Dr.Jit: A Just-In-Time Compiler for Differentiable Rendering

1.3.1 verified Fri May 01 auth: no python

Dr.Jit is a just-in-time compiler for numerical computation in differentiable rendering, supporting CUDA, LLVM, and scalar backends. It provides automatic differentiation, vectorized operations, and dynamic control flow. Version 1.3.1 is current; releases are active every few months. Requires Python >=3.8.

pip install drjit

Common errors

error ImportError: cannot import name 'if_stmt' from 'drjit' ↓

cause Attempting to import if_stmt from drjit (it is available in drjit but the import path may be incorrect if using old code).

fix

Use 'from drjit import if_stmt' (top-level, no submodule).

error RuntimeError: Array types from different backends cannot be mixed ↓

cause Using a CUDA array in an LLVM context, or vice versa.

fix

Ensure all arrays in a single operation originate from the same backend (e.g., all drjit.cuda.* or all drjit.llvm.*).

error AttributeError: module 'drjit' has no attribute 'plot' ↓

cause The plot function was removed in v1.0.0.

fix

Use matplotlib: import numpy as np; np_arr = dr.numpy(array); plt.plot(np_arr).

Warnings

breaking v1.0.0 introduced a new nanobind-based API, breaking backward compatibility with v0.x. The 'drjit.dynamic' module is removed; use top-level functions like dr.if_stmt, dr.while_loop, dr.cond. ↓

fix Update imports: replace 'from drjit.dynamic import if_stmt' with 'from drjit import if_stmt'.

deprecated dr.JitFlag.VCallRecord is deprecated; its use is discouraged and may be removed in future versions. ↓

fix Set dr.set_flag(dr.JitFlag.VCallRecord, False) to avoid warnings, or omit if not needed.

gotcha Mixing backends (e.g., CUDA arrays with LLVM operations) causes runtime errors. Dr.Jit arrays are backend-specific and non-interoperable. ↓

fix Always use types from a single backend: e.g., drjit.cuda.Float3 or drjit.llvm.Float3, not both.

deprecated drjit.plot function has been removed in v1.0.0. Use matplotlib or other plotting libraries for visualization. ↓

fix Export array to numpy with dr.numpy() and plot with matplotlib.

Imports

drjit
wrong
```
from drjit import *
```
correct
```
import drjit as dr
```
Wildcard imports are discouraged; use the 'dr' alias.
drjit.llvm.Array3f
wrong
```
from drjit.cuda import Array3f as Float3
```
correct
```
from drjit.llvm import Array3f as Float3
```
Mixing backends will cause runtime errors. Use drjit.llvm for CPU, drjit.cuda for GPU.
drjit.scatter_cas
wrong
```
import drjit; drjit.scatter_cas()
```
correct
```
from drjit import scatter_cas
```
Attribute is available at module level; using dr directly is fine, but ensure function name matches.
drjit.if_stmt
wrong
```
from drjit.dynamic import if_stmt
```
correct
```
from drjit import if_stmt, cond
```
In v1.0+, if_stmt and cond are top-level; the dynamic module is deprecated.

Quickstart

Compile a simple loop with Dr.Jit's syntax transform.

import drjit as dr

dr.set_flag(dr.JitFlag.VCallRecord, False)  # disable for simple loops

@dr.syntax
def f(x):
    y = dr.zeros(dr.int32, 4)
    for i in range(4):
        y[i] = x[i] + 1
    return y

x = dr.arange(dr.int32, 4)
result = f(x)
print(result)