makslevental

// Attention mask + max-reduce using named linalg ops.
// 3 kernels: broadcast mask → add to scores → max-reduce over j.
// After generalize + fuse: broadcast and add are inlined into the reduction.

func.func @fuse_mask_into_max_reduce(
    %scores: tensor<4x512x512xf32>,
    %mask: tensor<512x512xf32>,
    %neg_inf_init: tensor<4x512xf32>) -> tensor<4x512xf32> {

  %init3d = tensor.empty() : tensor<4x512x512xf32>

makslevental

No unity

./linkmlirnounity.sh  2.77s user 1.01s system 631% cpu 0.599 total
./linkmlirnounity.sh  2.86s user 0.97s system 670% cpu 0.572 total
./linkmlirnounity.sh  2.81s user 0.91s system 643% cpu 0.579 total
./linkmlirnounity.sh  2.92s user 0.96s system 656% cpu 0.591 total
./linkmlirnounity.sh  2.87s user 0.97s system 656% cpu 0.585 total
./linkmlirnounity.sh  2.88s user 0.92s system 658% cpu 0.578 total
./linkmlirnounity.sh  2.78s user 0.88s system 631% cpu 0.580 total
./linkmlirnounity.sh  2.83s user 0.91s system 648% cpu 0.575 total

makslevental

RTLD_LOCAL, on the other hand, adds the loaded library to a search list that is specific to the current ELF object (called the local scope or scope 1). RTLD_LOCAL is also transitive, meaning that the DT_NEEDED dependencies of a library opened with RTLD_LOCAL will themselves be loaded as RTLD_LOCAL and not added to the global scope.



maksimlevental@Maksims-MacBook-Pro-2 _mlir_libs % objdump -p MLIRPythonSupport-mlir.dll | less

SizeOfHeapCommit        0000000000001000
LoaderFlags             00000000
NumberOfRvaAndSizes     00000010

makslevental

; ModuleID = 'vector_add.air'
source_filename = "vector_add.metal"
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v24:32:32-v32:32:32-v48:64:64-v64:64:64-v96:128:128-v128:128:128-v192:256:256-v256:256:256-v512:512:512-v1024:1024:1024-n8:16:32"
target triple = "air64_v27-apple-macosx15.6.0"

; Function Attrs: mustprogress nounwind
define weak_odr void @mlir_vector_add(ptr addrspace(1) noundef "air-buffer-no-alias" %0, ptr addrspace(1) noundef "air-buffer-no-alias" %1, ptr addrspace(1) noundef "air-buffer-no-alias" %2, <3 x i32> noundef %3) local_unnamed_addr #0 {
  %5 = extractelement <3 x i32> %3, i64 0
  %6 = zext i32 %5 to i64
  %7 = getelementptr inbounds half, ptr addrspace(1) %0, i64 %6

makslevental

mlir_vector_add

makslevental

for i in {1..222} ; do
  curl 'https://lab.llvm.org/buildbot/api/v2/forceschedulers/force-build-scheduler' \
    -H 'Accept: application/json, text/plain, */*' \
    -H 'Accept-Language: en-US,en;q=0.9' \
    -H 'Connection: keep-alive' \
    -H 'Content-Type: application/json' \
    -b 'TWISTED_SESSION=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJ1c2VyX2luZm8iOnsiZnVsbF9uYW1lIjoiTWFrc2ltIExldmVudGFsIiwiZW1haWwiOiJtYWtzaW0ubGV2ZW50YWxAZ21haWwuY29tIiwidXNlcm5hbWUiOiJtYWtzbGV2ZW50YWwiLCJncm91cHMiOlsibGx2bSIsImxsdm0vTExWTSBDb21taXR0ZXJzIiwibGx2bS9ldWRzbCBBZG1pbnMiLCJsbHZtL2V1ZHNsIGNvbW1pdHRlcnMiLCJsbHZtL2V1ZHNsLWFkbWlucyIsImxsdm0vZXVkc2wtY29tbWl0dGVycyIsImxsdm0vaXNzdWUtc3Vic2NyaWJlcnMtbWxpci1weXRob24iLCJsbHZtL2lzc3VlLXN1YnNjcmliZXJzLW1saXI6cHl0aG9uIiwibGx2bS9saWdodGhvdXNlIGNvbW1pdHRlcnMiLCJsbHZtL2xpZ2h0aG91c2UtY29tbWl0dGVycyIsImxsdm0vbGx2bS1jb21taXR0ZXJzIl19LCJleHAiOjE3NjExMTA0MjN9.9kzWVd8PEVKx4lrKYmd4td4UDCenJzNV9wRFfvhj9gM; _ga=GA1.1.1617727655.1755182138; _ga_SBS5VNKHC1=GS2.1.s1755182138$o1$g1$t1755182143$j55$l0$h

makslevental

from collections.abc import Sequence

import mlir

from . import (
    ir as ir,
    passmanager as passmanager,
    rewrite as rewrite
)

makslevental

from collections.abc import Callable, Sequence
import enum
from typing import Any, overload

import WalkOrder
import mlir
import typing_extensions

from _mlir_libs import Context as Context, MLIRError as MLIRError

makslevental

diff --git a/mlir/lib/Bindings/Python/Pass.cpp b/mlir/lib/Bindings/Python/Pass.cpp
index 6ee85e8a3149..88557706bd04 100644
--- a/mlir/lib/Bindings/Python/Pass.cpp
+++ b/mlir/lib/Bindings/Python/Pass.cpp
@@ -59,6 +59,12 @@ void mlir::python::populatePassManagerSubmodule(nb::module_ &m) {
   //----------------------------------------------------------------------------
   // Mapping of the top-level PassManager
   //----------------------------------------------------------------------------
+
+  nb::class_<MlirExternalPass>(m, "ExternalPass");

makslevental

First I'll say that whoever is telling you you need to understand farkas lemma to understand loop carried dependencies is completely fucking with you. Even the people that should understand it probably don't (I can think of 5 people in my professional circle that are "polyhedral" people and I guarantee they don't remember/don't care about farkas).

Secondly, farkas clicked for me after thinking about barrier functions and duality. Bear with me because I haven't thought about this in a while:

Ax = b 

is a linear system of constraints and you're looking for a solution right? Constrained optimization is "hard" because in principle your search algo should always satisfy the constraints.

What's easier is unconstrained optimization i.e. gradient descent. How can you (sort of) turn Ax = b into an unconstrained optimization problem? Write down a penalty function that "penalizes" for violating each of the constraints and then just use gradient descent.