COMPMID-2609: Enable quantization with multiplier greater than 1 on OpenCL

Change-Id: I050f1f84e214e61f7cbb0197a672b68a4940edae
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2158
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>

4 files changed, 34 insertions, 19 deletions

diff --git a/tests/validation/reference/Convolution3d.h b/tests/validation/reference/Convolution3d.h
index 23918a4055..6ac5df93b3 100644
--- a/tests/validation/reference/Convolution3d.h
+++ b/tests/validation/reference/Convolution3d.h
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef __ARM_COMPUTE_TEST_VALIDATION_CONVOLUTION_H__
-#define __ARM_COMPUTE_TEST_VALIDATION_CONVOLUTION_H__
+#ifndef ARM_COMPUTE_TEST_VALIDATION_CONVOLUTION_H
+#define ARM_COMPUTE_TEST_VALIDATION_CONVOLUTION_H
 
 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
 #include "tests/validation/Helpers.h"
@@ -133,7 +133,7 @@ inline void convolution3d(const SimpleTensor<T> &in, const SimpleTensor<TW> &wei
     int         output_multiplier = 0;
     int         output_shift      = 0;
     const float multiplier        = input_scale * weights_scale / output_scale;
-    arm_compute::quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+    arm_compute::quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
 
     const int half_width_weights_start  = width_weights / 2;
     const int half_width_weights_end    = ((width_weights % 2) == 0) ? (half_width_weights_start - 1) : half_width_weights_start;
@@ -171,9 +171,8 @@ inline void convolution3d(const SimpleTensor<T> &in, const SimpleTensor<TW> &wei
     // Accumulate the bias
     acc += (*b_ptr);
 
-    acc = validation::asymm_rounding_divide_by_pow2(validation::asymm_int_mult(acc, output_multiplier), output_shift);
-    acc += output_offset;
-    acc = utility::clamp<int32_t>(acc, 0, 255);
+    // Quantize down
+    acc = validation::quantize_down_scale_by_fixedpoint(acc, output_multiplier, output_shift, output_offset, 0, 255);
 
     // Store the result
     *out_ptr = acc;
@@ -182,4 +181,4 @@ inline void convolution3d(const SimpleTensor<T> &in, const SimpleTensor<TW> &wei
 } // namespace convolution_3d
 } // namespace test
 } // namespace arm_compute
-#endif /*__ARM_COMPUTE_TEST_VALIDATION_CONVOLUTION_H__ */
+#endif /* ARM_COMPUTE_TEST_VALIDATION_CONVOLUTION_H */
diff --git a/tests/validation/reference/DepthwiseConvolutionLayer.cpp b/tests/validation/reference/DepthwiseConvolutionLayer.cpp
index 608093d381..0c7e92b8d0 100644
--- a/tests/validation/reference/DepthwiseConvolutionLayer.cpp
+++ b/tests/validation/reference/DepthwiseConvolutionLayer.cpp
@@ -197,7 +197,7 @@ SimpleTensor<T> depthwise_convolution_quantized(const SimpleTensor<T> &src, cons
                 int         output_shift      = 0;
                 const float weights_scale     = (is_quantized_per_channel) ? weights_scale_vec[out_z] : weights_scale_vec[0];
                 const float multiplier        = input_scale * weights_scale / output_scale;
-                arm_compute::quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+                arm_compute::quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
 
                 for(int y = minimum_y; y <= minimum_y + maximum_y; y += conv_info.stride().second)
                 {
@@ -220,9 +220,8 @@ SimpleTensor<T> depthwise_convolution_quantized(const SimpleTensor<T> &src, cons
                             }
                         }
                         val += bias_val;
-                        val = asymm_rounding_divide_by_pow2(asymm_int_mult(val, output_multiplier), output_shift);
-                        val += output_offset;
-                        val = utility::clamp<int32_t>(val, 0, 255);
+                        // Quantize down
+                        val = quantize_down_scale_by_fixedpoint(val, output_multiplier, output_shift, output_offset, 0, 255);
 
                         // Store the result
                         dst[out_pos++] = val;
diff --git a/tests/validation/reference/FullyConnectedLayer.cpp b/tests/validation/reference/FullyConnectedLayer.cpp
index cd84b9cfd1..261c6453b9 100644
--- a/tests/validation/reference/FullyConnectedLayer.cpp
+++ b/tests/validation/reference/FullyConnectedLayer.cpp
@@ -81,7 +81,7 @@ void vector_matrix_multiply(const SimpleTensor<T> &src, const SimpleTensor<T> &w
     int         output_multiplier = 0;
     int         output_shift      = 0;
     const float multiplier        = input_scale * weights_scale / output_scale;
-    arm_compute::quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+    arm_compute::quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
 
     for(int y = 0; y < rows_weights; ++y)
     {
@@ -96,9 +96,8 @@ void vector_matrix_multiply(const SimpleTensor<T> &src, const SimpleTensor<T> &w
         // Accumulate the bias
         acc += bias_ptr[y];
 
-        acc = asymm_rounding_divide_by_pow2(asymm_int_mult(acc, output_multiplier), output_shift);
-        acc += output_offset;
-        acc = utility::clamp<int32_t>(acc, 0, 255);
+        // Quantize down
+        acc = quantize_down_scale_by_fixedpoint(acc, output_multiplier, output_shift, output_offset, 0, 255);
 
         // Store the result
         dst_ptr[y] = static_cast<T>(acc);
diff --git a/tests/validation/reference/UtilsQuantizedAsymm.h b/tests/validation/reference/UtilsQuantizedAsymm.h
index b7b69d588a..444696c93a 100644
--- a/tests/validation/reference/UtilsQuantizedAsymm.h
+++ b/tests/validation/reference/UtilsQuantizedAsymm.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef __ARM_COMPUTE_TEST_VALIDATION_UTILS_QUANTIZED_ASYMM_H__
-#define __ARM_COMPUTE_TEST_VALIDATION_UTILS_QUANTIZED_ASYMM_H__
+#ifndef ARM_COMPUTE_TEST_VALIDATION_UTILS_QUANTIZED_ASYMM_H
+#define ARM_COMPUTE_TEST_VALIDATION_UTILS_QUANTIZED_ASYMM_H
 
 #include <cstdint>
 
@@ -51,7 +51,25 @@ inline int32_t asymm_int_mult(int32_t a, int32_t b)
     int32_t ab_x2_high32 = static_cast<int32_t>((ab_64 + nudge) / (1ll << 31));
     return overflow ? std::numeric_limits<int32_t>::max() : ab_x2_high32;
 }
+
+/** Quantize down the input value in range [min, max]. */
+inline int32_t quantize_down_scale_by_fixedpoint(int32_t val, int32_t result_mult_int, int32_t result_shift,
+                                                 int32_t result_offset_after_shift, int32_t min, int32_t max)
+{
+    int32_t res = 0;
+    if(result_shift < 0)
+    {
+        res = asymm_int_mult(val * (1 << (-result_shift)), result_mult_int);
+    }
+    else
+    {
+        res = asymm_rounding_divide_by_pow2(asymm_int_mult(val, result_mult_int), result_shift);
+    }
+    res += result_offset_after_shift;
+    res = utility::clamp<int32_t>(res, min, max);
+    return res;
+}
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
-#endif /* __ARM_COMPUTE_TEST_VALIDATION_UTILS_QUANTIZED_ASYMM_H__ */
+#endif /* ARM_COMPUTE_TEST_VALIDATION_UTILS_QUANTIZED_ASYMM_H */