AMD GPU not detected on Linux, undefined hidden symbols

[ottosch]

I always get this output:

INFO Using CPU backend for scanning (no GPU detected)

Since my OpenCL installation looked ok to me:

$ ldconfig -p | grep libOpenCL
libOpenCL.so.1 (libc6,x86-64) => /usr/lib/libOpenCL.so.1
libOpenCL.so (libc6,x86-64) => /usr/lib/libOpenCL.so

$ clinfo | grep "Number of"
Number of platforms                               1
Number of devices                                 1

I recompiled duckdb-ufsecp-extension, setting config.verbose = true, then replaced the compiled extension on frigate and got build errors this time:

2026-05-15 13:27:44,441 INFO Using home directory /home/otto/.frigate                                                                                                                                              
OpenCL Device: gfx90c:xnack+                                                                                                                                                                                       
  Vendor: Advanced Micro Devices, Inc.                                                                                                                                                                             
  Version: OpenCL 2.0                                                                                                                                                                                              
  Driver: 3581.0 (HSA1.1,LC)                                                                                                                                                                                       
  Global Memory: 16957 MB                                                                                                                                                                                          
  Local Memory: 64 KB                                                                                                                                                                                              
  Compute Units: 8                                                                                                                                                                                                 
  Max Clock: 1900 MHz                                                                                                                                                                                              
[DEBUG] Creating OpenCL context for device: gfx90c:xnack+                                                                                                                                                          
[DEBUG] Context created successfully                                                                                                                                                                               
[DEBUG] Command queue created successfully                                                                                                                                                                         
[DEBUG] build_program failed: Build failed: ld.lld: error: undefined hidden symbol: field_inv_impl                                                                                                                 
>>> referenced by /tmp/comgr-97681-7-3fe83e/input/linked.o:(field_inv)                                                                                                                                             
>>> referenced by /tmp/comgr-97681-7-3fe83e/input/linked.o:(field_inv)                                                                                                                                             
>>> referenced by /tmp/comgr-97681-7-3fe83e/input/linked.o:(field_inv)                                                                                                                                             
>>> referenced 11 more times                                                                                                                                                                                       
                                                                                                                                                                                                                   
ld.lld: error: undefined hidden symbol: scalar_mul_generator_glv_impl                                                                                                                                              
>>> referenced by /tmp/comgr-97681-7-3fe83e/input/linked.o:(scalar_mul_generator)                                                                                                                                  
>>> referenced by /tmp/comgr-97681-7-3fe83e/input/linked.o:(scalar_mul_generator)                                                                                                                                  
                                                                                                                                                                                                                   
ld.lld: error: undefined hidden symbol: point_add_impl                                                                                                                                                             
>>> referenced by /tmp/comgr-97681-7-3fe83e/input/linked.o:(point_add)                                                                                                                                             
>>> referenced by /tmp/comgr-97681-7-3fe83e/input/linked.o:(point_add)                                                                                                                                             
>>> referenced by /tmp/comgr-97681-7-3fe83e/input/linked.o:(msm_block_reduce_kernel)                                                                                                                               
>>> referenced 1 more times                                                                                                                                                                                        
                                                                                                                                                                                                                   
ld.lld: error: undefined hidden symbol: scalar_mul_glv_cl                                                                                                                                                          
>>> referenced by /tmp/comgr-97681-7-3fe83e/input/linked.o:(scalar_mul)
>>> referenced by /tmp/comgr-97681-7-3fe83e/input/linked.o:(scalar_mul)
Error: Creating the executable from LLVM IRs failed.

2026-05-15 13:27:44,984 INFO Using CPU backend for scanning (no GPU detected)

Maybe it's because UltrafastSecp256k1 was built without ROCm/HIP? Build log has:

-- +===========================================================+
-- |   UltrafastSecp256k1 Configuration                        |
-- +===========================================================+
--   Version:          3.66.0
--   Platform:         x86_64
--   C++ Standard:     20
--   Build Type:       Release
-- 
--   Components:
--     CPU:            ON
--     CUDA:           OFF
--     ROCm/HIP:       OFF
--     OpenCL:         ON
--     Metal:          OFF
--     Tests:          OFF
--     Benchmarks:     OFF
--     Examples:       OFF
--     Java JNI:       ON

I then tried to build UltrafastSecp256k1 with that support, but it would always fail with errors like:

[45/89] Building HIP object cuda_rocm/CMakeFiles/bench_bip324_cuda.dir/src/bench_bip324_transport.cu.o
FAILED: [code=1] cuda_rocm/CMakeFiles/bench_bip324_cuda.dir/src/bench_bip324_transport.cu.o
[46/89] Building HIP object cuda_rocm/CMakeFiles/gpu_bench_unified.dir/src/gpu_bench_unified.cu.o
FAILED: [code=1] cuda_rocm/CMakeFiles/gpu_bench_unified.dir/src/gpu_bench_unified.cu.o
[47/89] Building HIP object cuda_rocm/CMakeFiles/gpu_audit_runner.dir/src/gpu_audit_runner.cu.o
FAILED: [code=1] cuda_rocm/CMakeFiles/gpu_audit_runner.dir/src/gpu_audit_runner.cu.o

Even though I explicitly tried to disable benchmarks:

cmake -B build -DSECP256K1_BUILD_ROCM=ON -DSECP256K1_BUILD_CUDA=OFF -DSECP256K1_BUILD_TESTS=OFF -DSECP256K1_BUILD_BENCH=OFF -G Ninja

This was the config during this build attempt:

-- +===========================================================+
-- |   UltrafastSecp256k1 Configuration                        |
-- +===========================================================+
--   Version:          3.66.0
--   Platform:         x86_64
--   C++ Standard:     20
--   Build Type:       
-- 
--   Components:
--     CPU:            ON
--     CUDA:           OFF
--     ROCm/HIP:       ON
--     OpenCL:         OFF
--     Metal:          OFF
--     Tests:          OFF
--     Benchmarks:     OFF
--     Examples:       ON
--     Java JNI:       ON
-- 
--   Optimizations:
--     Assembly:       ON
--     Speed First:    OFF
--     Bitcoin Strict: ON

Since this could be specific to my GPU, will try building again if you have any suggestion.

[craigraw]

ROCm is not currently supported for AMD GPUs, only OpenCL. But the performance will be very similar.

Please provide the full output from clinfo.

[ottosch]

ROCm is not currently supported for AMD GPUs, only OpenCL. But the performance will be very similar.

oh, wasn't aware of that.

Output:

Number of platforms                               1
  Platform Name                                   AMD Accelerated Parallel Processing
  Platform Vendor                                 Advanced Micro Devices, Inc.
  Platform Version                                OpenCL 2.1 AMD-APP.dbg (3581.0)
  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                                     gfx90c:xnack+
  Device Vendor                                   Advanced Micro Devices, Inc.
  Device Vendor ID                                0x1002
  Device Version                                  OpenCL 2.0 
  Driver Version                                  3581.0 (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)                           0x164c
  Device Topology (AMD)                           PCI-E, 0000:03:00.0
  Device Profile                                  FULL_PROFILE
  Device Available                                Yes
  Compiler Available                              Yes
  Linker Available                                Yes
  Max compute units                               8
  SIMD per compute unit (AMD)                     4
  SIMD width (AMD)                                16
  SIMD instruction width (AMD)                    1
  Max clock frequency                             1900MHz
  Graphics IP (AMD)                               9.0
  Device Partition                                (core)
    Max number of sub-devices                     8
    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)     64
  Wavefront width (AMD)                           64
  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                              17781596160 (16.56GiB)
  Global free memory (AMD)                        17215296 (16.42GiB) 17215296 (16.42GiB)
  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                           15114356736 (14.08GiB)
  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                    15114356736 (14.08GiB)
  Preferred total size of global vars             17781596160 (16.56GiB)
  Global Memory cache type                        Read/Write
  Global Memory cache size                        16384 (16KiB)
  Global Memory cache line size                   64 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                            2229454848 (2.076GiB)
  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                        15114356736 (14.08GiB)
  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 (Wed Dec 31 21:00:00 1969)
  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)             8
  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                                   gfx90c:xnack+
  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                                   gfx90c:xnack+
  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                                   gfx90c:xnack+

ICD loader properties
  ICD loader Name                                 OpenCL ICD Loader
  ICD loader Vendor                               OCL Icd free software
  ICD loader Version                              2.3.4
  ICD loader Profile                              OpenCL 3.0

[craigraw]

Thanks for the info. It appears your OpenCL driver 3581.0 (HSA1.1,LC) (the LC is AMD's LLVM-based "Lightning Compiler") is unable to compile the kernel. That kernel compiles elsewhere (for example with NVIDIA and Intel OpenCL drivers) so it's a driver issue.

One thing worth trying is a different OpenCL driver for the same GPU, for example Mesa's Rusticl:

sudo apt install mesa-opencl-icd
mkdir -p ~/rusticl-icd
ln -s /etc/OpenCL/vendors/rusticl.icd ~/rusticl-icd/
export OCL_ICD_VENDORS=~/rusticl-icd
export RUSTICL_ENABLE=radeonsi
/opt/frigate/bin/frigate

[ottosch]

Sorry for the delay.
I tried that, but installing the opencl-mesa package for Arch Linux instead, but frigate got completely stuck at:

2026-06-01 21:32:57,679 INFO [main] c.s.f.Frigate Starting Frigate v1.5.2 on mainnet
2026-06-01 21:32:57,689 INFO [main] c.s.f.Frigate Using home directory /home/otto/.frigate

I had to pkill -9 frigate. After I unset OCL_ICD_VENDORS and started again, it ran normally (with no GPU detected).

[remcoros]

I experimented with rust-icl, but it freezes on both AMD Radeon APUs and Intel iGPUs (when forcing rusticl instead of using intel-opencl). Had an LLM investigate this, and it suspects that the rusticl/llvm compiler can't handle the complexity of the kernels.

I have near zero experience/knowledge of opencl, so not sure if this is even remotely true, but maybe it helps:

---

Part 3 introduces the hang. Part 3 contains scalar_mul_generator — a double-and-add loop over 256 bits, calling point_double_impl and point_add_mixed_impl at every step. Those inline field_mul_impl each time. At 256 iterations fully unrolled by LLVM, that's thousands of field muls — same explosion as the original field_inv_impl problem, just a different function.

So you're right to push back. The root cause is consistent: LLVM inlines everything aggressively and hits a combinatorial explosion. The fix that worked for field_inv_impl (loop-based squaring) worked because we removed the problematic call chain. The same fix is needed for scalar_mul_generator — but that one already uses a loop. The issue is LLVM unrolling that loop and inlining point_double_impl for each iteration.

Which means noinline on point_double_impl / point_add_mixed_impl needs to actually work. Let me verify if Rusticl really ignores noinline on those, or if it only ignored it on field_inv_impl for a different reason:

noinline is genuinely ignored by Rusticl. That's the Mesa bug — and it's the single root cause of everything we've been chasing.

The finding: Rusticl silently ignores attribute((noinline)) on OpenCL functions. This causes LLVM to fully inline all helper functions into kernels, which triggers compilation hangs on any non-trivial ECC/field arithmetic workload. field_inv_impl was the first symptom we hit; scalar_mul_generator inlining point_double_impl is another manifestation of the same bug.

This is the bug to file with Mesa. The fix is in Mesa/Rusticl, not in this extension, unless Craig wants to work around it by splitting into multiple separately-compiled programs (each small enough to stay under LLVM's pain threshold).