COMPMID-3236: Implement CLQLSTMLayer

COMPMID-3081: Extend CLQLSTMLayer with enhancements

Change-Id: Idb7aaaacdba957e5ad61e94edeab2e898730a109
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3057
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 17 insertions, 8 deletions

diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp
index dd4c55c2d8..ad675df7ea 100644
--- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp
@@ -56,7 +56,14 @@ Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1,
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
+    if(input0->data_type() == DataType::QASYMM8)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
+    }
+    else
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::QSYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL);
+    }
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
 
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
index 00cef56db7..066307c4b2 100644
--- a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
@@ -35,8 +35,6 @@
 
 #include "support/StringSupport.h"
 
-using namespace arm_compute;
-
 namespace arm_compute
 {
 namespace
@@ -98,9 +96,6 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 }
 } // namespace
 
-class Coordinates;
-} // namespace arm_compute
-
 CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel()
     : _input(nullptr), _bias(nullptr), _output(nullptr)
 {
@@ -184,3 +179,4 @@ void CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run(const Window
     }
     while(collapsed.slide_window_slice_3D(slice));
 }
+} // namespace arm_compute
diff --git a/src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp b/src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp
index e81ab2ffba..9fa253a55a 100644
--- a/src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp
@@ -36,7 +36,7 @@ namespace
 Status validate_arguments_matrix_a_reduction(const ITensorInfo *input, const ITensorInfo *output)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8);
 
     if(output->total_size() > 0)
     {
@@ -49,7 +49,7 @@ Status validate_arguments_matrix_a_reduction(const ITensorInfo *input, const ITe
 Status validate_arguments_matrix_b_reduction(const ITensorInfo *input, const ITensorInfo *output)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8);
 
     if(output->total_size() > 0)
     {
@@ -63,6 +63,9 @@ std::pair<Status, Window> validate_and_configure_window_matrix_b_reduction(ITens
 {
     constexpr unsigned int num_elems_processed_per_iteration = 16;
 
+    // Output auto initialization if not yet initialized
+    auto_init_if_empty(*output, TensorShape(input->dimension(0)), 1, DataType::S32);
+
     // Configure kernel window
     Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
 
@@ -94,6 +97,9 @@ void CLGEMMLowpMatrixAReductionKernel::configure(CLCompileContext &compile_conte
     ARM_COMPUTE_ERROR_ON_NULLPTR(mtx_a, vector_sum_row);
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_matrix_a_reduction(mtx_a->info(), vector_sum_row->info()));
 
+    // Output auto initialization if not yet initialized
+    auto_init_if_empty(*vector_sum_row->info(), TensorShape(mtx_a->info()->dimension(1)), 1, DataType::S32);
+
     _input  = mtx_a;
     _output = vector_sum_row;