From 42a31723ebe79895c9bb2297a9c2ef22c01a6f26 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Mon, 9 Jul 2018 14:35:32 +0100 Subject: COMPMID-1124 : Fixes in CLLSTM layer Change-Id: Ifc8e12c296d3ef2bf8e0f0bf1b87b7fd47a1fad7 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/139248 Tested-by: Jenkins Reviewed-by: Ruomei Yan Reviewed-by: Michalis Spyrou --- src/runtime/CL/functions/CLLSTMLayer.cpp | 142 ++++++++++++++----------------- 1 file changed, 63 insertions(+), 79 deletions(-) (limited to 'src/runtime/CL/functions/CLLSTMLayer.cpp') diff --git a/src/runtime/CL/functions/CLLSTMLayer.cpp b/src/runtime/CL/functions/CLLSTMLayer.cpp index a195ffa6b9..86e5eb9090 100644 --- a/src/runtime/CL/functions/CLLSTMLayer.cpp +++ b/src/runtime/CL/functions/CLLSTMLayer.cpp @@ -38,16 +38,15 @@ using namespace arm_compute; using namespace arm_compute::misc::shape_calculator; CLLSTMLayer::CLLSTMLayer(std::shared_ptr memory_manager) - : _memory_group(std::move(memory_manager)), _fully_connected_input_gate(), _gemm_input_gate1(), _gemm_input_gate2(), _transpose_input_gate1(), _transpose_input_gate2(), _accum_input_gate1(), - _accum_input_gate2(), _subtract_input_gate(), _activation_input_gate(), _fully_connected_forget_gate(), _gemm_forget_gate1(), _gemm_forget_gate2(), _transpose_forget_gate1(), - _transpose_forget_gate2(), _accum_forget_gate1(), _accum_forget_gate2(), _activation_forget_gate(), _fully_connected_cell_state(), _gemm_cell_state1(), _gemm_cell_state2(), _transpose_cell_state1(), - _accum_cell_state1(), _accum_cell_state2(), _pixelwise_mul_cell_state1(), _activation_cell_state(), _cell_clip(), _pixelwise_mul_cell_state2(), _fully_connected_output(), _gemm_output1(), - _gemm_output2(), _transpose_output1(), _transpose_output2(), _accum_output1(), _accum_output2(), _activation_output(), _activation_output_state(), _pixelwise_mul_output_state(), - _fully_connected_output_state(), _gemm_output_state(), _accum_output_state(), _projection_clip(), _copy_cell_state(), _copy_output(), _concat_scratch_buffer(), _input_gate_out1(), _input_gate_out2(), - _input_gate_out3(), _input_gate_out4(), _input_gate_out5(), _input_gate_out6(), _forget_gate_out1(), _forget_gate_out2(), _forget_gate_out3(), _forget_gate_out4(), _forget_gate_out5(), - _forget_gate_out6(), _cell_state_out1(), _cell_state_out2(), _cell_state_out3(), _cell_state_out4(), _cell_state_out5(), _output1(), _output2(), _output3(), _output4(), _output5(), _output6(), - _cell_state_activation(), _output_projection1(), _ones(), _run_peephole_opt(false), _run_cifg_opt(false), _perform_cell_clipping(false), _has_projection_weights(false), - _perform_projection_clipping(false) + : _memory_group(std::move(memory_manager)), _fully_connected_input_gate(), _gemm_input_gate(), _transpose_input_gate(), _accum_input_gate1(), _accum_input_gate2(), _subtract_input_gate(), + _pixelwise_mul_input_gate(), _activation_input_gate(), _fully_connected_forget_gate(), _gemm_forget_gate(), _transpose_forget_gate(), _accum_forget_gate1(), _accum_forget_gate2(), + _pixelwise_mul_forget_gate(), _activation_forget_gate(), _fully_connected_cell_state(), _gemm_cell_state1(), _gemm_cell_state2(), _transpose_cell_state(), _accum_cell_state1(), _accum_cell_state2(), + _pixelwise_mul_cell_state1(), _activation_cell_state(), _cell_clip(), _pixelwise_mul_cell_state2(), _fully_connected_output(), _gemm_output(), _pixelwise_mul_output_state1(), _transpose_output(), + _accum_output1(), _accum_output2(), _activation_output(), _activation_output_state(), _pixelwise_mul_output_state2(), _fully_connected_output_state(), _gemm_output_state(), _accum_output_state(), + _projection_clip(), _copy_cell_state(), _copy_output(), _concat_scratch_buffer(), _input_gate_out1(), _input_gate_out2(), _input_gate_out3(), _input_gate_out4(), _input_gate_out5(), + _forget_gate_out1(), _forget_gate_out2(), _forget_gate_out3(), _forget_gate_out4(), _forget_gate_out5(), _cell_state_out1(), _cell_state_out2(), _cell_state_out3(), _cell_state_out4(), + _cell_state_out5(), _output1(), _output2(), _output3(), _output4(), _output5(), _cell_state_activation(), _output_projection1(), _ones(), _run_peephole_opt(false), _run_cifg_opt(false), + _perform_cell_clipping(false), _has_projection_weights(false), _perform_projection_clipping(false) { } @@ -83,40 +82,34 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo const TensorShape cell_state_shape = cell_state->info()->tensor_shape(); TensorShape forget_gate1_shape = compute_transposed_shape(*recurrent_to_output_weights->info()); - TensorShape forget_gate2_shape = compute_transposed_shape(*forget_gate_bias->info()); - TensorShape forget_gate3_shape{ 1, output_state->info()->dimension(1) }; _forget_gate_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _forget_gate_out2.allocator()->init(TensorInfo(forget_gate1_shape, 1, input->info()->data_type())); _forget_gate_out3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); - _forget_gate_out6.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); + _forget_gate_out5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); // Configure block that calculates the forget gate - // forget_gate = Activation(input * input_to_forget_weights + output_state * recurrent_to_forget_weights + cell_state * cell_to_forget_weights + forget_gate_bias) + // forget_gate = Activation(input * input_to_forget_weights + output_state * recurrent_to_forget_weights + PixelWiseMul(cell_state, cell_to_forget_weights) + forget_gate_bias) _memory_group.manage(&_forget_gate_out1); _fully_connected_forget_gate.configure(input, input_to_forget_weights, forget_gate_bias, &_forget_gate_out1, true, false); _memory_group.manage(&_forget_gate_out2); - _transpose_forget_gate1.configure(recurrent_to_forget_weights, &_forget_gate_out2); + _transpose_forget_gate.configure(recurrent_to_forget_weights, &_forget_gate_out2); _memory_group.manage(&_forget_gate_out3); - _gemm_forget_gate1.configure(output_state, &_forget_gate_out2, nullptr, &_forget_gate_out3, 1.f, 0.f); + _gemm_forget_gate.configure(output_state, &_forget_gate_out2, nullptr, &_forget_gate_out3, 1.f, 0.f); _forget_gate_out2.allocator()->allocate(); - _memory_group.manage(&_forget_gate_out6); - _accum_forget_gate1.configure(&_forget_gate_out1, &_forget_gate_out3, &_forget_gate_out6, ConvertPolicy::SATURATE); - CLTensor *forget_gate_out = &_forget_gate_out6; + _memory_group.manage(&_forget_gate_out5); + _accum_forget_gate1.configure(&_forget_gate_out1, &_forget_gate_out3, &_forget_gate_out5, ConvertPolicy::SATURATE); + CLTensor *forget_gate_out = &_forget_gate_out5; if(lstm_params.has_peephole_opt()) { - _forget_gate_out4.allocator()->init(TensorInfo(forget_gate2_shape, 1, input->info()->data_type())); - _forget_gate_out5.allocator()->init(TensorInfo(forget_gate3_shape, 1, input->info()->data_type())); + _forget_gate_out4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _run_peephole_opt = true; _memory_group.manage(&_forget_gate_out4); - _transpose_forget_gate2.configure(lstm_params.cell_to_forget_weights(), &_forget_gate_out4); - _memory_group.manage(&_forget_gate_out5); - _gemm_forget_gate2.configure(cell_state, &_forget_gate_out4, nullptr, &_forget_gate_out5, 1.f, 0.f); + _pixelwise_mul_forget_gate.configure(cell_state, lstm_params.cell_to_forget_weights(), &_forget_gate_out4, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); + _accum_forget_gate2.configure(&_forget_gate_out5, &_forget_gate_out4, &_forget_gate_out3, ConvertPolicy::SATURATE); _forget_gate_out4.allocator()->allocate(); - _accum_forget_gate2.configure(&_forget_gate_out6, &_forget_gate_out5, &_forget_gate_out3, ConvertPolicy::SATURATE); _forget_gate_out5.allocator()->allocate(); - _forget_gate_out6.allocator()->allocate(); forget_gate_out = &_forget_gate_out3; } else @@ -126,12 +119,10 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo _activation_forget_gate.configure(forget_gate_out, &_forget_gate_out1, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); forget_gate_out->allocator()->allocate(); - TensorShape input_gate3_shape{ 1, output_state->info()->dimension(1) }; _input_gate_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); - _input_gate_out5.allocator()->init(TensorInfo(input_gate3_shape, 1, input->info()->data_type())); // Configure block that calculates the input gate - // input_gate = Activation(input * input_to_input_weights + output_state * recurrent_to_input_weights + cell_state * cell_to_input_weights + input_gate_bias), without CIFG + // 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 // input_gate = 1 - forget_gate, with CIFG if(lstm_params.has_cifg_opt()) { @@ -143,32 +134,28 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo } else { - TensorShape input_gate1_shape = compute_transposed_shape(*recurrent_to_output_weights->info()); - TensorShape input_gate2_shape = compute_transposed_shape(*lstm_params.cell_to_input_weights()->info()); + TensorShape input_gate_shape = compute_transposed_shape(*recurrent_to_output_weights->info()); - _input_gate_out2.allocator()->init(TensorInfo(input_gate1_shape, 1, input->info()->data_type())); + _input_gate_out2.allocator()->init(TensorInfo(input_gate_shape, 1, input->info()->data_type())); _input_gate_out3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); - _input_gate_out4.allocator()->init(TensorInfo(input_gate2_shape, 1, input->info()->data_type())); - _input_gate_out6.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); + _input_gate_out4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); + _input_gate_out5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _memory_group.manage(&_input_gate_out1); _fully_connected_input_gate.configure(input, lstm_params.input_to_input_weights(), lstm_params.input_gate_bias(), &_input_gate_out1, true, false); _memory_group.manage(&_input_gate_out2); - _transpose_input_gate1.configure(lstm_params.recurrent_to_input_weights(), &_input_gate_out2); + _transpose_input_gate.configure(lstm_params.recurrent_to_input_weights(), &_input_gate_out2); _memory_group.manage(&_input_gate_out3); - _gemm_input_gate1.configure(output_state, &_input_gate_out2, nullptr, &_input_gate_out3, 1.f, 0.f); + _gemm_input_gate.configure(output_state, &_input_gate_out2, nullptr, &_input_gate_out3, 1.f, 0.f); _input_gate_out2.allocator()->allocate(); _memory_group.manage(&_input_gate_out4); - _transpose_input_gate2.configure(lstm_params.cell_to_input_weights(), &_input_gate_out4); + _pixelwise_mul_input_gate.configure(cell_state, lstm_params.cell_to_input_weights(), &_input_gate_out4, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); _memory_group.manage(&_input_gate_out5); - _gemm_input_gate2.configure(cell_state, &_input_gate_out4, nullptr, &_input_gate_out5, 1.f, 0.f); - _input_gate_out4.allocator()->allocate(); - _memory_group.manage(&_input_gate_out6); - _accum_input_gate1.configure(&_input_gate_out1, &_input_gate_out3, &_input_gate_out6, ConvertPolicy::SATURATE); + _accum_input_gate1.configure(&_input_gate_out1, &_input_gate_out3, &_input_gate_out5, ConvertPolicy::SATURATE); _input_gate_out3.allocator()->allocate(); - _accum_input_gate2.configure(&_input_gate_out6, &_input_gate_out5, &_input_gate_out1, ConvertPolicy::SATURATE); + _accum_input_gate2.configure(&_input_gate_out5, &_input_gate_out4, &_input_gate_out1, ConvertPolicy::SATURATE); + _input_gate_out4.allocator()->allocate(); _input_gate_out5.allocator()->allocate(); - _input_gate_out6.allocator()->allocate(); _activation_input_gate.configure(&_input_gate_out1, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); } @@ -180,11 +167,11 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo _cell_state_out5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); // Configure block that calculates the cell state - // cell_state = Clip((RixelwiseMul(input_gate, Activation(input * input_to_cell_weights + output_state * recurrent_to_cell_weights + cell_bias)) + PixelwiseMul(forget_gate, cell_state)), cell_threshold) + // cell_state = Clip((PixelwiseMul(input_gate, Activation(input * input_to_cell_weights + output_state * recurrent_to_cell_weights + cell_bias)) + PixelwiseMul(forget_gate, cell_state)), cell_threshold) _memory_group.manage(&_cell_state_out1); _fully_connected_cell_state.configure(input, input_to_cell_weights, cell_bias, &_cell_state_out1, true, false); _memory_group.manage(&_cell_state_out2); - _transpose_cell_state1.configure(recurrent_to_cell_weights, &_cell_state_out2); + _transpose_cell_state.configure(recurrent_to_cell_weights, &_cell_state_out2); _memory_group.manage(&_cell_state_out3); _gemm_cell_state1.configure(output_state, &_cell_state_out2, nullptr, &_cell_state_out3, 1.f, 0.f); _cell_state_out2.allocator()->allocate(); @@ -209,42 +196,36 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo } TensorShape output1_shape = compute_transposed_shape(*recurrent_to_output_weights->info()); - TensorShape output2_shape = compute_transposed_shape(*cell_bias->info()); - TensorShape output3_shape{ 1, output_state->info()->dimension(1) }; _output1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _output2.allocator()->init(TensorInfo(output1_shape, 1, input->info()->data_type())); _output3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); - _output6.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); + _output5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); // Configure block that calculates the output - // output_gate = Activation(input * input_to_output_weights + output_state * recurrent_to_output_weights + cell_state * cell_to_output_weights + output_gate_bias) + // output_state = Activation(input * input_to_output_weights + output_state * recurrent_to_output_weights + PixelWiseMul(cell_state, cell_to_output_weights) + output_gate_bias) _memory_group.manage(&_output1); _fully_connected_output.configure(input, input_to_output_weights, output_gate_bias, &_output1, true, false); _memory_group.manage(&_output2); - _transpose_output1.configure(recurrent_to_output_weights, &_output2); + _transpose_output.configure(recurrent_to_output_weights, &_output2); _memory_group.manage(&_output3); - _gemm_output1.configure(output_state, &_output2, nullptr, &_output3, 1.f, 0.f); + _gemm_output.configure(output_state, &_output2, nullptr, &_output3, 1.f, 0.f); _output2.allocator()->allocate(); - _memory_group.manage(&_output6); - _accum_output1.configure(&_output1, &_output3, &_output6, ConvertPolicy::SATURATE); + _memory_group.manage(&_output5); + _accum_output1.configure(&_output1, &_output3, &_output5, ConvertPolicy::SATURATE); _output3.allocator()->allocate(); - CLTensor *output_gate_out = &_output6; + CLTensor *output_gate_out = &_output5; if(lstm_params.has_peephole_opt()) { - _output4.allocator()->init(TensorInfo(output2_shape, 1, input->info()->data_type())); - _output5.allocator()->init(TensorInfo(output3_shape, 1, input->info()->data_type())); + _output4.allocator()->init(TensorInfo(_cell_state_out1.info()->tensor_shape(), 1, input->info()->data_type())); _memory_group.manage(&_output4); - _transpose_output2.configure(lstm_params.cell_to_output_weights(), &_output4); - _memory_group.manage(&_output5); - _gemm_output2.configure(&_cell_state_out1, &_output4, nullptr, &_output5, 1.f, 0.f); - _accum_output2.configure(&_output6, &_output5, &_output1, ConvertPolicy::SATURATE); - _output6.allocator()->allocate(); + _pixelwise_mul_output_state1.configure(&_cell_state_out1, lstm_params.cell_to_output_weights(), &_output4, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); + _accum_output2.configure(&_output5, &_output4, &_output1, ConvertPolicy::SATURATE); + _output5.allocator()->allocate(); output_gate_out = &_output1; // Allocate intermediate buffers _output4.allocator()->allocate(); - _output5.allocator()->allocate(); } else { @@ -266,7 +247,7 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo */ _memory_group.manage(&_cell_state_activation); _activation_output_state.configure(&_cell_state_out1, &_cell_state_activation, activation_info); - _pixelwise_mul_output_state.configure(&_cell_state_activation, output, output_state, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); + _pixelwise_mul_output_state2.configure(&_cell_state_activation, output, output_state, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); _cell_state_activation.allocator()->allocate(); if(lstm_params.has_projection()) @@ -427,14 +408,13 @@ void CLLSTMLayer::run() _memory_group.acquire(); _fully_connected_forget_gate.run(); - CLScheduler::get().enqueue(_transpose_forget_gate1); - _gemm_forget_gate1.run(); + CLScheduler::get().enqueue(_transpose_forget_gate); + _gemm_forget_gate.run(); CLScheduler::get().enqueue(_accum_forget_gate1); if(_run_peephole_opt) { - CLScheduler::get().enqueue(_transpose_forget_gate2); - _gemm_forget_gate2.run(); + CLScheduler::get().enqueue(_pixelwise_mul_forget_gate); _accum_forget_gate2.run(); } CLScheduler::get().enqueue(_activation_forget_gate); @@ -442,24 +422,30 @@ void CLLSTMLayer::run() if(_run_cifg_opt) { _ones.map(true); - std::fill_n(_ones.buffer(), _ones.info()->total_size(), 1); + if(_ones.info()->data_type() == DataType::F16) + { + std::fill_n(reinterpret_cast(_ones.buffer()), _ones.info()->total_size() / _ones.info()->element_size(), 1); + } + else + { + std::fill_n(reinterpret_cast(_ones.buffer()), _ones.info()->total_size() / _ones.info()->element_size(), 1); + } _ones.unmap(); CLScheduler::get().enqueue(_subtract_input_gate); } else { _fully_connected_input_gate.run(); - CLScheduler::get().enqueue(_transpose_input_gate1); - _gemm_input_gate1.run(); - CLScheduler::get().enqueue(_transpose_input_gate2); - _gemm_input_gate2.run(); + CLScheduler::get().enqueue(_transpose_input_gate); + _gemm_input_gate.run(); + CLScheduler::get().enqueue(_pixelwise_mul_input_gate); CLScheduler::get().enqueue(_accum_input_gate1); _accum_input_gate2.run(); CLScheduler::get().enqueue(_activation_input_gate); } _fully_connected_cell_state.run(); - CLScheduler::get().enqueue(_transpose_cell_state1); + CLScheduler::get().enqueue(_transpose_cell_state); _gemm_cell_state1.run(); CLScheduler::get().enqueue(_accum_cell_state1); CLScheduler::get().enqueue(_activation_cell_state); @@ -473,21 +459,19 @@ void CLLSTMLayer::run() } _fully_connected_output.run(); - CLScheduler::get().enqueue(_transpose_output1); - _gemm_output1.run(); + CLScheduler::get().enqueue(_transpose_output); + _gemm_output.run(); CLScheduler::get().enqueue(_accum_output1); - CLScheduler::get().enqueue(_pixelwise_mul_output_state); if(_run_peephole_opt) { - CLScheduler::get().enqueue(_transpose_output2); - _gemm_output2.run(); + CLScheduler::get().enqueue(_pixelwise_mul_output_state1); _accum_output2.run(); } CLScheduler::get().enqueue(_activation_output); CLScheduler::get().enqueue(_activation_output_state); - CLScheduler::get().enqueue(_pixelwise_mul_output_state); + CLScheduler::get().enqueue(_pixelwise_mul_output_state2); if(_has_projection_weights) { -- cgit v1.2.1