COMPMID-1523: Fuse BN node with convolution.

Change-Id: I146936c9e98b343496a4b61cdbadf0eaa38e885a
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/154008
Reviewed-by: Michele DiGiorgio <michele.digiorgio@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Tested-by: bsgcomp <bsgcomp@arm.com>

3 files changed, 224 insertions, 10 deletions

diff --git a/tests/validation/CL/BatchNormalizationLayer.cpp b/tests/validation/CL/BatchNormalizationLayer.cpp
index 0d80ff7eb7..cbf3c7092d 100644
--- a/tests/validation/CL/BatchNormalizationLayer.cpp
+++ b/tests/validation/CL/BatchNormalizationLayer.cpp
@@ -25,16 +25,20 @@
 #include "arm_compute/runtime/CL/CLTensor.h"
 #include "arm_compute/runtime/CL/CLTensorAllocator.h"
 #include "arm_compute/runtime/CL/functions/CLBatchNormalizationLayer.h"
+#include "arm_compute/runtime/CL/functions/CLConvolutionLayer.h"
+#include "arm_compute/runtime/CL/functions/CLFuseBatchNormalization.h"
 #include "tests/CL/CLAccessor.h"
 #include "tests/PaddingCalculator.h"
+#include "tests/datasets/LargeConvolutionLayerDataset.h"
 #include "tests/datasets/RandomBatchNormalizationLayerDataset.h"
-#include "tests/datasets/ShapeDatasets.h"
+#include "tests/datasets/SmallConvolutionLayerDataset.h"
 #include "tests/framework/Asserts.h"
 #include "tests/framework/Macros.h"
 #include "tests/framework/datasets/Datasets.h"
 #include "tests/validation/Helpers.h"
 #include "tests/validation/Validation.h"
 #include "tests/validation/fixtures/BatchNormalizationLayerFixture.h"
+#include "tests/validation/fixtures/BatchNormalizationLayerFusionFixture.h"
 
 namespace arm_compute
 {
@@ -44,14 +48,20 @@ namespace validation
 {
 namespace
 {
-constexpr AbsoluteTolerance<float> tolerance_f32(0.00001f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
-constexpr AbsoluteTolerance<float> tolerance_f16(0.01f);    /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */
+RelativeTolerance<float>           rel_tolerance_f32(0.05f);    /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
+constexpr AbsoluteTolerance<float> abs_tolerance_f32(0.00001f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
+constexpr AbsoluteTolerance<float> tolerance_f16(0.01f);        /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */
 const auto                         act_infos = framework::dataset::make("ActivationInfo",
 {
     ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
     ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f),
     ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f),
 });
+
+const auto common_fusion_dataset = combine(combine(combine(framework::dataset::make("UseBias", { false, true }),
+                                                           framework::dataset::make("UseBeta", { false, true })),
+                                                   framework::dataset::make("UseGamma", { false, true })),
+                                           framework::dataset::make("Epsilon", { 0.001f }));
 } // namespace
 
 TEST_SUITE(CL)
@@ -150,9 +160,9 @@ FIXTURE_DATA_TEST_CASE(Random, CLBatchNormalizationLayerFixture<float>, framewor
                                                                                                                    framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
 {
     // Validate output
-    validate(CLAccessor(_target), _reference, tolerance_f32, 0);
+    validate(CLAccessor(_target), _reference, abs_tolerance_f32, 0);
 }
-TEST_SUITE_END()
+TEST_SUITE_END() //FP32
 
 TEST_SUITE(FP16)
 FIXTURE_DATA_TEST_CASE(Random, CLBatchNormalizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::RandomBatchNormalizationLayerDataset(),
@@ -165,11 +175,30 @@ FIXTURE_DATA_TEST_CASE(Random, CLBatchNormalizationLayerFixture<half>, framework
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f16, 0);
 }
-TEST_SUITE_END()
-TEST_SUITE_END()
+TEST_SUITE_END() // FP16
+TEST_SUITE_END() // Float
+
+TEST_SUITE_END() // BatchNormalizationLayer
+
+TEST_SUITE(BatchNormalizationLayerFusion)
+template <typename T>
+using CLBatchNormalizationLayerFusionFixture = BatchNormalizationLayerFusionValidationFixture<CLTensor, CLAccessor, CLConvolutionLayer, CLFuseBatchNormalization, T>;
+
+TEST_SUITE(Float)
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLBatchNormalizationLayerFusionFixture<float>, framework::DatasetMode::PRECOMMIT,
+                       combine(combine(combine(datasets::SmallConvolutionLayerDataset(), common_fusion_dataset),
+                                       framework::dataset::make("DataType", DataType::F32)),
+                               framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // Float
 
-TEST_SUITE_END()
-TEST_SUITE_END()
+TEST_SUITE_END() // BatchNormalizationLayerFusion
+TEST_SUITE_END() // CL
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/fixtures/BatchNormalizationLayerFusionFixture.h b/tests/validation/fixtures/BatchNormalizationLayerFusionFixture.h
new file mode 100644
index 0000000000..39c7d46114
--- /dev/null
+++ b/tests/validation/fixtures/BatchNormalizationLayerFusionFixture.h
@@ -0,0 +1,186 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_TEST_BATCH_NORMALIZATION_LAYER_FUSION_FIXTURE
+#define ARM_COMPUTE_TEST_BATCH_NORMALIZATION_LAYER_FUSION_FIXTURE
+
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "tests/AssetsLibrary.h"
+#include "tests/Globals.h"
+#include "tests/IAccessor.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Fixture.h"
+#include "tests/validation/Helpers.h"
+#include "tests/validation/reference/BatchNormalizationLayer.h"
+#include "tests/validation/reference/ConvolutionLayer.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+template <typename TensorType, typename AccessorType, typename ConvolutionFunctionType, typename FusionFunctionType, typename T>
+class BatchNormalizationLayerFusionValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape src_shape, TensorShape w_shape, TensorShape b_shape, TensorShape dst_shape, PadStrideInfo info, Size2D dilation,
+               bool use_conv_b, bool use_beta, bool use_gamma, float epsilon, DataType dt, DataLayout data_layout)
+    {
+        ARM_COMPUTE_UNUSED(dilation);
+
+        _data_type   = dt;
+        _data_layout = data_layout;
+        _use_conv_b  = use_conv_b;
+        _use_beta    = use_beta;
+        _use_gamma   = use_gamma;
+
+        _target    = compute_target(src_shape, w_shape, b_shape, dst_shape, info, epsilon);
+        _reference = compute_reference(src_shape, w_shape, b_shape, dst_shape, info, epsilon);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&src, U &&w_tensor, U &&b_tensor, U &&mean_tensor, U &&var_tensor, U &&beta_tensor, U &&gamma_tensor)
+    {
+        std::uniform_real_distribution<> distribution(-1.f, 1.f);
+        std::uniform_real_distribution<> distribution_gz(0, 1.f);
+
+        library->fill(src, distribution, 0);
+        library->fill(w_tensor, distribution, 1);
+        library->fill(mean_tensor, distribution, 2);
+        library->fill(var_tensor, distribution_gz, 3);
+        _use_conv_b ? library->fill(b_tensor, distribution, 4) : library->fill_tensor_value(b_tensor, 0.f);
+        _use_beta ? library->fill(beta_tensor, distribution, 5) : library->fill_tensor_value(beta_tensor, 0.f);
+        _use_gamma ? library->fill(gamma_tensor, distribution, 6) : library->fill_tensor_value(gamma_tensor, 1.f);
+    }
+
+    TensorType compute_target(TensorShape src_shape, TensorShape w_shape, TensorShape b_shape, TensorShape dst_shape, PadStrideInfo info, float epsilon)
+    {
+        if(_data_layout == DataLayout::NHWC)
+        {
+            permute(src_shape, PermutationVector(2U, 0U, 1U));
+            permute(w_shape, PermutationVector(2U, 0U, 1U));
+            permute(dst_shape, PermutationVector(2U, 0U, 1U));
+        }
+
+        // Create tensors
+        TensorType src      = create_tensor<TensorType>(src_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+        TensorType conv_w   = create_tensor<TensorType>(w_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+        TensorType conv_b   = create_tensor<TensorType>(b_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+        TensorType bn_mean  = create_tensor<TensorType>(b_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+        TensorType bn_var   = create_tensor<TensorType>(b_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+        TensorType bn_beta  = create_tensor<TensorType>(b_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+        TensorType bn_gamma = create_tensor<TensorType>(b_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+        TensorType fused_w  = create_tensor<TensorType>(w_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+        TensorType fused_b  = create_tensor<TensorType>(b_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+        TensorType dst      = create_tensor<TensorType>(dst_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+
+        // Create and configure function
+        FusionFunctionType      fuse_fn;
+        ConvolutionFunctionType conv_fn;
+        TensorType             *conv_b_ptr = _use_conv_b ? &conv_b : nullptr;
+        TensorType             *beta_ptr   = _use_beta ? &bn_beta : nullptr;
+        TensorType             *gamma_ptr  = _use_gamma ? &bn_gamma : nullptr;
+        fuse_fn.configure(&conv_w, &bn_mean, &bn_var, &fused_w, &fused_b, conv_b_ptr, beta_ptr, gamma_ptr, epsilon);
+        conv_fn.configure(&src, &fused_w, &fused_b, &dst, info);
+
+        ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(conv_w.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(conv_b.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(bn_mean.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(bn_var.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(bn_beta.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(bn_gamma.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(fused_w.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(fused_b.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        src.allocator()->allocate();
+        conv_w.allocator()->allocate();
+        conv_b.allocator()->allocate();
+        bn_mean.allocator()->allocate();
+        bn_var.allocator()->allocate();
+        bn_beta.allocator()->allocate();
+        bn_gamma.allocator()->allocate();
+        fused_w.allocator()->allocate();
+        fused_b.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!conv_w.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!conv_b.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!bn_mean.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!bn_var.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!bn_beta.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!bn_gamma.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!fused_w.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!fused_b.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(src),
+             AccessorType(conv_w), AccessorType(conv_b),
+             AccessorType(bn_mean), AccessorType(bn_var), AccessorType(bn_beta), AccessorType(bn_gamma));
+
+        // Compute function
+        fuse_fn.run();
+        conv_fn.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(TensorShape src_shape, TensorShape w_shape, TensorShape b_shape, TensorShape dst_shape, PadStrideInfo info, float epsilon)
+    {
+        // Create reference
+        SimpleTensor<T> src{ src_shape, _data_type, 1 };
+        SimpleTensor<T> conv_w{ w_shape, _data_type, 1 };
+        SimpleTensor<T> conv_b{ b_shape, _data_type, 1 };
+        SimpleTensor<T> bn_var{ b_shape, _data_type, 1 };
+        SimpleTensor<T> bn_mean{ b_shape, _data_type, 1 };
+        SimpleTensor<T> bn_beta{ b_shape, _data_type, 1 };
+        SimpleTensor<T> bn_gamma{ b_shape, _data_type, 1 };
+
+        // Fill reference
+        fill(src, conv_w, conv_b, bn_mean, bn_var, bn_beta, bn_gamma);
+
+        // Calculate Conv + BN
+        auto conv_res = reference::convolution_layer(src, conv_w, conv_b, dst_shape, info);
+        return reference::batch_normalization_layer(conv_res, bn_mean, bn_var, bn_beta, bn_gamma, epsilon, ActivationLayerInfo());
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+    DataType        _data_type{};
+    DataLayout      _data_layout{};
+    bool            _use_conv_b{};
+    bool            _use_beta{};
+    bool            _use_gamma{};
+};
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_BATCH_NORMALIZATION_LAYER_FUSION_FIXTURE */
diff --git a/tests/validation/reference/BatchNormalizationLayer.cpp b/tests/validation/reference/BatchNormalizationLayer.cpp
index 4ea3769c2c..37713c841d 100644
--- a/tests/validation/reference/BatchNormalizationLayer.cpp
+++ b/tests/validation/reference/BatchNormalizationLayer.cpp
@@ -77,7 +77,6 @@ template SimpleTensor<float> batch_normalization_layer(const SimpleTensor<float>
 template SimpleTensor<half> batch_normalization_layer(const SimpleTensor<half> &src, const SimpleTensor<half> &mean, const SimpleTensor<half> &var,
                                                       const SimpleTensor<half> &beta,
                                                       const SimpleTensor<half> &gamma, float epsilon, ActivationLayerInfo act_info);
-
 } // namespace reference
 } // namespace validation
 } // namespace test