Remove OpenCL padding CLFloorKernel

Use of proper vector size with boundary checking loads and stores

Resolves: COMPMID-3922

Change-Id: Ib631d499603b860fcfdbe3da903b866a125359a8
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4789
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 32 insertions, 38 deletions

diff --git a/src/core/CL/cl_kernels/floor.cl b/src/core/CL/cl_kernels/floor.cl
index 1988ba4e92..f6dd4edd2e 100644
--- a/src/core/CL/cl_kernels/floor.cl
+++ b/src/core/CL/cl_kernels/floor.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,10 +23,13 @@
  */
 #include "helpers.h"
 
+#if defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER)
+
 /** Perform a floor operation on an input tensor.
  *
- * @attention Data type can be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float
- * @attention Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
+ * @note Data type can be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float
+ * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
+ * @note The leftover size in the X dimension shoud be given as preprocessor argument using -DVEC_SIZE_LEFTOVER is; x_dimension % VEC_SIZE. e.g. -DVEC_SIZE_LEFTOVER=1
  * @note Can only take floating point data types.
  *
  * @param[in]  input_ptr                            Pointer to the source image. Supported data types: F16/F32
@@ -50,9 +53,16 @@ __kernel void floor_layer(
     TENSOR3D_DECLARATION(input),
     TENSOR3D_DECLARATION(output))
 {
-    Tensor3D input  = CONVERT_TO_TENSOR3D_STRUCT(input);
-    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
+    // Offset computation
+    const uint x_offs = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0);
+
+    // Address computation
+    __global uchar *input_addr  = input_ptr + input_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) + get_global_id(1) * input_stride_y + get_global_id(2) * input_stride_z;
+    __global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) + get_global_id(1) * output_stride_y + get_global_id(2) * output_stride_z;
+
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    data0 = floor(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input_addr));
 
-    VSTORE(VEC_SIZE)
-    (floor(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input.ptr)), 0, (__global DATA_TYPE *)output.ptr);
+    STORE_VECTOR_SELECT(data, DATA_TYPE, output_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0);
 }
+#endif // defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER)
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLFloorKernel.cpp b/src/core/CL/kernels/CLFloorKernel.cpp
index 2af0089bf0..6af1822656 100644
--- a/src/core/CL/kernels/CLFloorKernel.cpp
+++ b/src/core/CL/kernels/CLFloorKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -54,22 +54,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
 
     return Status{};
 }
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
-{
-    auto_init_if_empty(*output, *input);
-
-    const unsigned int num_elems_processed_per_iteration = 16 / input->element_size();
-
-    Window                 win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
-    AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
-    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
-    bool                   window_changed = update_window_and_padding(win, input_access, output_access);
-    output_access.set_valid_region(win, input->valid_region());
-
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, win);
-}
 } // namespace
 
 CLFloorKernel::CLFloorKernel()
@@ -86,22 +70,25 @@ void CLFloorKernel::configure(const CLCompileContext &compile_context, const ICL
 
     // Validate
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info()));
+    auto padding_info = get_padding_info({ input, output });
 
     _input  = input;
     _output = output;
 
-    const unsigned int    num_elems_processed_per_iteration = 16 / input->info()->element_size();
-    std::set<std::string> build_opts;
-    build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())));
-    build_opts.emplace(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)));
+    const unsigned int vec_size_x           = adjust_vec_size(max_cl_vector_width / input->info()->element_size(), input->info()->dimension(0));
+    const int          vec_size_x_leftovers = input->info()->dimension(0) % vec_size_x;
+    CLBuildOptions     build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x));
+    build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(vec_size_x_leftovers));
 
     // Create kernel
-    _kernel = create_kernel(compile_context, "floor_layer", build_opts);
+    _kernel = create_kernel(compile_context, "floor_layer", build_opts.options());
 
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input->info(), output->info());
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-    ICLKernel::configure_internal(win_config.second);
+    Window win = calculate_max_window(*input->info(), Steps(vec_size_x));
+    ICLKernel::configure_internal(win);
+    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
 }
 
 void CLFloorKernel::configure(const ICLTensor *input, ICLTensor *output)
@@ -112,8 +99,6 @@ void CLFloorKernel::configure(const ICLTensor *input, ICLTensor *output)
 Status CLFloorKernel::validate(const ITensorInfo *input, const ITensorInfo *output)
 {
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
-
     return Status{};
 }
 
diff --git a/tests/validation/CL/Floor.cpp b/tests/validation/CL/Floor.cpp
index 2961cfa3f2..a63f16b3c0 100644
--- a/tests/validation/CL/Floor.cpp
+++ b/tests/validation/CL/Floor.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -47,18 +47,16 @@ TEST_SUITE(Floor)
 // clang-format off
 DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(
         framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),  // Wrong data type
-                                                TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Window shrink
                                                 TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Invalid data type combination
                                                 TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching shapes
                                                 TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
         }),
         framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
-                                                TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
                                                 TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F16),
                                                 TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32),
                                                 TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
         })),
-        framework::dataset::make("Expected", { false, false, false, false, true })),
+        framework::dataset::make("Expected", { false, false, false, true })),
         input_info, output_info, expected)
 {
     const Status status = CLFloor::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false));
diff --git a/tests/validation/CL/UNIT/CompileContext.cpp b/tests/validation/CL/UNIT/CompileContext.cpp
index 06f7eae88e..a0541b6a99 100644
--- a/tests/validation/CL/UNIT/CompileContext.cpp
+++ b/tests/validation/CL/UNIT/CompileContext.cpp
@@ -56,6 +56,7 @@ TEST_CASE(CompileContextCache, framework::DatasetMode::ALL)
     std::set<std::string> build_opts;
     build_opts.emplace("-DDATA_TYPE=float");
     build_opts.emplace("-DVEC_SIZE=16");
+    build_opts.emplace("-DVEC_SIZE_LEFTOVER=0");
     compile_context.create_kernel(kernel_name, program_name, kernel_src.first, kernel_path, build_opts, kernel_src.second);
 
     // Check if the program is stored in the cache