Ensure that correct transformed matrices are used in CpuFullyConnected
Execution pack of CpuFullyConnected was altered explicitly with local
objects that were getting out of scope. Leading to incorrect results or
memory related issues.
Track transformed weights and register the weights matrix explicitly
during execution honoring the object lifetime scope.
Resolves: COMPMID-4762, COMPMID-4764
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I53449c377fb1cfccdf5e6f9505d963518748c318
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/349345
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6092
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
index d815a73..504200e 100644
--- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
+++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
@@ -44,7 +44,6 @@
const ITensor *original_weights{ nullptr };
ITensorPack run_pack{};
- ITensorPack prep_pack{};
WorkspaceData<Tensor> workspace{};
experimental::MemoryRequirements aux_mem_req{};
@@ -79,8 +78,7 @@
_impl->aux_mem_req = _impl->op->workspace();
_impl->run_pack = { { ACL_SRC_0, input }, { ACL_SRC_1, weights }, { ACL_SRC_2, biases }, { ACL_DST, output } };
- _impl->prep_pack = { { ACL_SRC_1, weights }, { ACL_SRC_2, biases } };
- _impl->workspace = manage_workspace<Tensor>(_impl->aux_mem_req, _impl->memory_group, _impl->run_pack, _impl->prep_pack);
+ _impl->workspace = manage_workspace<Tensor>(_impl->aux_mem_req, _impl->memory_group, _impl->run_pack, _impl->run_pack);
}
Status NEFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
@@ -101,20 +99,7 @@
{
if(!_impl->is_prepared)
{
- _impl->op->prepare(_impl->prep_pack);
-
- auto has_reshape = std::find_if(_impl->aux_mem_req.begin(),
- _impl->aux_mem_req.end(),
- [](const MemoryInfo & m) -> bool { return m.lifetime == MemoryLifetime::Persistent; });
-
- if(has_reshape != std::end(_impl->aux_mem_req))
- {
- _impl->original_weights->mark_as_unused();
- }
- else
- {
- _impl->run_pack.add_const_tensor(ACL_SRC_1, _impl->original_weights);
- }
+ _impl->op->prepare(_impl->run_pack);
// Release temporary tensors that are only used in prepare stage
release_temporaries<Tensor>(_impl->aux_mem_req, _impl->workspace);
diff --git a/src/runtime/cpu/operators/CpuFullyConnected.cpp b/src/runtime/cpu/operators/CpuFullyConnected.cpp
index e7808fb..eeabce0 100644
--- a/src/runtime/cpu/operators/CpuFullyConnected.cpp
+++ b/src/runtime/cpu/operators/CpuFullyConnected.cpp
@@ -150,9 +150,11 @@
_flattened_src(),
_converted_weights(),
_reshaped_weights(),
+ _trans_weights(),
+ _trans_weights_idx(AuxTensorIdx::Count),
_aux_mem(Count),
- _are_weights_converted(false),
- _are_weights_reshaped(false),
+ _needs_weights_conversion(false),
+ _needs_weights_reshape(false),
_is_fc_after_conv(false),
_is_quantized_asymmetric(false),
_is_prepared(false)
@@ -230,11 +232,13 @@
dst,
fc_info));
- _are_weights_converted = true;
- _are_weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
- _is_fc_after_conv = true;
- _is_quantized_asymmetric = is_data_type_quantized_asymmetric(src->data_type());
- _is_prepared = false;
+ _needs_weights_conversion = false;
+ _needs_weights_reshape = fc_info.transpose_weights ? !fc_info.are_weights_reshaped : false;
+ _needs_weights_reshape = _needs_weights_reshape && !fc_info.retain_internal_weights;
+ _is_fc_after_conv = true;
+ _is_quantized_asymmetric = is_data_type_quantized_asymmetric(src->data_type());
+ _is_prepared = false;
+ _trans_weights_idx = AuxTensorIdx::Count;
// With the Fully Connected layer we can have 4 different cases:
// 1) Convolution layer -> Fully Connected layer without batches
@@ -258,12 +262,13 @@
}
// Reshape weights if needed
- if(!_are_weights_reshaped)
+ if(_needs_weights_reshape)
{
// Reshape the weights
_transpose_weights = std::make_unique<kernels::CpuTransposeKernel>();
_transpose_weights->configure(weights, &_reshaped_weights);
- weights_to_use = &_reshaped_weights;
+ weights_to_use = &_reshaped_weights;
+ _trans_weights_idx = AuxTensorIdx::TransposedWeights;
}
// Convert weights if needed
@@ -276,8 +281,9 @@
src->tensor_shape(),
fc_info.weights_trained_layout);
- weights_to_use = &_converted_weights;
- _are_weights_converted = false;
+ weights_to_use = &_converted_weights;
+ _needs_weights_conversion = true;
+ _trans_weights_idx = AuxTensorIdx::ConvertedWeights;
}
if(_is_fc_after_conv)
@@ -291,7 +297,11 @@
configure_fc_fc(src, weights_to_use, biases, dst, fc_info.activation_info);
}
- _are_weights_reshaped = _are_weights_reshaped || fc_info.retain_internal_weights;
+ // Retain the tensorinfo with the weights to use
+ if(_needs_weights_reshape || _needs_weights_conversion)
+ {
+ _trans_weights = *weights_to_use;
+ }
// Set auxiliary memory requirements
auto gemm_mem_req = (_is_quantized_asymmetric) ? _mm_gemmlowp->workspace() : _mm_gemm->workspace();
@@ -308,7 +318,7 @@
}
else
{
- _aux_mem[TransposedWeights] = MemoryInfo(offset_int_vec(TransposedWeights), MemoryLifetime::Persistent, _reshaped_weights.total_size());
+ _aux_mem[TransposedWeights] = MemoryInfo(offset_int_vec(TransposedWeights), _needs_weights_conversion ? MemoryLifetime::Prepare : MemoryLifetime::Persistent, _reshaped_weights.total_size());
_aux_mem[ConvertedWeights] = MemoryInfo(offset_int_vec(ConvertedWeights), MemoryLifetime::Persistent, _converted_weights.total_size());
}
_aux_mem[FlattenedSrc] = MemoryInfo(offset_int_vec(FlattenedSrc), MemoryLifetime::Temporary, _flattened_src.total_size());
@@ -401,6 +411,7 @@
auto src = tensors.get_const_tensor(ACL_SRC_0);
CpuAuxTensorHandler flattened_src(offset_int_vec(FlattenedSrc), _flattened_src, tensors, false);
+ CpuAuxTensorHandler transformed_wei(offset_int_vec(_trans_weights_idx), _trans_weights, tensors, false);
// Linearize src if it comes from a convolutional layer
if(_is_fc_after_conv)
@@ -411,6 +422,10 @@
ITensorPack gemm_pack = tensors;
gemm_pack.add_const_tensor(ACL_SRC_0, (_is_fc_after_conv) ? flattened_src.get() : src);
+ if(_needs_weights_reshape || _needs_weights_conversion)
+ {
+ gemm_pack.add_const_tensor(ACL_SRC_1, transformed_wei.get());
+ }
// Run matrix multiply
if(_is_quantized_asymmetric)
@@ -436,7 +451,7 @@
const ITensor *cur_weights = weights;
// Reshape of the weights (happens only once)
- if(!_are_weights_reshaped)
+ if(_needs_weights_reshape)
{
// Run reshape weights kernel and mark weights as unused
ITensorPack transpose_pack{ { ACL_SRC, weights }, { ACL_DST, reshaped_weights.get() } };
@@ -444,32 +459,29 @@
cur_weights->mark_as_unused();
cur_weights = reshaped_weights.get();
-
- _are_weights_reshaped = true;
}
// Convert weights if needed (happens only once)
- if(!_are_weights_converted)
+ if(_needs_weights_conversion)
{
ITensorPack convert_pack{ { ACL_SRC, cur_weights }, { ACL_DST, converted_weights.get() } };
_convert_weights->run(convert_pack);
cur_weights->mark_as_unused();
cur_weights = converted_weights.get();
-
- _are_weights_converted = true;
}
- tensors.add_const_tensor(ACL_SRC_1, cur_weights);
+ ITensorPack gemm_pack = tensors;
+ gemm_pack.add_const_tensor(ACL_SRC_1, cur_weights);
// Prepare GEMM prepare and release unused weights
if(!_is_quantized_asymmetric)
{
- _mm_gemm->prepare(tensors);
+ _mm_gemm->prepare(gemm_pack);
}
else
{
- _mm_gemmlowp->prepare(tensors);
+ _mm_gemmlowp->prepare(gemm_pack);
}
_is_prepared = true;
diff --git a/src/runtime/cpu/operators/CpuFullyConnected.h b/src/runtime/cpu/operators/CpuFullyConnected.h
index 954a7b7..498ceae 100644
--- a/src/runtime/cpu/operators/CpuFullyConnected.h
+++ b/src/runtime/cpu/operators/CpuFullyConnected.h
@@ -128,14 +128,16 @@
std::unique_ptr<CpuGemm> _mm_gemm;
std::unique_ptr<CpuGemmLowpMatrixMultiplyCore> _mm_gemmlowp;
- TensorInfo _flattened_src;
- TensorInfo _converted_weights;
- TensorInfo _reshaped_weights;
+ TensorInfo _flattened_src;
+ TensorInfo _converted_weights;
+ TensorInfo _reshaped_weights;
+ TensorInfo _trans_weights;
+ AuxTensorIdx _trans_weights_idx;
experimental::MemoryRequirements _aux_mem;
- bool _are_weights_converted;
- bool _are_weights_reshaped;
+ bool _needs_weights_conversion;
+ bool _needs_weights_reshape;
bool _is_fc_after_conv;
bool _is_quantized_asymmetric;
bool _is_prepared;
diff --git a/src/runtime/cpu/operators/CpuGemmLowpMatrixMultiplyCore.cpp b/src/runtime/cpu/operators/CpuGemmLowpMatrixMultiplyCore.cpp
index 8adf704..f224468 100644
--- a/src/runtime/cpu/operators/CpuGemmLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/cpu/operators/CpuGemmLowpMatrixMultiplyCore.cpp
@@ -672,15 +672,6 @@
if(_asm_glue->is_configured())
{
_asm_glue->prepare(tensors);
-
- auto has_reshape = std::find_if(_aux_mem.begin(),
- _aux_mem.end(),
- [](const MemoryInfo & m) -> bool { return m.lifetime == MemoryLifetime::Persistent; });
-
- if(has_reshape != std::end(_aux_mem))
- {
- original_b->mark_as_unused();
- }
}
// Run non-assembly reshape
else if(_reshape_b_only_on_first_run && !_run_vector_matrix_multiplication && !_asm_glue->is_configured())
diff --git a/src/runtime/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp b/src/runtime/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp
index bbbd5ac..9786161 100644
--- a/src/runtime/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp
+++ b/src/runtime/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp
@@ -424,6 +424,8 @@
CpuAuxTensorHandler pretranspose(offset_int_vec(Pretranspose), _pretranspose_info, tensors, false);
ARM_COMPUTE_ERROR_ON(pretranspose.get()->buffer() == nullptr);
_gemm_kernel_asm->pretranspose_B_array(pretranspose.get()->buffer(), in1_ptr, ldb, multi_stride_b);
+
+ b->mark_as_unused();
}
if(_gemm_info.method == AsmConvMethod::Indirect)