cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

FangQ
Adept I

OpenCL CL_DEVICE_SIMD_PER_COMPUTE_UNIT_AMD returns 0 in amdgpu-pro 17.70

my opencl code needs to use the stream processor number to estimate a default workgroup/workitem configurations. Using CL_DEVICE_MAX_COMPUTE_UNITS and CL_DEVICE_SIMD_PER_COMPUTE_UNIT_AMD, I was able to estimate the core counts in AMD GPUs, but I found the CL_DEVICE_SIMD_PER_COMPUTE_UNIT_AMD call returns 0 in the latest Linux driver (amdgpu-pro 17.70), here is the code

mcxcl/mcx_host.cpp at master · fangq/mcxcl · GitHub

I am wondering if this is expected behavior. If CL_DEVICE_SIMD_PER_COMPUTE_UNIT_AMD is no longer valid, what other approach I can use to estimate core numbers?

0 Likes
7 Replies
dipak
Big Boss

Hi,

Thanks for reporting this. AFAIK, OCL run-time should support this query as long as cl_amd_device_attribute_query is reported. If the extension was reported, but the value was still 0, then there might be a problem with the driver. Please share the clinfo output and setup details.

Btw, did you observe the problem with earlier versions of amdgpu-pro or just for this version only?

Regards,

0 Likes

I did not have this issue with amdgpu-pro 16.30 on another Ubuntu 14.04 box, which has a R9 nano and RX 480.

A correction, the problematic amdgpu-pro driver is 17.40, the OS is Ubuntu 16.04. I uninstalled the amdgpu-pro driver yesterday and installed rocm 1.6 (which has a module rocm-amdgpu-pro with version 17.40-492261), I can see the same problem where CL_DEVICE_SIMD_PER_COMPUTE_UNIT_AMD returns 0.

the clinfo output is attached below:

fangq@pangu:$ clinfo

Number of platforms                               1

  Platform Name                                   AMD Accelerated Parallel Processing

  Platform Vendor                                 Advanced Micro Devices, Inc.

  Platform Version                                OpenCL 2.0 AMD-APP (2508.0)

  Platform Profile                                FULL_PROFILE

  Platform Extensions                             cl_khr_icd cl_amd_event_callback

  Platform Extensions function suffix             AMD

  Platform Name                                   AMD Accelerated Parallel Processing

Number of devices                                 1

  Device Name                                     gfx900

  Device Vendor                                   Advanced Micro Devices, Inc.

  Device Vendor ID                                0x1002

  Device Version                                  OpenCL 1.2

  Driver Version                                  1.1 (HSA,LC)

  Device OpenCL C Version                         OpenCL C 2.0

  Device Type                                     GPU

  Device Profile                                  FULL_PROFILE

  Max compute units                               64

  Max clock frequency                             1630MHz

  Device Partition                                (core)

    Max number of sub-devices                     64

    Supported partition types                     none specified

  Max work item dimensions                        3

  Max work item sizes                             1024x1024x1024

  Max work group size                             256

  Preferred work group size multiple              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                                     No

    Infinity and NANs                             No

    Round to nearest                              No

    Round to zero                                 No

    Round to infinity                             No

    IEEE754-2008 fused multiply-add               No

    Support is emulated in software               No

    Correctly-rounded divide and sqrt operations  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

    Correctly-rounded divide and sqrt operations  No

  Address bits                                    64, Little-Endian

  Global memory size                              8573157376 (7.984GiB)

  Error Correction support                        No

  Max memory allocation                           7287183769 (6.787GiB)

  Unified memory for Host and Device              No

  Minimum alignment for any data type             128 bytes

  Alignment of base address                       1024 bits (128 bytes)

  Global Memory cache type                        Read/Write

  Global Memory cache size                        16384

  Global Memory cache line                        64 bytes

  Image support                                   Yes

    Max number of samplers per kernel             26751

    Max size for 1D images from buffer            65536 pixels

    Max 1D or 2D image array size                 2048 images

    Max 2D image size                             16384x16384 pixels

    Max 3D image size                             2048x2048x2048 pixels

    Max number of read image args                 128

    Max number of write image args                8

  Local memory type                               Local

  Local memory size                               65536 (64KiB)

  Max constant buffer size                        7287183769 (6.787GiB)

  Max number of constant args                     8

  Max size of kernel argument                     1024

  Queue properties                              

    Out-of-order execution                        No

    Profiling                                     Yes

  Prefer user sync for interop                    Yes

  Profiling timer resolution                      1ns

  Execution capabilities                        

    Run OpenCL kernels                            Yes

    Run native kernels                            No

  printf() buffer size                            4194304 (4MiB)

  Built-in kernels                              

  Device Available                                Yes

  Compiler Available                              Yes

  Linker Available                                Yes

  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_media_ops cl_amd_media_ops2 cl_khr_subgroups cl_khr_depth_images cl_amd_liquid_flash cl_amd_copy_buffer_p2p

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_CPU)  No devices found in platform

  clCreateContextFromType(NULL, CL_DEVICE_TYPE_GPU)  Success (1)

    Platform Name                                 AMD Accelerated Parallel Processing

    Device Name                                   gfx900

  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                                   gfx900

ICD loader properties

  ICD loader Name                                 OpenCL ICD Loader

  ICD loader Vendor                               OCL Icd free software

  ICD loader Version                              2.2.8

  ICD loader Profile                              OpenCL 1.2

NOTE: your OpenCL library declares to support OpenCL 1.2,

but it seems to support up to OpenCL 2.1 too.

0 Likes

Thanks for sharing the above details.

As per the clinfo output, it looks like "cl_amd_device_attribute_query", which provides a means to query AMD specific device attributes, is missing from the device extensions. That might be the reason for returning 0 against CL_DEVICE_SIMD_PER_COMPUTE_UNIT_AMD. This extension should be checked before querying any such AMD specific device parameters.  Please check clinfo output of the earlier drivers where the query worked. I think, the extension was present there.

Regards,

0 Likes

sorry for the late reply. yes, I was able to see the cl_amd_device_attribute_query extension on a Ubuntu box with amdgpu-pro 16.30 as expected.

so does this mean the latest 17.40 amdgpu driver has a bug?

0 Likes

Technically, we can't say it as a bug because the required extension, which enables to query device specific attributes, was not reported by amdgpu-pro 17.40. If the extension was reported, but the value was still 0, then there might be a problem with the driver.

0 Likes

if it is not amdgpu-pro's bug, then what other module is responsible for enabling this extension?

0 Likes

First of all, let me clear my earlier point. cl_amd_device_attribute_query is a vendor specific optional extension, not a required one. So, it would be inappropriate to mark this as a bug if the optional extension is not reported. At this moment, I can't say the reason for not reporting the extension and/or when it will be enabled again. AFAIK, OCL runtime reports it as supported by the base driver. I'll check if I can get more information about its current status.

0 Likes