Repository navigation
POCL: support BFloat16 kernels when PoCL accepts SPV_KHR_bfloat16 - #946
Merged
Merged
Conversation
On Julia 1.11+, `BFloat16` is `Core.BFloat16`, which lowers to LLVM's `bfloat`, so any kernel touching a `BFloat16` array failed GPUCompiler's validation. Query the device's SPIR-V extensions (cl_khr_spirv_queries) and, when PoCL lists SPV_KHR_bfloat16 (from pocl_standalone_jll 7.2.3+3), enable it together with GPUCompiler's `supports_bfloat16`. Arithmetic is widened to `float` by GPUCompiler 2.20; older PoCL builds keep failing cleanly with an `InvalidIRError`. Closes #921.
Contributor
Benchmark resultsCompared with c10ff8cf4, the merge base. Times are the minimum time of every benchmark. A change is unusual when it exceeds both three times the benchmark's typical change between consecutive runs on the default branch and a minimum per category (performance 3%, latency 5%, applications 3%, other 3%). A benchmark that starts or stops allocating is listed however few the allocations. A single run can still be a fluke, so rerun before drawing conclusions.
No unusual changes. All benchmarks
|
BFloat16s.jl 0.6.4 rounds them correctly; the reason left is that vectorized ones crash LLVM on Julia 1.11 on AVX512-BF16 hosts.
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #946 +/- ##
==========================================
- Coverage 83.46% 82.83% -0.64%
==========================================
Files 31 31
Lines 2571 2534 -37
==========================================
- Hits 2146 2099 -47
- Misses 425 435 +10 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
On Julia 1.11+,
BFloat16isCore.BFloat16, which Julia lowers to LLVM'sbfloat. The POCL back end never set GPUCompiler'ssupports_bfloat16, so every kernel touching aBFloat16array, even a copy, failed withInvalidIRError: unsupported use of bfloat value(#921).This queries the device's SPIR-V extensions (
CL_DEVICE_SPIRV_EXTENSIONS_KHR, fromcl_khr_spirv_queries) and, when PoCL listsSPV_KHR_bfloat16, enables that extension andsupports_bfloat16. The pieces this relies on:SPV_KHR_bfloat16, plus a fix forabs(INT_MIN)that Julia 1.13's integer toBFloat16conversions hit: [pocl] Accept SPV_KHR_bfloat16 SPIR-V on the CPU devices JuliaPackaging/Yggdrasil#15429, shipped as pocl_standalone_jll 7.2.3+3. The PoCL version stayed the same, so compat can't require that build; the run-time query handles it, and older builds keep failing cleanly with anInvalidIRError.bfloatarithmetic tofloatfor SPIR-V (SPIR-V: work around missing bfloat arithmetic legalization in the back-ends GPUCompiler.jl#1076, in 2.20), as PoCL accepts no bfloat16 arithmetic extension. Hence the compat bump.Tests cover #921's copy kernel, arithmetic and conversions from
Float32/Float64/Int32/Int64/UInt64, checked against correctly rounded references. The references are computed inFloat32(arithmetic) andBigFloat(conversions) rather than with the host'sBFloat16operations: on Julia 1.11, vectorizedBFloat16code on the host either crashes LLVM (Cannot select ... X86ISD::VFPROUNDon AVX512-BF16 CPUs) or, forcopysign, gives wrong results. On Julia 1.10,BFloat16is still BFloat16s.jl's 16-bit primitive type and the tests pass as before.Closes #921