diff --git a/README.md b/README.md
index a78ee0e9..0ed9d602 100644
--- a/README.md
+++ b/README.md
@@ -56,11 +56,11 @@ FPS comparison of rendering [Franka](https://github.com/Genesis-Embodied-AI/Gene
Resolution: 128x128
- 
+ 
- 
+ 
## Install (Linux Only)
@@ -104,22 +104,6 @@ renderer = gs.options.renderers.BatchRenderer(
)
```
-### Performance Benchmark
-For comprehensive performance benchmarking across multiple renderers (Madrona, Omniverse, PyRender, ManiSkill), please refer to the detailed documentation in `scripts/perf_benchmark/README.md`.
-
-The benchmark suite includes:
-- Multi-renderer performance testing
-- Batch size and resolution scaling tests
-- Rasterizer vs raytracer comparisons
-- Automated report generation
-- Asset preprocessing utilities
-
-Quick start:
-```bash
-cd scripts/perf_benchmark
-python batch_benchmark.py -f benchmark_config_smoke_test.yml
-```
-
## Acknowledgments
The development of gs-madrona is actively supported by [Genesis AI](https://genesis-ai.company/).
diff --git a/docs/perf_franka_128x128_plot.png b/docs/perf_franka_128x128_plot.png
new file mode 100644
index 00000000..9d3e9e77
Binary files /dev/null and b/docs/perf_franka_128x128_plot.png differ
diff --git a/docs/perf_franka_128x128_table.png b/docs/perf_franka_128x128_table.png
new file mode 100644
index 00000000..eb00e7b8
Binary files /dev/null and b/docs/perf_franka_128x128_table.png differ
diff --git a/include/madrona/context.inl b/include/madrona/context.inl
index 153db67e..c62658da 100644
--- a/include/madrona/context.inl
+++ b/include/madrona/context.inl
@@ -112,12 +112,6 @@ inline void Context::iterateQuery(const Query &query, Fn &&fn)
std::forward(fn));
}
-#ifdef MADRONA_USE_JOB_SYSTEM
-JobID Context::currentJobID() const
-{
- return cur_job_id_;
-}
-#endif
#ifdef MADRONA_MW_MODE
WorldID Context::worldID() const
@@ -126,267 +120,5 @@ WorldID Context::worldID() const
}
#endif
-#if 0
-
-class Context {
- AllocContext mem;
-
- template
- inline ArchetypeRef archetype();
-
- template
- inline void clearArchetype();
-
- template
- inline void clearTemporaries();
-
- template
- inline Query query();
-
- template
- inline void forEach(const Query &query, Fn &&fn);
-
- template
- inline uint32_t numMatches(const Query &query);
-
- // Jobs
- template
- inline JobID submit(Fn &&fn, bool is_child = true,
- DepTs && ... dependencies);
-
- template
- inline JobID submitN(Fn &&fn, uint32_t num_invocations,
- bool is_child = true,
- DepTs && ... dependencies);
-
- // FIXME: currently this function requires that the query reference
- // is valid at least until the returned job is completed.
- template
- inline JobID parallelFor(const Query &query, Fn &&fn,
- bool is_child = true,
- DepTs && ... dependencies);
-
- template
- inline JobID ioRead(const char *path, Fn &&fn, bool is_child = true,
- DepTs && ... dependencies);
-
-#ifdef MADRONA_USE_JOB_SYSTEM
- inline JobID currentJobID() const;
-#endif
-
- template
- inline JobID submitImpl(Fn &&fn, bool is_child, DepTs && ... dependencies);
-
- template
- inline JobID submitNImpl(Fn &&fn, uint32_t num_invocations, bool is_child,
- DepTs && ... dependencies);
-
- template
- inline JobID parallelForImpl(const Query &query, Fn &&fn,
- bool is_child, DepTs && ... dependencies);
-
-private:
- template
- inline JobID submitNImpl(Fn &&fn, uint32_t num_invocations, JobID parent_id,
- DepTs && ... dependencies);
-
-#ifdef MADRONA_USE_JOB_SYSTEM
- JobManager * const job_mgr_;
- StateManager * const state_mgr_;
- StateCache * const state_cache_;
- IOManager * const io_mgr_;
- const int worker_idx_;
- JobID cur_job_id_;
-#endif
-
-friend class JobManager;
-};
-
-template
-ArchetypeRef Context::archetype()
-{
- return state_mgr_->archetype(
- MADRONA_MW_COND(cur_world_id_));
-}
-
-template
-void Context::clearArchetype()
-{
- state_mgr_->clear(MADRONA_MW_COND(cur_world_id_,)
- *state_cache_, false);
-}
-
-template
-void Context::clearTemporaries()
-{
-}
-
-template
-Query Context::query()
-{
- return state_mgr_->query();
-}
-
-template
-void Context::forEach(const Query &query, Fn &&fn)
-{
- state_mgr_->iterateQuery(MADRONA_MW_COND(cur_world_id_,) query,
- std::forward(fn));
-}
-
-template
-uint32_t Context::numMatches(const Query &query)
-{
- uint32_t num_entities = 0;
- state_mgr_->iterateArchetypes(MADRONA_MW_COND(cur_world_id_,) query,
- [&](int num_rows, auto ...) {
- num_entities += num_rows;
- });
-
- return num_entities;
-}
-
-template
-JobID Context::submit(Fn &&fn, bool is_child, Deps && ... dependencies)
-{
- return submitImpl(std::forward(fn), is_child,
- std::forward(dependencies)...);
-}
-
-template
-JobID Context::submitN(Fn &&fn, uint32_t num_invocations,
- bool is_child, Deps && ... dependencies)
-{
- return submitNImpl(std::forward(fn), num_invocations,
- is_child, std::forward(dependencies)...);
-}
-
-template
-JobID Context::parallelFor(const Query &query, Fn &&fn,
- bool is_child, Deps && ... dependencies)
-{
- return parallelForImpl(query, std::forward(fn), is_child,
- std::forward(dependencies)...);
-}
-
-
-template
-inline JobID Context::ioRead(const char *path, Fn &&fn,
- bool is_child, Deps && ... dependencies)
-{
- IOPromise promise = io_mgr_->makePromise();
- Job job = makeJob([promise, fn=std::move(fn), io_mgr=io_mgr_](
- Context &ctx) {
- fn(ctx, io_mgr->getBuffer(promise));
- });
-
- io_mgr_->load(promise, path, job);
-
- (void)is_child;
- ( (void)dependencies, ... );
-
- return JobID::none();
-}
-
-// FIXME: implement is_child, dependencies, num_invocations
-template
-JobID Context::submitImpl(Fn &&fn, bool is_child,
- Deps &&... dependencies)
-{
- JobID parent_id = is_child ? cur_job_id_ : JobID::none();
-
- return job_mgr_->queueJob(worker_idx_,
- std::forward(fn), 0, parent_id,
- MADRONA_MW_COND(cur_world_id_, ) JobPriority::Normal,
- std::forward(dependencies)...);
-}
-
-template
-JobID Context::submitNImpl(Fn &&fn, uint32_t num_invocations, bool is_child,
- Deps && ...dependencies)
-{
- assert(num_invocations > 0);
- JobID parent_id = is_child ? cur_job_id_ : JobID::none();
-
- return submitNImpl(std::forward(fn), num_invocations,
- parent_id, std::forward(dependencies)...);
-}
-
-template
-JobID Context::parallelForImpl(const Query &query, Fn &&fn,
- bool is_child, Deps && ... dependencies)
-{
- if (query.numMatchingArchetypes() == 0) {
- return JobID::none();
- }
-
- // FIXME: add isRunnable check in addition to no dependencies
- if constexpr (sizeof...(dependencies) == 0) {
- JobID parent_id = is_child ? cur_job_id_ : JobID::none();
-
- // Additional optimization: skip this proxy ID when only 1 archetype
- // is present (in fact for > 1 archetype might make sense to just
- // use the else codepath).
- JobID proxy_id = job_mgr_->reserveProxyJobID(worker_idx_, parent_id);
-
- state_mgr_->iterateArchetypes(MADRONA_MW_COND(cur_world_id_,)
- query, [this, &fn, proxy_id](int num_rows, auto ...ptrs) {
- if (num_rows == 0) {
- return;
- }
-
- // Clang complains this is unused without this->
- this->submitNImpl(
- [fn = Fn(fn), ptrs...](ContextT &ctx, uint32_t idx) {
- fn(ctx, ptrs[idx]...);
- }, num_rows, proxy_id);
- });
-
- // Note that even though we "relinquish" the id here, it is still safe
- // to return the ID, since the generation stored in the ID will simply
- // be invalid if the entire parallelFor job finishes, just like a normal
- // job id.
- job_mgr_->relinquishProxyJobID(worker_idx_, proxy_id);
-
- return proxy_id;
- } else {
- return submitImpl([fn = std::forward(fn), &query] (
- ContextT &ctx) {
- ctx.state_mgr_->iterateArchetypes(
- MADRONA_MW_COND(ctx.cur_world_id_,) query,
- [&ctx, &fn](int num_rows, auto ...ptrs) {
- if (num_rows == 0) {
- return;
- }
-
- // FIXME reconsider copying ptrs into the closure here
- // FIXME currently copies the user function's closure
- // Could allow making a fake jobs with data but not a function
- // by extending reserveProxyJobID - that job could be dependent
- // on the parallel for job and hold the user function closure.
- // If we allowed runtime determined # of dependencies, the
- // fast path (no dependencies above) could return the dependent
- // data-only job rather than using the fake ID as a parent
- ctx.template submitNImpl(
- [fn = Fn(fn), ptrs...](ContextT &ctx, uint32_t idx) {
- fn(ctx, ptrs[idx]...);
- }, num_rows, true);
- });
- }, is_child, dependencies...);
- }
-}
-
-template
-JobID Context::submitNImpl(Fn &&fn, uint32_t num_invocations, JobID parent_id,
- Deps && ...dependencies)
-{
- return job_mgr_->queueJob(worker_idx_,
- std::forward(fn), num_invocations, parent_id,
- MADRONA_MW_COND(cur_world_id_, ) JobPriority::Normal,
- std::forward(dependencies)...);
-}
-#endif
}
diff --git a/include/madrona/custom_context.inl b/include/madrona/custom_context.inl
index 09ad1285..10f35d30 100644
--- a/include/madrona/custom_context.inl
+++ b/include/madrona/custom_context.inl
@@ -15,55 +15,5 @@ CustomContext::CustomContext(DataT *world_data,
: Context(world_data, worker_init)
{}
-#if 0
-
-template
-class CustomContext : public Context {
-public:
- template
- inline JobID submit(Fn &&fn, bool is_child = true,
- Deps && ... dependencies);
-
- template
- inline JobID submitN(Fn &&fn, uint32_t num_invocations,
- bool is_child = true,
- Deps && ... dependencies);
-
- template
- inline JobID parallelFor(const Query &query, Fn &&fn,
- bool is_child = true,
- Deps && ... dependencies);
-}
-
-template
-template
-JobID CustomContext::submit(Fn &&fn, bool is_child,
- Deps && ... dependencies)
-{
- return submitImpl(std::forward(fn), is_child,
- std::forward(dependencies)...);
-}
-
-template
-template
-JobID CustomContext::submitN(Fn &&fn,
- uint32_t num_invocations, bool is_child, Deps && ... dependencies)
-{
- return submitNImpl(
- std::forward(fn), num_invocations, is_child,
- std::forward(dependencies)...);
-}
-
-template
-template
-JobID CustomContext::parallelFor(
- const Query &query, Fn &&fn, bool is_child,
- Deps && ... dependencies)
-{
- return parallelForImpl(query, std::forward(fn), is_child,
- std::forward(dependencies)...);
-}
-
-#endif
}
diff --git a/src/core/context.cpp b/src/core/context.cpp
index 1cd4190d..b005b7a8 100644
--- a/src/core/context.cpp
+++ b/src/core/context.cpp
@@ -13,14 +13,6 @@ namespace madrona {
Context::Context(WorldBase *world_data, const WorkerInit &init)
: data_(world_data)
-#ifdef MADRONA_USE_JOB_SYSTEM
- , job_mgr_(init.jobMgr),
- state_mgr_(init.stateMgr),
- state_cache_(init.stateCache),
- io_mgr_(nullptr),
- worker_idx_(init.workerIdx),
- cur_job_id_(JobID::none())
-#endif
#ifdef MADRONA_USE_TASK_GRAPH
, state_mgr_(init.stateMgr)
, state_cache_(init.stateCache)
diff --git a/src/core/worker_init.hpp b/src/core/worker_init.hpp
index 1b54ade5..c25a1b91 100644
--- a/src/core/worker_init.hpp
+++ b/src/core/worker_init.hpp
@@ -14,12 +14,6 @@
namespace madrona {
struct WorkerInit {
-#ifdef MADRONA_USE_JOB_SYSTEM
- JobManager *jobMgr;
- StateManager *stateMgr;
- StateCache *stateCache;
- int workerIdx;
-#endif
#ifdef MADRONA_USE_TASK_GRAPH
StateManager *stateMgr;
StateCache *stateCache;
diff --git a/src/mw/device/include/madrona/context.inl b/src/mw/device/include/madrona/context.inl
index 6618bcb8..0de45efa 100644
--- a/src/mw/device/include/madrona/context.inl
+++ b/src/mw/device/include/madrona/context.inl
@@ -126,205 +126,5 @@ inline void Context::iterateQuery(Query &query, Fn&& fn)
});
}
-#if 0
-
-class Context {
- template
- inline JobID submit(Fn &&fn, bool is_child = true,
- DepTs && ...dependencies);
-
- template
- inline JobID submitN(Fn &&fn, uint32_t num_invocations,
- bool is_child = true, DepTs && ...dependencies);
-
- template
- inline JobID parallelFor(const Query &query, Fn &&fn,
- bool is_child = true, DepTs && ... dependencies);
-
-#if 0
- template
- inline JobID submitImpl(Fn &&fn, bool is_child,
- DepTs && ... dependencies);
-
- template
- inline JobID submitNImpl(Fn &&fn, uint32_t num_invocations, bool is_child,
- DepTs && ... dependencies);
-
- template
- inline JobID parallelForImpl(const Query &query, Fn &&fn,
- bool is_child, DepTs && ... dependencies);
-#endif
-
- void markJobFinished();
-
- inline JobID currentJobID() const { return job_id_; }
-
- inline StateManager & state();
-
- WaveInfo computeWaveInfo();
-
- JobID waveSetupNewJob(uint32_t func_id, bool link_parent,
- uint32_t num_invocations, uint32_t bytes_per_job,
- void **thread_data_store);
-
- JobContainerBase * allocJob(uint32_t bytes_per_job, WaveInfo wave_info);
-
- inline void stageChildJob(uint32_t func_id, uint32_t num_combined_jobs,
- uint32_t bytes_per_job, void *containers);
-
- JobID job_id_;
-
- uint32_t lane_id_;
-}
-
-namespace mwGPU {
-
-// This function is executed at the thread-block granularity.
-// num_invocations <= consts::numMegakernelThreads
-template
-__attribute__((used, always_inline))
-inline void jobEntry(JobContainerBase *job_data,
- uint32_t *data_indices,
- uint32_t *invocation_offsets,
- uint32_t num_invocations,
- uint32_t grid_id)
-{
- uint32_t lane_id = threadIdx.x % consts::numWarpThreads;
-
- uint32_t invocation_idx =
- threadIdx.x + blockIdx.x * consts::numJobLaunchKernelThreads;
-
- if (invocation_idx >= num_invocations) {
- return;
- }
-
- uint32_t data_idx = data_indices[invocation_idx];
- uint32_t local_offset = invocation_offsets[invocation_idx];
-
- static_assert(std::is_trivially_destructible_v);
-
- ContainerT &job_container =
- static_cast(job_data)[data_idx];
-
- ContextT ctx = JobManager::makeContext(job_container.jobID,
- grid_id, job_container.worldID, lane_id);
-
- (job_container.fn)(ctx, local_offset);
-
- ctx.markJobFinished(num_invocations);
-}
-
-template
-struct JobFuncIDBase {
- static uint32_t id;
-};
-
-template ) =
- jobEntry>
-struct JobFuncID : JobFuncIDBase {};
-
-}
-
-Context::Context(WorldBase *world_data, WorkerInit &&init)
- : data_(world_data),
- job_id_(init.jobID),
- grid_id_(init.gridID),
- world_id_(init.worldID),
- lane_id_(init.laneID)
-{}
-
-StateManager & Context::state()
-{
- return *(StateManager *)mwGPU::GPUImplConsts::get().stateManagerAddr;
-}
-
-template
-JobID Context::submit(Fn &&fn, bool is_child, DepTs && ... dependencies)
-{
- return submitImpl(std::forward(fn), 1, is_child,
- std::forward(dependencies)...);
-}
-
-template
-JobID Context::submitN(Fn &&fn, uint32_t num_invocations,
- bool is_child, DepTs && ... dependencies)
-{
- return submitImpl(std::forward(fn), num_invocations, is_child,
- std::forward(dependencies)...);
-}
-
-template
-JobID Context::parallelFor(const Query &query, Fn &&fn,
- bool is_child, DepTs && ... dependencies)
-{
- return parallelForImpl(query, std::forward(fn), is_child,
- std::forward(dependencies)...);
-}
-
-
-template
-JobID Context::submitImpl(Fn &&fn, bool is_child, DepTs && ... dependencies)
-{
- auto wrapper = [fn = std::forward(fn)](ContextT &ctx, uint32_t) {
- fn(ctx);
- };
-
- return submitNImpl(std::move(wrapper), 1, is_child,
- std::forward(dependencies)...);
-}
-
-template
-JobID Context::submitNImpl(Fn &&fn, uint32_t num_invocations, bool is_child,
- DepTs &&... dependencies)
-{
- using namespace mwGPU;
-
- constexpr std::size_t num_deps = sizeof...(DepTs);
-
- using ContainerT = JobContainer;
-
- UserJobTracker *user_trackers =
- JobManager::get()->getUserJobTrackers();
-
- auto wave_info = computeWaveInfo();
-
- // Consolidate dependencies across warp
- // num_deps is guaranteed to be the same across activemask here
-#if 0
- uint32_t num_unique_dependencies = 0;
- auto countUniqueDeps = [&](JobID dep) {
- UserJobTracker &user_tracker = user_trackers[dep.id];
-
- uint32_t shared_id = dep.gen == user_tracker.gen ?
- user_tracker.sharedID : ~0_u32;
-
- uint32_t match_mask = __match_any_sync(wave_info.activeMask,
- shared_id);
-
- uint32_t top_match = getHighestSetBit(match_mask);
-
- uint32_t num_add = (top_match == lane_id_ && shared_id != ~0_u32) ?
- 1 : 0;
- num_unique_dependencies +=
- __reduce_add_sync(wave_info.activeMask, num_add);
- };
-#endif
-
- uint32_t func_id = mwGPU::JobFuncID::id;
-
- void *thread_data_store;
- JobID job_id = waveSetupNewJob(func_id, is_child, num_invocations,
- sizeof(ContainerT), &thread_data_store);
-
- new (thread_data_store) ContainerT(job_id, world_id_,
- num_invocations,
- std::forward(fn),
- std::forward(dependencies)...);
-
- return queue_job_id;
-}
-#endif
}
diff --git a/src/render/CMakeLists.txt b/src/render/CMakeLists.txt
index 73552b37..0c4016b4 100644
--- a/src/render/CMakeLists.txt
+++ b/src/render/CMakeLists.txt
@@ -22,7 +22,6 @@ add_library(madrona_render_core STATIC
render_ctx.hpp render_ctx.cpp
batch_renderer.hpp batch_renderer.cpp
render_common.hpp
- image_util.cpp
)
target_compile_definitions(madrona_render_core PUBLIC
diff --git a/src/render/image_util.cpp b/src/render/image_util.cpp
deleted file mode 100644
index f25fa1b3..00000000
--- a/src/render/image_util.cpp
+++ /dev/null
@@ -1,2 +0,0 @@
-#define STB_IMAGE_WRITE_IMPLEMENTATION
-#include