Implementation of Permute CL kernel to handle all permutations

This patch will add a generic permute cl-kernel to handle
all permutations available for tensors having rank upto 4.

Change-Id: I50eb555d9d45d5ad5f7fa9b0a3862dd17551d458
Signed-off-by: shubham <shub98.gupta@samsung.com>
Reviewed-on: https://review.mlplatform.org/449
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>

3 files changed, 40 insertions, 106 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 2bc2d06827..905a34a509 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -368,9 +368,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "NV21_to_RGBA8888_bt709", "color_convert.cl" },
     { "NV21_to_YUV444_bt709", "color_convert.cl" },
     { "output_stage_quantized", "direct_convolution_1x1_3x3_5x5_quantized.cl" },
-    { "permute_201", "permute.cl" },
-    { "permute_120", "permute.cl" },
-    { "permute_3201", "permute.cl" },
+    { "permute", "permute.cl" },
     { "pixelwise_mul_float", "pixelwise_mul_float.cl" },
     { "pixelwise_mul_int", "pixelwise_mul_int.cl" },
     { "pixelwise_mul_quantized", "pixelwise_mul_int.cl" },
diff --git a/src/core/CL/cl_kernels/permute.cl b/src/core/CL/cl_kernels/permute.cl
index 03fc15e4e8..77f03f7d5b 100644
--- a/src/core/CL/cl_kernels/permute.cl
+++ b/src/core/CL/cl_kernels/permute.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,11 +23,12 @@
  */
 #include "helpers.h"
 
-#if defined(DATA_TYPE) && defined(DEPTH_IN)
-/** Perform a DCHW -> DHWC permute operation on an input tensor.
+#if defined(DATA_TYPE) && defined(DEPTH_IN) && defined(P1) && defined(P2) && defined(P3) && defined(P4)
+/**Perform a permute operation on an input tensor of Shape DCHW.
  *
  * @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
+ * @attention Permutation vector is passed as a preprocessor arguement using -DP1, -DP2, -DP3 and -DP4=int, e.g. -DP1=2, -DP2=1, -DP3=0 and -DP4=3.
  *
  * @param[in]  input_ptr                            Pointer to the source image. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
  * @param[in]  input_stride_x                       Stride of the source image in X dimension (in bytes)
@@ -48,81 +49,26 @@
  * @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);
+__kernel void permute(TENSOR4D_DECLARATION(input),
+                      TENSOR4D_DECLARATION(output))
 
-    *((__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/QASYMM8/U16/S16/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);
-}
+    int out_index[4] = { 0 };
+    int in_index[4]  = { 0 };
 
-/** 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/QASYMM8/U16/S16/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);
+    in_index[0] = get_global_id(0);            // W
+    in_index[1] = get_global_id(1);            // H
+    in_index[2] = get_global_id(2) % DEPTH_IN; // C
+    in_index[3] = get_global_id(2) / DEPTH_IN; // B
+
+    out_index[0] = in_index[P1];
+    out_index[1] = in_index[P2];
+    out_index[2] = in_index[P3];
+    out_index[3] = in_index[P4];
 
-    *((__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);
+    *((__global DATA_TYPE *)tensor4D_offset(&out, out_index[0], out_index[1], out_index[2], out_index[3])) = *((__global DATA_TYPE *)in.ptr);
 }
-#endif // defined(DATA_TYPE) && defined(DEPTH_IN)
+#endif // defined(DATA_TYPE) && defined(DEPTH_IN) && defined(P1) && defined(P2) && defined(P3) && defined(P4)
diff --git a/src/core/CL/kernels/CLPermuteKernel.cpp b/src/core/CL/kernels/CLPermuteKernel.cpp
index a9a2c5c97a..6c44199f7a 100644
--- a/src/core/CL/kernels/CLPermuteKernel.cpp
+++ b/src/core/CL/kernels/CLPermuteKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -56,16 +56,20 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c
                                                          DataType::U16, DataType::S16,
                                                          DataType::U32, DataType::S32,
                                                          DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG((perm != PermutationVector{ 2U, 0U, 1U })
-                                    && (perm != PermutationVector{ 1U, 2U, 0U })
-                                    && (perm != PermutationVector{ 3U, 2U, 0U, 1U }),
-                                    "Only [2, 0, 1], [1, 2, 0] and [3, 2, 0, 1] permutation is supported");
 
-    const TensorShape output_shape = misc::shape_calculator::compute_permutation_output_shape(*input, perm);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() < 1 || input->num_dimensions() > 4,
+                                    "Permutation upto 4-D input tensor is supported");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(perm.num_dimensions() < 1 || perm.num_dimensions() > 4,
+                                    "Permutation vector size should be less than or equal to 4");
+    for(const auto &p : perm)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(p >= perm.num_dimensions(), "Permutation vector has invalid values");
+    }
 
     // Validate configured output
     if(output->total_size() != 0)
     {
+        const TensorShape output_shape = misc::shape_calculator::compute_permutation_output_shape(*input, perm);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     }
@@ -87,30 +91,16 @@ void CLPermuteKernel::configure(const ICLTensor *input, ICLTensor *output, const
     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 == PermutationVector{ 2U, 0U, 1U })
-    {
-        _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("permute_201", build_opts));
-    }
-    // Run [1, 2, 0] permute
-    else if(_perm == PermutationVector{ 1U, 2U, 0U })
-    {
-        _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("permute_120", build_opts));
-    }
-    // Run [3, 2, 0, 1] permute
-    else if(_perm == PermutationVector{ 3U, 2U, 0U, 1U })
-    {
-        _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("permute_3201", build_opts));
-    }
-    else
-    {
-        ARM_COMPUTE_ERROR("Not supported.");
-    }
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+    build_opts.add_option("-DDEPTH_IN=" + support::cpp11::to_string(input->info()->dimension(2)));
+    // New positions of  width(W), height(H), channel(C) and batch(D) based on permutation vector
+    build_opts.add_option("-DP1=" + support::cpp11::to_string((_perm.num_dimensions() >= 1) ? perm[0] : 0));
+    build_opts.add_option("-DP2=" + support::cpp11::to_string((_perm.num_dimensions() >= 2) ? perm[1] : 1));
+    build_opts.add_option("-DP3=" + support::cpp11::to_string((_perm.num_dimensions() >= 3) ? perm[2] : 2));
+    build_opts.add_option("-DP4=" + support::cpp11::to_string((_perm.num_dimensions() >= 4) ? perm[3] : 3));
+
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("permute", build_opts.options()));
 
     // Configure  kernel window
     Window win = calculate_max_window(*input->info(), Steps());