For encoded convolution, add check for when min_width would have been larger than in_width (#610)

* add special case when W == min_width.

* add test that fails without this fix

* add comments

* Make simple bug fixes before reworking structure

* Try to cleanup how we handle the narrow convolutions

* fix hardcoded returns

* update convolution templates for narrow encoded

* flake8 formatting

* update narrow for conv2d, fix test

* change narrow to unscaled, always do unscaled when in_W == min_W

* pre-commit fixes for nnet_conv1d_stream.h and nnet_conv2d_stream.h

* remove unneeded attributes

* change hardcoded instructions condition

Author: jmitrevs

hls4ml/backends/fpga/fpga_backend.py

@@ -394,6 +394,10 @@ def compute_conv1d_instructions(self, in_W, in_C, kernel_size=3, stride=1, pad=0
         else:
             min_W = (math.ceil(stride / kernel_size) - 1) * stride + kernel_size
 
+        # if the standard min_W is smaller than the in_W, then use unscaled
+        if min_W > in_W:
+            min_W = in_W
+
         min_oW = int((min_W - kernel_size) // stride + 1)
 
         out_W = int((in_W - kernel_size) // stride + 1)
@@ -442,11 +446,17 @@ def compute_conv2d_instructions(self, in_H, in_W, in_C, kernel_size=3, stride=1,
         else:
             min_H = (math.ceil(stride_height / kernel_height) - 1) * stride_height + kernel_height
 
+        if min_H > in_H:
+            min_H = in_H
+
         if kernel_width >= stride_width:
             min_W = (math.ceil(kernel_width / stride_width) - 1) * stride_width + kernel_width
         else:
             min_W = (math.ceil(stride_width / kernel_width) - 1) * stride_width + kernel_width
 
+        if min_W > in_W:
+            min_W = in_W
+
         min_oH = int((min_H - kernel_height) // stride_height + 1)
         min_oW = int((min_W - kernel_width) // stride_width + 1)
 
@@ -461,9 +471,25 @@ def compute_conv2d_instructions(self, in_H, in_W, in_C, kernel_size=3, stride=1,
             min_W += 1
 
         # Let's hardcode a few common cases:
-        if kernel_height == 1 and kernel_width == 1 and stride == 1 and scaled_H == in_H and scaled_W == in_W:
+        if (
+            min_H == 1
+            and min_W == 1
+            and kernel_height == 1
+            and kernel_width == 1
+            and stride == 1
+            and scaled_H == in_H
+            and scaled_W == in_W
+        ):
             return (1, 1, map(str, [1]))
-        if kernel_height == 3 and kernel_width == 3 and stride == 1 and scaled_H == in_H and scaled_W == in_W:
+        if (
+            min_H == 5
+            and min_W == 5
+            and kernel_height == 3
+            and kernel_width == 3
+            and stride == 1
+            and scaled_H == in_H
+            and scaled_W == in_W
+        ):
             return (
                 5,
                 5,
@@ -498,7 +524,15 @@ def compute_conv2d_instructions(self, in_H, in_W, in_C, kernel_size=3, stride=1,
                     ],
                 ),
             )
-        if kernel_height == 5 and kernel_width == 5 and stride == 1 and scaled_H == in_H and scaled_W == in_W:
+        if (
+            min_H == 9
+            and min_W == 9
+            and kernel_height == 5
+            and kernel_width == 5
+            and stride == 1
+            and scaled_H == in_H
+            and scaled_W == in_W
+        ):
             return (
                 9,
                 9,

hls4ml/backends/vivado/passes/conv_stream.py

@@ -1,8 +1,10 @@
+from hls4ml.model.layers import Conv1D, Conv2D, SeparableConv1D, SeparableConv2D
 from hls4ml.model.optimizer import OptimizerPass
-from hls4ml.model.layers import Conv1D, SeparableConv1D, Conv2D, SeparableConv2D
+
 
 class GenerateConvStreamingInstructions(OptimizerPass):
-    ''' Generates the instructions for streaming implementation of CNNs '''
+    '''Generates the instructions for streaming implementation of CNNs'''
+
     def match(self, node):
         return isinstance(node, (Conv1D, SeparableConv1D, Conv2D, SeparableConv2D))
 
@@ -13,19 +15,21 @@ def transform(self, model, node):
         elif '2D' in node_class:
             self._generate_2d_instructions(node)
         else:
-            raise Exception('Cannot generate instructions for node {} ({})'.format(node.name, node_class))
-    
+            raise Exception(f'Cannot generate instructions for node {node.name} ({node_class})')
+
     def _generate_1d_instructions(self, node):
         if node.model.config.get_config_value('IOType') == 'io_stream':
             min_w, instructions = node.model.config.backend.compute_conv1d_instructions(
                 node.get_input_variable().shape[0],
                 node.get_input_variable().shape[1],
                 node.get_attr('filt_width'),
-                node.get_attr('stride_width'))
+                node.get_attr('stride_width'),
+            )
             instructions_str = ','.join(str(i) for i in instructions)
             node.set_attr('min_width', min_w)
             node.set_attr('instructions', instructions_str)
         else:
+            # these are unused; just put dummy values
             node.set_attr('min_width', node.get_attr('in_width'))
             node.set_attr('instructions', '0')
 
@@ -36,7 +40,8 @@ def _generate_2d_instructions(self, node):
                 node.get_input_variable().shape[1],
                 node.get_input_variable().shape[2],
                 node.get_attr('filt_height'),
-                node.get_attr('stride_height'))
+                node.get_attr('stride_height'),
+            )
             instructions_str = ','.join(str(i) for i in instructions)
             node.set_attr('min_height', min_h)
             node.set_attr('min_width', min_w)