commit f55cca5f17da72004108f92047d3177b7cdb1a76
author Pablo Marquez Tello <pablo.tello@arm.com> Wed Apr 06 14:31:25 2022 +0100
committer Pablo Marquez Tello <pablo.tello@arm.com> Mon Apr 25 10:20:20 2022 +0000
tree 08a1d83f51a017b03603f6c724f660fd1f084b2e
parent fa6877f94b12ec80235e55bcfe5a9b6fdc009cf0

Add LU_BOUNDED_RELU support for QSYMM16

Partially resolves MLCE-604

Change-Id: Id585ab19fe5cd8f61c07a0aae6faac6ba5545d6d
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7379
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

src/cpu/kernels/activation/generic/neon/qsymm16.cpp

diff --git a/src/cpu/kernels/activation/generic/neon/qsymm16.cpp b/src/cpu/kernels/activation/generic/neon/qsymm16.cpp
index 865b9f1..ba14745 100644
--- a/src/cpu/kernels/activation/generic/neon/qsymm16.cpp
+++ b/src/cpu/kernels/activation/generic/neon/qsymm16.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020-2021 Arm Limited.
+ * Copyright (c) 2020-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -101,6 +101,22 @@
                 // Re-quantize to new output space
                 tmp = vquantize_int16(tmp_dep, qi_out.scale);
             }
+
+            else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+            {
+                // De-quantize
+                const auto vin_deq = vdequantize_int16(vin, qi_in.scale);
+                // Perform activation
+                const float32x4x2_t tmp_dep =
+                {
+                    {
+                        wrapper::vmin(va_f32, wrapper::vmax(vb_f32, vin_deq.val[0])),
+                        wrapper::vmin(va_f32, wrapper::vmax(vb_f32, vin_deq.val[1]))
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = vquantize_int16(tmp_dep, qi_out.scale);
+            }
             else
             {
                 ARM_COMPUTE_ERROR("Unsupported activation function");
@@ -125,6 +141,12 @@
                 tmp_f       = a_f32 * std::tanh(b_f32 * tmp_f);
                 tmp         = quantize_qsymm16(tmp_f, qi_out);
             }
+            else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+            {
+                float tmp_f = dequantize_qsymm16(in, qi_in.scale);
+                tmp_f       = std::min<float>(a_f32, std::max<float>(b_f32, tmp_f));
+                tmp         = quantize_qsymm16(tmp_f, qi_out);
+            }
             else
             {
                 ARM_COMPUTE_ERROR("Unsupported activation function");