commit 40b441905760846e9fdaca283a4a4de038a6ef0d
author Viet-Hoa Do <viet-hoa.do@arm.com> Thu Sep 22 10:24:23 2022 +0100
committer Viet-Hoa Do <viet-hoa.do@arm.com> Mon Oct 03 08:57:23 2022 +0000
tree 38a4f6b5122bfaf44a2a33e90b331a2e1a30b113
parent ff81de5a9a0f6b9331c3b112cc2aed552f0482a9

Optimize CPU add layer on quantized data

* Use fixed-point arithmetic where possible.
* Various optimization for the FP32-based implementation.
  This implementation is kept as the fall-back solution
  in case of unrealistic quantization parameters that exceed
  the range of fixed-point solution.

Resolves: COMPMID-5458
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I221d2d3801ecaae4fe0b7cf6ae8ef00ca3743665
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8317
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

tests/validation/NEON/ArithmeticAddition.cpp

diff --git a/tests/validation/NEON/ArithmeticAddition.cpp b/tests/validation/NEON/ArithmeticAddition.cpp
index f94e329..322cd7b 100644
--- a/tests/validation/NEON/ArithmeticAddition.cpp
+++ b/tests/validation/NEON/ArithmeticAddition.cpp
@@ -89,7 +89,7 @@
 }
 
 DATA_TEST_CASE(KernelSelection, framework::DatasetMode::ALL, concat(concat(
-               combine(combine(framework::dataset::make("CpuExt", std::string("NEON")),
+                combine(combine(combine(framework::dataset::make("CpuExt", std::string("NEON")),
                        framework::dataset::make("DataType", { DataType::F32,
                                                               DataType::F16,
                                                               DataType::U8,
@@ -100,21 +100,24 @@
                                                               DataType::QSYMM16
                                                             })),
                         framework::dataset::make("CanInterpretAs1D", {true, false})),
-                combine(combine(framework::dataset::make("CpuExt", std::string("SVE")),
+                        framework::dataset::make("CanUseFixedpoint", {true, false})),
+                combine(combine(combine(framework::dataset::make("CpuExt", std::string("SVE")),
                         framework::dataset::make("DataType", { DataType::F32,
                                                                DataType::F16,
                                                                DataType::U8,
                                                                DataType::S16,
                                                                DataType::S32
                                                              })),
-                        framework::dataset::make("CanInterpretAs1D", {true, false}))),
-                combine(combine(framework::dataset::make("CpuExt", std::string("SVE2")),
+                        framework::dataset::make("CanInterpretAs1D", {true, false})),
+                        framework::dataset::make("CanUseFixedpoint", {true, false}))),
+                combine(combine(combine(framework::dataset::make("CpuExt", std::string("SVE2")),
                         framework::dataset::make("DataType", { DataType::QASYMM8,
                                                                DataType::QASYMM8_SIGNED,
                                                                DataType::QSYMM16
                                                              })),
-                        framework::dataset::make("CanInterpretAs1D", {false}))),
-               cpu_ext, data_type, can_interpret_inputs_as_1d_array)
+                        framework::dataset::make("CanInterpretAs1D", {false})),
+                        framework::dataset::make("CanUseFixedpoint", {true, false}))),
+               cpu_ext, data_type, can_interpret_inputs_as_1d_array, can_use_fixedpoint)
 {
     using namespace cpu::kernels;
 
@@ -124,18 +127,23 @@
     cpu_isa.sve2 = (cpu_ext == "SVE2");
     cpu_isa.fp16 = (data_type == DataType::F16);
 
-    const auto *selected_impl = CpuAddKernel::get_implementation(CpuAddKernelDataTypeISASelectorData{data_type, cpu_isa, can_interpret_inputs_as_1d_array}, cpu::KernelSelectionType::Preferred);
+    const auto *selected_impl = CpuAddKernel::get_implementation(CpuAddKernelDataTypeISASelectorData{data_type, cpu_isa, can_interpret_inputs_as_1d_array, can_use_fixedpoint}, cpu::KernelSelectionType::Preferred);
 
     ARM_COMPUTE_ERROR_ON_NULLPTR(selected_impl);
 
     bool float_or_integer = (data_type == DataType::F32 || data_type == DataType::F16 || data_type == DataType::U8 ||
                              data_type == DataType::S16 || data_type == DataType::S32);
+    bool qasymm8_any = (data_type == DataType::QASYMM8 || data_type == DataType::QASYMM8_SIGNED);
 
     std::string expected;
     if(can_interpret_inputs_as_1d_array && float_or_integer)
     {
         expected = "neon_" + cpu_impl_dt(data_type) + "_add_as_1d_array";
     }
+    else if(qasymm8_any && can_use_fixedpoint)
+    {
+        expected = "neon_" + cpu_impl_dt(data_type) + "_add_fixedpoint";
+    }
     else
     {
         expected = lower_string(cpu_ext) + "_" + cpu_impl_dt(data_type) + "_add";