partitioner.cpp

/*
 * The MIT License (MIT)
 *
 * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
 *
 * 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 "migraphx/target_assignments.hpp"
#include <cstddef>
#include <limits>
#include <iterator>
#include <unordered_map>
#include <unordered_set>

#include <migraphx/env.hpp>
#include <migraphx/algorithm.hpp>
#include <migraphx/stringutils.hpp>
#include <migraphx/partitioner.hpp>
#include <migraphx/pass_manager.hpp>
#include <migraphx/dead_code_elimination.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/program.hpp>
#include <migraphx/make_op.hpp>
#include <migraphx/iterator_for.hpp>
#include <migraphx/ranges.hpp>

MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_DEBUG_PARTITIONER)
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {

static literal get_scalar(instruction_ref ins)
{
    if(ins->name() == "contiguous")
        return get_scalar(ins->inputs().front());
    const auto& s = ins->get_shape();
    if(s.elements() != 1 && not(s.scalar()))
        return {};
    if(not ins->can_eval())
        return {};
    auto e = ins->eval();
    literal r{};
    // needed for bool as visit_at invokes as() which promotes bool to int8
    // Without this we'll break type checks for logical ops that are fused.
    if(e.get_shape().type() == shape::bool_type)
    {
        r = literal{e.at<bool>()};
    }
    else
    {
        e.visit_at([&](auto x) { r = literal{x}; });
    }
    return r;
}

void update_tid_counter(std::size_t tid, std::unordered_map<std::size_t, std::size_t>& tid_counter)
{
    assert(tid != std::numeric_limits<std::size_t>::max());
    if(tid_counter.find(tid) != tid_counter.end())
    {
        tid_counter[tid]++;
    }
    else
    {
        tid_counter[tid] = 0;
    }
}

void generate_run_on_target_modules(migraphx::module_ref mm,
                                    migraphx::program& p,
                                    migraphx::instruction_ref ins,
                                    std::size_t& current_tid,
                                    const target_assignments& tass,
                                    std::unordered_set<instruction_ref>& skip_ins,
                                    std::unordered_map<std::size_t, std::size_t>& tid_counter,
                                    std::vector<instruction_ref>& same_tid_ins_vec,
                                    std::unordered_set<instruction_ref>& same_tid_ins_set)
{
    assert(same_tid_ins_vec.size() == same_tid_ins_set.size());
    if(same_tid_ins_vec.empty())
    {
        assert(current_tid == std::numeric_limits<std::size_t>::max());
        return;
    }
    // gather all parameters
    std::unordered_set<instruction_ref> params;
    // gather all return values
    std::unordered_set<instruction_ref> return_ins;
    for(auto tins : iterator_for(same_tid_ins_vec))
    {
        auto inputs  = (*tins)->inputs();
        auto outputs = (*tins)->outputs();
        transform_if(
            inputs.cbegin(),
            inputs.cend(),
            std::inserter(params, params.end()),
            [&](auto in_param) {
                return (params.count(in_param) == 0 and same_tid_ins_set.count(in_param) == 0);
            },
            [&](auto in_param) { return in_param; });
        if(std::any_of(outputs.begin(), outputs.end(), [&](const auto out_ins) {
               return same_tid_ins_set.count(out_ins) == 0;
           }))
        {
            return_ins.insert(*tins);
        }
    }
    if(enabled(MIGRAPHX_DEBUG_PARTITIONER{}))
    {
        std::cout << "params ins: \n";
        for(auto tmp : iterator_for(params))
        {
            (*tmp)->debug_print();
        }
        std::cout << "\n";
        std::cout << "return ins: \n";
        for(auto tmp : iterator_for(return_ins))
        {
            (*tmp)->debug_print();
        }
        std::cout << "\n";
    }

    auto* tmod = p.create_module("target_mod_" + std::to_string(current_tid) + "_" +
                                 std::to_string(tid_counter[current_tid]));
    update_tid_counter(current_tid, tid_counter);
    std::unordered_map<instruction_ref, instruction_ref> params_map;
    std::size_t param_counter = 0;
    std::vector<instruction_ref> rot_ins_params;
    for(auto pins : iterator_for(params))
    {
        auto scalar = get_scalar(*pins);
        if(scalar.empty())
        {
            params_map[*pins] = tmod->add_parameter("param:" + std::to_string(param_counter++),
                                                    (*pins)->get_shape());
            rot_ins_params.push_back(*pins);
        }
        else
        {
            params_map[*pins] = tmod->add_literal(scalar);
        }
    }
    // TODO: what if params_map is empty ?
    for(auto tins : iterator_for(same_tid_ins_vec))
    {
        auto inputs = (*tins)->inputs();
        std::vector<instruction_ref> new_inputs;
        std::transform(inputs.begin(),
                       inputs.end(),
                       std::back_inserter(new_inputs),
                       [&](auto input_ins) { return params_map.at(input_ins); });
        // [TODO]: what if it is has module args ?
        auto tmod_tins =
            tmod->add_instruction((*tins)->get_operator(), new_inputs, (*tins)->module_inputs());
        // add parameter to params map so that it can be looked up by other insturctions
        params_map[*tins] = tmod_tins;
    }
    std::vector<instruction_ref> rins;
    std::unordered_map<instruction_ref, std::size_t> return_ins_idx_map;
    for(auto ritr : iterator_for(return_ins))
    {
        rins.push_back(params_map.at(*ritr));
        return_ins_idx_map[*ritr] = std::distance(ritr, return_ins.begin());
    }
    tmod->add_return(rins);
    if(enabled(MIGRAPHX_DEBUG_PARTITIONER{}))
    {
        std::cout << "tmod: \n";
        tmod->debug_print();
    }
    // add run_on_target ins
    auto rot_ins = mm->insert_instruction(
        ins, make_op("run_on_target", {{"target_id", current_tid}}), rot_ins_params, {tmod});
    skip_ins.insert(rot_ins);
    // fetch return instructions from tuple
    for(auto mm_rins : iterator_for(return_ins))
    {
        auto tuple_elem_ins = mm->insert_instruction(
            ins, make_op("get_tuple_elem", {{"index", return_ins_idx_map.at(*mm_rins)}}), rot_ins);
        skip_ins.insert(tuple_elem_ins);
        // replace returns from tmod in main module
        mm->replace_instruction(*mm_rins, tuple_elem_ins);
    }
    dead_code_elimination{}.apply(*mm);
    // update current_tid
    same_tid_ins_set.clear();
    same_tid_ins_vec.clear();
    if(tass.find(ins) != tass.end())
    {
        current_tid = tass.at(ins);
        update_tid_counter(current_tid, tid_counter);
        same_tid_ins_set.insert(ins);
        same_tid_ins_vec.push_back(ins);
    }
    else
    {
        current_tid = std::numeric_limits<std::size_t>::max();
    }
    if(enabled(MIGRAPHX_DEBUG_PARTITIONER{}))
    {
        std::cout << "module after creation of tmod and rot: \n";
        mm->debug_print();
    }
}
/*
Given target assignments (tass) for the instructions, generate run_on_target modules subgraphs
automatically. Input graph should be uncompiled migraphx program. target assignments (tass) map
should have a map of instruction to target_id. Instructions that are not inside tass map are
considered to be targeted for the "Ref" by default. params, literals and other builtins shouldn't be
part of the tass, only compute and reshaper instructions should be part of tass. Copy, sync and
alloc instructions would be generated by compiler at later stage, so those shouldn't be considered.
(TODO): CustomOps may require special handling.

Step 1:
Identify subgraph boundaries
Ref is used for instructions that do not have assignments.
Boundaries can happen in following cases.
1.  Ref --> Target X --> Ref
2.  Ref --> Target X --> Target Y
3.  Target X --> Target Y --> Target Z , in this case target X and target Z can be same
4.  Target X --> "@return"
5.  Target X --> Ref --> "@return"

Each of those identified regions could have futher nested sub modules which needs to be handled
separately.

Step 2:
Collect parameters and return instructions for the subgraphs identified in Step 1.

Step 3:
Create modules using information collected in step 2 and insert run_on_target instructions.
*/

void partition(migraphx::module_ref mm,
               migraphx::program& p,
               const target_assignments& tass,
               std::unordered_map<std::size_t, std::size_t>& tid_counter)
{
    mm->sort();
    if(enabled(MIGRAPHX_DEBUG_PARTITIONER{}))
    {
        std::cout << "sorted module: \n";
        mm->debug_print();
    }
    std::vector<instruction_ref> same_tid_ins_vec;
    std::unordered_set<instruction_ref> same_tid_ins_set;
    // walk the graph in reverse-DFS order
    size_t current_tid = std::numeric_limits<std::size_t>::max();
    std::unordered_set<instruction_ref> skip_ins;
    for(auto ins : iterator_for(*mm))
    {
        // gather instructions belonging to the same target_id
        // for now, make sure that all the inputs to the insturctions are also from the same
        // target_id, if not create another module
        // skip all the builtins
        if(enabled(MIGRAPHX_DEBUG_PARTITIONER{}))
        {
            std::cout << "currently processing: \n";
            ins->debug_print();
            std::cout << "\n";
        }
        if(skip_ins.count(ins) == 0)
        {
            if(not ins->module_inputs().empty())
            {
                for(auto sub_mod : ins->module_inputs())
                {
                    partition(sub_mod, p, tass, tid_counter);
                }
                mm->replace_instruction(
                    ins, ins->get_operator(), ins->inputs(), ins->module_inputs());
            }
        }
        if(ins->name() == "@return")
        {
            generate_run_on_target_modules(mm,
                                           p,
                                           ins,
                                           current_tid,
                                           tass,
                                           skip_ins,
                                           tid_counter,
                                           same_tid_ins_vec,
                                           same_tid_ins_set);
        }
        // skip all params, literal and builitins other than return, skip "run_on_target_mod" ins
        else if(starts_with(ins->name(), "@") or skip_ins.count(ins) != 0)
        {
            continue;
        }
        else if(tass.find(ins) == tass.end())
        {
            generate_run_on_target_modules(mm,
                                           p,
                                           ins,
                                           current_tid,
                                           tass,
                                           skip_ins,
                                           tid_counter,
                                           same_tid_ins_vec,
                                           same_tid_ins_set);
        }
        else if(current_tid == std::numeric_limits<std::size_t>::max())
        {
            current_tid = tass.at(ins);
            update_tid_counter(current_tid, tid_counter);
            same_tid_ins_vec.push_back(ins);
            same_tid_ins_set.insert(ins);
        }
        else if(tass.at(ins) == current_tid)
        {
            same_tid_ins_vec.push_back(ins);
            same_tid_ins_set.insert(ins);
        }
        else if(tass.at(ins) != current_tid)
        {
            generate_run_on_target_modules(mm,
                                           p,
                                           ins,
                                           current_tid,
                                           tass,
                                           skip_ins,
                                           tid_counter,
                                           same_tid_ins_vec,
                                           same_tid_ins_set);
        }
        else
        {
            MIGRAPHX_THROW("Partition: this shouldn't occur");
        }
    }
}

void partition(migraphx::program& p, const target_assignments& tass)
{
    auto* mm = p.get_main_module();
    // sort the graph in reverse post order DFS order
    std::unordered_map<std::size_t, std::size_t> tid_counter;
    partition(mm, p, tass, tid_counter);
    dead_code_elimination{}.apply(p);
}

} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx