src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp

/*
 * Copyright (c) 2019-2021, 2023 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/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.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/CL/OpenCL.h"
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/core/utils/ActivationFunctionUtils.h"
#include "arm_compute/core/utils/StringUtils.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "src/core/AccessWindowStatic.h"
#include "src/core/CL/CLUtils.h"
#include "src/core/helpers/AutoConfiguration.h"
#include "src/core/helpers/WindowHelpers.h"
#include "src/core/utils/helpers/float_ops.h"
#include "support/Cast.h"
#include "support/StringSupport.h"

namespace arm_compute
{
namespace opencl
{
namespace kernels
{
namespace
{
using ElementsProcessed = Steps;

Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info,
                          const GEMMRHSMatrixInfo &rhs_info,
                          const GEMMKernelInfo    &gemm_info)
{
    ARM_COMPUTE_UNUSED(alpha);
    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src0, src1, dst);
    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src0, 1, DataType::F32, DataType::F16);
    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, src1);
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(src0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.k0 & (rhs_info.k0 - 1)) && rhs_info.k0 != 3), "Only 2,3,4,8,16 are supported for k0");
    ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 > 16);
    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 1 || lhs_info.m0 > 8);
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.n0 & (rhs_info.n0 - 1)) && rhs_info.n0 != 3), "Only 2,3,4,8,16 are supported for n0");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG((gemm_info.reinterpret_input_as_3d || gemm_info.depth_output_gemm3d != 0) && (src2 != nullptr)
                                    && (!gemm_info.broadcast_bias),
                                    "Bias addition only supported with broadcast mode in case the input or dst has to be reinterpreted as 3D");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision, "Mixed precision not supported");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.export_to_cl_image, "Export to CLImage not supported for GEMM native");

    const unsigned int m = gemm_info.m;
    const unsigned int n = gemm_info.n;
    const unsigned int k = gemm_info.k;

    ARM_COMPUTE_UNUSED(m);
    ARM_COMPUTE_UNUSED(n);
    ARM_COMPUTE_UNUSED(k);

    ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(0) != k);
    ARM_COMPUTE_RETURN_ERROR_ON(src1->dimension(0) != n);
    ARM_COMPUTE_RETURN_ERROR_ON(src1->dimension(1) != k);
    if(gemm_info.reinterpret_input_as_3d)
    {
        ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) * src0->dimension(2) != m);
    }
    else
    {
        ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) != m);
    }

    if(src2 != nullptr && !(helpers::float_ops::is_zero(beta)))
    {
        const unsigned int src2_dim0 = src2->dimension(0);
        const unsigned int src2_dim1 = src2->dimension(1);

        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src2, src1);
        if(gemm_info.broadcast_bias)
        {
            ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim1 != 1 || src2_dim0 != n), "Incorrect dimension of bias matrix which is to be broadcasted");
        }
        else
        {
            ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim0 != n || src2_dim1 != m), "Incorrect dimension of bias matrix");
        }
    }

    if(dst->total_size() != 0)
    {
        const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info));
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst);
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, dst);
    }

    return Status{};
}

std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info,
                                                        const GEMMRHSMatrixInfo &rhs_info,
                                                        const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed)
{
    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             = gemm_info.reinterpret_input_as_3d;
    bool          reinterpret_output_as_3d            = gemm_info.depth_output_gemm3d != 0;

    Window win{};
    Window win_out{};
    bool   window_changed = false;

    // In case both input and dst 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_output_as_3d = false;
    }

    // dst tensor auto initialization if not yet initialized
    auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)));

    TensorInfo tmp_info(*dst);

    if(reinterpret_output_as_3d)
    {
        // Since the dst 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(dst->tensor_shape());
        tmp_shape.collapse(2U, 1U);
        tmp_info.set_tensor_shape(tmp_shape);
    }

    // Configure kernel window
    num_elems_processed_per_iteration_x = rhs_info.n0;
    num_elems_processed_per_iteration_y = lhs_info.m0;

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

    AccessWindowStatic src0_access(src0, 0, 0,
                                   src0->dimension(0),
                                   src0->dimension(1));
    AccessWindowStatic src1_access(src1, 0, 0,
                                   ceil_to_multiple(src1->dimension(0), num_elems_processed_per_iteration_x),
                                   src1->dimension(1));
    AccessWindowStatic dst_access(dst, 0, 0,
                                  dst->dimension(0),
                                  dst->dimension(1));

    if(src2 != nullptr)
    {
        const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;

        AccessWindowStatic src2_access(src2, 0, 0,
                                       ceil_to_multiple(src2->dimension(0), bias_processed_per_iteration_x),
                                       src2->dimension(1));

        window_changed = update_window_and_padding(win, src0_access, src1_access, src2_access) || // window used by the execute_window_loop
                         update_window_and_padding(win_out, dst_access);                          // window used to update the padding requirements of dst tensor
    }
    else
    {
        window_changed = update_window_and_padding(win, src0_access, src1_access) || // window used by the execute_window_loop
                         update_window_and_padding(win_out, dst_access);             // window used to update the padding requirements of dst tensor
    }

    // 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>(dst->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

ClGemmMatrixMultiplyNativeKernel::ClGemmMatrixMultiplyNativeKernel()
{
    _type = CLKernelType::GEMM;
}

void ClGemmMatrixMultiplyNativeKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, float alpha,
                                                 float                    beta,
                                                 const GEMMLHSMatrixInfo &lhs_info,
                                                 const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
{
    ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);

    // dst tensor auto initialization if not yet initialized
    auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)));

    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info));

    auto padding_info         = get_padding_info({ src0, dst });
    _reinterpret_input_as_3d  = gemm_info.reinterpret_input_as_3d;
    _reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
    _use_dummy_work_items     = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
    _add_bias                 = src2 != nullptr;

    // In case both input and dst 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_src0 = src0->num_dimensions();
    _slide_matrix_b                        = (src1->num_dimensions() >= num_dimensions_src0);

    ElementsProcessed num_elements_processed{};

    // Configure kernel window
    auto win_config = validate_and_configure_window(src0, src1, src2 != nullptr ? src2 : nullptr, dst, lhs_info, rhs_info, gemm_info, num_elements_processed);
    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
    IClKernel::configure_internal(win_config.second);

    // If _reinterpret_input_as_3d = _reinterpret_output_as_3d = true,
    // we will dispatch a batched-GEMM to reduce the complexity of the address calculation within the OpenCL kernel.
    // This means that the actual m used by the kernel is given by dst->dimension(1) and not by gemm_info.m
    const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : dst->dimension(1);

    const unsigned int h_gemm_3d = _reinterpret_output_as_3d ? dst->dimension(1) : src0->dimension(1);
    const unsigned int d_gemm_3d = _reinterpret_output_as_3d ? dst->dimension(2) : src0->dimension(2);

    // Calculate partial (store instead of load) M0 and partial N0 for the partial blocks at the end of a row/column if any. This is to avoid padding.
    const unsigned int partial_store_m0 = internal_m % lhs_info.m0;
    const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0;

    // Shrink M0 to be always <= M (internal_m) to prevent out-of-bounds reads.
    // NOTE: This might have implications on heuristics and performance
    const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0);
    _m                             = internal_m;
    _n                             = gemm_info.n;
    _k                             = gemm_info.k;

    // Create build options
    CLBuildOptions build_opts;
    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src0->data_type()));
    build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
    build_opts.add_option_if(src2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
    build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA");
    build_opts.add_option_if(gemm_info.broadcast_bias, "-DBROADCAST_BIAS");
    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(h_gemm_3d));
    build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(d_gemm_3d));
    build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(src1->dimension(2)));
    build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
    build_opts.add_option("-DM0=" + support::cpp11::to_string(internal_m0));
    build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
    build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0));
    build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
    build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
    build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation())));
    build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a()));
    build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b()));

    std::string kernel_name("gemm_mm_native");

    // A macro guard to compile ONLY the kernel of interest
    build_opts.add_option("-D" + upper_string(kernel_name));

    // Create kernel
    _kernel = create_kernel(compile_context, kernel_name, build_opts.options());

    // Set config_id for enabling LWS tuning
    _config_id = kernel_name;
    _config_id += "_";
    _config_id += (_add_bias ? "add_bias_" : "");
    _config_id += (gemm_info.broadcast_bias ? "broadcast_bias_" : "");
    _config_id += (_reinterpret_input_as_3d ? "3di_" : "");
    _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
    _config_id += (gemm_info.activation_info.enabled() ? "fused_activation_" : "");
    _config_id += lower_string(string_from_data_type(src0->data_type()));
    _config_id += "_";
    _config_id += support::cpp11::to_string(dst->dimension(1));
    _config_id += "_";
    _config_id += support::cpp11::to_string(dst->dimension(0));
    _config_id += "_";
    _config_id += support::cpp11::to_string(gemm_info.k);
    _config_id += "_";
    _config_id += support::cpp11::to_string(dst->dimension(2));
    _config_id += "_";
    _config_id += support::cpp11::to_string(lhs_info.m0);
    _config_id += "_";
    _config_id += support::cpp11::to_string(rhs_info.n0);
    _config_id += "_";
    _config_id += support::cpp11::to_string(rhs_info.k0);

    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
}

Status ClGemmMatrixMultiplyNativeKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta,
                                                  const GEMMLHSMatrixInfo &lhs_info,
                                                  const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
{
    ElementsProcessed num_elements_processed{};
    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info));
    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src0->clone().get(),
                                                              src1->clone().get(),
                                                              src2 != nullptr ? src2->clone().get() : nullptr,
                                                              dst->clone().get(),
                                                              lhs_info,
                                                              rhs_info,
                                                              gemm_info,
                                                              num_elements_processed)
                                .first);

    return Status{};
}

void ClGemmMatrixMultiplyNativeKernel::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);

    const auto src0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
    const auto src1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
    const auto src2 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2));
    auto       dst  = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));

    ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
    ARM_COMPUTE_ERROR_ON(_add_bias && src2 == nullptr);

    if(src1->info()->num_dimensions() < 3)
    {
        // The stride_z for matrix B must be zero if we do not slice
        ARM_COMPUTE_ERROR_ON(src1->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
        unsigned int idx0;
        if(_add_bias)
        {
            idx0 = 4 * num_arguments_per_2D_tensor() + 7;
        }
        else
        {
            idx0 = 3 * num_arguments_per_2D_tensor() + 6;
        }
        const unsigned int total_cross_plane_pad = src0->info()->padding().top + src0->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 dst has to be reinterpreted as 3D tensor
        unsigned int idx0;
        if(_add_bias)
        {
            idx0 = 4 * num_arguments_per_2D_tensor() + 7 + (_reinterpret_input_as_3d ? 1 : 0);
        }
        else
        {
            idx0 = 3 * num_arguments_per_2D_tensor() + 6 + (_reinterpret_input_as_3d ? 1 : 0);
        }
        const unsigned int total_cross_plane_pad = dst->info()->padding().top + dst->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, src0, slice);
        add_2D_tensor_argument(idx, src1, slice_b);
        if(_add_bias)
        {
            add_2D_tensor_argument(idx, src2, slice);
        }
        add_2D_tensor_argument(idx, dst, slice);

        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src0->info()->strides_in_bytes()[2]));
        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src1->info()->strides_in_bytes()[2]));
        if(_add_bias)
        {
            _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src2->info()->strides_in_bytes()[2]));
        }
        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[2]));

        // Pass m, n and k at runtime
        _kernel.setArg<cl_int>(idx++, _m);
        _kernel.setArg<cl_int>(idx++, _n);
        _kernel.setArg<cl_int>(idx++, _k);

        enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
    }
    while(window.slide_window_slice_3D(slice));
}
} // namespace kernels
} // namespace opencl
} // namespace arm_compute