Mint i AMDGPU-PRO, OPENCL

[Jenot]

0

Sztuka nie udała się na Mincie, ale na Ubuntu 18.04, które ma teraz kernel 4.18 (niewspółpracujący z AMDGPU-PRO, przynajmniej w teorii), udało mi się wydłubać biblioteki OpenCL z debów AMDGPU-PRO 18.50. clinfo pokazuje:

Kod:

root@ubu-jenot:~/Pobrane/opencllib# clinfo
Number of platforms: 1
 Platform Profile: FULL_PROFILE
 Platform Version: OpenCL 2.1 AMD-APP (2766.4)
 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: AMD Accelerated Parallel Processing
Number of devices: 1
 Device Type: CL_DEVICE_TYPE_GPU
 Vendor ID: 1002h
 Board name: Radeon RX Vega
 Device Topology: PCI[ B#40, D#0, F#0 ]
 Max compute units: 64
 Max work items dimensions: 3
   Max work items[0]: 1024
   Max work items[1]: 1024
   Max work items[2]: 1024
 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: 1630Mhz
 Address bits: 64
 Max memory allocation: 4244635648
 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: 8573157376
 Constant buffer size: 4244635648
 Max number of constant args: 8
 Local memory type: Scratchpad
 Local memory size: 65536
 Max pipe arguments: 16
 Max pipe active reservations: 16
 Max pipe packet size: 4244635648
 Max global variable size: 3820172032
 Max global variable preferred total size: 8573157376
 Max read/write image args: 64
 Max on device events: 1024
 Queue on device max size: 8388608
 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: 0x7f8ffdfc9710
 Name: gfx900
 Vendor: Advanced Micro Devices, Inc.
 Device OpenCL C version: OpenCL C 2.0
 Driver version: 2766.4 (PAL,HSAIL)
 Profile: FULL_PROFILE
 Version: OpenCL 2.0 AMD-APP (2766.4)
 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_fp16 cl_khr_gl_sharing cl_khr_gl_depth_images 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 cl_khr_mipmap_image cl_khr_mipmap_image_writes

Czyli niby OK, zaraz sprawdzę na jakimś większym projekcie czy blender daje radę.

Edit: Wszystko działa. Teraz upgrade jądra mi niestraszny
Edit2: Gwoli wyjaśnienia - sterownik pozostał opensource'owy - ten który jestdomyślnie w distro, doinstalowane są tylko biblioteki OpenCL z AMDGPU-PRO. Działa OpenGL, OpenCL, po wydłubaniu biblioteki vulkana działa też i vulkan.