openvm_algebra_circuit/modular_chip/cuda/
addsub.rs

1use std::sync::Arc;
2
3use derive_new::new;
4use openvm_algebra_transpiler::Rv32ModularArithmeticOpcode;
5use openvm_circuit::arch::{AdapterCoreLayout, DenseRecordArena, RecordSeeker};
6use openvm_circuit_primitives::{
7    bitwise_op_lookup::BitwiseOperationLookupChipGPU, var_range::VariableRangeCheckerChipGPU,
8};
9use openvm_cuda_backend::{chip::get_empty_air_proving_ctx, prover_backend::GpuBackend, types::F};
10use openvm_cuda_common::copy::MemCopyH2D;
11use openvm_instructions::riscv::RV32_CELL_BITS;
12use openvm_mod_circuit_builder::{
13    ExprBuilderConfig, FieldExpressionChipGPU, FieldExpressionCoreAir, FieldExpressionMetadata,
14};
15use openvm_rv32_adapters::{Rv32VecHeapAdapterCols, Rv32VecHeapAdapterExecutor};
16use openvm_stark_backend::{prover::types::AirProvingContext, Chip};
17
18use crate::{modular_chip::addsub_expr, AlgebraRecord};
19
20#[derive(new)]
21pub struct ModularAddSubChipGpu<const BLOCKS: usize, const BLOCK_SIZE: usize> {
22    pub range_checker: Arc<VariableRangeCheckerChipGPU>,
23    pub bitwise_lookup: Arc<BitwiseOperationLookupChipGPU<RV32_CELL_BITS>>,
24    pub config: ExprBuilderConfig,
25    pub offset: usize,
26    pub pointer_max_bits: u32,
27    pub timestamp_max_bits: u32,
28}
29
30impl<const BLOCKS: usize, const BLOCK_SIZE: usize> Chip<DenseRecordArena, GpuBackend>
31    for ModularAddSubChipGpu<BLOCKS, BLOCK_SIZE>
32{
33    fn generate_proving_ctx(&self, arena: DenseRecordArena) -> AirProvingContext<GpuBackend> {
34        let range_bus = self.range_checker.cpu_chip.as_ref().unwrap().bus();
35        let (expr, is_add_flag, is_sub_flag) = addsub_expr(self.config.clone(), range_bus);
36
37        let total_input_limbs = expr.builder.num_input * expr.canonical_num_limbs();
38        let layout = AdapterCoreLayout::with_metadata(FieldExpressionMetadata::<
39            F,
40            Rv32VecHeapAdapterExecutor<2, BLOCKS, BLOCKS, BLOCK_SIZE, BLOCK_SIZE>,
41        >::new(total_input_limbs));
42
43        let record_size = RecordSeeker::<
44            DenseRecordArena,
45            AlgebraRecord<2, BLOCKS, BLOCK_SIZE>,
46            _,
47        >::get_aligned_record_size(&layout);
48
49        let records = arena.allocated();
50        if records.is_empty() {
51            return get_empty_air_proving_ctx::<GpuBackend>();
52        }
53        debug_assert_eq!(records.len() % record_size, 0);
54
55        let num_records = records.len() / record_size;
56
57        let local_opcode_idx = vec![
58            Rv32ModularArithmeticOpcode::ADD as usize,
59            Rv32ModularArithmeticOpcode::SUB as usize,
60            Rv32ModularArithmeticOpcode::SETUP_ADDSUB as usize,
61        ];
62        let opcode_flag_idx = vec![is_add_flag, is_sub_flag];
63
64        let air = FieldExpressionCoreAir::new(expr, self.offset, local_opcode_idx, opcode_flag_idx);
65
66        let adapter_width =
67            Rv32VecHeapAdapterCols::<F, 2, BLOCKS, BLOCKS, BLOCK_SIZE, BLOCK_SIZE>::width();
68
69        let d_records = records.to_device().unwrap();
70
71        let field_expr_chip = FieldExpressionChipGPU::new(
72            air,
73            d_records,
74            num_records,
75            record_size,
76            adapter_width,
77            BLOCKS,
78            self.range_checker.clone(),
79            self.bitwise_lookup.clone(),
80            self.pointer_max_bits,
81            self.timestamp_max_bits,
82        );
83
84        let d_trace = field_expr_chip.generate_field_trace();
85
86        AirProvingContext::simple_no_pis(d_trace)
87    }
88}
openvm_algebra_circuit/modular_chip/cuda/addsub.rs

openvm_algebra_circuit/modular_chip/cuda/
addsub.rs
