EGL surface switch loses draws with EXT_multisampled_render_to_texture on T4

Hey all - apologies for the report below - it was prepared by an LLM. I’ve reviewed it, but I am not an expert in graphics, this was a result of a reproduction of flashes found when using threejs in Chrome on EC2.

A standalone EGL/GLES program loses a draw after switching the same EGLContext to another pbuffer and back. Rebinding the already-bound FBO fixes it.

Environment

  • Tesla T4 (EC2 g4dn.2xlarge)
  • NVIDIA 615.71.09 and 615.78.08, proprietary kernel module
  • Amazon Linux 2023.12.20261102; kernel 6.1.189-236.400.amzn2023.x86_64

Reproduce

Extract repro.c from the attached ZIP.

cc -std=c11 -O2 -Wall -Wextra repro.c -lEGL -lGLESv2 -o repro
./repro --output bad.ppm
./repro --rebind --output good.ppm

Expected: red rectangle on the left, green on the right.
Actual: the right rectangle is missing, leaving the white background.

The sequence is:

  1. Bind an FBO with an RGBA8 texture attached using glFramebufferTexture2DMultisampleEXT (4 samples).
  2. Clear white, draw red, then glFlush().
  3. Call eglMakeCurrent with pbuffer B, then pbuffer A, using the same context both times.
  4. Update uniforms, draw green, and glReadPixels() from the still-bound FBO.

No FBO is deleted or rebound in this sequence. Both pbuffers remain valid; EGL calls succeed and no GL errors occur. There is no depth/stencil attachment or test.

Results

Same T4 and compiled binary before/after the driver upgrade:

Variant 615.71.09 615.78.08
Original 300/300 bad 300/300 bad
--rebind 0/300 bad 0/300 bad
--no-msaa 0/300 bad 0/300 bad
--no-switch 0/300 bad 0/300 bad

--rebind adds only glBindFramebuffer(GL_FRAMEBUFFER, sameFBO) after step 3. Using glUniform4fv instead of glProgramUniform4fvEXT also reproduces the failure.

The program checks pixels over 300 iterations and exits 0 for correct output, 1 for incorrect output, or 2 for setup/API errors.

Is this a known driver issue, and is there a fixed release?

Attachments

Source, comparison image, and test results:
nvidia-egl-repro-report.zip (9.4 KB)

Diagnostic collected on a fresh T4 with 615.71.09 immediately after reproducing the failure; nvidia-bug-report.sh exited 0:
nvidia-bug-report.log.gz (427.8 KB)