Add temporary tile support for dynamic fusion

* Multiple intermediate tensors can share the same tile.
  - A simple operator can reuse the input tensor for the result
    if the input tensor has the same shape, data type and it is
    only consumed by that operator.
  - The special case is a simple operator and an output operator
    consume the same tensor. However as the output operator
    doesn't change the content of the input tensor, it doesn't
    count as "consuming" the input tensor.
* These temporary tiles are declared automatically by the template
  writer. Individual operator doesn't need to generate output tile
  declaration.
* Cast is now simple operator.

Resolves: COMPMID-5778
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I232647ac976645e2d266a62e055b9eb48c356a8e
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8877
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 114 insertions, 94 deletions

diff --git a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateElementwiseBinary.cpp b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateElementwiseBinary.cpp
index 39cec6e31c..df8deee44f 100644
--- a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateElementwiseBinary.cpp
+++ b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateElementwiseBinary.cpp
@@ -65,94 +65,94 @@ std::string ClTemplateElementwiseBinary::get_component_code(const ComponentGroup
     std::string code;
     const bool  is_broadcast = _lhs->tensor_shape() != _rhs->tensor_shape();
     const bool  is_root      = (comp_group.get_root_component()->id() == this->id());
+    const bool  is_lhs_input = comp_group.is_input_tensor(_lhs);
+    const bool  is_rhs_input = comp_group.is_input_tensor(_rhs);
 
-    if(is_root)
-    {
-        code =
+    code =
 R"_(
     //------------------ START KERNEL {{meta_kernel_id}} ELTWISE_OP ---------------------
-)_"
-    // IN_0(LHS)            {{lhs}}
-    // IN_1(RHS)            {{rhs}}
-    // OUT(dst, accum)      {{dst}}
-    // dst = lhs + rhs (mix-precision, broadcast, boundary aware)
-R"_(
-    TILE({{DATA_TYPE}}, M0, N0, {{dst}});
-    TILE(uint, M0, 1, g_dst_indirect_y);
+)_";
+
+    if(is_root)
     {
-        TILE({{DATA_TYPE}}, M0, N0, lhs_tile);
-        TILE({{DATA_TYPE}}, M0, N0, rhs_tile);
-)_"
-        // Assuming un-collapsed window
+        code +=
 R"_(
-        {{lhs}}_offset_first_element_in_bytes += g_ind_2 * {{lhs}}_stride_z;
-        {{rhs}}_offset_first_element_in_bytes += g_ind_2 * {{rhs}}_stride_z;
-
-        T_LOAD({{DATA_TYPE}}, M0, N0, BUFFER, {{lhs}}, g_ind_0, g_ind_1, 1, {{lhs}}_stride_y, lhs_tile);
-        T_LOAD({{DATA_TYPE}}, {{rhs_m0}}, {{rhs_n0}}, BUFFER, {{rhs}}, {{rhs_start_ind_0}}, {{rhs_start_ind_1}}, 1, {{rhs}}_stride_y, rhs_tile);
+    TILE(uint, M0, 1, g_dst_indirect_y);
 )_";
-        if(is_broadcast)
-        {
-            code +=
+    }
+
+    if(is_lhs_input)
+    {
+        code +=
 R"_(
-        T_ELTWISE_BROADCAST_{{ELTWISE_OP}}_X({{DATA_TYPE}}, M0, N0, lhs_tile, rhs_tile, {{dst}});
+    TILE({{DATA_TYPE}}, M0, N0, {{lhs}});
 )_";
-        }
-        else
-        {
-            code +=
+    }
+
+    if(is_rhs_input)
+    {
+        code +=
 R"_(
-        T_ELTWISE_{{ELTWISE_OP}}({{DATA_TYPE}}, M0, N0, lhs_tile, rhs_tile, {{dst}});
+    TILE({{DATA_TYPE}}, M0, N0, {{rhs}});
 )_";
-        }
+    }
+
     code +=
-    // Calculate the destination indirect Y
 R"_(
-    LOOP_UNROLLING(int, i, 0, 1, M0,
     {
-        g_dst_indirect_y[i].v = (uint)min(g_ind_1 + i, (int)({{out}}_w * {{out}}_h) - 1);
-        g_dst_indirect_y[i].v += g_ind_2 * (int)({{out}}_w * {{out}}_h);
-    })
-    }
-    //------------------ END KERNEL {{meta_kernel_id}} ELTWISE_OP ---------------------
 )_";
-    }
 
-    else // non-root
+    if(is_lhs_input)
     {
-        code =
+        code +=
 R"_(
-    //------------------ START KERNEL {{meta_kernel_id}} ELTWISE_OP ---------------------
-)_"
-    // IN_0/Out(Accumulator)   {{acc}}
-    // IN_1(Operand)        {{operand}}
-    // acc = operand + acc (mix-precision, broadcast, boundary aware)
+        {{lhs}}_offset_first_element_in_bytes += g_ind_2 * {{lhs}}_stride_z;
+        T_LOAD({{DATA_TYPE}}, {{lhs_m0}}, {{lhs_n0}}, BUFFER, {{lhs}}, {{lhs_start_ind_0}}, {{lhs_start_ind_1}}, 1, {{lhs}}_stride_y, {{lhs}});
+)_";
+    }
+
+    if(is_rhs_input)
+    {
+        code +=
 R"_(
+        {{rhs}}_offset_first_element_in_bytes += g_ind_2 * {{rhs}}_stride_z;
+        T_LOAD({{DATA_TYPE}}, {{rhs_m0}}, {{rhs_n0}}, BUFFER, {{rhs}}, {{rhs_start_ind_0}}, {{rhs_start_ind_1}}, 1, {{rhs}}_stride_y, {{rhs}});
+)_";
+    }
+
+    if(is_broadcast)
+    {
+        code +=
+            R"_(
+        T_ELTWISE_BROADCAST_{{ELTWISE_OP}}_X({{DATA_TYPE}}, M0, N0, {{lhs}}, {{rhs}}, {{dst}});
+)_";
+    }
+    else
     {
-        TILE(DATA_TYPE, M0, N0, operand_tile);
-        T_LOAD({{DATA_TYPE}}, {{rhs_m0}}, {{rhs_n0}}, BUFFER, {{operand}}, {{rhs_start_ind_0}}, {{rhs_start_ind_1}}, 1, {{operand}}_stride_y, operand_tile);
+        code +=
+            R"_(
+        T_ELTWISE_{{ELTWISE_OP}}({{DATA_TYPE}}, M0, N0, {{lhs}}, {{rhs}}, {{dst}});
 )_";
+    }
 
-        if(is_broadcast)
-        {
-            code +=
+    if(is_root)
+    {
+        // Calculate the destination indirect Y
+        code +=
 R"_(
-        T_ELTWISE_BROADCAST_{{ELTWISE_OP}}_X({{DATA_TYPE}}, M0, N0, {{acc}}, operand_tile, {{acc}});
-)_";
-        }
-        else
+        LOOP_UNROLLING(int, i, 0, 1, M0,
         {
-            code +=
-R"_(
-        T_ELTWISE_{{ELTWISE_OP}}({{DATA_TYPE}}, M0, N0, {{acc}}, operand_tile, {{acc}});
+            g_dst_indirect_y[i].v = (uint)min(g_ind_1 + i, (int)({{arg_dst}}_w * {{arg_dst}}_h) - 1);
+            g_dst_indirect_y[i].v += g_ind_2 * (int)({{arg_dst}}_w * {{arg_dst}}_h);
+        })
 )_";
-        }
+    }
+
     code +=
 R"_(
     }
     //------------------ END KERNEL {{meta_kernel_id}} ELTWISE_OP ---------------------
 )_";
-    }
 
     return code;
 }
@@ -160,86 +160,105 @@ R"_(
 void ClTemplateElementwiseBinary::declare_variables(GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const
 {
     vtable.declare_variable(
+        comp_group,
         _lhs,
         GpuKernelArgumentInfo(common_tensor_type),
-        comp_group.is_intermediate_tensor(_lhs),
         "lhs");
 
     vtable.declare_variable(
+        comp_group,
         _rhs,
         GpuKernelArgumentInfo(common_tensor_type),
-        comp_group.is_intermediate_tensor(_rhs),
         "rhs");
 
     vtable.declare_variable(
+        comp_group,
         _dst,
         GpuKernelArgumentInfo(common_tensor_type),
-        comp_group.is_intermediate_tensor(_dst),
         "dst");
 }
 
 TagLUT ClTemplateElementwiseBinary::get_tag_lut(const GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const
 {
     TagLUT             lut{};
-    const ITensorInfo *accumulator = _lhs;
-    const ITensorInfo *operand     = _rhs;
 
     // Local build options
     lut["meta_kernel_id"] = id();
     lut["DATA_TYPE"]      = get_cl_type_from_data_type(_lhs->data_type());
     // Arguments and global shared variables
-    const bool is_root = (comp_group.get_root_component()->id() == this->id());
-    if(is_root)
+
+    lut["lhs"] = vtable.get_variable(_lhs);
+    lut["rhs"] = vtable.get_variable(_rhs);
+    lut["dst"] = vtable.get_variable(_dst);
+    lut["arg_dst"] = vtable.get_variable(comp_group.get_any_dst_tensor());
+
+    switch(_attributes.operation())
     {
-        lut["lhs"] = vtable.get_variable(_lhs);
-        lut["rhs"] = vtable.get_variable(_rhs);
-        lut["dst"] = vtable.get_variable(_dst);
-        lut["out"] = vtable.get_variable(comp_group.get_any_dst_tensor());
+        case Attributes::ElementwiseOp::ADD:
+            lut["ELTWISE_OP"] = "ADD";
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Arithmetic Operation not supported");
     }
-    else
+
+    ARM_COMPUTE_ERROR_ON(
+        comp_group.is_intermediate_tensor(_lhs) &&
+        detail::have_different_dimensions(_lhs->tensor_shape(), _dst->tensor_shape(), 0));
+    ARM_COMPUTE_ERROR_ON(
+        comp_group.is_intermediate_tensor(_rhs) &&
+        detail::have_different_dimensions(_rhs->tensor_shape(), _dst->tensor_shape(), 0));
+
+    // Set broadcast parameters
+    // PRE: All tensors are broadcast-compatible
+    if(_lhs->tensor_shape() != _dst->tensor_shape())
     {
-        // Determine which tensor is the accumulator
-        if(comp_group.is_intermediate_tensor(_lhs))
+        const auto is_broadcast_x = _lhs->dimension(0) == 1U && _dst->dimension(0) != 1U;
+        const auto is_broadcast_y = _lhs->dimension(1) == 1U && _dst->dimension(1) != 1U;
+        const auto is_broadcast_z = _lhs->dimension(2) == 1U && _dst->dimension(2) != 1U;
+
+        // Note that n0 maps to input tensor dimension 0, m0 maps to input dimensions 1 and 2 because of our collapse strategy
+        if(is_broadcast_x && is_broadcast_y && is_broadcast_z) // Broadcast in X, Y, Z: collapsed lhs win [M0xN0] = [1x1]
         {
-            accumulator = _lhs;
-            operand     = _rhs;
+            lut["lhs_m0"]          = "1";
+            lut["lhs_n0"]          = "1";
+            lut["lhs_start_ind_1"] = "0";
+            lut["lhs_start_ind_0"] = "0";
         }
-        else if(comp_group.is_intermediate_tensor(_rhs))
+        else if(is_broadcast_y && is_broadcast_z) // Broadcast in Y and Z: collapsed lhs win [M0xN0] = [1xN]
         {
-            accumulator = _rhs;
-            operand     = _lhs;
+            lut["lhs_m0"]          = "1";
+            lut["lhs_n0"]          = "N0";
+            lut["lhs_start_ind_1"] = "0";
+            lut["lhs_start_ind_0"] = "g_ind_0";
         }
         else
         {
-            ARM_COMPUTE_ERROR("Invalid elementwise component linking");
+            ARM_COMPUTE_ERROR("Only support lhs broadcasting in all X, Y, Z dimensions, or just in Y and Z dimensions");
         }
-        lut["acc"]     = vtable.get_variable(accumulator);
-        lut["operand"] = vtable.get_variable(operand);
     }
-    switch(_attributes.operation())
+    else
     {
-        case Attributes::ElementwiseOp::ADD:
-            lut["ELTWISE_OP"] = "ADD";
-            break;
-        default:
-            ARM_COMPUTE_ERROR("Arithmetic Operation not supported");
+        lut["lhs_m0"]          = "M0";
+        lut["lhs_n0"]          = "N0";
+        lut["lhs_start_ind_1"] = "g_ind_1";
+        lut["lhs_start_ind_0"] = "g_ind_0";
     }
-    ARM_COMPUTE_ERROR_ON_MSG(detail::have_different_dimensions(accumulator->tensor_shape(), _dst->tensor_shape(), 0), "Only the operand can be broadcast to match the accumulator's shape");
-    const bool is_broadcast = (operand->tensor_shape() != _dst->tensor_shape());
 
-    // Set broadcast parameters
-    // PRE: All tensors are broadcast-compatible
-    if(is_broadcast)
+    if(_rhs->tensor_shape() != _dst->tensor_shape())
     {
+        const auto is_broadcast_x = _rhs->dimension(0) == 1U && _dst->dimension(0) != 1U;
+        const auto is_broadcast_y = _rhs->dimension(1) == 1U && _dst->dimension(1) != 1U;
+        const auto is_broadcast_z = _rhs->dimension(2) == 1U && _dst->dimension(2) != 1U;
+
         // Note that n0 maps to input tensor dimension 0, m0 maps to input dimensions 1 and 2 because of our collapse strategy
-        if(operand->dimension(0) == 1U && operand->dimension(1) == 1U && operand->dimension(2) == 1U) // Broadcast in X, Y, Z: collapsed rhs win [M0xN0] = [1x1]
+        if(is_broadcast_x && is_broadcast_y && is_broadcast_z) // Broadcast in X, Y, Z: collapsed rhs win [M0xN0] = [1x1]
         {
             lut["rhs_m0"]          = "1";
             lut["rhs_n0"]          = "1";
             lut["rhs_start_ind_1"] = "0";
             lut["rhs_start_ind_0"] = "0";
         }
-        else if(operand->dimension(1) == 1U && operand->dimension(2) == 1U) // Broadcast in Y and Z: collapsed rhs win [M0xN0] = [1xN]
+        else if(is_broadcast_y && is_broadcast_z) // Broadcast in Y and Z: collapsed rhs win [M0xN0] = [1xN]
         {
             lut["rhs_m0"]          = "1";
             lut["rhs_n0"]          = "N0";
@@ -258,6 +277,7 @@ TagLUT ClTemplateElementwiseBinary::get_tag_lut(const GpuKernelVariableTable &vt
         lut["rhs_start_ind_1"] = "g_ind_1";
         lut["rhs_start_ind_0"] = "g_ind_0";
     }
+
     return lut;
 }