COMPMID-3916: Remove OpenCL padding CLRangeKernel

Change-Id: Id2cc77508b0f2fa36a298059476b01704cfbdcaf
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4580
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

2 files changed, 77 insertions, 118 deletions

diff --git a/src/core/CL/cl_kernels/range.cl b/src/core/CL/cl_kernels/range.cl
index 1e5c77b376..5569101615 100644
--- a/src/core/CL/cl_kernels/range.cl
+++ b/src/core/CL/cl_kernels/range.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 Arm Limited.
+ * Copyright (c) 2018-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,13 +23,29 @@
  */
 #include "helpers.h"
 
-#if defined(VECTOR_SIZE) && defined(START) && defined(STEP) && defined(DATA_TYPE)
+#if defined(VECTOR_SIZE) && defined(START) && defined(STEP) && defined(DATA_TYPE) && defined(VEC_SIZE_LEFTOVER)
+
+#if !defined(OFFSET_OUT) && !defined(SCALE_OUT)
+
+#if VECTOR_SIZE == 2
+#define STEP_VEC ((VEC_DATA_TYPE(DATA_TYPE, 2))(0, STEP))
+#elif VECTOR_SIZE == 3
+#define STEP_VEC ((VEC_DATA_TYPE(DATA_TYPE, 3))(0, STEP, 2 * STEP))
+#elif VECTOR_SIZE == 4
+#define STEP_VEC ((VEC_DATA_TYPE(DATA_TYPE, 4))(0, STEP, 2 * STEP, 3 * STEP))
+#elif VECTOR_SIZE == 8
+#define STEP_VEC ((VEC_DATA_TYPE(DATA_TYPE, 8))(0, STEP, 2 * STEP, 3 * STEP, 4 * STEP, 5 * STEP, 6 * STEP, 7 * STEP))
+#elif VECTOR_SIZE == 16
+#define STEP_VEC ((VEC_DATA_TYPE(DATA_TYPE, 16))(0, STEP, 2 * STEP, 3 * STEP, 4 * STEP, 5 * STEP, 6 * STEP, 7 * STEP, 8 * STEP, 9 * STEP, 10 * STEP, 11 * STEP, 12 * STEP, 13 * STEP, 14 * STEP, 15 * STEP))
+#endif // VECTOR_SIZE == 2
+
 /** Generates a sequence of numbers starting from START and extends by increments of 'STEP' up to but not including 'END'.
  *
  * @note starting value of the sequence must be given as a preprocessor argument using -DSTART=value. e.g. -DSTART=0
  * @note difference between consequtive elements of the sequence must be given as a preprocessor argument using -DSTEP=value. e.g. -DSTEP=1
  * @note Datatype must be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
  * @note vector size supported by the device must be given as a preprocessor argument using -DVECTOR_SIZE=value. e.g. -DDATA_TYPE=4
+ * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVECTOR_SIZE=3. It is defined as the remainder between the input's first dimension and VECTOR_SIZE
  *
  * @param[out] out_ptr                           Pointer to the destination tensor. Supported data types: U8/S8/U16/S16/U32/S32/F16/F32.
  * @param[in]  out_stride_x                      Stride of the destination tensor in X dimension (in bytes)
@@ -39,45 +55,34 @@
 __kernel void range(
     VECTOR_DECLARATION(out))
 {
-    uint           id      = get_global_id(0) * VECTOR_SIZE;
-    __global void *dst_ptr = out_ptr + out_offset_first_element_in_bytes + id * sizeof(DATA_TYPE);
+    uint     id                 = max((int)(get_global_id(0) * VECTOR_SIZE - (VECTOR_SIZE - VEC_SIZE_LEFTOVER) % VECTOR_SIZE), 0);
+    __global DATA_TYPE *dst_ptr = out_ptr + out_offset_first_element_in_bytes + id * sizeof(DATA_TYPE);
 #if VECTOR_SIZE == 1
     DATA_TYPE seq;
     seq = (DATA_TYPE)START + (DATA_TYPE)id * (DATA_TYPE)STEP;
 
-    *((__global DATA_TYPE *)dst_ptr) = seq;
-#else // VECTOR_SIZE == 1
+    *dst_ptr = seq;
+#else  // VECTOR_SIZE == 1
     VEC_DATA_TYPE(DATA_TYPE, VECTOR_SIZE)
-    seq;
-
-    seq.s0 = ((DATA_TYPE)START + (DATA_TYPE)id * (DATA_TYPE)STEP);
-#if VECTOR_SIZE > 1
-    seq.s1 = seq.s0 + (DATA_TYPE)STEP;
-#if VECTOR_SIZE > 2
-    seq.s2 = seq.s1 + (DATA_TYPE)STEP;
-#if VECTOR_SIZE > 3
-    seq.s3 = seq.s2 + (DATA_TYPE)STEP;
-#if VECTOR_SIZE > 4
-    seq.s4 = seq.s3 + (DATA_TYPE)STEP;
-#if VECTOR_SIZE > 5
-    seq.s5 = seq.s4 + (DATA_TYPE)STEP;
-#if VECTOR_SIZE > 6
-    seq.s6 = seq.s5 + (DATA_TYPE)STEP;
-#if VECTOR_SIZE > 7
-    seq.s7 = seq.s6 + (DATA_TYPE)STEP;
-#endif // VECTOR_SIZE > 7
-#endif // VECTOR_SIZE > 6
-#endif // VECTOR_SIZE > 5
-#endif // VECTOR_SIZE > 4
-#endif // VECTOR_SIZE > 3
-#endif // VECTOR_SIZE > 2
-#endif // VECTOR_SIZE > 1
-    VSTORE(VECTOR_SIZE)
-    (seq, 0, ((__global DATA_TYPE *)dst_ptr));
+    seq0 = ((DATA_TYPE)START + (DATA_TYPE)id * (DATA_TYPE)STEP);
+    seq0 = seq0 + STEP_VEC;
+    STORE_VECTOR_SELECT(seq, DATA_TYPE, dst_ptr, VECTOR_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
 #endif //VECTOR_SIZE == 1
 }
 
-#if defined(OFFSET_OUT) && defined(SCALE_OUT)
+#else // !defined(OFFSET_OUT) && !defined(SCALE_OUT)
+
+#if VECTOR_SIZE == 2
+#define STEP_VEC ((VEC_DATA_TYPE(float, 2))(0, STEP))
+#elif VECTOR_SIZE == 3
+#define STEP_VEC ((VEC_DATA_TYPE(float, 3))(0, STEP, 2 * STEP))
+#elif VECTOR_SIZE == 4
+#define STEP_VEC ((VEC_DATA_TYPE(float, 4))(0, STEP, 2 * STEP, 3 * STEP))
+#elif VECTOR_SIZE == 8
+#define STEP_VEC ((VEC_DATA_TYPE(float, 8))(0, STEP, 2 * STEP, 3 * STEP, 4 * STEP, 5 * STEP, 6 * STEP, 7 * STEP))
+#elif VECTOR_SIZE == 16
+#define STEP_VEC ((VEC_DATA_TYPE(float, 16))(0, STEP, 2 * STEP, 3 * STEP, 4 * STEP, 5 * STEP, 6 * STEP, 7 * STEP, 8 * STEP, 9 * STEP, 10 * STEP, 11 * STEP, 12 * STEP, 13 * STEP, 14 * STEP, 15 * STEP))
+#endif // VECTOR_SIZE == 2
 
 #define CONVERT_RTE(x, type) (convert_##type##_rte((x)))
 #define CONVERT_DOWN(x, type) CONVERT_RTE(x, type)
@@ -99,47 +104,25 @@ __kernel void range(
 __kernel void range_quantized(
     VECTOR_DECLARATION(out))
 {
-    size_t         id      = get_global_id(0) * VECTOR_SIZE;
-    __global void *dst_ptr = out_ptr + out_offset_first_element_in_bytes + id * sizeof(DATA_TYPE);
+    uint     id                 = max((int)(get_global_id(0) * VECTOR_SIZE - (VECTOR_SIZE - VEC_SIZE_LEFTOVER) % VECTOR_SIZE), 0);
+    __global DATA_TYPE *dst_ptr = out_ptr + out_offset_first_element_in_bytes + id * sizeof(DATA_TYPE);
 #if VECTOR_SIZE == 1
-    float seq;
-    seq                          = (float)START + (float)id * (float)STEP;
-    seq                          = (DATA_TYPE)(int)(seq / ((float)SCALE_OUT) + (float)OFFSET_OUT);
-    seq                          = max(0.0f, min(seq, 255.0f));
-    *((__global uchar *)dst_ptr) = CONVERT_SAT(CONVERT_DOWN(seq, int), uchar);
-#else // VECTOR_SIZE == 1
+    float               seq;
+    seq      = (float)START + (float)id * (float)STEP;
+    seq      = (DATA_TYPE)(int)(seq / ((float)SCALE_OUT) + (float)OFFSET_OUT);
+    seq      = max(0.0f, min(seq, 255.0f));
+    *dst_ptr = CONVERT_SAT(CONVERT_DOWN(seq, int), uchar);
+#else  // VECTOR_SIZE == 1
     VEC_DATA_TYPE(float, VECTOR_SIZE)
-    seq;
-    seq.s0 = (float)START + id * (float)STEP;
-#if VECTOR_SIZE > 1
-    seq.s1 = seq.s0 + (float)STEP;
-#if VECTOR_SIZE > 2
-    seq.s2 = seq.s1 + (float)STEP;
-#if VECTOR_SIZE > 3
-    seq.s3 = seq.s2 + (float)STEP;
-#if VECTOR_SIZE > 4
-    seq.s4 = seq.s3 + (float)STEP;
-#if VECTOR_SIZE > 5
-    seq.s5 = seq.s4 + (float)STEP;
-#if VECTOR_SIZE > 6
-    seq.s6 = seq.s5 + (float)STEP;
-#if VECTOR_SIZE > 7
-    seq.s7 = seq.s6 + (float)STEP;
-#endif // VECTOR_SIZE > 7
-#endif // VECTOR_SIZE > 6
-#endif // VECTOR_SIZE > 5
-#endif // VECTOR_SIZE > 4
-#endif // VECTOR_SIZE > 3
-#endif // VECTOR_SIZE > 2
-#endif // VECTOR_SIZE > 1
-    seq    = seq / ((VEC_DATA_TYPE(float, VECTOR_SIZE))((float)SCALE_OUT)) + ((VEC_DATA_TYPE(float, VECTOR_SIZE))((float)OFFSET_OUT));
-    seq    = max((VEC_DATA_TYPE(float, VECTOR_SIZE))(0.0f), min(seq, (VEC_DATA_TYPE(float, VECTOR_SIZE))(255.0f)));
+    seq = (float)START + id * (float)STEP;
+    seq = seq + STEP_VEC;
+    seq = seq / ((VEC_DATA_TYPE(float, VECTOR_SIZE))((float)SCALE_OUT)) + ((VEC_DATA_TYPE(float, VECTOR_SIZE))((float)OFFSET_OUT));
+    seq = max((VEC_DATA_TYPE(float, VECTOR_SIZE))(0.0f), min(seq, (VEC_DATA_TYPE(float, VECTOR_SIZE))(255.0f)));
     VEC_DATA_TYPE(uchar, VECTOR_SIZE)
-    res = CONVERT_SAT(CONVERT_DOWN(seq, VEC_DATA_TYPE(int, VECTOR_SIZE)), VEC_DATA_TYPE(uchar, VECTOR_SIZE));
-    VSTORE(VECTOR_SIZE)
-    (res, 0, ((__global DATA_TYPE *)dst_ptr));
+    res0 = CONVERT_SAT(CONVERT_DOWN(seq, VEC_DATA_TYPE(int, VECTOR_SIZE)), VEC_DATA_TYPE(uchar, VECTOR_SIZE));
+    STORE_VECTOR_SELECT(res, DATA_TYPE, dst_ptr, VECTOR_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
 #endif // VECTOR_SIZE == 1
 }
-#endif // defined(OFFSET_OUT) && defined(SCALE_OUT)
+#endif // !defined(OFFSET_OUT) && !defined(SCALE_OUT)
 
-#endif // defined(VECTOR_SIZE) && defined(START) && defined(STEP) && defined(DATA_TYPE)
+#endif // defined(VECTOR_SIZE) && defined(START) && defined(STEP) && defined(DATA_TYPE) && defined(VEC_SIZE_LEFTOVER)
diff --git a/src/core/CL/kernels/CLRangeKernel.cpp b/src/core/CL/kernels/CLRangeKernel.cpp
index 892f1c7c9f..85f79988c9 100644
--- a/src/core/CL/kernels/CLRangeKernel.cpp
+++ b/src/core/CL/kernels/CLRangeKernel.cpp
@@ -31,60 +31,37 @@
 #include "src/core/helpers/WindowHelpers.h"
 #include "support/StringSupport.h"
 
-using namespace arm_compute;
-
-namespace
+namespace arm_compute
 {
-unsigned int get_num_elems_processed_per_iteration(const DataType dt)
+namespace
 {
-    unsigned int num_elems_processed_per_iteration = preferred_vector_width(CLKernelLibrary::get().get_device(), dt);
-    if(num_elems_processed_per_iteration > 8)
-    {
-        num_elems_processed_per_iteration = 8; //kernel uses only 8 lanes.
-    }
-    return num_elems_processed_per_iteration;
-}
+constexpr unsigned int vector_size_byte_opencl = 16;
 
-Status validate_arguments(const ITensorInfo &output, const float start, const float end, const float step)
+Status validate_arguments(const ITensorInfo *output, const float start, const float end, const float step)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output,
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output,
                                                          1,
                                                          DataType::U8, DataType::S8, DataType::QASYMM8,
                                                          DataType::U16, DataType::S16,
                                                          DataType::U32, DataType::S32,
                                                          DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&output);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(output);
 
     ARM_COMPUTE_RETURN_ERROR_ON_MSG((start == end), "start of the requested sequence must not be equal to the end");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(((start < end) && (step <= 0)), "step must be greater than 0 when start < end");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(((start > end) && (step >= 0)), "step must be less than 0 when start > end");
 
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!check_value_range(start, output.data_type(), output.quantization_info()), "start value is outside the range of the data type");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!check_value_range(end, output.data_type(), output.quantization_info()), "end value is outside the range of the data type");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!check_value_range(step, output.data_type(), output.quantization_info()), "step value is outside the range of the data type");
-
     ARM_COMPUTE_RETURN_ERROR_ON_MSG((start == end), "start of the requested sequence must not be equal to the end");
 
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(output.num_dimensions() != 1, "Output has to be a 1-D tensor");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(output.tensor_shape().total_size() < num_of_elements_in_range(start, end, step), "Output tensor size is incorrect");
-
-    return Status{};
-}
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo &output, const float start, const float end, const float step)
-{
-    unsigned int num_elems_processed_per_iteration = get_num_elems_processed_per_iteration(output.data_type());
-    // Auto initialize output if not initialized
-    auto_init_if_empty(output, TensorShape(num_of_elements_in_range(start, end, step)), 1, output.data_type(), output.quantization_info());
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!check_value_range(start, output->data_type(), output->quantization_info()), "start value is outside the range of the data type");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!check_value_range(end, output->data_type(), output->quantization_info()), "end value is outside the range of the data type");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!check_value_range(step, output->data_type(), output->quantization_info()), "step value is outside the range of the data type");
 
-    // Configure kernel window
-    Window win = calculate_max_window(output, Steps(num_elems_processed_per_iteration));
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->num_dimensions() != 1, "Output has to be a 1-D tensor");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->tensor_shape().total_size() < num_of_elements_in_range(start, end, step), "Output tensor size is incorrect");
 
-    AccessWindowHorizontal output_access(&output, 0, num_elems_processed_per_iteration);
-    bool                   window_changed = update_window_and_padding(win, output_access);
-    output_access.set_valid_region(win, ValidRegion(Coordinates(), TensorShape(num_of_elements_in_range(start, end, step))));
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, win);
+    return Status{};
 }
 } // namespace
 
@@ -101,12 +78,13 @@ void CLRangeKernel::configure(ICLTensor *output, const float start, const float
 void CLRangeKernel::configure(const CLCompileContext &compile_context, ICLTensor *output, const float start, const float end, const float step)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
-
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*(output->info()), start, end, step));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(output->info(), start, end, step));
 
     // Configure kernel window
-    auto win_config = validate_and_configure_window(*(output->info()), start, end, step);
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    unsigned int num_elems_processed_per_iteration = adjust_vec_size(vector_size_byte_opencl / output->info()->element_size(), output->info()->dimension(0));
+    Window       win                               = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration));
+
+    auto padding_info = get_padding_info({ output });
 
     _start  = start;
     _end    = end;
@@ -115,11 +93,11 @@ void CLRangeKernel::configure(const CLCompileContext &compile_context, ICLTensor
 
     std::string kernel_name = "range";
 
-    unsigned int num_elems_processed_per_iteration = get_num_elems_processed_per_iteration(output->info()->data_type());
     // Set build options
     CLBuildOptions build_opts;
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(output->info()->data_type()));
     build_opts.add_option("-DVECTOR_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
+    build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(output->info()->dimension(0) % num_elems_processed_per_iteration));
     build_opts.add_option("-DSTART=" + support::cpp11::to_string(start));
     build_opts.add_option("-DSTEP=" + support::cpp11::to_string(step));
     if(is_data_type_quantized_asymmetric(output->info()->data_type()))
@@ -131,7 +109,7 @@ void CLRangeKernel::configure(const CLCompileContext &compile_context, ICLTensor
     }
     // Create kernel
     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
-    ICLKernel::configure_internal(win_config.second);
+    ICLKernel::configure_internal(win);
 
     // Set config_id for enabling LWS tuning
     _config_id = kernel_name;
@@ -139,15 +117,12 @@ void CLRangeKernel::configure(const CLCompileContext &compile_context, ICLTensor
     _config_id += lower_string(string_from_data_type(output->info()->data_type()));
     _config_id += "_";
     _config_id += support::cpp11::to_string(output->info()->dimension(0));
+    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
 }
 
 Status CLRangeKernel::validate(const ITensorInfo *output, const float start, const float end, const float step)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
-
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*output, start, end, step));
-    ARM_COMPUTE_RETURN_ON_ERROR((validate_and_configure_window(*(output->clone()), start, end, step)).first);
-
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(output, start, end, step));
     return Status{};
 }
 
@@ -160,3 +135,4 @@ void CLRangeKernel::run(const Window &window, cl::CommandQueue &queue)
 
     enqueue(queue, *this, window, lws_hint());
 }
+} // namespace arm_compute
\ No newline at end of file