Enable broadcast support for Equal_Ex op (#3431)

This patch enables broadcast support for Equal_Ex op.
Related issue: #3295.

Signed-off-by: prasannar <prasanna.r@samsung.com>

diff --git a/libs/ARMComputeEx/arm_compute/core/CL/kernels/CLEqualKernel.h b/libs/ARMComputeEx/arm_compute/core/CL/kernels/CLEqualKernel.h
index e20fda7..847beec 100644 (file)
--- a/libs/ARMComputeEx/arm_compute/core/CL/kernels/CLEqualKernel.h
+++ b/libs/ARMComputeEx/arm_compute/core/CL/kernels/CLEqualKernel.h
@@ -48,6 +48,8 @@ public:
   // Inherited methods overridden:
   void run(const Window &window, cl::CommandQueue &queue) override;
 
+  BorderSize border_size() const override;
+
 private:
   const ICLTensor *_input1;
   const ICLTensor *_input2;

diff --git a/libs/ARMComputeEx/src/core/CL/kernels/CLEqualKernel.cpp b/libs/ARMComputeEx/src/core/CL/kernels/CLEqualKernel.cpp
index 2348052..7777c59 100644 (file)
--- a/libs/ARMComputeEx/src/core/CL/kernels/CLEqualKernel.cpp
+++ b/libs/ARMComputeEx/src/core/CL/kernels/CLEqualKernel.cpp
@@ -22,23 +22,51 @@
 
 using namespace arm_compute;
 
+namespace
+{
+constexpr unsigned int num_elems_processed_per_iteration = 16;
+}
+
 CLEqualKernel::CLEqualKernel() : _input1(nullptr), _input2(nullptr), _output(nullptr) {}
 
+Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2,
+                          const ITensorInfo *output)
+{
+  const TensorShape &out_shape =
+      TensorShape::broadcast_shape(input1->tensor_shape(), input2->tensor_shape());
+
+  ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::QS8,
+                                                       DataType::QS16, DataType::S16, DataType::F16,
+                                                       DataType::F32, DataType::QASYMM8);
+  ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::QS8,
+                                                       DataType::QS16, DataType::S16, DataType::F16,
+                                                       DataType::F32, DataType::QASYMM8);
+
+  ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0,
+                                  "Inputs are not broadcast compatible");
+  // Validate in case of configured output
+  if (output->total_size() > 0)
+  {
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(
+        output, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16,
+        DataType::F32, DataType::QASYMM8);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(
+        detail::have_different_dimensions(out_shape, output->tensor_shape(), 0),
+        "Wrong shape for output");
+  }
+  return Status{};
+}
+
 void CLEqualKernel::configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output)
 {
-  ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(input1->info()->tensor_shape(),
-                                              input2->info()->tensor_shape());
-  ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(input1->info()->tensor_shape(),
-                                              output->info()->tensor_shape());
   ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input1, input2);
   ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input1, output);
+  ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1->info(), input2->info(), output->info()));
 
   _input1 = input1;
   _input2 = input2;
   _output = output;
 
-  constexpr unsigned int num_elems_processed_per_iteration = 16;
-
   // Create kernel
   std::string kernel_name = "equal";
   std::set<std::string> build_opts;
@@ -63,17 +91,44 @@ void CLEqualKernel::configure(const ICLTensor *input1, const ICLTensor *input2,
   _kernel =
       static_cast<cl::Kernel>(CLKernelLibraryEx::get().create_kernel(kernel_name, build_opts));
 
-  // Configure window
-  Window win = calculate_max_window(*input1->info(), Steps(num_elems_processed_per_iteration));
+  const std::pair<TensorShape, ValidRegion> broadcast_pair =
+      ITensorInfo::broadcast_shape_and_valid_region(*input1->info(), *input2->info());
+
+  const TensorShape &out_shape = broadcast_pair.first;
+  const ValidRegion &valid_region = broadcast_pair.second;
+
+  // Auto initialize output if not initialized
+  {
+    set_shape_if_empty(*output->info(), out_shape);
+
+    if (input1->info()->data_type() == DataType::S16 ||
+        input2->info()->data_type() == DataType::S16)
+    {
+      set_format_if_unknown(*output->info(), Format::S16);
+    }
+    else if (input1->info()->data_type() == DataType::F16 &&
+             input2->info()->data_type() == DataType::F16)
+    {
+      set_format_if_unknown(*output->info(), Format::F16);
+    }
+    else if (input1->info()->data_type() == DataType::F32 ||
+             input2->info()->data_type() == DataType::F32)
+    {
+      set_format_if_unknown(*output->info(), Format::F32);
+    }
+  }
+
+  Window win = calculate_max_window(valid_region, Steps(num_elems_processed_per_iteration));
+  Window win_input1 = win.broadcast_if_dimension_le_one(*input1->info());
+  Window win_input2 = win.broadcast_if_dimension_le_one(*input2->info());
 
   AccessWindowHorizontal input1_access(input1->info(), 0, num_elems_processed_per_iteration);
   AccessWindowHorizontal input2_access(input2->info(), 0, num_elems_processed_per_iteration);
   AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
 
-  ValidRegion valid_region =
-      intersect_valid_regions(input1->info()->valid_region(), input2->info()->valid_region());
-
-  update_window_and_padding(win, input1_access, input2_access, output_access);
+  bool window_changed = update_window_and_padding(win_input1, input1_access) ||
+                        update_window_and_padding(win_input2, input2_access) ||
+                        update_window_and_padding(win, output_access);
 
   output_access.set_valid_region(win, valid_region);
 
@@ -85,15 +140,55 @@ void CLEqualKernel::run(const Window &window, cl::CommandQueue &queue)
   ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
   ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
 
-  Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
+  const TensorShape &in_shape1 = _input1->info()->tensor_shape();
+  const TensorShape &in_shape2 = _input2->info()->tensor_shape();
+  const TensorShape &out_shape = _output->info()->tensor_shape();
+
+  bool can_collapse = true;
+  if (std::min(in_shape1.total_size(), in_shape2.total_size()) > 1)
+  {
+    can_collapse =
+        (std::min(in_shape1.num_dimensions(), in_shape2.num_dimensions()) > Window::DimZ);
+    for (size_t d = Window::DimZ; can_collapse && (d < out_shape.num_dimensions()); d++)
+    {
+      can_collapse = (in_shape1[d] == in_shape2[d]);
+    }
+  }
+
+  bool has_collapsed = false;
+  Window collapsed =
+      can_collapse ? window.collapse_if_possible(ICLKernel::window(), Window::DimZ, &has_collapsed)
+                   : window;
+
+  const TensorShape &in_shape1_collapsed =
+      has_collapsed ? in_shape1.collapsed_from(Window::DimZ) : in_shape1;
+  const TensorShape &in_shape2_collapsed =
+      has_collapsed ? in_shape2.collapsed_from(Window::DimZ) : in_shape2;
+
   Window slice = collapsed.first_slice_window_3D();
+  Window slice_input1 = slice.broadcast_if_dimension_le_one(in_shape1_collapsed);
+  Window slice_input2 = slice.broadcast_if_dimension_le_one(in_shape2_collapsed);
 
   do
   {
     unsigned int idx = 0;
-    add_3D_tensor_argument(idx, _input1, slice);
-    add_3D_tensor_argument(idx, _input2, slice);
+    add_3D_tensor_argument(idx, _input1, slice_input1);
+    add_3D_tensor_argument(idx, _input2, slice_input2);
     add_3D_tensor_argument(idx, _output, slice);
+
     enqueue(queue, *this, slice);
+
+    collapsed.slide_window_slice_3D(slice_input1);
+    collapsed.slide_window_slice_3D(slice_input2);
   } while (collapsed.slide_window_slice_3D(slice));
 }
+
+BorderSize CLEqualKernel::border_size() const
+{
+  const unsigned int replicateSize =
+      _output->info()->dimension(0) -
+      std::min(_input1->info()->dimension(0), _input2->info()->dimension(0));
+  const unsigned int border =
+      std::min<unsigned int>(num_elems_processed_per_iteration - 1U, replicateSize);
+  return BorderSize(0, border, 0, 0);
+}

diff --git a/libs/ARMComputeEx/src/runtime/CL/functions/CLEqual.cpp b/libs/ARMComputeEx/src/runtime/CL/functions/CLEqual.cpp
index 7881e3d..2b6a994 100644 (file)
--- a/libs/ARMComputeEx/src/runtime/CL/functions/CLEqual.cpp
+++ b/libs/ARMComputeEx/src/runtime/CL/functions/CLEqual.cpp
@@ -18,6 +18,10 @@
 
 #include "arm_compute/core/CL/kernels/CLEqualKernel.h"
 
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "support/ToolchainSupport.h"
+#include <utility>
+
 using namespace arm_compute;
 
 void CLEqual::configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output)
@@ -25,4 +29,14 @@ void CLEqual::configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output)
   auto k = arm_compute::support::cpp14::make_unique<CLEqualKernel>();
   k->configure(input1, input2, output);
   _kernel = std::move(k);
+
+  if (output->info()->dimension(0) > 1)
+  {
+    ICLTensor *broadcasted_info = (input1->info()->dimension(0) == 1) ? input1 : input2;
+
+    if (broadcasted_info->info()->dimension(0) == 1)
+    {
+      _border_handler.configure(broadcasted_info, _kernel->border_size(), BorderMode::REPLICATE);
+    }
+  }
 }

diff --git a/runtimes/pure_arm_compute/src/compilation.cc b/runtimes/pure_arm_compute/src/compilation.cc
index afbd8e7..eafd239 100644 (file)
--- a/runtimes/pure_arm_compute/src/compilation.cc
+++ b/runtimes/pure_arm_compute/src/compilation.cc
@@ -3576,6 +3576,16 @@ void Planner::visit(const ::internal::tflite::op::Equal::Node &node)
                           asTensorInfo(asTensorShape(_ctx.at(output_index).shape(), false),
                                        _ctx.at(output_index).type(), _ctx.at(output_index).scale(),
                                        _ctx.at(output_index).zeroPoint()));
+
+  if (!(_ctx.at(input1_index).shape() == _ctx.at(input2_index).shape()))
+  {
+    const auto broadcast_rank =
+        std::max(_ctx.at(input1_index).shape().rank(), _ctx.at(input2_index).shape().rank());
+    const_cast<::internal::tflite::operand::Shape &>(_ctx.at(input1_index).shape())
+        .extendRank(broadcast_rank);
+    const_cast<::internal::tflite::operand::Shape &>(_ctx.at(input2_index).shape())
+        .extendRank(broadcast_rank);
+  }
   _builder.addShapeConstr(input1_index,
                           asTensorInfo(asTensorShape(_ctx.at(input1_index).shape(), false),
                                        _ctx.at(input1_index).type(), _ctx.at(input1_index).scale(),