COMPMID-1719 CL: Implement RSqrt, Exp

Change-Id: I827b26239043a9e90d26c2583122648d2a45303a
Reviewed-on: https://review.mlplatform.org/317
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 263 insertions, 0 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 1361d02f74..da85472005 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -253,6 +253,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "elementwise_operation_MIN_quantized", "elementwise_operation_quantized.cl" },
     { "elementwise_operation_DIV_quantized", "elementwise_operation_quantized.cl" },
     { "elementwise_operation_SQUARED_DIFF_quantized", "elementwise_operation_quantized.cl" },
+    { "elementwise_unary", "elementwise_unary.cl" },
     { "erode", "erode.cl" },
     { "fast_corners", "fast_corners.cl" },
     { "flatten", "flatten.cl" },
@@ -651,6 +652,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/elementwise_operation_quantized.clembed"
     },
     {
+        "elementwise_unary.cl",
+#include "./cl_kernels/elementwise_unary.clembed"
+    },
+    {
         "erode.cl",
 #include "./cl_kernels/erode.clembed"
     },
diff --git a/src/core/CL/cl_kernels/elementwise_unary.cl b/src/core/CL/cl_kernels/elementwise_unary.cl
new file mode 100644
index 0000000000..bccb47ed1f
--- /dev/null
+++ b/src/core/CL/cl_kernels/elementwise_unary.cl
@@ -0,0 +1,87 @@
+/*
+ * 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.
+ */
+#include "helpers.h"
+#include "warp_helpers.h"
+
+#if defined(DATA_TYPE) && defined(VEC_SIZE) && defined(OPERATION)
+/** Calculate reverse square root
+ *
+ * @param[in] input Pointer to the first element.
+ *
+ * @return reverse square root
+ */
+inline VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) inverse_sqrt(const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) input)
+{
+    return rsqrt(input);
+}
+
+/** Calculate exponential
+ *
+ * @param[in] input Pointer to the first element.
+ *
+ * @return exponential
+ */
+inline VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) exponential(const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) input)
+{
+    return exp(input);
+}
+
+/** Applies element wise unary operator in a tensor.
+ *
+ * @param[in]  in_ptr                            Pointer to the source image. Supported data types: F16/32.
+ * @param[in]  in_stride_x                       Stride of the source image in X dimension (in bytes)
+ * @param[in]  in_step_x                         in_stride_x * number of elements along X processed per work item (in bytes)
+ * @param[in]  in_offset_first_element_in_bytes  Offset of the first element in the source image
+ * @param[out] out_ptr                           Pointer to the destination image. Supported data types: F16/32.
+ * @param[in]  out_stride_x                      Stride of the destination image in X dimension (in bytes)
+ * @param[in]  out_step_y                        out_stride_y * number of elements along Y processed per work item (in bytes)
+ * @param[in]  out_offset_first_element_in_bytes Offset of the first element in the destination image
+ */
+__kernel void elementwise_unary(
+    VECTOR_DECLARATION(in),
+    VECTOR_DECLARATION(out))
+{
+    Vector in  = CONVERT_TO_VECTOR_STRUCT(in);
+    Vector out = CONVERT_TO_VECTOR_STRUCT(out);
+
+#if defined(VEC_SIZE) && defined(LAST_ACCESSED_X)
+    // Check if access on width gets out of bounds
+    // If it does shift access vector to access elements within bounds
+    const int xi = (int)(get_global_id(0) * VEC_SIZE);
+    in.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * in_stride_x;
+    out.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * out_stride_x;
+
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in.ptr);
+
+    VSTORE(VEC_SIZE)
+    (OPERATION(data), 0, (__global DATA_TYPE *)out.ptr);
+#else  // !defined(VEC_SIZE) || !defined(LAST_ACCESSED_X)
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in.ptr);
+    VSTORE(VEC_SIZE)
+    (OPERATION(data), 0, (__global DATA_TYPE *)out.ptr);
+#endif // defined(VEC_SIZE) && defined(LAST_ACCESSED_X)
+}
+#endif // defined(DATA_TYPE) && defined(VEC_SIZE) && defined(OPERATION)
diff --git a/src/core/CL/kernels/CLElementWiseUnaryLayerKernel.cpp b/src/core/CL/kernels/CLElementWiseUnaryLayerKernel.cpp
new file mode 100644
index 0000000000..6d2105f3b1
--- /dev/null
+++ b/src/core/CL/kernels/CLElementWiseUnaryLayerKernel.cpp
@@ -0,0 +1,117 @@
+/*
+ * 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.
+ */
+#include "arm_compute/core/CL/kernels/CLElementWiseUnaryLayerKernel.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLValidate.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+
+using namespace arm_compute;
+
+namespace
+{
+Status validate_arguments(const ITensorInfo &input, const ITensorInfo &output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input, 1, DataType::F16, DataType::F32);
+
+    // Validate in case of configured output
+    if(output.total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&output);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::F16, DataType::F32);
+    }
+
+    return Status{};
+}
+} // namespace
+
+void CLElementWiseUnaryLayerKernel::configure(const ICLTensor *input, ICLTensor *output, const ElementWiseUnary &op)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input->info(), *output->info()));
+
+    // Configure kernel window
+    _input  = input;
+    _output = output;
+
+    const std::string kernel_name    = "elementwise_unary";
+    const int         vec_size_x     = 16 / output->info()->element_size();
+    const int         output_width_x = output->info()->tensor_shape().x();
+    const bool        multi_access_x = (output_width_x / vec_size_x > 0);
+
+    Window win = calculate_max_window(*output->info());
+    if(multi_access_x)
+    {
+        win.set(Window::DimX,
+                Window::Dimension(win.x().start(), ceil_to_multiple(win.x().end(), vec_size_x), vec_size_x));
+    }
+    ICLKernel::configure_internal(win);
+
+    // Set kernel build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+    build_opts.add_option_if(multi_access_x, "-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x));
+    build_opts.add_option_if(multi_access_x, "-DLAST_ACCESSED_X=" + support::cpp11::to_string(std::max<int>(output_width_x - vec_size_x, 0)));
+    switch(op)
+    {
+        case ElementWiseUnary::RSQRT:
+            build_opts.add_option("-DOPERATION=inverse_sqrt");
+            break;
+        case ElementWiseUnary::EXP:
+            build_opts.add_option("-DOPERATION=exponential");
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Not implemented");
+    }
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
+}
+
+Status CLElementWiseUnaryLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ElementWiseUnary &op)
+{
+    ARM_COMPUTE_UNUSED(op);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*input, *output));
+
+    return Status{};
+}
+
+void CLElementWiseUnaryLayerKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimX);
+
+    do
+    {
+        unsigned int idx = 0;
+        add_1D_tensor_argument(idx, _input, collapsed);
+        add_1D_tensor_argument(idx, _output, collapsed);
+        enqueue(queue, *this, collapsed);
+    }
+    while(window.slide_window_slice_1D(collapsed));
+}
\ No newline at end of file
diff --git a/src/runtime/CL/functions/CLElementWiseUnaryLayer.cpp b/src/runtime/CL/functions/CLElementWiseUnaryLayer.cpp
new file mode 100644
index 0000000000..b7e9a68680
--- /dev/null
+++ b/src/runtime/CL/functions/CLElementWiseUnaryLayer.cpp
@@ -0,0 +1,54 @@
+/*
+ * 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.
+ */
+#include "arm_compute/runtime/CL/functions/CLElementWiseUnaryLayer.h"
+
+#include "arm_compute/core/CL/kernels/CLElementWiseUnaryLayerKernel.h"
+#include "support/ToolchainSupport.h"
+
+#include <utility>
+
+namespace arm_compute
+{
+void CLRsqrtLayer::configure(const ICLTensor *input, ICLTensor *output)
+{
+    auto k = arm_compute::support::cpp14::make_unique<CLElementWiseUnaryLayerKernel>();
+    k->configure(input, output, ElementWiseUnary::RSQRT);
+    _kernel = std::move(k);
+}
+Status CLRsqrtLayer::validate(const ITensorInfo *input, const ITensorInfo *output)
+{
+    return CLElementWiseUnaryLayerKernel::validate(input, output, ElementWiseUnary::RSQRT);
+}
+
+void CLExpLayer::configure(const ICLTensor *input, ICLTensor *output)
+{
+    auto k = arm_compute::support::cpp14::make_unique<CLElementWiseUnaryLayerKernel>();
+    k->configure(input, output, ElementWiseUnary::EXP);
+    _kernel = std::move(k);
+}
+Status CLExpLayer::validate(const ITensorInfo *input, const ITensorInfo *output)
+{
+    return CLElementWiseUnaryLayerKernel::validate(input, output, ElementWiseUnary::EXP);
+}
+} // namespace arm_compute