COMPMID-1451: Perform CLOutputStage using floats.

Change-Id: Ic8312a5b6790aa7cd4468d42f08d557ad40e9441
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/154570
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

2 files changed, 18 insertions, 10 deletions

diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
index f41a12ae48..61180fd5d3 100644
--- a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
@@ -284,17 +284,14 @@ void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *
     {
         const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info();
 
-        float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale;
-        int   output_multiplier, output_shift;
-        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-
         if(!_skip_col2im)
         {
             _memory_group.manage(&_tmp_output);
             gemm_output_staged_to_use = &_tmp_output;
         }
 
-        _gemmlowp_output_stage.configure(gemm_output_to_use, biases, gemm_output_staged_to_use, output_multiplier, output_shift, output_quant_info.offset);
+        float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale;
+        _gemmlowp_output_stage.configure(gemm_output_to_use, biases, gemm_output_staged_to_use, multiplier, output_quant_info.offset);
     }
 
     if(!_skip_col2im)
@@ -448,17 +445,12 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
 
     if(is_quantized)
     {
-        float multiplier = input->quantization_info().scale * weights_to_use->quantization_info().scale / output->quantization_info().scale;
-        int   output_multiplier, output_shift;
-        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-
         if(!skip_col2im)
         {
             tmp_info = TensorInfo(gemm_output_to_use->tensor_shape(), 1, DataType::QASYMM8);
             tmp_info.set_quantization_info(output->quantization_info());
             gemm_output_staged_to_use = &tmp_info;
         }
-
         // Validate output stage for quantized case
         CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(gemm_output_to_use, biases, gemm_output_staged_to_use);
     }
diff --git a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
index b18d23fac9..f5dc655776 100644
--- a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
@@ -25,6 +25,7 @@
 
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h"
 #include "support/ToolchainSupport.h"
 
@@ -56,4 +57,19 @@ Status CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(const ITens
 {
     return CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(input, bias, output, min, max, output_3d_depth);
 }
+
+void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloat::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
+                                                               float multiplier, int offset,
+                                                               int min, int max, unsigned int output_3d_depth)
+{
+    auto k = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel>();
+    k->configure(input, bias, output, multiplier, offset, min, max, output_3d_depth);
+    _kernel = std::move(k);
+}
+
+Status CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloat::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output,
+                                                                int min, int max, unsigned int output_3d_depth)
+{
+    return CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::validate(input, bias, output, min, max, output_3d_depth);
+}
 } // namespace arm_compute
\ No newline at end of file