feat: fflonk gpu implementation (#26)

# What ❔

Implements fflonk protocol for the cuda.

## Why ❔


## Checklist


- [ ] PR title corresponds to the body of PR (we generate changelog
entries from PRs).
- [ ] Tests for the changes have been added / updated.
- [ ] Documentation comments have been added / updated.
- [ ] Code has been formatted via `zk fmt` and `zk lint`.

## Prerequisites
- [PR
matter-labs/zksync-crypto#11](matter-labs/zksync-crypto#11)
- [PR #38 ]

---------

Co-authored-by: Michael Carilli <[email protected]>
Co-authored-by: zksync-era-bot <[email protected]>
Co-authored-by: zksync-era-bot <[email protected]>
Co-authored-by: Robert Remen <[email protected]>
Co-authored-by: Daniyar Itegulov <[email protected]>

Cargo.toml

@@ -1,7 +1,6 @@
 [workspace]
-members = [
-    "crates/*"
-]
+members = ["crates/*"]
+exclude = ["crates/proof-compression"]
 resolver = "2"
 
 [workspace.package]
@@ -27,13 +26,17 @@ gpu-ffi-bindings-generator = { version = "=0.152.0", path = "crates/gpu-ffi", pa
 gpu-prover = { version = "=0.152.0", path = "crates/gpu-prover", package = "zksync-gpu-prover" }
 shivini = { version = "=0.152.0", path = "crates/shivini" }
 wrapper-prover = { version = "=0.152.0", path = "crates/wrapper-prover", package = "zksync-wrapper-prover" }
+# fflonk = { version = "=0.152.0", path = "crates/fflonk", package = "fflonk-gpu" }
+# proof-compression = { version = "=0.150.1", path = "crates/proof-compression", package = "proof-compression" }
 
 # These dependencies should be shared by all the crates.
-boojum = "=0.30.5"
-circuit_definitions = "=0.150.10"
-franklin-crypto = "=0.30.5"
-snark_wrapper = "=0.30.5"
-zkevm_test_harness = "=0.150.10"
+circuit_definitions = { version = "=0.150.11" }
+zkevm_test_harness = { version = "=0.150.11" }
+boojum = "=0.30.6"
+franklin-crypto = "=0.30.6"
+rescue_poseidon = "=0.30.6"
+snark_wrapper = "=0.30.6"
+fflonk-cpu = {package = "fflonk", version = "=0.30.6"}
 
 [profile.release]
 debug = "line-tables-only"

crates/fflonk/.gitignore

@@ -0,0 +1 @@
+data/

crates/fflonk/Cargo.toml

@@ -0,0 +1,28 @@
+[package]
+edition.workspace = true
+authors.workspace = true
+homepage.workspace = true
+license.workspace = true
+keywords.workspace = true
+categories.workspace = true
+repository.workspace = true
+version.workspace = true
+name = "fflonk-cuda"
+description = "CUDA implementation of the fflonk prover"
+exclude = ["/data"]
+
+[dependencies]
+fflonk-cpu = {workspace = true}
+circuit_definitions.workspace = true
+gpu-ffi.workspace = true
+rand = "0.4"
+derivative = "2.2"
+byteorder = "1"
+bincode = "1.3"
+serde = { version = "1", features = ["derive", "rc"] }
+serde_json = "1"
+serde_derive = "1"
+
+[features]
+default = ["sanity"]
+sanity = []

crates/fflonk/src/allocator/mod.rs

@@ -0,0 +1,9 @@
+use super::*;
+mod pinned;
+pub use pinned::*;
+
+mod pool;
+pub use pool::*;
+
+mod static_device;
+pub use static_device::*;

crates/fflonk/src/allocator/pinned.rs

@@ -0,0 +1,44 @@
+use super::*;
+
+// Both assembly and device setup has an ability to store data on the pinned memory
+// - Assembly uses for the variables(7487741), state and setup columns
+// - Device setup uses variable indexes and gate selectors
+static mut _STATIC_HOST_ALLOC: Option<GlobalHost> = None;
+
+#[derive(Clone, Debug, Default, Eq, PartialEq)]
+pub struct GlobalHost;
+
+impl GlobalHost {
+    pub fn init(domain_size: usize) -> CudaResult<Self> {
+        let num_variables = 0;
+        let num_cols = 3;
+
+        let size_of_indexes_in_bytes = 8 * num_cols * domain_size;
+        let size_of_vars_in_bytes = 32 * num_variables;
+
+        let total_size_in_bytes = size_of_indexes_in_bytes + size_of_vars_in_bytes;
+
+        todo!()
+    }
+}
+
+pub trait HostAllocator: Allocator + Default + Clone + Send + Sync + 'static {}
+
+unsafe impl Allocator for GlobalHost {
+    fn allocate(
+        &self,
+        layout: std::alloc::Layout,
+    ) -> Result<std::ptr::NonNull<[u8]>, std::alloc::AllocError> {
+        host_allocate(layout.size())
+            .map(|ptr| unsafe { std::ptr::NonNull::new_unchecked(ptr as _) })
+            .map(|ptr| std::ptr::NonNull::slice_from_raw_parts(ptr, layout.size()))
+            .map_err(|_| std::alloc::AllocError)
+    }
+
+    unsafe fn deallocate(&self, ptr: std::ptr::NonNull<u8>, layout: std::alloc::Layout) {
+        host_dealloc(ptr.as_ptr().cast()).expect("deallocate static buffer")
+    }
+}
+
+impl HostAllocator for GlobalHost {}
+impl HostAllocator for std::alloc::Global {}

crates/fflonk/src/allocator/pool.rs

@@ -0,0 +1,76 @@
+use super::*;
+
+pub trait DeviceAllocator: Default {
+    fn allocate(&self, layout: std::alloc::Layout) -> CudaResult<std::ptr::NonNull<[u8]>>;
+    fn allocate_zeroed(&self, layout: std::alloc::Layout) -> CudaResult<std::ptr::NonNull<[u8]>>;
+    fn deallocate(&self, ptr: std::ptr::NonNull<u8>, layout: std::alloc::Layout);
+    fn allocate_async(
+        &self,
+        layout: std::alloc::Layout,
+        pool: bc_mem_pool,
+        stream: bc_stream,
+    ) -> CudaResult<std::ptr::NonNull<[u8]>>;
+    fn deallocate_async(
+        &self,
+        ptr: std::ptr::NonNull<u8>,
+        layout: std::alloc::Layout,
+        stream: bc_stream,
+    );
+    fn allocate_zeroed_async(
+        &self,
+        layout: std::alloc::Layout,
+        pool: bc_mem_pool,
+        stream: bc_stream,
+    ) -> CudaResult<std::ptr::NonNull<[u8]>>;
+}
+
+#[derive(Clone, Debug, Default, Eq, PartialEq)]
+pub struct PoolAllocator;
+
+impl DeviceAllocator for PoolAllocator {
+    fn allocate(&self, layout: std::alloc::Layout) -> CudaResult<std::ptr::NonNull<[u8]>> {
+        unimplemented!("Pool allocator can't do static allocation/deallocation")
+    }
+
+    fn deallocate(&self, ptr: std::ptr::NonNull<u8>, layout: std::alloc::Layout) {
+        unimplemented!("Pool allocator can't do static allocation/deallocation")
+    }
+
+    fn allocate_async(
+        &self,
+        layout: std::alloc::Layout,
+        pool: bc_mem_pool,
+        stream: bc_stream,
+    ) -> CudaResult<std::ptr::NonNull<[u8]>> {
+        allocate_async_on(layout.size(), pool, stream)
+            .map(|ptr| unsafe { std::ptr::NonNull::new_unchecked(ptr as _) })
+            .map(|ptr| std::ptr::NonNull::slice_from_raw_parts(ptr, layout.size()))
+    }
+
+    fn deallocate_async(
+        &self,
+        ptr: std::ptr::NonNull<u8>,
+        _layout: std::alloc::Layout,
+        stream: bc_stream,
+    ) {
+        dealloc_async(ptr.as_ptr().cast(), stream).expect("deallocate")
+    }
+
+    fn allocate_zeroed(&self, layout: std::alloc::Layout) -> CudaResult<std::ptr::NonNull<[u8]>> {
+        let ptr = self.allocate(layout)?;
+        Ok(ptr)
+    }
+    fn allocate_zeroed_async(
+        &self,
+        layout: std::alloc::Layout,
+        pool: bc_mem_pool,
+        stream: bc_stream,
+    ) -> CudaResult<std::ptr::NonNull<[u8]>> {
+        allocate_zeroed_async_on(layout.size(), pool, stream)
+            .map(|ptr| unsafe { std::ptr::NonNull::new_unchecked(ptr as _) })
+            .map(|ptr| std::ptr::NonNull::slice_from_raw_parts(ptr, layout.size()))
+    }
+}
+
+unsafe impl Send for PoolAllocator {}
+unsafe impl Sync for PoolAllocator {}