Implement a CPU backend using POCL

[vchuravy]

TODO:

• POCL standalone build to remove dependency on OpenCL_jll
• LLVM 20 fixes pocl/pocl#1784 for support of exception handling

[github-actions]

Benchmark Results

	main	77f1ee0...	main/77f1ee03942cfc...
saxpy/default/Float16/1024	0.735 ± 0.0066 μs	0.0565 ± 0.026 ms	0.013
saxpy/default/Float16/1048576	0.177 ± 0.0072 ms	0.888 ± 0.023 ms	0.199
saxpy/default/Float16/16384	3.33 ± 0.031 μs	0.0636 ± 0.028 ms	0.0524
saxpy/default/Float16/2048	0.909 ± 0.01 μs	0.0497 ± 0.023 ms	0.0183
saxpy/default/Float16/256	0.592 ± 0.005 μs	0.0594 ± 0.026 ms	0.00997
saxpy/default/Float16/262144	0.0443 ± 0.00048 ms	0.271 ± 0.025 ms	0.164
saxpy/default/Float16/32768	6.02 ± 0.084 μs	0.0755 ± 0.028 ms	0.0797
saxpy/default/Float16/4096	1.31 ± 0.026 μs	0.0664 ± 0.026 ms	0.0197
saxpy/default/Float16/512	0.651 ± 0.0061 μs	0.0539 ± 0.026 ms	0.0121
saxpy/default/Float16/64	0.563 ± 0.0049 μs	0.0544 ± 0.028 ms	0.0104
saxpy/default/Float16/65536	11.6 ± 0.14 μs	0.109 ± 0.027 ms	0.107
saxpy/default/Float32/1024	0.655 ± 0.01 μs	0.0595 ± 0.026 ms	0.011
saxpy/default/Float32/1048576	0.231 ± 0.022 ms	0.473 ± 0.025 ms	0.489
saxpy/default/Float32/16384	2.79 ± 0.14 μs	0.0549 ± 0.025 ms	0.0508
saxpy/default/Float32/2048	0.776 ± 0.062 μs	0.0496 ± 0.023 ms	0.0156
saxpy/default/Float32/256	0.581 ± 0.0075 μs	0.0583 ± 0.027 ms	0.00997
saxpy/default/Float32/262144	0.0553 ± 0.0034 ms	0.163 ± 0.035 ms	0.34
saxpy/default/Float32/32768	5.36 ± 0.22 μs	0.0602 ± 0.027 ms	0.089
saxpy/default/Float32/4096	1.16 ± 0.1 μs	0.0605 ± 0.024 ms	0.0191
saxpy/default/Float32/512	0.618 ± 0.0093 μs	0.0583 ± 0.026 ms	0.0106
saxpy/default/Float32/64	0.571 ± 0.0069 μs	0.0583 ± 0.027 ms	0.00978
saxpy/default/Float32/65536	12.4 ± 0.67 μs	0.075 ± 0.029 ms	0.165
saxpy/default/Float64/1024	0.758 ± 0.035 μs	0.0592 ± 0.026 ms	0.0128
saxpy/default/Float64/1048576	0.505 ± 0.046 ms	0.511 ± 0.061 ms	0.989
saxpy/default/Float64/16384	5.34 ± 0.52 μs	0.0564 ± 0.026 ms	0.0945
saxpy/default/Float64/2048	1.14 ± 0.077 μs	0.0514 ± 0.024 ms	0.0222
saxpy/default/Float64/256	0.588 ± 0.0062 μs	0.055 ± 0.027 ms	0.0107
saxpy/default/Float64/262144	0.114 ± 0.012 ms	0.172 ± 0.03 ms	0.663
saxpy/default/Float64/32768	12.7 ± 1.1 μs	0.0637 ± 0.026 ms	0.199
saxpy/default/Float64/4096	1.7 ± 0.16 μs	0.0602 ± 0.024 ms	0.0282
saxpy/default/Float64/512	0.639 ± 0.01 μs	0.0534 ± 0.027 ms	0.012
saxpy/default/Float64/64	0.567 ± 0.0055 μs	0.0612 ± 0.027 ms	0.00925
saxpy/default/Float64/65536	28.1 ± 2.5 μs	0.0842 ± 0.026 ms	0.334
saxpy/static workgroup=(1024,)/Float16/1024	2.22 ± 0.025 μs	0.0482 ± 0.026 ms	0.0461
saxpy/static workgroup=(1024,)/Float16/1048576	0.157 ± 0.0036 ms	0.898 ± 0.026 ms	0.175
saxpy/static workgroup=(1024,)/Float16/16384	4.48 ± 0.089 μs	0.0609 ± 0.026 ms	0.0735
saxpy/static workgroup=(1024,)/Float16/2048	2.4 ± 0.026 μs	0.0571 ± 0.023 ms	0.042
saxpy/static workgroup=(1024,)/Float16/256	2.81 ± 0.031 μs	0.0554 ± 0.026 ms	0.0507
saxpy/static workgroup=(1024,)/Float16/262144	0.0419 ± 0.0011 ms	0.271 ± 0.027 ms	0.155
saxpy/static workgroup=(1024,)/Float16/32768	6.91 ± 0.17 μs	0.0732 ± 0.025 ms	0.0944
saxpy/static workgroup=(1024,)/Float16/4096	2.72 ± 0.035 μs	0.0613 ± 0.026 ms	0.0444
saxpy/static workgroup=(1024,)/Float16/512	3.26 ± 0.033 μs	0.0511 ± 0.026 ms	0.0638
saxpy/static workgroup=(1024,)/Float16/64	2.52 ± 0.22 μs	0.0524 ± 0.026 ms	0.0481
saxpy/static workgroup=(1024,)/Float16/65536	12.6 ± 0.32 μs	0.104 ± 0.026 ms	0.121
saxpy/static workgroup=(1024,)/Float32/1024	2.24 ± 0.036 μs	0.0424 ± 0.026 ms	0.0529
saxpy/static workgroup=(1024,)/Float32/1048576	0.233 ± 0.022 ms	0.459 ± 0.027 ms	0.507
saxpy/static workgroup=(1024,)/Float32/16384	4.48 ± 0.46 μs	0.0523 ± 0.024 ms	0.0857
saxpy/static workgroup=(1024,)/Float32/2048	2.4 ± 0.06 μs	0.0448 ± 0.023 ms	0.0534
saxpy/static workgroup=(1024,)/Float32/256	2.72 ± 0.051 μs	0.0468 ± 0.026 ms	0.058
saxpy/static workgroup=(1024,)/Float32/262144	0.0483 ± 0.0047 ms	0.158 ± 0.034 ms	0.306
saxpy/static workgroup=(1024,)/Float32/32768	7.64 ± 0.52 μs	0.0581 ± 0.026 ms	0.131
saxpy/static workgroup=(1024,)/Float32/4096	2.68 ± 0.076 μs	0.0521 ± 0.025 ms	0.0513
saxpy/static workgroup=(1024,)/Float32/512	2.74 ± 0.033 μs	0.0542 ± 0.026 ms	0.0504
saxpy/static workgroup=(1024,)/Float32/64	2.75 ± 5.5 μs	0.0542 ± 0.026 ms	0.0507
saxpy/static workgroup=(1024,)/Float32/65536	14.5 ± 1.3 μs	0.0729 ± 0.028 ms	0.199
saxpy/static workgroup=(1024,)/Float64/1024	2.33 ± 0.068 μs	0.0557 ± 0.026 ms	0.0419
saxpy/static workgroup=(1024,)/Float64/1048576	0.499 ± 0.049 ms	0.501 ± 0.051 ms	0.997
saxpy/static workgroup=(1024,)/Float64/16384	7.42 ± 0.4 μs	0.054 ± 0.025 ms	0.137
saxpy/static workgroup=(1024,)/Float64/2048	2.6 ± 0.065 μs	0.0484 ± 0.024 ms	0.0536
saxpy/static workgroup=(1024,)/Float64/256	2.64 ± 0.051 μs	0.0529 ± 0.025 ms	0.0499
saxpy/static workgroup=(1024,)/Float64/262144	0.0924 ± 0.008 ms	0.17 ± 0.031 ms	0.544
saxpy/static workgroup=(1024,)/Float64/32768	14.5 ± 1.5 μs	0.0616 ± 0.025 ms	0.235
saxpy/static workgroup=(1024,)/Float64/4096	3.16 ± 0.24 μs	0.0549 ± 0.025 ms	0.0575
saxpy/static workgroup=(1024,)/Float64/512	2.66 ± 0.074 μs	0.0567 ± 0.026 ms	0.047
saxpy/static workgroup=(1024,)/Float64/64	2.6 ± 0.06 μs	0.0559 ± 0.026 ms	0.0465
saxpy/static workgroup=(1024,)/Float64/65536	25.6 ± 2 μs	0.0808 ± 0.027 ms	0.317
time_to_load	0.316 ± 0.0096 s	1.09 ± 0.0058 s	0.289

Benchmark Plots

A plot of the benchmark results have been uploaded as an artifact to the workflow run for this PR.
Go to "Actions"->"Benchmark a pull request"->[the most recent run]->"Artifacts" (at the bottom).

[codecov]

Codecov Report

Attention: Patch coverage is 0% with 871 lines in your changes missing coverage. Please review.

Project coverage is 0.00%. Comparing base (b435bb2) to head (77f1ee0).

Files with missing lines	Patch %	Lines
src/pocl/nanoOpenCL.jl	0.00%	520 Missing ⚠️
src/pocl/device/array.jl	0.00%	101 Missing ⚠️
src/pocl/backend.jl	0.00%	93 Missing ⚠️
src/pocl/compiler/execution.jl	0.00%	43 Missing ⚠️
src/pocl/compiler/compilation.jl	0.00%	32 Missing ⚠️
src/pocl/device/quirks.jl	0.00%	24 Missing ⚠️
src/pocl/compiler/reflection.jl	0.00%	23 Missing ⚠️
src/pocl/pocl.jl	0.00%	20 Missing ⚠️
src/pocl/device/runtime.jl	0.00%	6 Missing ⚠️
src/nditeration.jl	0.00%	5 Missing ⚠️
... and 2 more

Additional details and impacted files

@@          Coverage Diff           @@
##            main    #556    +/-   ##
======================================
  Coverage   0.00%   0.00%            
======================================
  Files         12      21     +9     
  Lines        777    1503   +726     
======================================
- Misses       777    1503   +726     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[vchuravy]

• KernelIntrinsics #562
• Implement a CPU backend using POCL #556 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.