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>

2 files changed, 9 insertions, 6 deletions

diff --git a/src/core/CL/kernels/CLQLSTMLayerNormalizationKernel.cpp b/src/core/CL/kernels/CLQLSTMLayerNormalizationKernel.cpp
index b9767e8ec2..d9da3cb36e 100644
--- a/src/core/CL/kernels/CLQLSTMLayerNormalizationKernel.cpp
+++ b/src/core/CL/kernels/CLQLSTMLayerNormalizationKernel.cpp
@@ -31,11 +31,17 @@ namespace arm_compute
 {
 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 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 
     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 @@ void CLQLSTMLayerNormalizationKernel::configure(const ICLTensor *input, ICLTenso
     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 e966c6bdba..29ffee867b 100644
--- a/src/core/NEON/kernels/NEQLSTMLayerNormalizationKernel.cpp
+++ b/src/core/NEON/kernels/NEQLSTMLayerNormalizationKernel.cpp
@@ -172,10 +172,7 @@ void NEQLSTMLayerNormalizationKernel::run(const Window &window, const ThreadInfo
 
 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)