src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp

/*
 * Copyright (c) 2018-2023 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/cpu/operators/internal/CpuGemmAssemblyDispatch.h"

#include "arm_compute/runtime/NEON/NEScheduler.h"
#include "src/core/CPP/Validate.h"
#include "src/core/NEON/kernels/arm_gemm/utils.hpp"
#include "src/core/helpers/MemoryHelpers.h"
#include "src/core/utils/AssemblyUtils.h"
#include "src/cpu/kernels/assembly/CpuGemmAssemblyWrapperKernel.h"
#include "src/cpu/kernels/assembly/arm_gemm.hpp"
#include "src/cpu/utils/CpuAuxTensorHandler.h"

#include <arm_neon.h>

namespace arm_compute
{
namespace cpu
{
namespace
{
/** Run pretranspose_B_array in parallel (1D static scheduling)
 *
 * @tparam TypeInput
 * @tparam TypeOutput
 *
 * @param[in] gemm_asm         GemmCommon kernel to run
 * @param[in] dst              Pretransposed B array
 * @param[in] src              B array to be pretransposed
 * @param[in] src_ld           Stride in y
 * @param[in] src_multi_stride Stride in z ("multi")
 * @param[in] num_threads      Number of threads to run this method. Must be >= 1
 */
template <typename TypeInput, typename TypeOutput>
void run_parallel_pretranspose_B_array(arm_gemm::GemmCommon<TypeInput, TypeOutput> *gemm_asm, ITensor *dst, const TypeInput *src, int src_ld, int src_multi_stride, unsigned int num_threads)
{
    ARM_COMPUTE_ERROR_ON(gemm_asm == nullptr);
    ARM_COMPUTE_ERROR_ON(num_threads == 0);
    // The window size is also the total workload size
    const unsigned int wsize = gemm_asm->get_B_pretranspose_window_size();

    std::vector<IScheduler::Workload> workloads(num_threads);
    for(unsigned int t = 0; t < num_threads; ++t)
    {
        workloads[t] = [ = ](const ThreadInfo & info)
        {
            const unsigned int start = (info.thread_id * wsize) / num_threads;
            const unsigned int end   = ((info.thread_id + 1) * wsize) / num_threads;

            if(start < end)
            {
                gemm_asm->pretranspose_B_array_part(dst->buffer(), src, src_ld, src_multi_stride, start, end);
            }
        };
    }
    NEScheduler::get().run_tagged_workloads(workloads, "CpuGemmAssemblyDispatch/pretranspose_B_array");
}
} // namespace

using namespace arm_compute::experimental;

namespace
{
struct free_delete
{
    void operator()(void *x)
    {
        free(x);
    }
};

struct Params
{
    unsigned int M;
    unsigned int N;
    unsigned int K;
    unsigned int batches;
    unsigned int multis;
    unsigned int sections;
    bool         indirect;
};

Params extract_parameters(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *d, const AsmGemmInfo &info)
{
    ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, d);
    Params p;
    p.M        = d->tensor_shape().y();
    p.K        = a->tensor_shape().x();
    p.N        = d->tensor_shape().x();
    p.batches  = 1;
    p.multis   = 1;
    p.sections = 1;
    p.indirect = false;

    if(info.method == AsmConvMethod::Conv || info.method == AsmConvMethod::Indirect)
    {
        p.indirect = true;
        p.sections = b->tensor_shape()[2] * b->tensor_shape()[3];
    }
    else
    {
        p.multis  = b->tensor_shape().z();
        p.batches = d->tensor_shape().total_size_upper(2) / p.multis;
    }

    // Update M in case of GEMM3D for output
    if(info.depth_output_gemm3d != 0)
    {
        p.M       = d->tensor_shape().y() * d->tensor_shape().z();
        p.batches = d->tensor_shape().total_size_upper(3) / p.multis;
    }

    return p;
}

IScheduler::Hints scheduling_hint_heuristic(arm_gemm::GemmMethod method, DataType data_type)
{
    // Schedule assembly kernel
    const int         granule_threshold = 200;
    IScheduler::Hints scheduling_hint   = IScheduler::Hints(Window::DimX);
    if(method == arm_gemm::GemmMethod::GEMM_INTERLEAVED && data_type == DataType::F32)
    {
        scheduling_hint = IScheduler::Hints(Window::DimX, IScheduler::StrategyHint::DYNAMIC, granule_threshold);
    }
    else if(method == arm_gemm::GemmMethod::GEMM_INTERLEAVED_2D && (data_type == DataType::F32 || data_type == DataType::F16 || data_type == DataType::U8 || data_type == DataType::S8))
    {
        //GEMM_INTERLEAVED supports 2D parallelism, IScheduler::split_dimensions_all signals to parallelise over all window dimensions
        scheduling_hint = IScheduler::Hints(IScheduler::split_dimensions_all, IScheduler::StrategyHint::STATIC, granule_threshold);
    }
    else if(method == arm_gemm::GemmMethod::QUANTIZE_WRAPPER_2D && (data_type == DataType::QASYMM8 || data_type == DataType::QASYMM8_SIGNED))
    {
        //special case for QASYMM8 to support 2D parallelism, scheduler here may be tweaked differently compared to FP32 case
        scheduling_hint = IScheduler::Hints(IScheduler::split_dimensions_all, IScheduler::StrategyHint::STATIC, granule_threshold);
    }

    return scheduling_hint;
}

/** Fallback in case ACL doesn't have a function */
template <typename TypeInput, typename TypeOutput, class OutputStage = arm_gemm::Nothing>
class Fallback : public CpuGemmAssemblyDispatch::IFallback
{
public:
    /** Destructor */
    ~Fallback() = default;

    /** Initialise the functions's input and output.
     *
     * @param[in]  a         Input tensor containing the Matrix A.
     * @param[in]  b         Input tensor containing the Matrix B.
     * @param[in]  c         Input tensor containing the Matrix C.
     * @param[out] d         Output tensor to store the result of matrix multiplication.
     * @param[in]  args      Matrix multiplication information.
     * @param[in]  gemm_info GEMM meta-data
     * @param[in]  os        Output stage meta-data.
     */
    void configure(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, ITensorInfo *d,
                   arm_gemm::GemmArgs args, const AsmGemmInfo &gemm_info,
                   const OutputStage &os = {});

    /** Set requantization shifts to be used
     *
     * @param[in] shifts Requantization shifts
     *
     * @return Pointer to the shift data
     */
    /** Set requantization data to be used
      *
      *
      * @param shifts       Requantization shifts
      * @param multipliers  Requantization multipliers
      *
      * @return A tuple with the pointers to the shift and multiplier data respectively
      */
    std::tuple<bool, const int32_t *, const int32_t *, const int32_t *> set_requantize_data(const std::vector<int32_t> &shifts,
                                                                                            const std::vector<int32_t> &multipliers);

    // Inherited methods overridden:
    void run(ITensorPack &tensors) override;
    void prepare(ITensorPack &tensors) override;
    bool                             is_configured() const override;
    experimental::MemoryRequirements workspace() const override;
    bool                             isVarWeightsKernel() const override
    {
        if(!_gemm_kernel_asm)
            return false;
        const arm_compute::WeightFormat wf = assembly_utils::map_to_arm_compute_weight_format(_gemm_kernel_asm->get_config().weight_format);
        return wf != arm_compute::WeightFormat::UNSPECIFIED && wf != arm_compute::WeightFormat::ANY;
    }

private:
    enum AuxTensorIdx
    {
        AsmGemmWorkspace = 0,
        Pretranspose,
        Count
    };

    /** Configure the indirect buffer
     *
     * @param[in]  a    Input tensor containing the Matrix A.
     * @param[in]  b    Input tensor containing the Matrix B.
     * @param[out] d    Output tensor to store the result of matrix multiplication.
     * @param[in]  info GEMM meta-data
     */
    void configure_indirect(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *d, const AsmGemmInfo &info);
    /** Prepare the indirect buffer */
    void prepare_indirect_buffer(ITensorPack &tensors);

    /** Assembly Gemm kernel */
    std::shared_ptr<arm_gemm::GemmCommon<TypeInput, TypeOutput>> _gemm_kernel_asm{ nullptr };
    /** Optimised Arm® Neon™ kernel */
    std::unique_ptr<INEKernel> _optimised_kernel{ nullptr };
    /** Assembly GEMM workspace tensor info */
    TensorInfo _workspace_info{};
    /** Pre-transpose tensor info */
    TensorInfo _pretranspose_info{};
    /** Prepared flag */
    bool _is_prepared{ false };
    /** GEMM meta-data */
    AsmGemmInfo _gemm_info{};
    /** GEMM kernel description */
    arm_gemm::KernelDescription _kernel_info{};
    /** Per channel quantization shifts */
    std::vector<int32_t> _shifts{};
    std::vector<int32_t> right_shifts{};
    std::vector<int32_t> left_shifts{};
    /** Per channel quantization multipliers */
    std::vector<int32_t> _multipliers{};
    /** Indirect buffer */
    std::unique_ptr<const TypeInput *const *, free_delete> _indirect_arg{};
    std::unique_ptr<const TypeInput *, free_delete>        _indirect_buf{};
    std::vector<TypeInput>           _indirect_pad{};
    arm_gemm::ConvolutionParameters  _cp{};
    experimental::MemoryRequirements _aux_mem{ Count };
    bool                             _B_pretranspose_required{ false };
    bool                             _is_b_constant{ true };
    bool                             _is_c_constant{ true };
};

template <typename TypeInput, typename TypeOutput, class OutputStage>
std::tuple<bool, const int32_t *, const int32_t *, const int32_t *>
Fallback<TypeInput, TypeOutput, OutputStage>::set_requantize_data(const std::vector<int32_t> &shifts, const std::vector<int32_t> &multipliers)
{
    _multipliers   = multipliers;
    _shifts        = shifts;
    bool need_left = false;
    for(const auto s : _shifts)
    {
        left_shifts.push_back(std::max(-s, int32_t(0)));
        right_shifts.push_back(std::min(-s, int32_t(0)));
        if(s < 0 && !need_left)
        {
            need_left = true;
        }
    }
    return std::make_tuple(need_left, left_shifts.data(), right_shifts.data(), _multipliers.data());
}

template <typename TypeInput, typename TypeOutput, class OutputStage>
void Fallback<TypeInput, TypeOutput, OutputStage>::prepare_indirect_buffer(ITensorPack &tensors)
{
    auto             a              = tensors.get_const_tensor(TensorType::ACL_SRC_0);
    const TypeInput *A_ptr          = reinterpret_cast<TypeInput *>(a->buffer());
    const int        multis         = 1;
    const int        batches        = a->info()->tensor_shape().total_size_upper(3);
    const size_t     stride_A       = a->info()->strides_in_bytes().y() / sizeof(TypeInput);
    const size_t     batch_stride_A = a->info()->strides_in_bytes()[3] / sizeof(TypeInput);
    const size_t     multi_stride_A = a->info()->strides_in_bytes()[4] / sizeof(TypeInput);

    const size_t output_hw    = _cp.output_height * _cp.output_width;
    const int    batch_size   = _cp.kernel_height * _cp.kernel_width * output_hw * sizeof(TypeInput);
    const size_t batch_stride = batch_size / sizeof(TypeInput);
    const int    multi_size   = batch_size * batches;
    const size_t multi_stride = multi_size / sizeof(TypeInput);

    for(int64_t m = 0; m < multis; m++)
    {
        for(int64_t b = 0; b < batches; b++)
        {
            for(int64_t output_y = 0; output_y < _cp.output_height; output_y++)
            {
                for(int64_t output_x = 0; output_x < _cp.output_width; output_x++)
                {
                    int64_t output_xy = (output_y * _cp.output_width) + output_x;

                    for(int64_t kernel_y = 0; kernel_y < _cp.kernel_height; kernel_y++)
                    {
                        for(int64_t kernel_x = 0; kernel_x < _cp.kernel_width; kernel_x++)
                        {
                            int64_t input_x   = (output_x * _cp.output_stride_w) + kernel_x - _cp.padding_left;
                            int64_t input_y   = (output_y * _cp.output_stride_h) + kernel_y - _cp.padding_top;
                            int64_t kernel_xy = (kernel_y * _cp.kernel_width) + kernel_x;
                            int64_t input_xy  = (input_y * _cp.input_width) + input_x;

                            if(input_x < 0 || input_x >= _cp.input_width || input_y < 0 || input_y >= _cp.input_height)
                            {
                                _indirect_buf.get()[m * multi_stride + b * batch_stride + kernel_xy * output_hw + output_xy] = _indirect_pad.data();
                            }
                            else
                            {
                                _indirect_buf.get()[m * multi_stride + b * batch_stride + kernel_xy * output_hw + output_xy] =
                                    A_ptr + (m * multi_stride_A + b * batch_stride_A + input_xy * stride_A);
                            }
                        }
                    }
                }
            }
        }
    }
}

template <typename TypeInput, typename TypeOutput, class OutputStage>
void Fallback<TypeInput, TypeOutput, OutputStage>::configure_indirect(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *d, const AsmGemmInfo &info)
{
    ARM_COMPUTE_ERROR_ON(!(info.method == AsmConvMethod::Conv || info.method == AsmConvMethod::Indirect));

    float zeropad = 0.f;
    if(is_data_type_quantized(a->data_type()))
    {
        zeropad = a->quantization_info().uniform().offset;
    }

    const int64_t input_width    = static_cast<int64_t>(a->tensor_shape()[1]);
    const int64_t input_height   = static_cast<int64_t>(a->tensor_shape()[2]);
    const int64_t input_channels = static_cast<int64_t>(a->tensor_shape()[0]);
    const int64_t kernel_width   = static_cast<int64_t>(b->tensor_shape()[2]);
    const int64_t kernel_height  = static_cast<int64_t>(b->tensor_shape()[3]);
    const int64_t output_width   = static_cast<int64_t>(d->tensor_shape()[1]);
    const int64_t output_height  = static_cast<int64_t>(d->tensor_shape()[2]);

    _cp = { input_width, input_height, input_channels, kernel_width, kernel_height, output_width, output_height,
            info.ps_info.stride().first, info.ps_info.stride().second, info.padding_top, info.padding_left, zeropad
          };

    if(info.method == AsmConvMethod::Conv)
    {
        _gemm_kernel_asm->set_convolution_parameters(_cp);
    }

    if(info.method == AsmConvMethod::Indirect)
    {
        const unsigned int multis    = 1;
        const unsigned int batches   = a->tensor_shape().total_size_upper(3);
        const unsigned int kernel_hw = _cp.kernel_width * _cp.kernel_height;
        const unsigned int output_hw = _cp.output_width * _cp.output_height;

        using TypeInputPtr        = TypeInput *;
        const int    batch_size   = kernel_hw * output_hw * sizeof(TypeInputPtr);
        const size_t batch_stride = batch_size / sizeof(TypeInputPtr);
        const int    multi_size   = batch_size * batches;
        const size_t multi_stride = multi_size / sizeof(TypeInputPtr);

        _indirect_buf = std::unique_ptr<const TypeInput *, free_delete>(reinterpret_cast<const TypeInput **>(malloc(multi_size * multis)));
        _indirect_arg = std::unique_ptr<const TypeInput *const *, free_delete>(reinterpret_cast<const TypeInput *const **>(malloc(sizeof(TypeInput **) * kernel_hw * multis * batches)));
        _indirect_pad = std::vector<TypeInput>(_cp.input_channels, TypeInput(zeropad));

        // Set indirect argument
        int64_t pos = 0;
        for(int64_t m = 0; m < multis; m++)
        {
            for(int64_t b = 0; b < batches; b++)
            {
                for(int64_t kernel_xy = 0; kernel_xy < kernel_hw; kernel_xy++)
                {
                    (_indirect_arg.get())[pos++] = _indirect_buf.get() + m * multi_stride + b * batch_stride + kernel_xy * output_hw;
                }
            }
        }

        _gemm_kernel_asm->set_indirect_parameters(a->tensor_shape()[0], _indirect_arg.get());
    }
}

template <typename TypeInput, typename TypeOutput, class OutputStage>
void Fallback<TypeInput, TypeOutput, OutputStage>::configure(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, ITensorInfo *d,
                                                             arm_gemm::GemmArgs args, const AsmGemmInfo &gemm_info,
                                                             const OutputStage &os)
{
    ARM_COMPUTE_UNUSED(c);

    _is_b_constant = b->are_values_constant();
    _is_c_constant = c ? c->are_values_constant() : true;

    _gemm_kernel_asm = arm_gemm::gemm<TypeInput, TypeOutput, OutputStage>(args, os);
    if(_gemm_kernel_asm == nullptr)
    {
        //configuration not supported: Leave function unconfigured:
        return;
    }

    arm_gemm::GemmConfig gemm_cfg = _gemm_kernel_asm->get_config();

    // arm_compute wrapper for the Gemm object (see above)
    auto acl_gemm_wrapper = std::make_unique<kernel::CpuGemmAssemblyWrapperKernel<TypeInput, TypeOutput>>();
    ARM_COMPUTE_ERROR_ON(acl_gemm_wrapper == nullptr);
    acl_gemm_wrapper->configure(_gemm_kernel_asm.get(), gemm_cfg.filter);
    const size_t       workspace_size = _gemm_kernel_asm->get_working_size();
    const unsigned int alignment      = 4096;
    _workspace_info                   = TensorInfo(TensorShape(workspace_size), 1, DataType::U8);
    _aux_mem[AsmGemmWorkspace]        = MemoryInfo(offset_int_vec(AsmGemmWorkspace), MemoryLifetime::Temporary, workspace_size, alignment);

    //if we disable this code below in brackets then ConvLayer deadlocks when threads > 1 and
    //the shapes are In=1x1x1024 Weights=1x1x1024x1001 Biases=1001 Out=1x1x1001
    {
        const unsigned int window_size = _gemm_kernel_asm->get_window_size().total_size();
        if(window_size < static_cast<unsigned int>(args._maxthreads))
        {
            _gemm_kernel_asm->set_nthreads(window_size);
        }
    }

    _optimised_kernel = std::move(acl_gemm_wrapper);
    _gemm_info        = gemm_info;
    // Check for pre-transposed support
    if(_gemm_kernel_asm->B_pretranspose_required())
    {
        // Forcing 128-byte alignment (required by 32-bit kernels)
        const unsigned int alignment           = 128;
        const size_t       B_pretranspose_size = _gemm_kernel_asm->get_B_pretransposed_array_size();
        _pretranspose_info                     = TensorInfo(TensorShape(B_pretranspose_size), 1, DataType::U8);
        _aux_mem[Pretranspose]                 = MemoryInfo(offset_int_vec(Pretranspose), MemoryLifetime::Persistent, B_pretranspose_size, alignment);
        _B_pretranspose_required               = true;
    }

    // Handle indirect GEMM convolution
    if(gemm_info.method == AsmConvMethod::Conv || gemm_info.method == AsmConvMethod::Indirect)
    {
        configure_indirect(a, b, d, gemm_info);
    }
}

template <typename TypeInput, typename TypeOutput, class OutputStage>
void Fallback<TypeInput, TypeOutput, OutputStage>::prepare(ITensorPack &tensors)
{
    if(!_is_prepared)
    {
        auto b = tensors.get_const_tensor(TensorType::ACL_SRC_1);
        auto c = tensors.get_const_tensor(TensorType::ACL_SRC_2);

        // Setup up matrix bias in the assembly kernel, it's just a pointer to matrix C.
        if(c && c->info()->data_type() == DataType::S32)
        {
            _gemm_kernel_asm->set_quantized_bias(reinterpret_cast<const int32_t *>(c->buffer() + c->info()->offset_first_element_in_bytes()), 0);
        }

        // Pretranspose B if required
        if(_gemm_kernel_asm->B_pretranspose_required())
        {
            // Fixed format kernels need no pretranspose.
            ARM_COMPUTE_ERROR_ON(arm_compute::is_fixed_format(assembly_utils::map_to_arm_compute_weight_format(_gemm_kernel_asm->get_config().weight_format)));
            const int  ldb            = b->info()->strides_in_bytes().y() / b->info()->element_size();
            const auto in1_ptr        = reinterpret_cast<const TypeInput *>(b->buffer() + b->info()->offset_first_element_in_bytes());
            const int  multi_stride_b = b->info()->strides_in_bytes().z() / b->info()->element_size();

            CpuAuxTensorHandler pretranspose(offset_int_vec(Pretranspose), _pretranspose_info, tensors, false);
            ARM_COMPUTE_ERROR_ON(pretranspose.get()->buffer() == nullptr);
            run_parallel_pretranspose_B_array<TypeInput, TypeOutput>(_gemm_kernel_asm.get(), pretranspose.get(), in1_ptr, ldb, multi_stride_b, NEScheduler::get().num_threads());

            b->mark_as_unused();
        }

        if(_gemm_info.method == AsmConvMethod::Indirect)
        {
            prepare_indirect_buffer(tensors);
        }

        _is_prepared = true;
    }
}

template <typename TypeInput, typename TypeOutput, class OutputStage>
bool Fallback<TypeInput, TypeOutput, OutputStage>::is_configured() const
{
    return _optimised_kernel != nullptr;
}

template <typename TypeInput, typename TypeOutput, class OutputStage>
experimental::MemoryRequirements Fallback<TypeInput, TypeOutput, OutputStage>::workspace() const
{
    return _aux_mem;
}

template <typename TypeInput, typename TypeOutput, class OutputStage>
void Fallback<TypeInput, TypeOutput, OutputStage>::run(ITensorPack &tensors)
{
    auto a = tensors.get_const_tensor(TensorType::ACL_SRC_0);
    auto b = tensors.get_const_tensor(TensorType::ACL_SRC_1);
    auto c = tensors.get_const_tensor(TensorType::ACL_SRC_2);
    auto d = tensors.get_tensor(TensorType::ACL_DST);

    int       lda = a->info()->strides_in_bytes().y() / a->info()->element_size();
    int       ldb = 0;
    const int ldd = d->info()->strides_in_bytes().y() / d->info()->element_size();

    const size_t a_batch_idx = _gemm_info.reinterpret_input_as_3d != 0 ? 3 : 2;
    const size_t a_multi_idx = a_batch_idx + 1;
    const size_t d_batch_idx = _gemm_info.depth_output_gemm3d != 0 ? 3 : 2;
    const size_t d_multi_idx = d_batch_idx + 1;

    int       batch_stride_a = a->info()->strides_in_bytes()[a_batch_idx] / a->info()->element_size();
    const int batch_stride_d = d->info()->strides_in_bytes()[d_batch_idx] / d->info()->element_size();

    int       multi_stride_a = a->info()->strides_in_bytes()[a_multi_idx] / a->info()->element_size();
    int       multi_stride_b = 0;
    const int multi_stride_d = d->info()->strides_in_bytes()[d_multi_idx] / d->info()->element_size();

    auto             in0_ptr = reinterpret_cast<const TypeInput *>(a->buffer() + a->info()->offset_first_element_in_bytes());
    const TypeInput *in1_ptr = nullptr;
    auto             out_ptr = reinterpret_cast<TypeOutput *>(d->buffer() + d->info()->offset_first_element_in_bytes());

    // Check if B is pre-tranposed and de-reference if not
    if(!_gemm_kernel_asm->B_is_pretransposed())
    {
        ldb            = b->info()->strides_in_bytes().y() / b->info()->element_size();
        multi_stride_b = b->info()->strides_in_bytes().z() / b->info()->element_size();
        in1_ptr        = reinterpret_cast<const TypeInput *>(b->buffer() + b->info()->offset_first_element_in_bytes());
    }

    // If necessary, run pretranspose every time if either weights or biases are non-constant
    if((b && !_is_b_constant) || (c && !_is_c_constant && c->info()->data_type() == DataType::S32))
    {
        if(c && c->info()->data_type() == DataType::S32)
        {
            _gemm_kernel_asm->set_quantized_bias(reinterpret_cast<const int32_t *>(c->buffer() + c->info()->offset_first_element_in_bytes()), 0);
        }

        // Pretranspose B if required
        if(_B_pretranspose_required)
        {
            const int  ldb            = b->info()->strides_in_bytes().y() / b->info()->element_size();
            const auto b_ptr          = reinterpret_cast<const TypeInput *>(b->buffer() + b->info()->offset_first_element_in_bytes());
            const int  multi_stride_b = b->info()->strides_in_bytes().z() / b->info()->element_size();

            CpuAuxTensorHandler pretranspose(offset_int_vec(Pretranspose), _pretranspose_info, tensors, true);
            ARM_COMPUTE_ERROR_ON(pretranspose.get()->buffer() == nullptr);

            if(_is_b_constant)
            {
                _gemm_kernel_asm->requantize_bias(pretranspose.get()->buffer(), b_ptr, ldb, multi_stride_b);
            }
            else
            {
                run_parallel_pretranspose_B_array<TypeInput, TypeOutput>(_gemm_kernel_asm.get(), pretranspose.get(), b_ptr, ldb, multi_stride_b, NEScheduler::get().num_threads());
            }
        }
    }

    const auto scheduling_hint = scheduling_hint_heuristic(_kernel_info.method, d->info()->data_type());

    // Set workspace if needed and reset number of threads as buffer manager gets re-created with max_threads
    CpuAuxTensorHandler workspace(offset_int_vec(AsmGemmWorkspace), _workspace_info, tensors, false);
    if(workspace.get()->buffer() != nullptr)
    {
        _gemm_kernel_asm->set_working_space(reinterpret_cast<void *>(workspace.get()->buffer()));
        const unsigned int split_dim   = scheduling_hint.split_dimension();
        const unsigned int window_size = _gemm_kernel_asm->get_window_size().total_size();
        unsigned int       num_threads = NEScheduler::get().num_threads();
        if(window_size < num_threads)
        {
            num_threads = window_size;
        }
        if(split_dim != IScheduler::split_dimensions_all)
        {
            // Make sure the kernel does not expect more threads than we can actually spawn
            const unsigned int num_iterations = _optimised_kernel.get()->window().num_iterations(split_dim);
            num_threads                       = std::min(num_iterations, num_threads);
        }
        _gemm_kernel_asm->set_nthreads(num_threads);
    }

    // Prepare assembly kernel
    prepare(tensors);

    // Setup up matrix bias in the assembly kernel, it's just a pointer to matrix C.
    TypeOutput *bias = nullptr;
    if(c && c->info()->data_type() != DataType::S32)
    {
        bias = reinterpret_cast<TypeOutput *>(c->buffer() + c->info()->offset_first_element_in_bytes());
    }

    if(_gemm_info.method == AsmConvMethod::Indirect)
    {
        in0_ptr        = nullptr;
        lda            = 0;
        batch_stride_a = 0;
        multi_stride_a = 0;
    }

    // Set gemm parameters
    _gemm_kernel_asm->set_arrays(in0_ptr, lda, batch_stride_a, multi_stride_a,
                                 in1_ptr, ldb, multi_stride_b,
                                 out_ptr, ldd, batch_stride_d, multi_stride_d,
                                 bias, 0);
    // Schedule
    NEScheduler::get().schedule(_optimised_kernel.get(), scheduling_hint);
}

template <typename TypeInput, typename TypeOutput>
void create_arm_gemm(std::unique_ptr<CpuGemmAssemblyDispatch::IFallback> &arm_gemm,
                     const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, ITensorInfo *d,
                     arm_gemm::Activation activation, const AsmGemmInfo &info)
{
    Params         p           = extract_parameters(a, b, d, info);
    const CPUInfo &ci          = NEScheduler::get().cpu_info();
    unsigned int   num_threads = NEScheduler::get().num_threads();

    arm_gemm::GemmConfig cfg;
    cfg.weight_format = assembly_utils::map_to_arm_gemm_weight_format(info.weight_format);
    arm_gemm::GemmArgs args(&ci, p.M, p.N, p.K, p.sections, p.batches, p.multis, p.indirect, activation, num_threads, info.fixed_format, info.fast_mode, &cfg);

    // Create arm_gemm fallback
    auto fallback = std::make_unique<Fallback<TypeInput, TypeOutput>>();
    fallback->configure(a, b, c, d, args, info);
    arm_gemm = std::move(fallback);
}

template <typename TypeInput, typename TypeOutput>
void create_arm_gemm_quant(std::unique_ptr<CpuGemmAssemblyDispatch::IFallback> &arm_gemm,
                           const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, ITensorInfo *d,
                           arm_gemm::Activation activation, const AsmGemmInfo &info)
{
    ARM_COMPUTE_UNUSED(activation);
    Params             p           = extract_parameters(a, b, d, info);
    const CPUInfo     &ci          = NEScheduler::get().cpu_info();
    const unsigned int num_threads = NEScheduler::get().num_threads();

    arm_gemm::GemmConfig cfg;
    cfg.weight_format = assembly_utils::map_to_arm_gemm_weight_format(info.weight_format);
    arm_gemm::GemmArgs args(&ci, p.M, p.N, p.K, p.sections, p.batches, p.multis, p.indirect, activation, num_threads, info.fixed_format, info.fast_mode, &cfg);

    // Create arm_gemm fallback
    auto fallback = std::make_unique<Fallback<TypeInput, TypeOutput, arm_gemm::Requantize32>>();

    // Configure requantization info
    const int32_t                 negation = info.negated_offsets ? 1 : -1;
    const int32_t                 a_offset = -a->quantization_info().uniform().offset * negation;
    const int32_t                 b_offset = -b->quantization_info().uniform().offset * negation;
    const GEMMLowpOutputStageInfo os_info  = info.output_stage;

    arm_gemm::Requantize32 gemm_requant_info{};
    if(os_info.gemmlowp_shifts.size() > 1)
    {
        const auto requantize_data = fallback->set_requantize_data(os_info.gemmlowp_shifts, os_info.gemmlowp_multipliers);
        gemm_requant_info          = arm_gemm::Requantize32(nullptr, 0,
                                                            a_offset, b_offset, os_info.gemmlowp_offset,
                                                            (std::get<0>(requantize_data)) ? std::get<1>(requantize_data) : nullptr,
                                                            std::get<2>(requantize_data),
                                                            std::get<3>(requantize_data),
                                                            os_info.gemmlowp_min_bound, os_info.gemmlowp_max_bound);
    }
    else
    {
        gemm_requant_info = arm_gemm::Requantize32(nullptr, 0,
                                                   a_offset, b_offset, os_info.gemmlowp_offset,
                                                   -os_info.gemmlowp_shift, os_info.gemmlowp_multiplier,
                                                   os_info.gemmlowp_min_bound, os_info.gemmlowp_max_bound);
    }

    // Configure fallback
    fallback->configure(a, b, c, d, args, info, gemm_requant_info);
    arm_gemm = std::move(fallback);
}
} //namespace

CpuGemmAssemblyDispatch::CpuGemmAssemblyDispatch()
    : _arm_gemm(nullptr)
{
}

Status CpuGemmAssemblyDispatch::has_opt_impl(arm_compute::WeightFormat &expected_weight_format, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *d,
                                             const AsmGemmInfo &info)
{
    ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, d);
    ARM_COMPUTE_UNUSED(c);
    arm_gemm::Activation act         = assembly_utils::map_to_arm_gemm_activation(info.activation_info);
    Params               p           = extract_parameters(a, b, d, info);
    const CPUInfo       &ci          = NEScheduler::get().cpu_info();
    unsigned int         num_threads = NEScheduler::get().num_threads();
    arm_gemm::GemmConfig cfg;
    cfg.weight_format                           = assembly_utils::map_to_arm_gemm_weight_format(info.weight_format);
    arm_gemm::WeightFormat arm_gemm_expected_wf = assembly_utils::map_to_arm_gemm_weight_format(expected_weight_format);
    arm_gemm::GemmArgs     args(&ci, p.M, p.N, p.K, p.sections, p.batches, p.multis, p.indirect, act, num_threads, info.fixed_format, info.fast_mode, &cfg);
    switch(a->data_type())
    {
        case DataType::F32:
            ARM_COMPUTE_RETURN_ERROR_ON_MSG(!(arm_gemm::has_opt_gemm<float, float, arm_gemm::Nothing>(arm_gemm_expected_wf, args, {})),
                                            "We could not find an optimized kernel for F32 input");
            break;
#ifdef __aarch64__
        case DataType::U8:
        case DataType::QASYMM8:
            if(d->data_type() == DataType::S32)
            {
                ARM_COMPUTE_RETURN_ERROR_ON_MSG(!(arm_gemm::has_opt_gemm<uint8_t, uint32_t, arm_gemm::Nothing>(arm_gemm_expected_wf, args, {})),
                                                "We could not find an optimized kernel for U8/QASYMM8 input and U32 output");
            }
            else
            {
                ARM_COMPUTE_RETURN_ERROR_ON_MSG(!(arm_gemm::has_opt_gemm<uint8_t, uint8_t, arm_gemm::Requantize32>(arm_gemm_expected_wf, args, {})),
                                                "We could not find an optimized kernel for U8 input and U8 output");
            }
            break;
        case DataType::S8:
        case DataType::QASYMM8_SIGNED:
            if(d->data_type() == DataType::S32)
            {
                ARM_COMPUTE_RETURN_ERROR_ON_MSG(!(arm_gemm::has_opt_gemm<int8_t, int32_t, arm_gemm::Nothing>(arm_gemm_expected_wf, args, {})),
                                                "We could not find an optimized kernel for S8/QASYMM8_SIGNED input and S32 output");
            }
            else
            {
                ARM_COMPUTE_RETURN_ERROR_ON_MSG(!(arm_gemm::has_opt_gemm<int8_t, int8_t, arm_gemm::Requantize32>(arm_gemm_expected_wf, args, {})),
                                                "We could not find an optimized kernel for S8 input and S8 output");
            }
            break;
#endif /* __aarch64__ */
#if defined(ARM_COMPUTE_ENABLE_BF16)
        case DataType::BFLOAT16:
        {
            ARM_COMPUTE_RETURN_ERROR_ON_MSG(!(arm_gemm::has_opt_gemm<bfloat16, float, arm_gemm::Nothing>(arm_gemm_expected_wf, args, {})),
                                            "We could not find an optimized kernel for BFLOAT16 input and F32 output");
            break;
        }
#endif /* defined(ARM_COMPUTE_ENABLE_BF16) */
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
        case DataType::F16:
            ARM_COMPUTE_RETURN_ERROR_ON_MSG(!(arm_gemm::has_opt_gemm<float16_t, float16_t, arm_gemm::Nothing>(arm_gemm_expected_wf, args, {})),
                                            "We could not find an optimized kernel for F16 input and F16 output");
            break;
#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
        default:
            ARM_COMPUTE_RETURN_ERROR_ON_MSG(true, "Usupported type. Could not find a kernel");
            break;
    }
    expected_weight_format = assembly_utils::map_to_arm_compute_weight_format(arm_gemm_expected_wf);

    return Status{};
}

Status CpuGemmAssemblyDispatch::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *d, const AsmGemmInfo &info)
{
    ARM_COMPUTE_UNUSED(c, info);
    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(a, b, d);
    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(a);
    ARM_COMPUTE_RETURN_ERROR_ON_CPU_BF16_UNSUPPORTED(a);
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!(info.reshape_b_only_on_first_run), "Assembly kernel will not be executed when reshape_b_only_on_first_run is false");

#ifndef __aarch64__
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->element_size() == 1, "8bit integer types only supported for aarch64");
#endif /* __aarch64__ */
    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S8,
                                                         DataType::BFLOAT16, DataType::F16, DataType::F32);
    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(b, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL, DataType::S8,
                                                         DataType::BFLOAT16, DataType::F16, DataType::F32);
    if(is_data_type_quantized_per_channel(b->data_type()))
    {
        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QASYMM8_SIGNED, DataType::S8);
    }
    else if(is_fixed_format_fast_math(info.weight_format))
    {
        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(a, DataType::F32);
        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(b, DataType::BFLOAT16);
    }
    else
    {
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, b);
    }
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->data_type() == DataType::F32 && d->data_type() != DataType::F32, "Only F32 output supported for F32 input");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->data_type() == DataType::F16 && d->data_type() != DataType::F16, "Only F16 output supported for F16 input");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->data_type() == DataType::BFLOAT16 && d->data_type() != DataType::F32, "Only F32 output supported for BFLOAT16 input");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->data_type() == DataType::U8 && d->data_type() != DataType::U32, "Only U32 output supported for U8 input");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->data_type() == DataType::S8 && d->data_type() != DataType::S32, "Only S32 output supported for S8 input");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->data_type() == DataType::QASYMM8 && (d->data_type() != DataType::QASYMM8 && d->data_type() != DataType::S32),
                                    "Only QASYMM8/S32 output supported for QASYMM8 input");
    arm_compute::WeightFormat expected_weight_format;
    const Status              ret = CpuGemmAssemblyDispatch::has_opt_impl(expected_weight_format, a, b, c, d, info);
    if((bool)ret && expected_weight_format != arm_compute::WeightFormat::ANY)
    {
        // Correctness check: if the format expected by the kernel is
        // not "any", make sure that the one found matches the format
        // intended by the caller.
        ARM_COMPUTE_RETURN_ERROR_ON_MSG((expected_weight_format != info.weight_format),
                                        "The format expected by the kernel does not correspond with the one requested by the user.");
    }
    return ret;
}

bool CpuGemmAssemblyDispatch::is_activation_supported(const ActivationLayerInfo &activation)
{
    arm_gemm::Activation act = assembly_utils::map_to_arm_gemm_activation(activation);
    return act.type != arm_gemm::Activation::Type::None;
}

void CpuGemmAssemblyDispatch::configure(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, ITensorInfo *d, const AsmGemmInfo &info)
{
    ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, d);
    arm_gemm::Activation act = assembly_utils::map_to_arm_gemm_activation(info.activation_info);

    //If we don't support a combination of data types, silently return: it is the caller's responsibility to check if configure() was successful via is_configured()
    if(!CpuGemmAssemblyDispatch::validate(a, b, c, d, info))
    {
        return;
    }

    switch(a->data_type())
    {
        case DataType::F32:
            create_arm_gemm<float, float>(_arm_gemm, a, b, c, d, act, info);
            break;
#ifdef __aarch64__
        case DataType::U8:
        case DataType::QASYMM8:
            if(d->data_type() == DataType::S32)
            {
                create_arm_gemm<uint8_t, uint32_t>(_arm_gemm, a, b, c, d, act, info);
            }
            else
            {
                create_arm_gemm_quant<uint8_t, uint8_t>(_arm_gemm, a, b, c, d, act, info);
            }
            break;
        case DataType::S8:
        case DataType::QASYMM8_SIGNED:
            if(d->data_type() == DataType::S32)
            {
                create_arm_gemm<int8_t, int32_t>(_arm_gemm, a, b, c, d, act, info);
            }
            else
            {
                create_arm_gemm_quant<int8_t, int8_t>(_arm_gemm, a, b, c, d, act, info);
            }
            break;
#endif /* __aarch64__ */
#if defined(ARM_COMPUTE_ENABLE_BF16)
        case DataType::BFLOAT16:
            create_arm_gemm<bfloat16, float>(_arm_gemm, a, b, c, d, act, info);
            break;
#endif /* defined(ARM_COMPUTE_ENABLE_BF16) */
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
        case DataType::F16:
            create_arm_gemm<float16_t, float16_t>(_arm_gemm, a, b, c, d, act, info);
            break;
#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
        default:
            break;
    }
}

void CpuGemmAssemblyDispatch::prepare(ITensorPack &tensors)
{
    ARM_COMPUTE_ERROR_ON(_arm_gemm == nullptr);
    _arm_gemm->prepare(tensors);
}

bool CpuGemmAssemblyDispatch::is_configured() const
{
    return _arm_gemm && _arm_gemm->is_configured();
}

void CpuGemmAssemblyDispatch::run(ITensorPack &tensors)
{
    ARM_COMPUTE_ERROR_ON(_arm_gemm == nullptr);
    _arm_gemm->run(tensors);
}

experimental::MemoryRequirements CpuGemmAssemblyDispatch::workspace() const
{
    ARM_COMPUTE_ERROR_ON(_arm_gemm == nullptr);
    return _arm_gemm->workspace();
}
} // namespace cpu
} // namespace arm_compute