COMPMID-998: Release unused trainable parameters.
Change-Id: I361a520f34080016a25bc86e1e6789777c5152c1
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/124432
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
diff --git a/src/core/ITensor.cpp b/src/core/ITensor.cpp
index b65c4f4..eb5f072 100644
--- a/src/core/ITensor.cpp
+++ b/src/core/ITensor.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -151,3 +151,13 @@
}
}
}
+
+bool ITensor::is_used() const
+{
+ return _is_used;
+}
+
+void ITensor::mark_as_unused() const
+{
+ _is_used = false;
+}
diff --git a/src/graph2/GraphManager.cpp b/src/graph2/GraphManager.cpp
index e708dc6..a51ba61 100644
--- a/src/graph2/GraphManager.cpp
+++ b/src/graph2/GraphManager.cpp
@@ -82,6 +82,12 @@
// Finalize Graph context
ctx.finalize();
+
+ // Make first run
+ execute_graph(graph);
+
+ // Release all unused const nodes
+ detail::release_unused_tensors(graph);
}
void GraphManager::execute_graph(Graph &graph)
diff --git a/src/graph2/backends/CL/CLSubTensorHandle.cpp b/src/graph2/backends/CL/CLSubTensorHandle.cpp
index a001d57..65a1ba4 100644
--- a/src/graph2/backends/CL/CLSubTensorHandle.cpp
+++ b/src/graph2/backends/CL/CLSubTensorHandle.cpp
@@ -64,6 +64,11 @@
_sub_tensor.unmap();
}
+void CLSubTensorHandle::release_if_unused()
+{
+ // noop
+}
+
bool CLSubTensorHandle::is_subtensor() const
{
return true;
diff --git a/src/graph2/backends/CL/CLTensorHandle.cpp b/src/graph2/backends/CL/CLTensorHandle.cpp
index f515e0b..89678fb 100644
--- a/src/graph2/backends/CL/CLTensorHandle.cpp
+++ b/src/graph2/backends/CL/CLTensorHandle.cpp
@@ -60,6 +60,15 @@
_tensor.unmap();
}
+void CLTensorHandle::release_if_unused()
+{
+ // TODO (geopin01): Release tensor only if all sub-tensors are marked as not used
+ if(!_tensor.is_used())
+ {
+ _tensor.allocator()->free();
+ }
+}
+
bool CLTensorHandle::is_subtensor() const
{
return false;
diff --git a/src/graph2/backends/GLES/GCTensorHandle.cpp b/src/graph2/backends/GLES/GCTensorHandle.cpp
index aa9ac8c..2165cd2 100644
--- a/src/graph2/backends/GLES/GCTensorHandle.cpp
+++ b/src/graph2/backends/GLES/GCTensorHandle.cpp
@@ -60,6 +60,15 @@
_tensor.unmap();
}
+void GCTensorHandle::release_if_unused()
+{
+ // TODO (geopin01): Release tensor only if all sub-tensors are marked as not used
+ if(!_tensor.is_used())
+ {
+ _tensor.allocator()->free();
+ }
+}
+
bool GCTensorHandle::is_subtensor() const
{
return false;
diff --git a/src/graph2/backends/NEON/NESubTensorHandle.cpp b/src/graph2/backends/NEON/NESubTensorHandle.cpp
index 491cf82..1cd15be 100644
--- a/src/graph2/backends/NEON/NESubTensorHandle.cpp
+++ b/src/graph2/backends/NEON/NESubTensorHandle.cpp
@@ -61,6 +61,11 @@
// noop
}
+void NESubTensorHandle::release_if_unused()
+{
+ // noop
+}
+
bool NESubTensorHandle::is_subtensor() const
{
return true;
diff --git a/src/graph2/backends/NEON/NETensorHandle.cpp b/src/graph2/backends/NEON/NETensorHandle.cpp
index a4af8aa..0b901c3 100644
--- a/src/graph2/backends/NEON/NETensorHandle.cpp
+++ b/src/graph2/backends/NEON/NETensorHandle.cpp
@@ -59,6 +59,15 @@
{
}
+void NETensorHandle::release_if_unused()
+{
+ // TODO (geopin01): Release tensor only if all sub-tensors are marked as not used
+ if(!_tensor.is_used())
+ {
+ _tensor.allocator()->free();
+ }
+}
+
bool NETensorHandle::is_subtensor() const
{
return false;
diff --git a/src/graph2/detail/ExecutionHelpers.cpp b/src/graph2/detail/ExecutionHelpers.cpp
index ae214ad..3688d0b 100644
--- a/src/graph2/detail/ExecutionHelpers.cpp
+++ b/src/graph2/detail/ExecutionHelpers.cpp
@@ -135,6 +135,17 @@
return workload;
}
+void release_unused_tensors(Graph &g)
+{
+ for(auto &tensor : g.tensors())
+ {
+ if(tensor != nullptr && tensor->handle() != nullptr)
+ {
+ tensor->handle()->release_if_unused();
+ }
+ }
+}
+
void call_tensor_accessor(Tensor *tensor)
{
ARM_COMPUTE_ERROR_ON(!tensor);
diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
index 112af60..8d7c92b 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
@@ -64,7 +64,7 @@
CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayer()
: _im2col_kernel(), _weights_reshape_kernel(), _v2mm_kernel(), _vector_to_tensor_kernel(), _output_stage_kernel(), _v2mm_input_fill_border(), _v2mm_weights_fill_border(), _input_reshaped(),
- _weights_reshaped(), _v2mm_output(), _output_reshaped(), _is_quantized(false)
+ _weights_reshaped(), _v2mm_output(), _output_reshaped(), _is_first_run(true), _is_quantized(false), _original_weights(nullptr)
{
}
@@ -78,7 +78,9 @@
const size_t weights_h = weights->info()->dimension(1);
const size_t weights_z = weights->info()->dimension(2);
- _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
+ _is_first_run = true;
+ _original_weights = weights;
+ _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
bool append_bias = (biases != nullptr) && !_is_quantized;
const GPUTarget gpu_target = CLScheduler::get().target();
@@ -154,16 +156,23 @@
void CLDepthwiseConvolutionLayer::run()
{
+ // Run weights reshaping (Runs once for every configure)
+ if(_is_first_run)
+ {
+ ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
+
+ CLScheduler::get().enqueue(_weights_reshape_kernel);
+ CLScheduler::get().enqueue(_v2mm_weights_fill_border);
+ _is_first_run = false;
+
+ // Mark original weights tensor as unused
+ _original_weights->mark_as_unused();
+ }
+
CLScheduler::get().enqueue(_im2col_kernel);
-
- CLScheduler::get().enqueue(_weights_reshape_kernel);
-
CLScheduler::get().enqueue(_v2mm_input_fill_border);
- CLScheduler::get().enqueue(_v2mm_weights_fill_border);
CLScheduler::get().enqueue(_v2mm_kernel);
-
CLScheduler::get().enqueue(_vector_to_tensor_kernel);
-
if(_is_quantized)
{
CLScheduler::get().enqueue(_output_stage_kernel);
diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
index 2b4670b..676706f 100644
--- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
@@ -76,7 +76,7 @@
CLFullyConnectedLayer::CLFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager)
: _memory_group(memory_manager), _im2col_kernel(), _reshape_weights_kernel(), _mm_kernel(), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), _accumulate_biases_kernel(), _im2col_output(),
- _gemmlowp_output(), _reshape_weights_output(), _are_weights_reshaped(true), _is_fc_after_conv(true), _accumulate_biases(false), _is_quantized(false)
+ _gemmlowp_output(), _reshape_weights_output(), _are_weights_reshaped(true), _is_fc_after_conv(true), _accumulate_biases(false), _is_quantized(false), _original_weights(nullptr)
{
}
@@ -152,6 +152,7 @@
_is_fc_after_conv = true;
_accumulate_biases = false;
_is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
+ _original_weights = weights;
// Configure gemmlowp output
if(_is_quantized)
@@ -316,8 +317,13 @@
// Reshape of the weights (happens only once)
if(!_are_weights_reshaped)
{
+ ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
+
_are_weights_reshaped = true;
_reshape_weights_kernel.run();
+
+ // Mark original weights tensor as unused
+ _original_weights->mark_as_unused();
}
_memory_group.acquire();
diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
index e7ad62f..f43e100 100644
--- a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
@@ -90,8 +90,8 @@
}
CLGEMMConvolutionLayer::CLGEMMConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
- : _memory_group(memory_manager), _reshape_weights(), _im2col_kernel(), _mm_gemm(memory_manager), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), _col2im_kernel(), _im2col_output(),
- _weights_reshaped(), _gemm_output(), _tmp_output(), _is_quantized(false), _is_first_run(true)
+ : _memory_group(memory_manager), _reshape_weights(), _im2col_kernel(), _mm_gemm(memory_manager), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), _col2im_kernel(), _original_weights(nullptr),
+ _im2col_output(), _weights_reshaped(), _gemm_output(), _tmp_output(), _is_quantized(false), _is_first_run(true)
{
}
@@ -164,7 +164,9 @@
weights_info,
dilation));
- _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
+ _is_first_run = true;
+ _original_weights = weights;
+ _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
const DataType dt = input->info()->data_type();
@@ -349,9 +351,13 @@
// Run weights reshaping (Runs once for every configure)
if(_is_first_run)
{
- _reshape_weights.run();
+ ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
+ _reshape_weights.run();
_is_first_run = false;
+
+ // Mark original weights tensor as unused
+ _original_weights->mark_as_unused();
}
_memory_group.acquire();
diff --git a/src/runtime/CL/functions/CLLocallyConnectedLayer.cpp b/src/runtime/CL/functions/CLLocallyConnectedLayer.cpp
index a3eb501..986fe00 100644
--- a/src/runtime/CL/functions/CLLocallyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLLocallyConnectedLayer.cpp
@@ -73,7 +73,7 @@
CLLocallyConnectedLayer::CLLocallyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager)
: _memory_group(std::move(memory_manager)), _input_im2col_kernel(), _weights_reshape_kernel(), _mm_kernel(), _output_col2im_kernel(), _input_im2col_reshaped(), _weights_reshaped(), _gemm_output(),
- _is_first_run(false)
+ _is_first_run(false), _original_weights(nullptr)
{
}
@@ -126,8 +126,9 @@
ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
ARM_COMPUTE_ERROR_THROW_ON(CLLocallyConnectedLayer::validate(input->info(), weights->info(), biases == nullptr ? nullptr : biases->info(), output->info(), conv_info));
- bool _has_bias = (biases != nullptr);
- _is_first_run = true;
+ bool _has_bias = (biases != nullptr);
+ _original_weights = weights;
+ _is_first_run = true;
const unsigned int kernel_width = weights->info()->dimension(0);
const unsigned int kernel_height = weights->info()->dimension(1);
@@ -169,8 +170,13 @@
// Run weights reshaping (Runs once for every configure)
if(_is_first_run)
{
+ ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
+
_is_first_run = false;
CLScheduler::get().enqueue(_weights_reshape_kernel);
+
+ // Mark original weights tensor as unused
+ _original_weights->mark_as_unused();
}
_memory_group.acquire();
diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
index 95fcf88..f28ed71 100644
--- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
@@ -153,7 +153,7 @@
NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayer()
: _im2col_kernel(), _weights_reshape_kernel(), _v2mm_kernel(), _vector_to_tensor_kernel(), _output_stage_kernel(), _v2mm_input_fill_border(), _v2mm_weights_fill_border(), _input_reshaped(),
- _weights_reshaped(), _v2mm_output(), _output_reshaped(), _is_quantized(false)
+ _weights_reshaped(), _v2mm_output(), _output_reshaped(), _is_first_run(true), _is_quantized(false), _original_weights(nullptr)
{
}
@@ -167,7 +167,9 @@
const size_t weights_h = weights->info()->dimension(1);
const size_t weights_z = weights->info()->dimension(2);
- _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
+ _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
+ _is_first_run = true;
+ _original_weights = weights;
// Should bias be appended ?
bool append_bias = (biases != nullptr) && !_is_quantized;
@@ -241,10 +243,21 @@
void NEDepthwiseConvolutionLayer::run()
{
+ // Run weights reshaping (Runs once for every configure)
+ if(_is_first_run)
+ {
+ ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
+
+ NEScheduler::get().schedule(&_weights_reshape_kernel, Window::DimX);
+ NEScheduler::get().schedule(&_v2mm_weights_fill_border, Window::DimX);
+ _is_first_run = false;
+
+ // Mark original weights tensor as unused
+ _original_weights->mark_as_unused();
+ }
+
NEScheduler::get().schedule(&_im2col_kernel, Window::DimX);
- NEScheduler::get().schedule(&_weights_reshape_kernel, Window::DimX);
NEScheduler::get().schedule(&_v2mm_input_fill_border, Window::DimX);
- NEScheduler::get().schedule(&_v2mm_weights_fill_border, Window::DimX);
NEScheduler::get().schedule(&_v2mm_kernel, Window::DimX);
NEScheduler::get().schedule(&_vector_to_tensor_kernel, Window::DimX);
if(_is_quantized)
diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
index 26b7271..b310ad3 100644
--- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
+++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
@@ -132,7 +132,7 @@
NEFullyConnectedLayer::NEFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager)
: _memory_group(std::move(memory_manager)), _im2col_kernel(), _reshape_weights_kernel(), _interleave4x4_kernel(), _mm_kernel(), _accumulate_biases_kernel(), _im2col_output(), _interleave4x4_output(),
- _reshape_weights_output(), _are_weights_reshaped(false), _is_batched_fc_layer(false), _linearize_input(false), _accumulate_biases(false)
+ _reshape_weights_output(), _are_weights_reshaped(false), _is_batched_fc_layer(false), _linearize_input(false), _accumulate_biases(false), _original_weights(nullptr)
{
}
@@ -163,6 +163,7 @@
const int num_input_dimensions = input->info()->tensor_shape().num_dimensions() - num_batch_dimensions;
const size_t linear_input_size = input->info()->tensor_shape().total_size_lower(num_input_dimensions);
+ _original_weights = weights;
_linearize_input = (input->info()->tensor_shape().x() != linear_input_size) || (num_input_dimensions > 1 && linear_input_size == 1);
_are_weights_reshaped = are_weights_reshaped;
_accumulate_biases = biases != nullptr;
@@ -324,8 +325,13 @@
// Reshape of the weights (happens only once)
if(!_are_weights_reshaped)
{
+ ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
+
_are_weights_reshaped = true;
_reshape_weights_kernel.run();
+
+ // Mark original weights tensor as unused
+ _original_weights->mark_as_unused();
}
_memory_group.acquire();
diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
index d9707d9..b2dd022 100644
--- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
@@ -217,7 +217,7 @@
NEGEMMConvolutionLayer::NEGEMMConvolutionLayer(const std::shared_ptr<IMemoryManager> &memory_manager)
: _asm_glue(), _memory_group(memory_manager), _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(),
- _output_col2im_kernel(), _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), _tmp_output(), _workspace(), _append_bias(false),
+ _output_col2im_kernel(), _original_weights(nullptr), _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), _tmp_output(), _workspace(), _append_bias(false),
_is_fully_connected_convolution(false), _are_weights_reshaped(false), _is_quantized(false), _is_interleaved(false)
{
}
@@ -267,6 +267,7 @@
ARM_COMPUTE_ERROR_THROW_ON(status);
+ _original_weights = weights;
const unsigned int fixed_point_position = input->info()->fixed_point_position();
const ITensor *biases_to_use = (_append_bias) ? biases : nullptr;
@@ -549,8 +550,13 @@
// Run weights reshaping (Runs once for every configure)
if(!_are_weights_reshaped)
{
+ ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
+
_are_weights_reshaped = true;
_reshape_weights.run();
+
+ // Mark original weights tensor as unused
+ _original_weights->mark_as_unused();
}
_memory_group.acquire();
diff --git a/src/runtime/NEON/functions/NELocallyConnectedLayer.cpp b/src/runtime/NEON/functions/NELocallyConnectedLayer.cpp
index 9735594..913acf8 100644
--- a/src/runtime/NEON/functions/NELocallyConnectedLayer.cpp
+++ b/src/runtime/NEON/functions/NELocallyConnectedLayer.cpp
@@ -73,7 +73,7 @@
NELocallyConnectedLayer::NELocallyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager)
: _memory_group(std::move(memory_manager)), _input_im2col_kernel(), _weights_reshape_kernel(), _mm_kernel(), _output_col2im_kernel(), _input_im2col_reshaped(), _weights_reshaped(), _gemm_output(),
- _is_first_run(false)
+ _is_first_run(false), _original_weights(nullptr)
{
}
@@ -126,8 +126,9 @@
ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
ARM_COMPUTE_ERROR_THROW_ON(NELocallyConnectedLayer::validate(input->info(), weights->info(), biases == nullptr ? nullptr : biases->info(), output->info(), conv_info));
- bool _has_bias = (biases != nullptr);
- _is_first_run = true;
+ bool _has_bias = (biases != nullptr);
+ _is_first_run = true;
+ _original_weights = weights;
const unsigned int kernel_width = weights->info()->dimension(0);
const unsigned int kernel_height = weights->info()->dimension(1);
@@ -169,8 +170,13 @@
// Run weights reshaping (Runs once for every configure)
if(_is_first_run)
{
+ ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
+
_is_first_run = false;
NEScheduler::get().schedule(&_weights_reshape_kernel, 3);
+
+ // Mark original weights tensor as unused
+ _original_weights->mark_as_unused();
}
_memory_group.acquire();