diff options
| author | Michele Di Giorgio <michele.digiorgio@arm.com> | 2018-11-05 16:46:09 +0000 |
|---|---|---|
| committer | Michele Di Giorgio <michele.digiorgio@arm.com> | 2018-11-06 14:48:21 +0000 |
| commit | dd2619a777d8faaa17a7cd7c8f20c036903947ad (patch) | |
| tree | 8470acc52c782361471ddffa4bbe468ca0645f03 | |
| parent | 2709d61ec25156387e48a75d19a5ab5a035fc12b (diff) | |
| download | ComputeLibrary-dd2619a777d8faaa17a7cd7c8f20c036903947ad.tar.gz | |
COMPMID-1451: Fix order of allocations in CLLSTMLayer
ArmNN reported an issue with padding in CLLSTMLayer. This was due to the fact
that some tensors were allocated before they were passed to some configure
functions which attempted to change the padding requirement on already allocated
memory.
Also, increase tolerance on number of mismatches for CLBBoxTransform FP16.
Change-Id: Iad75b012be895693d0e553f3ab85f1ca7144e882
| -rw-r--r-- | src/runtime/CL/functions/CLLSTMLayer.cpp | 8 | ||||
| -rw-r--r-- | tests/validation/CL/BoundingBoxTransform.cpp | 2 |
2 files changed, 5 insertions, 5 deletions
diff --git a/src/runtime/CL/functions/CLLSTMLayer.cpp b/src/runtime/CL/functions/CLLSTMLayer.cpp index 3458135799..245016efcb 100644 --- a/src/runtime/CL/functions/CLLSTMLayer.cpp +++ b/src/runtime/CL/functions/CLLSTMLayer.cpp @@ -130,7 +130,6 @@ void CLLSTMLayer::configure(const ICLTensor *input, _forget_gate_out3.allocator()->allocate(); } _activation_forget_gate.configure(forget_gate_out, &_forget_gate_out1, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); - forget_gate_out->allocator()->allocate(); // Configure block that calculates the input gate // input_gate = Activation(input * input_to_input_weights + output_state * recurrent_to_input_weights + PixelWiseMul(cell_state, cell_to_input_weights) + input_gate_bias), without CIFG @@ -195,7 +194,6 @@ void CLLSTMLayer::configure(const ICLTensor *input, _activation_cell_state.configure(&_cell_state_out4, nullptr, activation_info); _memory_group.manage(&_cell_state_out5); _pixelwise_mul_cell_state1.configure(&_cell_state_out4, &_input_gate_out1, &_cell_state_out5, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); - _input_gate_out1.allocator()->allocate(); _cell_state_out4.allocator()->allocate(); _pixelwise_mul_cell_state2.configure(&_forget_gate_out1, cell_state_in, &_cell_state_out3, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); _forget_gate_out1.allocator()->allocate(); @@ -246,7 +244,6 @@ void CLLSTMLayer::configure(const ICLTensor *input, _output1.allocator()->allocate(); } _activation_output.configure(output_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); - output_gate_out->allocator()->allocate(); // Configure block that calculates the output state /** lstm_res = PixelwiseMul(output, Activation(cell_state)) @@ -281,7 +278,6 @@ void CLLSTMLayer::configure(const ICLTensor *input, // Copy cell state and output _copy_cell_state.configure(&_cell_state_out1, cell_state_out); - _cell_state_out1.allocator()->allocate(); _copy_output.configure(output_state_out, output); // Vector for holding the tensors to store in scratch buffer @@ -294,6 +290,10 @@ void CLLSTMLayer::configure(const ICLTensor *input, scratch_inputs.emplace_back(forget_gate_out); scratch_inputs.emplace_back(output_gate_out); _concat_scratch_buffer.configure(scratch_inputs, scratch_buffer); + _input_gate_out1.allocator()->allocate(); + _cell_state_out1.allocator()->allocate(); + forget_gate_out->allocator()->allocate(); + output_gate_out->allocator()->allocate(); } Status CLLSTMLayer::validate(const ITensorInfo *input, diff --git a/tests/validation/CL/BoundingBoxTransform.cpp b/tests/validation/CL/BoundingBoxTransform.cpp index b18f36adbc..dba3a0eabe 100644 --- a/tests/validation/CL/BoundingBoxTransform.cpp +++ b/tests/validation/CL/BoundingBoxTransform.cpp @@ -115,7 +115,7 @@ FIXTURE_DATA_TEST_CASE(BoundingBox, CLBoundingBoxTransformFixture<half>, framewo combine(combine(DeltaDataset, BboxInfoDataset), framework::dataset::make("DataType", { DataType::F16 }))) { // Validate output - validate(CLAccessor(_target), _reference, relative_tolerance_f16, 0.01f, absolute_tolerance_f16); + validate(CLAccessor(_target), _reference, relative_tolerance_f16, 0.03f, absolute_tolerance_f16); } TEST_SUITE_END() // FP16 TEST_SUITE_END() // Float |