Merge branch 're-arch-support' into tutorials

Author: dboyliao

utensor_cgen/backend/transformer.py

@@ -22,7 +22,10 @@ def __init__(self, config):
   def transform(self, ugraph):
     logger.info("Transforming graph: %s", ugraph.name)
     logger.info("Transform pipeline: %s", ' -> '.join(self.trans_methods))
-    self._check_non_quantized(ugraph)
+    if not self._check_generic(ugraph):
+      raise ValueError(
+        'the given graph is not generic:\n{}'.format(ugraph)
+      )
     new_ugraph = self.transformer.transform(ugraph)
     new_ugraph.name = ugraph.name
     logger.info('Graph transormation done')
@@ -35,26 +38,9 @@ def transform(self, ugraph):
     return new_ugraph
 
   @classmethod
-  def _check_non_quantized(cls, ugraph):
-    is_quantized = False
-    quant_ops = set([
-      "Dequantize", "QuantizedMaxPool",
-      "QuantizeV2", "QuantizedMatMul",
-      "QuantizedRelu", "QuantizedAdd",
-      "RequantizationRange",
-      "Requantize",
-      "QuantizedReshape",
-      "QuantizedConv2D"
-    ])
-    for op_info in ugraph.ops_info.values():
-      if op_info.op_type in quant_ops:
-        is_quantized = True
-        break
-    if is_quantized:
-      logger.warning((
-        "Expecting non-quantized graph, "
-        "graph transformation/optimization might not work properly"
-      ))
+  def _check_generic(cls, ugraph):
+    # TODO: do the real check once we have full list of generic ops
+    return True
 
   @class_property
   def default_config(cls):

utensor_cgen/backend/utensor/_graph_lower/_op_lower.py

@@ -56,6 +56,10 @@ def apply(cls, ugraph):
           op_info.op_type = 'QuantizedDepthwiseSeparableConvOperator'
         for op_info in ugraph.get_ops_by_type('FullyConnectedOperator'):
           op_info.op_type = 'QuantizedFullyConnectedOperator'
+      for op_info in ugraph.get_ops_by_type('DequantizeOperator'):
+        op_info.code_gen_attributes['namespaces'] = ('TFLM',)
+      for op_info in ugraph.get_ops_by_type('QuantizeOperator'):
+        op_info.code_gen_attributes['namespaces'] = ('TFLM',)
 
     @classmethod
     def _check_quantized(cls, ugraph):

utensor_cgen/backend/utensor/code_generator/rearch/_code_generator.py

@@ -1,18 +1,18 @@
 import re
+from collections import defaultdict
 from itertools import chain
 from pathlib import Path
-from collections import defaultdict
 
 from utensor_cgen.backend.base import BackendPart
+from utensor_cgen.backend.graph_lower.generic_graph_lower import \
+    TopoOrderTensorTimeslotPlanner
 from utensor_cgen.backend.utensor.snippets.composer import Composer
 from utensor_cgen.backend.utensor.snippets.legacy import (
     ContextGlobalArrayContainer, WeightSnippet)
 from utensor_cgen.backend.utensor.snippets.rearch import (
-    DeclareRamTensorSnippet, DeclareRomTensorSnippet, 
-    FreeTensorSnippet, SimpleContainer, TimeSlotContainer
-)
+    DeclareRamTensorSnippet, DeclareRomTensorSnippet, FreeTensorSnippet,
+    SimpleContainer, TimeSlotContainer)
 from utensor_cgen.backend.utensor.snippets.template_env import env
-from utensor_cgen.backend.graph_lower.generic_graph_lower import TopoOrderTensorTimeslotPlanner
 from utensor_cgen.logger import logger
 from utensor_cgen.utils import Configuration, class_property
 
@@ -183,8 +183,10 @@ def _time_slot_generate_files(
   ):
     template_vars = {}
     template_vars['model_name'] = ugraph.name
-    template_vars['meta_data_pool_size'] = self._compute_meta_data_size(ugraph)
-    template_vars['ram_data_pool_size'] = self._compute_ram_data_size(ugraph)
+    (template_vars['meta_data_pool_size'],
+     template_vars['meta_dtype']) = self._compute_meta_data_size(ugraph)
+    (template_vars['ram_data_pool_size'],
+     template_vars['ram_dtype']) = self._compute_ram_data_size(ugraph)
     template_vars['placeholders'] = placeholders
     template_vars['out_tensor_var_names'] = [
       tensor_var_map[tensor.name] for tensor in chain(*[
@@ -349,16 +351,23 @@ def default_config(cls):
     return config
 
   def _compute_meta_data_size(self, ugraph):
-    # TODO: if mem_optimizer is None, use a default mem optimizer
     if self.meta_data_pool_size == 'auto':
-      # TODO: compute actual meta data size with ugraph
-      size = 2048
+      # NOTE: simple heuristic, num of tensors * 64, maybe more or less depending on target platform
+      # NOTE: assuming user is using localCircularArenaAllocator
+      # TODO: target aware estimation
+      tensors = set()
+      for op_info in ugraph.ops_info.values():
+        tensors.update(op_info.input_tensors)
+        tensors.update(op_info.output_tensors)
+      size = len(tensors) * 64
     else:
       size = self.meta_data_pool_size
-    return size
+    dtype_str = self._get_mem_pool_dtype_str(size)
+    return size, dtype_str
 
   def _compute_ram_data_size(self, ugraph):
-    # TODO: if mem_optimizer is None, use a default mem optimizer
+    # TODO: if tensor alloc plan is None, use a default mem estimator
+    # NOTE: assuming user is using localCircularArenaAllocator
     if self.ram_data_pool_size == 'auto':
       # TODO: compute actual ram data size with ugraph
       if '_tensor_alloc' in ugraph.attributes:
@@ -367,4 +376,12 @@ def _compute_ram_data_size(self, ugraph):
         size = 256
     else:
       size = self.ram_data_pool_size
-    return size
+    dtype_str = self._get_mem_pool_dtype_str(size)
+    return size, dtype_str
+
+  @staticmethod
+  def _get_mem_pool_dtype_str(size):
+    # NOTE: assuming user is using localCircularArenaAllocator
+    if size > 2**15:
+      return 'uint32_t'
+    return 'uint16_t'

utensor_cgen/backend/utensor/code_generator/rearch/_operators/_impls.py

@@ -123,13 +123,14 @@ def get_eval_snippet(self, op_var_name, op_info, tensor_var_map):
 @OperatorFactory.register
 class _QuantizeOperator(_Operator):
   op_type = "QuantizeOperator"
+  namespaces = ('TFLM',)
 
   def get_declare_snippet(self, op_var_name, tensor_var_map):
     return DeclareOpSnippet(
       op=self,
       templ_dtypes=[self.out_dtypes[0], self.in_dtypes[0]],
       op_var_name=op_var_name,
-      nested_namespaces=['TFLM'],
+      nested_namespaces=self.namespaces,
     )
 
   def get_eval_snippet(self, op_var_name, op_info, tensor_var_map):
@@ -138,20 +139,21 @@ def get_eval_snippet(self, op_var_name, op_info, tensor_var_map):
       templ_dtypes=[self.out_dtypes[0], self.in_dtypes[0]],
       op_name=op_var_name,
       tensor_var_map=tensor_var_map,
-      nested_namespaces=['TFLM'],
+      nested_namespaces=self.namespaces,
     )
 
 
 @OperatorFactory.register
 class _DequantizeOperator(_Operator):
   op_type = "DequantizeOperator"
+  namespaces = ('TFLM',)
 
   def get_declare_snippet(self, op_var_name, tensor_var_map):
     return DeclareOpSnippet(
       op=self,
       templ_dtypes=[self.out_dtypes[0], self.in_dtypes[0]],
       op_var_name=op_var_name,
-      nested_namespaces=['TFLM'],
+      nested_namespaces=self.namespaces,
     )
 
   def get_eval_snippet(self, op_var_name, op_info, tensor_var_map):
@@ -160,7 +162,7 @@ def get_eval_snippet(self, op_var_name, op_info, tensor_var_map):
       templ_dtypes=[self.out_dtypes[0], self.in_dtypes[0]],
       op_name=op_var_name,
       tensor_var_map=tensor_var_map,
-      nested_namespaces=['TFLM'],
+      nested_namespaces=self.namespaces,
     )
 
 

utensor_cgen/backend/utensor/snippets/templates/snippets/rearch/simple.hpp

@@ -3,8 +3,8 @@
 
 #include "uTensor/core/tensor.hpp"
 
-// estimated ram usage: {{ram_data_pool_size}} bytes
-// estimated meta data uage: {{meta_data_pool_size}} bytes
+constexpr {{ram_dtype}} estimated_ram_usage = {{ram_data_pool_size}};
+constexpr {{meta_dtype}} estimated_meta_usage = {{meta_data_pool_size}};
 
 void compute_{{model_name}}({%for pl in placeholders%}uTensor::Tensor& {{pl}}, {%endfor%}{%for out_tensor in out_tensor_var_names%}uTensor::Tensor& {{out_tensor}}{%if not loop.last%}, {%endif%}{%endfor%});
 

utensor_cgen/cli/backend.py

@@ -43,8 +43,8 @@ def list_trans_methods(verbose):
 
 @cli.command(name='list-support-ops', help='list all supported op in the backend')
 @click.help_option('-h', '--help')
-@click.option('--target', default='utensor')
-@click.option('--config', default='utensor_cli.toml')
+@click.option('--target', default='utensor', show_default=True)
+@click.option('--config', default='utensor_cli.toml', show_default=True)
 def list_support_ops(target, config):
   from utensor_cgen.backend.api import BackendManager
   if os.path.exists(config):