Feat/neptune opencl

.circleci/config.yml

@@ -165,6 +165,45 @@ jobs:
             RUST_TEST_THREADS: 1
             FIL_PROOFS_USE_MULTICORE_SDR: true
 
+  # Running with `use_multicore_sdr=true` should be integrated directly into the test code. For now we
+  # just re-run the lifecycle tests to exercise the use_multicore_sdr code path with that setting set.
+  test_multicore_sdr_gpu2:
+    executor: default
+    environment: *setup-env
+    steps:
+      - checkout
+      - attach_workspace:
+          at: "."
+      - restore_cache:
+          keys:
+            - cargo-v28-b-{{ checksum "rust-toolchain" }}-{{ checksum "Cargo.toml" }}-{{ checksum "Cargo.lock" }}-{{ arch }}
+      - restore_parameter_cache
+      - run:
+          name: Test with use_multicore_sdr pairing enabled
+          command: |
+            ulimit -n 20000
+            ulimit -u 20000
+            ulimit -n 20000
+            cargo +<< pipeline.parameters.nightly-toolchain >> -Zpackage-features test --all --verbose --release lifecycle -- --ignored --nocapture
+            cargo +<< pipeline.parameters.nightly-toolchain >> -Zpackage-features test -p storage-proofs-porep --features single-threaded --release checkout_cores -- --test-threads=1
+          no_output_timeout: 30m
+          environment:
+            RUST_TEST_THREADS: 1
+            FIL_PROOFS_USE_MULTICORE_SDR: true
+
+      - run:
+          name: Test with use_multicore_sdr and blst enabled
+          command: |
+            ulimit -n 20000
+            ulimit -u 20000
+            ulimit -n 20000
+            cargo +<< pipeline.parameters.nightly-toolchain >> -Zpackage-features test --all --no-default-features --features gpu2,blst --verbose --release  lifecycle -- --ignored --nocapture
+          no_output_timeout: 30m
+          environment:
+            RUST_TEST_THREADS: 1
+            FIL_PROOFS_USE_MULTICORE_SDR: true
+
+
   test_no_gpu:
     executor: default
     environment: *setup-env
@@ -467,6 +506,11 @@ workflows:
             - cargo_fetch
             - ensure_groth_parameters_and_keys_linux
 
+      - test_multicore_sdr_gpu2:
+          requires:
+            - cargo_fetch
+            - ensure_groth_parameters_and_keys_linux
+
       - test_blst:
           name: test_blst_filecoin_proofs
           crate: "filecoin-proofs"

Cargo.toml

@@ -11,4 +11,3 @@ members = [
   "sha2raw",
   "filecoin-hashers",
 ]
-

fil-proofs-param/Cargo.toml

@@ -68,5 +68,6 @@ heap-profile = ["gperftools/heap"]
 simd = ["storage-proofs-core/simd"]
 asm = ["storage-proofs-core/asm"]
 gpu = ["storage-proofs-core/gpu", "storage-proofs-porep/gpu", "storage-proofs-post/gpu", "bellperson/gpu"]
+gpu2 = ["storage-proofs-core/gpu2", "storage-proofs-porep/gpu2", "storage-proofs-post/gpu2", "bellperson/gpu"]
 pairing = ["storage-proofs-core/pairing", "storage-proofs-porep/pairing", "storage-proofs-post/pairing", "bellperson/pairing"]
 blst = ["storage-proofs-core/blst", "storage-proofs-porep/blst", "storage-proofs-post/blst", "bellperson/blst"]

fil-proofs-tooling/Cargo.toml

@@ -62,6 +62,14 @@ gpu = [
     "bellperson/gpu",
     "filecoin-hashers/gpu",
 ]
+gpu2 = [
+    "storage-proofs-core/gpu2",
+    "storage-proofs-porep/gpu2",
+    "storage-proofs-post/gpu2",
+    "filecoin-proofs/gpu2",
+    "bellperson/gpu",
+    "filecoin-hashers/gpu2",
+]
 measurements = ["storage-proofs-core/measurements"]
 profile = ["storage-proofs-core/profile", "measurements"]
 pairing = [

filecoin-hashers/Cargo.toml

@@ -27,6 +27,7 @@ hex = "0.4.2"
 default = ["gpu", "pairing", "blake2s", "poseidon", "sha256"]
 
 gpu = ["bellperson/gpu", "neptune/gpu"]
+gpu2 = ["bellperson/gpu", "neptune/opencl"]
 
 pairing = ["bellperson/pairing", "neptune/pairing", "bellperson/pairing-serde"]
 blst = ["bellperson/blst", "neptune/blst", "bellperson/blst-serde"]

filecoin-proofs/Cargo.toml

@@ -62,6 +62,14 @@ gpu = [
     "filecoin-hashers/gpu",
     "fr32/gpu",
 ]
+gpu2 = [
+    "storage-proofs-core/gpu2",
+    "storage-proofs-porep/gpu2",
+    "storage-proofs-post/gpu2",
+    "bellperson/gpu",
+    "filecoin-hashers/gpu2",
+    "fr32/gpu",
+]
 pairing = [
     "storage-proofs-core/pairing",
     "storage-proofs-porep/pairing",

storage-proofs-core/Cargo.toml

@@ -64,6 +64,7 @@ measurements = ["cpu-time", "gperftools"]
 profile = ["measurements"]
 
 gpu = ["bellperson/gpu", "neptune/gpu", "filecoin-hashers/gpu", "fr32/gpu"]
+gpu2 = ["bellperson/gpu", "neptune/opencl", "filecoin-hashers/gpu2", "fr32/gpu"]
 pairing = ["bellperson/pairing", "neptune/pairing", "bellperson/pairing-serde", "filecoin-hashers/pairing", "fr32/pairing"]
 blst = ["bellperson/blst", "neptune/blst", "bellperson/blst-serde", "filecoin-hashers/blst", "fr32/blst"]
 

storage-proofs-porep/Cargo.toml

@@ -52,6 +52,7 @@ filecoin-hashers = { path = "../filecoin-hashers", version = "1.0.0", default-fe
 [features]
 default = ["pairing", "gpu"]
 gpu = ["storage-proofs-core/gpu", "filecoin-hashers/gpu", "neptune/gpu", "bellperson/gpu", "fr32/gpu"]
+gpu2 = ["storage-proofs-core/gpu2", "filecoin-hashers/gpu2", "neptune/opencl", "bellperson/gpu", "fr32/gpu"]
 pairing = ["storage-proofs-core/pairing", "bellperson/pairing", "neptune/pairing", "filecoin-hashers/pairing", "fr32/pairing"]
 blst = ["storage-proofs-core/blst", "bellperson/blst", "neptune/blst", "filecoin-hashers/blst", "fr32/blst"]
 single-threaded = []

storage-proofs-porep/src/stacked/vanilla/proof.rs

@@ -375,7 +375,7 @@ impl<'a, Tree: 'static + MerkleTreeTrait, G: 'static + Hasher> StackedDrg<'a, Tr
         Ok(tree)
     }
 
-    #[cfg(feature = "gpu")]
+    #[cfg(any(feature = "gpu", feature = "gpu2"))]
     fn generate_tree_c<ColumnArity, TreeArity>(
         layers: usize,
         nodes_count: usize,
@@ -406,7 +406,7 @@ impl<'a, Tree: 'static + MerkleTreeTrait, G: 'static + Hasher> StackedDrg<'a, Tr
         }
     }
 
-    #[cfg(not(feature = "gpu"))]
+    #[cfg(not(any(feature = "gpu", feature = "gpu2")))]
     fn generate_tree_c<ColumnArity, TreeArity>(
         layers: usize,
         nodes_count: usize,
@@ -428,7 +428,7 @@ impl<'a, Tree: 'static + MerkleTreeTrait, G: 'static + Hasher> StackedDrg<'a, Tr
     }
 
     #[allow(clippy::needless_range_loop)]
-    #[cfg(feature = "gpu")]
+    #[cfg(any(feature = "gpu", feature = "gpu2"))]
     fn generate_tree_c_gpu<ColumnArity, TreeArity>(
         layers: usize,
         nodes_count: usize,
@@ -549,7 +549,10 @@ impl<'a, Tree: 'static + MerkleTreeTrait, G: 'static + Hasher> StackedDrg<'a, Tr
                 s.spawn(move |_| {
                     let _gpu_lock = GPU_LOCK.lock().unwrap();
                     let mut column_tree_builder = ColumnTreeBuilder::<ColumnArity, TreeArity>::new(
+                        #[cfg(feature = "gpu")]
                         Some(BatcherType::GPU),
+                        #[cfg(feature = "gpu2")]
+                        Some(BatcherType::OpenCL),
                         nodes_count,
                         max_gpu_column_batch_size,
                         max_gpu_tree_batch_size,
@@ -731,7 +734,7 @@ impl<'a, Tree: 'static + MerkleTreeTrait, G: 'static + Hasher> StackedDrg<'a, Tr
         })
     }
 
-    #[cfg(feature = "gpu")]
+    #[cfg(any(feature = "gpu", feature = "gpu2"))]
     fn generate_tree_r_last<TreeArity>(
         data: &mut Data<'_>,
         nodes_count: usize,
@@ -764,7 +767,7 @@ impl<'a, Tree: 'static + MerkleTreeTrait, G: 'static + Hasher> StackedDrg<'a, Tr
         }
     }
 
-    #[cfg(not(feature = "gpu"))]
+    #[cfg(not(any(feature = "gpu", feature = "gpu2")))]
     fn generate_tree_r_last<TreeArity>(
         data: &mut Data<'_>,
         nodes_count: usize,
@@ -786,7 +789,7 @@ impl<'a, Tree: 'static + MerkleTreeTrait, G: 'static + Hasher> StackedDrg<'a, Tr
         )
     }
 
-    #[cfg(feature = "gpu")]
+    #[cfg(any(feature = "gpu", feature = "gpu2"))]
     fn generate_tree_r_last_gpu<TreeArity>(
         data: &mut Data<'_>,
         nodes_count: usize,
@@ -894,7 +897,10 @@ impl<'a, Tree: 'static + MerkleTreeTrait, G: 'static + Hasher> StackedDrg<'a, Tr
             s.spawn(move |_| {
                 let _gpu_lock = GPU_LOCK.lock().unwrap();
                 let mut tree_builder = TreeBuilder::<Tree::Arity>::new(
+                    #[cfg(feature = "gpu")]
                     Some(BatcherType::GPU),
+                    #[cfg(feature = "gpu2")]
+                    Some(BatcherType::OpenCL),
                     nodes_count,
                     max_gpu_tree_batch_size,
                     tree_r_last_config.rows_to_discard,
@@ -1306,7 +1312,7 @@ impl<'a, Tree: 'static + MerkleTreeTrait, G: 'static + Hasher> StackedDrg<'a, Tr
     // Assumes data is all zeros.
     // Replica path is used to create configs, but is not read.
     // Instead new zeros are provided (hence the need for replica to be all zeros).
-    #[cfg(feature = "gpu")]
+    #[cfg(any(feature = "gpu", feature = "gpu2"))]
     fn generate_fake_tree_r_last<TreeArity>(
         nodes_count: usize,
         tree_count: usize,
@@ -1341,7 +1347,10 @@ impl<'a, Tree: 'static + MerkleTreeTrait, G: 'static + Hasher> StackedDrg<'a, Tr
 
             let _gpu_lock = GPU_LOCK.lock().unwrap();
             let mut tree_builder = TreeBuilder::<Tree::Arity>::new(
+                #[cfg(feature = "gpu")]
                 Some(BatcherType::GPU),
+                #[cfg(feature = "gpu2")]
+                Some(BatcherType::OpenCL),
                 nodes_count,
                 max_gpu_tree_batch_size,
                 tree_r_last_config.rows_to_discard,
@@ -1436,7 +1445,7 @@ impl<'a, Tree: 'static + MerkleTreeTrait, G: 'static + Hasher> StackedDrg<'a, Tr
     // Assumes data is all zeros.
     // Replica path is used to create configs, but is not read.
     // Instead new zeros are provided (hence the need for replica to be all zeros).
-    #[cfg(not(feature = "gpu"))]
+    #[cfg(not(any(feature = "gpu", feature = "gpu2")))]
     fn generate_fake_tree_r_last<TreeArity>(
         nodes_count: usize,
         tree_count: usize,

storage-proofs-post/Cargo.toml

@@ -39,5 +39,6 @@ filecoin-hashers = { path = "../filecoin-hashers", version = "1.0.0", default-fe
 [features]
 default = ["pairing", "gpu"]
 gpu = ["storage-proofs-core/gpu", "filecoin-hashers/gpu", "fr32/gpu"]
+gpu2 = ["storage-proofs-core/gpu2", "filecoin-hashers/gpu2", "fr32/gpu"]
 pairing = ["storage-proofs-core/pairing", "bellperson/pairing", "neptune/pairing", "filecoin-hashers/pairing", "fr32/pairing"]
 blst = ["storage-proofs-core/blst", "bellperson/blst", "neptune/blst", "filecoin-hashers/blst", "fr32/blst"]