// Tests that the dot_general accepts i8 tensors and outputs an i32 tensor.
// Note: Argument quantization sequence omitted.
// CHECK: stablehlo.dot_general %{{.+}}, %{{.+}}, contracting_dims = [1] x [0] : (tensor<1x1024xi8>, tensor<1024x3xi8>) -> tensor<1x3xi32>

// Note: Result dequantization sequence omitted.
// CHECK: return %{{.+}} : tensor<1x3xf32>

// -----

}
// CHECK-LABEL: dot_general_srq
// CHECK-SAME: (%[[ARG_1:.+]]: tensor<1x1024x!quant.uniform<i8:f32, {{.*}}>) -> tensor<1x3x!quant.uniform<i8:f32, 3.000000e+00:-127>>
// CHECK-NOT: stablehlo.dot_general

// CHECK-SAME:                         %[[ARG_1:.*]]: tensor<?x?x1xi32>) -> tensor<?x?x1024xi8> {
// CHECK:            %[[VAL_0:.*]] = "tf.GatherNd"(%[[ARG_0]], %[[ARG_1]]) <{bad_indices_policy = ""}> : (tensor<256000x1024xi8>, tensor<?x?x1xi32>) -> tensor<?x?x1024xi8>
// CHECK:            return %[[VAL_0]] : tensor<?x?x1024xi8>
// CHECK:         }

  // Move binary op batched RHS before reshape:
  // binary(reshape(lhs), rhs) => reshape(binary(lhs, flatten(rhs)))
  // Pattern targetted here is as follows-
  // [input, lhr, rhs] == [<1x1024x128>, <1x1024x8x16>, <1x1x8x16xf32>]
  // This is valid only when the-
  // 1.last dimension of lhs is equal to the number of elements in constant rhs.
  // 2.Reduded shape of rhs, here <8x16> is equal to last dimensions of lhs.