proofman-common 1.3.1-alpha

use std::collections::{HashMap, HashSet};
use std::ffi::c_void;

use proofman_fields::PrimeField64;
use proofman_starks_lib_c::{expressions_bin_new_c, expressions_bin_free_c};

use crate::format_bytes;
use crate::load_const_pols;
use crate::{GlobalInfo, GlobalInfoAir};
use crate::ProofmanError;
use crate::ProofmanResult;
use crate::Setup;
use crate::exec_header;
use crate::ProofType;

/// The slot an air's packed const pols occupy in the GPU const buffer. Airs with identical
/// fixed columns -- same verkey, i.e. same const-tree root -- share one slot; `owner` is the
/// first of the group and the only one that uploads.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub struct FixedGroup {
    pub owner: (usize, usize),
}

/// Columns the recursion's HOST trace buffer needs for one air.
///
/// On GPU it only stages the exec map's width for `widenCompactWitnessGPU` to widen device-side. On
/// CPU it is the air's full width: since the solve and the scatter are fused, `generate_witness`
/// takes a pooled buffer before the prover buffer is in scope, so the fill can no longer go straight
/// into the prover's own cm1 slot. Must agree with `recursion_trace_stride`, which decides the fill
/// from the same exec header -- sizing narrow while the fill goes wide overruns the buffer.
///
/// One function because three call sites decide this, and the one that did it inline sized the whole
/// pool at the air's full width while the other two were compact.
pub fn recursion_staging_cols<F: PrimeField64>(setup: &Setup<F>, gpu: bool) -> u64 {
    let cm1 = setup.stark_info.map_sections_n["cm1"];
    if !gpu {
        return cm1;
    }
    setup.exec_data.as_deref().map(|e| exec_header(e).map_cols).filter(|&m| m > 0 && m < cm1).unwrap_or(cm1)
}

pub struct SetupsVadcop<F: PrimeField64> {
    pub sctx_compressor: Option<SetupCtx<F>>,
    pub sctx_recursive1: Option<SetupCtx<F>>,
    pub sctx_recursive2: Option<SetupCtx<F>>,
    pub setup_vadcop_final: Option<Setup<F>>,
    pub setup_vadcop_final_compressed: Option<Setup<F>>,
    pub max_compact_trace_size: usize,
    pub max_const_size: usize,
    pub max_const_tree_size: usize,
    pub max_prover_buffer_size: usize,
    pub max_prover_recursive_buffer_size: usize,
    pub max_prover_recursive2_buffer_size: usize,
    pub max_pinned_proof_size: usize,
    pub max_n_bits_ext: usize,
    pub total_const_pols_size: usize,
    pub recurser_const_slot_size: usize,
}

unsafe impl<F: PrimeField64> Send for SetupsVadcop<F> {}
unsafe impl<F: PrimeField64> Sync for SetupsVadcop<F> {}

impl<F: PrimeField64> SetupsVadcop<F> {
    pub fn new(
        global_info: &GlobalInfo,
        verify_constraints: bool,
        aggregation: bool,
        gpu: bool,
    ) -> ProofmanResult<Self> {
        if aggregation {
            let sctx_compressor = SetupCtx::new(global_info, &ProofType::Compressor, verify_constraints, gpu)?;
            let sctx_recursive1 = SetupCtx::new(global_info, &ProofType::Recursive1, verify_constraints, gpu)?;
            let sctx_recursive2 = SetupCtx::new(global_info, &ProofType::Recursive2, verify_constraints, gpu)?;
            let setup_vadcop_final = Setup::new(
                &global_info.get_setup_path("vadcop_final"),
                0,
                0,
                &GlobalInfoAir::new("VadcopFinal".to_string()),
                &ProofType::VadcopFinal,
                verify_constraints,
                gpu,
                None,
            )?;

            // Only if the key says it carries the stage. `Setup::new` reads the starkinfo from disk
            // and panics if it is absent, so a key built without the compressed final -- which is
            // the default for blake3, where it measured a 2% smaller proof for a whole extra
            // recursion layer -- would fail here, before proving even starts.
            let setup_vadcop_final_compressed = if global_info.has_compressed_final {
                Some(Setup::new(
                    &global_info.get_setup_path("vadcop_final_compressed"),
                    0,
                    0,
                    &GlobalInfoAir::new("VadcopFinalCompressed".to_string()),
                    &ProofType::VadcopFinalCompressed,
                    verify_constraints,
                    gpu,
                    None,
                )?)
            } else {
                None
            };

            let recurser_const_slot_size = if gpu {
                let n_constants = setup_vadcop_final.stark_info.n_constants as usize;
                let n_rows = 1usize << setup_vadcop_final.stark_info.stark_struct.n_bits;
                1 + n_constants + n_rows * n_constants
            } else {
                0
            };

            let total_const_pols_size = sctx_compressor.total_const_pols_size
                + sctx_recursive1.total_const_pols_size
                + sctx_recursive2.total_const_pols_size
                + setup_vadcop_final.const_pols_size_packed
                + setup_vadcop_final_compressed.as_ref().map_or(0, |s| s.const_pols_size_packed)
                + recurser_const_slot_size;

            let max_const_size = sctx_compressor
                .max_const_size
                .max(sctx_recursive1.max_const_size)
                .max(sctx_recursive2.max_const_size)
                .max(setup_vadcop_final.const_pols_size);
            let max_const_tree_size = sctx_compressor
                .max_const_tree_size
                .max(sctx_recursive1.max_const_tree_size)
                .max(sctx_recursive2.max_const_tree_size)
                .max(setup_vadcop_final.const_tree_size)
                .max(setup_vadcop_final_compressed.as_ref().map_or(0, |s| s.const_tree_size));
            let max_prover_buffer_size = sctx_compressor
                .max_prover_buffer_size
                .max(sctx_recursive1.max_prover_buffer_size)
                .max(sctx_recursive2.max_prover_buffer_size)
                .max(setup_vadcop_final.prover_buffer_size as usize)
                .max(setup_vadcop_final_compressed.as_ref().map_or(0, |s| s.prover_buffer_size as usize));

            // Recursive-capable buffers = prover buffer (mapTotalN) + witness tail: the room past the
            // proof layout where the recursive proof's COMPACT host trace lands (staging cols x N:
            // each repository's `max_compact_trace_size`, the finals' own staging width) before the device
            // widens it into cm1; gen_recursive_proof_gpu checks the room at launch. Sized exactly
            // to that trace -- 4e44d5549 sized it at the full trace, ~1 GB more than ever lands.
            let vadcop_final_tail = (recursion_staging_cols(&setup_vadcop_final, gpu)
                * (1 << setup_vadcop_final.stark_info.stark_struct.n_bits))
                as usize;
            // Its OWN staging width, not vadcop_final's: the two airs have had identical geometry so
            // far, which is why pairing them never showed; it would undersize the buffer the moment
            // they diverged.
            let vadcop_final_compressed_tail = setup_vadcop_final_compressed
                .as_ref()
                .map_or(0, |s| (recursion_staging_cols(s, gpu) * (1 << s.stark_info.stark_struct.n_bits)) as usize);
            let max_prover_recursive2_buffer_size = (sctx_recursive2.max_prover_buffer_size
                + sctx_recursive2.max_compact_trace_size)
                .max(sctx_recursive1.max_prover_buffer_size + sctx_recursive1.max_compact_trace_size);

            let max_prover_recursive_buffer_size = (sctx_recursive2.max_prover_buffer_size
                + sctx_recursive2.max_compact_trace_size)
                .max(sctx_recursive1.max_prover_buffer_size + sctx_recursive1.max_compact_trace_size)
                .max(sctx_compressor.max_prover_buffer_size + sctx_compressor.max_compact_trace_size)
                .max(setup_vadcop_final.prover_buffer_size as usize + vadcop_final_tail)
                .max(
                    setup_vadcop_final_compressed
                        .as_ref()
                        .map_or(0, |c| c.prover_buffer_size as usize + vadcop_final_compressed_tail),
                );

            // This floors every non-recursive GPU stream class, so which term dominates decides
            // whether a regular class fits at all.
            tracing::debug!(
                "Recursive buffer requirement: compressor {}, recursive1 {}, recursive2 {}, vadcop_final {}, vadcop_final_compressed {}",
                format_bytes((sctx_compressor.max_prover_buffer_size + sctx_compressor.max_compact_trace_size) as f64 * 8.0),
                format_bytes((sctx_recursive1.max_prover_buffer_size + sctx_recursive1.max_compact_trace_size) as f64 * 8.0),
                format_bytes((sctx_recursive2.max_prover_buffer_size + sctx_recursive2.max_compact_trace_size) as f64 * 8.0),
                format_bytes((setup_vadcop_final.prover_buffer_size as usize + vadcop_final_tail) as f64 * 8.0),
                format_bytes(
                    setup_vadcop_final_compressed
                        .as_ref()
                        .map_or(0, |c| c.prover_buffer_size as usize + vadcop_final_compressed_tail) as f64
                        * 8.0
                ),
            );

            let max_pinned_proof_size = sctx_compressor
                .max_pinned_proof_size
                .max(sctx_recursive1.max_pinned_proof_size)
                .max(sctx_recursive2.max_pinned_proof_size)
                .max(setup_vadcop_final.proof_size as usize)
                .max(setup_vadcop_final_compressed.as_ref().map_or(0, |c| c.proof_size as usize));

            let max_n_bits_ext = sctx_compressor
                .max_n_bits_ext
                .max(sctx_recursive1.max_n_bits_ext)
                .max(sctx_recursive2.max_n_bits_ext)
                .max(setup_vadcop_final.stark_info.stark_struct.n_bits_ext as usize);

            // Largest compact (staging-width) trace over every recursive proof kind, compressor
            // included: they share one host pool (see MemoryHandlerRecursive::trace), and it is also
            // the witness tail each recursive-capable stream class reserves past mapTotalN.
            let max_compact_trace_size = sctx_recursive1
                .max_compact_trace_size
                .max(sctx_recursive2.max_compact_trace_size)
                .max(sctx_compressor.max_compact_trace_size)
                // Their STAGING width, not `vadcop_final_trace_size` -- that one is the prover
                // buffer's and stays full. Sizing the shared pool from it put every buffer at the
                // air's full width, which is what made the pool 7x what it holds.
                .max(vadcop_final_tail)
                .max(vadcop_final_compressed_tail);

            Ok(SetupsVadcop {
                sctx_compressor: Some(sctx_compressor),
                sctx_recursive1: Some(sctx_recursive1),
                sctx_recursive2: Some(sctx_recursive2),
                setup_vadcop_final: Some(setup_vadcop_final),
                setup_vadcop_final_compressed,
                max_const_tree_size,
                max_const_size,
                max_prover_buffer_size,
                max_prover_recursive_buffer_size,
                max_prover_recursive2_buffer_size,
                max_pinned_proof_size,
                max_n_bits_ext,
                max_compact_trace_size,
                total_const_pols_size,
                recurser_const_slot_size,
            })
        } else {
            Ok(SetupsVadcop {
                sctx_compressor: None,
                sctx_recursive1: None,
                sctx_recursive2: None,
                setup_vadcop_final: None,
                setup_vadcop_final_compressed: None,
                total_const_pols_size: 0,
                recurser_const_slot_size: 0,
                max_const_tree_size: 0,
                max_const_size: 0,
                max_prover_buffer_size: 0,
                max_prover_recursive_buffer_size: 0,
                max_prover_recursive2_buffer_size: 0,
                max_pinned_proof_size: 0,
                max_n_bits_ext: 0,
                max_compact_trace_size: 0,
            })
        }
    }

    /// Size of every pooled signalValues buffer (in u64 elements): the largest `total_signal_no`
    /// across the recursive/compressor/final setups, so every circuit takes a pooled buffer and
    /// none allocates its own.
    ///
    /// Sizing from the second largest would spare `n_streams x (largest - second)` of host memory
    /// -- 328 MB per stream on the blake3 key, where the sole compressor (Keccakf) is 3.1x the next
    /// -- and it is not worth it: that made the compressor allocate and first-touch 485 MB per
    /// proof, measured at +37.6 ms on its witness (204.1 vs 166.5 ms, n=36 each). 0 if no setup
    /// exposes a `total_signal_no`, and the pool then stays off.
    pub fn signal_pool_cap(&self) -> usize {
        let mut sizes: Vec<usize> = Vec::new();
        for sctx in [self.sctx_compressor.as_ref(), self.sctx_recursive1.as_ref(), self.sctx_recursive2.as_ref()]
            .into_iter()
            .flatten()
        {
            sizes.extend(sctx.total_signal_nos());
        }
        for setup in
            [self.setup_vadcop_final.as_ref(), self.setup_vadcop_final_compressed.as_ref()].into_iter().flatten()
        {
            if let Some(n) = setup.total_signal_no {
                sizes.push(n as usize);
            }
        }
        sizes.into_iter().max().unwrap_or(0)
    }

    pub fn get_setup(&self, airgroup_id: usize, air_id: usize, setup_type: &ProofType) -> ProofmanResult<&Setup<F>> {
        match setup_type {
            ProofType::Compressor => self.sctx_compressor.as_ref().unwrap().get_setup(airgroup_id, air_id),
            ProofType::Recursive1 => self.sctx_recursive1.as_ref().unwrap().get_setup(airgroup_id, air_id),
            ProofType::Recursive2 => self.sctx_recursive2.as_ref().unwrap().get_setup(airgroup_id, air_id),
            ProofType::VadcopFinal => Ok(self.setup_vadcop_final.as_ref().unwrap()),
            // The only optional setup of the family: keys built without the compressed-final
            // stage (blake3's default) legitimately have none.
            ProofType::VadcopFinalCompressed => self.setup_vadcop_final_compressed.as_ref().ok_or_else(|| {
                ProofmanError::InvalidSetup("Proving key was built without the vadcop_final_compressed stage".into())
            }),
            _ => Err(ProofmanError::InvalidSetup("Invalid setup type".into())),
        }
    }
}

#[derive(Debug)]
pub struct SetupRepository<F: PrimeField64> {
    setups: HashMap<(usize, usize), Setup<F>>,
    max_const_tree_size: usize,
    max_const_size: usize,
    max_prover_buffer_size: usize,
    prover_buffer_sizes: Vec<((usize, usize), usize)>,
    max_prover_contributions_size: usize,
    max_pinned_proof_size: usize,
    max_compact_trace_size: usize,
    total_const_pols_size: usize,
    total_custom_commits_reserved_words: usize,
    fixed_groups: HashMap<(usize, usize), FixedGroup>,
    global_bin: Option<*mut c_void>,
    global_info_file: String,
    max_n_bits_ext: usize,
    max_const_pols_size_packed: usize,
    const_slot_cache_slots: usize,
}

unsafe impl<F: PrimeField64> Send for SetupRepository<F> {}
unsafe impl<F: PrimeField64> Sync for SetupRepository<F> {}

impl<F: PrimeField64> Drop for SetupRepository<F> {
    fn drop(&mut self) {
        if let Some(global_bin_ptr) = self.global_bin {
            expressions_bin_free_c(global_bin_ptr);
        }
    }
}

impl<F: PrimeField64> SetupRepository<F> {
    pub fn new(
        global_info: &GlobalInfo,
        setup_type: &ProofType,
        verify_constraints: bool,
        gpu: bool,
    ) -> ProofmanResult<Self> {
        let mut setups = HashMap::new();

        let global_bin = match setup_type == &ProofType::Basic {
            true => {
                let global_bin_path =
                    &global_info.get_proving_key_path().join("pilout.globalConstraints.bin").display().to_string();
                Some(expressions_bin_new_c(global_bin_path.as_str(), true, false))
            }
            false => None,
        };

        let global_info_path = &global_info.get_proving_key_path().join("pilout.globalInfo.json");
        let global_info_file = global_info_path.to_str().unwrap().to_string();

        let mut max_const_tree_size = 0;
        let mut max_const_size = 0;
        let mut max_n_bits_ext = 0;
        let mut max_prover_contributions_size = 0;
        let mut max_prover_buffer_size = 0;
        let mut prover_buffer_sizes: Vec<((usize, usize), usize)> = Vec::new();
        let mut max_pinned_proof_size = 0;
        let mut total_const_pols_size = 0;
        let mut total_custom_commits_reserved_words = 0;
        let mut max_compact_trace_size = 0;
        let mut max_const_pols_size_packed = 0;
        let mut n_const_slots = 0;

        // Airs in the order load_device_const_pols walks them, and the slot each verkey maps to.
        let mut sized_airs: Vec<(usize, usize)> = Vec::new();
        let mut groups: HashMap<Vec<u64>, (usize, usize)> = HashMap::new();

        // Initialize Hashmap for each airgroup_id, air_id

        for (airgroup_id, air_group) in global_info.airs.iter().enumerate() {
            for (air_id, _) in air_group.iter().enumerate() {
                let setup_path = global_info.get_air_setup_path(airgroup_id, air_id, setup_type);
                let setup = Setup::new(
                    &setup_path,
                    airgroup_id,
                    air_id,
                    &global_info.airs[airgroup_id][air_id],
                    setup_type,
                    verify_constraints,
                    gpu,
                    Some(&global_info.get_air_setup_path(airgroup_id, 0, &ProofType::Recursive2)),
                )?;
                if setup_type != &ProofType::Compressor || global_info.get_air_has_compressor(airgroup_id, air_id) {
                    let n = 1 << setup.stark_info.stark_struct.n_bits;
                    let n_bits_ext = setup.stark_info.stark_struct.n_bits_ext;
                    if max_const_tree_size < setup.const_tree_size {
                        max_const_tree_size = setup.const_tree_size;
                    }
                    if max_const_size < setup.const_pols_size {
                        max_const_size = setup.const_pols_size;
                    }
                    if max_prover_buffer_size < setup.prover_buffer_size {
                        max_prover_buffer_size = setup.prover_buffer_size;
                    }
                    prover_buffer_sizes.push(((airgroup_id, air_id), setup.prover_buffer_size as usize));
                    if max_prover_contributions_size < setup.contributions_size {
                        max_prover_contributions_size = setup.contributions_size;
                    }
                    if setup.gpu {
                        sized_airs.push((airgroup_id, air_id));
                        if !setup.verkey.is_empty() {
                            groups.entry(setup.get_vk()).or_insert((airgroup_id, air_id));
                        }
                    }
                    max_pinned_proof_size = max_pinned_proof_size.max(setup.pinned_proof_size);
                    max_n_bits_ext = max_n_bits_ext.max(n_bits_ext);
                    // Basic airs never land a compact witness (no exec map; their trace goes straight
                    // into cm1), so their repository reports 0 rather than a meaningless full width.
                    if setup_type != &ProofType::Basic {
                        max_compact_trace_size =
                            max_compact_trace_size.max((recursion_staging_cols(&setup, gpu) * n) as usize);
                    }
                }
                setups.insert((airgroup_id, air_id), setup);
                if setup_type == &ProofType::Recursive2 {
                    break;
                }
            }
        }

        // Second pass: the group owner is only settled once every member has been seen. Sizes
        // one slot per group, in the same order the loader assigns offsets -- must not drift.
        let mut fixed_groups: HashMap<(usize, usize), FixedGroup> = HashMap::new();
        let mut sized_slots: HashSet<(usize, usize)> = HashSet::new();
        let mut shared_airs = 0;
        let mut saved = 0;
        for air in sized_airs {
            let setup = &setups[&air];
            let group = match groups.get(&setup.get_vk()) {
                Some(&owner) => FixedGroup { owner },
                // Nothing to fingerprint with (no verkey): the air is its own group.
                None => FixedGroup { owner: air },
            };
            fixed_groups.insert(air, group);
            max_const_pols_size_packed = max_const_pols_size_packed.max(setup.const_pols_size_packed);
            if sized_slots.insert(group.owner) {
                n_const_slots += 1;
                total_const_pols_size += setup.const_pols_size_packed;
                // Custom commits ride the same buffer; the slot is filled later, when
                // register_custom_commits supplies the file path.
                total_const_pols_size += setup.custom_commits_reserved_words;
                total_custom_commits_reserved_words += setup.custom_commits_reserved_words;
            } else {
                shared_airs += 1;
                saved += setup.const_pols_size_packed;
                saved += setup.custom_commits_reserved_words;
            }
        }
        if shared_airs > 0 {
            let proof_type: &str = (*setup_type).into();
            tracing::info!(
                "Sharing GPU const pols: {shared_airs} {proof_type} airs reuse another air's fixed columns, \
                 saving {} MB",
                saved * std::mem::size_of::<u64>() / (1024 * 1024)
            );
        }

        prover_buffer_sizes.sort_by(|(ka, sa), (kb, sb)| sb.cmp(sa).then(ka.cmp(kb)));

        // Recursive1 const pols are not resident per setup: the loader carves a slot cache of
        // RECURSIVE1_CONST_SLOTS (or fewer when there are fewer setups) slots of the largest packed
        // set, filled at launch from host pinned copies (see DeviceCommitBuffers::constCache). The
        // custom commits of this repository would need resident slots; the recursion setups have
        // none, which is asserted here rather than assumed.
        let const_slot_cache_slots = if gpu && *setup_type == ProofType::Recursive1 && n_const_slots > 0 {
            assert!(
                total_custom_commits_reserved_words == 0,
                "recursive1 setups with custom commits cannot use the const slot cache"
            );
            RECURSIVE1_CONST_SLOTS.min(n_const_slots)
        } else {
            0
        };
        if const_slot_cache_slots > 0 {
            total_const_pols_size = const_slot_cache_slots * max_const_pols_size_packed;
        }

        Ok(Self {
            setups,
            fixed_groups,
            global_bin,
            global_info_file,
            max_const_tree_size,
            max_const_size,
            max_prover_contributions_size: max_prover_contributions_size as usize,
            max_prover_buffer_size: max_prover_buffer_size as usize,
            prover_buffer_sizes,
            max_pinned_proof_size: max_pinned_proof_size as usize,
            total_const_pols_size,
            total_custom_commits_reserved_words,
            max_compact_trace_size,
            max_n_bits_ext: max_n_bits_ext as usize,
            max_const_pols_size_packed,
            const_slot_cache_slots,
        })
    }
}

/// Air instance context for managing air instances (traces)
#[allow(dead_code)]
pub struct SetupCtx<F: PrimeField64> {
    setup_repository: SetupRepository<F>,
    pub max_const_tree_size: usize,
    pub max_const_size: usize,
    pub max_prover_contributions_size: usize,
    pub max_prover_buffer_size: usize,
    /// Per-air prover buffer size, largest first. See `SetupRepository::prover_buffer_sizes`.
    pub prover_buffer_sizes: Vec<((usize, usize), usize)>,
    pub max_pinned_proof_size: usize,
    pub max_compact_trace_size: usize,
    pub max_n_bits_ext: usize,
    pub total_const_pols_size: usize,
    /// Included in `total_const_pols_size`, and tracked separately because it stays resident even
    /// in no-const-buf mode: nothing stages custom commits per switch.
    pub total_custom_commits_reserved_words: usize,
    /// Largest packed const-pols set of the repository (elements): the slot size of the const cache.
    pub max_const_pols_size_packed: usize,
    /// Slots of the const slot cache this repository uses (recursive1 on GPU), 0 = resident slots.
    pub const_slot_cache_slots: usize,
    setup_type: ProofType,
}

/// Slots of the recursive1 const slot cache: a block uses ~20 distinct recursive1 setups, so 20
/// makes the second use of an air within a block a hit while freeing (44 - 20) x 100 MiB.
pub const RECURSIVE1_CONST_SLOTS: usize = 20;

impl<F: PrimeField64> SetupCtx<F> {
    pub fn new(
        global_info: &GlobalInfo,
        setup_type: &ProofType,
        verify_constraints: bool,
        gpu: bool,
    ) -> ProofmanResult<Self> {
        let setup_repository = SetupRepository::new(global_info, setup_type, verify_constraints, gpu)?;
        let max_const_tree_size = setup_repository.max_const_tree_size;
        let max_const_size = setup_repository.max_const_size;
        let max_prover_contributions_size = setup_repository.max_prover_contributions_size;
        let max_prover_buffer_size = setup_repository.max_prover_buffer_size;
        let prover_buffer_sizes = setup_repository.prover_buffer_sizes.clone();
        let max_pinned_proof_size = setup_repository.max_pinned_proof_size;
        let total_const_pols_size = setup_repository.total_const_pols_size;
        let total_custom_commits_reserved_words = setup_repository.total_custom_commits_reserved_words;
        let max_compact_trace_size = setup_repository.max_compact_trace_size;
        let max_n_bits_ext = setup_repository.max_n_bits_ext;
        let max_const_pols_size_packed = setup_repository.max_const_pols_size_packed;
        let const_slot_cache_slots = setup_repository.const_slot_cache_slots;
        Ok(SetupCtx {
            setup_repository,
            max_const_tree_size,
            max_const_size,
            max_prover_contributions_size,
            max_prover_buffer_size,
            prover_buffer_sizes,
            max_compact_trace_size,
            max_pinned_proof_size,
            max_n_bits_ext,
            total_const_pols_size,
            total_custom_commits_reserved_words,
            max_const_pols_size_packed,
            const_slot_cache_slots,
            setup_type: *setup_type,
        })
    }

    pub fn get_setup(&self, airgroup_id: usize, air_id: usize) -> ProofmanResult<&Setup<F>> {
        match self.setup_repository.setups.get(&(airgroup_id, air_id)) {
            Some(setup) => Ok(setup),
            None => Err(ProofmanError::InvalidSetup(format!(
                "Setup not found for airgroup_id: {}, air_id: {}",
                airgroup_id, air_id
            ))),
        }
    }

    /// `None` for airs never uploaded to the GPU (CPU setups, compressor-less airs).
    pub fn get_fixed_group(&self, airgroup_id: usize, air_id: usize) -> Option<FixedGroup> {
        self.setup_repository.fixed_groups.get(&(airgroup_id, air_id)).copied()
    }

    pub fn get_fixed(&self, airgroup_id: usize, air_id: usize) -> ProofmanResult<Vec<F>> {
        match self.setup_repository.setups.get(&(airgroup_id, air_id)) {
            Some(setup) => {
                let mut const_pols: Vec<F> = vec![F::ZERO; setup.const_pols_size];
                load_const_pols(setup, &mut const_pols);
                Ok(const_pols)
            }
            None => Err(ProofmanError::InvalidSetup(format!(
                "Setup not found for airgroup_id: {}, air_id: {}",
                airgroup_id, air_id
            ))),
        }
    }

    pub fn get_setups_list(&self) -> Vec<(usize, usize)> {
        self.setup_repository.setups.keys().cloned().collect()
    }

    /// `total_signal_no` (circom signalValues length, u64 elements) for every setup
    /// in this repository. Used to size the signalValues buffer pool.
    pub fn total_signal_nos(&self) -> Vec<usize> {
        self.setup_repository.setups.values().filter_map(|s| s.total_signal_no.map(|n| n as usize)).collect()
    }

    pub fn get_global_bin(&self) -> *mut c_void {
        self.setup_repository.global_bin.unwrap()
    }

    pub fn get_global_info_file(&self) -> String {
        self.setup_repository.global_info_file.clone()
    }
}

#[cfg(test)]
mod staging_tests {
    /// Every recursive air's staging width has to come from `recursion_staging_cols`, or one of them
    /// sizes the shared pool at its full width and the others' compaction buys nothing. The final
    /// airs were computed inline and did exactly that: 6 buffers at 1.07 GB instead of 0.075 GB.
    ///
    /// Grep rather than a call, because the defect is a call site that does the arithmetic ITSELF --
    /// it cannot be caught by testing the function everyone else already uses.
    #[test]
    fn no_call_site_sizes_the_recursive_pool_by_hand() {
        let src = include_str!("setup_ctx.rs");
        // Every `let` that sizes a compact trace must go through recursion_staging_cols and never
        // read cm1 directly: the SetupsVadcop max, the per-repository max, and the two finals'
        // tails the SetupsVadcop max is assembled from.
        for decl in ["let max_compact_trace_size", "let vadcop_final_tail", "let vadcop_final_compressed_tail"] {
            for body in src.split(decl).skip(1) {
                // Only real declarations (`let x = ...`), not this test's own string literals.
                if !body.trim_start().starts_with('=') {
                    continue;
                }
                let body = &body[..body.find(';').unwrap_or(body.len())];
                assert!(
                    !body.contains("map_sections_n"),
                    "{decl} reads cm1 directly; it must go through recursion_staging_cols:\n{body}"
                );
                let via_rule = body.contains("recursion_staging_cols");
                let via_tails = body.contains("vadcop_final_tail") || body.contains("max_compact_trace_size");
                assert!(via_rule || via_tails, "{decl} must use the shared rule:\n{body}");
            }
        }
    }
}