COMPMID-3710: Remove OpenCL padding: CLDepthConvertLayerKernel

Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: Iee1d4655012ce4cb699535697aeefec673f0bc63
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4157
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 45 insertions, 19 deletions

diff --git a/src/core/CL/cl_kernels/depth_convert.cl b/src/core/CL/cl_kernels/depth_convert.cl
index 75e6829cc0..046b26df01 100644
--- a/src/core/CL/cl_kernels/depth_convert.cl
+++ b/src/core/CL/cl_kernels/depth_convert.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2019 Arm Limited.
+ * Copyright (c) 2016-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -43,6 +43,7 @@
  * @note The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT:
  * e.g. -DDATA_TYPE_IN=uchar -DDATA_TYPE_OUT=short
  * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -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
  *
  * @param[in]  in_ptr                            Pointer to the source image. Supported data types: U8/S8/QSYMM8_PER_CHANNEL/U16/S16/U32/S32/F16/F32
  * @param[in]  in_stride_x                       Stride of the source image in X dimension (in bytes)
@@ -67,24 +68,27 @@ __kernel void convert_depth_down(
     TENSOR3D_DECLARATION(out),
     const int shift)
 {
-    // Get pixels pointer
-    Tensor3D in  = CONVERT_TO_TENSOR3D_STRUCT(in);
-    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out);
+    int x_offs = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0);
+
+    __global uchar *in_addr  = in_ptr + in_offset_first_element_in_bytes + sizeof(DATA_TYPE_IN) * x_offs + get_global_id(1) * in_stride_y + get_global_id(2) * in_stride_z;
+    __global uchar *out_addr = out_ptr + out_offset_first_element_in_bytes + sizeof(DATA_TYPE_OUT) * x_offs + get_global_id(1) * out_stride_y + get_global_id(2) * out_stride_z;
 
     // Load data
     VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE)
-    in_data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_IN *)in.ptr);
+    in_data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_IN *)in_addr);
 
 #if defined(IS_DATA_TYPE_QUANTIZED)
     in_data ^= (VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE))0x80;
 #endif // defined(IS_DATA_TYPE_QUANTIZED)
 
 #if defined(IS_DATA_TYPE_FLOAT)
-    VSTORE(VEC_SIZE)
-    (CONVERT_DOWN(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)), 0, (__global DATA_TYPE_OUT *)out.ptr);
+    VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)
+    res0 = CONVERT_DOWN(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE));
+    STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
 #else  /* defined(IS_DATA_TYPE_FLOAT) */
-    VSTORE(VEC_SIZE)
-    (CONVERT_DOWN(in_data >> shift, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)), 0, (__global DATA_TYPE_OUT *)out.ptr);
+    VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)
+    res0 = CONVERT_DOWN(in_data >> shift, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE));
+    STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
 #endif /* defined(IS_DATA_TYPE_FLOAT) */
 }
 
@@ -93,6 +97,7 @@ __kernel void convert_depth_down(
  * @note The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT:
  * e.g. -DDATA_TYPE_IN=uchar -DDATA_TYPE_OUT=short
  * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -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
  *
  * @param[in]  in_ptr                            Pointer to the source image. Supported data types: U8/S8/U16/S16/U32/S32/F16/F32
  * @param[in]  in_stride_x                       Stride of the source image in X dimension (in bytes)
@@ -117,19 +122,22 @@ __kernel void convert_depth_up(
     TENSOR3D_DECLARATION(out),
     const int shift)
 {
-    // Get pixels pointer
-    Tensor3D in  = CONVERT_TO_TENSOR3D_STRUCT(in);
-    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out);
+    int x_offs = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0);
+
+    __global uchar *in_addr  = in_ptr + in_offset_first_element_in_bytes + sizeof(DATA_TYPE_IN) * x_offs + get_global_id(1) * in_stride_y + get_global_id(2) * in_stride_z;
+    __global uchar *out_addr = out_ptr + out_offset_first_element_in_bytes + sizeof(DATA_TYPE_OUT) * x_offs + get_global_id(1) * out_stride_y + get_global_id(2) * out_stride_z;
 
     // Load data
     VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE)
-    in_data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_IN *)in.ptr);
+    in_data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_IN *)in_addr);
 
 #if defined(IS_DATA_TYPE_FLOAT)
-    VSTORE(VEC_SIZE)
-    (CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)), 0, (__global DATA_TYPE_OUT *)out.ptr);
+    VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)
+    res0 = CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE));
+    STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
 #else  /* defined(IS_DATA_TYPE_FLOAT) */
-    VSTORE(VEC_SIZE)
-    (CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)) << shift, 0, (__global DATA_TYPE_OUT *)out.ptr);
+    VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)
+    res0 = CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)) << shift;
+    STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
 #endif /* defined(IS_DATA_TYPE_FLOAT) */
 }
diff --git a/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp b/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp
index 24f638f8c4..452a14bd29 100644
--- a/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp
+++ b/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp
@@ -32,6 +32,7 @@
 #include "arm_compute/core/Validate.h"
 #include "src/core/CL/CLValidate.h"
 #include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/WindowHelpers.h"
 #include "support/StringSupport.h"
 
 #include <cstddef>
@@ -81,21 +82,27 @@ void CLDepthConvertLayerKernel::configure(const CLCompileContext &compile_contex
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
+    _input  = input;
+    _output = output;
+
     // Auto initialize output shape if not initialized (We can only auto-configure the shape, datatype must be given)
     set_shape_if_empty(*output->info(), input->info()->tensor_shape());
 
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), policy, shift));
 
+    auto padding_info = get_padding_info({ input, output });
+
     // Get data sizes
     const size_t input_size  = data_size_from_type(input->info()->data_type());
     const size_t output_size = data_size_from_type(output->info()->data_type());
 
     // Get number of elements to process per iterations
-    constexpr unsigned int num_elems_processed_per_iteration = 16;
+    const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / input->info()->element_size(), input->info()->dimension(0));
 
     // Set build options
     CLBuildOptions build_opts;
     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
+    build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(input->info()->dimension(0) % num_elems_processed_per_iteration));
     build_opts.add_option("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(input->info()->data_type()));
     build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type()));
     // Conversions from float always SATURATE as out-of-bounds conversion from float->integer is implementation defined
@@ -111,14 +118,25 @@ void CLDepthConvertLayerKernel::configure(const CLCompileContext &compile_contex
     unsigned int idx = 2 * num_arguments_per_3D_tensor(); // Skip the input and output parameters
     _kernel.setArg(idx++, shift);
 
+    // Since we have a leftover vector size calculated using the input tensor shape, it is required to
+    // have the input region equal to the tensor shape
+    ValidRegion input_valid_region = input->info()->valid_region();
+    input->info()->set_valid_region(ValidRegion(Coordinates(0, 0), input->info()->tensor_shape()));
+
     // Configure kernel
-    ICLSimple2DKernel::configure(input, output, num_elems_processed_per_iteration);
+    Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
+    ICLKernel::configure_internal(win);
 
     // Collapse window
     const Window &full_window      = window();
     Window        collapsed_window = full_window.collapse_if_possible(full_window, Window::DimZ);
     ICLKernel::configure_internal(collapsed_window);
 
+    // Restore the valid region
+    input->info()->set_valid_region(input_valid_region);
+
+    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
+
     // Set config_id for enabling LWS tuning
     _config_id = kernel_name;
     _config_id += "_";