COMPMID-3240: Add support for layer normalization to CLQLSTMLayer
Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: I45359a4ddb46c059097a2d77c008f802e8f4c143
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3065
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/arm_compute/core/CL/kernels/CLQLSTMLayerNormalizationKernel.h b/arm_compute/core/CL/kernels/CLQLSTMLayerNormalizationKernel.h
index 1a2f311..2d47072 100644
--- a/arm_compute/core/CL/kernels/CLQLSTMLayerNormalizationKernel.h
+++ b/arm_compute/core/CL/kernels/CLQLSTMLayerNormalizationKernel.h
@@ -73,7 +73,7 @@
*
* @return a status
*/
- static Status validate(const ITensorInfo *input, ITensorInfo *output, const ITensorInfo *weight, const ITensorInfo *bias);
+ static Status validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *weight, const ITensorInfo *bias);
// Inherited methods overridden:
void run(const Window &window, cl::CommandQueue &queue) override;
diff --git a/arm_compute/runtime/CL/functions/CLQLSTMLayer.h b/arm_compute/runtime/CL/functions/CLQLSTMLayer.h
index 72a61f8..722275e 100644
--- a/arm_compute/runtime/CL/functions/CLQLSTMLayer.h
+++ b/arm_compute/runtime/CL/functions/CLQLSTMLayer.h
@@ -27,6 +27,7 @@
#include "arm_compute/core/CL/kernels/CLElementwiseOperationKernel.h"
#include "arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h"
#include "arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h"
+#include "arm_compute/core/CL/kernels/CLQLSTMLayerNormalizationKernel.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/runtime/CL/functions/CLActivationLayer.h"
#include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
@@ -216,6 +217,16 @@
void prepare() override;
private:
+ enum class LayerNormGate : uint8_t
+ {
+ Forget,
+ Cell,
+ Input,
+ Output,
+ Count
+ };
+ static constexpr uint8_t _layer_norm_count = static_cast<uint8_t>(LayerNormGate::Count);
+
/** Internal method to configure matrix multiplication plus output stage of each gate.
*
* @param[in] compile_context The compile context to be used.
@@ -302,6 +313,7 @@
CLGEMMLowpOutputStage _projection_outstage{};
CLSaturatedArithmeticOperationKernel _accumulate_projection{};
CLActivationLayer _projection_clip{};
+ std::array<CLQLSTMLayerNormalizationKernel, _layer_norm_count> _layer_norms{ {} };
// Tensor pointers
const ICLTensor *_input_to_input_weights
@@ -317,6 +329,61 @@
const ICLTensor *_recurrent_to_cell_weights{ nullptr };
const ICLTensor *_recurrent_to_output_weights{ nullptr };
const ICLTensor *_projection_weights{ nullptr };
+ std::array<const ICLTensor *, _layer_norm_count> _layer_norm_weights{ {} };
+ std::array<const ICLTensor *, _layer_norm_count> _layer_norm_bias{ {} };
+
+ using LayerNormIndexType = typename std::underlying_type<LayerNormGate>::type;
+ inline LayerNormIndexType getGateIndex(LayerNormGate g)
+ {
+ return static_cast<LayerNormIndexType>(g);
+ }
+
+ inline void set_layer_norm_weight(const ICLTensor *t, LayerNormGate g)
+ {
+ _layer_norm_weights[getGateIndex(g)] = t;
+ }
+
+ inline void set_layer_norm_bias(const ICLTensor *t, LayerNormGate g)
+ {
+ _layer_norm_bias[getGateIndex(g)] = t;
+ }
+
+ inline const ICLTensor *get_layer_norm_weight(LayerNormGate g)
+ {
+ return _layer_norm_weights[getGateIndex(g)];
+ }
+
+ inline const ICLTensor *get_layer_norm_bias(LayerNormGate g)
+ {
+ return _layer_norm_bias[getGateIndex(g)];
+ }
+
+ inline CLQLSTMLayerNormalizationKernel &get_layer_norm(LayerNormGate g)
+ {
+ return _layer_norms[getGateIndex(g)];
+ }
+
+ inline void configure_layer_norm(LayerNormGate g, const ICLTensor *in)
+ {
+ ARM_COMPUTE_ERROR_ON(!_has_layer_norm);
+
+ CLTensor *out = &get_layer_norm_output(g);
+ _memory_group.manage(out);
+ out->allocator()->init(*(in->info()));
+
+ get_layer_norm(g).configure(in, out, get_layer_norm_weight(g), get_layer_norm_bias(g));
+ }
+
+ inline static Status validate_layer_norm(const ITensorInfo &in, const ITensorInfo &weight, const ITensorInfo &bias)
+ {
+ // Output quantization scale will be different, but ignored here
+ // since it will be configured at configure() stage.
+ const TensorInfo out
+ {
+ in
+ };
+ return CLQLSTMLayerNormalizationKernel::validate(&in, &out, &weight, &bias);
+ }
// Temporary tensors
CLTensor _input_to_forget_weights_transposed{ nullptr };
@@ -368,6 +435,12 @@
CLTensor _mm_projection_res{ nullptr };
CLTensor _projection_outstage_res{ nullptr };
CLTensor _ones{ nullptr };
+ std::array<CLTensor, _layer_norm_count> _layer_norm_output{ {} };
+
+ inline CLTensor &get_layer_norm_output(LayerNormGate g)
+ {
+ return _layer_norm_output[getGateIndex(g)];
+ }
bool _is_prepared{ false };
bool _has_cifg{ false };
@@ -375,6 +448,7 @@
bool _has_projection{ false };
bool _has_projection_clipping{ false };
bool _has_peephole{ false };
+ bool _has_layer_norm{ false };
};
} // namespace arm_compute
#endif /* ARM_COMPUTE_CLQLSTMLAYER_H */
diff --git a/src/core/CL/kernels/CLQLSTMLayerNormalizationKernel.cpp b/src/core/CL/kernels/CLQLSTMLayerNormalizationKernel.cpp
index b9767e8..d9da3cb 100644
--- a/src/core/CL/kernels/CLQLSTMLayerNormalizationKernel.cpp
+++ b/src/core/CL/kernels/CLQLSTMLayerNormalizationKernel.cpp
@@ -31,11 +31,17 @@
{
namespace
{
+QuantizationInfo compute_output_qinfo()
+{
+ return QuantizationInfo(1.f / 4096);
+}
+
std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input);
// Output auto inizialitation if not yet initialized
auto_init_if_empty(*output, *input);
+ output->set_quantization_info(compute_output_qinfo());
const uint32_t temp_num_elems_processed_per_iteration = max_cl_vector_width / input->element_size();
/* If width is less then step, then make step same as width to avoid global size being step instead of actual width. */
@@ -48,7 +54,7 @@
return std::make_pair(Status{}, win);
}
-Status validate_arguments(const ITensorInfo *input, ITensorInfo *output, const ITensorInfo *weight, const ITensorInfo *bias)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *weight, const ITensorInfo *bias)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weight, bias, output);
@@ -129,7 +135,7 @@
configure(CLKernelLibrary::get().get_compile_context(), input, output, weight, bias);
}
-Status CLQLSTMLayerNormalizationKernel::validate(const ITensorInfo *input, ITensorInfo *output, const ITensorInfo *weight, const ITensorInfo *bias)
+Status CLQLSTMLayerNormalizationKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *weight, const ITensorInfo *bias)
{
ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, weight, bias));
ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
diff --git a/src/core/NEON/kernels/NEQLSTMLayerNormalizationKernel.cpp b/src/core/NEON/kernels/NEQLSTMLayerNormalizationKernel.cpp
index e966c6b..29ffee8 100644
--- a/src/core/NEON/kernels/NEQLSTMLayerNormalizationKernel.cpp
+++ b/src/core/NEON/kernels/NEQLSTMLayerNormalizationKernel.cpp
@@ -172,10 +172,7 @@
inline QuantizationInfo NEQLSTMLayerNormalizationKernel::compute_output_qinfo()
{
- const UniformQuantizationInfo iq_info = _input->info()->quantization_info().uniform();
- const UniformQuantizationInfo wq_info = _weight->info()->quantization_info().uniform();
- const float output_scale = (wq_info.scale * iq_info.scale) * 1024;
- return QuantizationInfo(output_scale);
+ return QuantizationInfo(1.f / 4096);
}
inline std::pair<int64_t, int64_t> NEQLSTMLayerNormalizationKernel::sum_qsymm16(const int16_t *input_ptr)
diff --git a/src/runtime/CL/functions/CLQLSTMLayer.cpp b/src/runtime/CL/functions/CLQLSTMLayer.cpp
index 88c5f77..d9b5c7c 100644
--- a/src/runtime/CL/functions/CLQLSTMLayer.cpp
+++ b/src/runtime/CL/functions/CLQLSTMLayer.cpp
@@ -92,9 +92,6 @@
ICLTensor *cell_state_out, ICLTensor *output_state_out,
const LSTMParams<ICLTensor> &lstm_params)
{
- ARM_COMPUTE_UNUSED(forget_gate_bias);
- ARM_COMPUTE_UNUSED(cell_bias);
- ARM_COMPUTE_UNUSED(output_gate_bias);
ARM_COMPUTE_ERROR_ON_NULLPTR(input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
forget_gate_bias, cell_bias, output_gate_bias, cell_state_in, output_state_in, cell_state_out, output_state_out);
@@ -125,6 +122,21 @@
_recurrent_to_output_weights = recurrent_to_output_weights;
_projection_weights = lstm_params.projection_weights();
+ // Layer normalization
+ _has_layer_norm = lstm_params.use_layer_norm();
+ if(_has_layer_norm)
+ {
+ set_layer_norm_weight(lstm_params.forget_layer_norm_weights(), LayerNormGate::Forget);
+ set_layer_norm_weight(lstm_params.cell_layer_norm_weights(), LayerNormGate::Cell);
+ set_layer_norm_weight(lstm_params.input_layer_norm_weights(), LayerNormGate::Input);
+ set_layer_norm_weight(lstm_params.output_layer_norm_weights(), LayerNormGate::Output);
+
+ set_layer_norm_bias(forget_gate_bias, LayerNormGate::Forget);
+ set_layer_norm_bias(cell_bias, LayerNormGate::Cell);
+ set_layer_norm_bias(lstm_params.input_gate_bias(), LayerNormGate::Input);
+ set_layer_norm_bias(output_gate_bias, LayerNormGate::Output);
+ }
+
_has_cifg = lstm_params.has_cifg_opt();
_has_projection = lstm_params.has_projection();
_has_peephole = lstm_params.has_peephole_opt();
@@ -218,14 +230,23 @@
_cell_to_forget_outstage_res.allocator()->allocate();
}
+ CLTensor *forget_activation_input = &_recurrent_to_forget_outstage_res;
+
+ if(_has_layer_norm)
+ {
+ configure_layer_norm(LayerNormGate::Forget, &_recurrent_to_forget_outstage_res);
+ _recurrent_to_forget_outstage_res.allocator()->allocate();
+ forget_activation_input = &get_layer_norm_output(LayerNormGate::Forget);
+ }
+
// Output quantization info of Sigmoid and Tanh activations
const QuantizationInfo sigmoid_tanh_outqinfo(1.f / 32768.f, 0);
const TensorInfo forget_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo);
_memory_group.manage(&_forget_gate);
_forget_gate.allocator()->init(forget_gate_info);
- _forget_gate_sigmoid.configure(compile_context, &_recurrent_to_forget_outstage_res, &_forget_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
- _recurrent_to_forget_outstage_res.allocator()->allocate();
+ _forget_gate_sigmoid.configure(compile_context, forget_activation_input, &_forget_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
+ forget_activation_input->allocator()->allocate();
// Modulation gate.
const TensorInfo cell_outstage_info(mm_out_info.tensor_shape(), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.cell_intermediate_scale(), 0));
@@ -245,11 +266,20 @@
ConvertPolicy::SATURATE);
_input_to_cell_outstage_res.allocator()->allocate();
+ CLTensor *cell_activation_input = &_recurrent_to_cell_outstage_res;
+
+ if(_has_layer_norm)
+ {
+ configure_layer_norm(LayerNormGate::Cell, &_recurrent_to_cell_outstage_res);
+ _recurrent_to_cell_outstage_res.allocator()->allocate();
+ cell_activation_input = &get_layer_norm_output(LayerNormGate::Cell);
+ }
+
const TensorInfo cell_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo);
_memory_group.manage(&_cell_gate);
_cell_gate.allocator()->init(cell_gate_info);
- _cell_gate_tanh.configure(compile_context, &_recurrent_to_cell_outstage_res, &_cell_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f));
- _recurrent_to_cell_outstage_res.allocator()->allocate();
+ _cell_gate_tanh.configure(compile_context, cell_activation_input, &_cell_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f));
+ cell_activation_input->allocator()->allocate();
// Input gate.
const TensorInfo input_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo);
@@ -293,8 +323,17 @@
_cell_to_input_outstage_res.allocator()->allocate();
}
- _input_gate_tanh.configure(compile_context, &_recurrent_to_input_outstage_res, &_input_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f));
- _recurrent_to_input_outstage_res.allocator()->allocate();
+ CLTensor *input_activation_input = &_recurrent_to_input_outstage_res;
+
+ if(_has_layer_norm)
+ {
+ configure_layer_norm(LayerNormGate::Input, &_recurrent_to_input_outstage_res);
+ _recurrent_to_input_outstage_res.allocator()->allocate();
+ input_activation_input = &get_layer_norm_output(LayerNormGate::Input);
+ }
+
+ _input_gate_tanh.configure(compile_context, input_activation_input, &_input_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f));
+ input_activation_input->allocator()->allocate();
}
// Cell.
// TODO(COMPMID-3396): Perform multiplication in the quantized domain in CLPixelWiseMultiplicationKernel
@@ -344,11 +383,20 @@
_mul_cell_to_output_res.allocator()->allocate();
}
+ CLTensor *output_activation_input = &_recurrent_to_output_outstage_res;
+
+ if(_has_layer_norm)
+ {
+ configure_layer_norm(LayerNormGate::Output, &_recurrent_to_output_outstage_res);
+ _recurrent_to_output_outstage_res.allocator()->allocate();
+ output_activation_input = &get_layer_norm_output(LayerNormGate::Output);
+ }
+
const TensorInfo output_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo);
_memory_group.manage(&_output_gate);
_output_gate.allocator()->init(output_gate_info);
- _output_gate_sigmoid.configure(compile_context, &_recurrent_to_output_outstage_res, &_output_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
- _recurrent_to_output_outstage_res.allocator()->allocate();
+ _output_gate_sigmoid.configure(compile_context, output_activation_input, &_output_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
+ output_activation_input->allocator()->allocate();
// Hidden.
_hidden_tanh.configure(compile_context, cell_state_out, &_input_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f));
@@ -525,6 +573,8 @@
gemmlowp_info.gemmlowp_max_bound = std::numeric_limits<int16_t>::max();
gemmlowp_info.output_data_type = DataType::QSYMM16;
+ const bool has_layer_norm = lstm_params.use_layer_norm();
+
// Forget gate.
const TensorInfo forget_outstage_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.forget_intermediate_scale(), 0));
const TensorInfo mm_out_info(TensorShape(num_units, batch_size), 1, DataType::S32);
@@ -547,6 +597,13 @@
ARM_COMPUTE_RETURN_ON_ERROR(CLSaturatedArithmeticOperationKernel::validate(ArithmeticOperation::ADD, &forget_outstage_info, &forget_outstage_info, &forget_outstage_info, ConvertPolicy::SATURATE));
}
+ if(has_layer_norm)
+ {
+ const ITensorInfo *w_info = lstm_params.forget_layer_norm_weights();
+ const ITensorInfo *b_info = forget_gate_bias;
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_layer_norm(forget_outstage_info, *w_info, *b_info));
+ }
+
// Output quantization info of Sigmoid and Tanh activations
const QuantizationInfo sigmoid_tanh_outqinfo(1.f / 32768.f, 0);
@@ -563,6 +620,13 @@
ARM_COMPUTE_RETURN_ON_ERROR(CLSaturatedArithmeticOperationKernel::validate(ArithmeticOperation::ADD, &cell_outstage_info, &cell_outstage_info, &cell_outstage_info, ConvertPolicy::SATURATE));
+ if(has_layer_norm)
+ {
+ const ITensorInfo *w_info = lstm_params.cell_layer_norm_weights();
+ const ITensorInfo *b_info = cell_bias;
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_layer_norm(cell_outstage_info, *w_info, *b_info));
+ }
+
const TensorInfo cell_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo);
ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_outstage_info, &cell_gate_info, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f)));
@@ -602,6 +666,13 @@
ARM_COMPUTE_RETURN_ON_ERROR(CLSaturatedArithmeticOperationKernel::validate(ArithmeticOperation::ADD, &input_outstage_info, &input_outstage_info, &input_outstage_info, ConvertPolicy::SATURATE));
}
+ if(has_layer_norm)
+ {
+ const ITensorInfo *w_info = lstm_params.input_layer_norm_weights();
+ const ITensorInfo *b_info = lstm_params.input_gate_bias();
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_layer_norm(cell_outstage_info, *w_info, *b_info));
+ }
+
ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&input_outstage_info, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f)));
}
// Cell.
@@ -634,6 +705,13 @@
ARM_COMPUTE_RETURN_ON_ERROR(CLSaturatedArithmeticOperationKernel::validate(ArithmeticOperation::ADD, &output_outstage_info, &output_outstage_info, &output_outstage_info, ConvertPolicy::SATURATE));
}
+ if(has_layer_norm)
+ {
+ const ITensorInfo *w_info = lstm_params.output_layer_norm_weights();
+ const ITensorInfo *b_info = output_gate_bias;
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_layer_norm(output_outstage_info, *w_info, *b_info));
+ }
+
const TensorInfo output_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo);
ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&output_outstage_info, &output_gate_info, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
@@ -715,6 +793,11 @@
CLScheduler::get().enqueue(_accumulate_cell_forget);
}
+ if(_has_layer_norm)
+ {
+ CLScheduler::get().enqueue(get_layer_norm(LayerNormGate::Forget));
+ }
+
_forget_gate_sigmoid.run();
// Modulation gate.
@@ -725,6 +808,11 @@
_recurrent_to_cell_outstage.run();
CLScheduler::get().enqueue(_accumulate_input_recurrent_modulation);
+ if(_has_layer_norm)
+ {
+ CLScheduler::get().enqueue(get_layer_norm(LayerNormGate::Cell));
+ }
+
_cell_gate_tanh.run();
// Input gate
@@ -747,6 +835,11 @@
CLScheduler::get().enqueue(_accumulate_cell_input);
}
+ if(_has_layer_norm)
+ {
+ CLScheduler::get().enqueue(get_layer_norm(LayerNormGate::Input));
+ }
+
_input_gate_tanh.run();
}
@@ -771,6 +864,11 @@
CLScheduler::get().enqueue(_accumulate_cell_to_output);
}
+ if(_has_layer_norm)
+ {
+ CLScheduler::get().enqueue(get_layer_norm(LayerNormGate::Output));
+ }
+
_output_gate_sigmoid.run();
// Hidden.