[ExecuTorch][Llama] Use Eigen blas for custom sdpa

kimishpatel · kimishpatel · commit 8c0113d33304 · 2024-10-15T07:09:38.000-07:00
OpenBlas's implementation is not thread safe. Thus when used within parallel_for, it produces incorrect output. This has been documented in a few places like here OpenMathLib/OpenBLAS#1441 and OpenMathLib/OpenBLAS#2543. I tried few options to disable openblas's multithreading but none of them seemed to work. It is possible that upstream openblas has fixed this. Will validate this by pulling in latest, but in the meanwhile using eigen_blas to unblock. Differential Revision: [D64398816](https://our.internmc.facebook.com/intern/diff/D64398816/) [ghstack-poisoned]
diff --git a/kernels/optimized/lib_defs.bzl b/kernels/optimized/lib_defs.bzl
@@ -144,7 +144,7 @@ def define_libs():
                 (
                     "^android-arm64.*$",
                     [
-                        "fbsource//third-party/openblas:openblas",
+                        "fbsource//arvr/third-party/eigen:eigen3_blas",
                     ],
                 ),
             ],

Original file line number	Diff line number	Diff line change
`@@ -144,7 +144,7 @@ def define_libs():`
`144`	`144`	`(`
`145`	`145`	`"^android-arm64.*$",`
`146`	`146`	`[`
`147`		`- "fbsource//third-party/openblas:openblas",`
	`147`	`+ "fbsource//arvr/third-party/eigen:eigen3_blas",`
`148`	`148`	`],`
`149`	`149`	`),`
`150`	`150`	`],`