diff options
author | Viet-Hoa Do <viet-hoa.do@arm.com> | 2022-12-16 14:45:57 +0000 |
---|---|---|
committer | Viet-Hoa Do <viet-hoa.do@arm.com> | 2022-12-30 13:59:23 +0000 |
commit | 3558c5840e7c973e2b1a86ae3a9335b44cad59d4 (patch) | |
tree | b5f14b344ff8bc03e5143a54a5f3480263db543e /src/dynamic_fusion/sketch/gpu/template_writer/GpuKernelVariableTable.cpp | |
parent | 9d3bd41030366326e9c8afe5db3a5812a76b135b (diff) | |
download | ComputeLibrary-3558c5840e7c973e2b1a86ae3a9335b44cad59d4.tar.gz |
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>
Diffstat (limited to 'src/dynamic_fusion/sketch/gpu/template_writer/GpuKernelVariableTable.cpp')
-rw-r--r-- | src/dynamic_fusion/sketch/gpu/template_writer/GpuKernelVariableTable.cpp | 47 |
1 files changed, 26 insertions, 21 deletions
diff --git a/src/dynamic_fusion/sketch/gpu/template_writer/GpuKernelVariableTable.cpp b/src/dynamic_fusion/sketch/gpu/template_writer/GpuKernelVariableTable.cpp index 13c0b141a5..2eafe62bfa 100644 --- a/src/dynamic_fusion/sketch/gpu/template_writer/GpuKernelVariableTable.cpp +++ b/src/dynamic_fusion/sketch/gpu/template_writer/GpuKernelVariableTable.cpp @@ -24,6 +24,7 @@ #include "GpuKernelVariableTable.h" #include "arm_compute/core/CL/CLHelpers.h" #include "arm_compute/core/ITensorInfo.h" +#include "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.h" namespace arm_compute { @@ -31,44 +32,48 @@ namespace experimental { namespace dynamic_fusion { -void GpuKernelVariableTable::declare_variable(const ITensorInfo *tensor, GpuKernelArgumentInfo argument_info, bool is_interm, const std::string &alias) +void GpuKernelVariableTable::declare_variable(const GpuKernelComponentGroup &comp_group, const ITensorInfo *tensor, GpuKernelArgumentInfo argument_info, const std::string &alias) { ARM_COMPUTE_ERROR_ON_MSG(!tensor->has_valid_id(), "Tensor info with valid id expected"); + // Do not re-declare if the variable associated with the tensor has already been declared - if(get_variable(tensor).has_valid_id()) + auto it = _vars.find(tensor->id()); + + if(it != _vars.end()) { - ARM_COMPUTE_ERROR_ON(!(get_variable(tensor).kernel_argument_info == argument_info)); + ARM_COMPUTE_ERROR_ON(!(it->second.kernel_argument_info == argument_info)); return; } - // Declare variable associated with the tensor - std::stringstream ss; - ss << alias << "_t" << tensor->id(); - const auto uniq_name = ss.str(); - TensorVariable var{ tensor->id(), uniq_name, argument_info }; - if(is_interm) + const auto target = comp_group.get_tile_for_tensor(tensor); + + if(target != tensor) { - _interm_var = var; - _interm_tensors.insert(tensor->id()); + // If the tensor uses a shared tile, don't declare another variable. + it = _vars.find(target->id()); + + ARM_COMPUTE_ERROR_ON_MSG( + it == _vars.end(), + "The variable used for this tensor must have been declared."); + + _vars[tensor->id()] = it->second; } else { + // Declare variable associated with the tensor + std::stringstream ss; + ss << alias << "_t" << tensor->id(); + const auto uniq_name = ss.str(); + TensorVariable var{ tensor->id(), uniq_name, argument_info }; + _vars.emplace(tensor->id(), var); } } GpuKernelVariableTable::TensorVariable GpuKernelVariableTable::get_variable(const ITensorInfo *tensor) const { - const TensorVariable empty_var{}; - if(_vars.find(tensor->id()) != _vars.end()) - { - return _vars.at(tensor->id()); - } - if(_interm_tensors.find(tensor->id()) != _interm_tensors.end()) - { - return _interm_var; - } - return empty_var; + const auto var = _vars.at(tensor->id()); + return var; } GpuKernelVariableTable::VariableList GpuKernelVariableTable::get_variable_list(const std::vector<const ITensorInfo *> &tensors) const |