feat: support CUDA GPU Compute Capability 9.0 and 10.0

.github/runs-on.yml

@@ -7,5 +7,4 @@ images:
 runners:
   gpu-nvidia:
     cpu: [32, 64]
-    family: ["g6"]
-    hdd: 200
+    family: ["g6"]

.github/workflows/docker.yml

@@ -12,7 +12,8 @@ jobs:
       "runs-on",
       "runner=32cpu-linux-x64",
       "run-id=${{ github.run_id }}",
-      "spot=false"
+      "spot=false",
+      "disk=large"
     ]
     steps:
 

Dockerfile.server

@@ -1,5 +1,5 @@
 # Start with a base image that includes CUDA
-FROM nvidia/cuda:12.5.1-devel-ubuntu22.04 AS builder
+FROM nvidia/cuda:12.9.0-devel-ubuntu22.04 AS builder
 
 # Install necessary packages
 RUN apt-get update && apt-get install -y \
@@ -39,6 +39,10 @@ RUN mkdir -p /root/.ssh && chmod 700 /root/.ssh \
     && ssh-keyscan github.com >> /root/.ssh/known_hosts \
     && git config --global url."[email protected]:".insteadOf "https://github.com/"
 
+RUN nvcc --version
+
+RUN df -h
+
 # Build the Rust project
 RUN --mount=type=ssh CARGO_NET_GIT_FETCH_WITH_CLI=true cargo build -p moongate-server --release
 

Dockerfile.toolchain

@@ -1,5 +1,5 @@
 # Start with a base image that includes CUDA
-FROM nvidia/cuda:12.5.1-devel-ubuntu22.04 AS builder
+FROM nvidia/cuda:12.9.0-devel-ubuntu22.04 AS builder
 
 # Install necessary packages
 RUN apt-get update && DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC  apt-get install -y \
@@ -15,6 +15,8 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC  apt-get install
 RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
 ENV PATH="/root/.cargo/bin:${PATH}"
 
+RUN nvcc --version
+
 # Set the working directory in the container
 WORKDIR /usr/src/app
 

core/build.rs

@@ -188,7 +188,10 @@ fn main() {
         cc_builder.flag("--expt-relaxed-constexpr");
         cc_builder.flag("-gencode=arch=compute_86,code=sm_86");
         cc_builder.flag("-gencode=arch=compute_89,code=sm_89");
-        cc_builder.flag("-gencode=arch=compute_89,code=compute_89");
+        cc_builder.flag("-gencode=arch=compute_90,code=sm_90");
+        cc_builder.flag("-gencode=arch=compute_100,code=sm_100");
+        cc_builder.flag("-gencode=arch=compute_120,code=sm_120");
+        cc_builder.flag("-gencode=arch=compute_120,code=compute_120");
 
         env::set_var("DEP_SPPARK_ROOT", "../sppark");
         if let Some(include) = env::var_os("DEP_SPPARK_ROOT") {

core/src/challenger/grinding_challenger.rs

@@ -109,7 +109,7 @@
     use sp1_stark::{inner_perm, InnerChallenger};
 
     use crate::{
         challenger::grinding_challenger::DeviceGrindingChallenger, cuda_runtime::stream::CudaStream,
     };
 
     #[test]
@@ -117,7 +117,7 @@
         (1..20).for_each(|bits| {
             let mut challenger = InnerChallenger::new(inner_perm());
 
             let stream = CudaStream::default();
 
             // Observe 7 elements to make the input buffer almost full and trigger duplexing on
             // device.
@@ -136,12 +136,23 @@
             // Clone the original challenger because after grinding on device the internal state
             // of `challenger` will change.
             let mut original_challenger = challenger.clone();
-            let result = challenger.grind_device(bits, &stream);
+            let result = challenger.grind(bits);
 
             assert!(original_challenger.check_witness(bits, result));
 
-            let mut original_challenger_2 = challenger_2.clone();
-            let result_2 = challenger_2.grind_device(bits, &stream);
+            let mut original_challenger_2: p3_challenger::DuplexChallenger<
+                BabyBear,
+                p3_poseidon2::Poseidon2<
+                    BabyBear,
+                    p3_poseidon2::Poseidon2ExternalMatrixGeneral,
+                    p3_baby_bear::DiffusionMatrixBabyBear,
+                    16,
+                    7,
+                >,
+                16,
+                8,
+            > = challenger_2.clone();
+            let result_2 = challenger_2.grind(bits);
 
             assert!(original_challenger_2.check_witness(bits, result_2));
 

core/src/fri/open.rs

@@ -7,13 +7,13 @@
 use sp1_stark::Com;
 
 use p3_baby_bear::BabyBear;
-use p3_challenger::{CanObserve, CanSample, CanSampleBits};
+use p3_challenger::{CanObserve, CanSample, CanSampleBits, GrindingChallenger};
 use p3_field::{AbstractExtensionField, AbstractField, TwoAdicField};
 use p3_fri::{BatchOpening, CommitPhaseProofStep, FriProof, QueryProof};
 use sp1_stark::Challenger;
 
 use crate::{
     challenger::grinding_challenger::DeviceGrindingChallenger,
     cuda_runtime::stream::CudaStream,
     device::{
         memory::{ToDevice, ToHost},
@@ -84,7 +84,7 @@
             .in_scope(|| commit_phase(committer, input, log_max_height, challenger));
 
         let pow_witness = tracing::debug_span!("pow witness")
-            .in_scope(|| challenger.grind_device(config.proof_of_work_bits, main_stream));
+            .in_scope(|| challenger.grind(config.proof_of_work_bits));
 
         let query_indices: Vec<usize> =
             (0..config.num_queries).map(|_| challenger.sample_bits(log_max_height)).collect();

perf/run_s3.sh

@@ -19,6 +19,7 @@ aws s3 cp s3://sp1-testing-suite/$s3_path/stdin.bin stdin.bin
 export RUST_LOG=debug
 export FIX_CORE_SHAPES=true
 export FIX_RECURSION_SHAPES=true
+export SP1_ALLOW_DEPRECATED_HOOKS=true
 
 # Run moongate-perf
 RUST_BACKTRACE=full cargo run -p moongate-perf --release -- --program-path program.bin --stdin-path stdin.bin --stage $stage --trace nvtx