opencv/modules/gapi/src/compiler/gcompiler.cpp

515 lines
19 KiB
C++

// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
//
// Copyright (C) 2018-2020 Intel Corporation
#include "precomp.hpp"
#include <vector>
#include <stack>
#include <unordered_map>
#include <ade/util/algorithm.hpp> // any_of
#include <ade/util/zip_range.hpp> // zip_range, indexed
#include <ade/graph.hpp>
#include <ade/passes/check_cycles.hpp>
#include "api/gcomputation_priv.hpp"
#include "api/gnode_priv.hpp" // FIXME: why it is here?
#include "api/gproto_priv.hpp" // FIXME: why it is here?
#include "api/gcall_priv.hpp" // FIXME: why it is here?
#include "api/gbackend_priv.hpp" // Backend basic API (newInstance, etc)
#include "compiler/gmodel.hpp"
#include "compiler/gmodelbuilder.hpp"
#include "compiler/gcompiler.hpp"
#include "compiler/gcompiled_priv.hpp"
#include "compiler/gstreaming_priv.hpp"
#include "compiler/passes/passes.hpp"
#include "compiler/passes/pattern_matching.hpp"
#include "executor/gexecutor.hpp"
#include "executor/gstreamingexecutor.hpp"
#include "backends/common/gbackend.hpp"
// <FIXME:>
#if !defined(GAPI_STANDALONE)
#include <opencv2/gapi/cpu/core.hpp> // Also directly refer to Core,
#include <opencv2/gapi/cpu/imgproc.hpp> // ...Imgproc
#include <opencv2/gapi/cpu/video.hpp> // ...and Video kernel implementations
#include <opencv2/gapi/render/render.hpp> // render::ocv::backend()
#endif // !defined(GAPI_STANDALONE)
// </FIXME:>
#include <opencv2/gapi/gcompoundkernel.hpp> // compound::backend()
#include "logger.hpp"
namespace
{
cv::gapi::GKernelPackage getKernelPackage(cv::GCompileArgs &args)
{
auto withAuxKernels = [](const cv::gapi::GKernelPackage& pkg) {
cv::gapi::GKernelPackage aux_pkg;
for (const auto &b : pkg.backends()) {
aux_pkg = combine(aux_pkg, b.priv().auxiliaryKernels());
}
return combine(pkg, aux_pkg);
};
auto has_use_only = cv::gapi::getCompileArg<cv::gapi::use_only>(args);
if (has_use_only)
return withAuxKernels(has_use_only.value().pkg);
static auto ocv_pkg =
#if !defined(GAPI_STANDALONE)
combine(cv::gapi::core::cpu::kernels(),
cv::gapi::imgproc::cpu::kernels(),
cv::gapi::video::cpu::kernels(),
cv::gapi::render::ocv::kernels());
#else
cv::gapi::GKernelPackage();
#endif // !defined(GAPI_STANDALONE)
auto user_pkg = cv::gapi::getCompileArg<cv::gapi::GKernelPackage>(args);
auto user_pkg_with_aux = withAuxKernels(user_pkg.value_or(cv::gapi::GKernelPackage{}));
return combine(ocv_pkg, user_pkg_with_aux);
}
cv::gapi::GNetPackage getNetworkPackage(cv::GCompileArgs &args)
{
return cv::gapi::getCompileArg<cv::gapi::GNetPackage>(args)
.value_or(cv::gapi::GNetPackage{});
}
cv::util::optional<std::string> getGraphDumpDirectory(cv::GCompileArgs& args)
{
auto dump_info = cv::gapi::getCompileArg<cv::graph_dump_path>(args);
if (!dump_info.has_value())
{
const char* path = std::getenv("GRAPH_DUMP_PATH");
return path
? cv::util::make_optional(std::string(path))
: cv::util::optional<std::string>();
}
else
{
return cv::util::make_optional(dump_info.value().m_dump_path);
}
}
template<typename C>
cv::gapi::GKernelPackage auxKernelsFrom(const C& c) {
cv::gapi::GKernelPackage result;
for (const auto &b : c) {
result = cv::gapi::combine(result, b.priv().auxiliaryKernels());
}
return result;
}
// Creates ADE graph from input/output proto args
std::unique_ptr<ade::Graph> makeGraph(const cv::GProtoArgs &ins, const cv::GProtoArgs &outs) {
std::unique_ptr<ade::Graph> pG(new ade::Graph);
ade::Graph& g = *pG;
cv::gimpl::GModel::Graph gm(g);
cv::gimpl::GModel::init(gm);
cv::gimpl::GModelBuilder builder(g);
auto proto_slots = builder.put(ins, outs);
// Store Computation's protocol in metadata
cv::gimpl::Protocol p;
std::tie(p.inputs, p.outputs, p.in_nhs, p.out_nhs) = proto_slots;
gm.metadata().set(p);
return pG;
}
using adeGraphs = std::vector<std::unique_ptr<ade::Graph>>;
// Creates ADE graphs (patterns and substitutes) from pkg's transformations
void makeTransformationGraphs(const cv::gapi::GKernelPackage& pkg,
adeGraphs& patterns,
adeGraphs& substitutes) {
const auto& transforms = pkg.get_transformations();
const auto size = transforms.size();
if (0u == size) return;
// pre-generate all required graphs
patterns.resize(size);
substitutes.resize(size);
for (auto it : ade::util::zip(ade::util::toRange(transforms),
ade::util::toRange(patterns),
ade::util::toRange(substitutes))) {
const auto& t = std::get<0>(it);
auto& p = std::get<1>(it);
auto& s = std::get<2>(it);
auto pattern_comp = t.pattern();
p = makeGraph(pattern_comp.priv().m_ins, pattern_comp.priv().m_outs);
auto substitute_comp = t.substitute();
s = makeGraph(substitute_comp.priv().m_ins, substitute_comp.priv().m_outs);
}
}
void checkTransformations(const cv::gapi::GKernelPackage& pkg,
const adeGraphs& patterns,
const adeGraphs& substitutes) {
const auto& transforms = pkg.get_transformations();
const auto size = transforms.size();
if (0u == size) return;
GAPI_Assert(size == patterns.size());
GAPI_Assert(size == substitutes.size());
const auto empty = [] (const cv::gimpl::SubgraphMatch& m) {
return m.inputDataNodes.empty() && m.startOpNodes.empty()
&& m.finishOpNodes.empty() && m.outputDataNodes.empty()
&& m.inputTestDataNodes.empty() && m.outputTestDataNodes.empty();
};
// **FIXME**: verify other types of endless loops. now, only checking if pattern exists in
// substitute within __the same__ transformation
for (size_t i = 0; i < size; ++i) {
const auto& p = patterns[i];
const auto& s = substitutes[i];
auto matchInSubstitute = cv::gimpl::findMatches(*p, *s);
if (!empty(matchInSubstitute)) {
std::stringstream ss;
ss << "Error: (in transformation with description: '"
<< transforms[i].description
<< "') pattern is detected inside substitute";
throw std::runtime_error(ss.str());
}
}
}
} // anonymous namespace
// GCompiler implementation ////////////////////////////////////////////////////
cv::gimpl::GCompiler::GCompiler(const cv::GComputation &c,
GMetaArgs &&metas,
GCompileArgs &&args)
: m_c(c), m_metas(std::move(metas)), m_args(std::move(args))
{
using namespace std::placeholders;
auto kernels_to_use = getKernelPackage(m_args);
auto networks_to_use = getNetworkPackage(m_args);
std::unordered_set<cv::gapi::GBackend> all_backends;
const auto take = [&](std::vector<cv::gapi::GBackend> &&v) {
all_backends.insert(v.begin(), v.end());
};
take(kernels_to_use.backends());
take(networks_to_use.backends());
m_all_kernels = cv::gapi::combine(kernels_to_use,
auxKernelsFrom(all_backends));
// NB: The expectation in the line above is that
// NN backends (present here via network package) always add their
// inference kernels via auxiliary...()
// sanity check transformations
{
adeGraphs patterns, substitutes;
// FIXME: returning vectors of unique_ptrs from makeTransformationGraphs results in
// compile error (at least) on GCC 9 with usage of copy-ctor of std::unique_ptr, so
// using initialization by lvalue reference instead
makeTransformationGraphs(m_all_kernels, patterns, substitutes);
checkTransformations(m_all_kernels, patterns, substitutes);
// NB: saving generated patterns to m_all_patterns to be used later in passes
m_all_patterns = std::move(patterns);
}
auto dump_path = getGraphDumpDirectory(m_args);
m_e.addPassStage("init");
m_e.addPass("init", "check_cycles", ade::passes::CheckCycles());
m_e.addPass("init", "apply_transformations",
std::bind(passes::applyTransformations, _1, std::cref(m_all_kernels),
std::cref(m_all_patterns))); // Note: and re-using patterns here
m_e.addPass("init", "expand_kernels",
std::bind(passes::expandKernels, _1,
m_all_kernels)); // NB: package is copied
m_e.addPass("init", "topo_sort", ade::passes::TopologicalSort());
m_e.addPass("init", "init_islands", passes::initIslands);
m_e.addPass("init", "check_islands", passes::checkIslands);
// TODO:
// - Check basic graph validity (i.e., all inputs are connected)
// - Complex dependencies (i.e. parent-child) unrolling
// - etc, etc, etc
// Remove GCompoundBackend to avoid calling setupBackend() with it in the list
m_all_kernels.remove(cv::gapi::compound::backend());
m_e.addPassStage("kernels");
m_e.addPass("kernels", "bind_net_params",
std::bind(passes::bindNetParams, _1,
networks_to_use));
m_e.addPass("kernels", "resolve_kernels",
std::bind(passes::resolveKernels, _1,
std::ref(m_all_kernels))); // NB: and not copied here
// (no compound backend present here)
m_e.addPass("kernels", "check_islands_content", passes::checkIslandsContent);
//Input metas may be empty when a graph is compiled for streaming
m_e.addPassStage("meta");
if (!m_metas.empty())
{
m_e.addPass("meta", "initialize", std::bind(passes::initMeta, _1, std::ref(m_metas)));
m_e.addPass("meta", "propagate", std::bind(passes::inferMeta, _1, false));
m_e.addPass("meta", "finalize", passes::storeResultingMeta);
// moved to another stage, FIXME: two dumps?
// m_e.addPass("meta", "dump_dot", passes::dumpDotStdout);
}
// Special stage for backend-specific transformations
// FIXME: document passes hierarchy and order for backend developers
m_e.addPassStage("transform");
m_e.addPassStage("exec");
m_e.addPass("exec", "fuse_islands", passes::fuseIslands);
m_e.addPass("exec", "sync_islands", passes::syncIslandTags);
// FIXME: Since a set of passes is shared between
// GCompiled/GStreamingCompiled, this pass is added here unconditionally
// (even if it is not actually required to produce a GCompiled).
// FIXME: add a better way to do that!
m_e.addPass("exec", "add_streaming", passes::addStreaming);
// Note: Must be called after addStreaming as addStreaming pass
// can possibly add new nodes to the IslandModel
m_e.addPass("exec", "sort_islands", passes::topoSortIslands);
if (dump_path.has_value())
{
m_e.addPass("exec", "dump_dot", std::bind(passes::dumpGraph, _1,
dump_path.value()));
}
// FIXME: This should be called for "ActiveBackends" only (see metadata).
// However, ActiveBackends are known only after passes are actually executed.
// At these stage, they are not executed yet.
ade::ExecutionEngineSetupContext ectx(m_e);
auto backends = m_all_kernels.backends();
for (auto &b : backends)
{
b.priv().addBackendPasses(ectx);
if (!m_metas.empty())
{
b.priv().addMetaSensitiveBackendPasses(ectx);
}
}
}
void cv::gimpl::GCompiler::validateInputMeta()
{
if (m_metas.size() != m_c.priv().m_ins.size())
{
util::throw_error(std::logic_error
("COMPILE: GComputation interface / metadata mismatch! "
"(expected " + std::to_string(m_c.priv().m_ins.size()) + ", "
"got " + std::to_string(m_metas.size()) + " meta arguments)"));
}
const auto meta_matches = [](const GMetaArg &meta, const GProtoArg &proto) {
switch (proto.index())
{
// FIXME: Auto-generate methods like this from traits:
case GProtoArg::index_of<cv::GMat>():
case GProtoArg::index_of<cv::GMatP>():
case GProtoArg::index_of<cv::GFrame>():
return util::holds_alternative<cv::GMatDesc>(meta);
case GProtoArg::index_of<cv::GScalar>():
return util::holds_alternative<cv::GScalarDesc>(meta);
case GProtoArg::index_of<cv::detail::GArrayU>():
return util::holds_alternative<cv::GArrayDesc>(meta);
case GProtoArg::index_of<cv::detail::GOpaqueU>():
return util::holds_alternative<cv::GOpaqueDesc>(meta);
default:
GAPI_Assert(false);
}
return false; // should never happen
};
for (const auto &meta_arg_idx : ade::util::indexed(ade::util::zip(m_metas, m_c.priv().m_ins)))
{
const auto &meta = std::get<0>(ade::util::value(meta_arg_idx));
const auto &proto = std::get<1>(ade::util::value(meta_arg_idx));
if (!meta_matches(meta, proto))
{
const auto index = ade::util::index(meta_arg_idx);
util::throw_error(std::logic_error
("GComputation object type / metadata descriptor mismatch "
"(argument " + std::to_string(index) + ")"));
// FIXME: report what we've got and what we've expected
}
}
// All checks are ok
}
void cv::gimpl::GCompiler::validateOutProtoArgs()
{
for (const auto &out_pos : ade::util::indexed(m_c.priv().m_outs))
{
const auto &node = proto::origin_of(ade::util::value(out_pos)).node;
if (node.shape() != cv::GNode::NodeShape::CALL)
{
auto pos = ade::util::index(out_pos);
util::throw_error(std::logic_error
("Computation output " + std::to_string(pos) +
" is not a result of any operation"));
}
}
}
cv::gimpl::GCompiler::GPtr cv::gimpl::GCompiler::generateGraph()
{
if (!m_metas.empty())
{
// Metadata may be empty if we're compiling our graph for streaming
validateInputMeta();
}
validateOutProtoArgs();
auto g = makeGraph(m_c.priv().m_ins, m_c.priv().m_outs);
if (!m_metas.empty())
{
GModel::Graph(*g).metadata().set(OriginalInputMeta{m_metas});
}
return g;
}
void cv::gimpl::GCompiler::runPasses(ade::Graph &g)
{
m_e.runPasses(g);
GAPI_LOG_INFO(NULL, "All compiler passes are successful");
}
void cv::gimpl::GCompiler::compileIslands(ade::Graph &g)
{
compileIslands(g, m_args);
}
void cv::gimpl::GCompiler::compileIslands(ade::Graph &g, const cv::GCompileArgs &args)
{
GModel::Graph gm(g);
std::shared_ptr<ade::Graph> gptr(gm.metadata().get<IslandModel>().model);
GIslandModel::Graph gim(*gptr);
GIslandModel::compileIslands(gim, g, args);
}
cv::GCompiled cv::gimpl::GCompiler::produceCompiled(GPtr &&pg)
{
// This is the final compilation step. Here:
// - An instance of GExecutor is created. Depending on the platform,
// build configuration, etc, a GExecutor may be:
// - a naive single-thread graph interpreter;
// - a std::thread-based thing
// - a TBB-based thing, etc.
// - All this stuff is wrapped into a GCompiled object and returned
// to user.
// Note: this happens in the last pass ("compile_islands"):
// - Each GIsland of GIslandModel instantiates its own,
// backend-specific executable object
// - Every backend gets a subgraph to execute, and builds
// an execution plan for it (backend-specific execution)
// ...before call to produceCompiled();
const auto &outMetas = GModel::ConstGraph(*pg).metadata()
.get<OutputMeta>().outMeta;
std::unique_ptr<GExecutor> pE(new GExecutor(std::move(pg)));
// FIXME: select which executor will be actually used,
// make GExecutor abstract.
GCompiled compiled;
compiled.priv().setup(m_metas, outMetas, std::move(pE));
return compiled;
}
cv::GStreamingCompiled cv::gimpl::GCompiler::produceStreamingCompiled(GPtr &&pg)
{
GStreamingCompiled compiled;
GMetaArgs outMetas;
// FIXME: the whole below construct is ugly, need to revise
// how G*Compiled learns about its meta.
if (!m_metas.empty())
{
outMetas = GModel::ConstGraph(*pg).metadata().get<OutputMeta>().outMeta;
}
std::unique_ptr<GStreamingExecutor> pE(new GStreamingExecutor(std::move(pg),
m_args));
if (!m_metas.empty() && !outMetas.empty())
{
compiled.priv().setup(m_metas, outMetas, std::move(pE));
}
else if (m_metas.empty() && outMetas.empty())
{
// Otherwise, set it up with executor object only
compiled.priv().setup(std::move(pE));
}
else GAPI_Assert(false && "Impossible happened -- please report a bug");
return compiled;
}
cv::GCompiled cv::gimpl::GCompiler::compile()
{
std::unique_ptr<ade::Graph> pG = generateGraph();
runPasses(*pG);
compileIslands(*pG);
return produceCompiled(std::move(pG));
}
cv::GStreamingCompiled cv::gimpl::GCompiler::compileStreaming()
{
// FIXME: self-note to DM: now keep these compile()/compileStreaming() in sync!
std::unique_ptr<ade::Graph> pG = generateGraph();
GModel::Graph(*pG).metadata().set(Streaming{});
runPasses(*pG);
if (!m_metas.empty())
{
// If the metadata has been passed, compile our islands!
compileIslands(*pG);
}
return produceStreamingCompiled(std::move(pG));
}
void cv::gimpl::GCompiler::runMetaPasses(ade::Graph &g, const cv::GMetaArgs &metas)
{
auto pass_ctx = ade::passes::PassContext{g};
cv::gimpl::passes::initMeta(pass_ctx, metas);
cv::gimpl::passes::inferMeta(pass_ctx, true);
cv::gimpl::passes::storeResultingMeta(pass_ctx);
// Also run meta-sensitive backend-specific passes, if there's any.
// FIXME: This may be hazardous if our backend are not very robust
// in their passes -- how can we guarantee correct functioning in the
// future?
ade::ExecutionEngine engine;
engine.addPassStage("exec"); // FIXME: Need a better decision on how we replicate
// our main compiler stages here.
ade::ExecutionEngineSetupContext ectx(engine);
// NB: &&b or &b doesn't work here since "backends" is a set. Nevermind
for (auto b : GModel::Graph(g).metadata().get<ActiveBackends>().backends)
{
b.priv().addMetaSensitiveBackendPasses(ectx);
}
engine.runPasses(g);
}