IGpuKernelComponent.h"
+#include <memory>
+
+namespace arm_compute
+{
+/** Forward declaration */
+class ITensorInfo;
+namespace experimental
+{
+namespace dynamic_fusion
+{
+/** Forward declaration */
+template <typename T>
+class ArgumentPack;
+class DepthwiseConv2dAttributes;
+
+/** Component specific settings
+ */
+class ClComponentDepthwiseConv2dSettings
+{
+public:
+    /** Set export_input_to_cl_image flag */
+    ClComponentDepthwiseConv2dSettings &export_input_to_cl_image(bool cl_image);
+    /** Get export_input_to_cl_image flag */
+    bool export_input_to_cl_image() const;
+
+    /** Set export_weights_to_cl_image flag */
+    ClComponentDepthwiseConv2dSettings &export_weights_to_cl_image(bool cl_image);
+    /** Get export_weights_to_cl_image flag */
+    bool export_weights_to_cl_image() const;
+
+    /** Set fast_relaxed_math flag */
+    ClComponentDepthwiseConv2dSettings &fast_relaxed_math(bool fast_relaxed_math);
+    /** Get fast_relaxed_math flag */
+    bool fast_relaxed_math() const;
+
+    /** Set is_fma_available flag */
+    ClComponentDepthwiseConv2dSettings &is_fma_available(bool is_fma_available);
+    /** Get is_fma_available flag */
+    bool is_fma_available() const;
+
+    /** Set N0: number of columns processed by each thread */
+    ClComponentDepthwiseConv2dSettings &n0(unsigned int n0);
+    /** Get N0: number of columns processed by each thread */
+    unsigned int n0() const;
+
+    /** Set M0: number of rows processed by each thread */
+    ClComponentDepthwiseConv2dSettings &m0(unsigned int m0);
+    /** Set M0: number of rows processed by each thread */
+    unsigned int m0() const;
+
+private:
+    bool         _export_input_to_cl_image{ false };   /**< Export input to cl_image */
+    bool         _export_weights_to_cl_image{ false }; /**< Export the weights to cl_image */
+    bool         _fast_relaxed_math{ true };           /**< Enable/disable -cl-fast-relaxed-math flag */
+    bool         _is_fma_available{ false };           /**< Is fma instruction available */
+    unsigned int _n0{ 0 };                             /**< Number of columns processed by each thread */
+    unsigned int _m0{ 0 };                             /**< Number of rows processed by each thread */
+};
+
+/** Forward declaration */
+class ClTemplateDepthwiseConv2d;
+
+class ClComponentDepthwiseConv2d final : public IGpuKernelComponent
+{
+public:
+    /** Attributes are a set of backend-agnostic parameters that define what a component does */
+    using Attributes = DepthwiseConv2dAttributes;
+    /** Settings are a set of backend-specific parameters that influence the implementation of a component */
+    using Settings = ClComponentDepthwiseConv2dSettings;
+
+public:
+    /** Validate the component
+     *
+     * @param[in]     properties Component properties @ref Properties
+     * @param[in,out] tensors    Tensor arguments to the component
+     * @param[in]     attributes Component attributes @ref Attributes
+     * @param[in]     settings   Component settings @ref Settings
+     *
+     * @return Status       Validation results
+     *
+     * Tensor argument names:
+     * - ACL_SRC_0: Input
+     * - ACL_SRC_1: Weight
+     * - ACL_SRC_2: Bias (Optional)
+     * - ACL_DST_0: Output
+     *
+     * Tensor argument constness:
+     * - ACL_SRC_0: Const
+     * - ACL_SRC_1: Const
+     * - ACL_SRC_2: Const
+     * - ACL_DST_0: Const
+     *
+     * Valid data layouts:
+     * - NHWC
+     *
+     * Valid data type configurations:
+     * |ACL_SRC_0      |ACL_SRC_1      |ACL_SRC_2      |ACL_DST_0      |
+     * |:--------------|:--------------|:--------------|:--------------|
+     * |F16            |F16            |F16            |F16            |
+     * |F32            |F32            |F32            |F32            |
+     */
+    static Status validate(
+        const Properties                &properties,
+        const ArgumentPack<ITensorInfo> &tensors,
+        const Attributes                &attributes,
+        const Settings                  &settings);
+
+    /** Constructor
+     *
+     * Similar to @ref ClComponentDepthwiseConv2d::validate()
+     */
+    ClComponentDepthwiseConv2d(
+        ComponentId                      id,
+        const Properties                &properties,
+        const ArgumentPack<ITensorInfo> &tensors,
+        const Attributes                &attributes,
+        const Settings                  &settings);
+
+    /** Destructor */
+    ~ClComponentDepthwiseConv2d() override;
+    /** Prevent instances of this class from being copy constructed */
+    ClComponentDepthwiseConv2d(const ClComponentDepthwiseConv2d &component) = delete;
+    /** Prevent instances of this class from being copied */
+    ClComponentDepthwiseConv2d &operator=(const ClComponentDepthwiseConv2d &component) = delete;
+    /** Allow instances of this class to be move constructed */
+    ClComponentDepthwiseConv2d(ClComponentDepthwiseConv2d &&component) = default;
+    /** Allow instances of this class to be moved */
+    ClComponentDepthwiseConv2d &operator=(ClComponentDepthwiseConv2d &&component) = default;
+    /** Get template writer for the component */
+    const IGpuTemplateComponentWriter *template_writer() const override;
+    /** Get component type */
+    GpuComponentType type() const override
+    {
+        return GpuComponentType::Complex;
+    }
+
+private:
+    std::unique_ptr<ClTemplateDepthwiseConv2d> _component_writer;
+};
+} // namespace dynamic_fusion
+} // namespace experimental
+} // namespace arm_compute
+#endif /* SRC_DYNAMIC_FUSION_SKETCH_GPU_COMPONENTS_CL_CLCOMPONENTDEPTHWISECONV2D */
diff --git a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp
index e94cfd1581..dc05825500 100644
--- a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp
+++ b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp
@@ -35,26 +35,24 @@ namespace experimental
 {
 namespace dynamic_fusion
 {
-using Settings = ClComponentDirectConv2dSettings;
-
-Settings &Settings::export_to_cl_image(bool cl_image)
+ClComponentDirectConv2dSettings &ClComponentDirectConv2dSettings::export_to_cl_image(bool cl_image)
 {
     _export_to_cl_image = cl_image;
     return *this;
 }
 
-bool Settings::export_to_cl_image() const
+bool ClComponentDirectConv2dSettings::export_to_cl_image() const
 {
     return _export_to_cl_image;
 }
 
-Settings &Settings::fast_relaxed_math(bool fast_relaxed_math)
+ClComponentDirectConv2dSettings &ClComponentDirectConv2dSettings::fast_relaxed_math(bool fast_relaxed_math)
 {
     _fast_relaxed_math = fast_relaxed_math;
     return *this;
 }
 
-bool Settings::fast_relaxed_math() const
+bool ClComponentDirectConv2dSettings::fast_relaxed_math() const
 {
     return _fast_relaxed_math;
 }