diff --git a/python/tutorials/03-matrix-multiplication.py b/python/tutorials/03-matrix-multiplication.py
index ff9daa08c..e35cd0147 100644
--- a/python/tutorials/03-matrix-multiplication.py
+++ b/python/tutorials/03-matrix-multiplication.py
@@ -104,7 +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 %c64 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>
@@ -138,7 +138,40 @@ module attributes {"triton_gpu.num-warps" = 8 : i32} {
       %116 = arith.addi %arg21, %c1_i32 : i32
       %117 = tensor.extract_slice %113[0, 0] [128, 16] [1, 1] : tensor<128x64xf16, #shared> to tensor<128x16xf16, #shared>
       %119 = tensor.extract_slice %114[0, 0] [16, 256] [1, 1] : tensor<64x256xf16, #shared> to tensor<16x256xf16, #shared>
-      scf.yield %99, %100, %101, %108, %110, %113, %114, %111, %112, %102, %115, %116, %117, %119 : 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>
+      %691 = triton_gpu.convert_layout %117 : (tensor<128x16xf16, #shared>) -> tensor<128x16xf16, #triton_gpu.dot_op<{opIdx = 0, parent = #mma}>>
+      %711 = triton_gpu.convert_layout %119 : (tensor<16x256xf16, #shared>) -> tensor<16x256xf16, #triton_gpu.dot_op<{opIdx = 1, parent = #mma}>>
+      %891 = tt.dot %691, %711, %99 {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>
+      %901 = tensor.extract_slice %113[0, 16] [128, 32] [1, 1] : tensor<128x64xf16, #shared> to tensor<128x32xf16, #shared>
+      %911 = triton_gpu.convert_layout %901 : (tensor<128x32xf16, #shared>) -> tensor<128x32xf16, #triton_gpu.dot_op<{opIdx = 0, parent = #mma}>>
+      %921 = tensor.extract_slice %114[16, 0] [32, 256] [1, 1] : tensor<64x256xf16, #shared> to tensor<32x256xf16, #shared>
+      %931 = triton_gpu.convert_layout %921 : (tensor<32x256xf16, #shared>) -> tensor<32x256xf16, #triton_gpu.dot_op<{opIdx = 1, parent = #mma}>>
+      %941 = tt.dot %911, %931, %891 {allowTF32 = true, transA = false, transB = false} : tensor<128x32xf16, #triton_gpu.dot_op<{opIdx = 0, parent = #mma}>> * tensor<32x256xf16, #triton_gpu.dot_op<{opIdx = 1, parent = #mma}>> -> tensor<128x256xf32, #mma>
+      %951 = tensor.extract_slice %113[0, 48] [128, 16] [1, 1] : tensor<128x64xf16, #shared> to tensor<128x16xf16, #shared>
+      %961 = triton_gpu.convert_layout %951 : (tensor<128x16xf16, #shared>) -> tensor<128x16xf16, #triton_gpu.dot_op<{opIdx = 0, parent = #mma}>>
+      %971 = tensor.extract_slice %114[48, 0] [16, 256] [1, 1] : tensor<64x256xf16, #shared> to tensor<16x256xf16, #shared>
+      %981 = triton_gpu.convert_layout %971 : (tensor<16x256xf16, #shared>) -> tensor<16x256xf16, #triton_gpu.dot_op<{opIdx = 1, parent = #mma}>>
+      %991 = tt.dot %961, %981, %941 {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>
+      %1001 = tt.addptr %100, %cst : tensor<128x64x!tt.ptr<f16>, #blocked0>
+      %1011 = tt.addptr %101, %41 : tensor<64x256x!tt.ptr<f16>, #blocked1>
+      %1021 = arith.addi %102, %c64 : index
+      %1031 = arith.cmpi slt, %1021, %39 : index
+      %1041 = arith.remsi %115, %c3_i32 : i32
+      %1051 = arith.remsi %116, %c3_i32 : i32
+      %1061 = arith.index_cast %1051 : i32 to index
+      %1071 = tt.splat %1031 : (i1) -> tensor<128x64xi1, #blocked0>
+      %1081 = triton_gpu.insert_slice_async %111, %108, %1041, %1071 {axis = 0 : i32, cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<128x64x!tt.ptr<f16>, #blocked0> -> tensor<3x128x64xf16, #shared>
+      %1091 = tt.splat %1031 : (i1) -> tensor<64x256xi1, #blocked1>
+      %1101 = triton_gpu.insert_slice_async %112, %110, %1041, %1091 {axis = 0 : i32, cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<64x256x!tt.ptr<f16>, #blocked1> -> tensor<3x64x256xf16, #shared>
+      %1111 = tt.addptr %111, %cst : tensor<128x64x!tt.ptr<f16>, #blocked0>
+      %1121 = tt.addptr %112, %41 : tensor<64x256x!tt.ptr<f16>, #blocked1>
+      triton_gpu.async_wait {num = 2 : i32}
+      %1131 = tensor.extract_slice %1081[%1061, 0, 0] [1, 128, 64] [1, 1, 1] : tensor<3x128x64xf16, #shared> to tensor<128x64xf16, #shared>
+      %1141 = tensor.extract_slice %1101[%1061, 0, 0] [1, 64, 256] [1, 1, 1] : tensor<3x64x256xf16, #shared> to tensor<64x256xf16, #shared>
+      %1151 = arith.addi %115, %c1_i32 : i32
+      %1161 = arith.addi %116, %c1_i32 : i32
+      %1171 = tensor.extract_slice %1131[0, 0] [128, 16] [1, 1] : tensor<128x64xf16, #shared> to tensor<128x16xf16, #shared>
+      %1191 = tensor.extract_slice %1141[0, 0] [16, 256] [1, 1] : tensor<64x256xf16, #shared> to tensor<16x256xf16, #shared>
+      scf.yield %991, %1001, %1011, %1081, %1101, %1131, %1141, %1111, %1121, %1021, %1151, %1161, %1171, %1191 : 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>
     }
     triton_gpu.async_wait {num = 0 : i32}
     %73 = triton_gpu.convert_layout %72#0 : (tensor<128x256xf32, #mma>) -> tensor<128x256xf32, #blocked1>