Lwjgl3: Multiple calls to clCompileProgram/clLinkProgram cause crash

[kurtzmarc]

Environment

• LWJGL version: 3.2.1
• LWJGL build #: release
• Java version: 1.8
• Platform: Windows x64
• Module: opencl

Description

We are using the OpenCL bindings to build and run OpenCL kernels. Instead of using the one-step clBuildProgram() call, we are using the two-step process where we call clCompileProgram() followed by clLinkProgram() to build our kernel. If we do this twice in a row, we get a crash on the second clLinkProgram call. I am attaching an isolated test case and a crash log. One very odd feature of this bug is that it seems to be affected by the Java stack. If I add random int variables, the crash will trigger/not-trigger. So you may need to add/remove random int variables depending on your system. I have tried it on four different OpenCL devices across two different Windows systems and found this bug to occur on all.

package org.test;

import org.junit.Test;
import org.lwjgl.PointerBuffer;
import org.lwjgl.system.MemoryStack;

import java.nio.IntBuffer;

import static org.lwjgl.opencl.CL10.*;
import static org.lwjgl.opencl.CL12.clCompileProgram;
import static org.lwjgl.opencl.CL12.clLinkProgram;
import static org.lwjgl.system.MemoryStack.stackPush;

public class CompileLinkBugTest
{
    @Test
    public void compileLinkBug()
    {
        try (MemoryStack stack = stackPush()) {
            PointerBuffer platforms = stack.mallocPointer(1);
            clGetPlatformIDs(platforms, (int[]) null);
            long platformPtr = platforms.get(0);

            PointerBuffer devices = stack.mallocPointer(1);
            clGetDeviceIDs(platformPtr, CL_DEVICE_TYPE_ALL, devices, (int[]) null);
            long devicePtr = devices.get(0);

            IntBuffer errcode_ret = stack.callocInt(1);

            PointerBuffer propeties = null;
            long contextPtr = clCreateContext(propeties, devicePtr, null, 0, errcode_ret);

            createProgramPtr(contextPtr, devicePtr, "__kernel void a() { return; }");
            createProgramPtr(contextPtr, devicePtr, "__kernel void b() { return; }");
        }
    }

    private long createProgramPtr(long contextPtr, long devicePtr, String src)
    {
        try (MemoryStack stack = stackPush()) {
            //int a = 0;  // NOTE - you may need to uncomment this to trigger crash
            if (contextPtr == 0) {
                throw new RuntimeException("Device is not loaded.");
            }
            long ptr = clCreateProgramWithSource(contextPtr, src, null);

            PointerBuffer deviceList = stack.pointers(devicePtr);

            clCompileProgram(ptr, deviceList, "-cl-std=CL2.0", null, null, null, 0);
            ptr = clLinkProgram(contextPtr, deviceList, "", ptr, null, 0);
            if (ptr == 0) {
                throw new RuntimeException("CL Linking failed\n");
            }
            return ptr;
        }
    }
}

hs_err_pid17432.log

Tested on:

Name: AMD Accelerated Parallel Processing 
Vendor: Advanced Micro Devices, Inc. 
Version: OpenCL 2.1 AMD-APP (2906.10) 
     Name: Ellesmere 
     Vendor: Advanced Micro Devices, Inc. 
     Version: OpenCL 2.0 AMD-APP (2906.10) 
     Device type: GPU

Name: NVIDIA CUDA 
Vendor: NVIDIA Corporation 
Version: OpenCL 1.2 CUDA 10.2.108 
     Name: Quadro M2000M 
     Vendor: NVIDIA Corporation 
     Version: OpenCL 1.2 CUDA 
     Device type: GPU

Name: Intel(R) OpenCL 
Vendor: Intel(R) Corporation 
Version: OpenCL 2.0  
     Name: Intel(R) HD Graphics P530 
     Vendor: Intel(R) Corporation 
     Version: OpenCL 2.0  
     Device type: GPU

     Name: Intel(R) Xeon(R) CPU E3-1535M v5 @ 2.90GHz 
     Vendor: Intel(R) Corporation 
     Version: OpenCL 2.0 (Build 10) 
     Device type: CPU

[Spasi]

Thank you for reporting this!

[kurtzmarc]

Thanks for the fix! Is there a date set for the next release?
Also does this same problem exist with clCreateSubDevices?

[httpdigest]

When I look at the Khronos specification for clCreateSubDevices and then at the LWJGL3 method, I don't see any differences. Both signatures match.

[kurtzmarc]

You are correct. This appears to be a discrepancy just within the Khronos specs which mention:

Otherwise, it returns a NULL value with the following error values returned in errcode_ret

However there is no errcode_ret parameter or way to get an error code from clCreateSubDevices as far as I can see. It followed a similar pattern to what I saw with clBuildProgram so I wrongly assumed it had the same issue.

[httpdigest]

Oh indeed! I didn't notice that, thanks for pointing out. That actually is a bug in their documentation.

[mikailgedik]

Has the issue been resolved? This one's critical for me since this bug is a death note for my program

[Spasi]

Yes, the fix is available in the 3.2.4 snapshot.