[Issue]: PoCL - OpenVX Conformance failures

[kiritigowda]

Problem Description

PoCL is a portable open source (MIT-licensed) implementation of the OpenCL standard. In addition to being an easily portable/layered multi-device open-source OpenCL implementation, a major goal of this project is improving interoperability of diversity of OpenCL-capable devices by integrating them to a single centrally orchestrated platform. Also one of the key goals is to enhance performance portability of OpenCL programs across device types utilizing runtime and compiler techniques.

PoCL Conformance failures

• CPU - Intel CPU
• iGPU - Intel GPU OpenCL driver

[ ======== ]
[ ALL DONE ] 5820 test(s) from 69 test case(s) ran
[ PASSED   ] 5807 test(s)
[ FAILED   ] 13 test(s), listed below:
[ FAILED   ] SmokeTestBase.vxReleaseReferenceBase
[ FAILED   ] SmokeTestBase.vxLoadKernels
[ FAILED   ] SmokeTestBase.vxUnloadKernels
[ FAILED   ] vxConvertDepth.BitExact/6/WRAP/15x17 U8->S16
[ FAILED   ] vxConvertDepth.BitExact/15/WRAP/1x1 S16->U8
[ FAILED   ] GaussianPyramid.GraphProcessing/1/randomInput/VX_BORDER_UNDEFINED/sz=16x16/VX_SCALE_PYRAMID_ORB
[ FAILED   ] GaussianPyramid.GraphProcessing/3/randomInput/VX_BORDER_UNDEFINED/sz=256x256/VX_SCALE_PYRAMID_ORB
[ FAILED   ] GaussianPyramid.GraphProcessing/5/randomInput/VX_BORDER_UNDEFINED/sz=640x480/VX_SCALE_PYRAMID_ORB
[ FAILED   ] GaussianPyramid.GraphProcessing/37/lena/VX_BORDER_UNDEFINED/VX_SCALE_PYRAMID_ORB
[ FAILED   ] GaussianPyramid.ImmediateProcessing/1/randomInput/VX_BORDER_UNDEFINED/sz=16x16/VX_SCALE_PYRAMID_ORB
[ FAILED   ] GaussianPyramid.ImmediateProcessing/3/randomInput/VX_BORDER_UNDEFINED/sz=256x256/VX_SCALE_PYRAMID_ORB
[ FAILED   ] GaussianPyramid.ImmediateProcessing/5/randomInput/VX_BORDER_UNDEFINED/sz=640x480/VX_SCALE_PYRAMID_ORB
[ FAILED   ] GaussianPyramid.ImmediateProcessing/37/lena/VX_BORDER_UNDEFINED/VX_SCALE_PYRAMID_ORB
[ DISABLED ] 8190 test(s)

Operating System

ANY

CPU

Intel CPU

GPU

Other

Other

Intel iGPU

ROCm Version

ROCm 6.0.0

ROCm Component

MIVisionX

Steps to Reproduce

No response

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response

[kiritigowda]

@pjaaskel @karolherbst - capturing the current TOT develop issues - Can we reproduce these in PoCL CI?

[pjaaskel]

Added pocl/pocl#1959.

Also please note that also Intel's proprietary CPU driver and GPU driver produce (some of) these issues, thus it's not PoCL only. @karolherbst, what was the end result of the CTS run via Rusticl?

[karolherbst]

@karolherbst, what was the end result of the CTS run via Rusticl?

There is a virtual memory leak (some driver internal reason) and the run aborted at some point, but no fails up to that point.. Need to figure out where it's leaking, but I suspect queue recreation is partly responsible for this.

[karolherbst]

Okay.. I figured out what's going on. Something is leaking cl_event objects, which then keeps the cl_queue object alive internally, which then keeps GPU contexts alive, and we do have some VRAM allocations per context...

Not entirely sure if it's a rusticl internal bug or not, but it's quite unlikely with event objects. I'll try to dig a bit deeper, but it could be that something isn't calling clReleaseEvent often enough.

[karolherbst]

Yeah, I confirmed that there aren't enough clReleaseEvent calls to clean up events returned to the applications (ignoring calls where the pointer us NULL to return them). So it's probably something for you to figure out.

I'll try to mitigate this problem, but so far it seems like an issue inside MIVisionX. Should this be tracked in a new issue?

[karolherbst]

Anyway, with that worked around:

[ ======== ]
[ ALL DONE ] 5820 test(s) from 69 test case(s) ran
[ PASSED   ] 5761 test(s)
[ FAILED   ] 59 test(s), listed below:
[ FAILED   ] SmokeTestBase.vxReleaseReferenceBase
[ FAILED   ] SmokeTestBase.vxLoadKernels
[ FAILED   ] SmokeTestBase.vxUnloadKernels
[ FAILED   ] vxuMultiply.Fuzzy/326/WRAP/TO_ZERO 640x480 S16*S16*(1/2^0)=S16
[ FAILED   ] vxuMultiply.Fuzzy/327/WRAP/TO_ZERO 640x480 S16*S16*(1/2^1)=S16
[ FAILED   ] vxuMultiply.Fuzzy/328/WRAP/TO_ZERO 640x480 S16*S16*(1/2^2)=S16
[ FAILED   ] vxuMultiply.Fuzzy/329/WRAP/TO_ZERO 640x480 S16*S16*(1/2^3)=S16
[ FAILED   ] vxuMultiply.Fuzzy/330/WRAP/TO_ZERO 640x480 S16*S16*(1/2^4)=S16
[ FAILED   ] vxuMultiply.Fuzzy/331/WRAP/TO_ZERO 640x480 S16*S16*(1/2^5)=S16
[ FAILED   ] vxuMultiply.Fuzzy/332/WRAP/TO_ZERO 640x480 S16*S16*(1/2^6)=S16
[ FAILED   ] vxuMultiply.Fuzzy/333/WRAP/TO_ZERO 640x480 S16*S16*(1/2^7)=S16
[ FAILED   ] vxuMultiply.Fuzzy/334/WRAP/TO_ZERO 640x480 S16*S16*(1/2^8)=S16
[ FAILED   ] vxuMultiply.Fuzzy/335/WRAP/TO_ZERO 640x480 S16*S16*(1/2^9)=S16
[ FAILED   ] vxuMultiply.Fuzzy/336/WRAP/TO_ZERO 640x480 S16*S16*(1/2^10)=S16
[ FAILED   ] vxuMultiply.Fuzzy/337/WRAP/TO_ZERO 640x480 S16*S16*(1/2^11)=S16
[ FAILED   ] vxuMultiply.Fuzzy/338/WRAP/TO_ZERO 640x480 S16*S16*(1/2^12)=S16
[ FAILED   ] vxuMultiply.Fuzzy/339/WRAP/TO_ZERO 640x480 S16*S16*(1/2^13)=S16
[ FAILED   ] vxuMultiply.Fuzzy/340/WRAP/TO_ZERO 640x480 S16*S16*(1/2^14)=S16
[ FAILED   ] vxuMultiply.Fuzzy/341/WRAP/TO_ZERO 640x480 S16*S16*(1/2^15)=S16
[ FAILED   ] vxMultiply.Inference/69/WRAP/TO_ZERO S16*S16*(1/2^0)=S16
[ FAILED   ] vxMultiply.Inference/70/WRAP/TO_ZERO S16*S16*(1/2^1)=S16
[ FAILED   ] vxMultiply.Inference/71/WRAP/TO_ZERO S16*S16*(1/2^2)=S16
[ FAILED   ] vxMultiply.Inference/72/WRAP/TO_ZERO S16*S16*(1/2^3)=S16
[ FAILED   ] vxMultiply.Inference/73/WRAP/TO_ZERO S16*S16*(1/2^4)=S16
[ FAILED   ] vxMultiply.Inference/74/WRAP/TO_ZERO S16*S16*(1/2^5)=S16
[ FAILED   ] vxMultiply.Inference/75/WRAP/TO_ZERO S16*S16*(1/2^6)=S16
[ FAILED   ] vxMultiply.Inference/76/WRAP/TO_ZERO S16*S16*(1/2^7)=S16
[ FAILED   ] vxMultiply.Inference/77/WRAP/TO_ZERO S16*S16*(1/2^8)=S16
[ FAILED   ] vxMultiply.Inference/78/WRAP/TO_ZERO S16*S16*(1/2^9)=S16
[ FAILED   ] vxMultiply.Inference/79/WRAP/TO_ZERO S16*S16*(1/2^10)=S16
[ FAILED   ] vxMultiply.Inference/80/WRAP/TO_ZERO S16*S16*(1/2^11)=S16
[ FAILED   ] vxMultiply.Inference/81/WRAP/TO_ZERO S16*S16*(1/2^12)=S16
[ FAILED   ] vxMultiply.Inference/82/WRAP/TO_ZERO S16*S16*(1/2^13)=S16
[ FAILED   ] vxMultiply.Inference/83/WRAP/TO_ZERO S16*S16*(1/2^14)=S16
[ FAILED   ] vxMultiply.Inference/84/WRAP/TO_ZERO S16*S16*(1/2^15)=S16
[ FAILED   ] vxMultiply.Fuzzy/326/WRAP/TO_ZERO 640x480 S16*S16*(1/2^0)=S16
[ FAILED   ] vxMultiply.Fuzzy/327/WRAP/TO_ZERO 640x480 S16*S16*(1/2^1)=S16
[ FAILED   ] vxMultiply.Fuzzy/328/WRAP/TO_ZERO 640x480 S16*S16*(1/2^2)=S16
[ FAILED   ] vxMultiply.Fuzzy/329/WRAP/TO_ZERO 640x480 S16*S16*(1/2^3)=S16
[ FAILED   ] vxMultiply.Fuzzy/330/WRAP/TO_ZERO 640x480 S16*S16*(1/2^4)=S16
[ FAILED   ] vxMultiply.Fuzzy/331/WRAP/TO_ZERO 640x480 S16*S16*(1/2^5)=S16
[ FAILED   ] vxMultiply.Fuzzy/332/WRAP/TO_ZERO 640x480 S16*S16*(1/2^6)=S16
[ FAILED   ] vxMultiply.Fuzzy/333/WRAP/TO_ZERO 640x480 S16*S16*(1/2^7)=S16
[ FAILED   ] vxMultiply.Fuzzy/334/WRAP/TO_ZERO 640x480 S16*S16*(1/2^8)=S16
[ FAILED   ] vxMultiply.Fuzzy/335/WRAP/TO_ZERO 640x480 S16*S16*(1/2^9)=S16
[ FAILED   ] vxMultiply.Fuzzy/336/WRAP/TO_ZERO 640x480 S16*S16*(1/2^10)=S16
[ FAILED   ] vxMultiply.Fuzzy/337/WRAP/TO_ZERO 640x480 S16*S16*(1/2^11)=S16
[ FAILED   ] vxMultiply.Fuzzy/338/WRAP/TO_ZERO 640x480 S16*S16*(1/2^12)=S16
[ FAILED   ] vxMultiply.Fuzzy/339/WRAP/TO_ZERO 640x480 S16*S16*(1/2^13)=S16
[ FAILED   ] vxMultiply.Fuzzy/340/WRAP/TO_ZERO 640x480 S16*S16*(1/2^14)=S16
[ FAILED   ] vxMultiply.Fuzzy/341/WRAP/TO_ZERO 640x480 S16*S16*(1/2^15)=S16
[ FAILED   ] GaussianPyramid.GraphProcessing/1/randomInput/VX_BORDER_UNDEFINED/sz=16x16/VX_SCALE_PYRAMID_ORB
[ FAILED   ] GaussianPyramid.GraphProcessing/3/randomInput/VX_BORDER_UNDEFINED/sz=256x256/VX_SCALE_PYRAMID_ORB
[ FAILED   ] GaussianPyramid.GraphProcessing/5/randomInput/VX_BORDER_UNDEFINED/sz=640x480/VX_SCALE_PYRAMID_ORB
[ FAILED   ] GaussianPyramid.GraphProcessing/37/lena/VX_BORDER_UNDEFINED/VX_SCALE_PYRAMID_ORB
[ FAILED   ] GaussianPyramid.ImmediateProcessing/1/randomInput/VX_BORDER_UNDEFINED/sz=16x16/VX_SCALE_PYRAMID_ORB
[ FAILED   ] GaussianPyramid.ImmediateProcessing/3/randomInput/VX_BORDER_UNDEFINED/sz=256x256/VX_SCALE_PYRAMID_ORB
[ FAILED   ] GaussianPyramid.ImmediateProcessing/5/randomInput/VX_BORDER_UNDEFINED/sz=640x480/VX_SCALE_PYRAMID_ORB
[ FAILED   ] GaussianPyramid.ImmediateProcessing/37/lena/VX_BORDER_UNDEFINED/VX_SCALE_PYRAMID_ORB
[ DISABLED ] 8190 test(s)

[simonCatBot]

Analysis and Proposed Fixes for PoCL Conformance Failures

Looking at the 13 failing tests, they fall into 3 categories:

1. SmokeTestBase failures (3 tests)

• , ,
• Likely cause: PoCL-specific OpenCL context/kernel handling
• Fix: Check PoCL version compatibility (PoCL 3.0+ recommended for full OpenCL conformance)

2. vxConvertDepth.BitExact failures (2 tests)

• U8->S16 and S16->U8 conversions with WRAP mode
• Likely cause: PoCL handles overflow/wrap edge cases differently
• Fix: Review conversion kernel clamping logic for PoCL compatibility

3. GaussianPyramid failures (8 tests)

• All with VX_BORDER_UNDEFINED and VX_SCALE_PYRAMID_ORB
• Likely cause: Work-group size limits differ on PoCL
• Fix: Add runtime work-group size detection:

0 ""

0 ""

0 ""

1 "/usr/include/stdc-predef.h" 1 3 4

0 "" 2

1 ""

Recommended Investigation Steps:

1. Test with latest PoCL version (4.0+)
2. Compare work-group sizes between AMD OpenCL and PoCL
3. Check for PoCL-specific image format limitations
4. Add PoCL detection and conditional workarounds if needed

[simonCatBot]

Detailed Analysis: PoCL OpenVX Conformance Failures

Test Failure Breakdown

The 13 failures fall into 3 categories:

Category 1: SmokeTestBase (3 tests)

Analysis: These test the OpenVX base reference counting and kernel loading. PoCL may implement these differently than proprietary OpenCL drivers.

Possible Fix: Add PoCL-specific reference handling:

0 ""

0 ""

0 ""

1 "/usr/include/stdc-predef.h" 1 3 4

0 "" 2

1 ""

Category 2: vxConvertDepth.BitExact (2 tests)

• with WRAP mode (15x17)
• with WRAP mode (1x1)

Analysis: Bit-exact conversions with wrap mode are sensitive to implementation details. PoCL may use different SIMD paths or rounding.

Proposed Investigation:

0 ""

0 ""

0 ""

1 "/usr/include/stdc-predef.h" 1 3 4

0 "" 2

1 ""

Category 3: GaussianPyramid (8 tests)

All failures are with:

• (scale factor)

This is the most significant cluster of failures.

Root Cause Hypothesis:
PoCL has different work-group size limits than AMD OpenCL. The ORB pyramid scale calculation may produce work-group sizes incompatible with PoCL.

Detailed Fix Proposal:

0 ""

0 ""

0 ""

1 "/usr/include/stdc-predef.h" 1 3 4

0 "" 2

1 ""

PoCL-Specific Debug Code

Add runtime PoCL detection for targeted fixes:

0 ""

0 ""

0 ""

1 "/usr/include/stdc-predef.h" 1 3 4

0 "" 2

1 ""

Kernel-Specific Fixes for GaussianPyramid

The ORB scale uses = 0.832050294... (approx)

0 ""

0 ""

0 ""

1 "/usr/include/stdc-predef.h" 1 3 4

0 "" 2

1 ""

Recommended Next Steps

1. PoCL Version Check: Verify PoCL version:

2. Compare Work-Group Limits:

3. Isolate a Single Failure:

4. Kernel Compilation Check:

PoCL Known Issues

Based on PoCL's GitHub, common compatibility issues include:

• Image format support differences (CL_UNORM_INT8 vs raw)
• Memory alignment requirements
• Built-in function precision

Consider filing upstream PoCL issue if the problem is in their OpenCL implementation, not MIVisionX code.