remove C_POST_CONVOLUTION_BITWIDTH

Instead calculate that bitwidth when needed.
marph91 · Feb 21, 2021 · 36f43d4 · 36f43d4
1 parent 59d65c6
commit 36f43d4
Show file tree

Hide file tree

Showing 5 changed files with 15 additions and 24 deletions.
diff --git a/sim/test_convolution.py b/sim/test_convolution.py
@@ -40,9 +40,7 @@ def input_weights_int(self) -> int:
  def output_data(self) -> int:
  ones_count = 0
  for act, weight in zip(self.input_activations, self.input_weights):
- ones_count = ones_count + (
- act and weight
- ) # TODO: should be xnor: (not (act ^ weight))
+ ones_count = ones_count + (not (act ^ weight))
  return ones_count
 
  input_channel = dut.C_INPUT_CHANNEL.value.integer

diff --git a/sim/test_window_convolution_activation.py b/sim/test_window_convolution_activation.py
@@ -42,7 +42,6 @@ async def run_test(dut):
  dut.C_INPUT_CHANNEL.value.integer,
  )
  output_channel = dut.C_OUTPUT_CHANNEL.value.integer
- post_convolution_bitwidth = dut.C_POST_CONVOLUTION_BITWIDTH.value.integer
 
  # define the reference model
  batch_shape = (1,) + image_shape
@@ -120,7 +119,6 @@ def get_threshold(self):
  # use batch normalization as activation
  # see also: https://arxiv.org/pdf/1612.07119.pdf, 4.2.2 Batchnorm-activation as Threshold
  threshold_batchnorm = mean - (beta / (variance * epsilon - gamma))
- print(threshold_batchnorm)
 
  # Conversion from LARQ format [-1, 1] to pocket-bnn format [0, 1] (positive only):
  # make the threshold compatible to positive only values
@@ -131,7 +129,7 @@ def get_threshold(self):
  threshold.append(int(threshold_pos))
 
  return concatenate_integers(
- self.replace_minus(threshold), bitwidth=post_convolution_bitwidth
+ self.replace_minus(threshold), bitwidth=math.ceil(math.log2(kernel_size[0] ** 2 * image_shape[2] + 1))
  )
 
  cases = (
@@ -220,7 +218,6 @@ def test_window_convolution_activation(record_waveform, kernel_size):
  "C_OUTPUT_CHANNEL": 8,
  "C_IMG_WIDTH": 4,
  "C_IMG_HEIGHT": 4,
- "C_POST_CONVOLUTION_BITWIDTH": 8,
  }
  run(
  vhdl_sources=get_files(

diff --git a/src/convolution.vhd b/src/convolution.vhd
@@ -3,22 +3,23 @@ library ieee;
  use ieee.std_logic_1164.all;
  use ieee.numeric_std.all;
 
+library util;
+ use util.math_pkg.all;
+
 entity convolution is
  generic (
  -- TODO: input bitwidth, for now = 1
 
  C_KERNEL_SIZE : integer range 2 to 7 := 2;
- C_INPUT_CHANNEL : integer := 1;
-
- C_POST_CONVOLUTION_BITWIDTH : integer := 8
+ C_INPUT_CHANNEL : integer := 1
  );
  port (
  isl_clk : in std_logic;
  isl_valid : in std_logic;
  islv_data : in std_logic_vector(C_KERNEL_SIZE * C_KERNEL_SIZE * C_INPUT_CHANNEL - 1 downto 0);
  -- maybe weights + 1 for bias
  islv_weights : in std_logic_vector(C_KERNEL_SIZE * C_KERNEL_SIZE * C_INPUT_CHANNEL - 1 downto 0);
- oslv_data : out std_logic_vector(C_POST_CONVOLUTION_BITWIDTH - 1 downto 0);
+ oslv_data : out std_logic_vector(log2(C_KERNEL_SIZE * C_KERNEL_SIZE * C_INPUT_CHANNEL + 1) - 1 downto 0);
  osl_valid : out std_logic
  );
 end entity convolution;
@@ -35,7 +36,6 @@ begin
 
  proc_convolution : process (isl_clk) is
 
- -- TODO: log2(islv_data'range)
  variable usig_ones_count : unsigned(oslv_data'range);
 
  begin

diff --git a/src/processing_element.vhd b/src/processing_element.vhd
@@ -10,10 +10,8 @@ entity processing_element is
  -- TODO: input bitwidth, for now = 1
 
  C_KERNEL_SIZE : integer range 2 to 3 := 2;
- C_INPUT_CHANNEL : integer;
+ C_INPUT_CHANNEL : integer
  -- C_OUTPUT_CHANNEL : integer; --> 1 output channel
-
- C_POST_CONVOLUTION_BITWIDTH : integer := 8
  );
  port (
  isl_clk : in std_logic;

diff --git a/src/window_convolution_activation.vhd b/src/window_convolution_activation.vhd
@@ -4,6 +4,8 @@ library ieee;
  use ieee.numeric_std.all;
 
 library cnn_lib;
+library util;
+ use util.math_pkg.all;
 
 library window_ctrl_lib;
 
@@ -18,9 +20,7 @@ entity window_convolution_activation is
  C_OUTPUT_CHANNEL : integer;
 
  C_IMG_WIDTH : integer;
- C_IMG_HEIGHT : integer;
-
- C_POST_CONVOLUTION_BITWIDTH : integer := 8
+ C_IMG_HEIGHT : integer
  );
  port (
  isl_clk : in std_logic;
@@ -29,7 +29,7 @@ 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(C_POST_CONVOLUTION_BITWIDTH * 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);
  osl_valid : out std_logic
  );
@@ -44,6 +44,7 @@ architecture behavioral of window_convolution_activation is
 
  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);
  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';
@@ -98,8 +99,7 @@ begin
  i_convolution : entity cnn_lib.convolution
  generic map (
  C_KERNEL_SIZE => C_KERNEL_SIZE,
- C_INPUT_CHANNEL => C_INPUT_CHANNEL,
- C_POST_CONVOLUTION_BITWIDTH => C_POST_CONVOLUTION_BITWIDTH
+ C_INPUT_CHANNEL => C_INPUT_CHANNEL
  )
  port map (
  isl_clk => isl_clk,
@@ -109,11 +109,9 @@ begin
  oslv_data => a_data_convolution(output_channel),
  osl_valid => sl_valid_convolution
  );
-
- -- TODO: output channel increments fastest, not slowest
+ -- output channel increments fastest
  a_weights(output_channel) <= get_fastest_increment(islv_weights, output_channel, C_OUTPUT_CHANNEL);
 
- -- one batch normalization for each output channel
  i_batch_normalization : entity cnn_lib.batch_normalization
  generic map (
  C_POST_CONVOLUTION_BITWIDTH => C_POST_CONVOLUTION_BITWIDTH