faster devectorizer with one line

* external benchmark schedule in 5 sec (codex slop)

* prune

* real?

* delete

.github/workflows/sticky_comment.py

@@ -1,34 +0,0 @@
-#!/usr/bin/env python3
-# Sticky PR comment via the REST API: find an existing comment containing MARKER and PATCH it, or POST a new one.
-# Works on GitHub and Gitea (stdlib only, replaces marocchino/sticky-pull-request-comment which needs GraphQL).
-# Env vars: GITHUB_TOKEN, GITHUB_API_URL, GITHUB_REPOSITORY (set by the runner), PR_NUMBER, MARKER, and BODY_FILE or MESSAGE.
-import json, os, sys, urllib.request
-
-api, repo = os.environ["GITHUB_API_URL"], os.environ["GITHUB_REPOSITORY"]
-pr, marker = os.environ["PR_NUMBER"], os.environ["MARKER"]
-body = open(os.environ["BODY_FILE"]).read() if os.environ.get("BODY_FILE") else os.environ["MESSAGE"]
-
-if not body.strip():
-  print("comment body is empty, not posting")
-  sys.exit(0)
-
-def req(url, method="GET", payload=None):
-  r = urllib.request.Request(url, data=None if payload is None else json.dumps(payload).encode(), method=method,
-    headers={"Authorization": f"token {os.environ['GITHUB_TOKEN']}", "Accept": "application/json", "Content-Type": "application/json"})
-  return json.load(urllib.request.urlopen(r))
-
-# find the latest sticky comment (paginate, 100 comments per page)
-existing, page = None, 1
-while True:
-  comments = req(f"{api}/repos/{repo}/issues/{pr}/comments?per_page=100&page={page}")
-  stickies = [c for c in comments if marker in (c.get("body") or "")]
-  if stickies: existing = stickies[-1]
-  if not comments or len(comments) < 100: break
-  page += 1
-
-if existing is not None and existing["body"] == body:
-  print("comment is already up to date")
-  sys.exit(0)
-url = f"{api}/repos/{repo}/issues/comments/{existing['id']}" if existing is not None else f"{api}/repos/{repo}/issues/{pr}/comments"
-resp = req(url, 'PATCH' if existing is not None else 'POST', {'body': body})
-print(f"{'updated' if existing is not None else 'created'} comment {resp['id']}")

.github/workflows/szdiff.yml

@@ -26,14 +26,14 @@ jobs:
       - name: Check whether branch is up-to-date
         id: brstat
         run: |
-          # fetch master from the base repo (tinygrad/tinygrad on GitHub, the mirror on Gitea), not the PR head remote
-          git fetch "${{ github.event.pull_request.base.repo.clone_url }}" master
+          git remote add tinygrad https://github.com/tinygrad/tinygrad
+          git fetch tinygrad master
           echo "${{ github.event.pull_request.head.sha }}"
-          git rev-list --left-right --count FETCH_HEAD...${{ github.event.pull_request.head.sha }} | awk '{print "Behind "$1" - Ahead "$2""}'
-          count=$(git rev-list --left-right --count FETCH_HEAD...${{ github.event.pull_request.head.sha }} | awk '{print $1}')
+          git rev-list --left-right --count  tinygrad/master...${{ github.event.pull_request.head.sha }} | awk '{print "Behind "$1" - Ahead "$2""}'
+          count=$(git rev-list --left-right --count  tinygrad/master...${{ github.event.pull_request.head.sha }} | awk '{print $1}')
           if [ $count -gt 0 ]
           then
-            echo "Current branch is behind ${{ github.event.pull_request.base.repo.full_name }} master branch!"
+            echo "Current branch is behind tinygrad master branch!"
             echo "stat=true" >> "$GITHUB_OUTPUT"
           else
             echo "stat=false" >> "$GITHUB_OUTPUT"
@@ -75,13 +75,13 @@ jobs:
           python sz.py "$BASE" "$PR" > loc_content.txt
       - name: Comment Code Line Diff
         continue-on-error: false
-        env:
+        uses: marocchino/sticky-pull-request-comment@v3
+        with:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          PR_NUMBER: ${{ github.event.pull_request.number }}
-          MARKER: "### Changes"
-          BODY_FILE: loc_content.txt
-        # note: run the script from the base checkout, never from the PR checkout
-        run: python3 "$GITHUB_WORKSPACE/base/.github/workflows/sticky_comment.py"
+          ignore_empty: true
+          skip_unchanged: true
+          recreate: true
+          path: loc_content.txt
 
   rebase:
     name: Core Library Line Difference
@@ -91,14 +91,12 @@ jobs:
     needs: checkbranch
     if: needs.checkbranch.outputs.branchstat == 'true'
     steps:
-      # pull_request_target: a plain checkout gets the base repo, so no PR code is executed
-      - uses: actions/checkout@v6
       - name: Comment Rebase
         continue-on-error: false
-        env:
+        uses: marocchino/sticky-pull-request-comment@v3
+        with:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          PR_NUMBER: ${{ github.event.pull_request.number }}
-          MARKER: "line count difference bot is disabled"
-          MESSAGE: |
-            This branch currently is behind ${{ github.event.pull_request.base.repo.full_name }} master. The line count difference bot is disabled.
-        run: python3 .github/workflows/sticky_comment.py
+          skip_unchanged: true
+          recreate: true
+          message: |
+            This branch currently is behind tinygrad/master. The line count difference bot is disabled.

tinygrad/codegen/__init__.py

@@ -125,8 +125,7 @@ def do_devectorize(b:UOp):
   # broadcasting needs to be already unpacked, Invalid matches any dtype and shape
   if not all(x.shape == b.shape or x.base.is_invalid for x in b.src): return None
   src = []
-  for idx in itertools.product(*[range(x) for x in b.shape]):
-    idx_c = [UOp.const(i) for i in idx]
+  for idx_c in itertools.product(*[[UOp.const(i) for i in range(x)] for x in b.shape]):
     src.append(b.replace(dtype=None, src=tuple(x.base if x.base.is_invalid else x.index(*idx_c) for x in b.src)))
   return UOp.stack(*src).reshape(b.shape) if b.op is not Ops.STORE else UOp.group(*src)
 
@@ -330,10 +329,7 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
   sink = graph_rewrite(sink, symbolic_simple+pm_expand_broadcast+pm_add_loads, name="*** expand broadcast / add loads")
 
   # devectorize
-  sink = graph_rewrite(sink, symbolic_simple+devectorizer2, ctx=ren, name="devectorize2")
-
-  # simplify indexing
-  sink = graph_rewrite(sink, indexing_simplify, name="simplify load/store indexing")
+  sink = graph_rewrite(sink, symbolic_simple+devectorizer2+indexing_simplify, ctx=ren, name="devectorize2")
 
   # some coalescing misses without this
   sink = graph_rewrite(sink, sym, name="early symbolic")

tinygrad/helpers.py

@@ -620,5 +620,3 @@ class count:
     cur = self.n
     self.n += self.step
     return cur
-
-# test change for the szdiff bot