- [🐛 bug report] LoopRecord

[ziyi-zhang]

Summary

Simple rejection sampling loop does not work.

System configuration

• Platform: Ubuntu2004
• Compiler: clang12
• Python version: 3.8.8
• Mitsuba 3 version: 9a588fab5405cf7e7adf9536643795de5f361735 (latest)
• Compiled variants:
  • llvm_ad_rgb
  • cuda_ad_rgb

Description

When LoopRecord is disabled, both LLVM and CUDA work. When enabled, llvm complains

Critical Dr.Jit compiler failure: jit_llvm_compile(): parsing failed. drjit_561bf61d2e73fe582c243e53d5d7a37c:130:30: error: use of undefined value '%f93_final' %f93 = phi <8 x float> [ %f93_final, %l_90_tail ], [ %f89, %l_90_start ]

Cuda can run without warning, but the output dir_world is empty (while count is correct).

Steps to reproduce

import drjit as dr
import mitsuba as mi
mi.set_variant('llvm_ad_rgb')

N_sample = 10
sampler = mi.load_dict({"type": "independent"})
sampler.seed(7, N_sample)
active = mi.Bool(True)
count = mi.UInt32(0)
dir_world = dr.zero(mi.Vector3f)

with dr.scoped_set_flag(dr.JitFlag.LoopRecord, True):
    loop = mi.Loop("Trapezoid_effe_rejection_sampling", lambda: (active, dir_world, count, sampler))
    while loop(active):
        sample = sampler.next_2d(active)
        r = sample.x
        theta = dr.lerp(0.0, 3.0, sample.y)

        s, c = dr.sincos(theta)
        dir_world = mi.Vector3f(r * c, r * s, dr.safe_sqrt(dr.fnmadd(r, r, 1.0)))
        count += mi.UInt32(1)
        active &= (dir_world.x < 0.2) & (count < 1000)

print(count)
print(dir_world)