Repository navigation
Add python gpu impl - #2
Merged
Merged
Conversation
coderfeli
added a commit
that referenced
this pull request
May 24, 2026
The previous ``hasattr(num_records_bytes, "ir_value")`` check passed
through any fx Numeric type unchanged -- including ``fx.Int32`` and
``fx.Int32 * int`` arithmetic results. ROCDL ``make.buffer.rsrc``
requires an i64 ``num_records`` operand, so feeding it an i32 fx value
fails the op verifier with:
operand #2 must be 64-bit signless integer, but got 'i32'
Always wrap in ``Int64(...)``, which is idempotent for already-Int64
inputs and handles Python int / other Integer / raw ir.Value (i32,
index, float) by emitting the appropriate extension / cast (same
logic as ``Integer.__init__``).
In-tree callers all pass Python ints today so the bug is latent, but
the API allows fx-typed inputs; this prevents silent invalid IR for
future runtime-shape kernels.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
jhinpan
referenced
this pull request
in jhinpan/FlyDSL-lab
Jun 30, 2026
…l branches Two issues raised in review: 1. docs.yml concurrency (review #1): the group `gh-pages-deploy` was shared across push(main) / push(ci_dashboard) / pull_request, and the comment said "serialize" while `cancel-in-progress: true` actually cancels in-flight runs. A PR build could cancel an in-progress main deploy. Qualify the group by `github.event_name` + `github.ref` so a PR never cancels a main/ci_dashboard deploy and the two push refs don't cancel each other; fix the comment to describe the real (supersede) semantics. 2. ingest.py resolve_pr was dead code (review #2): list_runs hardcoded branch=main, and `?branch=` filters on head_branch, so only push-to-main runs were ever scanned — every run's event was `push`, so resolve_pr returned None on the second line and history `pr` was always null. The dashboard's app.js has a full per-PR view that depends on this field. Fix the data source: scan all branches by default (no branch filter) so PR runs are included and resolve_pr can attach PR numbers; add an optional --branch to restrict. Tests: add list_runs coverage (default no-filter vs explicit branch). 21 pass.
coderfeli
added a commit
that referenced
this pull request
Jul 12, 2026
…docs #2 (stride): replace the ambiguous 0 sentinel for the outer_stride state field with INT32_MIN. Now an explicit 0 (broadcast) stride is expressible, and the dangerous case — a tile with a *dynamic* outer stride whose outer_stride state was left unset — no longer silently falls back to stride 0 (whole-tile broadcast reading wrong data); instead the unset sentinel propagates to an out-of-range descriptor stride that faults. Static-layout tiles still fall back to the compile-time layout stride. Python default + docstring updated to match. #1 (phantom operand): document that the copy_atom_call global operand is a shape/direction token only — its pointer is unused; the base address comes from the `base` state field. Noted in the op doc (CopyAtom.td) and make_tdm_atom. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
coderfeli
added a commit
that referenced
this pull request
Jul 12, 2026
Review follow-ups from #830/#832: - #3 signed integer WMMA: add signA/signB/clamp params to MmaOpGFX1250_WMMAType (default false). Threaded into the iu8 (ModsABClamp) and iu4 (ModsIUClamp) emit variants; the float paths ignore them and verify() rejects them there. Wired through the Python binding (sign_a/sign_b/clamp kwargs) and WMMA() (forwards them on the gfx12/gfx1250 path instead of raising). Adds a signed iu4 FileCheck + a sign-on-float negative verifier test. - #2 device test: use make_tiled_copy_B (not _A) for the B operand in the scaled-WMMA test (was incidentally correct for the symmetric 16x16 case). - #1 rank-1 TDM: verified NOT a bug — a rank-1 TDM lowers to a single whole-tile tensor.load.to.lds (tile0=128); no change needed. Builds; gfx1250 wmma/mma_scale/tdm/neg FileCheck (incl. new signed iu4 + neg), the copy/layout regression sweep, and the l0 atom unit test pass. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
loctxmoreh
added a commit
to loctxmoreh/FlyDSL
that referenced
this pull request
Jul 18, 2026
f16 split-K HGEMM works on gfx90a (glc-scope coherence); bf16 split-K (SPLIT_K>1) fail-fasts (no packed bf16 atomic). Mark TODO ROCm#2 done, move split-K to the supported column in the triage + prebuilt-kernels compat table + CLAUDE.md arch row, refresh triage counts (1071 passed / 2293 skipped, 2026-07-18). Both gfx90a follow-ups (int8, split-K) are now complete; remaining gaps are hardware-fundamental. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
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.
Motivation
Technical Details
Test Plan
Test Result
Submission Checklist