COMPMID-2767 [NEON] add support for QASYMM8_SIGNED to DequantizationLayer

Change-Id: If5b21d1e656b21baf39346c2fd74e8edc75007f5
Signed-off-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2429
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

6 files changed, 75 insertions, 19 deletions

diff --git a/arm_compute/core/NEON/NEAsymm.h b/arm_compute/core/NEON/NEAsymm.h
index 234d48882c..67adcef9b1 100644
--- a/arm_compute/core/NEON/NEAsymm.h
+++ b/arm_compute/core/NEON/NEAsymm.h
@@ -451,6 +451,30 @@ inline float32x4x4_t vdequantize(const uint8x16_t &qv, float scale, int32_t offs
     return vdequantized_input;
 }
 
+/** Dequantize a vector of 16 values stored as signed asymmetric.
+ *
+ * @param[in] qv     Input values to be dequantized.
+ * @param[in] scale  Quantization scaling factor.
+ * @param[in] offset Zero quantization offset.
+ *
+ * @return Dequantized values in a neon vector
+ */
+inline float32x4x4_t vdequantize(const int8x16_t &qv, float scale, int32_t offset)
+{
+    const int32x4_t     voffset = vdupq_n_s32(offset);
+    const float32x4_t   vscale  = vdupq_n_f32(scale);
+    const float32x4x4_t vdequantized_input =
+    {
+        {
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(qv)))), voffset)), vscale),
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(qv)))), voffset)), vscale),
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(qv)))), voffset)), vscale),
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(qv)))), voffset)), vscale),
+        }
+    };
+    return vdequantized_input;
+}
+
 /** Dequantize following symmetric quantization scheme a neon vector holding 16 quantized values.
  *
  * @param[in] qv     Input values to be dequantized.
diff --git a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp
index f555df3828..947f257bcb 100644
--- a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp
@@ -43,7 +43,7 @@ namespace
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QSYMM8_PER_CHANNEL, DataType::QSYMM8, DataType::QSYMM16);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL, DataType::QSYMM8, DataType::QSYMM16);
 
     if(output->tensor_shape().total_size() > 0)
     {
@@ -116,7 +116,7 @@ inline void store_result<float16_t>(float16_t *ptr, const float32x4x2_t &v)
 }
 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
 
-template <typename T>
+template <typename TOut, typename TIn>
 void run_dequantization_qasymm8(const ITensor *input, ITensor *output, const Window &window)
 {
     const UniformQuantizationInfo &qinfo  = input->info()->quantization_info().uniform();
@@ -137,8 +137,8 @@ void run_dequantization_qasymm8(const ITensor *input, ITensor *output, const Win
 
     execute_window_loop(win_collapsed, [&](const Coordinates &)
     {
-        const auto in_ptr  = reinterpret_cast<const uint8_t *>(in.ptr());
-        const auto out_ptr = reinterpret_cast<T *>(out.ptr());
+        const auto in_ptr  = reinterpret_cast<const TIn *>(in.ptr());
+        const auto out_ptr = reinterpret_cast<TOut *>(out.ptr());
 
         int x = window_start_x;
         for(; x <= (window_end_x - window_step_x); x += window_step_x)
@@ -146,14 +146,14 @@ void run_dequantization_qasymm8(const ITensor *input, ITensor *output, const Win
             const auto vin  = wrapper::vloadq(in_ptr + x);
             const auto vdeq = vdequantize(vin, scale, offset);
 
-            store_result<T>(reinterpret_cast<T *>(out_ptr + x), vdeq);
+            store_result(reinterpret_cast<TOut *>(out_ptr + x), vdeq);
         }
 
         // Compute left-over elements
         for(; x < window_end_x; ++x)
         {
-            uint8_t val    = *(in_ptr + x);
-            *(out_ptr + x) = static_cast<T>(dequantize(val, scale, offset));
+            auto val       = *(in_ptr + x);
+            *(out_ptr + x) = static_cast<TOut>(Qasymm8QuantizationHelper<TIn>::dequantize(val, qinfo));
         }
     },
     in, out);
@@ -340,7 +340,10 @@ void run_dequantization_core(const ITensor *input, ITensor *output, const Window
     switch(input->info()->data_type())
     {
         case DataType::QASYMM8:
-            run_dequantization_qasymm8<T>(input, output, window);
+            run_dequantization_qasymm8<T, uint8_t>(input, output, window);
+            break;
+        case DataType::QASYMM8_SIGNED:
+            run_dequantization_qasymm8<T, int8_t>(input, output, window);
             break;
         case DataType::QSYMM8_PER_CHANNEL:
             input->info()->data_layout() == DataLayout::NHWC ? run_dequantization_qsymm8_per_channel_nhwc<T>(input, output, window) : run_dequantization_qsymm8_per_channel_nchw<T>(input, output, window);
diff --git a/tests/AssetsLibrary.h b/tests/AssetsLibrary.h
index 29d9cc6d49..c4892748f4 100644
--- a/tests/AssetsLibrary.h
+++ b/tests/AssetsLibrary.h
@@ -846,6 +846,7 @@ void AssetsLibrary::fill_tensor_uniform(T &&tensor, std::random_device::result_t
         }
         case DataType::S8:
         case DataType::QSYMM8:
+        case DataType::QASYMM8_SIGNED:
         {
             ARM_COMPUTE_ERROR_ON(!(std::is_same<int8_t, D>::value));
             std::uniform_int_distribution<int8_t> distribution_s8(low, high);
diff --git a/tests/validation/NEON/DequantizationLayer.cpp b/tests/validation/NEON/DequantizationLayer.cpp
index 0dce76a933..4389419d73 100644
--- a/tests/validation/NEON/DequantizationLayer.cpp
+++ b/tests/validation/NEON/DequantizationLayer.cpp
@@ -55,6 +55,14 @@ const auto dataset_quant_f32 = combine(combine(combine(datasets::SmallShapes(),
 const auto dataset_quant_f16 = combine(combine(combine(datasets::SmallShapes(), datasets::QuantizedTypes()),
                                                framework::dataset::make("DataType", DataType::F16)),
                                        framework::dataset::make("DataLayout", { DataLayout::NCHW }));
+const auto dataset_quant_asymm_signed_f32 = combine(combine(combine(datasets::SmallShapes(),
+                                                                    framework::dataset::make("QuantizedTypes", { DataType::QASYMM8_SIGNED })),
+                                                            framework::dataset::make("DataType", DataType::F32)),
+                                                    framework::dataset::make("DataLayout", { DataLayout::NCHW }));
+const auto dataset_quant_asymm_signed_f16 = combine(combine(combine(datasets::SmallShapes(),
+                                                                    framework::dataset::make("QuantizedTypes", { DataType::QASYMM8_SIGNED })),
+                                                            framework::dataset::make("DataType", DataType::F16)),
+                                                    framework::dataset::make("DataLayout", { DataLayout::NCHW }));
 const auto dataset_quant_per_channel_f32 = combine(combine(combine(datasets::SmallShapes(), datasets::QuantizedPerChannelTypes()),
                                                            framework::dataset::make("DataType", DataType::F32)),
                                                    framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }));
@@ -73,6 +81,12 @@ const auto dataset_quant_per_channel_nightly_f32 = combine(combine(combine(datas
 const auto dataset_quant_per_channel_nightly_f16 = combine(combine(combine(datasets::LargeShapes(), datasets::QuantizedPerChannelTypes()),
                                                                    framework::dataset::make("DataType", DataType::F16)),
                                                            framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }));
+
+const auto dataset_precommit_f16 = concat(concat(dataset_quant_f16, dataset_quant_per_channel_f16), dataset_quant_asymm_signed_f16);
+const auto dataset_precommit_f32 = concat(concat(dataset_quant_f32, dataset_quant_per_channel_f32), dataset_quant_asymm_signed_f32);
+const auto dataset_nightly_f16   = concat(dataset_quant_f16, dataset_quant_per_channel_f16);
+const auto dataset_nightly_f32   = concat(dataset_quant_f32, dataset_quant_per_channel_f32);
+
 } // namespace
 
 TEST_SUITE(NEON)
@@ -86,14 +100,16 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(
                                                 TensorInfo(TensorShape(16U, 16U, 2U, 5U), 1, DataType::QASYMM8),   // Missmatching shapes
                                                 TensorInfo(TensorShape(17U, 16U, 16U, 5U), 1, DataType::QASYMM8),  // Valid
                                                 TensorInfo(TensorShape(16U, 16U, 16U, 5U), 1, DataType::QASYMM8),  // Valid
+                                                TensorInfo(TensorShape(16U, 16U, 16U, 5U), 1, DataType::QASYMM8_SIGNED),  // Valid
         }),
         framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(16U, 16U, 16U, 5U), 1, DataType::F32),
                                                 TensorInfo(TensorShape(16U, 16U, 16U, 5U), 1, DataType::U8),
                                                 TensorInfo(TensorShape(16U, 16U, 16U, 5U), 1, DataType::F32),
                                                 TensorInfo(TensorShape(17U, 16U, 16U, 5U), 1, DataType::F32),
                                                 TensorInfo(TensorShape(16U, 16U, 16U, 5U), 1, DataType::F32),
+                                                TensorInfo(TensorShape(16U, 16U, 16U, 5U), 1, DataType::F32),
         })),
-        framework::dataset::make("Expected", { false, false, false, true, true})),
+        framework::dataset::make("Expected", { false, false, false, true, true, true })),
         input_info, output_info, expected)
 {
     ARM_COMPUTE_EXPECT(bool(NEDequantizationLayer::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false))) == expected, framework::LogLevel::ERRORS);
@@ -132,12 +148,12 @@ using NEDequantizationLayerFixture = DequantizationValidationFixture<Tensor, Acc
 
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEDequantizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, concat(dataset_quant_f16, dataset_quant_per_channel_f16))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEDequantizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, dataset_precommit_f16)
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, NEDequantizationLayerFixture<half>, framework::DatasetMode::NIGHTLY, concat(dataset_quant_nightly_f16, dataset_quant_per_channel_nightly_f16))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEDequantizationLayerFixture<half>, framework::DatasetMode::NIGHTLY, dataset_nightly_f16)
 {
     // Validate output
     validate(Accessor(_target), _reference);
@@ -146,12 +162,12 @@ TEST_SUITE_END() // FP16
 #endif           /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
 
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEDequantizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, concat(dataset_quant_f32, dataset_quant_per_channel_f32))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEDequantizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, dataset_precommit_f32)
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, NEDequantizationLayerFixture<float>, framework::DatasetMode::NIGHTLY, concat(dataset_quant_nightly_f32, dataset_quant_per_channel_nightly_f32))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEDequantizationLayerFixture<float>, framework::DatasetMode::NIGHTLY, dataset_nightly_f32)
 {
     // Validate output
     validate(Accessor(_target), _reference);
diff --git a/tests/validation/fixtures/DequantizationLayerFixture.h b/tests/validation/fixtures/DequantizationLayerFixture.h
index f44f8658c2..3699613a39 100644
--- a/tests/validation/fixtures/DequantizationLayerFixture.h
+++ b/tests/validation/fixtures/DequantizationLayerFixture.h
@@ -106,6 +106,7 @@ protected:
                 fill(src);
                 return reference::dequantization_layer<T>(src);
             }
+            case DataType::QASYMM8_SIGNED:
             case DataType::QSYMM8_PER_CHANNEL:
             case DataType::QSYMM8:
             {
@@ -149,6 +150,8 @@ protected:
             }
             case DataType::QASYMM8:
                 return QuantizationInfo(1.f / distribution_scale_q8(gen), distribution_offset_q8(gen));
+            case DataType::QASYMM8_SIGNED:
+                return QuantizationInfo(1.f / distribution_scale_q8(gen), -distribution_offset_q8(gen));
             default:
                 ARM_COMPUTE_ERROR("Unsupported data type");
         }
diff --git a/tests/validation/reference/DequantizationLayer.cpp b/tests/validation/reference/DequantizationLayer.cpp
index 16f25c4427..7dd36402b3 100644
--- a/tests/validation/reference/DequantizationLayer.cpp
+++ b/tests/validation/reference/DequantizationLayer.cpp
@@ -36,18 +36,27 @@ namespace reference
 namespace
 {
 template <typename TOut>
-TOut dequantize(int8_t val, const UniformQuantizationInfo qinfo)
+TOut dequantize(int8_t val, const UniformQuantizationInfo qinfo, DataType dt)
 {
-    return static_cast<TOut>(dequantize_qsymm8(val, qinfo));
+    if(dt == DataType::QSYMM8 || dt == DataType::QSYMM8_PER_CHANNEL)
+    {
+        return static_cast<TOut>(dequantize_qsymm8(val, qinfo));
+    }
+    else
+    {
+        return static_cast<TOut>(dequantize_qasymm8_signed(val, qinfo));
+    }
 }
 template <typename TOut>
-TOut dequantize(uint8_t val, const UniformQuantizationInfo qinfo)
+TOut dequantize(uint8_t val, const UniformQuantizationInfo qinfo, DataType dt)
 {
+    ARM_COMPUTE_UNUSED(dt);
     return static_cast<TOut>(dequantize_qasymm8(val, qinfo));
 }
 template <typename TOut>
-TOut dequantize(int16_t val, const UniformQuantizationInfo qinfo)
+TOut dequantize(int16_t val, const UniformQuantizationInfo qinfo, DataType dt)
 {
+    ARM_COMPUTE_UNUSED(dt);
     return static_cast<TOut>(dequantize_qsymm16(val, qinfo));
 }
 } // namespace
@@ -77,7 +86,7 @@ SimpleTensor<TOut> dequantization_layer(const SimpleTensor<TIn> &src)
                 // Dequantize slice
                 for(int s = 0; s < WH; ++s)
                 {
-                    dst[idx + s] = dequantize<TOut>(static_cast<TIn>(src[idx + s]), channel_qinfo);
+                    dst[idx + s] = dequantize<TOut>(static_cast<TIn>(src[idx + s]), channel_qinfo, src_data_type);
                 }
             }
         }
@@ -89,7 +98,7 @@ SimpleTensor<TOut> dequantization_layer(const SimpleTensor<TIn> &src)
 
         for(int i = 0; i < src.num_elements(); ++i)
         {
-            dst[i] = static_cast<TOut>(dequantize<TOut>(static_cast<TIn>(src[i]), quantization_info));
+            dst[i] = static_cast<TOut>(dequantize<TOut>(static_cast<TIn>(src[i]), quantization_info, src_data_type));
         }
     }