{
  "tvm_version": "0.27.0.post1",
  "target": "llvm",
  "original_tensor_program": "# from tvm.script import tirx as T\n# from tvm.tirx.layout import Axis\n\n@T.prim_func(s_tir=True)\ndef affine_tail(A: T.Buffer((37,), \"float32\"), B: T.Buffer((37,), \"float32\")):\n    T.func_attr({\"tir.noalias\": True})\n    # with T.sblock(\"root\"):\n    for i in range(37):\n        with T.sblock(\"affine\"):\n            vi = T.axis.spatial(37, i)\n            T.reads(A[vi])\n            T.writes(B[vi])\n            B[vi] = A[vi] * T.float32(2.0) + T.float32(1.0)",
  "scheduled_tile8": "# from tvm.script import ir as I\n# from tvm.script import tirx as T\n# from tvm.tirx.layout import Axis\n\n@I.ir_module\nclass Module:\n    @T.prim_func(s_tir=True)\n    def main(A: T.Buffer((37,), \"float32\"), B: T.Buffer((37,), \"float32\")):\n        T.func_attr({\"global_symbol\": \"affine_tail\", \"tir.noalias\": True})\n        # with T.sblock(\"root\"):\n        for i_0 in range(5):\n            for i_1 in T.vectorized(8):\n                with T.sblock(\"affine\"):\n                    vi = T.axis.spatial(37, i_0 * 8 + i_1)\n                    T.where(i_0 * 8 + i_1 < 37)\n                    T.reads(A[vi])\n                    T.writes(B[vi])\n                    B[vi] = A[vi] * T.float32(2.0) + T.float32(1.0)",
  "relax_original": "# from tvm.script import ir as I\n# from tvm.script import relax as R\n\n@I.ir_module\nclass Module:\n    @R.function\n    def forward(x: R.Tensor((3, 7), dtype=\"float32\"), w: R.Tensor((7, 5), dtype=\"float32\"), bias: R.Tensor((5,), dtype=\"float32\")) -> R.Tensor((3, 5), dtype=\"float32\"):\n        R.func_attr({\"num_input\": 3})\n        with R.dataflow():\n            matmul: R.Tensor((3, 5), dtype=\"float32\") = R.matmul(x, w, out_dtype=None)\n            add: R.Tensor((3, 5), dtype=\"float32\") = R.add(matmul, bias)\n            relu: R.Tensor((3, 5), dtype=\"float32\") = R.nn.relu(add)\n            gv: R.Tensor((3, 5), dtype=\"float32\") = relu\n            R.output(gv)\n        return gv",
  "relax_lowered": "# from tvm.script import ir as I\n# from tvm.script import tirx as T\n# from tvm.tirx.layout import Axis\n# from tvm.script import relax as R\n\n@I.ir_module\nclass Module:\n    @T.prim_func(private=True, s_tir=True)\n    def fused_matmul_add_relu(x: T.Buffer((T.int64(3), T.int64(7)), \"float32\"), w: T.Buffer((T.int64(7), T.int64(5)), \"float32\"), bias: T.Buffer((T.int64(5),), \"float32\"), compute_intermediate: T.Buffer((T.int64(3), T.int64(5)), \"float32\")):\n        T.func_attr({\"tirx.noalias\": True})\n        # with T.sblock(\"root\"):\n        matmul_intermediate = T.sblock_alloc_buffer((T.int64(3), T.int64(5)))\n        T_add_intermediate = T.sblock_alloc_buffer((T.int64(3), T.int64(5)))\n        for i0, i1, k in T.grid(T.int64(3), T.int64(5), T.int64(7)):\n            with T.sblock(\"matmul\"):\n                v_i0, v_i1, v_k = T.axis.remap(\"SSR\", [i0, i1, k])\n                T.reads(x[v_i0, v_k], w[v_k, v_i1])\n                T.writes(matmul_intermediate[v_i0, v_i1])\n                with T.init():\n                    matmul_intermediate[v_i0, v_i1] = T.float32(0.0)\n                matmul_intermediate[v_i0, v_i1] = matmul_intermediate[v_i0, v_i1] + x[v_i0, v_k] * w[v_k, v_i1]\n        for ax0, ax1 in T.grid(T.int64(3), T.int64(5)):\n            with T.sblock(\"T_add\"):\n                v_ax0, v_ax1 = T.axis.remap(\"SS\", [ax0, ax1])\n                T.reads(matmul_intermediate[v_ax0, v_ax1], bias[v_ax1])\n                T.writes(T_add_intermediate[v_ax0, v_ax1])\n                T_add_intermediate[v_ax0, v_ax1] = matmul_intermediate[v_ax0, v_ax1] + bias[v_ax1]\n        for i0, i1 in T.grid(T.int64(3), T.int64(5)):\n            with T.sblock(\"compute\"):\n                v_i0, v_i1 = T.axis.remap(\"SS\", [i0, i1])\n                T.reads(T_add_intermediate[v_i0, v_i1])\n                T.writes(compute_intermediate[v_i0, v_i1])\n                compute_intermediate[v_i0, v_i1] = T.max(T_add_intermediate[v_i0, v_i1], T.float32(0.0))\n\n    @R.function\n    def forward(x: R.Tensor((3, 7), dtype=\"float32\"), w: R.Tensor((7, 5), dtype=\"float32\"), bias: R.Tensor((5,), dtype=\"float32\")) -> R.Tensor((3, 5), dtype=\"float32\"):\n        R.func_attr({\"num_input\": 3})\n        cls = Module\n        with R.dataflow():\n            gv = R.call_tir(cls.fused_matmul_add_relu, (x, w, bias), out_ty=R.Tensor((3, 5), dtype=\"float32\"))\n            R.output(gv)\n        return gv"
}
