COMPMID-1937: Adds support for DequantizationLayer for NEON/CL.

Change-Id: I4b73edd176a277294e0e42e642460bc61210778a
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/744
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>

2 files changed, 96 insertions, 82 deletions

diff --git a/src/core/CL/cl_kernels/dequantization_layer.cl b/src/core/CL/cl_kernels/dequantization_layer.cl
index 4908bb0b31..7307700473 100644
--- a/src/core/CL/cl_kernels/dequantization_layer.cl
+++ b/src/core/CL/cl_kernels/dequantization_layer.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,51 +23,68 @@
  */
 #include "helpers.h"
 
+#if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(SCALE) && defined(OFFSET)
+
 /** This performs the dequantization of 8-bit unsigned integers to floating point.
  *
- * @param[in]  input_ptr                             Pointer to the source image. Supported data types: F16/F32
- * @param[in]  input_stride_x                        Stride of the source image in X dimension (in bytes)
- * @param[in]  input_step_x                          input_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  input_stride_y                        Stride of the source image in Y dimension (in bytes)
- * @param[in]  input_step_y                          input_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  input_stride_z                        Stride of the source tensor in Z dimension (in bytes)
- * @param[in]  input_step_z                          input_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  input_offset_first_element_in_bytes   The offset of the first element in the source image
- * @param[out] output_ptr                            Pointer to the destination image. Supported data types: same as @p input_ptr
- * @param[in]  output_stride_x                       Stride of the destination image in X dimension (in bytes)
- * @param[in]  output_step_x                         output_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  output_stride_y                       Stride of the destination image in Y dimension (in bytes)
- * @param[in]  output_step_y                         output_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  output_stride_z                       Stride of the source tensor in Z dimension (in bytes)
- * @param[in]  output_step_z                         output_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  output_offset_first_element_in_bytes  The offset of the first element in the destination image
- * @param[in]  min_max_ptr                           Pointer to the min/max vector. Minimum value in position 0, maximum value in position 1. Suppported data types: F32.
- * @param[in]  min_max_stride_x                      Stride of the min/max vector in X dimension (in bytes)
- * @param[in]  min_max_step_x                        min_max_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  min_max_offset_first_element_in_bytes The offset of the first element in the min/max vector
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
+ * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
+ * @note Quantization scale of input tensor is passed in with -DSCALE=scale.
+ * @note Quantization offset of input tensor is passed in with -DOFFSET=offset.
+ *
+ * @param[in]  input_ptr                            Pointer to the source tensor. Supported data types: QASYMM8
+ * @param[in]  input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  input_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out] output_ptr                           Pointer to the destination tensor. Supported data types: F16/F32
+ * @param[in]  output_stride_x                      Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes The offset of the first element in the destination tensor
  */
 __kernel void dequantization_layer(
     TENSOR3D_DECLARATION(input),
-    TENSOR3D_DECLARATION(output),
-    VECTOR_DECLARATION(min_max))
+    TENSOR3D_DECLARATION(output))
 {
     // Get pixels pointer
-    Tensor3D input   = CONVERT_TO_TENSOR3D_STRUCT(input);
-    Tensor3D output  = CONVERT_TO_TENSOR3D_STRUCT(output);
-    Vector   min_max = CONVERT_TO_VECTOR_STRUCT(min_max);
-
-    // min_max_value.s0 = min, min_max_value.s1 = max
-    const float2 min_max_value = vload2(0, (__global float *)min_max.ptr);
+    Tensor3D input  = CONVERT_TO_TENSOR3D_STRUCT(input);
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
 
-    const float4 vmin  = (float4)min_max_value.s0;
-    const float4 scale = (float4)((min_max_value.s1 - min_max_value.s0) / 255.0f);
+#if defined(VEC_SIZE) && defined(LAST_ACCESSED_X)
+    // Check if access on width gets out of bounds
+    // If it does shift access vector to access elements within bounds
+    const int xi = (int)(get_global_id(0) * VEC_SIZE);
+    input.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * input_stride_x;
+    output.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * output_stride_x;
 
     // Load data
-    const uchar4 data = vload4(0, (__global uchar *)input.ptr);
+    VEC_DATA_TYPE(int, VEC_SIZE)
+    val = CONVERT(VLOAD(VEC_SIZE)(0, (__global uchar *)input.ptr), VEC_DATA_TYPE(int, VEC_SIZE));
+
+    // Create scale and offset vectors
+    const VEC_DATA_TYPE(float, VEC_SIZE)
+    vscale = SCALE;
+
+    const VEC_DATA_TYPE(int, VEC_SIZE)
+    voffset = OFFSET;
 
     // Dequantize
-    const float4 res = convert_float4(data) * scale + vmin;
+    VEC_DATA_TYPE(float, VEC_SIZE)
+    res = vscale * CONVERT((val - voffset), VEC_DATA_TYPE(float, VEC_SIZE));
 
     // Store result
-    vstore4(res, 0, (__global float *)output.ptr);
+    VSTORE(VEC_SIZE)
+    (CONVERT(res, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)), 0, (__global DATA_TYPE *)output.ptr);
+#else  // !defined(VEC_SIZE) || !defined(LAST_ACCESSED_X)
+    *((__global DATA_TYPE *)(output.ptr)) = (DATA_TYPE)((float)((int)(*((__global uchar *)(input.ptr))) - (int)(OFFSET)) * (float)(SCALE));
+#endif // defined(VEC_SIZE) && defined(LAST_ACCESSED_X)
 }
+
+#endif // defined(VEC_SIZE) && defined(DATA_TYPE) && defined(SCALE) && defined(OFFSET)
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLDequantizationLayerKernel.cpp b/src/core/CL/kernels/CLDequantizationLayerKernel.cpp
index d4c1bec5f4..78cc5596dd 100644
--- a/src/core/CL/kernels/CLDequantizationLayerKernel.cpp
+++ b/src/core/CL/kernels/CLDequantizationLayerKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,6 +26,7 @@
 #include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/CLValidate.h"
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Utils.h"
@@ -36,74 +37,78 @@ using namespace arm_compute;
 
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *min_max)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output, min_max);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
-    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() < 3);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8);
 
     if(output->tensor_shape().total_size() > 0)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32);
+        ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(output);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F16, DataType::F32);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
     }
 
     return Status{};
 }
 
-std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, ITensorInfo *min_max)
+std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
 {
+    // Configure kernel window
+    Window win = calculate_max_window(*input, Steps());
+
     // Output tensor auto initialization if not yet initialized
     auto_init_if_empty(*output, input->tensor_shape(), 1, DataType::F32);
 
-    constexpr unsigned int num_elems_processed_per_iteration = 4;
-
-    // Configure window
-    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, 0, num_elems_processed_per_iteration);
-    AccessWindowStatic     min_max_access(min_max, 0, 0, 2, min_max->dimension(1));
-
-    // Update window and padding
-    bool window_changed = update_window_and_padding(win, input_access, output_access, min_max_access);
+    // CLDequantizationLayerKernel doesn't need padding so update_window_and_padding() can be skipped
+    Coordinates coord;
+    coord.set_num_dimensions(output->num_dimensions());
+    output->set_valid_region(ValidRegion(coord, output->tensor_shape()));
 
-    output_access.set_valid_region(win, input->valid_region());
-
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_tuple(err, win);
+    return std::make_tuple(Status{}, win);
 }
 } // namespace
 
 CLDequantizationLayerKernel::CLDequantizationLayerKernel()
-    : _input(nullptr), _output(nullptr), _min_max(nullptr)
+    : _input(nullptr), _output(nullptr)
 {
 }
 
-void CLDequantizationLayerKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *min_max)
+void CLDequantizationLayerKernel::configure(const ICLTensor *input, ICLTensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, min_max);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), min_max->info()));
-
-    _input   = input;
-    _output  = output;
-    _min_max = min_max;
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info()));
 
-    // Create kernel
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("dequantization_layer"));
+    _input  = input;
+    _output = output;
 
-    // Configure kernel window
-    auto win_config = validate_and_configure_window(input->info(), output->info(), min_max->info());
+    const int  vec_size_x     = 16 / output->info()->element_size();
+    const int  output_width_x = output->info()->tensor_shape().x();
+    const bool multi_access_x = (output_width_x / vec_size_x > 0);
 
-    ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
+    // Create and update the window (if needed)
+    Window win = calculate_max_window(*output->info());
+    if(multi_access_x)
+    {
+        win.set(Window::DimX,
+                Window::Dimension(win.x().start(), ceil_to_multiple(win.x().end(), vec_size_x), vec_size_x));
+    }
+    ICLKernel::configure_internal(win);
 
-    ICLKernel::configure_internal(std::get<1>(win_config));
+    // Create kernel
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DSCALE=" + float_to_string_with_full_precision(input->info()->quantization_info().scale));
+    build_opts.add_option("-DOFFSET=" + support::cpp11::to_string(input->info()->quantization_info().offset));
+    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x));
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(output->info()->data_type()));
+    build_opts.add_option_if(multi_access_x, "-DLAST_ACCESSED_X=" + support::cpp11::to_string(std::max<int>(output_width_x - vec_size_x, 0)));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("dequantization_layer", build_opts.options()));
 }
 
-Status CLDequantizationLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *min_max)
+Status CLDequantizationLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, min_max));
-    ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), output->clone().get(), min_max->clone().get())));
-
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
+    ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), output->clone().get())));
     return Status{};
 }
 
@@ -115,20 +120,12 @@ void CLDequantizationLayerKernel::run(const Window &window, cl::CommandQueue &qu
     Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), 3);
     Window slice            = window_collapsed.first_slice_window_3D();
 
-    Window min_max_window = window;
-    min_max_window.set(Window::DimX, Window::Dimension(0, 0, 0));
-    min_max_window.set(Window::DimY, Window::Dimension(0, _min_max->info()->dimension(1), 1));
-    min_max_window.set(Window::DimZ, Window::Dimension(0, 0, 0));
-
-    Window min_max_slice = min_max_window.first_slice_window_1D();
-
     do
     {
         unsigned int idx = 0;
         add_3D_tensor_argument(idx, _input, slice);
         add_3D_tensor_argument(idx, _output, slice);
-        add_1D_tensor_argument(idx, _min_max, min_max_slice);
         enqueue(queue, *this, slice);
     }
-    while(window_collapsed.slide_window_slice_3D(slice) && min_max_window.slide_window_slice_1D(min_max_slice));
+    while(window_collapsed.slide_window_slice_3D(slice));
 }