commit ee4833d0a16cd65ad4fa9de087ec92080716e996
author Sang-Hoon Park <sang-hoon.park@arm.com> Wed May 20 09:13:32 2020 +0100
committer Sang-Hoon Park <sang-hoon.park@arm.com> Wed May 20 10:30:25 2020 +0000
tree cb7fbd8d320ea28666d91942e76006e3cd3b02e4
parent 463f976813a99b1eda347c619c6b7baf098757b8

COMPMID-3497: Add checks for zero scale values to QLSTMLayer

Checks are added to validate() function to check both for
configuration and validation call.

Change-Id: I2ae9a92a5d90112f5b41befc4ce655ff9451d150
Signed-off-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3227
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

src/runtime/NEON/functions/NEQLSTMLayer.cpp

diff --git a/src/runtime/NEON/functions/NEQLSTMLayer.cpp b/src/runtime/NEON/functions/NEQLSTMLayer.cpp
index beb180f..083e3fd 100644
--- a/src/runtime/NEON/functions/NEQLSTMLayer.cpp
+++ b/src/runtime/NEON/functions/NEQLSTMLayer.cpp
@@ -644,6 +644,7 @@
     const bool has_layer_norm = lstm_params.use_layer_norm();
 
     // Forget gate.
+    ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.forget_intermediate_scale() == 0);
     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);
     const float      input_to_forget_scale = input_to_forget_weights->quantization_info().uniform().scale * qinput.scale / lstm_params.forget_intermediate_scale();
@@ -679,6 +680,7 @@
     ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&forget_outstage_info, &forget_gate_info, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
 
     // Modulation gate.
+    ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_intermediate_scale() == 0);
     const TensorInfo cell_outstage_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.cell_intermediate_scale(), 0));
     const float      input_to_cell_scale = input_to_cell_weights->quantization_info().uniform().scale * qinput.scale / lstm_params.cell_intermediate_scale();
     ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemmlowp_info, input, &input_weights_transposed, &eff_bias_info, input_to_cell_scale, &mm_out_info, &cell_outstage_info));
@@ -714,6 +716,7 @@
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(forget_gate_bias, lstm_params.input_gate_bias());
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(forget_gate_bias, lstm_params.input_gate_bias());
 
+        ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_intermediate_scale() == 0);
         const TensorInfo input_outstage_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.input_intermediate_scale(), 0));
         const float      input_to_input_scale = lstm_params.input_to_input_weights()->quantization_info().uniform().scale * qinput.scale / lstm_params.input_intermediate_scale();
         ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemmlowp_info, input, &input_weights_transposed, &eff_bias_info, input_to_input_scale, &mm_out_info, &input_outstage_info));
@@ -752,6 +755,7 @@
                                                                                                              quantized_cell_clip)));
     }
     // Output gate.
+    ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.output_intermediate_scale() == 0);
     const TensorInfo output_outstage_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.output_intermediate_scale(), 0));
     const float      input_to_output_scale = input_to_output_weights->quantization_info().uniform().scale * qinput.scale / lstm_params.output_intermediate_scale();
     ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemmlowp_info, input, &input_weights_transposed, &eff_bias_info, input_to_output_scale, &mm_out_info, &output_outstage_info));
@@ -787,6 +791,8 @@
     const TensorInfo hidden_mul_res(TensorShape(num_units, batch_size), 1, DataType::S32);
     const TensorInfo hidden_out_info(TensorShape(num_units, batch_size), 1, DataType::QASYMM8_SIGNED);
     ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplicationKernel::validate(&output_gate_info, &input_gate_info, &hidden_mul_res, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
+
+    ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.hidden_state_scale() == 0);
     const float hidden_state_scale = std::pow(2, -15) / lstm_params.hidden_state_scale() * std::pow(2, -15);
     ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(hidden_state_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift, /* ignore_epsilon */ true));
     gemmlowp_info.gemmlowp_offset = lstm_params.hidden_state_zero();
@@ -799,6 +805,7 @@
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(recurrent_to_forget_weights, lstm_params.projection_weights());
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(forget_gate_bias, lstm_params.projection_bias());
+        ARM_COMPUTE_RETURN_ERROR_ON(qoutput_state_in.scale == 0);
 
         const UniformQuantizationInfo qprojection      = lstm_params.projection_weights()->quantization_info().uniform();
         const float                   projection_scale = qprojection.scale * lstm_params.hidden_state_scale() / qoutput_state_in.scale;