COMPMID-617: Add validate support for NEON BatchNormalizationLayer.

Change-Id: I037ec6df7eee06bdd1381e908677803426fa614c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110788
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

5 files changed, 156 insertions, 36 deletions

diff --git a/arm_compute/core/NEON/kernels/NEBatchNormalizationLayerKernel.h b/arm_compute/core/NEON/kernels/NEBatchNormalizationLayerKernel.h
index 1dfe075310..f5f818c083 100644
--- a/arm_compute/core/NEON/kernels/NEBatchNormalizationLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEBatchNormalizationLayerKernel.h
@@ -62,6 +62,24 @@ public:
      * @param[in]      epsilon Small value to avoid division with zero.
      */
     void configure(ITensor *input, ITensor *output, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, float epsilon);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEBatchNormalizationLayerKernel
+     *
+     * @param[in] input   Source tensor info. In case of @p output tensor = nullptr, this tensor will store the result.
+     *                    3 lower dimensions represent a single input with dimensions [width, height, FM].
+     *                    The rest are optional and used for representing batches. Data types supported: QS8/QS16/F16/F32.
+     * @param[in] output  Destination tensor info. Output will have the same number of dimensions as input. Data type supported: same as @p input
+     * @param[in] mean    Mean values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
+     * @param[in] var     Variance values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
+     * @param[in] beta    Beta values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
+     * @param[in] gamma   Gamma values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
+     * @param[in] epsilon Small value to avoid division with zero.
+     *
+     * @return an error status
+     */
+    static Error validate(const ITensorInfo *input, const ITensorInfo *output,
+                          const ITensorInfo *mean, const ITensorInfo *var,
+                          const ITensorInfo *beta, const ITensorInfo *gamma,
+                          float epsilon);
 
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
diff --git a/arm_compute/runtime/NEON/functions/NEBatchNormalizationLayer.h b/arm_compute/runtime/NEON/functions/NEBatchNormalizationLayer.h
index b2de7162f1..b3110883cc 100644
--- a/arm_compute/runtime/NEON/functions/NEBatchNormalizationLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEBatchNormalizationLayer.h
@@ -58,6 +58,24 @@ public:
      * @param[in]      epsilon Small value to avoid division with zero.
      */
     void configure(ITensor *input, ITensor *output, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, float epsilon);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEBatchNormalizationLayer
+     *
+     * @param[in] input   Source tensor info. In case of @p output tensor = nullptr, this tensor will store the result.
+     *                    3 lower dimensions represent a single input with dimensions [width, height, FM].
+     *                    The rest are optional and used for representing batches. Data types supported: QS8/QS16/F16/F32.
+     * @param[in] output  Destination tensor info. Output will have the same number of dimensions as input. Data type supported: same as @p input
+     * @param[in] mean    Mean values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
+     * @param[in] var     Variance values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
+     * @param[in] beta    Beta values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
+     * @param[in] gamma   Gamma values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
+     * @param[in] epsilon Small value to avoid division with zero.
+     *
+     * @return an error status
+     */
+    static Error validate(const ITensorInfo *input, const ITensorInfo *output,
+                          const ITensorInfo *mean, const ITensorInfo *var,
+                          const ITensorInfo *beta, const ITensorInfo *gamma,
+                          float epsilon);
 
     // Inherited methods overridden:
     void run() override;
diff --git a/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp b/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp
index 1123f2c9ca..4bbf67d13b 100644
--- a/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp
@@ -33,9 +33,39 @@
 
 using namespace arm_compute;
 
-NEBatchNormalizationLayerKernel::NEBatchNormalizationLayerKernel()
-    : _func(nullptr), _input(nullptr), _output(nullptr), _mean(nullptr), _var(nullptr), _gamma(nullptr), _beta(nullptr), _epsilon()
+namespace
+{
+Error validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *mean, const ITensorInfo *var, const ITensorInfo *beta, const ITensorInfo *gamma, float epsilon)
 {
+    ARM_COMPUTE_UNUSED(epsilon);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
+
+    if(nullptr != output)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+    }
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, mean, var, beta, gamma);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, mean, var, beta, gamma);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mean, var, beta, gamma);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(2) != mean->dimension(0));
+
+    return Error{};
+}
+
+std::pair<Error, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
+{
+    unsigned int num_elems_processed_per_iteration = 16 / input->element_size();
+
+    Window                 win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
+    AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
+    bool                   window_changed = update_window_and_padding(win, input_access, output_access);
+    output_access.set_valid_region(win, input->valid_region());
+    Error err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Error{};
+    return std::make_pair(err, win);
 }
 
 void batch_normalization_q8(ITensor *in, ITensor *out, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, float epsilon, const Window &window)
@@ -213,10 +243,28 @@ void batch_normalization_fp16(ITensor *in, ITensor *out, const ITensor *mean, co
     input, output);
 }
 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+} // namespace
+
+NEBatchNormalizationLayerKernel::NEBatchNormalizationLayerKernel()
+    : _func(nullptr), _input(nullptr), _output(nullptr), _mean(nullptr), _var(nullptr), _gamma(nullptr), _beta(nullptr), _epsilon()
+{
+}
 
 void NEBatchNormalizationLayerKernel::configure(ITensor *input, ITensor *output, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, float epsilon)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, mean, var, beta, gamma);
+
+    ITensorInfo *output_info = nullptr;
+
+    if(nullptr != output)
+    {
+        // Output tensor auto initialization if not yet initialized
+        auto_init_if_empty(*output->info(), *input->info());
+
+        output_info = output->info();
+    }
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output_info, mean->info(), var->info(), beta->info(), gamma->info(), epsilon));
 
     _input   = input;
     _output  = input;
@@ -228,39 +276,23 @@ void NEBatchNormalizationLayerKernel::configure(ITensor *input, ITensor *output,
 
     if(output != nullptr)
     {
-        // Output tensor auto initialization if not yet initialized
-        auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position());
-
-        ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output);
-
         _output = output;
     }
 
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, mean, var, beta, gamma);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output, mean, var, beta, gamma);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(mean, var, beta, gamma);
-    ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != mean->info()->dimension(0));
-
-    unsigned int num_elems_processed_per_iteration = 0;
-
     switch(input->info()->data_type())
     {
         case DataType::QS8:
-            _func                             = &batch_normalization_q8;
-            num_elems_processed_per_iteration = 16;
+            _func = &batch_normalization_q8;
             break;
         case DataType::QS16:
-            _func                             = &batch_normalization_q16;
-            num_elems_processed_per_iteration = 8;
+            _func = &batch_normalization_q16;
             break;
         case DataType::F32:
-            _func                             = &batch_normalization_fp32;
-            num_elems_processed_per_iteration = 4;
+            _func = &batch_normalization_fp32;
             break;
         case DataType::F16:
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            _func                             = &batch_normalization_fp16;
-            num_elems_processed_per_iteration = 8;
+            _func = &batch_normalization_fp16;
             break;
 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
         default:
@@ -268,19 +300,19 @@ void NEBatchNormalizationLayerKernel::configure(ITensor *input, ITensor *output,
             break;
     }
 
-    Window                 win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
-    AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration);
-    if(output != nullptr)
-    {
-        AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
-        update_window_and_padding(win, input_access, output_access);
-        output_access.set_valid_region(win, input->info()->valid_region());
-    }
-    else
-    {
-        update_window_and_padding(win, input_access);
-    }
-    INEKernel::configure(win);
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), output_info);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    INEKernel::configure(win_config.second);
+}
+
+Error NEBatchNormalizationLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *mean, const ITensorInfo *var, const ITensorInfo *beta, const ITensorInfo *gamma,
+                                                float epsilon)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, mean, var, beta, gamma, epsilon));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output ? output->clone().get() : nullptr).first);
+
+    return Error{};
 }
 
 void NEBatchNormalizationLayerKernel::run(const Window &window, const ThreadInfo &info)
diff --git a/src/runtime/NEON/functions/NEBatchNormalizationLayer.cpp b/src/runtime/NEON/functions/NEBatchNormalizationLayer.cpp
index ef79b02048..cfab12c33b 100644
--- a/src/runtime/NEON/functions/NEBatchNormalizationLayer.cpp
+++ b/src/runtime/NEON/functions/NEBatchNormalizationLayer.cpp
@@ -43,6 +43,12 @@ void NEBatchNormalizationLayer::configure(ITensor *input, ITensor *output, const
     _norm_kernel.configure(input, output, mean, var, beta, gamma, epsilon);
 }
 
+Error NEBatchNormalizationLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *mean, const ITensorInfo *var, const ITensorInfo *beta, const ITensorInfo *gamma,
+                                          float epsilon)
+{
+    return NEBatchNormalizationLayerKernel::validate(input, output, mean, var, beta, gamma, epsilon);
+}
+
 void NEBatchNormalizationLayer::run()
 {
     NEScheduler::get().schedule(&_norm_kernel, Window::DimY);
diff --git a/tests/validation/NEON/BatchNormalizationLayer.cpp b/tests/validation/NEON/BatchNormalizationLayer.cpp
index a1421d0b17..806d3b3f20 100644
--- a/tests/validation/NEON/BatchNormalizationLayer.cpp
+++ b/tests/validation/NEON/BatchNormalizationLayer.cpp
@@ -80,6 +80,52 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(datasets::Ran
     validate(dst.info()->valid_region(), valid_region);
 }
 
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
+               framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),    // Window shrink
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),    // Mismatching data types
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),    // Mismatching data types
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),    // Invalid mean/var/beta/gamma shape
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 2), // Mismatching fixed point position
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 2),
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 2),
+                                                     }),
+               framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F16),
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 3),
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 2),
+                                                       TensorInfo(),
+                                                     })),
+               framework::dataset::make("MVBGInfo",{ TensorInfo(TensorShape(2U), 1, DataType::F32),
+                                                     TensorInfo(TensorShape(2U), 1, DataType::F32),
+                                                     TensorInfo(TensorShape(2U), 1, DataType::F16),
+                                                     TensorInfo(TensorShape(2U), 1, DataType::F32),
+                                                     TensorInfo(TensorShape(5U), 1, DataType::F32),
+                                                     TensorInfo(TensorShape(2U), 1, DataType::QS8, 2),
+                                                     TensorInfo(TensorShape(2U), 1, DataType::QS8, 2),
+                                                     TensorInfo(TensorShape(2U), 1, DataType::QS8, 2),
+                                                   })),
+               framework::dataset::make("Expected", { false, true, true, true, true, true, false, false})),
+               input_info, output_info, mvbg_info, expected)
+{
+    const auto &mean_info = mvbg_info;
+    const auto &var_info = mvbg_info;
+    const auto &beta_info = mvbg_info;
+    const auto &gamma_info = mvbg_info;
+    bool has_error = bool(NEBatchNormalizationLayer::validate(
+            &input_info.clone()->set_is_resizable(false), output_info.total_size() ? &output_info.clone()->set_is_resizable(false) : nullptr,
+            &mean_info.clone()->set_is_resizable(false), &var_info.clone()->set_is_resizable(false),
+            &beta_info.clone()->set_is_resizable(false), &gamma_info.clone()->set_is_resizable(false), 1.f));
+    ARM_COMPUTE_EXPECT(has_error == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
 TEST_SUITE(Float)
 FIXTURE_DATA_TEST_CASE(Random, NEBatchNormalizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::RandomBatchNormalizationLayerDataset(),
                                                                                                                    framework::dataset::make("DataType", DataType::F32)))