COMPMID-838 Implement CLPermute

Change-Id: I6d97b649f1ebc289c9e6f8949e67740a6b3cbcb2
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/116636
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

5 files changed, 271 insertions, 3 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 352b89baa5..22a328bcaf 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2018 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -291,6 +291,9 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "NV21_to_RGB888_bt709", "color_convert.cl" },
     { "NV21_to_RGBA8888_bt709", "color_convert.cl" },
     { "NV21_to_YUV444_bt709", "color_convert.cl" },
+    { "permute_201", "permute.cl" },
+    { "permute_120", "permute.cl" },
+    { "permute_3201", "permute.cl" },
     { "pixelwise_mul_float", "pixelwise_mul_float.cl" },
     { "pixelwise_mul_int", "pixelwise_mul_int.cl" },
     { "pooling_layer_2", "pooling_layer.cl" },
@@ -574,6 +577,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/optical_flow_pyramid_lk.clembed"
     },
     {
+        "permute.cl",
+#include "./cl_kernels/permute.clembed"
+    },
+    {
         "pixelwise_mul_float.cl",
 #include "./cl_kernels/pixelwise_mul_float.clembed"
     },
diff --git a/src/core/CL/ICLKernel.cpp b/src/core/CL/ICLKernel.cpp
index 3eb94b7ddc..ee02129bba 100644
--- a/src/core/CL/ICLKernel.cpp
+++ b/src/core/CL/ICLKernel.cpp
@@ -184,6 +184,12 @@ size_t ICLKernel::get_max_workgroup_size()
 
 cl::NDRange ICLKernel::gws_from_window(const Window &window)
 {
+    // Make sure that dimensions > Z are 1
+    for(unsigned int i = 3; i < Coordinates::num_max_dimensions; ++i)
+    {
+        ARM_COMPUTE_ERROR_ON((window[i].end() - window[i].start()) != 1);
+    }
+
     if((window.x().end() - window.x().start()) == 0 || (window.y().end() - window.y().start()) == 0)
     {
         return cl::NullRange;
diff --git a/src/core/CL/cl_kernels/helpers.h b/src/core/CL/cl_kernels/helpers.h
index 768f7ee434..02c6c4cc41 100644
--- a/src/core/CL/cl_kernels/helpers.h
+++ b/src/core/CL/cl_kernels/helpers.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -123,7 +123,7 @@
 
 #define CONVERT_TO_TENSOR4D_STRUCT(name, mod_size)                                                                                                 \
     update_tensor4D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
-                                 name##_stride_z, name##_step_z, name##_stride_w, name##_step_z, mod_size)
+                                 name##_stride_z, name##_step_z, name##_stride_w, name##_step_w, mod_size)
 
 #define CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(name, mod_size) \
     update_tensor4D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, 0, name##_stride_w, 0, mod_size)
diff --git a/src/core/CL/cl_kernels/permute.cl b/src/core/CL/cl_kernels/permute.cl
new file mode 100644
index 0000000000..6f978c9b70
--- /dev/null
+++ b/src/core/CL/cl_kernels/permute.cl
@@ -0,0 +1,128 @@
+/*
+ * 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"
+
+#if defined(DATA_TYPE) && defined(DEPTH_IN)
+/** Perform a DCHW -> DHWC permute operation on an input tensor.
+ *
+ * @attention Data type can be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float
+ * @attention Input tensor depth should be given as a preprocessor argument using -DDEPTH_IN=size. e.g. -DDEPTH_IN=16
+ *
+ * @param[in]  input_ptr                            Pointer to the source image. Supported data types: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32
+ * @param[in]  input_stride_x                       Stride of the source image in X dimension (in bytes)
+ * @param[in]  input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  input_stride_y                       Stride of the source image in Y dimension (in bytes)
+ * @param[in]  input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  input_offset_first_element_in_bytes  The offset of the first element in the source image
+ * @param[out] output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      Stride of the destination image in X dimension (in bytes)
+ * @param[in]  output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      Stride of the destination image in Y dimension (in bytes)
+ * @param[in]  output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_stride_w                      Stride of the source tensor in W dimension (in bytes)
+ * @param[in]  output_step_w                        output_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes The offset of the first element in the destination image
+ */
+__kernel void permute_201(
+    TENSOR4D_DECLARATION(input),
+    TENSOR4D_DECLARATION(output))
+{
+    Tensor4D in  = CONVERT_TO_TENSOR4D_STRUCT(input, DEPTH_IN);
+    Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0);
+
+    *((__global DATA_TYPE *)tensor4D_offset(&out, (get_global_id(2) % DEPTH_IN), get_global_id(0), get_global_id(1), (get_global_id(2) / DEPTH_IN))) = *((__global DATA_TYPE *)in.ptr);
+}
+
+/** Perform a DCHW -> DWCH permute operation on an input tensor.
+ *
+ * @attention Data type can be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float
+ * @attention Input tensor depth should be given as a preprocessor argument using -DDEPTH_IN=size. e.g. -DDEPTH_IN=16
+ *
+ * @param[in]  input_ptr                            Pointer to the source image. Supported data types: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32
+ * @param[in]  input_stride_x                       Stride of the source image in X dimension (in bytes)
+ * @param[in]  input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  input_stride_y                       Stride of the source image in Y dimension (in bytes)
+ * @param[in]  input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  input_offset_first_element_in_bytes  The offset of the first element in the source image
+ * @param[out] output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      Stride of the destination image in X dimension (in bytes)
+ * @param[in]  output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      Stride of the destination image in Y dimension (in bytes)
+ * @param[in]  output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_stride_w                      Stride of the source tensor in W dimension (in bytes)
+ * @param[in]  output_step_w                        output_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes The offset of the first element in the destination image
+ */
+__kernel void permute_120(
+    TENSOR4D_DECLARATION(input),
+    TENSOR4D_DECLARATION(output))
+{
+    Tensor4D in  = CONVERT_TO_TENSOR4D_STRUCT(input, DEPTH_IN);
+    Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0);
+
+    *((__global DATA_TYPE *)tensor4D_offset(&out, get_global_id(1), (get_global_id(2) % DEPTH_IN), get_global_id(0), (get_global_id(2) / DEPTH_IN))) = *((__global DATA_TYPE *)in.ptr);
+}
+
+/** Perform a DCHW -> HWCD permute operation on an input tensor.
+ *
+ * @attention Data type can be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float
+ * @attention Input tensor depth should be given as a preprocessor argument using -DDEPTH_IN=size. e.g. -DDEPTH_IN=16
+ *
+ * @param[in]  input_ptr                            Pointer to the source image. Supported data types: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32
+ * @param[in]  input_stride_x                       Stride of the source image in X dimension (in bytes)
+ * @param[in]  input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  input_stride_y                       Stride of the source image in Y dimension (in bytes)
+ * @param[in]  input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  input_offset_first_element_in_bytes  The offset of the first element in the source image
+ * @param[out] output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      Stride of the destination image in X dimension (in bytes)
+ * @param[in]  output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      Stride of the destination image in Y dimension (in bytes)
+ * @param[in]  output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_stride_w                      Stride of the source tensor in W dimension (in bytes)
+ * @param[in]  output_step_w                        output_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes The offset of the first element in the destination image
+ */
+__kernel void permute_3201(
+    TENSOR4D_DECLARATION(input),
+    TENSOR4D_DECLARATION(output))
+{
+    Tensor4D in  = CONVERT_TO_TENSOR4D_STRUCT(input, DEPTH_IN);
+    Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0);
+
+    *((__global DATA_TYPE *)tensor4D_offset(&out, (get_global_id(2) / DEPTH_IN), (get_global_id(2) % DEPTH_IN), get_global_id(0), get_global_id(1))) = *((__global DATA_TYPE *)in.ptr);
+}
+#endif // defined(DATA_TYPE) && defined(DEPTH_IN)
diff --git a/src/core/CL/kernels/CLPermuteKernel.cpp b/src/core/CL/kernels/CLPermuteKernel.cpp
new file mode 100644
index 0000000000..7001c5565b
--- /dev/null
+++ b/src/core/CL/kernels/CLPermuteKernel.cpp
@@ -0,0 +1,127 @@
+/*
+ * 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/CLPermuteKernel.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/CL/OpenCL.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Types.h"
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+
+CLPermuteKernel::CLPermuteKernel()
+    : _input(nullptr), _output(nullptr), _perm()
+{
+}
+namespace
+{
+TensorShape get_output_shape(const ITensorInfo *input, const PermutationVector &perm)
+{
+    TensorShape output_shape = input->tensor_shape();
+    permute(output_shape, perm);
+    return output_shape;
+}
+}
+void CLPermuteKernel::configure(const ICLTensor *input, ICLTensor *output, const PermutationVector &perm)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::QASYMM8,
+                                                  DataType::U16, DataType::S16, DataType::QS16,
+                                                  DataType::U32, DataType::S32,
+                                                  DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MSG(input->info()->num_dimensions() < 3, "Invalid input size!");
+    ARM_COMPUTE_ERROR_ON_MSG(
+        (perm.num_dimensions() != 3 && ((perm[0] != 2 && perm[1] != 0 && perm[2] != 1) || (perm[0] != 1 && perm[1] != 2 && perm[2] != 0))) && (perm.num_dimensions() != 4 && ((perm[0] != 2 && perm[1] != 0
+                && perm[2] != 1)
+                || (perm[0] != 1 && perm[1] != 2 && perm[2] != 0))),
+        "Only [2, 0, 1],[1, 2, 0] and [3, 2, 0, 1] permutation is supported");
+
+    _input  = input;
+    _output = output;
+    _perm   = perm;
+
+    const TensorShape output_shape = get_output_shape(input->info(), perm);
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
+
+    // Create kernel
+    std::set<std::string> build_opts;
+
+    build_opts.emplace("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+    build_opts.emplace("-DDEPTH_IN=" + support::cpp11::to_string(input->info()->dimension(2)));
+
+    // Run [2, 0, 1] permute
+    if(_perm[0] == 2 && _perm[1] == 0 && _perm[2] == 1)
+    {
+        _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("permute_201", build_opts));
+    }
+    // Run [1, 2, 0] permute
+    else if(_perm[0] == 1 && _perm[1] == 2 && _perm[2] == 0)
+    {
+        _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("permute_120", build_opts));
+    }
+    // Run [3, 2, 0, 1] permute
+    else if(_perm[0] == 3 && _perm[1] == 2 && _perm[2] == 0 && _perm[3] == 1)
+    {
+        _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("permute_3201", build_opts));
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Not supported.");
+    }
+
+    // Configure  kernel window
+    Window win = calculate_max_window(*input->info(), Steps());
+
+    ICLKernel::configure(win);
+}
+
+void CLPermuteKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
+
+    Window slice_in = window.first_slice_window_4D();
+    Window slice_out(slice_in);
+
+    // Setup output slice
+    slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
+    slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
+    slice_out.set(Window::DimZ, Window::Dimension(0, 0, 0));
+    slice_out.set(3, Window::Dimension(0, 0, 0));
+
+    do
+    {
+        auto         collapsed_slice_in  = slice_in.collapse(ICLKernel::window(), 2);
+        auto         collapsed_slice_out = slice_out.collapse(ICLKernel::window(), 2);
+        unsigned int idx                 = 0;
+        add_4D_tensor_argument(idx, _input, collapsed_slice_in);
+        add_4D_tensor_argument(idx, _output, collapsed_slice_out);
+        enqueue(queue, *this, collapsed_slice_in);
+    }
+    while(window.slide_window_slice_4D(slice_in) && window.slide_window_slice_4D(slice_out));
+}