add support of variable output bitwidth

I. e. make the batch normalization optional. This is done to get viable class
scores at the output of the bnn.

Also fixed a small synthesis error, caused by multiple assignments at the
batch normalization signals.

sim/test_window_convolution_activation.py

@@ -154,7 +154,6 @@ def get_threshold(self):
  ),
  # mixed
  Testcase(
- # choice([-1, 1])
  [choice([-1, 1]) for _ in range(math.prod(image_shape))],
  [
  choice([-1, 1])
@@ -234,6 +233,7 @@ def get_threshold(self):
 def test_window_convolution_activation(
  record_waveform, kernel_size, stride, input_channel, output_channel
 ):
+ # TODO: Add test for output bitwidth /= 1.
  generics = {
  "C_KERNEL_SIZE": kernel_size,
  "C_STRIDE": stride,

src/window_convolution_activation.vhd

@@ -17,8 +17,9 @@ entity window_convolution_activation is
  C_KERNEL_SIZE : integer range 1 to 7 := 3;
  C_STRIDE : integer := 1;
 
- C_INPUT_CHANNEL : integer := 4;
- C_OUTPUT_CHANNEL : integer := 8;
+ C_INPUT_CHANNEL : integer := 4;
+ C_OUTPUT_CHANNEL : integer := 8;
+ C_OUTPUT_CHANNEL_BITWIDTH : integer range 1 to 32 := 1;
 
  C_IMG_WIDTH : integer := 4;
  C_IMG_HEIGHT : integer := 4
@@ -30,8 +31,8 @@ entity window_convolution_activation is
  islv_data : in std_logic_vector(C_INPUT_CHANNEL - 1 downto 0);
  -- islv_weights and islv_threshold are constants
  islv_weights : in std_logic_vector(C_KERNEL_SIZE * C_KERNEL_SIZE * C_INPUT_CHANNEL * C_OUTPUT_CHANNEL - 1 downto 0);
- islv_threshold : in std_logic_vector(log2(C_KERNEL_SIZE * C_KERNEL_SIZE * C_INPUT_CHANNEL + 1) * C_OUTPUT_CHANNEL - 1 downto 0);
- oslv_data : out std_logic_vector(C_OUTPUT_CHANNEL - 1 downto 0);
+ islv_threshold : in std_logic_vector(C_OUTPUT_CHANNEL * log2(C_KERNEL_SIZE ** 2 * C_INPUT_CHANNEL + 1) - 1 downto 0);
+ oslv_data : out std_logic_vector(C_OUTPUT_CHANNEL * C_OUTPUT_CHANNEL_BITWIDTH - 1 downto 0);
  osl_valid : out std_logic
  );
 end entity window_convolution_activation;
@@ -41,16 +42,16 @@ architecture behavioral of window_convolution_activation is
  signal sl_valid_window_ctrl : std_logic := '0';
  signal slv_data_window_ctrl : std_logic_vector(C_KERNEL_SIZE * C_KERNEL_SIZE * C_INPUT_CHANNEL - 1 downto 0);
 
- signal sl_valid_convolution : std_logic := '0';
+ signal slv_valid_convolution : std_logic_vector(C_OUTPUT_CHANNEL - 1 downto 0);
 
  type t_slv_array_1d is array(natural range <>) of std_logic_vector;
 
- constant C_POST_CONVOLUTION_BITWIDTH : integer := log2(C_KERNEL_SIZE * C_KERNEL_SIZE * C_INPUT_CHANNEL + 1);
+ constant C_POST_CONVOLUTION_BITWIDTH : integer := log2(C_KERNEL_SIZE ** 2 * C_INPUT_CHANNEL + 1);
  signal a_data_convolution : t_slv_array_1d(0 to C_OUTPUT_CHANNEL - 1)(C_POST_CONVOLUTION_BITWIDTH - 1 downto 0);
 
- signal sl_valid_batch_normalization : std_logic := '0';
- signal slv_data_batch_normalization : std_logic_vector(C_OUTPUT_CHANNEL * 1 - 1 downto 0);
- signal a_data_batch_normalization : t_slv_array_1d(0 to C_OUTPUT_CHANNEL - 1)(1 - 1 downto 0);
+ signal slv_valid_batch_normalization : std_logic_vector(C_OUTPUT_CHANNEL - 1 downto 0);
+ signal slv_data_batch_normalization  : std_logic_vector(C_OUTPUT_CHANNEL * C_OUTPUT_CHANNEL_BITWIDTH - 1 downto 0);
+ signal a_data_batch_normalization  : t_slv_array_1d(0 to C_OUTPUT_CHANNEL - 1)(C_OUTPUT_CHANNEL_BITWIDTH - 1 downto 0);
 
  signal sl_valid_out : std_logic := '0';
  signal slv_data_out : std_logic_vector(oslv_data'range);
@@ -108,31 +109,38 @@ begin
  islv_data => slv_data_window_ctrl,
  islv_weights => a_weights(output_channel),
  oslv_data => a_data_convolution(output_channel),
- osl_valid => sl_valid_convolution
+ osl_valid => slv_valid_convolution(output_channel)
  );
 
  -- output channel increments fastest
  a_weights(output_channel) <= get_fastest_increment(islv_weights, output_channel, C_OUTPUT_CHANNEL);
 
- i_batch_normalization : entity cnn_lib.batch_normalization
- generic map (
- C_POST_CONVOLUTION_BITWIDTH => C_POST_CONVOLUTION_BITWIDTH
- )
- port map (
- isl_clk => isl_clk,
- isl_valid => sl_valid_convolution,
- islv_data => a_data_convolution(output_channel),
- islv_threshold => a_threshold(output_channel),
- oslv_data => a_data_batch_normalization(output_channel),
- osl_valid => sl_valid_batch_normalization
- );
+ gen_batch_normalization : if C_OUTPUT_CHANNEL_BITWIDTH = 1 generate
+
+ i_batch_normalization : entity cnn_lib.batch_normalization
+ generic map (
+ C_POST_CONVOLUTION_BITWIDTH => C_POST_CONVOLUTION_BITWIDTH
+ )
+ port map (
+ isl_clk => isl_clk,
+ isl_valid => slv_valid_convolution(output_channel),
+ islv_data => a_data_convolution(output_channel),
+ islv_threshold => a_threshold(output_channel),
+ oslv_data => a_data_batch_normalization(output_channel),
+ osl_valid => slv_valid_batch_normalization(output_channel)
+ );
+
+ -- TODO: output channel increments fastest, not slowest
+ a_threshold(output_channel) <= get_slice(islv_threshold, output_channel, C_POST_CONVOLUTION_BITWIDTH);
+ slv_data_batch_normalization((output_channel + 1) * C_OUTPUT_CHANNEL_BITWIDTH - 1 downto output_channel * C_OUTPUT_CHANNEL_BITWIDTH) <= a_data_batch_normalization(output_channel);
+ else generate
+ slv_data_batch_normalization((output_channel + 1) * C_OUTPUT_CHANNEL_BITWIDTH - 1 downto output_channel * C_OUTPUT_CHANNEL_BITWIDTH) <= std_logic_vector(resize(unsigned(a_data_convolution(output_channel)), C_OUTPUT_CHANNEL_BITWIDTH));
+ slv_valid_batch_normalization(output_channel) <= slv_valid_convolution(output_channel);
+ end generate gen_batch_normalization;
 
- -- TODO: output channel increments fastest, not slowest
- a_threshold(output_channel) <= get_slice(islv_threshold, output_channel, C_POST_CONVOLUTION_BITWIDTH);
- slv_data_batch_normalization((output_channel + 1) * 1 - 1 downto output_channel * 1) <= a_data_batch_normalization(output_channel);
  end generate gen_convolution;
 
  oslv_data <= slv_data_batch_normalization;
- osl_valid <= sl_valid_batch_normalization;
+ osl_valid <= slv_valid_batch_normalization(0);
 
 end architecture behavioral;

src/window_ctrl/window_ctrl.vhd

@@ -13,7 +13,7 @@ library window_ctrl_lib;
 entity window_ctrl is
  generic (
  -- global data bitwidth
- C_BITWIDTH : integer range 1 to 16 := 8;
+ C_BITWIDTH : integer := 8;
 
  -- image properties
  C_CH_IN : integer range 1 to 512 := 1; -- input channel