From dfc8e7fb95eb976dbfe6567c8c003c703c6709d0 Mon Sep 17 00:00:00 2001
From: Jokeren <kerenzhou@openai.com>
Date: Mon, 28 Nov 2022 13:47:13 -0800
Subject: [PATCH] Fix

---
 python/tutorials/03-matrix-multiplication.py | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/python/tutorials/03-matrix-multiplication.py b/python/tutorials/03-matrix-multiplication.py
index 837911fc4..e35cd0147 100644
--- a/python/tutorials/03-matrix-multiplication.py
+++ b/python/tutorials/03-matrix-multiplication.py
@@ -24,7 +24,6 @@ module attributes {"triton_gpu.num-warps" = 8 : i32} {
     %c2_i32 = arith.constant 2 : i32
     %c64 = arith.constant 64 : index
     %c128 = arith.constant 128 : index
-    %c256 = arith.constant 256 : index
     %cst = arith.constant dense<64> : tensor<128x64xi32, #blocked0>
     %cst_0 = arith.constant dense<0.000000e+00> : tensor<128x256xf32, #mma>
     %c8_i32 = arith.constant 8 : i32
@@ -105,7 +104,7 @@ module attributes {"triton_gpu.num-warps" = 8 : i32} {
     %67 = tensor.extract_slice %63[0, 0, 0] [1, 64, 256] [1, 1, 1] : tensor<3x64x256xf16, #shared> to tensor<64x256xf16, #shared>
     %68 = tensor.extract_slice %66[0, 0] [128, 16] [1, 1] : tensor<128x64xf16, #shared> to tensor<128x16xf16, #shared>
     %70 = tensor.extract_slice %67[0, 0] [16, 256] [1, 1] : tensor<64x256xf16, #shared> to tensor<16x256xf16, #shared>
-    %72:14 = scf.for %arg9 = %c0 to %39 step %c256 iter_args(%arg10 = %cst_0, %arg11 = %45, %arg12 = %49, %arg13 = %61, %arg14 = %63, %arg15 = %66, %arg16 = %67, %arg17 = %64, %arg18 = %65, %arg19 = %c64, %arg20 = %c2_i32, %arg21 = %c1_i32, %arg22 = %68, %arg23 = %70) -> (tensor<128x256xf32, #mma>, tensor<128x64x!tt.ptr<f16>, #blocked0>, tensor<64x256x!tt.ptr<f16>, #blocked1>, tensor<3x128x64xf16, #shared>, tensor<3x64x256xf16, #shared>, tensor<128x64xf16, #shared>, tensor<64x256xf16, #shared>, tensor<128x64x!tt.ptr<f16>, #blocked0>, tensor<64x256x!tt.ptr<f16>, #blocked1>, index, i32, i32, tensor<128x16xf16, #shared>, tensor<16x256xf16, #shared>) {
+    %72:14 = scf.for %arg9 = %c0 to %39 step %c128 iter_args(%arg10 = %cst_0, %arg11 = %45, %arg12 = %49, %arg13 = %61, %arg14 = %63, %arg15 = %66, %arg16 = %67, %arg17 = %64, %arg18 = %65, %arg19 = %c64, %arg20 = %c2_i32, %arg21 = %c1_i32, %arg22 = %68, %arg23 = %70) -> (tensor<128x256xf32, #mma>, tensor<128x64x!tt.ptr<f16>, #blocked0>, tensor<64x256x!tt.ptr<f16>, #blocked1>, tensor<3x128x64xf16, #shared>, tensor<3x64x256xf16, #shared>, tensor<128x64xf16, #shared>, tensor<64x256xf16, #shared>, tensor<128x64x!tt.ptr<f16>, #blocked0>, tensor<64x256x!tt.ptr<f16>, #blocked1>, index, i32, i32, tensor<128x16xf16, #shared>, tensor<16x256xf16, #shared>) {
       %69 = triton_gpu.convert_layout %arg22 : (tensor<128x16xf16, #shared>) -> tensor<128x16xf16, #triton_gpu.dot_op<{opIdx = 0, parent = #mma}>>
       %71 = triton_gpu.convert_layout %arg23 : (tensor<16x256xf16, #shared>) -> tensor<16x256xf16, #triton_gpu.dot_op<{opIdx = 1, parent = #mma}>>
       %89 = tt.dot %69, %71, %arg10 {allowTF32 = true, transA = false, transB = false} : tensor<128x16xf16, #triton_gpu.dot_op<{opIdx = 0, parent = #mma}>> * tensor<16x256xf16, #triton_gpu.dot_op<{opIdx = 1, parent = #mma}>> -> tensor<128x256xf32, #mma>