diff --git a/python/pyxir/contrib/target/components/DPUCZDX8G/dnnc_output.py b/python/pyxir/contrib/target/components/DPUCZDX8G/dnnc_output.py index 026ad97..616c933 100644 --- a/python/pyxir/contrib/target/components/DPUCZDX8G/dnnc_output.py +++ b/python/pyxir/contrib/target/components/DPUCZDX8G/dnnc_output.py @@ -12,9 +12,10 @@ # See the License for the specific language governing permissions and # limitations under the License. -""" Module wrapping parsing DNNC compiler output """ +"""Module wrapping parsing DNNC compiler output""" import abc +import re class BaseDNNCOutput(object): @@ -38,47 +39,46 @@ def get_output_nodes(self): class DNNCOutput(BaseDNNCOutput): - def _parse(self, output): # type: (str) -> dict """ Parse DNNC v4.0 compiler output """ d = { - 'Boundary Input Tensors': {}, - 'Boundary Output Tensors': {}, - 'Boundary Output Tensors Shapes': {}, - 'Input Nodes': {}, - 'Output Nodes': {} + "Boundary Input Tensors": {}, + "Boundary Output Tensors": {}, + "Boundary Output Tensors Shapes": {}, + "Input Nodes": {}, + "Output Nodes": {}, } - lines = output.split('\n')[0].split('\\n') + lines = output.split("\n")[0].split("\\n") for idx, line in enumerate(lines): split_line = line.lstrip().rstrip().split(" : ") - if split_line[0] == 'Kernel ID': - d['Kernel ID'] = lines[idx + 1].split(" : ")[0] - d['Name'] = lines[idx + 1].split(" : ")[1] - elif split_line[0] == 'Kernel Name': - d['Kernel Name'] = split_line[1] - elif split_line[0] == 'Kernel Type': - d['Kernel Type'] = split_line[1] - elif split_line[0] == 'Code Size': - d['Code Size'] = split_line[1] - elif split_line[0] == 'Param Size': - d['Param Size'] = split_line[1] - elif split_line[0] == 'Workload MACs': - d['Workload MACs'] = split_line[1] - elif split_line[0] == 'IO Memory Space': - d['IO Memory Space'] = split_line[1] - elif split_line[0] == 'Mean Value': - d['Mean Value'] = split_line[1].split(',') - elif split_line[0] == 'Node Count': - d['Node Count'] = split_line[1] - elif split_line[0] == 'Tensor Count': - d['Tensor Count'] = split_line[1] - elif split_line[0] == 'Total Tensor Count': - d['Tensor Count'] = split_line[1] - elif split_line[0] == 'Boundary Input Tensor(s) (H*W*C)': + if split_line[0] == "Kernel ID": + d["Kernel ID"] = lines[idx + 1].split(" : ")[0] + d["Name"] = lines[idx + 1].split(" : ")[1] + elif split_line[0] == "Kernel Name": + d["Kernel Name"] = split_line[1] + elif split_line[0] == "Kernel Type": + d["Kernel Type"] = split_line[1] + elif split_line[0] == "Code Size": + d["Code Size"] = split_line[1] + elif split_line[0] == "Param Size": + d["Param Size"] = split_line[1] + elif split_line[0] == "Workload MACs": + d["Workload MACs"] = split_line[1] + elif split_line[0] == "IO Memory Space": + d["IO Memory Space"] = split_line[1] + elif split_line[0] == "Mean Value": + d["Mean Value"] = split_line[1].split(",") + elif split_line[0] == "Node Count": + d["Node Count"] = split_line[1] + elif split_line[0] == "Tensor Count": + d["Tensor Count"] = split_line[1] + elif split_line[0] == "Total Tensor Count": + d["Tensor Count"] = split_line[1] + elif split_line[0] == "Boundary Input Tensor(s) (H*W*C)": for i in range(idx + 1, len(lines)): split_line_i = lines[i].lstrip().rstrip().split(" : ") @@ -87,12 +87,14 @@ def _parse(self, output): name, shape = split_line_i - if shape in d['Boundary Input Tensors']: - raise ValueError("DNNC compiler cannot handle multiple" - " inputs with the same shape") + if shape in d["Boundary Input Tensors"]: + raise ValueError( + "DNNC compiler cannot handle multiple" + " inputs with the same shape" + ) - d['Boundary Input Tensors'][shape] = name.split(":")[0] - elif split_line[0] == 'Boundary Output Tensor(s) (H*W*C)': + d["Boundary Input Tensors"][shape] = name.split(":")[0] + elif split_line[0] == "Boundary Output Tensor(s) (H*W*C)": for i in range(idx + 1, len(lines)): split_line_i = lines[i].lstrip().rstrip().split(" : ") @@ -105,13 +107,12 @@ def _parse(self, output): # if shape in d['Boundary Output Tensors']: # raise ValueError("DNNC compiler cannot handle multiple" # " outputs with the same shape") - - d['Boundary Output Tensors'][name] = name + ':0' - d['Boundary Output Tensors Shapes'][shape] = name + ':0' - elif split_line[0] == 'Total Node Count': - d['Total Node Count'] = split_line[1] - elif split_line[0] in ['Input Node(s) (H*W*C)', - 'Input Node(s)(H*W*C)']: + + d["Boundary Output Tensors"][name] = name + ":0" + d["Boundary Output Tensors Shapes"][shape] = name + ":0" + elif split_line[0] == "Total Node Count": + d["Total Node Count"] = split_line[1] + elif split_line[0] in ["Input Node(s) (H*W*C)", "Input Node(s)(H*W*C)"]: for i in range(idx + 1, len(lines)): split_line_i = lines[i].lstrip().rstrip().split(" : ") @@ -120,14 +121,15 @@ def _parse(self, output): name, shape = split_line_i - if shape in d['Input Nodes']: - raise ValueError("DNNC compiler cannot handle multiple" - " inputs with the same shape") + if shape in d["Input Nodes"]: + raise ValueError( + "DNNC compiler cannot handle multiple" + " inputs with the same shape" + ) - d['Input Nodes'][shape] = name[:-3] + d["Input Nodes"][shape] = name[:-3] - elif split_line[0] in ['Output Node(s) (H*W*C)', - 'Output Node(s)(H*W*C)']: + elif split_line[0] in ["Output Node(s) (H*W*C)", "Output Node(s)(H*W*C)"]: for i in range(idx + 1, len(lines)): split_line_i = lines[i].lstrip().rstrip().split(" : ") @@ -140,22 +142,22 @@ def _parse(self, output): # raise ValueError("DNNC compiler cannot handle multiple" # " outputs with the same shape") - d['Output Nodes'][name[:-3]] = name[:-3] - d['Output Nodes'][shape] = name[:-3] + d["Output Nodes"][name[:-3]] = name[:-3] + d["Output Nodes"][shape] = name[:-3] return d def get_input_nodes(self): # type: () -> Dict[str, str] - return self.d['Input Nodes'] + return self.d["Input Nodes"] def get_output_nodes(self): # type: () -> Dict[str, str] - return self.d['Boundary Output Tensors'] + return self.d["Boundary Output Tensors"] def get_output_nodes_on_shapes(self): # type: () -> Dict[str, str] - return self.d['Boundary Output Tensors Shapes'] + return self.d["Boundary Output Tensors Shapes"] def get_dnnc_str(self, str_value: str) -> str: - return str_value.replace('-', '_') + return re.sub(r"[^A-Za-z0-9]", "_", str_value) diff --git a/tests/unit/contrib/compilation_infra.py b/tests/unit/contrib/compilation_infra.py index cde6a9c..630b791 100644 --- a/tests/unit/contrib/compilation_infra.py +++ b/tests/unit/contrib/compilation_infra.py @@ -67,6 +67,8 @@ def _create_conv2d_pool2d_nhwc_oihw( conv_invalid=False, kernel_layout="OIHW", target="DPUCZDX8G-zcu104", + conv_name="conv1", + pool_name="pool1", ) -> XGraph: kernel_w, kernel_h = w_shape[2], w_shape[3] @@ -76,7 +78,7 @@ def _create_conv2d_pool2d_nhwc_oihw( x1 = px.ops.input("in1", shape=list(in_shape)) w1 = px.ops.constant("weight", W) conv1 = px.ops.conv2d( - op_name="conv1", + op_name=conv_name, input_layer=x1, weights_layer=w1, kernel_size=[kernel_w, kernel_h], @@ -87,7 +89,7 @@ def _create_conv2d_pool2d_nhwc_oihw( data_layout="NHWC", ) pool1 = px.ops.pool2d( - op_name="pool1", + op_name=pool_name, input_layer=conv1, pool_type=pool_type, pool_size=list(pool_size), @@ -221,6 +223,52 @@ def inputs_func(iter): shutil.rmtree(build_dir) +def conv2d_pool2d_naming_test(conv_names, pool_names) -> None: + + in_shape = (1, 4, 4, 1) + target = "DPUCZDX8G-zcu104" + for conv_name, pool_name in zip(conv_names, pool_names): + xgraph = _create_conv2d_pool2d_nhwc_oihw( + in_shape, + (2, 1, 2, 2), + [0, 0], + [1, 1], + [1, 1], + "Max", + [2, 2], + [0, 0], + [1, 1], + 1, + False, + "OIHW", + target, + conv_name, + pool_name, + ) + + def inputs_func(iter): + inputs = np.ones(in_shape, dtype=np.float32) + return {"in1": inputs} + + work_dir = os.path.join(FILE_PATH, "work") + build_dir = os.path.join(FILE_PATH, "build") + quantize_func = TARGET_REGISTRY.get_target_quantizer(target) + q_xgraph = quantize_func(xgraph, inputs_func, work_dir=work_dir) + opt_xgraph = px.optimize(q_xgraph, target) + c_xgraph = px.compile( + opt_xgraph, target, work_dir=work_dir, build_dir=build_dir + ) + c_output = c_xgraph.get_compiler_output() + + assert list(c_output.keys()) == ["xp0"] + assert c_output.get_in_map("xp0") == {"xinput0": "xinput0:0"} + # assert c_output.get_out_map("xp0") == {pool_name: "pool1:0"} + assert len(c_output.get_code_files("xp0")) == 1 + + shutil.rmtree(work_dir) + shutil.rmtree(build_dir) + + def _create_scale_conv2d_nhwc_oihw( in_shape, w_shape, @@ -620,7 +668,7 @@ def inputs_func(iter): c_xgraph = px.compile( opt_xgraph, target, work_dir=work_dir, build_dir=build_dir ) - + g = xir.Graph.deserialize(os.path.join(build_dir, "xp0.xmodel")) # TODO subgraphs[1].get_attr("device") -> *** RuntimeError: bad any_cast subgraphs = get_child_subgraphs(g) @@ -628,4 +676,96 @@ def inputs_func(iter): len(subgraphs) == expected_nb_subgraphs ), "Expected {0} subgraphs but got: {1}".format( expected_nb_subgraphs, len(subgraphs) - ) \ No newline at end of file + ) + + +def _create_multi_output_conv2d_nhwc_oihw( + in_shape, + w_shape, + conv_padding, + conv_strides, + conv_dilation, + kernel_layout="OIHW", + target="DPUCZDX8G-zcu104", + out_names=["out1", "out2"], +) -> XGraph: + + kernel_w, kernel_h = w_shape[2], w_shape[3] + W = np.random.randint(-100, 100, size=w_shape).astype(np.float32) + # B = np.array([1., -1.], dtype=np.float32) + out_ch = w_shape[kernel_layout.index("O")] + + x1 = px.ops.input("in1", shape=list(in_shape)) + w1 = px.ops.constant("weight", W) + conv1 = px.ops.conv2d( + op_name="conv1", + input_layer=x1, + weights_layer=w1, + kernel_size=[kernel_w, kernel_h], + strides=list(conv_strides), + padding_hw=list(conv_padding), + dilation=list(conv_dilation), + data_layout="NHWC", + ) + r1 = px.ops.relu("r1", [conv1]) + + W2 = np.random.randint(-10, 10, size=(10, out_ch, 1, 1)).astype(np.float32) + w2 = px.ops.constant("weight2", W2) + conv2 = px.ops.conv2d( + op_name=out_names[0], + input_layer=r1, + weights_layer=w2, + kernel_size=[1, 1], + strides=[1, 1], + padding_hw=[0, 0], + dilation=[1, 1], + data_layout="NHWC", + ) + + W3 = np.random.randint(-10, 10, size=(10, out_ch, 1, 1)).astype(np.float32) + w3 = px.ops.constant("weight2", W3) + conv3 = px.ops.conv2d( + op_name="conv3", + input_layer=r1, + weights_layer=w3, + kernel_size=[1, 1], + strides=[1, 1], + padding_hw=[0, 0], + dilation=[1, 1], + data_layout="NHWC", + ) + r3 = px.ops.relu(out_names[1], [conv3]) + + net = [x1, conv1, r1, conv2, conv3, r3] + xgraph = XGRAPH_FACTORY.build_from_xlayer(net) + xgraph = px.partition(xgraph, [target]) + return xgraph + + +def multi_output_conv2d_naming_test(out_names) -> None: + + in_shape = (1, 20, 20, 10) + target = "DPUCZDX8G-zcu104" + xgraph = _create_multi_output_conv2d_nhwc_oihw( + in_shape, (10, 10, 2, 2), [0, 0], [1, 1], [1, 1], "OIHW", target, out_names, + ) + + def inputs_func(iter): + inputs = np.ones(in_shape, dtype=np.float32) + return {"in1": inputs} + + work_dir = os.path.join(FILE_PATH, "work") + build_dir = os.path.join(FILE_PATH, "build") + quantize_func = TARGET_REGISTRY.get_target_quantizer(target) + q_xgraph = quantize_func(xgraph, inputs_func, work_dir=work_dir) + opt_xgraph = px.optimize(q_xgraph, target) + c_xgraph = px.compile(opt_xgraph, target, work_dir=work_dir, build_dir=build_dir) + c_output = c_xgraph.get_compiler_output() + + assert list(c_output.keys()) == ["xp0"] + assert c_output.get_in_map("xp0") == {"xinput0": "xinput0:0"} + # assert c_output.get_out_map("xp0") == {pool_name: "pool1:0"} + assert len(c_output.get_code_files("xp0")) == 1 + + shutil.rmtree(work_dir) + shutil.rmtree(build_dir) diff --git a/tests/unit/contrib/test_dpuczdx8g.py b/tests/unit/contrib/test_dpuczdx8g.py index fa5798b..6992f06 100644 --- a/tests/unit/contrib/test_dpuczdx8g.py +++ b/tests/unit/contrib/test_dpuczdx8g.py @@ -16,7 +16,7 @@ import os import unittest - +import logging import numpy as np from pyxir import partition @@ -34,8 +34,14 @@ xcompiler_conv2d_leaky_relu_nhwc_oihw_test, conv2d_pool2d_nhwc_oihw_test, conv2d_leaky_relu_nhwc_oihw_test, + conv2d_pool2d_naming_test, + multi_output_conv2d_naming_test, ) +# logging.basicConfig() +# logger = logging.getLogger("pyxir") +# logger.setLevel(logging.DEBUG) + try: import tensorflow as tf @@ -408,7 +414,7 @@ def test_compile_conv2d_leaky_relu_dnnc(self): ], ) - @unittest.skipIf(not is_dpuczdx8g_vart_flow_enabled(), "DPUCZDX8G DNNC/DNNDK test") + @unittest.skipIf(not is_dpuczdx8g_vart_flow_enabled(), "DPUCZDX8G VART test") def test_compile_conv2d_leaky_relu(self): xcompiler_conv2d_leaky_relu_nhwc_oihw_test( (1, 4, 4, 1), @@ -424,6 +430,12 @@ def test_compile_conv2d_leaky_relu(self): ], ) + @unittest.skipIf(is_dpuczdx8g_vart_flow_enabled(), "DPUCZDX8G DNNC/DNNDK test") + def test_dnnc_out_names(self): + multi_output_conv2d_naming_test(["nn.conv-134", "nn.relu-22"]) + multi_output_conv2d_naming_test(["nn..con.v--1.", "n.n.-relu-2-"]) + conv2d_pool2d_naming_test(["conv1", "nn.conv-1"], ["nn.pool1", "nn.pool-1"]) + def test_supported_ops(self): ultra96_ops = TestDPUCZDX8G.target_registry.get_supported_op_check_names( "dpuv2-ultra96"