COMPMID-706 - Add GEMMLowp output stage for scaling by a fixed point number

DoD:
- Implement NEON kernel for quantizing down the gemmlowp result. The
  result should be scaled by a fixedpoint number
- Implement OpenCL kernel for quantizing down the gemmlowp result. The
  result should be scaled by a fixedpoint number
- Add test for validating the result

Required for:
- Integration of GEMMLowp in Android NN
- Convolution quantized
- Fully connected quantized

Change-Id: Ia963d25d695471e963961fb49a5600e78374ac4f
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110981
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

2 files changed, 71 insertions, 1 deletions

diff --git a/tests/validation/CPP/GEMMLowp.cpp b/tests/validation/CPP/GEMMLowp.cpp
index 35b8a6486e..92878947c8 100644
--- a/tests/validation/CPP/GEMMLowp.cpp
+++ b/tests/validation/CPP/GEMMLowp.cpp
@@ -24,6 +24,9 @@
 #include "GEMMLowp.h"
 
 #include "arm_compute/core/Types.h"
+#include "tests/validation/CPP/UtilsQuantizedAsymm.h"
+
+#include <limits>
 
 namespace arm_compute
 {
@@ -43,13 +46,15 @@ void quantize_down_int32_to_uint8_scale(const SimpleTensor<T> *in, const SimpleT
 
     for(int i = 0; i < in->num_elements(); ++i)
     {
-        int32_t result = ((*in)[i] + result_offset) * result_mult_int;
+        int32_t result = ((*in)[i] + result_offset);
 
         if(bias != nullptr)
         {
             result += (*bias)[i % cols_in];
         }
 
+        result *= result_mult_int;
+
         result >>= result_shift;
 
         // Bounded ReLu
@@ -61,6 +66,35 @@ void quantize_down_int32_to_uint8_scale(const SimpleTensor<T> *in, const SimpleT
         (*dst)[i] = static_cast<uint8_t>(std::max(0, std::min(255, result)));
     }
 }
+
+template <typename T>
+void quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<T> *in, const SimpleTensor<T> *bias, SimpleTensor<uint8_t> *dst, int32_t result_fixedpoint_multiplier, int32_t result_shift,
+                                                      int32_t result_offset_after_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);
+        result += result_offset_after_shift;
+
+        // Bounded ReLu
+        if(min != max)
+        {
+            result = std::max(min, std::min(max, result));
+        }
+
+        (*dst)[i] = static_cast<uint8_t>(std::max(0, std::min(255, result)));
+    }
+}
 } // namespace
 
 template <typename T_out, typename T_in>
@@ -133,6 +167,33 @@ SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTe
     return dst;
 }
 
+template <typename T>
+SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<T> &in, int32_t result_fixedpoint_multiplier, int32_t result_shift,
+                                                                                int32_t result_offset_after_shift, int32_t min,
+                                                                                int32_t max)
+{
+    SimpleTensor<uint8_t> dst(in.shape(), DataType::QASYMM8);
+
+    quantize_down_int32_to_uint8_scale_by_fixedpoint<T>(&in, nullptr, &dst, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max);
+
+    return dst;
+}
+
+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, int32_t max)
+{
+    SimpleTensor<uint8_t> dst(in.shape(), DataType::QASYMM8);
+
+    quantize_down_int32_to_uint8_scale_by_fixedpoint<T>(&in, &bias, &dst, result_fixedpoint_multiplier, result_shift, result_offset_after_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<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/CPP/GEMMLowp.h b/tests/validation/CPP/GEMMLowp.h
index 6c72b56e7a..a3d0bebe3f 100644
--- a/tests/validation/CPP/GEMMLowp.h
+++ b/tests/validation/CPP/GEMMLowp.h
@@ -49,6 +49,15 @@ SimpleTensor<T1> gemmlowp(const SimpleTensor<T2> &a, const SimpleTensor<T2> &b);
 template <typename T>
 SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<T> &in, const SimpleTensor<T> &bias, int32_t result_offset, int32_t result_mult_int, int32_t result_shift,
                                                                   int32_t min = 0, int32_t max = 0);
+
+template <typename T>
+SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<T> &in, 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<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);
 } // namespace reference
 } // namespace validation
 } // namespace test