COMPMID-1722 : CL: Implement Range

Change-Id: I88da6eb5289c303b1dc91606c1560ce629746058
Reviewed-on: https://review.mlplatform.org/381
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

5 files changed, 385 insertions, 0 deletions

diff --git a/src/core/CL/CLHelpers.cpp b/src/core/CL/CLHelpers.cpp
index 0947d58973..924fb1d322 100644
--- a/src/core/CL/CLHelpers.cpp
+++ b/src/core/CL/CLHelpers.cpp
@@ -230,4 +230,30 @@ bool cl_winograd_convolution_layer_supported(const Size2D &output_tile, const Si
         return (std::find(winograd_configs_nhwc.begin(), winograd_configs_nhwc.end(), p) != winograd_configs_nhwc.end());
     }
 }
+
+size_t preferred_vector_width(const cl::Device &device, const DataType dt)
+{
+    switch(dt)
+    {
+        case DataType::U8:
+        case DataType::S8:
+        case DataType::QASYMM8:
+            return device.getInfo<CL_DEVICE_PREFERRED_VECTOR_WIDTH_CHAR>();
+        case DataType::U16:
+        case DataType::S16:
+            return device.getInfo<CL_DEVICE_PREFERRED_VECTOR_WIDTH_SHORT>();
+        case DataType::U32:
+        case DataType::S32:
+            return device.getInfo<CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT>();
+        case DataType::F16:
+        case DataType::F32:
+            return device.getInfo<CL_DEVICE_PREFERRED_VECTOR_WIDTH_FLOAT>();
+        case DataType::U64:
+        case DataType::S64:
+            return device.getInfo<CL_DEVICE_PREFERRED_VECTOR_WIDTH_LONG>();
+        default:
+            return 1;
+    }
+}
+
 } // namespace arm_compute
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 03bc8d15db..ea5342c19d 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -382,6 +382,8 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "pooling_layer_MxN_quantized_nchw", "pooling_layer_quantized.cl" },
     { "prior_box_layer_nchw", "prior_box_layer.cl" },
     { "quantization_layer", "quantization_layer.cl" },
+    { "range", "range.cl" },
+    { "range_quantized", "range.cl" },
     { "reduction_operation_x", "reduction_operation.cl" },
     { "reduction_operation_non_parallel_x", "reduction_operation.cl" },
     { "reduction_operation_y", "reduction_operation.cl" },
@@ -817,6 +819,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/quantization_layer.clembed"
     },
     {
+        "range.cl",
+#include "./cl_kernels/range.clembed"
+    },
+    {
         "reduction_operation.cl",
 #include "./cl_kernels/reduction_operation.clembed"
     },
diff --git a/src/core/CL/cl_kernels/range.cl b/src/core/CL/cl_kernels/range.cl
new file mode 100644
index 0000000000..d122c9ac89
--- /dev/null
+++ b/src/core/CL/cl_kernels/range.cl
@@ -0,0 +1,145 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "helpers.h"
+
+#if defined(VECTOR_SIZE) && defined(START) && defined(STEP) && defined(DATA_TYPE)
+/** 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
+ *
+ * @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)
+ * @param[in]  out_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  out_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__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);
+#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
+    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));
+#endif //VECTOR_SIZE == 1
+}
+
+#if defined(OFFSET_OUT) && defined(SCALE_OUT)
+
+#define CONVERT_RTE(x, type) (convert_##type##_rte((x)))
+#define CONVERT_DOWN(x, type) CONVERT_RTE(x, type)
+
+/** 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=vector_size. e.g. -DDATA_TYPE=4
+ * @note The quantization offset of the output must be passed at compile time using -DOFFSET_OUT, i.e. -DOFFSET_OUT=10
+ * @note The quantization scale of the output must be passed at compile time using -DSCALE_OUT, i.e. -DSCALE_OUT=10
+ *
+ * @param[out] out_ptr                           Pointer to the destination tensor. Supported data types: QASYMM8.
+ * @param[in]  out_stride_x                      Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  out_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  out_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__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);
+#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
+    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)));
+    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));
+#endif // VECTOR_SIZE == 1
+}
+#endif // defined(OFFSET_OUT) && defined(SCALE_OUT)
+
+#endif // defined(VECTOR_SIZE) && defined(START) && defined(STEP) && defined(DATA_TYPE)
diff --git a/src/core/CL/kernels/CLRangeKernel.cpp b/src/core/CL/kernels/CLRangeKernel.cpp
new file mode 100644
index 0000000000..f53fe87a70
--- /dev/null
+++ b/src/core/CL/kernels/CLRangeKernel.cpp
@@ -0,0 +1,159 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/CL/kernels/CLRangeKernel.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLValidate.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "utils/Utils.h"
+
+using namespace arm_compute;
+
+namespace
+{
+size_t num_of_elements_in_range(const float start, const float end, const float step)
+{
+    ARM_COMPUTE_ERROR_ON_MSG(step == 0, "CLRange Step cannot be 0");
+    return size_t(std::ceil((end - start) / step));
+}
+
+unsigned int get_num_elems_processed_per_iteration(const DataType dt)
+{
+    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;
+}
+
+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,
+                                                         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_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(!utils::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(!utils::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(!utils::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());
+
+    // Configure kernel window
+    Window win = calculate_max_window(output, Steps(num_elems_processed_per_iteration));
+
+    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);
+}
+} // namespace
+
+CLRangeKernel::CLRangeKernel()
+    : _start(0), _end(1), _step(1), _output(nullptr)
+{
+}
+
+void CLRangeKernel::configure(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));
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(*(output->info()), start, end, step);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+
+    _start  = start;
+    _end    = end;
+    _step   = step;
+    _output = output;
+
+    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("-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()))
+    {
+        build_opts.add_option("-DOFFSET_OUT=" + support::cpp11::to_string(output->info()->quantization_info().offset));
+        build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(output->info()->quantization_info().scale));
+        kernel_name += "_quantized";
+    }
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
+    ICLKernel::configure_internal(win_config.second);
+
+    // Set config_id for enabling LWS tuning
+    _config_id = kernel_name;
+    _config_id += "_";
+    _config_id += lower_string(string_from_data_type(output->info()->data_type()));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(0));
+}
+
+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);
+
+    return Status{};
+}
+
+void CLRangeKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+    unsigned int idx = 0;
+    add_1D_tensor_argument(idx, _output, window);
+
+    enqueue(queue, *this, window, lws_hint());
+}
diff --git a/src/runtime/CL/functions/CLRange.cpp b/src/runtime/CL/functions/CLRange.cpp
new file mode 100644
index 0000000000..b2cd472d02
--- /dev/null
+++ b/src/runtime/CL/functions/CLRange.cpp
@@ -0,0 +1,49 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/CL/functions/CLRange.h"
+
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/CL/kernels/CLRangeKernel.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/runtime/CL/CLScheduler.h"
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+
+void CLRange::configure(ICLTensor *output, const float start, const float end, const float step)
+{
+    auto k = arm_compute::support::cpp14::make_unique<CLRangeKernel>();
+    k->set_target(CLScheduler::get().target());
+    k->configure(output, start, end, step);
+    _kernel = std::move(k);
+
+    // Tune kernels
+    CLScheduler::get().tune_kernel_static(*_kernel);
+}
+
+Status CLRange::validate(const ITensorInfo *output, const float start, const float end, const float step)
+{
+    return CLRangeKernel::validate(output, start, end, step);
+}