[BUG]: cudaGraphicsGLRegisterImage returns null handle with cuda-bindings 12.9.3

[pijyoi]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct

Type of Bug

Silent Failure

Component

cuda.bindings

Describe the bug

On cuda-bindings 12.9.3, cudaGraphicsGLRegisterImage reports cudaSuccess but returns a null handle.
This does not happen with cuda-bindings 12.9.2

How to Reproduce

PySide6 is used to create the OpenGL context and texture.

from cuda.bindings import runtime as cudart
from PySide6 import QtGui, QtOpenGL

class GLWidget(QtOpenGL.QOpenGLWindow):
    def initializeGL(self):
        self.m_texture = QtOpenGL.QOpenGLTexture(QtOpenGL.QOpenGLTexture.Target.Target2D)
        self.m_texture.setFormat(QtOpenGL.QOpenGLTexture.TextureFormat.RGBA8_UNorm)
        self.m_texture.setSize(512, 512)
        self.m_texture.allocateStorage()

        err, self.gfx_resource = cudart.cudaGraphicsGLRegisterImage(
            self.m_texture.textureId(), self.m_texture.target().value,
            cudart.cudaGraphicsRegisterFlags.cudaGraphicsRegisterFlagsWriteDiscard
        )
        error_name = cudart.cudaGetErrorName(err)[1].decode()
        print(error_name, self.gfx_resource)

app = QtGui.QGuiApplication([])
win = GLWidget()
win.show()

# 12.9.2: cudaSuccess <cudaGraphicsResource_t 0x1b75dca8e90>
# 12.9.3: cudaSuccess <cudaGraphicsResource_t 0x0>

Expected behavior

cudaGraphicsGLRegisterImage should return a valid handle upon cudaSuccess

Operating System

No response

nvidia-smi output

No response

[leofang]

Thanks for reporting, @pijyoi!

@mdboom It seems all graphics APIs were commented out in #914 and not forwarding to cudart, e.g.

cuda-python/cuda_bindings/cuda/bindings/_lib/cyruntime/cyruntime.pxi

Lines 100 to 107 in 21db67c

	cdef cudaError_t _cudaGraphicsGLRegisterImage(cudaGraphicsResource** resource, cyruntime.GLuint image, cyruntime.GLenum target, unsigned int flags) except ?cudaErrorCallRequiresNewerDriver nogil:
	cdef cudaError_t err = cudaSuccess
	# cudaFree(0) is a NOP operations that initializes the context state
	err = cudaFree(<void*>0)
	if err != cudaSuccess:
	return err
	# err = <cudaError_t>cydriver._cuGraphicsGLRegisterImage(<cydriver.CUgraphicsResource*>resource, image, target, flags)
	return err

Could you take a look?

[pijyoi]

A lighter-weight (wheel-size wise) MWE that may be better for CI.
PySide6-Essentials weighs in at 75MB whereas pyglet is only 1MB.

from cuda.bindings import runtime as cudart
import pyglet

tex = pyglet.image.Texture.create(512, 512)

err, gfx_resource = cudart.cudaGraphicsGLRegisterImage(
    tex.id, tex.target,
    cudart.cudaGraphicsRegisterFlags.cudaGraphicsRegisterFlagsWriteDiscard
)
error_name = cudart.cudaGetErrorName(err)[1].decode()
print(error_name, gfx_resource)