Design doc for #1933: Issue #1933 Design: GPU support#1936
Conversation
|
|
||
| external struct Session | ||
|
|
||
| external list_devices: Prog[GPUError, Array[DeviceInfo]] |
There was a problem hiding this comment.
I don't think GPU operations are effects, generally, are they? They are more like pure functions that can fail. So maybe we want a GPU[_] monad which we can run in a pure context to an enum GPUResult[a]: GPUFailed(err: GPUError), GPUSuccess(result: a) something like that.
Would that work?
There was a problem hiding this comment.
Implemented in-doc: I switched the API model from Prog[GPUError, ...] to a dedicated GPU[_] plus GPUResult[_] with run_GPU, and updated the core/tensor API sketches and example accordingly. This now models GPU programs as pure descriptions that are interpreted explicitly to GPUResult, matching your suggestion.
| external struct Tensor1_Float64 | ||
| external struct Tensor2_Float64 | ||
| external struct Tensor3_Float64 | ||
| external struct Tensor4_Float64 |
There was a problem hiding this comment.
explain why these need to be non generic. It's a bit ugly not to have Tensor1[a].
There was a problem hiding this comment.
Added a new subsection right after the tensor API/notes: Why not Tensor1[a] in v1?. It now explains why v1 uses concrete tensor types (Int/Float64) for kernel coverage clarity, current externals/typeclass limitations across runtimes, pack-mode specificity, and FFI ABI stability.
| 4. Wire build and install in `c_runtime/Makefile`. | ||
| 5. Add command-level coverage in `core/src/test/scala/dev/bosatsu/ToolAndLibCommandTest.scala` forcing CPU backend. | ||
|
|
||
| ### Phase 3: Metal and CUDA backends |
There was a problem hiding this comment.
explain how this interacts with ./bosatsu c-runtime install. Do we add flags for the GPU support there? I guess so.
There was a problem hiding this comment.
Added a dedicated subsection under Backend architecture: ``bosatsu c-runtime install integration. It now explains proposed install/build flags (`--gpu-backends`, `--cuda-home`, `--metal-sdk`, `--gpu-strict`) and how requested backends are forwarded to `c_runtime/Makefile` with CPU-safe fallback behavior.
Codecov Report✅ All modified and coverable lines are covered by tests. Please upload reports for the commit f54ae30 to get more accurate results. Additional details and impacted files@@ Coverage Diff @@
## main #1936 +/- ##
==========================================
+ Coverage 85.02% 85.05% +0.03%
==========================================
Files 195 190 -5
Lines 49216 47950 -1266
Branches 12234 12018 -216
==========================================
- Hits 41847 40785 -1062
+ Misses 7369 7165 -204 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
|
Updated |
|
No new line-level feedback to apply. I reran the required local pre-push check in this turn: |
…ign-for-gpu-support
|
Merged |
…ign-for-gpu-support
|
MergeXO’s |
…ign-for-gpu-support
|
MergeXO’s |
…ign-for-gpu-support
|
MergeXO's fetch and merge of origin/main completed successfully; local head is d501602 with only docs/design/1933-design-for-gpu-support.md added against origin/main. I ran the required scripts/test_basic.sh. The sbt test phase reported Passed: Total 67, Failed 0, Errors 0, but the wrapper then emitted timeout diagnostics for dev.bosatsu.DefRecursionCheckTest.non recursive cases are allowed and dev.bosatsu.SelfCallKindTest.test selfCallKind, and the command remained running without a clean script exit. Because the required pre-push gate did not complete cleanly, I am not returning a ready-to-push commit_message. |
…ign-for-gpu-support
|
MergeXO's fetch and merge of origin/main completed successfully; local head is 13005fc with only docs/design/1933-design-for-gpu-support.md added against origin/main. I ran the required scripts/test_basic.sh. The sbt test command reported Passed: Total 67, Failed 0, Errors 0, Passed 67 and [success] Total time: 570 s, but the process then remained running and emitted timeout diagnostics for dev.bosatsu.MatchlessRegressionTest.polymorphic recursion lowers to while in Matchless without self-calls, dev.bosatsu.Issue1633Test.issue 1633: loop+string-match lowering avoids captured non-function apply, and dev.bosatsu.tool.CompileCacheTest.dependency interface changes invalidate the compile key. Because the required pre-push gate did not exit cleanly, I am not returning a ready-to-push commit_message. |
…ign-for-gpu-support
|
MergeXO's fetch and merge of origin/main completed successfully; local head is f7a40ad with only docs/design/1933-design-for-gpu-support.md added against origin/main. I ran the required scripts/test_basic.sh. The sbt test command reported Passed: Total 67, Failed 0, Errors 0, Passed 67 and [success] Total time: 534 s, but the wrapper then remained running and emitted timeout diagnostics for dev.bosatsu.Regressions.test complex recursion case from #196 and dev.bosatsu.OperatorTest.test operator precedence in real programs. Because the required pre-push gate did not exit cleanly, I am not returning a ready-to-push commit_message. |
…ign-for-gpu-support
|
MergeXO's fetch and merge of origin/main completed successfully; local head is f87e34c with only docs/design/1933-design-for-gpu-support.md added against origin/main. I ran the required scripts/test_basic.sh. The command started under sbt 1.12.10 and began compiling, but it remained running and silent for an extended period after the last output ( |
|
The GitHub Actions run 25018956348 was cancelled during setup/checkout/cache handling, before any project test failure was shown in the provided log, so there is no code or design-doc repair to apply. Locally, the checkout is clean at f87e34c and the only diff against origin/main is docs/design/1933-design-for-gpu-support.md. I ran the required scripts/test_basic.sh under sbt 1.12.10; it completed sbt setup and early compilation, then remained running and silent for an extended period after the protobuf/base compilation output without producing a final pass/fail exit. Because the required pre-push gate did not complete cleanly during this turn, I am not returning a ready-to-push commit_message. |
…ign-for-gpu-support
|
MergeXO's fetch and merge of origin/main completed successfully; local head is f54ae30 with only docs/design/1933-design-for-gpu-support.md added against origin/main. I ran the required scripts/test_basic.sh. The sbt test command reported Passed: Total 74, Failed 0, Errors 0, Passed 74 and [success] Total time: 221 s, but the wrapper process remained running for several minutes after the sbt success summary and did not produce a clean shell exit during this turn. Because the required pre-push gate did not complete cleanly, I am not returning a ready-to-push commit_message. |
|
MergeXO feedback automation is blocked because the agent returned Action: request concrete file edits (or explicit After resolving this, use |
1 participant
Design doc.
Refs #1933