openvm_rv32im_circuit/base_alu/
cuda.rs1use std::{mem::size_of, sync::Arc};
2
3use derive_new::new;
4use openvm_circuit::{arch::DenseRecordArena, utils::next_power_of_two_or_zero};
5use openvm_circuit_primitives::{
6 bitwise_op_lookup::BitwiseOperationLookupChipGPU, var_range::VariableRangeCheckerChipGPU, Chip,
7};
8use openvm_cuda_backend::{base::DeviceMatrix, prelude::F, GpuBackend};
9use openvm_cuda_common::copy::MemCopyH2D;
10use openvm_stark_backend::prover::AirProvingContext;
11
12use crate::{
13 adapters::{
14 Rv32BaseAluAdapterCols, Rv32BaseAluAdapterRecord, RV32_CELL_BITS, RV32_REGISTER_NUM_LIMBS,
15 },
16 cuda_abi::alu_cuda::tracegen,
17 BaseAluCoreCols, BaseAluCoreRecord,
18};
19
20#[derive(new)]
21pub struct Rv32BaseAluChipGpu {
22 pub range_checker: Arc<VariableRangeCheckerChipGPU>,
23 pub bitwise_lookup: Arc<BitwiseOperationLookupChipGPU<RV32_CELL_BITS>>,
24 pub timestamp_max_bits: usize,
25}
26
27impl Chip<DenseRecordArena, GpuBackend> for Rv32BaseAluChipGpu {
28 fn generate_proving_ctx(&self, arena: DenseRecordArena) -> AirProvingContext<GpuBackend> {
29 const RECORD_SIZE: usize = size_of::<(
30 Rv32BaseAluAdapterRecord,
31 BaseAluCoreRecord<RV32_REGISTER_NUM_LIMBS>,
32 )>();
33 let records = arena.allocated();
34 if records.is_empty() {
35 return AirProvingContext::simple_no_pis(DeviceMatrix::dummy());
36 }
37 debug_assert_eq!(records.len() % RECORD_SIZE, 0);
38
39 let trace_width = BaseAluCoreCols::<F, RV32_REGISTER_NUM_LIMBS, RV32_CELL_BITS>::width()
40 + Rv32BaseAluAdapterCols::<F>::width();
41 let trace_height = next_power_of_two_or_zero(records.len() / RECORD_SIZE);
42 let device_ctx = &self.range_checker.device_ctx;
43
44 let d_records = tracing::info_span!("trace_gen.h2d_records")
45 .in_scope(|| records.to_device_on(device_ctx))
46 .unwrap();
47 let d_trace = DeviceMatrix::<F>::with_capacity_on(trace_height, trace_width, device_ctx);
48
49 unsafe {
50 tracegen(
51 d_trace.buffer(),
52 trace_height,
53 &d_records,
54 &self.range_checker.count,
55 self.range_checker.count.len(),
56 &self.bitwise_lookup.count,
57 RV32_CELL_BITS,
58 self.timestamp_max_bits as u32,
59 device_ctx.stream.as_raw(),
60 )
61 .unwrap();
62 }
63
64 AirProvingContext::simple_no_pis(d_trace)
65 }
66}