From ed571b63a9bce88177da11cff8a992ca201477d1 Mon Sep 17 00:00:00 2001
From: Qin Jiajia <jiajia.qin@intel.com>
Date: Sat, 12 Oct 2024 13:41:20 +0800
Subject: [PATCH] tune workgroup size

---
 js/web/lib/wasm/jsep/webgpu/ops/matmulnbits.ts | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/js/web/lib/wasm/jsep/webgpu/ops/matmulnbits.ts b/js/web/lib/wasm/jsep/webgpu/ops/matmulnbits.ts
index 032c1f8270223..b2c2f4ebe8d84 100644
--- a/js/web/lib/wasm/jsep/webgpu/ops/matmulnbits.ts
+++ b/js/web/lib/wasm/jsep/webgpu/ops/matmulnbits.ts
@@ -288,8 +288,8 @@ export const createMatMulNBitsBlockSize32ProgramInfo = (
   const components = 1;
   const outputShape = batchDims.concat([dimAOuter, dimBOuter]);
 
-  const workgroupSize = 64;
-  const workgroupY = 4;
+  const workgroupSize = 128;
+  const workgroupY = 8;
   const workgroupX = workgroupSize / workgroupY;
   const tileSize = workgroupX * bComponents * 8; // each uint32 has 8 data.
   const aLengthPerTile = tileSize / aComponents;