COMPMID-3793: Remove OpenCL padding: CLWidthConcatenateLayerKernel

Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: I705044a9429bb9a08268368b09463c2af85616d5
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4253
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 24 insertions, 36 deletions

diff --git a/arm_compute/core/CL/kernels/CLWidthConcatenateLayerKernel.h b/arm_compute/core/CL/kernels/CLWidthConcatenateLayerKernel.h
index 32e90af404..16cf167b25 100644
--- a/arm_compute/core/CL/kernels/CLWidthConcatenateLayerKernel.h
+++ b/arm_compute/core/CL/kernels/CLWidthConcatenateLayerKernel.h
@@ -69,9 +69,6 @@ public:
 
     // Inherited methods overridden:
     void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
-
-private:
-    unsigned int _width_offset;
 };
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_CLWIDTHCONCATENATELAYERKERNEL_H */
diff --git a/src/core/CL/cl_kernels/concatenate.cl b/src/core/CL/cl_kernels/concatenate.cl
index 0f4b5afe2c..7c6c8d211a 100644
--- a/src/core/CL/cl_kernels/concatenate.cl
+++ b/src/core/CL/cl_kernels/concatenate.cl
@@ -278,11 +278,12 @@ __kernel void concatenate_width_x4(
 #endif /* defined(INPUT1_WIDTH) */
 #endif /* defined(DEPTH) && defined(ELEMENT_SIZE) */
 
-#if defined(WIDTH_OFFSET) && defined(DEPTH)
+#if defined(WIDTH_OFFSET) && defined(DEPTH) && defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER)
 /** This kernel concatenates the input tensor into the output tensor along the first dimension
  *
  * @note The data type has to be passed at compile time using -DDATA_TYPE. i.e. -DDATA_TYPE=float
  * @note Vector size has to be passed at compile time using -DVEC_SIZE. i.e. -DVEC_SIZE=16
+ * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE
  * @note The offset for the first spatial dimension has to be passed at compile time using -DWIDTH_OFFSET. i.e. -DWIDTH_OFFSET=128
  * @note Tensor depth should be given as a preprocessor argument using -DDEPTH=size. e.g. -DDEPTH=16
  *
@@ -312,23 +313,27 @@ __kernel void concatenate_width(
     TENSOR4D_DECLARATION(src),
     TENSOR4D_DECLARATION(dst))
 {
-    Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, DEPTH);
-    Tensor4D dst = CONVERT_TO_TENSOR4D_STRUCT(dst, DEPTH);
+    // Calculate input indices
+    const int x = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0);
+    const int y = get_global_id(1);
+    const int z = get_global_id(2) % (int)DEPTH;
+    const int w = get_global_id(2) / (int)DEPTH;
+
+    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) + y * src_stride_y + z * src_stride_z + w * src_stride_w;
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) + y * dst_stride_y + z * dst_stride_z + w * dst_stride_w;
 
     VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
-    source_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr);
+    source_values0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src_addr);
 
 #if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT)
-    const VEC_QUANT out = requantize(source_values, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT);
-    VSTORE(VEC_SIZE)
-    (out, 0, (__global DATA_TYPE *)(dst.ptr) + WIDTH_OFFSET);
+    const VEC_QUANT out0 = requantize(source_values0, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT);
+    STORE_VECTOR_SELECT(out, DATA_TYPE, dst_addr + WIDTH_OFFSET * sizeof(DATA_TYPE), VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
 #else  /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
-    VSTORE(VEC_SIZE)
-    (source_values, 0, (__global DATA_TYPE *)(dst.ptr) + WIDTH_OFFSET);
+    STORE_VECTOR_SELECT(source_values, DATA_TYPE, dst_addr + WIDTH_OFFSET * sizeof(DATA_TYPE), VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
 #endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
 }
 
-#endif /* defined(WIDTH_OFFSET) && defined(DEPTH) */
+#endif /* defined(WIDTH_OFFSET) && defined(DEPTH) && defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER)*/
 
 #if defined(VEC_SIZE_LEFTOVER)
 
diff --git a/src/core/CL/kernels/CLWidthConcatenateLayerKernel.cpp b/src/core/CL/kernels/CLWidthConcatenateLayerKernel.cpp
index 8b0aaf3227..a9a601dc8e 100644
--- a/src/core/CL/kernels/CLWidthConcatenateLayerKernel.cpp
+++ b/src/core/CL/kernels/CLWidthConcatenateLayerKernel.cpp
@@ -38,21 +38,6 @@ namespace arm_compute
 {
 namespace
 {
-constexpr unsigned int num_elems_processed_per_iteration = 16;
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, unsigned int width_offset, ITensorInfo *output)
-{
-    // The window needs to be based on input as we copy all the widths of input
-    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, width_offset, num_elems_processed_per_iteration);
-    bool                   window_changed = update_window_and_padding(win, input_access, output_access);
-
-    Window win_collapsed = win.collapse(win, Window::DimZ);
-
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, win_collapsed);
-}
 Status validate_arguments(const ITensorInfo *input, unsigned int width_offset, const ITensorInfo *output)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
@@ -73,14 +58,12 @@ Status validate_arguments(const ITensorInfo *input, unsigned int width_offset, c
 } // namespace
 
 CLWidthConcatenateLayerKernel::CLWidthConcatenateLayerKernel()
-    : _width_offset(0)
 {
 }
 
 Status CLWidthConcatenateLayerKernel::validate(const ITensorInfo *input, unsigned int width_offset, const ITensorInfo *output)
 {
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, width_offset, output));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), width_offset, output->clone().get()).first);
     return Status{};
 }
 
@@ -89,13 +72,16 @@ void CLWidthConcatenateLayerKernel::configure(const CLCompileContext &compile_co
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input, width_offset, output));
 
-    _width_offset = width_offset;
+    auto padding_info = get_padding_info({ input, output });
+
+    const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16, input->dimension(0));
 
     // Add build options
     CLBuildOptions build_opts;
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->data_type()));
     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
-    build_opts.add_option("-DWIDTH_OFFSET=" + support::cpp11::to_string(_width_offset));
+    build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(input->dimension(0) % num_elems_processed_per_iteration));
+    build_opts.add_option("-DWIDTH_OFFSET=" + support::cpp11::to_string(width_offset));
     build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input->dimension(2)));
 
     if(is_data_type_quantized_asymmetric(input->data_type()) && input->quantization_info() != output->quantization_info())
@@ -112,13 +98,13 @@ void CLWidthConcatenateLayerKernel::configure(const CLCompileContext &compile_co
     // Create kernel
     _kernel = create_kernel(compile_context, "concatenate_width", build_opts.options());
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input, width_offset, output);
-    ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
-
-    ICLKernel::configure_internal(std::get<1>(win_config));
+    Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
+    ICLKernel::configure_internal(win.collapse(win, Window::DimZ));
 
     // Set output valid region
     output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
+
+    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
 }
 
 void CLWidthConcatenateLayerKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)