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>
diff --git a/src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.cpp b/src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.cpp
index 0d25749..81c3f0c 100644
--- a/src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.cpp
+++ b/src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.cpp
@@ -133,8 +133,9 @@
_finalized = true;
- std::set<const ITensorInfo *> input_tensors;
std::set<const ITensorInfo *> output_tensors;
+ std::map<const ITensorInfo *, std::vector<const ITensorInfo *>> possible_tile_map;
+ std::map<const ITensorInfo *, int32_t> tile_usages;
for(auto component : _components)
{
@@ -156,26 +157,139 @@
}
else if(_interm_tensors.find(tensor) == _interm_tensors.end())
{
- input_tensors.insert(tensor);
+ _input_tensors.insert(tensor);
+
+ tile_usages[tensor] = 0;
+ possible_tile_map.emplace(tensor, std::vector<const ITensorInfo *>());
}
}
for(auto tensor : dst_tensors)
{
- ARM_COMPUTE_ERROR_ON(input_tensors.find(tensor) != input_tensors.end());
+ ARM_COMPUTE_ERROR_ON(_input_tensors.find(tensor) != _input_tensors.end());
ARM_COMPUTE_ERROR_ON(output_tensors.find(tensor) != output_tensors.end());
ARM_COMPUTE_ERROR_ON(_interm_tensors.find(tensor) != _interm_tensors.end());
output_tensors.insert(tensor);
+
+ tile_usages[tensor] = 0;
+ possible_tile_map.emplace(tensor, std::vector<const ITensorInfo *>());
+ }
+
+ // Check if the output can overwrite the input tile.
+ const auto component_type = component->type();
+ if(component_type == GpuComponentType::Simple || component_type == GpuComponentType::Output)
+ {
+ ARM_COMPUTE_ERROR_ON(dst_tensors.size() != 1);
+
+ const auto dst_tensor = dst_tensors[0];
+ const auto &dst_shape = dst_tensor->tensor_shape();
+ const auto &dst_type = dst_tensor->data_type();
+
+ tile_usages[dst_tensor] = 0;
+
+ for(auto src_tensor : src_tensors)
+ {
+ const auto &src_shape = src_tensor->tensor_shape();
+ const auto &src_type = src_tensor->data_type();
+
+ if(src_shape == dst_shape && src_type == dst_type)
+ {
+ const auto tile_usages_it = tile_usages.find(src_tensor);
+ ARM_COMPUTE_ERROR_ON(tile_usages_it == tile_usages.end());
+
+ if(component_type == GpuComponentType::Simple || tile_usages_it->second > 0)
+ {
+ // Increase the number of tile usages unless this component is an output
+ // and the tile has not been shared with any component.
+ // (Reason: output component doesn't change the content of the tile)
+ ++tile_usages_it->second;
+ }
+
+ possible_tile_map[dst_tensor].push_back(src_tensor);
+ }
+ }
+ }
+ else
+ {
+ // Outputs of complex and unfusable components need dedicated tile.
+ for(auto tensor : dst_tensors)
+ {
+ tile_usages[tensor] = 0;
+ }
+ }
+ }
+
+ // Find the smallest list of tiles that the intermediate tensors need to write to.
+ for(auto tensor : _input_tensors)
+ {
+ _tile_map[tensor] = tensor;
+ }
+
+ for(auto component : _components)
+ {
+ const auto dst_tensors = component->tensors().get_const_dst_tensors();
+
+ for(auto tensor : dst_tensors)
+ {
+ const auto target_tiles = possible_tile_map.at(tensor);
+ _tile_map[tensor] = tensor;
+
+ for(auto target : target_tiles)
+ {
+ const auto num_usage = tile_usages[target];
+
+ if(num_usage <= 1)
+ {
+ // The target tile is consumed by only this operator, so we can reuse it
+ // for the destination tensor data.
+ _tile_map[tensor] = _tile_map.at(target);
+ break;
+ }
+ }
+ }
+ }
+
+ for(auto tensor : output_tensors)
+ {
+ _tile_map[tensor] = tensor;
+ }
+
+ // All intermediate tensors that cannot be shared with any previous tensor
+ // will need to be declared as tile variable.
+ for(auto tensor_tile : _tile_map)
+ {
+ if(tensor_tile.first == tensor_tile.second &&
+ _interm_tensors.find(tensor_tile.first) != _interm_tensors.end())
+ {
+ _tiles.push_back(tensor_tile.first);
}
}
std::set_union(
- input_tensors.begin(), input_tensors.end(),
+ _input_tensors.begin(), _input_tensors.end(),
output_tensors.begin(), output_tensors.end(),
std::back_inserter(_argument_tensors));
_any_output_tensor = *output_tensors.begin();
}
+std::vector<const ITensorInfo *> GpuKernelComponentGroup::get_tiles() const
+{
+ ARM_COMPUTE_ERROR_ON_MSG(!_finalized, "The component group must have been finalized.");
+ return _tiles;
+}
+
+const ITensorInfo *GpuKernelComponentGroup::get_tile_for_tensor(const ITensorInfo *tensor) const
+{
+ ARM_COMPUTE_ERROR_ON_MSG(!_finalized, "The component group must have been finalized.");
+
+ if(_tile_map.find(tensor) != _tile_map.end())
+ {
+ return _tile_map.at(tensor);
+ }
+
+ return tensor;
+}
+
const ITensorInfo *GpuKernelComponentGroup::get_any_dst_tensor() const
{
ARM_COMPUTE_ERROR_ON_MSG(!_finalized, "The component group must have been finalized.");
@@ -203,6 +317,12 @@
return _interm_tensors.find(tensor) != _interm_tensors.end();
}
+bool GpuKernelComponentGroup::is_input_tensor(const ITensorInfo *tensor) const
+{
+ ARM_COMPUTE_ERROR_ON_MSG(!_finalized, "The component group must have been finalized.");
+ return _input_tensors.find(tensor) != _input_tensors.end();
+}
+
size_t GpuKernelComponentGroup::size() const
{
return _components.size();