Remove OpenCL padding: CLPadLayerKernel

Resolves: COMPMID-3912

Change-Id: I1f8bd3bfec263ebfd70bc96f9183ccdc3089db13
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4754
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

4 files changed, 113 insertions, 127 deletions

diff --git a/src/core/CL/cl_kernels/helpers.h b/src/core/CL/cl_kernels/helpers.h
index 372ccd91fb..df3b4937b2 100644
--- a/src/core/CL/cl_kernels/helpers.h
+++ b/src/core/CL/cl_kernels/helpers.h
@@ -110,18 +110,22 @@
  * @{
  */
 #define ROT1_0(x) ((x))
+#define ROT1_1(x) ((x))
 
 #define ROT2_0(x) ((x))
 #define ROT2_1(x) ((x).s10)
+#define ROT2_2(x) ((x))
 
 #define ROT3_0(x) ((x))
 #define ROT3_1(x) ((x).s201)
 #define ROT3_2(x) ((x).s120)
+#define ROT3_3(x) ((x))
 
 #define ROT4_0(x) ((x))
 #define ROT4_1(x) ((x).s3012)
 #define ROT4_2(x) ((x).s2301)
 #define ROT4_3(x) ((x).s1230)
+#define ROT4_4(x) ((x))
 
 #define ROT8_0(x) ((x))
 #define ROT8_1(x) ((x).s70123456)
@@ -131,6 +135,7 @@
 #define ROT8_5(x) ((x).s34567012)
 #define ROT8_6(x) ((x).s23456701)
 #define ROT8_7(x) ((x).s12345670)
+#define ROT8_8(x) ((x))
 
 #define ROT16_0(x) ((x))
 #define ROT16_1(x) ((x).sF0123456789ABCDE)
@@ -148,6 +153,7 @@
 #define ROT16_13(x) ((x).s3456789ABCDEF012)
 #define ROT16_14(x) ((x).s23456789ABCDEF01)
 #define ROT16_15(x) ((x).s123456789ABCDEF0)
+#define ROT16_16(x) ((x))
 /** @} */ // end of group ROTs_n
 
 /** Circular-right-shift (rotate-right) the given vector by the given amount.
diff --git a/src/core/CL/cl_kernels/pad_layer.cl b/src/core/CL/cl_kernels/pad_layer.cl
index fe71b5d119..903e924a2f 100644
--- a/src/core/CL/cl_kernels/pad_layer.cl
+++ b/src/core/CL/cl_kernels/pad_layer.cl
@@ -23,14 +23,15 @@
  */
 #include "helpers.h"
 
-#if defined(DATA_TYPE) && defined(VEC_SIZE) && defined(PAD_X_BEFORE) && defined(SRC_WIDTH)
+#if defined(DATA_TYPE) && defined(VEC_SIZE) && defined(PAD_X_BEFORE) && defined(SRC_WIDTH) && defined(PAD_X_BEFORE_REMAINDER) && defined(VEC_SIZE_LEFTOVER_WRITE)
 
 #define VEC_TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
 #define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE)
 #define VEC_SELECT SELECT_VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
 #define OFFSETS VEC_OFFS(SELECT_DATA_TYPE(DATA_TYPE), VEC_SIZE)
+#define SCALAR_COND(x) CONVERT((VEC_SELECT)x == (VEC_SELECT)1, VEC_SELECT)
 
-#if defined(CONST_VAL)
+#if defined(CONST_VAL) && defined(VEC_SIZE_LEFTOVER_READ)
 /** Perform a pad operation when PaddingMode is CONSTANT
  *
  * @note Data type can be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float
@@ -39,7 +40,9 @@
  * @note Pad to add to the left must be passed using the -DPAD_X_BEFORE compile flag, e.g. -DPAD_X_BEFORE=5
  * @note Input tensor's width must be passed using the -DSRC_WIDTH compile flag, e.g. -DSRC_WIDTH=224
  * @note In case pad left is more than the vector size, the number of threads to skip along the X axis must be passed using the
- *       -DNUM_THREADS_TO_SKIP_X compile flag, e.g. -DNUM_THREADS_TO_SKIP_X=1. This is defined as (PAD_X_BEFORE / VEC_SIZE)
+ *       -DTHREADS_TO_SKIP_BEFORE compile flag, e.g. -DTHREADS_TO_SKIP_BEFORE=1. This is defined as (PAD_X_BEFORE / VEC_SIZE)
+ * @note In case pad left is more than the vector size, the thread from which to skip along the X axis for pad right must be passed using the
+ *       -DTHREADS_TO_SKIP_AFTER compile flag, e.g. -THREADS_TO_SKIP_AFTER=1. This is defined as ((SRC_WIDTH + PAD_X_BEFORE) / VEC_SIZE)
  * @note If pad also needs to be added to the top of the tensor, the following compile flags must be passed at compile time:
  *       -# -DPAD_Y_BEFORE: Pad to add to the top of the input tensor (e.g. -DPAD_Y_BEFORE=3)
  *       -# -DSRC_HEIGHT: Input tensor's height (e.g. -DSRC_HEIGHT=127)
@@ -76,67 +79,77 @@ __kernel void pad_layer_constant(TENSOR3D_DECLARATION(src),
 #endif // defined(PAD_W_BEFORE)
                                 )
 {
-    const int x = get_global_id(0);
-    const int y = get_global_id(1);
-    const int z = get_global_id(2);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
 
-    uint cond = 0;
+    int x = get_global_id(0);
+    int y = get_global_id(1);
+    int z = get_global_id(2);
 
-#if defined(PAD_W_BEFORE)
-    cond |= batch < PAD_W_BEFORE || batch >= (SRC_BATCH + PAD_W_BEFORE);
-#endif // defined(PAD_W_BEFORE)
+    // If true, write only padding values; no reads performed
+    uint cond = 0;
+#if defined(THREADS_TO_SKIP_BEFORE)
+    cond |= x < THREADS_TO_SKIP_BEFORE || x > THREADS_TO_SKIP_AFTER;
+#endif // defined(THREADS_TO_SKIP_BEFORE)
+#if defined(PAD_Y_BEFORE)
+    cond |= y < PAD_Y_BEFORE || y >= (SRC_HEIGHT + PAD_Y_BEFORE);
+#endif // defined(PAD_Y_BEFORE)
 #if defined(PAD_Z_BEFORE)
     cond |= z < PAD_Z_BEFORE || z >= (SRC_DEPTH + PAD_Z_BEFORE);
 #endif // defined(PAD_Z_BEFORE)
+#if defined(PAD_W_BEFORE)
+    cond |= batch < PAD_W_BEFORE || batch >= (SRC_BATCH + PAD_W_BEFORE);
+#endif // defined(PAD_W_BEFORE)
 
     if(cond)
     {
-        Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
-        VSTORE(VEC_SIZE)
-        ((VEC_TYPE)CONST_VAL, 0, (__global DATA_TYPE *)dst.ptr);
+        VEC_TYPE const_vals0 = (VEC_TYPE)CONST_VAL;
+        STORE_VECTOR_SELECT(const_vals, DATA_TYPE, dst.ptr, VEC_SIZE, VEC_SIZE_LEFTOVER_WRITE, get_global_id(0) == (get_global_size(0) - 1));
     }
     else
     {
-        Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
-        Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
-
-#if defined(NUM_THREADS_TO_SKIP_X)
-        /* In case the pad left is greater than the vector size, and we are past the threads operating solely on pad values,
-         * the input pointer must be brought back along the X axis to start from the first non-pad values.
-         *
-         * E.g. with VEC_SIZE=2, PAD_X_BEFORE=5, CONST_VAL=0 and 1D input |1 2 3 4 5 6|:
-         *  -# The first thread will compute the output values |0 0| since it detects (x_outs == (0, 1)) < PAD_X_BEFORE
-         *  -# The second thread will compute the output values |0 0| since it detects (x_outs == (2, 3)) < PAD_X_BEFORE
-         *  -# The third thread should compute |0 1|, however the input pointer is now ahead of ((x * VEC_SIZE) == 4) values, reading |4 5|
-         *  -# To detect this, we use ((PAD_X_BEFORE / VEC_SIZE) == NUM_THREADS_TO_SKIP_X == 2) and check that it is >= to the current x
-         *  -# So, we bring the pointer back of NUM_THREADS_TO_SKIP_X threads, which means multiplying this constant by the input's step along the X axis
-         *  -# Now that the pointer is back of ((NUM_THREADS_TO_SKIP_X * src_step_x) == 4) values, it will read the desired values |0 1|
-         */
-        src.ptr -= select(0u, NUM_THREADS_TO_SKIP_X * src_step_x, x >= NUM_THREADS_TO_SKIP_X);
-#endif // defined(NUM_THREADS_TO_SKIP_X)
+        // Calculate input's coordinates based on output's
+        int w = 0;
+#if defined(THREADS_TO_SKIP_BEFORE)
+        x -= THREADS_TO_SKIP_BEFORE;
+#endif // defined(THREADS_TO_SKIP_BEFORE)
+#if defined(PAD_Y_BEFORE)
+        y -= PAD_Y_BEFORE;
+#endif // defined(PAD_Y_BEFORE)
 #if defined(PAD_Z_BEFORE)
-        src.ptr -= PAD_Z_BEFORE * src_step_z;
+        z -= PAD_Z_BEFORE;
 #endif // defined(PAD_Z_BEFORE)
 #if defined(PAD_W_BEFORE)
-        src.ptr -= PAD_W_BEFORE * SRC_DEPTH * src_step_z;
+        w -= PAD_W_BEFORE * SRC_DEPTH;
 #endif // defined(PAD_W_BEFORE)
+        x *= VEC_SIZE;
+        x -= PAD_X_BEFORE_REMAINDER;
 
-        VEC_TYPE src_vals = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr);
+        // Check for out of bound reads and clamp X coordinate
+        uint cond_left  = x < 0;
+        uint cond_right = (x + VEC_SIZE) > SRC_WIDTH;
+        x               = clamp(x, 0, (SRC_WIDTH - VEC_SIZE));
 
-        VEC_INT xs_out = (VEC_INT)(x * VEC_SIZE) + CONVERT(OFFSETS, VEC_INT);
-        VEC_INT cond   = xs_out < (VEC_INT)PAD_X_BEFORE || xs_out >= (VEC_INT)(SRC_WIDTH + PAD_X_BEFORE);
-#if defined(PAD_Y_BEFORE)
-        cond |= (VEC_INT)y < (VEC_INT)PAD_Y_BEFORE || (VEC_INT)y >= (VEC_INT)(SRC_HEIGHT + PAD_Y_BEFORE);
-#endif // defined(PAD_Y_BEFORE)
-        VSTORE(VEC_SIZE)
-        (select(src_vals, (VEC_TYPE)CONST_VAL, CONVERT(cond, VEC_SELECT)), 0, (__global DATA_TYPE *)dst.ptr);
+        // Calculate input's address
+        __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * src_stride_x + y * src_stride_y + z * src_stride_z + w * (int)src_stride_z;
+
+        // Read values and rotate them properly if they would have been across paddings
+        VEC_TYPE src_vals0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src_addr);
+        src_vals0          = select(src_vals0, ROTATE(src_vals0, VEC_SIZE, PAD_X_BEFORE_REMAINDER), SCALAR_COND(cond_left));
+        src_vals0          = select(src_vals0, ROTATE(src_vals0, VEC_SIZE, VEC_SIZE_LEFTOVER_READ), SCALAR_COND(cond_right));
+
+        // Check what values would be padding and replace them with the constant value
+        VEC_INT xs_out = (VEC_INT)(get_global_id(0) * VEC_SIZE) + VEC_OFFS(int, VEC_SIZE);
+        VEC_INT conds  = xs_out < (VEC_INT)PAD_X_BEFORE || xs_out >= (VEC_INT)(SRC_WIDTH + PAD_X_BEFORE);
+        src_vals0      = select(src_vals0, (VEC_TYPE)CONST_VAL, CONVERT(conds, VEC_SELECT));
+
+        // Store values in bounds
+        STORE_VECTOR_SELECT(src_vals, DATA_TYPE, dst.ptr, VEC_SIZE, VEC_SIZE_LEFTOVER_WRITE, get_global_id(0) == (get_global_size(0) - 1));
     }
 }
-#endif // defined(CONST_VAL)
+#endif // defined(CONST_VAL) && defined(VEC_SIZE_LEFTOVER_READ)
 
-#if defined(PAD_X_BEFORE_REMAINDER) && defined(PAD_X_AFTER_REMAINDER) && defined(PAD_X_BEFORE_REMAINDER_REFL) && defined(PAD_X_AFTER_REMAINDER_REFL) && defined(AFTER_PAD_FACT_X)
+#if defined(IS_REFLECT) && defined(PAD_X_AFTER_REMAINDER) && defined(PAD_X_BEFORE_REMAINDER_REFL) && defined(PAD_X_AFTER_REMAINDER_REFL) && defined(AFTER_PAD_FACT_X)
 
-#define SCALAR_COND(x) (VEC_SELECT) x == (VEC_SELECT)1
 #define ROTATE_REVERSE(x, n) ROTATE(REVERSE(x, VEC_SIZE), VEC_SIZE, n)
 #define SYMM_REFL_LEFT(x, n0, n1) select(ROTATE_REVERSE(x, n1), ROTATE(x, VEC_SIZE, n0), OFFSETS >= (VEC_SELECT)n0)
 #define SYMM_REFL_RIGHT(x, n0, n1) select(ROTATE(x, VEC_SIZE, n0), ROTATE_REVERSE(x, n1), OFFSETS >= (VEC_SELECT)n0)
@@ -232,20 +245,19 @@ __kernel void pad_layer_symmetric_reflect(TENSOR3D_DECLARATION(src),
     ((VEC_TYPE)(*(__global DATA_TYPE *)src_addr), 0, (__global DATA_TYPE *)dst.ptr);
 #else // SRC_WIDTH == 1
 
-    VEC_TYPE src_vals = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src_addr);
+    VEC_TYPE src_vals0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src_addr);
 
     // Choose rearrangement policy based on the defined conditions
-    src_vals = select(src_vals, SYMM_REFL_LEFT(src_vals, PAD_X_BEFORE_REMAINDER, PAD_X_BEFORE_REMAINDER_REFL), SCALAR_COND(is_across_pad_left));
-    src_vals = select(src_vals, SYMM_REFL_RIGHT(src_vals, PAD_X_AFTER_REMAINDER, PAD_X_AFTER_REMAINDER_REFL), SCALAR_COND(is_across_pad_right));
-    src_vals = select(src_vals, REVERSE(src_vals, VEC_SIZE), SCALAR_COND((is_before_pad_left || is_after_pad_right)));
+    src_vals0 = select(src_vals0, SYMM_REFL_LEFT(src_vals0, PAD_X_BEFORE_REMAINDER, PAD_X_BEFORE_REMAINDER_REFL), SCALAR_COND(is_across_pad_left));
+    src_vals0 = select(src_vals0, SYMM_REFL_RIGHT(src_vals0, PAD_X_AFTER_REMAINDER, PAD_X_AFTER_REMAINDER_REFL), SCALAR_COND(is_across_pad_right));
+    src_vals0 = select(src_vals0, REVERSE(src_vals0, VEC_SIZE), SCALAR_COND((is_before_pad_left || is_after_pad_right)));
 #if defined(AFTER_PAD_REM)
-    src_vals = select(src_vals, ROTATE(src_vals, VEC_SIZE, AFTER_PAD_REM), SCALAR_COND(neg_offs));
+    src_vals0 = select(src_vals0, ROTATE(src_vals0, VEC_SIZE, AFTER_PAD_REM), SCALAR_COND(neg_offs));
 #endif // defined(AFTER_PAD_REM)
 
-    // Store
-    VSTORE(VEC_SIZE)
-    (src_vals, 0, (__global DATA_TYPE *)dst.ptr);
+    // Store values in bounds
+    STORE_VECTOR_SELECT(src_vals, DATA_TYPE, dst.ptr, VEC_SIZE, VEC_SIZE_LEFTOVER_WRITE, get_global_id(0) == (get_global_size(0) - 1));
 #endif // SRC_WIDTH == 1
 }
-#endif // defined(PAD_X_BEFORE_REMAINDER) && defined(PAD_X_AFTER_REMAINDER) && defined(PAD_X_BEFORE_REMAINDER_REFL) && defined(PAD_X_AFTER_REMAINDER_REFL) && defined(AFTER_PAD_FACT_X)
-#endif // defined(DATA_TYPE) && defined(VEC_SIZE) && defined(PAD_X_BEFORE) && defined(SRC_WIDTH)
+#endif // defined(IS_REFLECT) && defined(PAD_X_AFTER_REMAINDER) && defined(PAD_X_BEFORE_REMAINDER_REFL) && defined(PAD_X_AFTER_REMAINDER_REFL) && defined(AFTER_PAD_FACT_X)
+#endif // defined(DATA_TYPE) && defined(VEC_SIZE) && defined(PAD_X_BEFORE) && defined(SRC_WIDTH) && defined(PAD_X_BEFORE_REMAINDER) && defined(VEC_SIZE_LEFTOVER_WRITE)
diff --git a/src/core/CL/kernels/CLPadLayerKernel.cpp b/src/core/CL/kernels/CLPadLayerKernel.cpp
index 485676667c..2f54b390d5 100644
--- a/src/core/CL/kernels/CLPadLayerKernel.cpp
+++ b/src/core/CL/kernels/CLPadLayerKernel.cpp
@@ -39,7 +39,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_UNUSED(constant_value);
     ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
-    ARM_COMPUTE_RETURN_ERROR_ON(padding.size() > input->num_dimensions());
+    ARM_COMPUTE_RETURN_ERROR_ON((padding.size() < 1) || (padding.size() > input->num_dimensions()));
     if(mode == PaddingMode::REFLECT || mode == PaddingMode::SYMMETRIC)
     {
         ARM_COMPUTE_RETURN_ERROR_ON(padding.size() > 3);
@@ -62,40 +62,10 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c
 
     return Status{};
 }
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const PaddingList &padding, PixelValue constant_value, PaddingMode mode,
-                                                        unsigned int &num_elems_processed_per_iteration)
-{
-    ARM_COMPUTE_UNUSED(constant_value, mode);
-
-    const TensorShape padded_shape = misc::shape_calculator::compute_padded_shape(input->tensor_shape(), padding);
-    auto_init_if_empty(*output, input->clone()->set_tensor_shape(padded_shape));
-
-    num_elems_processed_per_iteration = std::min(16U, 32U / static_cast<unsigned int>(element_size_from_data_type(input->data_type())));
-    if(input->dimension(0) < num_elems_processed_per_iteration)
-    {
-        num_elems_processed_per_iteration = 1 << static_cast<unsigned int>(std::log2(input->dimension(0)));
-    }
-
-    // Configure kernel window
-    Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
-
-    const int input_start_x = mode == PaddingMode::CONSTANT ? -(padding.at(0).first % num_elems_processed_per_iteration) : 0;
-    const int input_start_y = (mode == PaddingMode::CONSTANT && padding.size() > 1) ? -padding.at(1).first : 0;
-
-    AccessWindowRectangle  input_access(input, input_start_x, input_start_y, num_elems_processed_per_iteration, 1);
-    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
-
-    const bool window_changed = update_window_and_padding(win, input_access, output_access);
-    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
-
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, win);
-}
 } // namespace
 
 CLPadLayerKernel::CLPadLayerKernel()
-    : _input(nullptr), _output(nullptr), _input_start_x(0), _input_start_y(0), _4d_enabled(false)
+    : _input(nullptr), _output(nullptr), _4d_enabled(false)
 {
 }
 
@@ -106,40 +76,36 @@ void CLPadLayerKernel::configure(const ICLTensor *input, ICLTensor *output, cons
 
 void CLPadLayerKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const PaddingList &padding, PixelValue constant_value, PaddingMode mode)
 {
-    // Perform validation step
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(misc::shape_calculator::compute_padded_shape(input->info()->tensor_shape(), padding)));
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), padding, constant_value, mode));
 
+    auto padding_info = get_padding_info({ input, output });
+
     _input      = input;
     _output     = output;
     _4d_enabled = (mode == PaddingMode::CONSTANT) && (padding.size() > 3);
 
-    // Configure window
-    unsigned int vec_size;
-    auto         win_config = validate_and_configure_window(input->info(), output->info(), padding, constant_value, mode, vec_size);
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-    ICLKernel::configure_internal(win_config.second);
-
     // Set build options
-    std::string kernel_name = "pad_layer_";
-
-    const DataType    &data_type       = input->info()->data_type();
-    const unsigned int input_width     = input->info()->dimension(0);
-    const unsigned int input_height    = input->info()->dimension(1);
-    const unsigned int input_depth     = input->info()->dimension(2);
-    const unsigned int pad_x_before    = padding.at(0).first;
-    const unsigned int pad_y_before    = padding.size() > 1 ? padding.at(1).first : 0;
-    const unsigned int pad_z_before    = padding.size() > 2 ? padding.at(2).first : 0;
-    const unsigned int pad_right_start = input_width + pad_x_before;
-
-    _input_start_x = mode == PaddingMode::CONSTANT ? -(pad_x_before % vec_size) : 0;
-    _input_start_y = (mode == PaddingMode::CONSTANT && padding.size() > 1) ? -padding.at(1).first : 0;
+    const DataType    &data_type               = input->info()->data_type();
+    const unsigned int input_width             = input->info()->dimension(0);
+    const unsigned int input_height            = input->info()->dimension(1);
+    const unsigned int input_depth             = input->info()->dimension(2);
+    const unsigned int pad_x_before            = padding.at(0).first;
+    const unsigned int pad_y_before            = padding.size() > 1 ? padding.at(1).first : 0;
+    const unsigned int pad_z_before            = padding.size() > 2 ? padding.at(2).first : 0;
+    const unsigned int vec_size                = adjust_vec_size(std::min(16U, 32U / static_cast<unsigned int>(element_size_from_data_type(input->info()->data_type()))), input_width);
+    const unsigned int pad_right_start         = input_width + pad_x_before;
+    const unsigned int pad_x_before_remainder  = pad_x_before % vec_size;
+    const unsigned int vec_size_leftover_write = vec_size - (ceil_to_multiple(output->info()->dimension(0), vec_size) - output->info()->dimension(0));
 
     CLBuildOptions build_opts;
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size));
     build_opts.add_option("-DPAD_X_BEFORE=" + support::cpp11::to_string(pad_x_before));
     build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input_width));
+    build_opts.add_option("-DPAD_X_BEFORE_REMAINDER=" + support::cpp11::to_string(pad_x_before_remainder));
+    build_opts.add_option("-DVEC_SIZE_LEFTOVER_WRITE=" + support::cpp11::to_string(vec_size_leftover_write));
     if(padding.size() > 1)
     {
         build_opts.add_option("-DPAD_Y_BEFORE=" + support::cpp11::to_string(pad_y_before));
@@ -152,15 +118,23 @@ void CLPadLayerKernel::configure(const CLCompileContext &compile_context, const
         }
     }
 
+    std::string kernel_name = "pad_layer_";
     switch(mode)
     {
         case PaddingMode::CONSTANT:
         {
             kernel_name += "constant";
 
+            const unsigned int vec_size_leftover_read = vec_size - (ceil_to_multiple(pad_right_start, vec_size) - pad_right_start);
+
             build_opts.add_option("-DCONST_VAL=" + string_from_pixel_value(constant_value, data_type));
-            build_opts.add_option_if(pad_x_before >= vec_size, "-DNUM_THREADS_TO_SKIP_X=" + support::cpp11::to_string(pad_x_before / vec_size));
+            build_opts.add_option("-DVEC_SIZE_LEFTOVER_READ=" + support::cpp11::to_string(vec_size_leftover_read));
 
+            if(pad_x_before >= vec_size)
+            {
+                build_opts.add_option("-DTHREADS_TO_SKIP_BEFORE=" + support::cpp11::to_string(pad_x_before / vec_size));
+                build_opts.add_option("-DTHREADS_TO_SKIP_AFTER=" + support::cpp11::to_string(pad_right_start / vec_size));
+            }
             if(_4d_enabled)
             {
                 build_opts.add_option("-DPAD_W_BEFORE=" + support::cpp11::to_string(padding.at(3).first));
@@ -176,13 +150,11 @@ void CLPadLayerKernel::configure(const CLCompileContext &compile_context, const
 
             const auto is_reflect = static_cast<unsigned int>(mode == PaddingMode::REFLECT);
 
-            const unsigned int pad_x_before_remainder = pad_x_before % vec_size;
-            const unsigned int pad_x_after_remainder  = pad_right_start % vec_size;
-            const unsigned int after_pad_fact_x       = (2 * input_width + pad_x_before) - is_reflect;
-            const unsigned int output_last_x          = ceil_to_multiple(pad_right_start + padding.at(0).second, vec_size);
+            const unsigned int pad_x_after_remainder = pad_right_start % vec_size;
+            const unsigned int after_pad_fact_x      = (2 * input_width + pad_x_before) - is_reflect;
+            const unsigned int output_last_x         = ceil_to_multiple(pad_right_start + padding.at(0).second, vec_size);
 
             build_opts.add_option("-DIS_REFLECT=" + support::cpp11::to_string(is_reflect));
-            build_opts.add_option("-DPAD_X_BEFORE_REMAINDER=" + support::cpp11::to_string(pad_x_before_remainder));
             build_opts.add_option("-DPAD_X_AFTER_REMAINDER=" + support::cpp11::to_string(pad_x_after_remainder));
             build_opts.add_option("-DPAD_X_BEFORE_REMAINDER_REFL=" + support::cpp11::to_string((pad_x_before_remainder + is_reflect) % vec_size));
             build_opts.add_option("-DPAD_X_AFTER_REMAINDER_REFL=" + support::cpp11::to_string((pad_x_after_remainder - is_reflect) % vec_size));
@@ -197,14 +169,17 @@ void CLPadLayerKernel::configure(const CLCompileContext &compile_context, const
 
     // Create kernel
     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
+
+    // Configure window
+    Window win = calculate_max_window(*output->info(), Steps(vec_size));
+    ICLKernel::configure_internal(win);
+
+    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
 }
 
 Status CLPadLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding, PixelValue constant_value, PaddingMode mode)
 {
-    unsigned int vec_size;
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, padding, constant_value, mode));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), padding, constant_value, mode, vec_size).first);
-
     return Status{};
 }
 
@@ -213,25 +188,20 @@ void CLPadLayerKernel::run(const Window &window, cl::CommandQueue &queue)
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
 
-    Window win_in = window;
-    win_in.adjust(Window::DimX, _input_start_x, true);
-    win_in.adjust(Window::DimY, _input_start_y, true);
-
-    Window       slice_out = window.first_slice_window_3D();
-    Window       slice_in  = win_in.first_slice_window_3D();
-    unsigned int batch     = 0;
+    Window       slice = window.first_slice_window_3D();
+    unsigned int batch = 0;
     do
     {
         unsigned int idx = 0;
-        add_3D_tensor_argument(idx, _input, slice_in);
-        add_3D_tensor_argument(idx, _output, slice_out);
+        add_3D_tensor_argument(idx, _input, slice);
+        add_3D_tensor_argument(idx, _output, slice);
         if(_4d_enabled)
         {
             add_argument<unsigned int>(idx, batch++);
         }
 
-        enqueue(queue, *this, slice_out, lws_hint());
+        enqueue(queue, *this, slice, lws_hint());
     }
-    while(window.slide_window_slice_3D(slice_out) && win_in.slide_window_slice_3D(slice_in));
+    while(window.slide_window_slice_3D(slice));
 }
 } // namespace arm_compute
diff --git a/src/core/CL/kernels/CLPadLayerKernel.h b/src/core/CL/kernels/CLPadLayerKernel.h
index 2b0abb18df..90af337f94 100644
--- a/src/core/CL/kernels/CLPadLayerKernel.h
+++ b/src/core/CL/kernels/CLPadLayerKernel.h
@@ -88,8 +88,6 @@ public:
 private:
     const ICLTensor *_input;
     ICLTensor       *_output;
-    int              _input_start_x;
-    int              _input_start_y;
     bool             _4d_enabled;
 };
 } // namespace arm_compute