Re:Для них внезапно нужен работающий OpenCL. На который даже у линукс-сообщества нет ресурсов

ubuntu@host:~$ uname -a
Linux ubuntu 5.5.5-050505-lowlatency #202002191432 SMP PREEMPT Wed Feb 19 19:47:19 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux

ubuntu@host:~$ clinfo
Number of platforms:                 1
  Platform Profile:                 FULL_PROFILE
  Platform Version:                 OpenCL 2.1 AMD-APP (3052.0)
  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:                     Vega 10 XT [Radeon RX Vega 64]
  Device Topology:                 PCI[ B#9, D#0, F#0 ]
  Max compute units:                 56
  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:                 1301Mhz
  Address bits:                     64
  Max memory allocation:             7287183769
  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:             26751
  Max size of kernel argument:             1024
  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:                 8573157376
  Constant buffer size:                 7287183769
  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:                 2992216473
  Max global variable size:             7287183769
  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:                     0x7fcfc4305d50
  Name:                         gfx900
  Vendor:                     Advanced Micro Devices, Inc.
  Device OpenCL C version:             OpenCL C 2.0
  Driver version:                 3052.0 (HSA1.1,LC)
  Profile:                     FULL_PROFILE
  Version:                     OpenCL 2.0
  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

Более того, Linux еще и HPC могет:

ubuntu@host:~$ rocminfo
ROCk module is loaded
ubuntu is member of video group
=====================    
HSA System Attributes    
=====================    
Runtime Version:         1.1
System Timestamp Freq.:  1000.000000MHz
Sig. Max Wait Duration:  18446744073709551615 (0xFFFFFFFFFFFFFFFF) (timestamp count)
Machine Model:           LARGE                              
System Endianness:       LITTLE                            

==========              
HSA Agents              
==========              
*******                  
Agent 1                  
*******                  
  Name:                    AMD Ryzen 7 2700 Eight-Core Processor
  Marketing Name:          AMD Ryzen 7 2700 Eight-Core Processor
  Vendor Name:             CPU                                
  Feature:                 None specified                    
  Profile:                 FULL_PROFILE                      
  Float Round Mode:        NEAR                              
  Max Queue Number:        0(0x0)                            
  Queue Min Size:          0(0x0)                            
  Queue Max Size:          0(0x0)                            
  Queue Type:              MULTI                              
  Node:                    0                                  
  Device Type:             CPU                                
  Cache Info:              
    L1:                      32768(0x8000) KB                  
  Chip ID:                 0(0x0)                            
  Cacheline Size:          64(0x40)                          
  Max Clock Freq. (MHz):   3200                              
  BDFID:                   0                                  
  Internal Node ID:        0                                  
  Compute Unit:            16                                
  SIMDs per CU:            0                                  
  Shader Engines:          0                                  
  Shader Arrs. per Eng.:   0                                  
  WatchPts on Addr. Ranges:1                                  
  Features:                None
  Pool Info:              
    Pool 1                  
      Segment:                 GLOBAL; FLAGS: KERNARG, FINE GRAINED
      Size:                    49319512(0x2f08e58) KB            
      Allocatable:             TRUE                              
      Alloc Granule:           4KB                                
      Alloc Alignment:         4KB                                
      Acessible by all:        TRUE                              
    Pool 2                  
      Segment:                 GLOBAL; FLAGS: COARSE GRAINED      
      Size:                    49319512(0x2f08e58) KB            
      Allocatable:             TRUE                              
      Alloc Granule:           4KB                                
      Alloc Alignment:         4KB                                
      Acessible by all:        TRUE                              
  ISA Info:                
    N/A                      
*******                  
Agent 2                  
*******                  
  Name:                    gfx900                            
  Marketing Name:          Vega 10 XT [Radeon RX Vega 64]    
  Vendor Name:             AMD                                
  Feature:                 KERNEL_DISPATCH                    
  Profile:                 BASE_PROFILE                      
  Float Round Mode:        NEAR                              
  Max Queue Number:        128(0x80)                          
  Queue Min Size:          4096(0x1000)                      
  Queue Max Size:          131072(0x20000)                    
  Queue Type:              MULTI                              
  Node:                    1                                  
  Device Type:             GPU                                
  Cache Info:              
    L1:                      16(0x10) KB                        
  Chip ID:                 26751(0x687f)                      
  Cacheline Size:          64(0x40)                          
  Max Clock Freq. (MHz):   1301                              
  BDFID:                   2304                              
  Internal Node ID:        1                                  
  Compute Unit:            56                                
  SIMDs per CU:            4                                  
  Shader Engines:          4                                  
  Shader Arrs. per Eng.:   1                                  
  WatchPts on Addr. Ranges:4                                  
  Features:                KERNEL_DISPATCH
  Fast F16 Operation:      FALSE                              
  Wavefront Size:          64(0x40)                          
  Workgroup Max Size:      1024(0x400)                        
  Workgroup Max Size per Dimension:
    x                        1024(0x400)                        
    y                        1024(0x400)                        
    z                        1024(0x400)                        
  Max Waves Per CU:        40(0x28)                          
  Max Work-item Per CU:    2560(0xa00)                        
  Grid Max Size:           4294967295(0xffffffff)            
  Grid Max Size per Dimension:
    x                        4294967295(0xffffffff)            
    y                        4294967295(0xffffffff)            
    z                        4294967295(0xffffffff)            
  Max fbarriers/Workgrp:   32                                
  Pool Info:              
    Pool 1                  
      Segment:                 GLOBAL; FLAGS: COARSE GRAINED      
      Size:                    8372224(0x7fc000) KB              
      Allocatable:             TRUE                              
      Alloc Granule:           4KB                                
      Alloc Alignment:         4KB                                
      Acessible by all:        FALSE                              
    Pool 2                  
      Segment:                 GROUP                              
      Size:                    64(0x40) KB                        
      Allocatable:             FALSE                              
      Alloc Granule:           0KB                                
      Alloc Alignment:         0KB                                
      Acessible by all:        FALSE                              
  ISA Info:                
    ISA 1                    
      Name:                    amdgcn-amd-amdhsa--gfx900          
      Machine Models:          HSA_MACHINE_MODEL_LARGE            
      Profiles:                HSA_PROFILE_BASE                  
      Default Rounding Mode:   NEAR                              
      Default Rounding Mode:   NEAR                              
      Fast f16:                TRUE                              
      Workgroup Max Size:      1024(0x400)                        
      Workgroup Max Size per Dimension:
        x                        1024(0x400)                        
        y                        1024(0x400)                        
        z                        1024(0x400)                        
      Grid Max Size:           4294967295(0xffffffff)            
      Grid Max Size per Dimension:
        x                        4294967295(0xffffffff)            
        y                        4294967295(0xffffffff)            
        z                        4294967295(0xffffffff)            
      FBarrier Max Size:       32                                
*** Done ***            
...

Пилить вам шура свою бздю еще лет 100 пока она догонит сегодняшний Linux