Port OpenCL Quantization to new API

Partially resolves: COMPMID-4193

Change-Id: Ie8367769c690442a0e30383c67851b50ab7c6742
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5231
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 253 insertions, 0 deletions

diff --git a/src/core/gpu/cl/kernels/ClQuantizationKernel.cpp b/src/core/gpu/cl/kernels/ClQuantizationKernel.cpp
new file mode 100644
index 0000000000..ea56289157
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClQuantizationKernel.cpp
@@ -0,0 +1,180 @@
+/*
+ * Copyright (c) 2017-2021 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 "src/core/gpu/cl/kernels/ClQuantizationKernel.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
+#include "src/core/AccessWindowStatic.h"
+#include "src/core/CL/CLValidate.h"
+#include "src/core/helpers/WindowHelpers.h"
+#include "support/Cast.h"
+#include "support/StringSupport.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+namespace kernels
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F32, DataType::F16);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
+
+    // Output must always be initialized
+    ARM_COMPUTE_RETURN_ERROR_ON(dst->tensor_shape().total_size() == 0);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QASYMM16);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst);
+
+    return Status{};
+}
+} // namespace
+
+ClQuantizationKernel::ClQuantizationKernel()
+{
+}
+
+void ClQuantizationKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+
+    auto padding_info = get_padding_info({ src, dst });
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst));
+
+    const int  vec_size_x     = 16 / src->element_size();
+    const int  input_width_x  = src->tensor_shape().x();
+    const bool multi_access_x = (input_width_x / vec_size_x > 0);
+
+    const UniformQuantizationInfo qinfo            = dst->quantization_info().uniform();
+    const DataType                output_data_type = dst->data_type();
+
+    float   scale_to_apply  = qinfo.scale;
+    int32_t offset_to_apply = qinfo.offset;
+    if(is_data_type_quantized_asymmetric(src->data_type()))
+    {
+        /*
+         * In case of requantization of a quantized input tensor to an output tensor with another quantization
+         * instead of of apply dequantization and then a quantization functions, we just compute new scale and
+         * offset to apply.
+         *
+         * Assuming:
+         *   - q_i as input quantized value
+         *   - q_o as output quantized value
+         *   - z_i as input quantization offset value
+         *   - z_o as output quantization offset value
+         *   - s_i as input quantization scale value
+         *   - s_o as output quantization scale value
+         *   - z_n as new quantization offset value
+         *   - s_n as new quantization scale value
+         *
+         * q_o = ( q_i - z_i ) * s_i / s_o + z_o
+         *
+         * We can rewrite the formula as:
+         *
+         * q_o = ( q_i * s_i / s_o ) - z_i * s_i / s_o + z_o
+         *
+         * q_o = q_i / s_n + z_n
+         *
+         * Where:
+         *
+         * s_n = s_o / s_i
+         *
+         * z_n = - z_i * s_i / s_o + z_o
+         *
+         */
+        const UniformQuantizationInfo qinfo_in = src->quantization_info().uniform();
+        scale_to_apply /= qinfo_in.scale;
+        // In order to minimize flooring we convert the offset to a float,
+        // then compute the new offset in the float domain,
+        // finally we convert it back as int32_t
+        offset_to_apply -= static_cast<int32_t>(static_cast<float>(qinfo_in.offset) * qinfo_in.scale / qinfo.scale);
+    }
+
+    // Create kernel
+    CLBuildOptions build_opts;
+    build_opts.add_option_if(is_data_type_float(src->data_type()), "-DIS_FLOAT");
+    build_opts.add_option("-DSCALE=" + float_to_string_with_full_precision(scale_to_apply));
+    build_opts.add_option("-DOFFSET=" + support::cpp11::to_string(offset_to_apply));
+    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x));
+    build_opts.add_option("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(src->data_type()));
+    build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output_data_type));
+    build_opts.add_option_if(multi_access_x, "-DLAST_ACCESSED_X=" + support::cpp11::to_string(std::max<int>(input_width_x - vec_size_x, 0)));
+    std::pair<int, int> min_max_quant_values = quantization::get_min_max_values_from_quantized_data_type(output_data_type);
+    build_opts.add_option("-DMIN_QUANT_VAL=" + support::cpp11::to_string(min_max_quant_values.first));
+    build_opts.add_option("-DMAX_QUANT_VAL=" + support::cpp11::to_string(min_max_quant_values.second));
+
+    _kernel = create_kernel(compile_context, "quantization_layer", build_opts.options());
+
+    // Configure kernel window
+    Window win = calculate_max_window(*src, Steps());
+    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);
+
+    dst->set_valid_region(ValidRegion(Coordinates(), dst->tensor_shape()));
+
+    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
+}
+
+Status ClQuantizationKernel::validate(const ITensorInfo *src, const ITensorInfo *dst)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst));
+    return Status{};
+}
+
+void ClQuantizationKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
+    auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
+
+    Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), 3);
+    Window slice            = window_collapsed.first_slice_window_3D();
+
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, src, slice);
+        add_3D_tensor_argument(idx, dst, slice);
+        enqueue(queue, *this, slice, lws_hint());
+    }
+    while(window_collapsed.slide_window_slice_3D(slice));
+}
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
diff --git a/src/core/gpu/cl/kernels/ClQuantizationKernel.h b/src/core/gpu/cl/kernels/ClQuantizationKernel.h
new file mode 100644
index 0000000000..20822cf9c9
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClQuantizationKernel.h
@@ -0,0 +1,73 @@
+/*
+ * Copyright (c) 2017-2021 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.
+ */
+#ifndef ARM_COMPUTE_CL_QUANTIZATION_KERNEL_H
+#define ARM_COMPUTE_CL_QUANTIZATION_KERNEL_H
+
+#include "src/core/common/Macros.h"
+#include "src/core/gpu/cl/ClCompileContext.h"
+#include "src/core/gpu/cl/IClKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+namespace opencl
+{
+namespace kernels
+{
+/** Interface for the quantization layer kernel.
+ *
+ * @note The implementation supports only 3D input tensors.
+ */
+class ClQuantizationKernel : public IClKernel
+{
+public:
+    /** Default constructor */
+    ClQuantizationKernel();
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClQuantizationKernel);
+    /** Set the input, output.
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  src             Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F32/F16.
+     * @param[out] dst             Destination tensor info with the same dimensions of input. Data types supported: QASYMM8/QASYMM8_SIGNED/QASYMM16.
+     *
+     * @note Output auto initialization is not supported by this kernel
+     */
+    void configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst);
+    /** Static function to check if given info will lead to a valid configuration of @ref ClQuantizationKernel
+     *
+     * @param[in] src Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F32/F16.
+     * @param[in] dst Destination tensor info with the same dimensions of input. Data types supported: QASYMM8/QASYMM8_SIGNED/QASYMM16.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *dst);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
+};
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
+#endif /*ARM_COMPUTE_CL_QUANTIZATION_KERNEL_H */