Implement FP32/16 Depthwise Conv2d operator in dynamic fusion

This patch adds Depthwise Conv2d operator into dynamic fusion interface and adds the associated tests.

Resolves: COMPMID-5517
Change-Id: I385c94dff7fd40c72b8337ef797e508df4499a82
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8678
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 699 insertions, 0 deletions

diff --git a/tests/validation/dynamic_fusion/gpu/cl/DepthwiseConv2d.cpp b/tests/validation/dynamic_fusion/gpu/cl/DepthwiseConv2d.cpp
new file mode 100644
index 0000000000..f08cc60ea2
--- /dev/null
+++ b/tests/validation/dynamic_fusion/gpu/cl/DepthwiseConv2d.cpp
@@ -0,0 +1,477 @@
+/*
+ * Copyright (c) 2022 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.
+ */
+
+#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuDepthwiseConv2d.h"
+
+#include "tests/CL/CLAccessor.h"
+#include "tests/datasets/DepthwiseConvolutionLayerDataset.h"
+#include "tests/datasets/DilatedDepthwiseConvolutionLayerDataset.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Fixture.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/dynamic_fusion/gpu/cl/DepthwiseConv2dFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+const auto depth_multipliers       = framework::dataset::make("DepthMultiplier", { 1U, 4U });
+const auto large_depth_multipliers = framework::dataset::make("DepthMultiplier", { 1, 2, 5, 8 });
+
+TEST_SUITE(CL)
+TEST_SUITE(DYNAMIC_FUSION)
+TEST_SUITE(DEPTHWISE_CONV2D)
+
+RelativeTolerance<float>            tolerance_f32(0.01f);                 /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
+RelativeTolerance<half_float::half> tolerance_f16(half_float::half(0.1)); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */
+constexpr float                     tolerance_num = 0.02f;                /**< Tolerance number */
+
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zip(zip(
+                framework::dataset::make("InputInfo", { TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC),    // Mismatching data type input/weights
+                                                        TensorInfo(TensorShape(3U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC),    // Mismatching input feature maps
+                                                        TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC),    // Mismatching depth multiplier
+                                                        TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC),    // Invalid biases size
+                                                        TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC),    // Invalid biases dimensions
+                                                        TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC),    // Invalid output size
+                                                        TensorInfo(TensorShape(8U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC),    // patch size bigger than input width
+                                                        TensorInfo(TensorShape(8U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC),    // dilation < 1
+                                                        TensorInfo(TensorShape(8U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC),
+                                                        TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::QASYMM8, DataLayout::NHWC),    // Unsupported data type
+                                                        TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::QASYMM8_SIGNED, DataLayout::NHWC),    // Unsupported data type
+                                                        TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::QSYMM16, DataLayout::NHWC),    // Unsupported data type
+                                                        TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::QSYMM8, DataLayout::NHWC),    // Unsupported data type
+                                                        TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::QSYMM8_PER_CHANNEL, DataLayout::NHWC),    // Unsupported data type
+                                                        TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::QASYMM16, DataLayout::NHWC),    // Unsupported data type
+                                                        TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::U8, DataLayout::NHWC),    // Unsupported data type
+                                                        TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::S8, DataLayout::NHWC),    // Unsupported data type
+                                                        TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::U16, DataLayout::NHWC),    // Unsupported data type
+                                                        TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::S16, DataLayout::NHWC),    // Unsupported data type
+                                                        TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::U32, DataLayout::NHWC),    // Unsupported data type
+                                                        TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::S32, DataLayout::NHWC),    // Unsupported data type
+                                                        TensorInfo(TensorShape(32U, 13U, 8U), 1, DataType::F32, DataLayout::NCHW),    // Unsupported data layout
+                                                        TensorInfo(TensorShape(8U, 32U, 13U, 4U), 1, DataType::F32, DataLayout::NHWC),
+                                                        TensorInfo(TensorShape(8U, 32U, 13U, 4U), 1, DataType::F32, DataLayout::NHWC), // weight dimension > 3
+                                                        TensorInfo(TensorShape(8U, 32U, 13U, 4U), 1, DataType::F32, DataLayout::NHWC),
+                                                        TensorInfo(TensorShape(8U, 32U, 13U, 4U), 1, DataType::F32, DataLayout::NHWC),
+                                                        TensorInfo(TensorShape(8U, 32U, 13U, 4U), 1, DataType::F32, DataLayout::NHWC),
+                                                      }),
+                framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(2U, 3U, 3U, 2U), 1, DataType::F16, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(2U, 3U, 3U, 2U), 1, DataType::F32, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(2U, 3U, 3U, 2U), 1, DataType::F32, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(2U, 3U, 3U, 2U), 1, DataType::F32, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(2U, 3U, 3U, 2U), 1, DataType::F32, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(2U, 3U, 3U, 2U), 1, DataType::F32, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(16U, 3U, 3U), 1, DataType::F32, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(16U, 3U, 3U), 1, DataType::F32, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(16U, 3U, 3U), 1, DataType::F32, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::QASYMM8, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::QASYMM8_SIGNED, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::QSYMM16, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::QSYMM8, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::QSYMM8_PER_CHANNEL, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::QASYMM16, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::U8, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::S8, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::U16, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::S16, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::U32, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::S32, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(3U, 3U, 24U), 1, DataType::F32, DataLayout::NCHW),
+                                                          TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::F32, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(24U, 3U, 3U, 5U), 1, DataType::F32, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::F32, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::F32, DataLayout::NHWC),
+                                                          TensorInfo(TensorShape(24U, 4U, 3U), 1, DataType::F32, DataLayout::NHWC),
+                                                        })),
+                framework::dataset::make("BiasesInfo", { TensorInfo(TensorShape(2U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(2U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(2U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(2U, 2U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(2U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(16U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(16U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(16U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NCHW),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U), 1, DataType::F32, DataLayout::NHWC),
+                                                       })),
+                framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(2U, 25U, 11U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(2U, 25U, 11U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(2U, 25U, 11U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(2U, 25U, 11U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(2U, 25U, 11U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(16U, 25U, 11U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(16U, 25U, 11U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(16U, 25U, 11U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U, 32U, 11U), 1, DataType::QASYMM8, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U, 32U, 11U), 1, DataType::QASYMM8_SIGNED, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U, 32U, 11U), 1, DataType::QSYMM16, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U, 32U, 11U), 1, DataType::QSYMM8, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U, 32U, 11U), 1, DataType::QSYMM8_PER_CHANNEL, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U, 32U, 11U), 1, DataType::QASYMM16, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U, 32U, 11U), 1, DataType::U8, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U, 32U, 11U), 1, DataType::S8, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U, 32U, 11U), 1, DataType::U16, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U, 32U, 11U), 1, DataType::S16, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U, 32U, 11U), 1, DataType::U32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U, 32U, 11U), 1, DataType::S32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(32U, 11U, 24U), 1, DataType::F32, DataLayout::NCHW),
+                                                         TensorInfo(TensorShape(24U, 32U, 11U, 4U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U, 32U, 11U, 4U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U, 33U, 14U, 4U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U, 17U, 5U, 4U), 1, DataType::F32, DataLayout::NHWC),
+                                                         TensorInfo(TensorShape(24U, 15U, 4U, 4U), 1, DataType::F32, DataLayout::NHWC),
+                                                       })),
+                framework::dataset::make("Padding", {  Padding2D(0, 0, 0, 0),
+                                                       Padding2D(0, 0, 0, 0),
+                                                       Padding2D(0, 0, 0, 0),
+                                                       Padding2D(0, 0, 0, 0),
+                                                       Padding2D(0, 0, 0, 0),
+                                                       Padding2D(0, 0, 0, 0),
+                                                       Padding2D(0, 0, 0, 0),
+                                                       Padding2D(0, 0, 0, 0),
+                                                       Padding2D(0, 0, 0, 0),
+                                                       Padding2D(1, 1, 0, 0),
+                                                       Padding2D(1, 1, 0, 0),
+                                                       Padding2D(1, 1, 0, 0),
+                                                       Padding2D(1, 1, 0, 0),
+                                                       Padding2D(1, 1, 0, 0),
+                                                       Padding2D(1, 1, 0, 0),
+                                                       Padding2D(1, 1, 0, 0),
+                                                       Padding2D(1, 1, 0, 0),
+                                                       Padding2D(1, 1, 0, 0),
+                                                       Padding2D(1, 1, 0, 0),
+                                                       Padding2D(1, 1, 0, 0),
+                                                       Padding2D(1, 1, 0, 0),
+                                                       Padding2D(1, 1, 0, 0),
+                                                       Padding2D(1, 1, 0, 0),
+                                                       Padding2D(1, 1, 0, 0),
+                                                       Padding2D(2, 1, 2, 1),
+                                                       Padding2D(2, 1, 2, 1),
+                                                       Padding2D(2, 1, 2, 1),
+                                                      })),
+                framework::dataset::make("Stride", {   Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(1, 1),
+                                                       Size2D(2, 3),
+                                                       Size2D(2, 3),
+                                                      })),
+                framework::dataset::make("DepthMultiplier", { 1,
+                                                              1,
+                                                              3,
+                                                              1,
+                                                              1,
+                                                              1,
+                                                              2,
+                                                              2,
+                                                              2,
+                                                              3,
+                                                              3,
+                                                              3,
+                                                              3,
+                                                              3,
+                                                              3,
+                                                              3,
+                                                              3,
+                                                              3,
+                                                              3,
+                                                              3,
+                                                              3,
+                                                              3,
+                                                              3,
+                                                              3,
+                                                              3,
+                                                              3,
+                                                              3,
+                                                             })),
+                       framework::dataset::make("Dilation", { Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(20U, 1U),
+                                                              Size2D(0U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(2U, 3U),
+                                                             })),
+                framework::dataset::make("Expected", { false, false, false, false, false, false, false, false, true, false,
+                                                       false, false, false, false, false, false, false, false, false, false,
+                                                       false, false, true, false, true, true, true })),
+                input_info, weights_info, biases_info, output_info, padding, stride, depth_multiplier, dilation, expected)
+{
+    CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
+    GpuWorkloadContext gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
+    GpuWorkloadSketch sketch{ &gpu_ctx };
+
+    const TensorInfo sketch_input_info   = sketch.create_tensor_info(input_info);
+    const TensorInfo sketch_weights_info = sketch.create_tensor_info(weights_info);
+    const TensorInfo sketch_biases_info  = sketch.create_tensor_info(biases_info);
+    const TensorInfo sketch_output_info  = sketch.create_tensor_info(output_info);
+
+    DepthwiseConv2dAttributes attributes {};
+    attributes.pad(padding)
+              .stride(stride)
+              .dilation(dilation)
+              .depth_multiplier(depth_multiplier);
+
+    const Status status = GpuDepthwiseConv2d::validate_op(sketch, &sketch_input_info, &sketch_weights_info, &sketch_biases_info, &sketch_output_info, attributes);
+    const bool res = bool(status);
+    ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
+template <typename T>
+using DynamicFusionGpuDepthwiseConv2dFixture = DynamicFusionGpuDepthwiseConv2dValidationFixture<CLTensor, CLAccessor, GpuDepthwiseConv2d, T>;
+
+TEST_SUITE(Float)
+TEST_SUITE(FP16)
+TEST_SUITE(W3x3)
+FIXTURE_DATA_TEST_CASE(RunSmall, DynamicFusionGpuDepthwiseConv2dFixture<half>, framework::DatasetMode::ALL,
+                       combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(),
+                                               depth_multipliers),
+                                       framework::dataset::make("DataType", DataType::F16)),
+                               framework::dataset::make("DataLayout", DataLayout::NHWC)))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, DynamicFusionGpuDepthwiseConv2dFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(),
+                                                                                                                        large_depth_multipliers),
+                                                                                                                        framework::dataset::make("DataType", DataType::F16)),
+                                                                                                                        framework::dataset::make("DataLayout", DataLayout::NHWC)))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+TEST_SUITE(Dilation)
+FIXTURE_DATA_TEST_CASE(RunSmall, DynamicFusionGpuDepthwiseConv2dFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(),
+                                                                                                                    depth_multipliers),
+                                                                                                                    framework::dataset::make("DataType", DataType::F16)),
+                                                                                                                    framework::dataset::make("DataLayout", { DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, DynamicFusionGpuDepthwiseConv2dFixture<half>, framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(),
+                                               large_depth_multipliers),
+                                       framework::dataset::make("DataType", DataType::F16)),
+                               framework::dataset::make("DataLayout", { DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+TEST_SUITE_END() // Dilation
+TEST_SUITE_END() // W3x3
+
+TEST_SUITE(Generic)
+FIXTURE_DATA_TEST_CASE(RunSmall, DynamicFusionGpuDepthwiseConv2dFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(),
+                                                                                                                    depth_multipliers),
+                                                                                                                    framework::dataset::make("DataType", DataType::F16)),
+                                                                                                                    framework::dataset::make("DataLayout", { DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, DynamicFusionGpuDepthwiseConv2dFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset(),
+                                                                                                                        large_depth_multipliers),
+                                                                                                                        framework::dataset::make("DataType", DataType::F16)),
+                                                                                                                        framework::dataset::make("DataLayout", { DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num);
+}
+
+TEST_SUITE(Dilation)
+FIXTURE_DATA_TEST_CASE(RunSmall, DynamicFusionGpuDepthwiseConv2dFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(),
+                                                                                                                    depth_multipliers),
+                                                                                                                    framework::dataset::make("DataType", DataType::F16)),
+                                                                                                                    framework::dataset::make("DataLayout", { DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, DynamicFusionGpuDepthwiseConv2dFixture<half>, framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset(),
+                                               large_depth_multipliers),
+                                       framework::dataset::make("DataType", DataType::F16)),
+                               framework::dataset::make("DataLayout", { DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num);
+}
+TEST_SUITE_END() // Dilation
+TEST_SUITE_END() // Generic
+TEST_SUITE_END() // FP16
+
+TEST_SUITE(FP32)
+TEST_SUITE(W3x3)
+FIXTURE_DATA_TEST_CASE(RunSmall, DynamicFusionGpuDepthwiseConv2dFixture<float>, framework::DatasetMode::ALL,
+                       combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(),
+                                               depth_multipliers),
+                                       framework::dataset::make("DataType", DataType::F32)),
+                               framework::dataset::make("DataLayout", DataLayout::NHWC)))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, DynamicFusionGpuDepthwiseConv2dFixture<float>, framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(),
+                                               large_depth_multipliers),
+                                       framework::dataset::make("DataType", DataType::F32)),
+                               framework::dataset::make("DataLayout", DataLayout::NHWC)))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+
+TEST_SUITE(Dilation)
+
+FIXTURE_DATA_TEST_CASE(RunSmall, DynamicFusionGpuDepthwiseConv2dFixture<float>, framework::DatasetMode::ALL,
+                       combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(),
+                                               depth_multipliers),
+                                       framework::dataset::make("DataType", DataType::F32)),
+                               framework::dataset::make("DataLayout", DataLayout::NHWC)))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, DynamicFusionGpuDepthwiseConv2dFixture<float>, framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(),
+                                               large_depth_multipliers),
+                                       framework::dataset::make("DataType", DataType::F32)),
+                               framework::dataset::make("DataLayout", DataLayout::NHWC)))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END() // Dilation
+TEST_SUITE_END() // W3x3
+
+TEST_SUITE(Generic)
+FIXTURE_DATA_TEST_CASE(RunSmall, DynamicFusionGpuDepthwiseConv2dFixture<float>, framework::DatasetMode::ALL,
+                       combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(),
+                                               depth_multipliers),
+                                       framework::dataset::make("DataType", DataType::F32)),
+                               framework::dataset::make("DataLayout", { DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, DynamicFusionGpuDepthwiseConv2dFixture<float>, framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset(),
+                                               large_depth_multipliers),
+                                       framework::dataset::make("DataType", DataType::F32)),
+                               framework::dataset::make("DataLayout", { DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLargeKernelSize, DynamicFusionGpuDepthwiseConv2dFixture<float>, framework::DatasetMode::ALL,
+                       combine(combine(combine(datasets::LargeKernelSizeDepthwiseConvolutionLayerNHWCDataset(),
+                                               framework::dataset::make("DepthMultiplier", { 1 })),
+                                       framework::dataset::make("DataType", DataType::F32)),
+                               framework::dataset::make("DataLayout", { DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+
+TEST_SUITE(Dilation)
+FIXTURE_DATA_TEST_CASE(RunSmall, DynamicFusionGpuDepthwiseConv2dFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(),
+                                                                                                                     depth_multipliers),
+                                                                                                                     framework::dataset::make("DataType", DataType::F32)),
+                                                                                                                     framework::dataset::make("DataLayout", { DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, DynamicFusionGpuDepthwiseConv2dFixture<float>, framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(),
+                                               large_depth_multipliers),
+                                       framework::dataset::make("DataType", DataType::F32)),
+                               framework::dataset::make("DataLayout", { DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END() // Dilation
+TEST_SUITE_END() // Generic
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // Float
+TEST_SUITE_END() // DEPTHWISE_CONV2D
+TEST_SUITE_END() // DYNAMIC_FUSION
+TEST_SUITE_END() // CL
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/fixtures/dynamic_fusion/gpu/cl/DepthwiseConv2dFixture.h b/tests/validation/fixtures/dynamic_fusion/gpu/cl/DepthwiseConv2dFixture.h
new file mode 100644
index 0000000000..c7600e082e
--- /dev/null
+++ b/tests/validation/fixtures/dynamic_fusion/gpu/cl/DepthwiseConv2dFixture.h
@@ -0,0 +1,222 @@
+/*
+ * Copyright (c) 2022 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 TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DEPTHWISECONV2DFIXTURE
+#define TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DEPTHWISECONV2DFIXTURE
+
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h"
+#include "arm_compute/dynamic_fusion/sketch/attributes/DepthwiseConv2dAttributes.h"
+#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h"
+#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuDepthwiseConv2d.h"
+
+#include "tests/CL/CLAccessor.h"
+
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Fixture.h"
+#include "tests/framework/Macros.h"
+
+#include "tests/validation/Validation.h"
+#include "tests/validation/reference/DepthwiseConvolutionLayer.h"
+
+using namespace arm_compute::experimental::dynamic_fusion;
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class DynamicFusionGpuDepthwiseConv2dValidationGenericFixture : public framework::Fixture
+{
+public:
+    using TBias = typename std::conditional < std::is_same<typename std::decay<T>::type, uint8_t>::value
+                  || std::is_same<typename std::decay<T>::type, int8_t>::value,
+                  int32_t, T >::type; // If T: uint8_t or int8_t then TBias: int32_t, otherwise TBias: T
+
+    template <typename...>
+    void setup(TensorShape input_shape, Size2D kernel_size, const PadStrideInfo &pad_stride, const Size2D &dilation,
+               const unsigned int depth_multiplier, const DataType data_type, const DataLayout data_layout)
+    {
+        ARM_COMPUTE_ERROR_ON(data_layout != DataLayout::NHWC); // Dynamic fusion depthwise conv2d only supports NHWC layout
+
+        DepthwiseConv2dAttributes dwc_conv2d_attr;
+        const Padding2D           padding_2d(pad_stride.pad_left(), pad_stride.pad_right(), pad_stride.pad_top(), pad_stride.pad_bottom());
+        dwc_conv2d_attr.pad(padding_2d)
+        .stride(Size2D(pad_stride.stride().first, pad_stride.stride().second))
+        .dilation(dilation)
+        .depth_multiplier(depth_multiplier)
+        .dimension_rounding_type(pad_stride.round());
+
+        // Calculate Output and Weight Shapes
+        TensorShape weights_shape = TensorShape(kernel_size.width, kernel_size.height);
+
+        const TensorInfo in_info(input_shape, 1, data_type);
+        const TensorInfo we_info(weights_shape, 1, data_type);
+
+        const ConvolutionInfo info{ pad_stride, depth_multiplier, ActivationLayerInfo(), dilation };
+        const TensorShape     output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(in_info, we_info, info);
+
+        weights_shape.set(2, output_shape.z());
+        const TensorShape bias_shape = TensorShape(weights_shape[2]);
+
+        _data_type   = data_type;
+        _data_layout = data_layout;
+        _target      = compute_target(input_shape, weights_shape, bias_shape, dwc_conv2d_attr);
+        _reference   = compute_reference(input_shape, weights_shape, bias_shape, output_shape, dwc_conv2d_attr);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        switch(tensor.data_type())
+        {
+            case DataType::F16:
+            {
+                arm_compute::utils::uniform_real_distribution_16bit<half> distribution{ -1.0f, 1.0f };
+                library->fill(tensor, distribution, i);
+                break;
+            }
+            case DataType::F32:
+            {
+                std::uniform_real_distribution<float> distribution(-1.0f, 1.0f);
+                library->fill(tensor, distribution, i);
+                break;
+            }
+            default:
+                library->fill_tensor_uniform(tensor, i);
+        }
+    }
+
+    // Given input is in nchw format
+    TensorType compute_target(TensorShape input_shape, TensorShape weights_shape, const TensorShape &bias_shape, const DepthwiseConv2dAttributes dwc_conv2d_attr)
+    {
+        ARM_COMPUTE_ERROR_ON(_data_layout != DataLayout::NHWC);
+
+        // Our test shapes are assumed in NCHW data layout, thus the permutation
+        permute(input_shape, PermutationVector(2U, 0U, 1U));
+        permute(weights_shape, PermutationVector(2U, 0U, 1U));
+
+        // Create a new workload sketch
+        auto              cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
+        auto              gpu_ctx        = GpuWorkloadContext{ &cl_compile_ctx };
+        GpuWorkloadSketch sketch{ &gpu_ctx };
+
+        // Create sketch tensors
+        auto input_info  = sketch.create_tensor_info(TensorInfo(input_shape, 1, _data_type, _data_layout));
+        auto weight_info = sketch.create_tensor_info(TensorInfo(weights_shape, 1, _data_type, _data_layout));
+        auto bias_info   = sketch.create_tensor_info(TensorInfo(bias_shape, 1, _data_type, _data_layout));
+        auto dst_info    = sketch.create_tensor_info();
+        FunctionType::create_op(sketch, &input_info, &weight_info, &bias_info, &dst_info, dwc_conv2d_attr);
+
+        // Configure runtime
+        ClWorkloadRuntime runtime;
+        runtime.configure(sketch);
+
+        // (Important) Allocate auxiliary tensor memory if there are any
+        for(auto &data : runtime.get_auxiliary_tensors())
+        {
+            auto       tensor      = data.first;
+            const auto aux_mem_req = data.second;
+            tensor->allocator()->init(*data.first->info(), aux_mem_req.alignment);
+            tensor->allocator()->allocate();
+        }
+
+        // Construct user tensors
+        TensorType t_input{};
+        TensorType t_weight{};
+        TensorType t_bias{};
+        TensorType t_dst{};
+
+        // Initialize user tensors
+        t_input.allocator()->init(input_info);
+        t_weight.allocator()->init(weight_info);
+        t_bias.allocator()->init(bias_info);
+        t_dst.allocator()->init(dst_info);
+
+        // Allocate and fill user tensors
+        t_input.allocator()->allocate();
+        t_weight.allocator()->allocate();
+        t_bias.allocator()->allocate();
+        t_dst.allocator()->allocate();
+
+        fill(AccessorType(t_input), 0);
+        fill(AccessorType(t_weight), 1);
+        fill(AccessorType(t_bias), 2);
+
+        // Run runtime
+        runtime.run({ &t_input, &t_weight, &t_bias, &t_dst });
+        return t_dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape,
+                                      const TensorShape &output_shape, DepthwiseConv2dAttributes dwc_conv2d_attr)
+    {
+        // Create reference
+        SimpleTensor<T>     src{ input_shape, _data_type, 1 };
+        SimpleTensor<T>     weight{ weights_shape, _data_type, 1 };
+        SimpleTensor<TBias> bias{ bias_shape, _data_type, 1 };
+
+        fill(src, 0);
+        fill(weight, 1);
+        fill(bias, 2);
+
+        auto src_nchw          = src;
+        auto weights_nchw      = weight;
+        auto bias_nchw         = bias;
+        auto output_shape_nchw = output_shape;
+
+        PadStrideInfo legacy_pad_stride(dwc_conv2d_attr.stride().x(), dwc_conv2d_attr.stride().y(), dwc_conv2d_attr.pad().left, dwc_conv2d_attr.pad().right, dwc_conv2d_attr.pad().top,
+                                        dwc_conv2d_attr.pad().bottom,
+                                        DimensionRoundingType{});
+        auto dst_nchw = reference::depthwise_convolution(src_nchw, weights_nchw, bias_nchw, output_shape_nchw, legacy_pad_stride, dwc_conv2d_attr.depth_multiplier(), dwc_conv2d_attr.dilation());
+        return dst_nchw;
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+    DataType        _data_type{};
+    DataLayout      _data_layout{};
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class DynamicFusionGpuDepthwiseConv2dValidationFixture : public DynamicFusionGpuDepthwiseConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, Size2D kernel_size, const PadStrideInfo &info, const Size2D &dilation, const unsigned int depth_multiplier, DataType data_type, DataLayout data_layout)
+    {
+        DynamicFusionGpuDepthwiseConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, kernel_size, info, dilation,
+                                                                                                                  depth_multiplier, data_type, data_layout);
+    }
+};
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+#endif /* TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DEPTHWISECONV2DFIXTURE */