Add output operator for dynamic fusion
* The output of the fused operator must be explicitly specified
using GpuOutput operator.
* Any temporary tensors used to connect the output of an operator
to the input of another operator will be marked as no-alloc
and won't be allocated as a tensor in the memory.
Resolves: COMPMID-5771
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I5ae8e800f8f737db23a055a92b01c4f1d78c3bb8
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8794
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/dynamic_fusion/sketch/gpu/GpuKernelComponentGraph.cpp b/src/dynamic_fusion/sketch/gpu/GpuKernelComponentGraph.cpp
index 6e6422c..1f90aab 100644
--- a/src/dynamic_fusion/sketch/gpu/GpuKernelComponentGraph.cpp
+++ b/src/dynamic_fusion/sketch/gpu/GpuKernelComponentGraph.cpp
@@ -42,15 +42,24 @@
*/
MemoryDescriptorMap assign_memory_descriptors(const std::map<ITensorInfo::Id, const ITensorInfo *> tensors, const DependencyGraph &graph)
{
+ const auto all_tensors = graph.all_tensors();
+ const auto src_tensors = graph.global_src_tensors();
+ const auto dst_tensors = graph.global_dst_tensors();
+ const auto interm_tensors = graph.intermediate_tensors();
+
MemoryDescriptorMap mem_map{};
- for(auto t_id : graph.all_tensors())
+ for(auto t_id : all_tensors)
{
const auto &tensor = tensors.at(t_id);
// Only global src and dst tensors to the entire component graph are "User" tensors, which are user-specified memories
- if(is_in(t_id, graph.global_src_tensors()) || is_in(t_id, graph.global_dst_tensors()))
+ if(is_in(t_id, src_tensors) || is_in(t_id, dst_tensors))
{
mem_map[t_id] = MemoryDescriptor{ MemoryType::User };
}
+ else if(is_in(t_id, interm_tensors))
+ {
+ mem_map[t_id] = MemoryDescriptor { MemoryType::NoAlloc };
+ }
else
{
AuxMemoryInfo aux_mem_info{ tensor->total_size() };