fix linalg op

Author: yifeizh2

include/gc/Dialect/Linalgx/LinalgxStructuredOps.td

@@ -328,9 +328,14 @@ def Linalgx_ScaledDotProductAttentionOp
       Variadic<TensorOrMemref>:$inputs,
       Variadic<TensorOrMemref>:$outputs);
   let results = (outs Variadic<TensorOrMemref>:$results);
-  let regions = (region AnyRegion:$region);
 
   let hasVerifier = 1;
+  let assemblyFormat = [{
+    attr-dict
+    `ins` `(` $inputs `:` type($inputs) `)`
+    `outs` `(` $outputs `:` type($outputs) `)`
+    (`->` type($results)^)?
+  }];
 }
 
 #endif // LINALGX_STRUCTURED_OPS

test/gc/Transform/flashAttention.mlir

@@ -0,0 +1,7 @@
+// RUN: gc-opt --split-input-file --flash-attention-conversion %s
+
+func.func @flash_attention(%arg0: tensor<1x16x384x64xf32>, %arg1: tensor<1x16x384x64xf32>, %arg2: tensor<1x16x384x64xf32>, %arg3: tensor<1x16x384x384xf32>) -> tensor<1x16x384x64xf32> {
+    %0 = tensor.empty() : tensor<1x16x384x64xf32>
+    %1 = linalgx.scaled_dot_product_attention ins(%arg0, %arg1, %arg2, %arg3: tensor<1x16x384x64xf32>, tensor<1x16x384x64xf32>, tensor<1x16x384x64xf32>, tensor<1x16x384x384xf32>) outs(%0 : tensor<1x16x384x64xf32>)  -> tensor<1x16x384x64xf32>
+    return %1 : tensor<1x16x384x64xf32>
+}