Archives Discussions

[ekondis]

Hello,

I've been using a Linux machine (Ubuntu 14.04.3 64bit) for OpenCL experiments and research through remote access (ssh). It is equipped with 2 different vendor GPU cards (GTX-660 and R7-260X). The GTX-660 works as an OpenCL 1.2 device and the R7-260X as an OpenCL 2.0 device. However, after I upgraded to the Crimson driver (15.30) from 15.20 two issues were raised:

1. The R7-260X GPU is not recognised as an OpenCL 2.0 device any more but rather as a 1.2 device.
2. The CPU has disappeared from the list of valid OpenCL devices!

The computer is based on a quite old CPU (Pentium 4 3.06GHz CPU with 2GB RAM) but as a host machine for GPU computing purpose it works ok.

So, have the requirements for OpenCL 2.0 changed in any way for Crimson driver? Or is it maybe the small amount of RAM? It would be good to have an official anwer in this question.

Here I provide the clinfo output for Crimson driver

Number of platforms:                 2
  Platform Profile:                 FULL_PROFILE
  Platform Version:                 OpenCL 1.2 CUDA 7.5.23
  Platform Name:                 NVIDIA CUDA
  Platform Vendor:                 NVIDIA Corporation
  Platform Extensions:                 cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_copy_opts 
  Platform Profile:                 FULL_PROFILE
  Platform Version:                 OpenCL 2.0 AMD-APP (1912.5)
  Platform Name:                 AMD Accelerated Parallel Processing
  Platform Vendor:                 Advanced Micro Devices, Inc.
  Platform Extensions:                 cl_khr_icd cl_amd_event_callback cl_amd_offline_devices 
  Platform Name:                 NVIDIA CUDA
Number of devices:                 1
  Device Type:                     CL_DEVICE_TYPE_GPU
  Vendor ID:                     10deh
  Max compute units:                 5
  Max work items dimensions:             3
    Max work items[0]:                 1024
    Max work items[1]:                 1024
    Max work items[2]:                 64
  Max work group size:                 1024
  Preferred vector width char:             1
  Preferred vector width short:             1
  Preferred vector width int:             1
  Preferred vector width long:             1
  Preferred vector width float:             1
  Preferred vector width double:         1
  Native vector width char:             1
  Native vector width short:             1
  Native vector width int:             1
  Native vector width long:             1
  Native vector width float:             1
  Native vector width double:             1
  Max clock frequency:                 1097Mhz
  Address bits:                     64
  Max memory allocation:             535478272
  Image support:                 Yes
  Max number of images read arguments:         256
  Max number of images write arguments:         16
  Max image 2D width:                 16384
  Max image 2D height:                 16384
  Max image 3D width:                 4096
  Max image 3D height:                 4096
  Max image 3D depth:                 4096
  Max samplers within kernel:             32
  Max size of kernel argument:             4352
  Alignment (bits) of base address:         4096
  Minimum alignment (bytes) for any datatype:     128
  Single precision floating point capability
    Denorms:                     Yes
    Quiet NaNs:                     Yes
    Round to nearest even:             Yes
    Round to zero:                 Yes
    Round to +ve and infinity:             Yes
    IEEE754-2008 fused multiply-add:         Yes
  Cache type:                     Read/Write
  Cache line size:                 128
  Cache size:                     81920
  Global memory size:                 2141913088
  Constant buffer size:                 65536
  Max number of constant args:             9
  Local memory type:                 Scratchpad
  Local memory size:                 49152
  Kernel Preferred work group size multiple:     32
  Error correction support:             0
  Unified memory for Host and Device:         0
  Profiling timer resolution:             1000
  Device endianess:                 Little
  Available:                     Yes
  Compiler available:                 Yes
  Execution capabilities:                 
    Execute OpenCL kernels:             Yes
    Execute native function:             No
  Queue on Host properties:                 
    Out-of-Order:                 Yes
    Profiling :                     Yes
  Platform ID:                     0x1051a30
  Name:                         GeForce GTX 660
  Vendor:                     NVIDIA Corporation
  Device OpenCL C version:             OpenCL C 1.2 
  Driver version:                 352.63
  Profile:                     FULL_PROFILE
  Version:                     OpenCL 1.2 CUDA
  Extensions:                     cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_copy_opts  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_fp64 
  Platform Name:                 AMD Accelerated Parallel Processing
Number of devices:                 1
  Device Type:                     CL_DEVICE_TYPE_GPU
  Vendor ID:                     1002h
  Board name:                     
  Device Topology:                 PCI[ B#4, D#0, F#0 ]
  Max compute units:                 14
  Max work items dimensions:             3
    Max work items[0]:                 256
    Max work items[1]:                 256
    Max work items[2]:                 256
  Max work group size:                 256
  Preferred vector width char:             4
  Preferred vector width short:             2
  Preferred vector width int:             1
  Preferred vector width long:             1
  Preferred vector width float:             1
  Preferred vector width double:         1
  Native vector width char:             4
  Native vector width short:             2
  Native vector width int:             1
  Native vector width long:             1
  Native vector width float:             1
  Native vector width double:             1
  Max clock frequency:                 1175Mhz
  Address bits:                     64
  Max memory allocation:             1404198912
  Image support:                 Yes
  Max number of images read arguments:         128
  Max number of images write arguments:         8
  Max image 2D width:                 16384
  Max image 2D height:                 16384
  Max image 3D width:                 2048
  Max image 3D height:                 2048
  Max image 3D depth:                 2048
  Max samplers within kernel:             16
  Max size of kernel argument:             1024
  Alignment (bits) of base address:         2048
  Minimum alignment (bytes) for any datatype:     128
  Single precision floating point capability
    Denorms:                     No
    Quiet NaNs:                     Yes
    Round to nearest even:             Yes
    Round to zero:                 Yes
    Round to +ve and infinity:             Yes
    IEEE754-2008 fused multiply-add:         Yes
  Cache type:                     Read/Write
  Cache line size:                 64
  Cache size:                     16384
  Global memory size:                 2131105664
  Constant buffer size:                 65536
  Max number of constant args:             8
  Local memory type:                 Scratchpad
  Local memory size:                 32768
  Max pipe arguments:                 0
  Max pipe active reservations:             0
  Max pipe packet size:                 0
  Max global variable size:             0
  Max global variable preferred total size:     0
  Max read/write image args:             0
  Max on device events:                 0
  Queue on device max size:             0
  Max on device queues:                 0
  Queue on device preferred size:         0
  SVM capabilities:                 
    Coarse grain buffer:             No
    Fine grain buffer:                 No
    Fine grain system:                 No
    Atomics:                     No
  Preferred platform atomic alignment:         0
  Preferred global atomic alignment:         0
  Preferred local atomic alignment:         0
  Kernel Preferred work group size multiple:     64
  Error correction support:             0
  Unified memory for Host and Device:         0
  Profiling timer resolution:             1
  Device endianess:                 Little
  Available:                     Yes
  Compiler available:                 Yes
  Execution capabilities:                 
    Execute OpenCL kernels:             Yes
    Execute native function:             No
  Queue on Host properties:                 
    Out-of-Order:                 No
    Profiling :                     Yes
  Queue on Device properties:                 
    Out-of-Order:                 No
    Profiling :                     No
  Platform ID:                     0x7f7f516c4a18
  Name:                         Bonaire
  Vendor:                     Advanced Micro Devices, Inc.
  Device OpenCL C version:             OpenCL C 1.2 
  Driver version:                 1912.5 (VM)
  Profile:                     FULL_PROFILE
  Version:                     OpenCL 1.2 AMD-APP (1912.5)
  Extensions:                     cl_khr_fp64 cl_amd_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_gl_sharing cl_ext_atomic_counters_32 cl_amd_device_attribute_query cl_amd_vec3 cl_amd_printf cl_amd_media_ops cl_amd_media_ops2 cl_amd_popcnt cl_khr_image2d_from_buffer cl_khr_spir cl_khr_gl_event 

And here is clinfo output if I replace the libamdocl64.so and libamdocl12cl64.so files with the ones supplied from the previous driver (15.20):

Number of platforms:                 2
  Platform Profile:                 FULL_PROFILE
  Platform Version:                 OpenCL 1.2 CUDA 7.5.23
  Platform Name:                 NVIDIA CUDA
  Platform Vendor:                 NVIDIA Corporation
  Platform Extensions:                 cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_copy_opts 
  Platform Profile:                 FULL_PROFILE
  Platform Version:                 OpenCL 2.0 AMD-APP (1800.11)
  Platform Name:                 AMD Accelerated Parallel Processing
  Platform Vendor:                 Advanced Micro Devices, Inc.
  Platform Extensions:                 cl_khr_icd cl_amd_event_callback cl_amd_offline_devices 
  Platform Name:                 NVIDIA CUDA
Number of devices:                 1
  Device Type:                     CL_DEVICE_TYPE_GPU
  Vendor ID:                     10deh
  Max compute units:                 5
  Max work items dimensions:             3
    Max work items[0]:                 1024
    Max work items[1]:                 1024
    Max work items[2]:                 64
  Max work group size:                 1024
  Preferred vector width char:             1
  Preferred vector width short:             1
  Preferred vector width int:             1
  Preferred vector width long:             1
  Preferred vector width float:             1
  Preferred vector width double:         1
  Native vector width char:             1
  Native vector width short:             1
  Native vector width int:             1
  Native vector width long:             1
  Native vector width float:             1
  Native vector width double:             1
  Max clock frequency:                 1097Mhz
  Address bits:                     64
  Max memory allocation:             535478272
  Image support:                 Yes
  Max number of images read arguments:         256
  Max number of images write arguments:         16
  Max image 2D width:                 16384
  Max image 2D height:                 16384
  Max image 3D width:                 4096
  Max image 3D height:                 4096
  Max image 3D depth:                 4096
  Max samplers within kernel:             32
  Max size of kernel argument:             4352
  Alignment (bits) of base address:         4096
  Minimum alignment (bytes) for any datatype:     128
  Single precision floating point capability
    Denorms:                     Yes
    Quiet NaNs:                     Yes
    Round to nearest even:             Yes
    Round to zero:                 Yes
    Round to +ve and infinity:             Yes
    IEEE754-2008 fused multiply-add:         Yes
  Cache type:                     Read/Write
  Cache line size:                 128
  Cache size:                     81920
  Global memory size:                 2141913088
  Constant buffer size:                 65536
  Max number of constant args:             9
  Local memory type:                 Scratchpad
  Local memory size:                 49152
  Kernel Preferred work group size multiple:     32
  Error correction support:             0
  Unified memory for Host and Device:         0
  Profiling timer resolution:             1000
  Device endianess:                 Little
  Available:                     Yes
  Compiler available:                 Yes
  Execution capabilities:                 
    Execute OpenCL kernels:             Yes
    Execute native function:             No
  Queue on Host properties:                 
    Out-of-Order:                 Yes
    Profiling :                     Yes
  Platform ID:                     0x2694a30
  Name:                         GeForce GTX 660
  Vendor:                     NVIDIA Corporation
  Device OpenCL C version:             OpenCL C 1.2 
  Driver version:                 352.63
  Profile:                     FULL_PROFILE
  Version:                     OpenCL 1.2 CUDA
  Extensions:                     cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_copy_opts  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_fp64 
  Platform Name:                 AMD Accelerated Parallel Processing
Number of devices:                 2
  Device Type:                     CL_DEVICE_TYPE_GPU
  Vendor ID:                     1002h
  Board name:                     
  Device Topology:                 PCI[ B#4, D#0, F#0 ]
  Max compute units:                 14
  Max work items dimensions:             3
    Max work items[0]:                 256
    Max work items[1]:                 256
    Max work items[2]:                 256
  Max work group size:                 256
  Preferred vector width char:             4
  Preferred vector width short:             2
  Preferred vector width int:             1
  Preferred vector width long:             1
  Preferred vector width float:             1
  Preferred vector width double:         1
  Native vector width char:             4
  Native vector width short:             2
  Native vector width int:             1
  Native vector width long:             1
  Native vector width float:             1
  Native vector width double:             1
  Max clock frequency:                 1175Mhz
  Address bits:                     64
  Max memory allocation:             1404567552
  Image support:                 Yes
  Max number of images read arguments:         128
  Max number of images write arguments:         64
  Max image 2D width:                 16384
  Max image 2D height:                 16384
  Max image 3D width:                 2048
  Max image 3D height:                 2048
  Max image 3D depth:                 2048
  Max samplers within kernel:             16
  Max size of kernel argument:             1024
  Alignment (bits) of base address:         2048
  Minimum alignment (bytes) for any datatype:     128
  Single precision floating point capability
    Denorms:                     No
    Quiet NaNs:                     Yes
    Round to nearest even:             Yes
    Round to zero:                 Yes
    Round to +ve and infinity:             Yes
    IEEE754-2008 fused multiply-add:         Yes
  Cache type:                     Read/Write
  Cache line size:                 64
  Cache size:                     16384
  Global memory size:                 2130706432
  Constant buffer size:                 65536
  Max number of constant args:             8
  Local memory type:                 Scratchpad
  Local memory size:                 32768
  Max pipe arguments:                 16
  Max pipe active reservations:             16
  Max pipe packet size:                 1404567552
  Max global variable size:             1264110592
  Max global variable preferred total size:     2130706432
  Max read/write image args:             64
  Max on device events:                 1024
  Queue on device max size:             524288
  Max on device queues:                 1
  Queue on device preferred size:         262144
  SVM capabilities:                 
    Coarse grain buffer:             Yes
    Fine grain buffer:                 Yes
    Fine grain system:                 No
    Atomics:                     No
  Preferred platform atomic alignment:         0
  Preferred global atomic alignment:         0
  Preferred local atomic alignment:         0
  Kernel Preferred work group size multiple:     64
  Error correction support:             0
  Unified memory for Host and Device:         0
  Profiling timer resolution:             1
  Device endianess:                 Little
  Available:                     Yes
  Compiler available:                 Yes
  Execution capabilities:                 
    Execute OpenCL kernels:             Yes
    Execute native function:             No
  Queue on Host properties:                 
    Out-of-Order:                 No
    Profiling :                     Yes
  Queue on Device properties:                 
    Out-of-Order:                 Yes
    Profiling :                     Yes
  Platform ID:                     0x7f6ffde01430
  Name:                         Bonaire
  Vendor:                     Advanced Micro Devices, Inc.
  Device OpenCL C version:             OpenCL C 2.0 
  Driver version:                 1800.11 (VM)
  Profile:                     FULL_PROFILE
  Version:                     OpenCL 2.0 AMD-APP (1800.11)
  Extensions:                     cl_khr_fp64 cl_amd_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_gl_sharing cl_khr_gl_depth_images cl_ext_atomic_counters_32 cl_amd_device_attribute_query cl_amd_vec3 cl_amd_printf cl_amd_media_ops cl_amd_media_ops2 cl_amd_popcnt cl_khr_image2d_from_buffer cl_khr_spir cl_khr_subgroups cl_khr_gl_event cl_khr_depth_images 
  Device Type:                     CL_DEVICE_TYPE_CPU
  Vendor ID:                     1002h
  Board name:                     
  Max compute units:                 1
  Max work items dimensions:             3
    Max work items[0]:                 1024
    Max work items[1]:                 1024
    Max work items[2]:                 1024
  Max work group size:                 1024
  Preferred vector width char:             16
  Preferred vector width short:             8
  Preferred vector width int:             4
  Preferred vector width long:             2
  Preferred vector width float:             4
  Preferred vector width double:         2
  Native vector width char:             16
  Native vector width short:             8
  Native vector width int:             4
  Native vector width long:             2
  Native vector width float:             4
  Native vector width double:             2
  Max clock frequency:                 3066Mhz
  Address bits:                     64
  Max memory allocation:             2096975872
  Image support:                 Yes
  Max number of images read arguments:         128
  Max number of images write arguments:         64
  Max image 2D width:                 8192
  Max image 2D height:                 8192
  Max image 3D width:                 2048
  Max image 3D height:                 2048
  Max image 3D depth:                 2048
  Max samplers within kernel:             16
  Max size of kernel argument:             4096
  Alignment (bits) of base address:         1024
  Minimum alignment (bytes) for any datatype:     128
  Single precision floating point capability
    Denorms:                     Yes
    Quiet NaNs:                     Yes
    Round to nearest even:             Yes
    Round to zero:                 Yes
    Round to +ve and infinity:             Yes
    IEEE754-2008 fused multiply-add:         Yes
  Cache type:                     Read/Write
  Cache line size:                 64
  Cache size:                     16384
  Global memory size:                 2096975872
  Constant buffer size:                 65536
  Max number of constant args:             8
  Local memory type:                 Global
  Local memory size:                 32768
  Max pipe arguments:                 16
  Max pipe active reservations:             16
  Max pipe packet size:                 2096975872
  Max global variable size:             1879048192
  Max global variable preferred total size:     1879048192
  Max read/write image args:             64
  Max on device events:                 0
  Queue on device max size:             0
  Max on device queues:                 0
  Queue on device preferred size:         0
  SVM capabilities:                 
    Coarse grain buffer:             No
    Fine grain buffer:                 No
    Fine grain system:                 No
    Atomics:                     No
  Preferred platform atomic alignment:         0
  Preferred global atomic alignment:         0
  Preferred local atomic alignment:         0
  Kernel Preferred work group size multiple:     1
  Error correction support:             0
  Unified memory for Host and Device:         1
  Profiling timer resolution:             1
  Device endianess:                 Little
  Available:                     Yes
  Compiler available:                 Yes
  Execution capabilities:                 
    Execute OpenCL kernels:             Yes
    Execute native function:             Yes
  Queue on Host properties:                 
    Out-of-Order:                 No
    Profiling :                     Yes
  Queue on Device properties:                 
    Out-of-Order:                 No
    Profiling :                     No
  Platform ID:                     0x7f6ffde01430
  Name:                         Intel(R) Pentium(R) 4 CPU 3.06GHz
  Vendor:                     GenuineIntel
  Device OpenCL C version:             OpenCL C 1.2 
  Driver version:                 1800.11 (sse2)
  Profile:                     FULL_PROFILE
  Version:                     OpenCL 1.2 AMD-APP (1800.11)
  Extensions:                     cl_khr_fp64 cl_amd_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_gl_sharing cl_ext_device_fission cl_amd_device_attribute_query cl_amd_vec3 cl_amd_printf cl_amd_media_ops cl_amd_media_ops2 cl_amd_popcnt cl_khr_spir cl_khr_gl_event 

I intend to upgrade to a more capable AMD GPU for my research so I need to have OpenCL 2.0 support.

[ekondis]

I also tried the runtime libraries (libamdocl64.so and libamdocl12cl64.so) on three other systems all running Ubuntu 14.04 64bit. The two having old CPUs, Pentium D and AthlonX2 64, did not expose the respective CPU OpenCL devices. In contrast, the more recent Core i7-2500 normally exposed the OpenCL device without problems. Therefore I suspect that some additional instruction set requirements have been set for the Crimson driver.

If this is true then what are these requirements for the runtime to expose the CPU as an OpenCL device?

And secondly, is this issue related with my problem of having an R7-260X GPU recognised as an OpenCL 1.2 device and not as 2.0?

[dipak · Accepted solution]

Hi,

Thanks for reporting the issue.

Just confirmed from the driver team that this clinfo issue is indeed due to shortage of RAM. As I've come to know that, on Crimson Driver, neither CPU, nor iGPU report OpenCL support on 2GB systems and dGPU is reported as only OCL 1.2 capable regardless the ASIC family. The team is in process of including this information in the release note. Sorry for the inconvenience.

Regards,

[ekondis]

Thank you for your response.

So the what it the minimum RAM size of a system in order to support OpenCL 2.0? 4GB or 3GB?

[dipak]

It's 4GB.

[ekondis]

After adding 2GB more RAM I can confirm that a total of 4GB RAM enable both OpenCL 2.0 on R7 260X and CPU OpenCL support.