Implement address precalculation for indirect conv2d - OpenCL

- Implement kernel (ClIndirectConv2dAddressPrecalculationKernel)
- Implement OpenCL kernel (indirect_convolution.cl)
- Add test

Resolves COMPMID-5708

Change-Id: If7408e37cbc6f9ad8506ff3334bc574e5d6763fb
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8661
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 355 insertions, 0 deletions

diff --git a/tests/validation/CL/IndirectConv2dAddressPrecalculation.cpp b/tests/validation/CL/IndirectConv2dAddressPrecalculation.cpp
new file mode 100644
index 0000000000..67f70685d1
--- /dev/null
+++ b/tests/validation/CL/IndirectConv2dAddressPrecalculation.cpp
@@ -0,0 +1,89 @@
+/*
+ * 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/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/CLTensorAllocator.h"
+#include "src/gpu/cl/kernels/ClIndirectConv2dAddressPrecalculationKernel.h"
+#include "tests/CL/CLAccessor.h"
+#include "tests/CL/Helper.h"
+#include "tests/PaddingCalculator.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/IndirectConv2dAddressPrecalculationFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+using namespace arm_compute::misc::shape_calculator;
+using namespace arm_compute::opencl::kernels;
+
+using CLIndirectConv2dAddressPrecalculation = CLSynthetizeOperator<ClIndirectConv2dAddressPrecalculationKernel>;
+
+using CLIndirectConv2dAddressPrecalculationFixture = IndirectConv2dAddressPrecalculationValidationFixture<CLTensor, CLAccessor, CLIndirectConv2dAddressPrecalculation>;
+
+// *INDENT-OFF*
+// clang-format off
+/** Data types */
+
+namespace
+{
+const auto src_w_values  = framework::dataset::make("src_w", {91});
+const auto src_h_values  = framework::dataset::make("src_h", {103});
+const auto src_b_values  = framework::dataset::make("src_b", {1, 2});
+const auto wei_w_values  = framework::dataset::make("wei_w", {3, 5});
+const auto wei_h_values  = framework::dataset::make("wei_h", {1, 6});
+const auto pad_values    = framework::dataset::make("pad", {1, 2, 3});
+const auto stride_values = framework::dataset::make("stride", {1, 2});
+const auto m0_values     = framework::dataset::make("M0", { 1, 2, 4, 5, 7 });
+} // namespace
+
+TEST_SUITE(CL)
+TEST_SUITE(IndirectConv2dAddressPrecalculation)
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLIndirectConv2dAddressPrecalculationFixture, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(src_w_values,
+                                                                        src_h_values),
+                                                                        src_b_values),
+                                                                        wei_w_values),
+                                                                        wei_h_values),
+                                                                        pad_values),
+                                                                        stride_values),
+                                                                        m0_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+TEST_SUITE_END() // IndirectConv2dAddressPrecalculation
+TEST_SUITE_END() // CL
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/fixtures/IndirectConv2dAddressPrecalculationFixture.h b/tests/validation/fixtures/IndirectConv2dAddressPrecalculationFixture.h
new file mode 100644
index 0000000000..f595a8663d
--- /dev/null
+++ b/tests/validation/fixtures/IndirectConv2dAddressPrecalculationFixture.h
@@ -0,0 +1,123 @@
+/*
+ * 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 ARM_COMPUTE_TEST_INDIRECT_CONV2D_ADDRESS_PRECALCULATION_FIXTURE
+#define ARM_COMPUTE_TEST_INDIRECT_CONV2D_ADDRESS_PRECALCULATION_FIXTURE
+
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "tests/Globals.h"
+#include "tests/framework/Fixture.h"
+#include "tests/validation/Helpers.h"
+#include "tests/validation/reference/IndirectConv2dAddressPrecalculation.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+using namespace arm_compute::misc::shape_calculator;
+
+template <typename TensorType, typename AccessorType, typename OperatorType>
+class IndirectConv2dAddressPrecalculationValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(unsigned int src_w,
+               unsigned int src_h,
+               unsigned int src_b,
+               unsigned int wei_w,
+               unsigned int wei_h,
+               unsigned int pad,
+               unsigned int stride,
+               unsigned int m0)
+    {
+        DirectConvComputeKernelInfo desc;
+        desc.m0                         = m0;
+        desc.n0                         = 1;     // Not used by the kernel
+        desc.k0                         = 1;     // Not used by the kernel
+        desc.export_weights_to_cl_image = false; // Not used by the kernel
+
+        const PadStrideInfo conv_info(stride, stride, pad, pad);
+
+        const TensorShape shape_conv_src(23, // The input channels are not used by the kernel
+                                         src_w,
+                                         src_h,
+                                         src_b);
+
+        const TensorShape shape_conv_wei(23, // The input channels are not used by the kernel
+                                         wei_w,
+                                         wei_h,
+                                         23 // The output channels are not used by the kernel
+                                        );
+
+        // The result of the kernel does not change with the datatype. Hence, we can fix it to Fp16 for validation purposes
+        const DataType data_type = DataType::F16;
+
+        _target    = compute_target(shape_conv_src, shape_conv_wei, data_type, conv_info, desc);
+        _reference = compute_reference(shape_conv_src, shape_conv_wei, data_type, conv_info, desc);
+    }
+
+protected:
+    TensorType compute_target(TensorShape shape_conv_src, TensorShape shape_conv_wei, DataType data_type, const PadStrideInfo &conv_info, const DirectConvComputeKernelInfo &desc)
+    {
+        TensorInfo src_conv_info(shape_conv_src, 1, data_type, DataLayout::NHWC);
+        TensorInfo wei_conv_info(shape_conv_wei, 1, data_type, DataLayout::NHWC);
+        TensorType dst;
+
+        // The output tensor will be auto-initialized within the function
+
+        // Create and configure function
+        OperatorType func;
+        func.configure(&src_conv_info, &wei_conv_info, dst.info(), conv_info, desc);
+
+        add_padding_x({ &dst });
+
+        // Allocate tensors
+        dst.allocator()->allocate();
+
+        // Compute GEMM LHS matrix reshape function
+        ITensorPack tensors = { { ACL_DST, &dst } };
+        func.run(tensors);
+
+        return dst;
+    }
+
+    SimpleTensor<int32_t> compute_reference(TensorShape shape_conv_src, TensorShape shape_conv_wei, DataType data_type, const PadStrideInfo &conv_info, const DirectConvComputeKernelInfo &desc)
+    {
+        ARM_COMPUTE_UNUSED(data_type);
+        TensorShape shape_out         = compute_indirect_buffer_shape(shape_conv_src, DataLayout::NHWC, shape_conv_wei, conv_info, desc);
+        TensorShape output_conv_shape = compute_deep_convolution_shape(shape_conv_src, DataLayout::NHWC, shape_conv_wei, conv_info);
+
+        return reference::indirect_conv2d_addr_precalculation(shape_conv_src, shape_conv_wei, output_conv_shape, shape_out, conv_info);
+    }
+
+    TensorType            _target{};
+    SimpleTensor<int32_t> _reference{};
+};
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_INDIRECT_CONV2D_ADDRESS_PRECALCULATION_FIXTURE */
\ No newline at end of file
diff --git a/tests/validation/reference/IndirectConv2dAddressPrecalculation.cpp b/tests/validation/reference/IndirectConv2dAddressPrecalculation.cpp
new file mode 100644
index 0000000000..240dcb686c
--- /dev/null
+++ b/tests/validation/reference/IndirectConv2dAddressPrecalculation.cpp
@@ -0,0 +1,99 @@
+/*
+ * 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 "IndirectConv2dAddressPrecalculation.h"
+
+#include "arm_compute/core/Types.h"
+
+#include "tests/validation/Helpers.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace reference
+{
+SimpleTensor<int32_t> indirect_conv2d_addr_precalculation(const TensorShape &shape_conv_src, const TensorShape &shape_conv_wei, const TensorShape &shape_conv_dst, const TensorShape &shape_dst,
+                                                          const PadStrideInfo &conv_info)
+{
+    SimpleTensor<int32_t> out{ shape_dst, DataType::S32 };
+
+    constexpr unsigned int width_idx = 1;
+    constexpr unsigned int heigh_idx = 2;
+
+    const int src_conv_width  = static_cast<int32_t>(shape_conv_src[width_idx]); // NHWC
+    const int src_conv_height = static_cast<int32_t>(shape_conv_src[heigh_idx]); // NHWC
+    const int dst_conv_width  = static_cast<int32_t>(shape_conv_dst[width_idx]); // NHWC
+    const int wei_conv_width  = static_cast<int32_t>(shape_conv_wei[width_idx]); // NHWC
+    const int wei_conv_height = static_cast<int32_t>(shape_conv_wei[heigh_idx]); // NHWC
+    const int dst_width       = static_cast<int32_t>(shape_dst[0]);
+    const int dst_height      = static_cast<int32_t>(shape_dst[1]);
+    const int dst_batch       = static_cast<int32_t>(shape_dst[2]);
+    const int ks              = wei_conv_width * wei_conv_height;
+    const int stride_x        = static_cast<int32_t>(conv_info.stride().first);
+    const int stride_y        = static_cast<int32_t>(conv_info.stride().second);
+    const int pad_left        = static_cast<int32_t>(conv_info.pad_left());
+    const int pad_top         = static_cast<int32_t>(conv_info.pad_top());
+
+    const int m0 = dst_width / ks;
+
+    for(int z = 0; z < dst_batch; ++z)
+    {
+        for(int y = 0; y < dst_height; ++y)
+        {
+            const int mout = y * m0;
+            for(int ki = 0; ki < ks; ++ki)
+            {
+                const int xk = ki % wei_conv_width;
+                const int yk = ki / wei_conv_width;
+                for(int mi = 0; mi < m0; ++mi)
+                {
+                    int xi = ((mout + mi) % dst_conv_width) * stride_x;
+                    int yi = ((mout + mi) / dst_conv_width) * stride_y;
+                    xi -= pad_left;
+                    yi -= pad_top;
+                    const int x_s = xi + xk;
+                    const int y_s = yi + yk;
+                    int       my  = x_s + y_s * src_conv_width;
+                    my            = my + z * src_conv_width * src_conv_height;
+                    my            = x_s >= 0 ? my : -1;
+                    my            = x_s < src_conv_width ? my : -1;
+                    my            = y_s >= 0 ? my : -1;
+                    my            = y_s < src_conv_height ? my : -1;
+
+                    const unsigned int addr_out = mi + ki * m0 + y * (dst_width) + z * (dst_width * dst_height);
+
+                    out[addr_out] = my;
+                }
+            }
+        }
+    }
+
+    return out;
+}
+} // namespace reference
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
\ No newline at end of file
diff --git a/tests/validation/reference/IndirectConv2dAddressPrecalculation.h b/tests/validation/reference/IndirectConv2dAddressPrecalculation.h
new file mode 100644
index 0000000000..f4a90dfd9f
--- /dev/null
+++ b/tests/validation/reference/IndirectConv2dAddressPrecalculation.h
@@ -0,0 +1,44 @@
+/*
+ * 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 ARM_COMPUTE_TEST_INDIRECT_CONV2D_ADDRESS_PRECALCULATION_H
+#define ARM_COMPUTE_TEST_INDIRECT_CONV2D_ADDRESS_PRECALCULATION_H
+
+#include "tests/SimpleTensor.h"
+#include "tests/validation/Helpers.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace reference
+{
+SimpleTensor<int32_t> indirect_conv2d_addr_precalculation(const TensorShape &shape_conv_src, const TensorShape &shape_conv_wei, const TensorShape &shape_conv_out, const TensorShape &shape_out,
+                                                          const PadStrideInfo &conv_info);
+} // namespace reference
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_INDIRECT_CONV2D_ADDRESS_PRECALCULATION_H */
\ No newline at end of file