COMPMID-941 Add NEON broadcast multiply support

Change-Id: I1f808c25750461bec9a28b2f6615fbd0f624117a
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/122262
Reviewed-by: Michele DiGiorgio <michele.digiorgio@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

5 files changed, 121 insertions, 54 deletions

diff --git a/arm_compute/core/NEON/kernels/NEPixelWiseMultiplicationKernel.h b/arm_compute/core/NEON/kernels/NEPixelWiseMultiplicationKernel.h
index 2a82d6ed00..8c245569a5 100644
--- a/arm_compute/core/NEON/kernels/NEPixelWiseMultiplicationKernel.h
+++ b/arm_compute/core/NEON/kernels/NEPixelWiseMultiplicationKernel.h
@@ -86,6 +86,7 @@ public:
 
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
+    BorderSize border_size() const override;
 
 private:
     /** Common signature for all the specialised multiplication functions with integer scaling factor
diff --git a/arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h b/arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h
index 7d22500c5e..7d5e397e80 100644
--- a/arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h
+++ b/arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,14 +37,17 @@ class NEPixelWiseMultiplication : public INESimpleFunction
 public:
     /** Initialise the kernel's inputs, output and convertion policy.
      *
-     * @param[in]  input1          First tensor input. Data types supported: U8/QS8/S16/F32.
-     * @param[in]  input2          Second tensor input. Data types supported: U8/QS8/S16/F32.
-     * @param[out] output          Output tensor. Data types supported: U8/QS8/S16/F32.
-     * @param[in]  scale           Scale to apply after multiplication. Must be positive.
-     * @param[in]  overflow_policy Overflow policy.
-     * @param[in]  rounding_policy Rounding policy.
+     * @param[in, out] input1          An input tensor. Data types supported: U8/QS8/S16/F32.
+     *                                 The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
+     * @param[in, out] input2          An input tensor. Data types supported: same as @p input1.
+     *                                 The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
+     * @param[out]     output          Output tensor. Data types supported: U8/QS8/S16/F32.
+     * @param[in]      scale           Scale to apply after multiplication.
+     *                                 Scale must be positive and its value must be either 1/255 or 1/2^n where n is between 0 and 15. For QS8 and QS16 scale must be 1.
+     * @param[in]      overflow_policy Overflow policy.
+     * @param[in]      rounding_policy Rounding policy.
      */
-    void configure(const ITensor *input1, const ITensor *input2, ITensor *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy);
+    void configure(ITensor *input1, ITensor *input2, ITensor *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy);
     /** Static function to check if given info will lead to a valid configuration of @ref NEPixelWiseMultiplication
      *
      * @param[in] input1          First tensor info input. Data types supported: U8/QS8/S16/F32.
diff --git a/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp b/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp
index c271032e54..193ca3799c 100644
--- a/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp
+++ b/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -54,18 +54,23 @@ const float       scale255_constant      = 1.f / 255.f;
 const float32x4_t scale255_constant_f32q = vdupq_n_f32(scale255_constant);
 const float32x4_t positive_round_f32q    = vdupq_n_f32(0.5f);
 
+constexpr unsigned int num_elems_processed_per_iteration = 16;
+
 inline Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy)
 {
     ARM_COMPUTE_UNUSED(overflow_policy);
     ARM_COMPUTE_UNUSED(rounding_policy);
 
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, input2, output);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::U8 && (input1->data_type() != DataType::U8 || input2->data_type() != DataType::U8),
                                     "Output can only be U8 if both inputs are U8");
 
+    const TensorShape &out_shape = TensorShape::broadcast_shape(input1->tensor_shape(), input2->tensor_shape());
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output->tensor_shape(), 0), "Wrong shape for output");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible");
+
     if(is_data_type_fixed_point(input1->data_type()) || is_data_type_fixed_point(input2->data_type()) || is_data_type_fixed_point(output->data_type()))
     {
         // Check that all data types are the same and all fixed-point positions are the same
@@ -96,19 +101,44 @@ inline Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *i
 
 inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output)
 {
-    constexpr unsigned int num_elems_processed_per_iteration = 16;
+    const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(*input1, *input2);
+    const ValidRegion &valid_region = broadcast_pair.second;
+
+    // Auto initialize output if not initialized
+    {
+        set_shape_if_empty(*output, input1->tensor_shape());
+
+        if(input1->data_type() == DataType::S16 || input2->data_type() == DataType::S16)
+        {
+            set_format_if_unknown(*output, Format::S16);
+        }
+        else if(input1->data_type() == DataType::F32 || input2->data_type() == DataType::F32)
+        {
+            set_format_if_unknown(*output, Format::F32);
+        }
+        else if(input1->data_type() == DataType::F16 || input2->data_type() == DataType::F16)
+        {
+            set_format_if_unknown(*output, Format::F16);
+        }
+        else if(input1->data_type() == DataType::QS8 && input2->data_type() == DataType::QS8)
+        {
+            set_data_type_if_unknown(*output, DataType::QS8);
+            set_fixed_point_position_if_zero(*output, input1->fixed_point_position());
+        }
+    }
 
     // Configure kernel window
-    Window                 win = calculate_max_window(*input1, Steps(num_elems_processed_per_iteration));
-    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
+    Window win        = calculate_max_window(valid_region, Steps(num_elems_processed_per_iteration));
+    Window win_input1 = win.broadcast_if_dimension_le_one(*input1);
+    Window win_input2 = win.broadcast_if_dimension_le_one(*input2);
 
-    bool window_changed = update_window_and_padding(win,
-                                                    AccessWindowHorizontal(input1, 0, num_elems_processed_per_iteration),
-                                                    AccessWindowHorizontal(input2, 0, num_elems_processed_per_iteration),
-                                                    output_access);
+    AccessWindowHorizontal input1_access(input1, 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal input2_access(input2, 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
 
-    ValidRegion valid_region = intersect_valid_regions(input1->valid_region(),
-                                                       input2->valid_region());
+    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);
 
@@ -508,31 +538,12 @@ void NEPixelWiseMultiplicationKernel::configure(const ITensor *input1, const ITe
     ARM_COMPUTE_UNUSED(rounding_policy);
     ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
 
-    // Auto initialize output if not initialized
-    {
-        set_shape_if_empty(*output->info(), input1->info()->tensor_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::F32 || input2->info()->data_type() == DataType::F32)
-        {
-            set_format_if_unknown(*output->info(), Format::F32);
-        }
-        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::QS8 && input2->info()->data_type() == DataType::QS8)
-        {
-            set_data_type_if_unknown(*output->info(), DataType::QS8);
-            set_fixed_point_position_if_zero(*output->info(), input1->info()->fixed_point_position());
-        }
-    }
-
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1->info(), input2->info(), output->info(), scale, overflow_policy, rounding_policy));
 
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input1->info(), input2->info(), output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+
     _input1         = input1;
     _input2         = input2;
     _output         = output;
@@ -656,15 +667,13 @@ void NEPixelWiseMultiplicationKernel::configure(const ITensor *input1, const ITe
         ARM_COMPUTE_ERROR("You called with the wrong img formats");
     }
 
-    // Configure kernel window
-    auto win_config = validate_and_configure_window(input1->info(), input2->info(), output->info());
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
     INEKernel::configure(win_config.second);
 }
 
 Status NEPixelWiseMultiplicationKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float scale, ConvertPolicy overflow_policy,
                                                  RoundingPolicy rounding_policy)
 {
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input1, input2, output, scale, overflow_policy, rounding_policy));
     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input1->clone().get(), input2->clone().get(), output->clone().get()).first);
 
@@ -677,34 +686,71 @@ void NEPixelWiseMultiplicationKernel::run(const Window &window, const ThreadInfo
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
 
-    Iterator input1(_input1, window);
-    Iterator input2(_input2, window);
-    Iterator output(_output, window);
+    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(INEKernel::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);
+
+    Iterator input1(_input1, slice_input1);
+    Iterator input2(_input2, slice_input2);
+    Iterator output(_output, slice);
 
     if(_func_int != nullptr)
     {
-        execute_window_loop(window, [&](const Coordinates & id)
+        execute_window_loop(collapsed, [&](const Coordinates & id)
         {
             (*_func_int)(input1.ptr(), input2.ptr(), output.ptr(), _scale_exponent);
+            collapsed.slide_window_slice_3D(slice_input1);
+            collapsed.slide_window_slice_3D(slice_input2);
         },
         input1, input2, output);
     }
     else if(_func_q_int != nullptr)
     {
         int fixed_point_position = _input1->info()->fixed_point_position();
-        execute_window_loop(window, [&](const Coordinates & id)
+        execute_window_loop(collapsed, [&](const Coordinates & id)
         {
             (*_func_q_int)(input1.ptr(), input2.ptr(), output.ptr(), _scale_exponent, fixed_point_position);
+            collapsed.slide_window_slice_3D(slice_input1);
+            collapsed.slide_window_slice_3D(slice_input2);
         },
         input1, input2, output);
     }
     else
     {
         ARM_COMPUTE_ERROR_ON(_func_float == nullptr);
-        execute_window_loop(window, [&](const Coordinates & id)
+        execute_window_loop(collapsed, [&](const Coordinates & id)
         {
             (*_func_float)(input1.ptr(), input2.ptr(), output.ptr(), _scale);
+            collapsed.slide_window_slice_3D(slice_input1);
+            collapsed.slide_window_slice_3D(slice_input2);
         },
         input1, input2, output);
     }
 }
+
+BorderSize NEPixelWiseMultiplicationKernel::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);
+}
\ No newline at end of file
diff --git a/src/runtime/NEON/functions/NEPixelWiseMultiplication.cpp b/src/runtime/NEON/functions/NEPixelWiseMultiplication.cpp
index 5a474e4a83..cf6b984717 100644
--- a/src/runtime/NEON/functions/NEPixelWiseMultiplication.cpp
+++ b/src/runtime/NEON/functions/NEPixelWiseMultiplication.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,6 +23,7 @@
  */
 #include "arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h"
 
+#include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/NEON/kernels/NEPixelWiseMultiplicationKernel.h"
 #include "support/ToolchainSupport.h"
 
@@ -30,11 +31,21 @@
 
 using namespace arm_compute;
 
-void NEPixelWiseMultiplication::configure(const ITensor *input1, const ITensor *input2, ITensor *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy)
+void NEPixelWiseMultiplication::configure(ITensor *input1, ITensor *input2, ITensor *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy)
 {
     auto k = arm_compute::support::cpp14::make_unique<NEPixelWiseMultiplicationKernel>();
     k->configure(input1, input2, output, scale, overflow_policy, rounding_policy);
     _kernel = std::move(k);
+
+    if(output->info()->dimension(0) > 1)
+    {
+        ITensor *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);
+        }
+    }
 }
 Status NEPixelWiseMultiplication::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy)
 {
diff --git a/tests/validation/NEON/PixelWiseMultiplication.cpp b/tests/validation/NEON/PixelWiseMultiplication.cpp
index 44b4ff289c..9304c8bad0 100644
--- a/tests/validation/NEON/PixelWiseMultiplication.cpp
+++ b/tests/validation/NEON/PixelWiseMultiplication.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -117,6 +117,8 @@ template <typename T>
 using NEPixelWiseMultiplicationToQS8Fixture = PixelWiseMultiplicationValidationFixture<Tensor, Accessor, NEPixelWiseMultiplication, T, qint8_t>;
 template <typename T>
 using NEPixelWiseMultiplicationToQS16Fixture = PixelWiseMultiplicationValidationFixture<Tensor, Accessor, NEPixelWiseMultiplication, T, qint16_t>;
+template <typename T>
+using NEPixelWiseMultiplicationBroadcastFixture = PixelWiseMultiplicationBroadcastValidationFixture<Tensor, Accessor, NEPixelWiseMultiplication, T, float>;
 
 TEST_SUITE(NEON)
 TEST_SUITE(PixelWiseMultiplication)
@@ -262,6 +264,10 @@ TEST_SUITE_END() // ScaleOther
 
 TEST_SUITE_END() // F32toF32
 
+TEST_SUITE(Broadcast)
+PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, BroadcastFixture<float>, PRECOMMIT, SmallShapesBroadcast(), F32, F32, scale_255, TO_NEAREST_UP, VALIDATE(float, 1.f))
+TEST_SUITE_END() // Broadcast
+
 TEST_SUITE_END()
 TEST_SUITE_END()
 } // namespace validation