COMPMID-2406: Create a new GEMMLowpQuantizeDownInt32ToInt16ScaleKernel for NEON

Change-Id: I3f3e247728fd6dafca066e41835f0ef9442d9b7a
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1379
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 63 insertions, 2 deletions

diff --git a/tests/validation/reference/GEMMLowp.cpp b/tests/validation/reference/GEMMLowp.cpp
index 9a7e409e8a..97d05327e7 100644
--- a/tests/validation/reference/GEMMLowp.cpp
+++ b/tests/validation/reference/GEMMLowp.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -95,6 +95,34 @@ void quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<T> *in,
         (*dst)[i] = static_cast<uint8_t>(std::max(0, std::min(255, result)));
     }
 }
+
+template <typename T>
+void quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<T> *in, const SimpleTensor<T> *bias, SimpleTensor<int16_t> *dst, int32_t result_fixedpoint_multiplier, int32_t result_shift,
+                                                      int32_t min, int32_t max)
+{
+    const int cols_in = in->shape().x();
+
+    for(int i = 0; i < in->num_elements(); ++i)
+    {
+        int32_t result = (*in)[i];
+
+        if(bias != nullptr)
+        {
+            result += (*bias)[i % cols_in];
+        }
+
+        // Fixed point multiplication
+        result = asymm_rounding_divide_by_pow2(asymm_int_mult(result, result_fixedpoint_multiplier), result_shift);
+
+        // Bounded ReLu
+        if(min != max)
+        {
+            result = std::max(min, std::min(max, result));
+        }
+
+        (*dst)[i] = static_cast<int16_t>(std::max(-32768, std::min(32767, result)));
+    }
+}
 } // namespace
 
 template <typename T_out, typename T_in>
@@ -201,10 +229,36 @@ SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(
     return dst;
 }
 
+template <typename T>
+SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<T> &in, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t min,
+                                                                                int32_t max)
+{
+    SimpleTensor<int16_t> dst(in.shape(), DataType::QSYMM16);
+
+    quantize_down_int32_to_int16_scale_by_fixedpoint<T>(&in, nullptr, &dst, result_fixedpoint_multiplier, result_shift, min, max);
+
+    return dst;
+}
+
+template <typename T>
+SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<T> &in, const SimpleTensor<T> &bias, int32_t result_fixedpoint_multiplier, int32_t result_shift,
+                                                                                int32_t min, int32_t max)
+{
+    SimpleTensor<int16_t> dst(in.shape(), DataType::QSYMM16);
+
+    quantize_down_int32_to_int16_scale_by_fixedpoint<T>(&in, &bias, &dst, result_fixedpoint_multiplier, result_shift, min, max);
+
+    return dst;
+}
+
 template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, int32_t result_fixedpoint_multiplier, int32_t result_shift,
                                                                                          int32_t result_offset_after_shift, int32_t min, int32_t max);
 template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b, int32_t result_fixedpoint_multiplier,
                                                                                          int32_t result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max);
+template SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, int32_t result_fixedpoint_multiplier, int32_t result_shift,
+                                                                                         int32_t min, int32_t max);
+template SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b, int32_t result_fixedpoint_multiplier,
+                                                                                         int32_t result_shift, int32_t min, int32_t max);
 template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<int32_t> &a, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min,
                                                                            int32_t max);
 template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b, int32_t result_offset, int32_t result_mult_int,
diff --git a/tests/validation/reference/GEMMLowp.h b/tests/validation/reference/GEMMLowp.h
index 4396155b96..5581f67652 100644
--- a/tests/validation/reference/GEMMLowp.h
+++ b/tests/validation/reference/GEMMLowp.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -58,6 +58,13 @@ SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(
 template <typename T>
 SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<T> &in, const SimpleTensor<T> &bias, int32_t result_fixedpoint_multiplier, int32_t result_shift,
                                                                                 int32_t result_offset_after_shift, int32_t min = 0, int32_t max = 0);
+
+template <typename T>
+SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<T> &in, int32_t result_fixedpoint_multiplier, int32_t result_shift,
+                                                                                int32_t min, int32_t max);
+template <typename T>
+SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<T> &in, const SimpleTensor<T> &bias, int32_t result_fixedpoint_multiplier,
+                                                                                int32_t result_shift, int32_t min, int32_t max);
 } // namespace reference
 } // namespace validation
 } // namespace test