src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp

/*
 * Copyright (c) 2017-2019 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/CLGEMMMatrixMultiplyKernel.h"

#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/CL/OpenCL.h"
#include "arm_compute/core/Error.h"
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/Utils.h"
#include "arm_compute/core/Window.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"

#include <set>
#include <string>

namespace arm_compute
{
using namespace arm_compute::misc::shape_calculator;

namespace
{
using ElementsProcessed = Steps;

inline Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info,
                                 bool fp_mixed_precision)
{
    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output);
    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input0);
    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32);
    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
    ARM_COMPUTE_RETURN_ERROR_ON_MSG((fp_mixed_precision && (input0->data_type() != DataType::F16)), "Mixed precision floating point is supported only for F16 data");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the matrix A must be <= 4");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the matrix B must be <= 3");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_interleaved_transposed && reshape_info.reinterpret_input_as_3d(), "The input tensor cannot be reinterpreted as 3D if is_interleaved_transposed is true");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 2 && reshape_info.reinterpret_input_as_3d(), "The input1 tensor cannot have more than 2 dimensions if input0 has to be reinterpreted as 3D");

    if(!is_interleaved_transposed)
    {
        ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != input1->dimension(1));
    }
    else
    {
        GEMMRHSMatrixInfo rhs_info;
        GEMMLHSMatrixInfo lhs_info;
        const int         m                         = reshape_info.m();
        const int         n                         = reshape_info.n();
        const int         k                         = reshape_info.k();
        const int         mult_transpose1xW_width   = reshape_info.mult_transpose1xW_width();
        const int         mult_interleave4x4_height = reshape_info.mult_interleave4x4_height();
        rhs_info.n0                                 = 16 / input1->element_size();
        rhs_info.k0                                 = 1;
        rhs_info.h0                                 = mult_transpose1xW_width;
        rhs_info.interleave                         = false;
        rhs_info.transpose                          = false;
        lhs_info.m0                                 = 4;
        lhs_info.k0                                 = 4;
        lhs_info.v0                                 = mult_interleave4x4_height;
        lhs_info.interleave                         = true;
        lhs_info.transpose                          = true;

        TensorShape tensor_shape0{ input0->tensor_shape() };
        tensor_shape0.set(0, k);
        tensor_shape0.set(1, m);

        TensorShape tensor_shape1{ input1->tensor_shape() };
        tensor_shape1.set(0, n);
        tensor_shape1.set(1, k);

        const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0);
        const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1);

        const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_lhs_reshaped_shape(tensor_info0, lhs_info));
        const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info));

        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info_reshaped0);
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1);
    }

    if(output->total_size() != 0)
    {
        const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, is_interleaved_transposed, reshape_info));
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output);
    }

    return Status{};
}

inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output,
                                                               bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info, GPUTarget gpu_target,
                                                               ElementsProcessed &num_elements_processed)
{
    bool   window_changed = false;
    Window win{};
    Window win_out{};

    const DataType data_type                           = input0->data_type();
    unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
    unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
    bool           reinterpret_input_as_3d             = reshape_info.reinterpret_input_as_3d();
    bool           reinterpret_output_as_3d            = (reshape_info.depth_output_gemm3d() != 0);

    // In case both input and output have to be reinterpreted as 3D tensors,
    // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
    if(reinterpret_input_as_3d == reinterpret_output_as_3d)
    {
        reinterpret_input_as_3d  = false;
        reinterpret_output_as_3d = false;
    }

    // Output tensor auto inizialitation if not yet initialized
    auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, is_interleaved_transposed, reshape_info)));

    TensorInfo tmp_info(*output);

    if(reinterpret_output_as_3d)
    {
        // Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
        // the window needs to be constructed on the 2D collapsed version of the tensor
        TensorShape tmp_shape(output->tensor_shape());
        tmp_shape.collapse(2U, 1U);
        tmp_info.set_tensor_shape(tmp_shape);
    }

    if(is_interleaved_transposed)
    {
        // reinterpret_input_as_3d is not supported if is_interleaved_transposed is set
        ARM_COMPUTE_ERROR_ON(reshape_info.reinterpret_input_as_3d());

        // Configure kernel window
        num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(data_type);
        num_elems_processed_per_iteration_y = 4;

        // Note: bottom paddings are calculated manually as the output can be reinterpreted as 3D tensor
        // The only way to set properly the paddings, it is to set those explicitly through the AccessWindowStatic
        const int m          = reshape_info.m();
        const int bottom_pad = (num_elems_processed_per_iteration_y - (m % num_elems_processed_per_iteration_y)) % num_elems_processed_per_iteration_y;

        win     = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
        win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));

        AccessWindowRectangle input0_access(input0, 0, 0, num_elems_processed_per_iteration_y, 1, 1.f, 0.25f);
        AccessWindowStatic    input1_access(input1, 0, 0,
                                            ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x),
                                            ceil_to_multiple(input1->dimension(1), num_elems_processed_per_iteration_y));
        AccessWindowStatic output_access(output, 0, 0,
                                         ceil_to_multiple(output->dimension(0), num_elems_processed_per_iteration_x),
                                         output->dimension(1) + bottom_pad);

        window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop
                         update_window_and_padding(win_out, output_access);              // window used to update the padding requirements of output tensor

        output_access.set_valid_region(win_out, ValidRegion(Coordinates(0, 0), output->tensor_shape()));
    }
    else // The input tensors have not been reshaped
    {
        // Special case for 1xN, 2xN, 3xN and 4xN input0 tensor. num_elems_processed_per_iteration_x is set up for the default case.
        num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(data_type);
        num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->dimension(1)), 4);

        // Note: bottom paddings are calculated manually as the output can be reinterpreted as 3D tensor
        // The only way to set properly the paddings, it is to set those explicitly through the AccessWindowStatic
        const int m          = reinterpret_input_as_3d ? input0->tensor_shape()[1] * input0->tensor_shape()[2] : input0->tensor_shape()[1];
        const int bottom_pad = (num_elems_processed_per_iteration_y - (m % num_elems_processed_per_iteration_y)) % num_elems_processed_per_iteration_y;

        // Create kernels according to the architecture, data type and input size.
        GPUTarget arch_target = get_arch_from_target(gpu_target);
        if(arch_target == GPUTarget::BIFROST && data_type == DataType::F32)
        {
            num_elems_processed_per_iteration_x = (input1->dimension(0) <= 1000 && input0->num_dimensions() == 1) ? 2 : 4;
        }

        // Configure window
        win     = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
        win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));

        AccessWindowStatic input0_access(input0, 0, 0, input0->dimension(0), input0->dimension(1) + bottom_pad);
        AccessWindowStatic input1_access(input1, 0, 0, ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x), input1->dimension(1));
        AccessWindowStatic output_access(output, 0, 0,
                                         ceil_to_multiple(output->dimension(0), num_elems_processed_per_iteration_x),
                                         output->dimension(1) + bottom_pad);

        window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop
                         update_window_and_padding(win_out, output_access);              // window used to update the padding requirements of output tensor

        Coordinates coord;
        coord.set_num_dimensions(output->num_dimensions());
        output_access.set_valid_region(win_out, ValidRegion(coord, output->tensor_shape()));
    }

    // Collapse along the Z direction
    // This collapse needs to be here in order to tune the Z dimension of LWS
    Window             collapsed             = win;
    const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u);
    collapsed                                = win.collapse(win, dimension_to_collapse);

    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
    return std::make_pair(err, collapsed);
}
} // namespace

CLGEMMMatrixMultiplyKernel::CLGEMMMatrixMultiplyKernel()
    : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_input_as_3d(false), _reinterpret_output_as_3d(false)
{
}

void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha, bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info,
                                           bool fp_mixed_precision)
{
    ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);

    // Perform validate step
    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), is_interleaved_transposed, reshape_info, fp_mixed_precision));

    _input0                   = input0;
    _input1                   = input1;
    _output                   = output;
    _reinterpret_input_as_3d  = reshape_info.reinterpret_input_as_3d();
    _reinterpret_output_as_3d = (reshape_info.depth_output_gemm3d() != 0);

    // In case both input and output have to be reinterpreted as 3D tensors,
    // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
    if(_reinterpret_input_as_3d == _reinterpret_output_as_3d)
    {
        _reinterpret_input_as_3d  = false;
        _reinterpret_output_as_3d = false;
    }

    // Check if we need to slide the matrix B
    const unsigned int num_dimensions_input0 = _reinterpret_input_as_3d ? _input0->info()->num_dimensions() - 1 : _input0->info()->num_dimensions();

    _slide_matrix_b = (_input1->info()->num_dimensions() >= num_dimensions_input0);

    const DataType data_type = input0->info()->data_type();

    // Get target architecture
    GPUTarget gpu_target = get_target();

    ElementsProcessed num_elements_processed{};

    // Configure kernel window
    auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), is_interleaved_transposed, reshape_info, gpu_target, num_elements_processed);
    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
    ICLKernel::configure_internal(win_config.second);

    // Create build options
    CLBuildOptions build_opts;

    // Only define ALPHA when alpha is not 1.0f. This avoids performing unnecessary multiplications.
    if(std::abs(1.0f - alpha) > 0.00001f)
    {
        build_opts.add_option("-DALPHA=" + float_to_string_with_full_precision(alpha));
    }
    build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
    build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
    build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1)));
    build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(2)));

    // Do not slide matrix B if _slide_matrix_b = false
    build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2)));

    const bool is_bifrost = get_arch_from_target(gpu_target) == GPUTarget::BIFROST;

    std::string kernel_name;
    if(is_interleaved_transposed)
    {
        const int mult_transpose1xW_width   = reshape_info.mult_transpose1xW_width();
        const int mult_interleave4x4_height = reshape_info.mult_interleave4x4_height();

        build_opts.add_option("-DCOLS_B=" + support::cpp11::to_string(input1->info()->dimension(0)));
        build_opts.add_option("-DMULT_TRANSPOSE1XW_WIDTH=" + support::cpp11::to_string(mult_transpose1xW_width));
        build_opts.add_option("-DMULT_INTERLEAVE4X4_HEIGHT=" + support::cpp11::to_string(mult_interleave4x4_height));

        if(is_data_type_float(data_type) && is_bifrost)
        {
            kernel_name = "gemm_mm_interleaved_transposed_" + lower_string(string_from_data_type(data_type)) + "_bifrost";
        }
        else
        {
            kernel_name = "gemm_mm_interleaved_transposed_" + lower_string(string_from_data_type(data_type));
            if(fp_mixed_precision && data_type == DataType::F16)
            {
                // currently wider accumulator is only supported for fp16 kernels.
                kernel_name += "_acc32";
            }
        }
    }
    else // The input tensors have not been reshaped
    {
        build_opts.add_option("-DCOLS_A=" + support::cpp11::to_string(input0->info()->dimension(0)));
        build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));

        // Create kernels according to the architecture, data type and input size.
        if(is_data_type_float(data_type) && is_bifrost)
        {
            kernel_name = "gemm_mm_floating_point";

            if(input0->info()->num_dimensions() != 1)
            {
                kernel_name += "_" + lower_string(string_from_data_type(data_type)) + "_bifrost";
                if(fp_mixed_precision && data_type == DataType::F16)
                {
                    // currently wider accumulator is only supported for fp16 kernels.
                    kernel_name += "_acc32";
                }
            }
            else if(input1->info()->dimension(0) <= 1000 && data_type == DataType::F32)
            {
                // The first kernel is optimized for the case of 1000 or less output elements (e.g. FC8 of AlexNet and VGG-16, and
                // FC1 of Inception v3). The second kernel is optimized for the case of greater than 1000 output elements (e.g.
                // FC6 and FC7 of AlexNet and VGG-16).
                kernel_name += "_" + lower_string(string_from_data_type(data_type)) + "_bifrost_1000";
            }

            // The work-group size equal to the Bifrost quad size has been proved to be optimal for these kernels
            // via exhaustive autotuning over a range of representative layer configurations.
            set_lws_hint(cl::NDRange(4));
        }
        else // (MIDGARD and F32) or (F16)
        {
            kernel_name = "gemm_mm_floating_point";
        }
        build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_Y=" + support::cpp11::to_string(num_elements_processed.y()));
        build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_X=" + support::cpp11::to_string(num_elements_processed.x()));
    }

    // Create kernel
    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));

    // Set config_id for enabling LWS tuning
    _config_id = "gemm_";
    _config_id += (is_interleaved_transposed ? "reshaped_" : "");
    _config_id += (fp_mixed_precision ? "fp_mixed_" : "");
    _config_id += (_reinterpret_input_as_3d ? "3di_" : "");
    _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
    _config_id += lower_string(string_from_data_type(input0->info()->data_type()));
    _config_id += "_";
    _config_id += support::cpp11::to_string(output->info()->dimension(1));
    _config_id += "_";
    _config_id += support::cpp11::to_string(output->info()->dimension(0));
    _config_id += "_";
    _config_id += support::cpp11::to_string(output->info()->dimension(2));
    _config_id += "_";
    _config_id += support::cpp11::to_string(output->info()->dimension(3));
    _config_id += "_";
    _config_id += (is_interleaved_transposed ? support::cpp11::to_string(input1->info()->dimension(0)) : support::cpp11::to_string(input1->info()->dimension(1)));
}

Status CLGEMMMatrixMultiplyKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, bool is_interleaved_transposed,
                                            const GEMMReshapeInfo &reshape_info, GPUTarget gpu_target, bool fp_mixed_precision)
{
    // Note: num_elements_processed will be set in validate_and_configure_window()
    ElementsProcessed num_elements_processed{};
    ARM_COMPUTE_UNUSED(alpha);
    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, is_interleaved_transposed, reshape_info, fp_mixed_precision));
    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
                                                              input1->clone().get(),
                                                              output->clone().get(),
                                                              is_interleaved_transposed,
                                                              reshape_info,
                                                              gpu_target,
                                                              num_elements_processed)
                                .first);

    return Status{};
}

void CLGEMMMatrixMultiplyKernel::run(const Window &window, cl::CommandQueue &queue)
{
    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);

    if(_input1->info()->num_dimensions() < 3)
    {
        // The stride_z for matrix B must be zero if we do not slice
        ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0);
    }

    Window slice          = window.first_slice_window_3D();
    Window slice_matrix_b = slice;

    slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
    slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));

    if(_reinterpret_input_as_3d)
    {
        // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor
        const unsigned int idx0                  = 3 * num_arguments_per_2D_tensor() + 3;
        const unsigned int total_cross_plane_pad = _input0->info()->padding().top + _input0->info()->padding().bottom;
        _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
    }

    if(_reinterpret_output_as_3d)
    {
        // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor
        const unsigned int idx0                  = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0);
        const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom;
        _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
    }

    do
    {
        Window slice_b = slice;
        // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
        // This scenario can happen when the matrix multiplication is used to perform a convolution operation
        if(!_slide_matrix_b)
        {
            slice_b = slice_matrix_b;
        }

        unsigned int idx = 0;
        add_2D_tensor_argument(idx, _input0, slice);
        add_2D_tensor_argument(idx, _input1, slice_b);
        add_2D_tensor_argument(idx, _output, slice);
        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));
        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));
        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
        enqueue(queue, *this, slice, lws_hint());
    }
    while(window.slide_window_slice_3D(slice));
}
} // namespace arm_compute