improved convolution timing

Also removed duplicated testcase and added default generics for out-of-context
synthesis.

sim/test_window_convolution_activation.py

@@ -225,7 +225,6 @@ def get_threshold(self):
  (3, 1, 4, 8),
  (3, 1, 1, 4),
  (3, 1, 3, 8),
- (3, 1, 4, 8),
  (3, 1, 8, 16),
  (3, 2, 4, 8),
  (5, 1, 4, 8),

src/convolution.vhd

@@ -2,6 +2,7 @@
 library ieee;
  use ieee.std_logic_1164.all;
  use ieee.numeric_std.all;
+ use ieee.math_real.all;
 
 library util;
  use util.math_pkg.all;
@@ -10,7 +11,7 @@ entity convolution is
  generic (
  -- TODO: input bitwidth, for now = 1
 
- C_KERNEL_SIZE : integer range 1 to 7 := 2;
+ C_KERNEL_SIZE : integer range 1 to 7 := 3;
  C_INPUT_CHANNEL : integer := 1
  );
  port (
@@ -26,7 +27,33 @@ end entity convolution;
 
 architecture behavioral of convolution is
 
+ function find_best_parallelity return integer is
+ begin
+
+ if (C_INPUT_CHANNEL > C_KERNEL_SIZE) then
+ if (C_INPUT_CHANNEL mod 4 = 0) then
+ if (C_INPUT_CHANNEL mod 8 = 0) then
+ if (C_INPUT_CHANNEL mod 16 = 0) then
+ return 16;
+ end if;
+ return 8;
+ end if;
+ return 4;
+ end if;
+ end if;
+
+ return C_KERNEL_SIZE;
+ end function find_best_parallelity;
+
+ constant C_PARALLEL_POPCOUNT : integer := find_best_parallelity;
+ constant C_SPLIT : integer := integer(ceil(real(islv_data'length) / real(C_PARALLEL_POPCOUNT)));
+
+ type t_ones_count is array(natural range<>) of unsigned(oslv_data'range);
+
+ signal a_ones_count : t_ones_count(0 to C_SPLIT);
+
  signal sl_add : std_logic := '0';
+ signal sl_popcount : std_logic := '0';
  signal slv_multiplication_result : std_logic_vector(islv_data'range);
 
  signal sl_valid_out : std_logic := '0';
@@ -36,31 +63,49 @@ begin
 
  proc_convolution : process (isl_clk) is
 
- variable usig_ones_count : unsigned(oslv_data'range);
+ variable v_usig_popcount : unsigned(oslv_data'range);
+ variable v_usig_popcount_total : unsigned(oslv_data'range);
 
  begin
 
  if (rising_edge(isl_clk)) then
- sl_valid_out <= '0';
+ assert (C_INPUT_CHANNEL mod C_PARALLEL_POPCOUNT = 0) or
+ (C_KERNEL_SIZE mod C_PARALLEL_POPCOUNT = 0) severity failure;
+
+ sl_popcount <= '0';
  sl_add <= '0';
+ sl_valid_out <= '0';
 
  if (isl_valid = '1') then
  -- or map directly to hardware (islv_weights as constant)
  slv_multiplication_result <= islv_data xnor islv_weights;
+ a_ones_count <= (others => (others => '0'));
+ sl_popcount <= '1';
+ end if;
+
+ -- TODO: The split/adder can be improved.
+ if (sl_popcount = '1') then
+ for split in 0 to C_SPLIT - 1 loop
+ v_usig_popcount := (others => '0');
+ for i in 0 to C_PARALLEL_POPCOUNT - 1 loop
+ if (slv_multiplication_result(i + split * C_PARALLEL_POPCOUNT) = '1') then
+ v_usig_popcount := v_usig_popcount + 1;
+ end if;
+ end loop;
+ a_ones_count(split) <= v_usig_popcount;
+ end loop;
 
  sl_add <= '1';
  end if;
 
  if (sl_add = '1') then
- usig_ones_count := (others => '0');
- for i in slv_multiplication_result'range loop
- if (slv_multiplication_result(i) = '1') then
- usig_ones_count := usig_ones_count + 1;
- end if;
+ v_usig_popcount_total := (others => '0');
+ for split in 0 to C_SPLIT - 1 loop
+ v_usig_popcount_total := v_usig_popcount_total + a_ones_count(split);
  end loop;
 
- slv_data_out <= std_logic_vector(usig_ones_count);
  sl_valid_out <= '1';
+ slv_data_out <= std_logic_vector(v_usig_popcount_total);
  end if;
  end if;
 

src/window_convolution_activation.vhd

@@ -14,14 +14,14 @@ entity window_convolution_activation is
  generic (
  -- TODO: input bitwidth, for now = 1
 
- C_KERNEL_SIZE : integer range 1 to 7 := 2;
+ C_KERNEL_SIZE : integer range 1 to 7 := 3;
  C_STRIDE : integer := 1;
 
- C_INPUT_CHANNEL : integer;
- C_OUTPUT_CHANNEL : integer;
+ C_INPUT_CHANNEL : integer := 4;
+ C_OUTPUT_CHANNEL : integer := 8;
 
- C_IMG_WIDTH : integer;
- C_IMG_HEIGHT : integer
+ C_IMG_WIDTH : integer := 4;
+ C_IMG_HEIGHT : integer := 4
  );
  port (
  isl_clk : in std_logic;