COMPMID-1336: Add CLArithmeticAddition support for QASYMM8

Change-Id: Ice2bb644841fdea4e776872ff5481eb927e66bd1
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/137714
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

3 files changed, 115 insertions, 4 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 712a1179a6..aa11edf9ec 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -149,6 +149,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "accumulate_weighted", "accumulate.cl" },
     { "activation_layer", "activation_layer.cl" },
     { "activation_layer_qa8", "activation_layer_qa8.cl" },
+    { "arithmetic_add_quantized", "arithmetic_op_quantized.cl" },
     { "arithmetic_add", "arithmetic_op.cl" },
     { "arithmetic_sub", "arithmetic_op.cl" },
     { "arithmetic_div", "arithmetic_op.cl" },
@@ -416,6 +417,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/arithmetic_op.clembed"
     },
     {
+        "arithmetic_op_quantized.cl",
+#include "./cl_kernels/arithmetic_op_quantized.clembed"
+    },
+    {
         "bitwise_op.cl",
 #include "./cl_kernels/bitwise_op.clembed"
     },
diff --git a/src/core/CL/cl_kernels/arithmetic_op_quantized.cl b/src/core/CL/cl_kernels/arithmetic_op_quantized.cl
new file mode 100644
index 0000000000..82e92e32a8
--- /dev/null
+++ b/src/core/CL/cl_kernels/arithmetic_op_quantized.cl
@@ -0,0 +1,89 @@
+/*
+ * Copyright (c) 2016-2018 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"
+
+#ifdef SATURATE
+#define ADD(x, y) add_sat((x), (y))
+#define SUB(x, y) sub_sat((x), (y))
+#else /* SATURATE */
+#define ADD(x, y) (x) + (y)
+#define SUB(x, y) (x) - (y)
+#endif /* SATURATE */
+
+#if defined(OFFSET)
+
+/** This function adds two tensors.
+ *
+ * @attention The quantization offset must be passed at compile time using -DOFFSET, i.e. -DOFFSET=10
+ * @attention To perform saturating operation -DSATURATE has to be passed to the compiler otherwise wrapping policy will be used.
+ *
+ * @param[in]  in1_ptr                           Pointer to the source tensor. Supported data types: QASYMM8
+ * @param[in]  in1_stride_x                      Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  in1_step_x                        in1_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  in1_stride_y                      Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  in1_step_y                        in1_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  in1_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  in1_step_z                        in1_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  in1_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[in]  in2_ptr                           Pointer to the source tensor. Supported data types: same as @p in1_ptr
+ * @param[in]  in2_stride_x                      Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  in2_step_x                        in2_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  in2_stride_y                      Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  in2_step_y                        in2_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  in2_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  in2_step_z                        in2_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  in2_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] out_ptr                           Pointer to the destination tensor. Supported data types: same as @p in1_ptr
+ * @param[in]  out_stride_x                      Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  out_step_x                        out_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  out_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  out_step_y                        out_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  out_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  out_step_z                        out_stride_z * number of elements along Z 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 arithmetic_add_quantized(
+    TENSOR3D_DECLARATION(in1),
+    TENSOR3D_DECLARATION(in2),
+    TENSOR3D_DECLARATION(out))
+{
+    // Get pixels pointer
+    Tensor3D in1 = CONVERT_TO_TENSOR3D_STRUCT(in1);
+    Tensor3D in2 = CONVERT_TO_TENSOR3D_STRUCT(in2);
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out);
+
+    // Load values
+    const short16 in_a   = CONVERT(vload16(0, (__global uchar *)in1.ptr), short16);
+    const short16 in_b   = CONVERT(vload16(0, (__global uchar *)in2.ptr), short16);
+    const short16 offset = OFFSET;
+
+    // Calculate result
+    short16 res = ADD(in_a, SUB(in_b, offset));
+
+    res = max((short16)0, min(res, (short16)255));
+
+    // Store result
+    vstore16(CONVERT(res, uchar16), 0, (__global uchar *)out.ptr);
+}
+#endif /* defined(OFFSET) */
diff --git a/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp b/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp
index b0177ab8b6..011807ad88 100644
--- a/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp
+++ b/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp
@@ -37,9 +37,15 @@ Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2,
 {
     ARM_COMPUTE_UNUSED(policy);
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input1);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QASYMM8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input2);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QASYMM8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
+
+    const bool is_qasymm = is_data_type_quantized_asymmetric(input1.data_type()) || is_data_type_quantized_asymmetric(input2.data_type());
+    if(is_qasymm)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &input2);
+    }
 
     const TensorShape out_shape = TensorShape::broadcast_shape(input1.tensor_shape(), input2.tensor_shape());
 
@@ -50,12 +56,16 @@ Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2,
     if(output.total_size() > 0)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&output);
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
         ARM_COMPUTE_RETURN_ERROR_ON_MSG((output.data_type() == DataType::U8) && ((input1.data_type() != DataType::U8) || (input2.data_type() != DataType::U8)),
                                         "Output can only be U8 if both inputs are U8");
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output.tensor_shape(), 0),
                                         "Wrong shape for output");
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(&input1, &output);
+        if(is_qasymm)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &output);
+        }
     }
 
     return Status{};
@@ -124,6 +134,8 @@ void CLArithmeticAdditionKernel::configure(const ICLTensor *input1, const ICLTen
 
     const bool has_float_out = is_data_type_float(output->info()->data_type());
 
+    std::string kernel_name = "arithmetic_add";
+
     // Set kernel build options
     std::set<std::string> build_opts;
     build_opts.emplace((policy == ConvertPolicy::WRAP || has_float_out) ? "-DWRAP" : "-DSATURATE");
@@ -134,9 +146,14 @@ void CLArithmeticAdditionKernel::configure(const ICLTensor *input1, const ICLTen
     {
         build_opts.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input1->info()->fixed_point_position()));
     }
+    else if(is_data_type_quantized_asymmetric(input1->info()->data_type()))
+    {
+        build_opts.emplace("-DOFFSET=" + support::cpp11::to_string(input1->info()->quantization_info().offset));
+        kernel_name += "_quantized";
+    }
 
     // Create kernel
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("arithmetic_add", build_opts));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts));
 
     ICLKernel::configure(win_config.second);
 }