COMPMID-665 - NEON: Add QASYMM8 in place Activation layer

- Added min and max arguments for QuantizeDownInt32ToUint8Scale in order
  to apply bounded relu

- Added support for int32_t biases

- Extended tests

Change-Id: I015dae17faa7284766b5435ca33bcf593c1b2b69
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/96512
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

4 files changed, 152 insertions, 32 deletions

diff --git a/tests/validation/CPP/GEMMLowp.cpp b/tests/validation/CPP/GEMMLowp.cpp
index 8670a22a66..bf002cf2b5 100644
--- a/tests/validation/CPP/GEMMLowp.cpp
+++ b/tests/validation/CPP/GEMMLowp.cpp
@@ -33,6 +33,36 @@ namespace validation
 {
 namespace reference
 {
+namespace
+{
+template <typename T>
+void quantize_down_int32_to_uint8_scale(const SimpleTensor<T> *in, const SimpleTensor<T> *bias, SimpleTensor<uint8_t> *dst, int32_t result_offset, int32_t result_mult_int, int32_t result_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] + result_offset) * result_mult_int;
+
+        if(bias != nullptr)
+        {
+            result += (*bias)[i % cols_in];
+        }
+
+        result >>= result_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>
 SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<T> &a, const SimpleTensor<T> &b, int32_t a_offset, int32_t b_offset)
 {
@@ -80,21 +110,31 @@ SimpleTensor<int32_t> gemmlowp(const SimpleTensor<int8_t> &a, const SimpleTensor
 }
 
 template <typename T>
-SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<T> &in, int32_t result_offset, int32_t result_mult_int, int32_t result_shift)
+SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<T> &in, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min, int32_t max)
 {
     SimpleTensor<uint8_t> dst(in.shape(), DataType::QASYMM8);
 
-    for(int i = 0; i < in.num_elements(); ++i)
-    {
-        const int32_t result = ((in[i] + result_offset) * result_mult_int) >> result_shift;
-        dst[i]               = static_cast<uint8_t>(std::max(0, std::min(255, result)));
-    }
+    quantize_down_int32_to_uint8_scale<T>(&in, nullptr, &dst, result_offset, result_mult_int, result_shift, min, max);
+
+    return dst;
+}
+
+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, int32_t max)
+{
+    SimpleTensor<uint8_t> dst(in.shape(), DataType::QASYMM8);
+
+    quantize_down_int32_to_uint8_scale<T>(&in, &bias, &dst, result_offset, result_mult_int, result_shift, min, max);
 
     return dst;
 }
 
 template SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<uint8_t> &a, const SimpleTensor<uint8_t> &b, int32_t a_offset, int32_t b_offset);
-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);
+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,
+                                                                           int32_t result_shift, int32_t min, int32_t max);
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/CPP/GEMMLowp.h b/tests/validation/CPP/GEMMLowp.h
index cbed2206e3..ee33d8e0c0 100644
--- a/tests/validation/CPP/GEMMLowp.h
+++ b/tests/validation/CPP/GEMMLowp.h
@@ -35,14 +35,17 @@ namespace validation
 {
 namespace reference
 {
+SimpleTensor<int32_t> gemmlowp(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b);
+
 template <typename T>
 SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<T> &a, const SimpleTensor<T> &b, int32_t a_offset, int32_t b_offset);
 
 template <typename T>
-SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<T> &in, int32_t result_offset, int32_t result_mult_int, int32_t result_shift);
-
-SimpleTensor<int32_t> gemmlowp(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b);
+SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<T> &in, 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(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);
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/NEON/GEMMLowp.cpp b/tests/validation/NEON/GEMMLowp.cpp
index ba91ced443..078096a0dd 100644
--- a/tests/validation/NEON/GEMMLowp.cpp
+++ b/tests/validation/NEON/GEMMLowp.cpp
@@ -131,34 +131,55 @@ TEST_SUITE(OutputStage)
 
 TEST_SUITE(QuantizeDownInt32ToUint8Scale)
 
-using NEGEMMLowpQuantizeDownInt32ToUint8ScaleFixture = GEMMLowpQuantizeDownInt32ToUint8ScaleValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToUint8Scale>;
+const auto quantize_down_int32_to_uint8_scale_cases = framework::dataset::make("result_offset", -2, 2) * framework::dataset::make("result_mult_int", 1, 2) * framework::dataset::make("result_shift", 2,
+                                                      3)
+                                                      * framework::dataset::make("min", 0) * framework::dataset::make("max", 0) * framework::dataset::make("addBias", { false, true });
+
+const auto quantize_down_int32_to_uint8_scale_relu_cases = framework::dataset::make("result_offset", -2, 2) * framework::dataset::make("result_mult_int", 1,
+                                                           2)
+                                                           * framework::dataset::make("result_shift", 2, 3) * framework::dataset::make("min", 0, 2) * framework::dataset::make("max", 171, 174) * framework::dataset::make("addBias", { false, true });
 
-const auto quantize_down_int32_to_uint8_scale_cases = framework::dataset::make("result_offset", -4, 4) * framework::dataset::make("result_mult_int", 1, 3) * framework::dataset::make("result_shift", 2,
-                                                      4);
+using NEGEMMLowpQuantizeDownInt32ToUint8ScaleFixture = GEMMLowpQuantizeDownInt32ToUint8ScaleValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToUint8Scale>;
 
 DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()), quantize_down_int32_to_uint8_scale_cases),
-               shape, result_offset, result_mult_int, result_shift)
+               shape, result_offset, result_mult_int, result_shift, min, max, add_bias)
 {
+    TensorShape shape_bias(shape[0]);
+
     // Create tensors
-    Tensor in  = create_tensor<Tensor>(shape, DataType::S32);
-    Tensor out = create_tensor<Tensor>(shape, DataType::QASYMM8);
+    Tensor in   = create_tensor<Tensor>(shape, DataType::S32);
+    Tensor bias = create_tensor<Tensor>(shape_bias, DataType::S32);
+    Tensor out  = create_tensor<Tensor>(shape, DataType::QASYMM8);
 
     ARM_COMPUTE_EXPECT(in.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
     ARM_COMPUTE_EXPECT(out.info()->is_resizable(), framework::LogLevel::ERRORS);
 
     // Create and configure function
     NEGEMMLowpQuantizeDownInt32ToUint8Scale output_stage;
-    output_stage.configure(&in, &out, result_offset, result_mult_int, result_shift);
+    output_stage.configure(&in, add_bias ? &bias : nullptr, &out, result_offset, result_mult_int, result_shift, min, max);
 
-    // Validate valid region
+    // Validate valid region input and output
     const ValidRegion valid_region = shape_to_valid_region(shape);
     validate(in.info()->valid_region(), valid_region);
     validate(out.info()->valid_region(), valid_region);
 
+    // Validate valid region bias
+    if(add_bias)
+    {
+        const ValidRegion valid_region_bias = shape_to_valid_region(shape_bias);
+        validate(bias.info()->valid_region(), valid_region_bias);
+    }
+
     // Validate padding
     const PaddingSize padding = PaddingCalculator(shape.x(), 16).required_padding();
     validate(in.info()->padding(), padding);
     validate(out.info()->padding(), padding);
+
+    if(add_bias)
+    {
+        validate(bias.info()->padding(), padding);
+    }
 }
 
 FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToUint8ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_cases))
@@ -173,8 +194,35 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpQuantizeDownInt32ToUint8ScaleFixture,
     validate(Accessor(_target), _reference);
 }
 
-TEST_SUITE_END() // QuantizeDownInt32ToUint8Scale
+TEST_SUITE(BoundedReLu)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToUint8ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_relu_cases))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
 
+FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpQuantizeDownInt32ToUint8ScaleFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), quantize_down_int32_to_uint8_scale_relu_cases))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END() // BoundedReLu
+
+TEST_SUITE(AddBias)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToUint8ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_relu_cases))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpQuantizeDownInt32ToUint8ScaleFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), quantize_down_int32_to_uint8_scale_relu_cases))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END() // AddBias
+
+TEST_SUITE_END() // QuantizeDownInt32ToUint8Scale
 TEST_SUITE_END() // OutputStage
 
 TEST_SUITE_END() // GEMMLowp
diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h
index f9b0dbd959..a99e9323c8 100644
--- a/tests/validation/fixtures/GEMMLowpFixture.h
+++ b/tests/validation/fixtures/GEMMLowpFixture.h
@@ -122,10 +122,10 @@ class GEMMLowpQuantizeDownInt32ToUint8ScaleValidationFixture : public framework:
 {
 public:
     template <typename...>
-    void setup(TensorShape shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift)
+    void setup(TensorShape shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min, int32_t max, bool add_bias)
     {
-        _target    = compute_target(shape, result_offset, result_mult_int, result_shift);
-        _reference = compute_reference(shape, result_offset, result_mult_int, result_shift);
+        _target    = compute_target(shape, result_offset, result_mult_int, result_shift, min, max, add_bias);
+        _reference = compute_reference(shape, result_offset, result_mult_int, result_shift, min, max, add_bias);
     }
 
 protected:
@@ -136,43 +136,72 @@ protected:
         library->fill(tensor, distribution, i);
     }
 
-    TensorType compute_target(const TensorShape &shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift)
+    TensorType compute_target(const TensorShape &shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min, int32_t max, bool add_bias)
     {
+        TensorShape shape_bias(shape[0]);
+
         // Create tensors
         TensorType a = create_tensor<TensorType>(shape, DataType::S32, 1);
-        TensorType b = create_tensor<TensorType>(shape, DataType::QASYMM8, 1);
+        TensorType b = create_tensor<TensorType>(shape_bias, DataType::S32, 1);
+        TensorType c = create_tensor<TensorType>(shape, DataType::QASYMM8, 1);
 
         // Create and configure function
         FunctionType output_stage;
-        output_stage.configure(&a, &b, result_offset, result_mult_int, result_shift);
+        output_stage.configure(&a, add_bias ? &b : nullptr, &c, result_offset, result_mult_int, result_shift, min, max);
 
         ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS);
-        ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS);
 
         // Allocate tensors
         a.allocator()->allocate();
-        b.allocator()->allocate();
+        c.allocator()->allocate();
 
         ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS);
-        ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS);
 
-        // Fill tensors
+        // Fill tensor
         fill(AccessorType(a), 0);
 
+        if(add_bias)
+        {
+            ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+            // Allocate bias tensor
+            b.allocator()->allocate();
+
+            ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+            // Fill tensor
+            fill(AccessorType(b), 1);
+        }
+
         // Compute GEMM function
         output_stage.run();
-        return b;
+        return c;
     }
 
-    SimpleTensor<uint8_t> compute_reference(const TensorShape &shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift)
+    SimpleTensor<uint8_t> compute_reference(const TensorShape &shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min, int32_t max, bool add_bias)
     {
         // Create reference
+        TensorShape shape_bias(shape[0]);
+
         SimpleTensor<int32_t> a{ shape, DataType::S32, 1 };
+        SimpleTensor<int32_t> b{ shape_bias, DataType::S32, 1 };
 
         // Fill reference
         fill(a, 0);
 
-        return reference::gemmlowp_quantize_down_int32_to_uint8_scale<int32_t>(a, result_offset, result_mult_int, result_shift);
+        if(add_bias)
+        {
+            // Fill bias
+            fill(b, 1);
+
+            return reference::gemmlowp_quantize_down_int32_to_uint8_scale<int32_t>(a, b, result_offset, result_mult_int, result_shift, min, max);
+        }
+        else
+        {
+            return reference::gemmlowp_quantize_down_int32_to_uint8_scale<int32_t>(a, result_offset, result_mult_int, result_shift, min, max);
+        }
     }
 
     TensorType            _target{};