diff --git a/python/tvm/topi/adreno/conv2d_alter_op.py b/python/tvm/topi/adreno/conv2d_alter_op.py
index 6cf749a62b27..cf72cc2a846e 100644
--- a/python/tvm/topi/adreno/conv2d_alter_op.py
+++ b/python/tvm/topi/adreno/conv2d_alter_op.py
@@ -143,8 +143,11 @@ def _alter_conv2d_layout(attrs, inputs, tinfos, out_type):
         CO, _, KH, KW = get_const_tuple(kernel_tensor.shape)
 
         # pre-compute weight transformation in winograd
+        # alpha, alpha, CO, CI
         weight = relay.nn.contrib_conv2d_winograd_weight_transform(inputs[1], tile_size=tile_size)
         weight = relay.transpose(weight, axes=[2, 3, 0, 1])  # HWOI -> OIHW
+        # (oc, ic, h, w) -> (h, w, ic, oc)
+        new_attrs["kernel_layout"] = "HWIO"
         new_attrs["tile_size"] = tile_size
         new_attrs["channels"] = CO
 
diff --git a/tests/python/relay/opencl_texture/test_conv2d_nchw_texture.py b/tests/python/relay/opencl_texture/test_conv2d_nchw_texture.py
index 504a2b4e3ed3..4e813b1ffe84 100644
--- a/tests/python/relay/opencl_texture/test_conv2d_nchw_texture.py
+++ b/tests/python/relay/opencl_texture/test_conv2d_nchw_texture.py
@@ -1029,3 +1029,36 @@ def test_conv2d_different_lowering_same_op(target, dtype):
     ]
 
     build_run_compare(mod, params1, {"data": input_shape}, dtype, target, static_memory_scope)
+
+
+@tvm.testing.requires_opencl
+@tvm.testing.parametrize_targets("opencl -device=adreno")
+def test_conv2d_winograd_non_rect(target, dtype):
+    input_shape = (1, 771, 36, 64)
+    A = relay.var("data", shape=input_shape, dtype=dtype)
+    filter_shape = (128, 771, 3, 3)
+    B = relay.var("weight", shape=filter_shape, dtype=dtype)
+    D = relay.nn.conv2d(
+        A, B, padding=[1, 1, 1, 1], channels=128, kernel_size=[3, 3], out_dtype=dtype
+    )
+
+    mod = relay.Function([A, B], D)
+    np.random.seed(1)
+    initializer = relay.testing.init.Xavier()
+    filter_data = np.zeros(filter_shape).astype(dtype)
+    initializer("weight", filter_data)
+    params1 = {
+        "weight": tvm.nd.array(filter_data),
+    }
+
+    temp = utils.tempdir()
+    stat_file = temp.relpath("stat.log")
+    with open(stat_file, "w") as f:
+        f.write(
+            f'{{"input": ["opencl -keys=adreno,opencl,gpu -device=adreno -max_num_threads=256 -texture_spatial_limit=16384 -thread_warp_size=1", "conv2d_nchw_winograd.image2d", [["TENSOR", [1, 771, 36, 64], "{dtype}"], ["TENSOR", [128, 771, 3, 3], "{dtype}"], [1, 1], [1, 1, 1, 1], [1, 1], "{dtype}"], {{}}], "config": {{"index": 5399, "code_hash": null, "entity": [["auto_unroll_max_step", "ot", 16], ["tile_y", "sp", [-1, 1, 32]], ["tile_x", "sp", [-1, 4, 8]], ["tile_rc", "sp", [-1, 193]]]}}, "result": [[0.0037244], 0, 7.06374192237854, 1653898629.7427933], "version": 0.2, "tvm_version": "0.8.dev0"}}\n'
+        )
+    graph = build_run_compare(
+        mod, params1, {"data": input_shape}, dtype, target, stat_file=stat_file
+    )
+    matches = re.findall("winograd", graph)
+    assert len(matches) > 0
diff --git a/tests/python/relay/opencl_texture/test_conv2d_nhwc_texture.py b/tests/python/relay/opencl_texture/test_conv2d_nhwc_texture.py
index 37c22137f035..0b89e3dc9c7f 100644
--- a/tests/python/relay/opencl_texture/test_conv2d_nhwc_texture.py
+++ b/tests/python/relay/opencl_texture/test_conv2d_nhwc_texture.py
@@ -581,3 +581,104 @@ def test_conv2d_vgg16_winograd_4d(target, dtype):
     )
     matches = re.findall("winograd", graph)
     assert len(matches) > 0
+
+
+@tvm.testing.requires_opencl
+@tvm.testing.parametrize_targets("opencl -device=adreno")
+def test_conv2d_winograd_conv(target, dtype):
+    input_shape = (1, 3, 3, 4)
+    A = relay.var("data", shape=input_shape, dtype=dtype)
+    filter_shape3 = (3, 3, 4, 8)
+    bias_shape3 = (1, 1, 1, 8)
+    B3 = relay.var("weight3", shape=filter_shape3, dtype=dtype)
+    D = relay.nn.conv2d(
+        A,
+        B3,
+        data_layout="NHWC",
+        kernel_layout="HWIO",
+        padding=[1, 1, 1, 1],
+        channels=8,
+        kernel_size=[3, 3],
+        out_dtype=dtype,
+    )
+
+    filter_shape4 = (3, 3, 8, 8)
+    bias_shape4 = (1, 1, 1, 8)
+    B4 = relay.var("weight4", shape=filter_shape4, dtype=dtype)
+    D = relay.nn.conv2d(
+        D,
+        B4,
+        data_layout="NHWC",
+        kernel_layout="HWIO",
+        padding=[1, 1, 1, 1],
+        channels=8,
+        kernel_size=[3, 3],
+        out_dtype=dtype,
+    )
+    mod = relay.Function([A, B3, B4], D)
+    np.random.seed(1)
+    initializer = relay.testing.init.Xavier()
+    filter_data3 = np.zeros(filter_shape3).astype(dtype)
+    bias_data3 = np.zeros(bias_shape3).astype(dtype)
+    filter_data4 = np.zeros(filter_shape4).astype(dtype)
+    bias_data4 = np.zeros(bias_shape4).astype(dtype)
+    initializer("weight", filter_data3)
+    initializer("bias", bias_data3)
+    initializer("weight", filter_data4)
+    initializer("bias", bias_data4)
+    params1 = {
+        "weight3": tvm.nd.array(filter_data3),
+        "weight4": tvm.nd.array(filter_data4),
+    }
+
+    temp = utils.tempdir()
+    stat_file = temp.relpath("stat.log")
+    with open(stat_file, "w") as f:
+        f.write(
+            f'{{"input": ["opencl -keys=adreno,opencl,gpu -device=adreno -max_num_threads=256", "conv2d_nhwc_winograd.image2d", [["TENSOR", [1, 3, 3, 4], "{dtype}"], ["TENSOR", [3, 3, 4, 8], "{dtype}"], [1, 1], [1, 1, 1, 1], [1, 1], "{dtype}"], {{}}], "config": {{"index": 1591, "code_hash": null, "entity": [["auto_unroll_max_step", "ot", 4], ["tile_y", "sp", [-1, 1, 32]], ["tile_x", "sp", [-1, 4, 2]], ["tile_rc", "sp", [-1, 8]]]}}, "result": [[0.0037244], 0, 7.06374192237854, 1653898629.7427933], "version": 0.2, "tvm_version": "0.8.dev0"}}\n'
+        )
+    graph = build_run_compare(
+        mod, params1, {"data": input_shape}, dtype, target, stat_file=stat_file
+    )
+    matches = re.findall("winograd", graph)
+    assert len(matches) > 0
+
+
+@tvm.testing.requires_opencl
+@tvm.testing.parametrize_targets("opencl -device=adreno")
+def test_conv2d_winograd_non_rect(target, dtype):
+    input_shape = (1, 36, 64, 771)
+    A = relay.var("data", shape=input_shape, dtype=dtype)
+    filter_shape = (3, 3, 771, 128)
+    B = relay.var("weight", shape=filter_shape, dtype=dtype)
+    D = relay.nn.conv2d(
+        A,
+        B,
+        data_layout="NHWC",
+        kernel_layout="HWIO",
+        padding=[1, 1, 1, 1],
+        channels=128,
+        kernel_size=[3, 3],
+        out_dtype=dtype,
+    )
+
+    mod = relay.Function([A, B], D)
+    np.random.seed(1)
+    initializer = relay.testing.init.Xavier()
+    filter_data = np.zeros(filter_shape).astype(dtype)
+    initializer("weight", filter_data)
+    params1 = {
+        "weight": tvm.nd.array(filter_data),
+    }
+
+    temp = utils.tempdir()
+    stat_file = temp.relpath("stat.log")
+    with open(stat_file, "w") as f:
+        f.write(
+            f'{{"input": ["opencl -keys=adreno,opencl,gpu -device=adreno -max_num_threads=256 -texture_spatial_limit=16384 -thread_warp_size=1", "conv2d_nhwc_winograd.image2d", [["TENSOR", [1, 36, 64, 771], "{dtype}"], ["TENSOR", [3, 3, 771, 128], "{dtype}"], [1, 1], [1, 1, 1, 1], [1, 1], "{dtype}"], {{}}], "config": {{"index": 5399, "code_hash": null, "entity": [["auto_unroll_max_step", "ot", 16], ["tile_y", "sp", [-1, 1, 32]], ["tile_x", "sp", [-1, 4, 8]], ["tile_rc", "sp", [-1, 193]]]}}, "result": [[0.0037244], 0, 7.06374192237854, 1653898629.7427933], "version": 0.2, "tvm_version": "0.8.dev0"}}\n'
+        )
+    graph = build_run_compare(
+        mod, params1, {"data": input_shape}, dtype, target, stat_file=stat_file
+    )
+    matches = re.findall("winograd", graph)
+    assert len(matches) > 0