Add Dynamic Fusion Tests with BugFixes

- Allow fusing arbitrary number of existing elementwise operators
- Fix issues with 3D and 4D tensors in Elementwise Addition and Floor components
    - Collapse the 3D/4D window in the same way as that used by Conv2d,
      i.e. collapse dim 1 and dim 2 together
- Fix Floor component issues when used after other components
- Add Dynamic Fusion Tests (Floor + Div, Conv2d + Add + Div)
- Add Addition ElementWise Broadcasting Test

Resolves: [COMPMID-5356]
Change-Id: I58b93a90175bb0440d43531d18cac94b5f5c2689
Signed-off-by: Mohammed Suhail Munshi <MohammedSuhail.Munshi@arm.com>
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/433956
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7957
Reviewed-by: SiCong Li <sicong.li@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

6 files changed, 130 insertions, 31 deletions

diff --git a/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/Common.h b/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/Common.h
index 57ac70aa22..04919acb83 100644
--- a/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/Common.h
+++ b/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/Common.h
@@ -371,6 +371,7 @@ public:
     {
         return Window{};
     }
+
     /** Get the tag look-up table used to instantiate the component code.
      *
      * @param vtable
@@ -557,7 +558,7 @@ public:
 
     std::string build_code()
     {
-        ARM_COMPUTE_ERROR_ON_MSG(_graph_root < 0, "No root found in the component graph");
+        ARM_COMPUTE_ERROR_ON_MSG(_graph_root == -1, "No root found in the component graph");
 
         // These data structures will hold the data from all the components in the blueprint
         std::set<std::string>    headers_list{};
@@ -666,9 +667,10 @@ public:
         return _tile_info;
     }
 
+    // Get the global execution window, i.e. that of the root component
     Window get_execution_window() const
     {
-        ARM_COMPUTE_ERROR_ON_MSG(_graph_root < 0, "No root found in the component graph");
+        ARM_COMPUTE_ERROR_ON_MSG(_graph_root == -1, "No root found in the component graph");
         ARM_COMPUTE_ERROR_ON_MSG(_dst_id == -1, "Destination Tensor Id should be ready before calling get_execution_window()");
 
         return _components.find(_graph_root)->second->get_window();
@@ -925,4 +927,4 @@ private:
 } // namespace experimental
 } // namespace arm_compute
 #endif //ARM_COMPUTE_EXPERIMENTAL_DYNAMICFUSION_IMPL_COMMON_H
-#endif /* ENABLE_EXPERIMENTAL_DYNAMIC_FUSION */
\ No newline at end of file
+#endif /* ENABLE_EXPERIMENTAL_DYNAMIC_FUSION */
diff --git a/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClElementwiseKernelComponent.cpp b/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClElementwiseKernelComponent.cpp
index 24a9eee9a3..7515aec27a 100644
--- a/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClElementwiseKernelComponent.cpp
+++ b/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClElementwiseKernelComponent.cpp
@@ -24,6 +24,7 @@
 #ifdef ENABLE_EXPERIMENTAL_DYNAMIC_FUSION
 
 #include "src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClElementwiseKernelComponent.h"
+#include "arm_compute/core/Error.h"
 #include "arm_compute/core/Validate.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/WindowHelpers.h"
@@ -57,9 +58,13 @@ Window ClElementwiseKernelComponent::get_window() const
 
     auto_init_if_empty(*dst_info, out_shape, 1, lhs_info->data_type());
 
+    TensorShape output_shape = dst_info->tensor_shape();
+    // Collapse Dim 1 (W) and Dim 2 (H) together, leave Dim 0 (C) and upper dimensions unchanged
+    // This is in line with the collapsing convention used by Conv2d
+    output_shape.collapse(2U, 1U);
     const unsigned int vector_size_byte_opencl           = 16;
     const unsigned int num_elems_processed_per_iteration = adjust_vec_size(vector_size_byte_opencl / dst_info->element_size(), dst_info->dimension(0));
-    Window             win                               = calculate_max_window(*dst_info, Steps(num_elems_processed_per_iteration));
+    Window             win                               = calculate_max_window(output_shape, Steps(num_elems_processed_per_iteration));
 
     return win;
 }
@@ -83,8 +88,12 @@ std::string ClElementwiseKernelComponent::get_component_code() const
         TILE({{DATA_TYPE}}, M0, N0, lhs_tile);
         TILE({{DATA_TYPE}}, M0, N0, rhs_tile);
 
+        // Since mout maps to dimensions 1 (y) and dimension 2 (z) of the input tensor because of the collapsed window, bout maps to dimension 3 (w)
+        {{lhs}}_offset_first_element_in_bytes += bout * {{lhs}}_stride_w;
+        {{rhs}}_offset_first_element_in_bytes += bout * {{rhs}}_stride_w;
+
         T_LOAD({{DATA_TYPE}}, M0, N0, BUFFER, {{lhs}}, cout, mout, 1, {{lhs}}_stride_y, lhs_tile);
-        T_LOAD({{DATA_TYPE}}, M0, N0, BUFFER, {{rhs}}, cout, mout, 1, {{rhs}}_stride_y, rhs_tile);
+        T_LOAD({{DATA_TYPE}}, {{rhs_m0}}, {{rhs_n0}}, BUFFER, {{rhs}}, {{rhs_start_x}}, {{rhs_start_y}}, 1, {{rhs}}_stride_y, rhs_tile);
 
 #if defined(IS_BROADCAST)
         T_ELTWISE_BROADCAST_{{ELTWISE_OP}}_X({{DATA_TYPE}}, M0, N0, lhs_tile, rhs_tile, {{dst}});
@@ -107,7 +116,7 @@ std::string ClElementwiseKernelComponent::get_component_code() const
     {
         TILE({{DATA_TYPE}}, M0, N0, addend_tile);
 
-        T_LOAD({{DATA_TYPE}}, M0, N0, BUFFER, {{addend}}, cout, mout, 1, {{addend}}_stride_y, addend_tile);
+        T_LOAD({{DATA_TYPE}}, {{rhs_m0}}, {{rhs_n0}}, BUFFER, {{addend}}, {{rhs_start_x}}, {{rhs_start_y}}, 1, {{addend}}_stride_y, addend_tile);
 
 #if defined(IS_BROADCAST)
         T_ELTWISE_BROADCAST_{{ELTWISE_OP}}_X({{DATA_TYPE}}, M0, N0, {{acc}}, addend_tile, {{acc}});
@@ -122,16 +131,18 @@ std::string ClElementwiseKernelComponent::get_component_code() const
 
 CLBuildOptions ClElementwiseKernelComponent::generate_build_options() const
 {
-    const auto t_src_info = _blueprint->impl().get_kernel_argument_info(_rhs.arg_id);
+    const auto t_rhs_info = _blueprint->impl().get_kernel_argument_info(_rhs.arg_id);
     const auto t_dst_info = _blueprint->impl().get_kernel_argument_info(_blueprint->impl().get_dst_id());
 
-    CLBuildOptions build_opts{};
-    const auto     n0           = _blueprint->impl().get_execution_window().x().step();
-    const auto     m0           = _blueprint->impl().get_execution_window().y().step();
-    const bool     is_broadcast = t_src_info->tensor_shape() != t_dst_info->tensor_shape();
+    CLBuildOptions     build_opts{};
+    const auto         n0               = _blueprint->impl().get_execution_window().x().step();
+    const auto         m0               = _blueprint->impl().get_execution_window().y().step();
+    const unsigned int partial_store_n0 = t_dst_info->dimension(0) % n0;
+    const bool         is_broadcast     = t_rhs_info->tensor_shape() != t_dst_info->tensor_shape();
 
     build_opts.add_option("-DM0=" + support::cpp11::to_string(m0));
     build_opts.add_option("-DN0=" + support::cpp11::to_string(n0));
+    build_opts.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_store_n0));
     build_opts.add_option_if(is_broadcast, "-DIS_BROADCAST");
 
     return build_opts;
@@ -166,6 +177,7 @@ ClElementwiseKernelComponent::TagLUT ClElementwiseKernelComponent::get_tag_lut(c
 {
     TagLUT     lut{};
     const auto t_dst_info = _blueprint->impl().get_kernel_argument_info(_blueprint->impl().get_dst_id());
+    const auto t_rhs_info = _blueprint->impl().get_kernel_argument_info(_rhs.arg_id);
     // Arguments and global shared variables
     const bool is_root = _blueprint->impl().group(_lhs.arg_id) == SharedVarGroup::Argument && _blueprint->impl().group(_rhs.arg_id) == SharedVarGroup::Argument;
     if(is_root)
@@ -211,6 +223,39 @@ ClElementwiseKernelComponent::TagLUT ClElementwiseKernelComponent::get_tag_lut(c
         default:
             ARM_COMPUTE_ERROR("Arithmetic Operation not supported");
     }
+
+    // Set broadcast parameters
+    // PRE: All tensors are broadcast-compatible
+    const bool is_broadcast = t_rhs_info->tensor_shape() != t_dst_info->tensor_shape();
+    if(is_broadcast)
+    {
+        // Note that n0 maps to input tensor dimension 0, m0 maps to input dimensions 1 and 2 because of our collapse strategy
+        if(t_rhs_info->dimension(0) == 1U && t_rhs_info->dimension(1) == 1U && t_rhs_info->dimension(2) == 1U) // Broadcast in X, Y, Z: collapsed rhs win [M0xN0] = [1x1]
+        {
+            lut["rhs_m0"]      = "1";
+            lut["rhs_n0"]      = "1";
+            lut["rhs_start_y"] = "0";
+            lut["rhs_start_x"] = "0";
+        }
+        else if(t_rhs_info->dimension(1) == 1U && t_rhs_info->dimension(2) == 1U) // Broadcast in Y and Z: collapsed rhs win [M0xN0] = [1xN]
+        {
+            lut["rhs_m0"]      = "1";
+            lut["rhs_n0"]      = "N0";
+            lut["rhs_start_y"] = "0";
+            lut["rhs_start_x"] = "cout";
+        }
+        else
+        {
+            ARM_COMPUTE_ERROR("Only support rhs broadcasting in all X, Y, Z dimensions, or just in Y and Z dimensions");
+        }
+    }
+    else
+    {
+        lut["rhs_m0"]      = "M0";
+        lut["rhs_n0"]      = "N0";
+        lut["rhs_start_y"] = "mout";
+        lut["rhs_start_x"] = "cout";
+    }
     return lut;
 }
 } // namespace dynamic_fusion
diff --git a/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClElementwiseKernelComponent.h b/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClElementwiseKernelComponent.h
index 91b14ffafa..f8377457d3 100644
--- a/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClElementwiseKernelComponent.h
+++ b/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClElementwiseKernelComponent.h
@@ -37,6 +37,20 @@ namespace dynamic_fusion
 class ClElementwiseKernelComponent : public IClKernelComponent
 {
 public:
+    /** Construct a new Cl Elementwise Kernel Component object
+     *
+     * @param[in]  blueprint Blueprint to which this component is added
+     * @param[in]  desc      Component descriptor
+     * @param[in]  lhs       Link to LHS tensor
+     * @param[in]  rhs       Link to RHS tensor
+     * @param[out] dst       Link to DST tensor
+     *
+     * Support Level
+     * Data Type:       F16, F32
+     * Tensor Shape:    Any shape of arbitrary dimension >= 1 and <= 4
+     * Value Range:     All
+     * Broadcasting:    Only RHS tensor can be broadcasted into LHS. Only support broadcasting in dimension 1 and dimension 2 or all dimension 0, 1 and 2
+     */
     ClElementwiseKernelComponent(ClKernelBlueprint *blueprint, const ClElementwiseKernelDescriptor &desc, const Link &lhs, const Link &rhs, const Link &dst)
         : IClKernelComponent(blueprint), _desc{ desc }, _lhs{ lhs }, _rhs{ rhs }, _dst{ dst }
     {
diff --git a/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClFloorKernelComponent.cpp b/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClFloorKernelComponent.cpp
index 87cc110561..0a20a8f600 100644
--- a/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClFloorKernelComponent.cpp
+++ b/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClFloorKernelComponent.cpp
@@ -21,9 +21,10 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifdef ENABLE_EXPERIMENTAL_DYNAMIC_FUSION
 
+#ifdef ENABLE_EXPERIMENTAL_DYNAMIC_FUSION
 #include "src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClFloorKernelComponent.h"
+#include "arm_compute/core/Error.h"
 #include "arm_compute/core/Validate.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/WindowHelpers.h"
@@ -38,12 +39,10 @@ ComponentType ClFloorKernelComponent::get_component_type() const
 {
     return ComponentType::Simple;
 }
-
 std::set<std::string> ClFloorKernelComponent::get_headers_list() const
 {
     return std::set<std::string> { "common/experimental/gemm_fused_post_ops/fp_mixed_precision_helpers.h", "tile_helpers.h" };
 }
-
 Window ClFloorKernelComponent::get_window() const
 {
     const ITensorInfo *src_info = _blueprint->impl().get_kernel_argument_info(_src.arg_id);
@@ -52,16 +51,22 @@ Window ClFloorKernelComponent::get_window() const
     ARM_COMPUTE_ERROR_ON_NULLPTR(src_info, dst_info);
     auto_init_if_empty(*dst_info, src_info->tensor_shape(), 1, src_info->data_type());
 
+    TensorShape output_shape = dst_info->tensor_shape();
+    // Collapse Dim 1 (W) and Dim 2 (H) together, leave Dim 0 (C) and upper dimensions unchanged
+    // This is in line with the collapsing convention used by Conv2d
+    output_shape.collapse(2U, 1U);
     const unsigned int vector_size_byte_opencl           = 16;
     const unsigned int num_elems_processed_per_iteration = adjust_vec_size(vector_size_byte_opencl / dst_info->element_size(), dst_info->dimension(0));
-    Window             win                               = calculate_max_window(*dst_info, Steps(num_elems_processed_per_iteration));
+    Window             win                               = calculate_max_window(output_shape, Steps(num_elems_processed_per_iteration));
 
     return win;
 }
-
 std::string ClFloorKernelComponent::get_component_code() const
 {
-    return R"_(
+    bool is_root = _blueprint->impl().group(_src.arg_id) == SharedVarGroup::Argument;
+    if(is_root)
+    {
+        return R"_(
     //------------------ START KERNEL {{meta_kernel_id}} FLOOR ---------------------
     // IN_0(src)            {{src}}
     // OUT(dst, accum)      {{dst}}
@@ -69,30 +74,40 @@ std::string ClFloorKernelComponent::get_component_code() const
     {
         TILE({{DATA_TYPE}}, M0, N0, src_tile);
 
+        // Since mout maps to dimensions 1 (y) and dimension 2 (z) of the input tensor because of the collapsed window, bout maps to dimension 3 (w)
+        {{src}}_offset_first_element_in_bytes += bout * {{src}}_stride_w;
         T_LOAD({{DATA_TYPE}}, M0, N0, BUFFER, {{src}}, cout, mout, 1, {{src}}_stride_y, src_tile);
+
         T_FLOOR({{DATA_TYPE}}, M0, N0, src_tile, {{dst}});
     }
-
     //------------------ END KERNEL {{meta_kernel_id}} FLOOR ---------------------
 )_";
+    }
+    else
+    {
+        return R"_(
+    //------------------ START KERNEL {{meta_kernel_id}} FLOOR ---------------------
+    // IN_0/Out(Accumulator)        {{acc}}
+    // output = floor(input)
+    {
+        T_FLOOR({{DATA_TYPE}}, M0, N0, {{acc}}, {{acc}});
+    }
+    //------------------ END KERNEL {{meta_kernel_id}} FLOOR ---------------------
+)_";
+    }
 }
-
 CLBuildOptions ClFloorKernelComponent::generate_build_options() const
 {
-    CLBuildOptions build_opts{};
-
-    const auto n0 = _blueprint->impl().get_execution_window().x().step();
-    const auto m0 = _blueprint->impl().get_execution_window().y().step();
-
+    CLBuildOptions     build_opts{};
+    const auto         n0               = _blueprint->impl().get_execution_window().x().step();
+    const auto         m0               = _blueprint->impl().get_execution_window().y().step();
     const auto         dst_info         = _blueprint->impl().get_kernel_argument_info(_blueprint->impl().get_dst_id());
     const unsigned int partial_store_n0 = dst_info->dimension(0) % n0;
     build_opts.add_option("-DM0=" + support::cpp11::to_string(m0));
     build_opts.add_option("-DN0=" + support::cpp11::to_string(n0));
     build_opts.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_store_n0));
-
     return build_opts;
 }
-
 std::string ClFloorKernelComponent::generate_config_id() const
 {
     auto        t_dst_info = _blueprint->impl().get_kernel_argument_info(_blueprint->impl().get_dst_id());
@@ -106,20 +121,28 @@ std::string ClFloorKernelComponent::generate_config_id() const
     config_id += lower_string(string_from_data_layout(t_dst_info->data_layout()));
     return config_id;
 }
-
 void ClFloorKernelComponent::allocate_shared_vars(SharedVarTable &vtable) const
 {
     vtable.add(_src, _blueprint->impl().group(_src.arg_id), ClKernelArgDescriptor(_src.arg_id, ClKernelTensorArgType::Tensor_4D_t_Buffer), "src");
     vtable.add(_dst, _blueprint->impl().group(_dst.arg_id), ClKernelArgDescriptor(_dst.arg_id, ClKernelTensorArgType::Tensor_4D_t_Buffer), "dst");
 }
-
 ClFloorKernelComponent::TagLUT ClFloorKernelComponent::get_tag_lut(const SharedVarTable &vtable) const
 {
     TagLUT     lut{};
     const auto t_dst_info = _blueprint->impl().get_kernel_argument_info(_blueprint->impl().get_dst_id());
     // Arguments and global shared variables
-    lut["src"]            = vtable.get(_src);
-    lut["dst"]            = vtable.get(_dst);
+    const bool is_root = _blueprint->impl().group(_src.arg_id) == SharedVarGroup::Argument;
+
+    if(is_root)
+    {
+        lut["src"] = vtable.get(_src);
+        lut["dst"] = vtable.get(_dst);
+    }
+    else
+    {
+        lut["acc"] = vtable.get(_src);
+    }
+
     lut["meta_kernel_id"] = id();
     lut["DATA_TYPE"]      = get_cl_type_from_data_type(t_dst_info->data_type());
     return lut;
diff --git a/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClFloorKernelComponent.h b/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClFloorKernelComponent.h
index 5463e233d4..e791b36382 100644
--- a/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClFloorKernelComponent.h
+++ b/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClFloorKernelComponent.h
@@ -37,6 +37,17 @@ namespace dynamic_fusion
 class ClFloorKernelComponent : public IClKernelComponent
 {
 public:
+    /** Construct a new Cl Floor Kernel Component object
+     *
+     * @param blueprint  Blueprint to which this component is added
+     * @param src        Link to SRC tensor
+     * @param dst        Link to DST tensor
+     *
+     * Support Level
+     * Data Type:       F16, F32
+     * Tensor Shape:    Any shape of arbitrary dimension >= 1 and <= 4
+     * Value Range:     All
+     */
     ClFloorKernelComponent(ClKernelBlueprint *blueprint, const Link &src, const Link &dst)
         : IClKernelComponent(blueprint), _src{ src }, _dst{ dst }
     {
@@ -71,4 +82,4 @@ private:
 } // namespace experimental
 } // namespace arm_compute
 #endif // ARM_COMPUTE_EXPERIMENTAL_DYNAMICFUSION_IMPL_COMPONENTS_CLFLOORKERNELCOMPONENT_H
-#endif /* ENABLE_EXPERIMENTAL_DYNAMIC_FUSION */
\ No newline at end of file
+#endif /* ENABLE_EXPERIMENTAL_DYNAMIC_FUSION */
diff --git a/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClStoreKernelComponents.cpp b/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClStoreKernelComponents.cpp
index 4ac27e007f..7c805d5368 100644
--- a/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClStoreKernelComponents.cpp
+++ b/src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClStoreKernelComponents.cpp
@@ -108,6 +108,10 @@ std::string ClStoreIndirectWidthSelectKernelComponent::get_component_code() cons
     return R"_(
     //------------------ START KERNEL {{meta_kernel_id}} STORE ---------------------
     {
+    // This also follows NHWC layout
+    // cout maps to global_id(0) maps to Channel
+    // mout maps to global_id(1) maps to Height and Weight (Collapsed Window)
+    // bout maps to global_id(3) maps to N / Batch
     #define _IDST_WIDTH {{dst}}_w
     #define _IDST_HEIGHT {{dst}}_h
         TILE(uint, M0, 1, dst_indirect_y);