OpenCL Runtime Error on Metis PCIe with Voyager SDK v1.5.1

HI ​@FUKU55 ! (I moved your reply out into its own post so you can follow the replies more easily 🙂)

Glad to hear you’re enjoying your new Metis PCIe card! The OpenCL runtime error you’re seeing with Voyager SDK v1.5.1 sounds like it could be related to a breaking change introduced in this version (all models and pipelines must be recompiled using the new SDK). Are you running anything that was compiled under a previous version (like v1.4)? That could explain the crash.

If so, as a first step, try recompiling your model now you’re on v1.5.1, and let me know if that clears it up or if it persists, and we can dig deeper 👍

Thank you for your reply!

I'm testing from a new directory for v1.5.1. The model is empty.

Therefore, the model is being recompiled each time I launch the Python executable.

I am using the previous Python source, and there are also issues with the v1.5.1 source.

I'm still investigating the cause.
Thank you！

Ah, so that rules out the breaking change. Could be an OpenCL config or compatibility issue (we have seen that from time to time). Do you have a few more details on your host system? Maybe that’ll shed some light?

Hello!

I tried running `./inference.py yolov5s-v7-coco dataset --no-display`, but got an error.

I did download the model.

INFO : Using dataset val

INFO : Dataset ‘COCO2017’ split ‘val’ downloaded successfully to /home/tomo/.cache/axelera/data/coco

INFO : Dataset ‘COCO2017’ split ‘labels’ downloaded successfully to /home/tomo/.cache/axelera/data/coco

INFO : Dataset ‘COCO2017’ split ‘annotations’ downloaded successfully to /home/tomo/.cache/axelera/data/coco

Creating new label cache: /home/tomo/.cache/axelera/data/coco/labels/val2017/val_coco_objdet.cache

Labels found: 4952, corrupt images: 0

Background images: 48, missing label files: 48, empty label files: 0

terminate called after throwing an instance of ‘std::runtime_error’

what(): No functional OpenCL platform of type ‘’ found. Available platform may be installed but not working correctly.

Next, I tried this.

./inference.py yolov5s-v7-coco dataset --no-display --disable-opencl

INFO : Using dataset val

INFO : Model: yolov5s-v7-coco

INFO : Dataset: CocoDataset-COCO2017

INFO : Date: 2025-12-09 17:39:53.639316

INFO : Inference Time: 10251.38ms

INFO : Evaluation Time: 3196.38ms

INFO : Evaluation Metrics:

INFO : ==========================
INFO    : | mAP_box       | 36.41% |
INFO    : | mAP50_box     | 55.65% |
INFO    : | precision_box      | 67.01% |
INFO    : | recall_box       | 50.79% |
INFO    : ==========================
INFO    : Key Metric (mAP_box): 36.41%
Core Temp  : 30.0°C                                                                                                               
CPU %      : 9.3%
End-to-end : 487.2fps
Latency    : 105.6ms (min:33.7 max:126.6 σ:6.2 x̄:105.5)ms

This itself seems to be working fine.

Running `./inference.py yolov5s-v7-coco media/traffic1_1080p.mp4`
results in an error.
terminate called after throwing an instance of ‘std::runtime_error’
  what():  No functional OpenCL platform of type ‘’ found. Available platform may be installed but not working correctly.
 

The display uses the Ryzen integrated graphics processor.

Translated with DeepL.com (free version)

./inference.py yolov5s-v7-coco media/traffic1_1080p.mp4 -vv

The output was as follows:

DEBUG   :axelera.app.utils: $ vainfo
TRACE   :axelera.app.utils: Found VA-API in vainfo output: vainfo: Supported profile and entrypoints
DEBUG   :axelera.app.utils: Found OpenCL GPU devices for platform AMD Accelerated Parallel Processing: gfx1036
DEBUG   :axelera.app.device_manager: Using devices metis-0:5:0, metis-0:4:0, metis-0:3:0, metis-0:6:0
DEBUG   :axelera.app.network: Create network from ax_models/zoo/yolo/object_detection/yolov5s-v7-coco.yaml
TRACE   :axelera.app.utils: Changing class_path from $AXELERA_FRAMEWORK/ax_models/yolo/ax_yolo.py to /home/tomo/voyager150/ax_models/yolo/ax_yolo.py
TRACE   :axelera.app.utils: Changing weight_path from weights/yolov5s-v7.pt to /home/tomo/voyager150/ax_models/zoo/yolo/object_detection/weights/yolov5s-v7.pt
TRACE   :axelera.app.utils: Changing class_path from $AXELERA_FRAMEWORK/ax_datasets/objdataadapter.py to /home/tomo/voyager150/ax_datasets/objdataadapter.py
TRACE   :axelera.app.utils: Changing labels_path from $AXELERA_FRAMEWORK/ax_datasets/labels/coco.names to /home/tomo/voyager150/ax_datasets/labels/coco.names
TRACE   :axelera.app.utils: Changing class_path from $AXELERA_FRAMEWORK/ax_models/decoders/yolo.py to /home/tomo/voyager150/ax_models/decoders/yolo.py
TRACE   :axelera.app.utils: Changing weight_path from weights/yolov5s-v7.pt to /home/tomo/.cache/axelera/weights/yolov5s-v7-coco/yolov5s-v7.pt
TRACE   :axelera.app.utils: Changing class_path from $AXELERA_FRAMEWORK/ax_models/yolo/ax_yolo.py to /home/tomo/voyager150/ax_models/yolo/ax_yolo.py
TRACE   :axelera.app.utils: Changing weight_path from weights/yolov5s-v7.pt to /home/tomo/voyager150/ax_models/zoo/yolo/object_detection/weights/yolov5s-v7.pt
TRACE   :axelera.app.utils: Changing class_path from $AXELERA_FRAMEWORK/ax_datasets/objdataadapter.py to /home/tomo/voyager150/ax_datasets/objdataadapter.py
TRACE   :axelera.app.utils: Changing labels_path from $AXELERA_FRAMEWORK/ax_datasets/labels/coco.names to /home/tomo/voyager150/ax_datasets/labels/coco.names
TRACE   :axelera.app.utils: Changing class_path from $AXELERA_FRAMEWORK/ax_models/decoders/yolo.py to /home/tomo/voyager150/ax_models/decoders/yolo.py
TRACE   :axelera.app.utils: Changing class_path from $AXELERA_FRAMEWORK/ax_models/yolo/ax_yolo.py to /home/tomo/voyager150/ax_models/yolo/ax_yolo.py
TRACE   :axelera.app.utils: Changing class_path from $AXELERA_FRAMEWORK/ax_datasets/objdataadapter.py to /home/tomo/voyager150/ax_datasets/objdataadapter.py
TRACE   :axelera.app.utils: Changing labels_path from $AXELERA_FRAMEWORK/ax_datasets/labels/coco.names to /home/tomo/voyager150/ax_datasets/labels/coco.names
TRACE   :axelera.app.utils: Changing class_path from $AXELERA_FRAMEWORK/ax_models/decoders/yolo.py to /home/tomo/voyager150/ax_models/decoders/yolo.py
DEBUG   :axelera.app.network: Register custom operator 'decodeyolo' with class DecodeYolo from ax_models/decoders/yolo.py
TRACE   :axelera.app.pipeline: The source is not clearly declared, default as full frame
TRACE   :axelera.app.network: 1. build/yolov5s-v7-coco/yolov5s-v7-coco/4/manifest.json: Available
DEBUG   :axelera.app.device_manager: Reconfiguring devices with device_firmware=1, mvm_utilisation_core_0=100%, clock_profile_core_0=800MHz, mvm_utilisation_core_1=100%, clock_profile_core_1=800MHz, mvm_utilisation_core_2=100%, clock_profile_core_2=800MHz, mvm_utilisation_core_3=100%, clock_profile_core_3=800MHz
DEBUG   :axelera.app.pipe.manager: 
DEBUG   :axelera.app.pipe.manager: --- EXECUTION VIEW ---
DEBUG   :axelera.app.pipe.manager: Input
DEBUG   :axelera.app.pipe.manager:   └─detections
DEBUG   :axelera.app.pipe.manager: 
DEBUG   :axelera.app.pipe.manager: --- RESULT VIEW ---
DEBUG   :axelera.app.pipe.manager: Input
DEBUG   :axelera.app.pipe.manager:   └─detections
DEBUG   :axelera.app.pipe.manager: Network type: NetworkType.SINGLE_MODEL
TRACE   :axelera.app.transforms: composite_expansion transformed:
TRACE   :axelera.app.transforms:   TorchToTensor(input_layout='NHWC', output_layout='NCHW', datatype='float32', scale=True)
TRACE   :axelera.app.transforms:  to:
TRACE   :axelera.app.transforms:   ToTensor()
TRACE   :axelera.app.transforms:   PermuteChannels(input_layout=<TensorLayout.NHWC: 2>, output_layout=<TensorLayout.NCHW: 1>)
TRACE   :axelera.app.transforms:   TypeCast(datatype='float32')
TRACE   :axelera.app.transforms:   Normalize(mean='0', std='255.0', tensor_layout=<TensorLayout.NCHW: 1>, format='RGB')
TRACE   :axelera.app.transforms: opencl_letterbox_with_normalize transformed:
TRACE   :axelera.app.transforms:   Letterbox(height=640, width=640, scaleup=True, half_pixel_centers=False, pad_val=114, image_width=1000000, image_height=1000000)
TRACE   :axelera.app.transforms:   ToTensor()
TRACE   :axelera.app.transforms:   PermuteChannels(input_layout=<TensorLayout.NHWC: 2>, output_layout=<TensorLayout.NCHW: 1>)
TRACE   :axelera.app.transforms:   TypeCast(datatype='float32')
TRACE   :axelera.app.transforms:   Normalize(mean='0', std='255.0', tensor_layout=<TensorLayout.NCHW: 1>, format='RGB')
TRACE   :axelera.app.transforms:  to:
TRACE   :axelera.app.transforms:   OpenCLetterBoxToTensorAndNormalize(width=640, height=640, scaleup=True, pad_val=114, mean=[0.0], std=[1.0], datatype='float32')
DEBUG   :yolo: Model Type: YoloFamily.YOLOv5 (YOLOv5 pattern:
DEBUG   :yolo: - 3 output tensors (anchor-based)
DEBUG   :yolo: - Each output has 255 channels
DEBUG   :yolo:   = 3 anchors × (4 box + 1 obj + 80 classes)
DEBUG   :yolo:   = 3 × (4 + 1 + 80) = 255
DEBUG   :yolo: - Shapes: [[4, 40, 40, 255], [4, 20, 20, 255], [4, 80, 80, 255]])
DEBUG   :axelera.app.pipe.io: New source 0: media/traffic1_1080p.mp4 (VIDEO_FILE)
DEBUG   :axelera.app.pipe.io: FPS of media/traffic1_1080p.mp4: 60
DEBUG   :axelera.app.operators.inference: Using inferencenet name=inference-task0 model=/home/tomo/voyager150/build/yolov5s-v7-coco/yolov5s-v7-coco/4/model.json devices=metis-0:5:0,metis-0:4:0,metis-0:3:0,metis-0:6:0 double_buffer=True dmabuf_inputs=True dmabuf_outputs=True num_children=0
TRACE   :axelera.app.pipe.gst: environment at gst pipeline construction:
TRACE   :axelera.app.pipe.gst: {'AXELERA_DEVICE_DIR': '/opt/axelera/device-1.5.1-1/omega',
TRACE   :axelera.app.pipe.gst:  'AXELERA_FRAMEWORK': '/home/tomo/voyager150',
TRACE   :axelera.app.pipe.gst:  'AXELERA_RISCV_TOOLCHAIN_DIR': '/opt/axelera/riscv-gnu-newlib-toolchain-409b951ba662-7',
TRACE   :axelera.app.pipe.gst:  'AXELERA_RUNTIME_DIR': '/opt/axelera/runtime-1.5.1-1',
TRACE   :axelera.app.pipe.gst:  'AXE_PROFILING_CONFIG': ''}
DEBUG   :axelera.app.pipe.gst: Started building gst pipeline
TRACE   :axelera.app.pipe.gst_helper: Creating filesrc0 filesrc(location=***)
TRACE   :axelera.app.pipe.gst_helper: Creating decodebin0 decodebin(force-sw-decoders=True, caps=video/x-raw(ANY), expose-all-streams=False)
TRACE   :axelera.app.pipe.gst_helper: Creating decodebin-link0 axinplace(lib=libinplace_addstreamid.so, mode=meta, options=stream_id:0)
TRACE   :axelera.app.pipe.gst_helper: Creating axtransform-colorconvert-cl0 axtransform(lib=libtransform_colorconvert_cl.so, options=format:rgb)
terminate called after throwing an instance of 'std::runtime_error'
  what():  No functional OpenCL platform of type '' found. Available platform may be installed but not working correctly.
中止 (コアダンプ)
 

clinfo displays the following:

Number of platforms                               1
  Platform Name                                   AMD Accelerated Parallel Processing
  Platform Vendor                                 Advanced Micro Devices, Inc.
  Platform Version                                OpenCL 2.1 AMD-APP (3649.0)
  Platform Profile                                FULL_PROFILE
  Platform Extensions                             cl_khr_icd cl_amd_event_callback 
  Platform Extensions function suffix             AMD
  Platform Host timer resolution                  1ns

  Platform Name                                   AMD Accelerated Parallel Processing
Number of devices                                 1
  Device Name                                     gfx1036
  Device Vendor                                   Advanced Micro Devices, Inc.
  Device Vendor ID                                0x1002
  Device Version                                  OpenCL 2.0 
  Driver Version                                  3649.0 (HSA1.1,LC)
  Device OpenCL C Version                         OpenCL C 2.0 
  Device Type                                     GPU
  Device Board Name (AMD)                         AMD Radeon Graphics
  Device PCI-e ID (AMD)                           0x164e
  Device Topology (AMD)                           PCI-E, 0000:11:00.0
  Device Profile                                  FULL_PROFILE
  Device Available                                Yes
  Compiler Available                              Yes
  Linker Available                                Yes
  Max compute units                               1
  SIMD per compute unit (AMD)                     4
  SIMD width (AMD)                                32
  SIMD instruction width (AMD)                    1
  Max clock frequency                             2200MHz
  Graphics IP (AMD)                               10.3
  Device Partition                                (core)
    Max number of sub-devices                     1
    Supported partition types                     None
    Supported affinity domains                    (n/a)
  Max work item dimensions                        3
  Max work item sizes                             1024x1024x1024
  Max work group size                             256
  Preferred work group size (AMD)                 256
  Max work group size (AMD)                       1024
  Preferred work group size multiple (kernel)     32
  Wavefront width (AMD)                           32
  Preferred / native vector sizes                 
    char                                                 4 / 4       
    short                                                2 / 2       
    int                                                  1 / 1       
    long                                                 1 / 1       
    half                                                 1 / 1        (cl_khr_fp16)
    float                                                1 / 1       
    double                                               1 / 1        (cl_khr_fp64)
  Half-precision Floating-point support           (cl_khr_fp16)
    Denormals                                     Yes
    Infinity and NANs                             Yes
    Round to nearest                              Yes
    Round to zero                                 Yes
    Round to infinity                             Yes
    IEEE754-2008 fused multiply-add               Yes
    Support is emulated in software               No
  Single-precision Floating-point support         (core)
    Denormals                                     Yes
    Infinity and NANs                             Yes
    Round to nearest                              Yes
    Round to zero                                 Yes
    Round to infinity                             Yes
    IEEE754-2008 fused multiply-add               Yes
    Support is emulated in software               No
    Correctly-rounded divide and sqrt operations  Yes
  Double-precision Floating-point support         (cl_khr_fp64)
    Denormals                                     Yes
    Infinity and NANs                             Yes
    Round to nearest                              Yes
    Round to zero                                 Yes
    Round to infinity                             Yes
    IEEE754-2008 fused multiply-add               Yes
    Support is emulated in software               No
  Address bits                                    64, Little-Endian
  Global memory size                              33266290688 (30.98GiB)
  Global free memory (AMD)                        32318916 (30.82GiB) 32318916 (30.82GiB)
  Global memory channels (AMD)                    4
  Global memory banks per channel (AMD)           4
  Global memory bank width (AMD)                  256 bytes
  Error Correction support                        No
  Max memory allocation                           28276347080 (26.33GiB)
  Unified memory for Host and Device              Yes
  Shared Virtual Memory (SVM) capabilities        (core)
    Coarse-grained buffer sharing                 Yes
    Fine-grained buffer sharing                   Yes
    Fine-grained system sharing                   No
    Atomics                                       No
  Minimum alignment for any data type             128 bytes
  Alignment of base address                       2048 bits (256 bytes)
  Preferred alignment for atomics                 
    SVM                                           0 bytes
    Global                                        0 bytes
    Local                                         0 bytes
  Max size for global variable                    28276347080 (26.33GiB)
  Preferred total size of global vars             33266290688 (30.98GiB)
  Global Memory cache type                        Read/Write
  Global Memory cache size                        16384 (16KiB)
  Global Memory cache line size                   128 bytes
  Image support                                   Yes
    Max number of samplers per kernel             16
    Max size for 1D images from buffer            134217728 pixels
    Max 1D or 2D image array size                 8192 images
    Base address alignment for 2D image buffers   256 bytes
    Pitch alignment for 2D image buffers          256 pixels
    Max 2D image size                             16384x16384 pixels
    Max 3D image size                             16384x16384x8192 pixels
    Max number of read image args                 128
    Max number of write image args                8
    Max number of read/write image args           64
  Max number of pipe args                         16
  Max active pipe reservations                    16
  Max pipe packet size                            2506543304 (2.334GiB)
  Local memory type                               Local
  Local memory size                               65536 (64KiB)
  Local memory size per CU (AMD)                  65536 (64KiB)
  Local memory banks (AMD)                        32
  Max number of constant args                     8
  Max constant buffer size                        28276347080 (26.33GiB)
  Preferred constant buffer size (AMD)            16384 (16KiB)
  Max size of kernel argument                     1024
  Queue properties (on host)                      
    Out-of-order execution                        No
    Profiling                                     Yes
  Queue properties (on device)                    
    Out-of-order execution                        Yes
    Profiling                                     Yes
    Preferred size                                262144 (256KiB)
    Max size                                      8388608 (8MiB)
  Max queues on device                            1
  Max events on device                            1024
  Prefer user sync for interop                    Yes
  Number of P2P devices (AMD)                     0
  Profiling timer resolution                      1ns
  Profiling timer offset since Epoch (AMD)        0ns (Thu Jan  1 09:00:00 1970)
  Execution capabilities                          
    Run OpenCL kernels                            Yes
    Run native kernels                            No
    Thread trace supported (AMD)                  No
    Number of async queues (AMD)                  8
    Max real-time compute queues (AMD)            8
    Max real-time compute units (AMD)             1
  printf() buffer size                            4194304 (4MiB)
  Built-in kernels                                (n/a)
  Device Extensions                               cl_khr_fp64 cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_gl_sharing cl_amd_device_attribute_query cl_amd_media_ops cl_amd_media_ops2 cl_khr_image2d_from_buffer cl_khr_subgroups cl_khr_depth_images cl_amd_copy_buffer_p2p cl_amd_assembly_program 

NULL platform behavior
  clGetPlatformInfo(NULL, CL_PLATFORM_NAME, ...)  AMD Accelerated Parallel Processing
  clGetDeviceIDs(NULL, CL_DEVICE_TYPE_ALL, ...)   Success [AMD]
  clCreateContext(NULL, ...) [default]            Success [AMD]
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_DEFAULT)  Success (1)
    Platform Name                                 AMD Accelerated Parallel Processing
    Device Name                                   gfx1036
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_CPU)  No devices found in platform
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_GPU)  Success (1)
    Platform Name                                 AMD Accelerated Parallel Processing
    Device Name                                   gfx1036
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_ACCELERATOR)  No devices found in platform
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_CUSTOM)  No devices found in platform
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_ALL)  Success (1)
    Platform Name                                 AMD Accelerated Parallel Processing
    Device Name                                   gfx1036

ICD loader properties
  ICD loader Name                                 OpenCL ICD Loader
  ICD loader Vendor                               OCL Icd free software
  ICD loader Version                              2.2.14
  ICD loader Profile                              OpenCL 3.0
 

Hi ​@FUKU55 !

Looks like the gfx1036 is being detected but maybe OpenCL isn't properly initialised?

Running clinfo in your terminal might shed some light on things. It'll show us if the AMD OpenCL platform is actually functional or just visible but broken

In the meantime, --disable-opencl is a perfectly valid workaround - just a bit slower on the preprocessing side 👍

Hello, Spanner!

Thanks for getting in touch.

I tried replying to clinfo again, but it still doesn't seem to be reflecting.

I'll try reinstalling the GPU driver,

or in some cases, I might rebuild the system.

Hello,Spanner！

This issue may be due to differences between Voyager 1.5 and 1.4.
When I installed Voyager 1.4 (in a separate directory), it worked without errors.
The board's firmware is the latest Voyager 1.5 version.

Voyager 1.5 seems to cause STD runtime errors.
I'll keep studying.

I was just wondering,

but when I looked at the source /operators/axstreamer/src/AxOpenCl.cpp

std::vector<std::string>
get_order_preference(std::string_view preference, Ax::Logger &logger)
{
  using namespace std::string_literals;
  if (preference == "intel") {
    return { "Intel"s };
  } else if (preference == "arm") {
    return { "ARM"s, "rusticl"s };
  } else if (preference == "cpu") {
    return { "Portable Computing Language"s };
  } else if (preference == "gpu") {
    return { "NVIDIA"s, "Intel"s, "ARM"s };
  } else if (preference == "nvidia") {
    return { "NVIDIA"s };
  } else if (preference != "auto") { // AUTO or any other value
    logger(AX_WARN) << "Unknown OpenCL preference: " << preference
                    << ", using auto" << std::endl;
  }
  return std::vector<std::string>{
    "Intel",
    "ARM",
    "rusticl",
    "NVIDIA",
    "Portable Computing Language",
  };
}

it seems the GPU section doesn't include RADEON or anything similar.
Is it supported?
Is there another way to recognize it?
I'm currently using a RADEON GPU, so I plan to try out an NVIDIA GPU.
 

I can’t remember if I asked this already, but did you recompile the model(s) for use in v1.5.x ? That’s a breaking change that was required - bit inconvenient at the moment, but it’ll make things a lot better going forward 😄

Just wondering if that’s why 1.4 worked where 1.5 didn’t?

Thank you.

That is indeed something we should be mindful of.

I've been installing and running it in each version's directory, and I assumed the model was being recompiled each time.

I'll double-check again.

Sometimes you think you're running it, but you're not, right?

Hello,Spanner !!!

I have good news to report!

I wiped the system clean and installed Ubuntu.
After installing Voyager SDK fresh from the 1.5 repository.
it worked perfectly!!

It still throws errors without --disable-opencl, but it runs.

./inference.py yolov5s-v7-coco media/traffic1_1080p.mp4
INFO    : Could not exec vainfo: Command ‘['vainfo’]' returned non-zero exit status 3.
WARNING : Failed to get OpenCL platforms : clGetPlatformIDs failed: PLATFORM_NOT_FOUND_KHR
WARNING : Please check the documentation for installation instructions
Core Temp  : 32.0°C                                                                                                                          
CPU %      : 22.6%
End-to-end : 414.2fps
Latency    : 133.7ms (min:21.1 max:234.3 σ:18.6 x̄:128.0)ms

Yeah, gotta reinstall.
I'll set up the environment and test it out. Thanks!

Translated with DeepL.com (free version)

Ah, great work ​@FUKU55 ! Unfortunate that it took a hard reset, but sometimes that’s just the best option with so many devices and systems! 😄