[Bug] [TIRx] VectorizeLoop loses a let binding when scalarizing a conditional

Abierto
  • ¿Algo no está bien?
#20,880 0 comentarios 0 reacciones 0 asignados Ver en GitHub

Los mantenedores suelen responder en 1 día

Nadie ha tomado este issue todavía.

Evaluación

Dificultad
4/5
Tiempo estimado
1-2 días
Aptitud para principiantes
25/100
Tipo de issue
Error
Claridad
Bien especificado
Estado de actividad
Activo
Stack tecnológico
python
Área
compilers

Línea de trabajo

The failure is an assertion in src/tirx/transform/make_packed_api.cc at line 279, reached after tirx.transform.VectorizeLoop() runs. Start by running the standalone script with T.serial(4) and with T.vectorized(4), then apply only VectorizeLoop to see the lost let binding. Done means the vectorized loop compiles and the three input checks pass, with a regression test for the scalarized conditional.

Escrito por el modelo de indexación a partir del texto del issue.

Descripción

needs-triage type: bug
Expected behavior

The vectorized loop below should compile and produce the same result as its serial equivalent. The first two elements should become A[i] + 3, and the last two should become A[i] + 4.

The input passes both tvm.tirx.analysis.verify_ssa and verify_well_formed.

Actual behavior

Compiling the vectorized version with the default LLVM pipeline fails in MakePackedAPI:

src/tirx/transform/make_packed_api.cc:279
TVM_FFI_ICHECK_EQ(undefined.size(), 0)

tvm.error.InternalError: Check failed: undefined.size() == 0 (1 vs. 0) :
In PrimFunc main variables (v,) are used, but are not passed in as API arguments

Applying tirx.transform.VectorizeLoop() in isolation shows the lost binding. The vector value remains bound outside the scalar fallback loop, but the conditional stores refer to a different, unbound scalar variable:

v: T.let[T.float32x4] = A[T.Ramp(0, 1, 4)] + T.Broadcast(T.float32(1.0), 4)
for i_s in range(4):
    v_1 = T.float32()  # Declaration only; no lane value is assigned.
    if i_s < 2:
        A[i_s] = v_1 + T.float32(2.0)
    else:
        A[i_s] = v_1 + T.float32(3.0)

The standalone script below failed with the same undefined-variable assertion in three fresh processes. Replacing only T.vectorized(4) with T.serial(4) compiled successfully in three fresh processes; each serial process checked three input arrays against the expected result.

No custom pipeline or disabled passes are needed.

Environment
  • Apache TVM 0.26.dev0+source.8f328e8 (source snapshot; no .git directory)
  • Ubuntu 22.04.5, Linux 6.8.0-52-generic, x86-64
  • Python 3.11.16
  • LLVM 15.0.7
  • Target llvm, opt_level=2
  • CPU-only reproducer; no GPU is required
Steps to reproduce

Save the following script as repro_vectorize_let.py and run:

python repro_vectorize_let.py

For the passing control, change T.vectorized(4) to T.serial(4) and run the script again.

import numpy as np
import tvm
from tvm.script import ir as I, tirx as T


@I.ir_module
class Module:
    @T.prim_func(s_tir=True)
    def main(A: T.Buffer((4,), "float32")):
        for i in T.vectorized(4):
            v: T.let[T.float32] = A[i] + T.float32(1)
            if i < 2:
                A[i] = v + T.float32(2)
            else:
                A[i] = v + T.float32(3)


assert tvm.tirx.analysis.verify_ssa(Module["main"])
assert tvm.tirx.analysis.verify_well_formed(Module["main"])
print("input validation passed", flush=True)

with tvm.transform.PassContext(opt_level=2):
    executable = tvm.compile(Module, target="llvm")
print("compiled", flush=True)

for seed in [42, 9001, 20261101]:
    rng = np.random.default_rng(seed)
    x = rng.integers(-16, 17, size=4).astype("float32") / 64
    a = tvm.runtime.tensor(x.copy(), tvm.cpu(0))
    executable["main"](a)
    expected = x + np.array([3, 3, 4, 4], dtype="float32")
    np.testing.assert_array_equal(a.numpy(), expected)
print("three inputs passed", flush=True)
Triage
  • needs-triage
  • type: bug
Lenguaje dominante
Python
Estrellas
13.9k
Forks
4k
Merge medio
7 h 59 min
PR fusionados (30 d)
159

Preparar el entorno

Este proyecto no incluye contenedor de desarrollo, Dockerfile ni guía de contribución, así que la configuración corre por tu cuenta: empieza por su README y consulta nuestra guía para la primera contribución para los pasos generales.

Primeros pasos

  1. Lee el issue completo y luego la guía de contribución del proyecto.
  2. Comenta en el issue que vas a ocuparte — evita que dos personas hagan lo mismo.
  3. Haz un fork del repositorio y trabaja en una rama.
  4. Abre un pull request que haga referencia al número del issue.

Más de apache/tvm

Todos los issues de apache/tvm

Issues similares

Más issues de Python

Recibe los nuevos issues en tu correo

Un resumen breve de issues de GitHub para principiantes.