Rework DMA to support single-beat transfers

fpga/rtl/dma_axi_master.vhd

@@ -1,6 +1,13 @@
 --
 --  AXI3 master for multi-channel DMA controller.
 --
+--  Note: DMA transfers may not cross a 4 kB address boundary.
+--
+--  It is recommended to use only so-called regular transactions.
+--  A transaction is regular if its length is 1, 2, 4, 8 or 16 beats,
+--  and its address is aligned to the burst length.
+--  Regular transactions never cross a 4 kB boundary.
+--
 --  Joris van Rantwijk 2024
 --
 
@@ -14,9 +21,6 @@ use work.puzzlefw_pkg.all;
 entity dma_axi_master is
 
     generic (
-        -- Number of beats per transfer.
-        transfer_size:      integer range 1 to 16 := 16;
-
         -- Number of read channels.
         num_read_channels:  integer range 0 to 16;
 
@@ -57,10 +61,11 @@ entity dma_axi_master is
 
         -- Read channels.
         -- The client passes a read command for a specified address via valid/ready handshake.
-        -- Some time later, the controller pushes (transfer_size) data words out to the channel.
+        -- Some time later, the controller pushes (read_cmd_length + 1) data words out to the channel.
         -- The client must be ready to accept all data words.
         -- Multiple transfers can be in flight for the channel.
         read_cmd_addr:      in  dma_address_array(0 to num_read_channels-1);
+        read_cmd_length:    in  dma_burst_length_array(0 to num_read_channels-1);
         read_cmd_valid:     in  std_logic_vector(num_read_channels-1 downto 0);
         read_cmd_ready:     out std_logic_vector(num_read_channels-1 downto 0);
         read_data:          out dma_data_array(0 to num_read_channels-1);
@@ -68,11 +73,12 @@ entity dma_axi_master is
 
         -- Write channels.
         -- The client passes a write command for a specified address via valid/ready handshake.
-        -- Some time later, the controller pulls (transfer_size) data words from the channel.
+        -- Some time later, the controller pulls (write_cmd_length + 1) data words from the channel.
         -- The client must supply these data words promptly.
         -- Some more time later, the controller pulses write_finished to indicate that the write has finished.
         -- Multiple transfers can be in flight for the channel.
         write_cmd_addr:     in  dma_address_array(0 to num_write_channels-1);
+        write_cmd_length:   in  dma_burst_length_array(0 to num_write_channels-1);
         write_cmd_valid:    in  std_logic_vector(num_write_channels-1 downto 0);
         write_cmd_ready:    out std_logic_vector(num_write_channels-1 downto 0);
         write_data:         in  dma_data_array(0 to num_write_channels-1);
@@ -131,11 +137,13 @@ architecture arch of dma_axi_master is
 
         -- Registered output signals to AXI bus.
         awaddr:             std_logic_vector(31 downto 3);
+        awlen:              std_logic_vector(3 downto 0);
         awvalid:            std_logic;
         wdata:              std_logic_vector(63 downto 0);
         wlast:              std_logic;
         wvalid:             std_logic;
         araddr:             std_logic_vector(31 downto 3);
+        arlen:              std_logic_vector(3 downto 0);
         arvalid:            std_logic;
 
         -- Registered output signals to read channels.
@@ -171,11 +179,13 @@ architecture arch of dma_axi_master is
 
     constant regs_init: regs_type := (
         awaddr          => (others => '0'),
+        awlen           => (others => '0'),
         awvalid         => '0',
         wdata           => (others => '0'),
         wlast           => '0',
         wvalid          => '0',
         araddr          => (others => '0'),
+        arlen           => (others => '0'),
         arvalid         => '0',
         read_cmd_ready  => (others => '0'),
         read_data       => (others => '0'),
@@ -204,11 +214,12 @@ architecture arch of dma_axi_master is
 
     -- Check that the DMA transfer fits inside the address window if it starts at the specified address offset 
     function is_valid_dma_address(addr: std_logic_vector(31 downto 3);
+                                  len: std_logic_vector(3 downto 0);
                                   limit: std_logic_vector(31 downto 12))
         return boolean
     is begin
         return (unsigned(limit) /= 0)
-                and (unsigned(addr) <= shift_left(resize(unsigned(limit), 29), 9) - transfer_size);
+                and (unsigned(addr) < shift_left(resize(unsigned(limit), 29), 9) - unsigned(len));
     end function;
 
     -- Calculate tha AXI address for a DMA transfer by adding the address offset from the client
@@ -224,8 +235,6 @@ architecture arch of dma_axi_master is
 begin
 
     -- Drive fixed output signals to AXI bus.
-    m_axi_awlen     <= std_logic_vector(to_unsigned(transfer_size - 1, 4));  -- use fixed burst length
-    m_axi_arlen     <= std_logic_vector(to_unsigned(transfer_size - 1, 4));  -- use fixed burst length
     m_axi_awsize    <= "011";   -- always use 64-bit transfers
     m_axi_arsize    <= "011";   -- always use 64-bit transfers
     m_axi_awburst   <= "01";    -- always use incrementing burst
@@ -245,6 +254,7 @@ begin
     -- Drive variable output signals to AXI bus.
     m_axi_awid      <= std_logic_vector(to_unsigned(r.write_channel, 6));
     m_axi_awaddr    <= r.awaddr & "000";  -- addresses are 8-byte aligned
+    m_axi_awlen     <= r.awlen;
     m_axi_awvalid   <= r.awvalid;
     m_axi_wid       <= std_logic_vector(to_unsigned(r.write_channel, 6));
     m_axi_wdata     <= r.wdata;
@@ -252,6 +262,7 @@ begin
     m_axi_wvalid    <= r.wvalid;
     m_axi_arid      <= std_logic_vector(to_unsigned(r.read_channel, 6));
     m_axi_araddr    <= r.araddr & "000";  -- addresses are 8-byte aligned
+    m_axi_arlen     <= r.arlen;
     m_axi_arvalid   <= r.arvalid;
 
     -- Drive output signals to read channels.
@@ -340,23 +351,26 @@ begin
                 when WRITE_STATE_START =>
                     -- Calculate the AXI address by adding the client address offset to the window base address.
                     v.awaddr := calc_dma_address(write_cmd_addr(r.write_channel), window_base_addr);
+                    v.awlen := write_cmd_length(r.write_channel);
 
                     -- Setup first data word.
                     v.wdata := write_data(r.write_channel);
-                    v.cnt_write_beat := (others => '0');
+                    v.cnt_write_beat := unsigned(write_cmd_length(r.write_channel));
-                    if transfer_size = 1 then
+                    if unsigned(write_cmd_length(r.write_channel)) = 0 then
                         v.wlast := '1';
                     else
                         v.wlast := '0';
                     end if;
 
                     -- Check address.
-                    if is_valid_dma_address(write_cmd_addr(r.write_channel), window_size) then
+                    if is_valid_dma_address(write_cmd_addr(r.write_channel),
+                                            write_cmd_length(r.write_channel),
+                                            window_size) then
                         -- Setup AXI write burst.
                         v.awvalid := '1';
                         v.wvalid := '1';
                         v.cnt_write_start := r.cnt_write_start + 1;
-                        if transfer_size = 1 then
+                        if unsigned(write_cmd_length(r.write_channel)) = 0 then
                             v.write_state := WRITE_STATE_WAIT;
                         else
                             v.write_state := WRITE_STATE_FLOW;
@@ -385,14 +399,14 @@ begin
                     -- Push data words to interconnect.
                     if m_axi_wready = '1' then
                         v.wdata := write_data(r.write_channel);
-                        if r.cnt_write_beat = transfer_size - 2 then
+                        if r.cnt_write_beat = 1 then
                             -- This will be the last beat of the transfer.
                             v.wlast := '1';
                             v.write_state := WRITE_STATE_WAIT;
                         else
                             v.wlast := '0';
                         end if;
-                        v.cnt_write_beat := r.cnt_write_beat + 1;
+                        v.cnt_write_beat := r.cnt_write_beat - 1;
                     end if;
 
                 when WRITE_STATE_WAIT =>
@@ -470,9 +484,12 @@ begin
 
                     -- Calculate the AXI address by adding the client address offset to the window base address.
                     v.araddr := calc_dma_address(read_cmd_addr(r.read_channel), window_base_addr);
+                    v.arlen := read_cmd_length(r.read_channel);
 
                     -- Check address.
-                    if is_valid_dma_address(read_cmd_addr(r.read_channel), window_size) then
+                    if is_valid_dma_address(read_cmd_addr(r.read_channel),
+                                            read_cmd_length(r.read_channel),
+                                            window_size) then
                         -- Setup AXI read burst.
                         v.arvalid := '1';
                         v.read_state := READ_STATE_WAIT;

fpga/rtl/dma_write_channel.vhd

@@ -1,8 +1,6 @@
 --
 --  Management of DMA transfers from FPGA to memory.
 --
---  The AXI master MUST be configured for 16 beats per transfer.
---
 --  Joris van Rantwijk 2024
 --
 
@@ -16,8 +14,16 @@ use work.puzzlefw_pkg.all;
 entity dma_write_channel is
 
     generic (
+        -- Preferred DMA burst length as the 2-log of the number of beats.
+        -- Transactions will use the preferred burst length unless there
+        -- are fewer words available in the buffer.
+        transfer_size_bits: integer range 0 to 4 := 4;
+
         -- Size of the input data queue as 2-log of the number of words.
-        queue_size_bits:    integer range 5 to 16 := 10
+        queue_size_bits:    integer range 5 to 16 := 10;
+
+        -- Number of clock cycles to wait before starting a small DMA burst.
+        idle_timeout:       natural := 256
     );
 
     port (
@@ -32,6 +38,10 @@ entity dma_write_channel is
         -- no new transfer will be started.
         channel_en:         in  std_logic;
 
+        -- High if there uncompleted DMA transfers are in progress.
+        -- If a DMA error occurs, the channel may get stuck in a busy state.
+        channel_busy:       out std_logic;
+
         -- High to initialize the channel.
         --
         -- This resets the write pointer to the start of the buffer and deletes
@@ -39,7 +49,7 @@ entity dma_write_channel is
         --
         -- This function must be used whenever the buffer address range is changed.
         -- This function must not be used while uncompleted DMA transfers
-        -- are in progress.
+        -- are in progress, except to recover from an error.
         channel_init:       in  std_logic;
 
         -- Start and end address of the DMA buffer.
@@ -55,10 +65,10 @@ entity dma_write_channel is
 
         -- An interrupt is triggered when the write pointer equals the
         -- interrupt address.
-        addr_interrupt:     in  std_logic_vector(31 downto 7);
+        addr_interrupt:     in  std_logic_vector(31 downto 3);
 
         -- Write pointer.
-        addr_pointer:       out std_logic_vector(31 downto 7);
+        addr_pointer:       out std_logic_vector(31 downto 3);
 
         -- High to enable interrupt on reaching a configured address.
         intr_en:            in  std_logic;
@@ -77,6 +87,7 @@ entity dma_write_channel is
 
         -- Signals to AXI master.
         write_cmd_addr:     out dma_address_type;
+        write_cmd_length:   out dma_burst_length_type;
         write_cmd_valid:    out std_logic;
         write_cmd_ready:    in  std_logic;
         write_data:         out dma_data_type;
@@ -88,37 +99,56 @@ end entity;
 
 architecture arch of dma_write_channel is
 
-    -- Number of beats per DMA transfer.
+    constant transfer_size: integer range 1 to 16 := 2**transfer_size_bits;
-    -- The AXI master must use the same transfer size.
-    -- Address alignments must match the transfer size.
-    constant transfer_size: integer := 16;
 
-    type state_type is (STATE_IDLE, STATE_START, STATE_FLOW);
+    type state_type is (STATE_SINGLE_IDLE, STATE_SINGLE_START, STATE_SINGLE_DATA,
+                        STATE_BURST_PREPARE,
+                        STATE_BURST_IDLE, STATE_BURST_START, STATE_BURST_DATA);
 
     type regs_type is record
         cmd_valid:          std_logic;
-        cmd_addr:           std_logic_vector(31 downto 7);
+        cmd_addr:           std_logic_vector(31 downto 3);
-        addr_pointer:       std_logic_vector(31 downto 7);
+        cmd_full_burst:     std_logic;
+        addr_pointer:       std_logic_vector(31 downto 3);
         intr_out:           std_logic;
+        channel_busy:       std_logic;
         state:              state_type;
-        pending_beats:      unsigned(4 downto 0);
+        pending_beats:      unsigned(3 downto 0);
+        pending_transfers:  unsigned(3 downto 0);
+        idle_timer:         integer range 0 to idle_timeout - 1;
     end record;
 
     constant regs_init: regs_type := (
         cmd_valid           => '0',
         cmd_addr            => (others => '0'),
+        cmd_full_burst      => '0',
         addr_pointer        => (others => '0'),
         intr_out            => '0',
-        state           => STATE_IDLE,
+        channel_busy        => '0',
-        pending_beats   => (others => '0')
+        state               => STATE_SINGLE_IDLE,
+        pending_beats       => (others => '0'),
+        pending_transfers   => (others => '0'),
+        idle_timer          => 0
     );
 
     signal r: regs_type := regs_init;
     signal rnext: regs_type;
 
     signal s_fifo_reset:    std_logic;
+    signal s_fifo_valid:    std_logic;
     signal s_fifo_length:   unsigned(queue_size_bits - 1 downto 0);
 
+    -- Return true if address is aligned to a full-size burst.
+    function is_address_aligned(addr: std_logic_vector(31 downto 3))
+        return boolean
+    is begin
+        if transfer_size = 1 then
+            return true;
+        else
+            return (unsigned(addr(transfer_size_bits + 2 downto 3)) = 0);
+        end if;
+    end function;
+
 begin
 
     --
@@ -134,15 +164,18 @@ begin
             in_valid            => in_valid,
             in_ready            => in_ready,
             in_data             => in_data,
-            out_valid           => open,
+            out_valid           => s_fifo_valid,
             out_ready           => write_data_ready,
             out_data            => write_data,
             queue_length        => s_fifo_length ); 
 
     -- Drive output ports.
+    channel_busy        <= r.channel_busy;
     addr_pointer        <= r.addr_pointer;
     intr_out            <= r.intr_out;
-    write_cmd_addr  <= r.cmd_addr & "0000";
+    write_cmd_addr      <= r.cmd_addr;
+    write_cmd_length    <= "0000" when (r.cmd_full_burst = '0') else
+                           std_logic_vector(to_unsigned(transfer_size - 1, 4));
     write_cmd_valid     <= r.cmd_valid;
 
     -- Drive reset signal to FIFO.
@@ -160,36 +193,137 @@ begin
         -- State machine.
         case r.state is
 
-            when STATE_IDLE =>
+            when STATE_SINGLE_IDLE =>
-                -- Issue a new transfer when possible.
+
-                -- Data for the whole transfer must be ready in the FIFO.
+                v.cmd_full_burst := '0';
-                if (channel_en = '1')
+
-                        and (s_fifo_length >= transfer_size)
+                -- Clear busy flag.
-                        and (r.cmd_addr /= addr_limit) then
+                if r.pending_transfers = 0 then
-                    v.cmd_valid := '1';
+                    v.channel_busy := '0';
-                    v.state := STATE_START;
-                    v.pending_beats := to_unsigned(transfer_size, v.pending_beats'length);
                 end if;
 
-            when STATE_START =>
+                -- Update idle timer.
+                if (s_fifo_valid = '1') and (r.idle_timer /= idle_timeout - 1) then
+                    v.idle_timer := r.idle_timer + 1;
+                end if;
+
+                -- Switch to burst mode if the FIFO has enough data,
+                -- and the address is aligned to a full burst.
+                if (transfer_size > 1)
+                        and (s_fifo_length >= transfer_size)
+                        and is_address_aligned(r.cmd_addr) then
+
+                    v.state := STATE_BURST_PREPARE;
+
+                -- Otherwise, start a single-beat transfer when possible.
+                elsif (channel_en = '1')
+                        and (s_fifo_valid = '1')
+                        and (r.cmd_addr(31 downto 7) /= addr_limit)
+                        and (r.pending_transfers /= 15)
+                        and ((transfer_size = 1)
+                             or (not is_address_aligned(r.cmd_addr))
+                             or (r.idle_timer = idle_timeout - 1)) then
+
+                    v.cmd_valid := '1';
+                    v.channel_busy := '1';
+                    v.pending_beats := to_unsigned(0, v.pending_beats'length);
+                    v.state := STATE_SINGLE_START;
+
+                end if;
+
+            when STATE_SINGLE_START =>
+                -- Wait until start of transfer.
+
+                if write_cmd_ready = '1' then
+                    v.cmd_valid := '0';
+
+                    -- Update pending transfers.
+                    v.pending_transfers := r.pending_transfers + 1;
+
+                    -- Update command address.
+                    v.cmd_addr := std_logic_vector(unsigned(r.cmd_addr) + 1);
+                    if v.cmd_addr(31 downto 7) = addr_end then
+                        -- Wrap at end of buffer.
+                        v.cmd_addr := addr_start & "0000";
+                    end if;
+
+                    if write_data_ready = '1' then
+                        v.state := STATE_SINGLE_IDLE;
+                    else
+                        v.state := STATE_SINGLE_DATA;
+                    end if;
+                end if;
+
+                -- Reset idle counter.
+                v.idle_timer := 0;
+
+            when STATE_SINGLE_DATA =>
+
+                -- Wait until data accepted.
+                if (write_data_ready = '1') and (r.pending_beats = 0) then
+                    v.state := STATE_SINGLE_IDLE;
+                end if;
+
+            when STATE_BURST_PREPARE =>
+
+                -- Wait until pending transfers are completed.
+                if r.pending_transfers = 0 then
+                    v.state := STATE_BURST_IDLE;
+                end if;
+
+                -- Reset idle counter.
+                v.idle_timer := 0;
+
+            when STATE_BURST_IDLE =>
+
+                v.cmd_full_burst := '1';
+
+                -- Clear busy flag.
+                if r.pending_transfers = 0 then
+                    v.channel_busy := '0';
+                end if;
+
+                -- Switch to single transfer mode if FIFO has insufficient data.
+                if (s_fifo_length < transfer_size)
+                        and (r.pending_transfers = 0) then
+                    v.state := STATE_SINGLE_IDLE;
+                end if;
+
+                -- Start a burst transfer if possible.
+                if (channel_en = '1')
+                        and (s_fifo_length >= transfer_size)
+                        and (r.cmd_addr(31 downto 7) /= addr_limit)
+                        and (r.pending_transfers /= 15) then
+                    v.cmd_valid := '1';
+                    v.channel_busy := '1';
+                    v.pending_beats := to_unsigned(transfer_size - 1, v.pending_beats'length);
+                    v.state := STATE_BURST_START;
+                end if;
+
+            when STATE_BURST_START =>
+
                 -- Wait until start of transfer.
                 if write_cmd_ready = '1' then
                     v.cmd_valid := '0';
 
+                    -- Update pending transfers.
+                    v.pending_transfers := r.pending_transfers + 1;
+
                     -- Update command address.
-                    v.cmd_addr := std_logic_vector(unsigned(r.cmd_addr) + 1);
+                    v.cmd_addr := std_logic_vector(unsigned(r.cmd_addr) + transfer_size);
-                    if v.cmd_addr = addr_end then
+                    if v.cmd_addr(31 downto 7) = addr_end then
                         -- Wrap at end of buffer.
-                        v.cmd_addr := addr_start;
+                        v.cmd_addr := addr_start & "0000";
                     end if;
 
-                    v.state := STATE_FLOW;
+                    v.state := STATE_BURST_DATA;
                 end if;
 
-            when STATE_FLOW =>
+            when STATE_BURST_DATA =>
+
                 -- Wait until last beat.
-                if (write_data_ready = '1') and (r.pending_beats = 1) then
+                if (write_data_ready = '1') and (r.pending_beats = 0) then
-                    v.state := STATE_IDLE;
+                    v.state := STATE_BURST_IDLE;
                 end if;
 
         end case;
@@ -200,17 +334,29 @@ begin
         end if;
 
         -- Update pointer on write completion.
-        if write_finished = '1' then
+        if (write_finished = '1') and (r.pending_transfers /= 0) then
-            v.addr_pointer := std_logic_vector(unsigned(r.addr_pointer) + 1);
+            v.pending_transfers := v.pending_transfers - 1;
-            if v.addr_pointer = addr_end then
+
-                -- Wrap at end of buffer.
+            if r.cmd_full_burst = '1' then
-                v.addr_pointer := addr_start;
+                -- Completion of a full burst.
+                v.addr_pointer := std_logic_vector(unsigned(r.addr_pointer) + transfer_size);
+
+                -- Raise interrupt when pointer passes threshold.
+                if (intr_en = '1')
+                    and (r.addr_pointer(31 downto 3 + transfer_size_bits)
+                         = addr_interrupt(31 downto 3 + transfer_size_bits)) then
+                    v.intr_out := '1';
                 end if;
+            else
+                -- Completion of single transfer.
+                v.addr_pointer := std_logic_vector(unsigned(r.addr_pointer) + 1);
+            end if;
+
+            -- Wrap at end of buffer.
+            if v.addr_pointer(31 downto 7) = addr_end then
+                v.addr_pointer := addr_start & "0000";
             end if;
 
-        -- Clear interrupt.
-        if intr_clear = '1' then
-            v.intr_out := '0';
         end if;
 
         -- Raise interrupt when pointer equals threshold.
@@ -218,13 +364,18 @@ begin
             v.intr_out := '1';
         end if;
 
-        -- Initialize channel.
+        -- Clear interrupt.
-        if channel_init = '1' then
+        if intr_clear = '1' then
-            v.cmd_valid := '0';
-            v.cmd_addr := addr_start;
-            v.addr_pointer := addr_start;
             v.intr_out := '0';
-            v.state := STATE_IDLE;
+        end if;
+
+        -- Initialize pointers.
+        if channel_init = '1' then
+            v.cmd_addr := addr_start & "0000";
+            v.addr_pointer := addr_start & "0000";
+            v.state := STATE_SINGLE_IDLE;
+            v.pending_transfers := (others => '0');
+            v.idle_timer := 0;
         end if;
 
         -- Synchronous reset.

fpga/rtl/puzzlefw_pkg.vhd

@@ -15,10 +15,27 @@ package puzzlefw_pkg is
     subtype dma_address_type is std_logic_vector(31 downto 3);
     type dma_address_array is array(natural range <>) of dma_address_type;
 
+    -- Burst length for DMA on AXI bus as number of beats minus 1.
+    subtype dma_burst_length_type is std_logic_vector(3 downto 0);
+    type dma_burst_length_array is array (natural range <>) of dma_burst_length_type;
+
     -- 64-bit data for DMA on AXI bus.
     subtype dma_data_type is std_logic_vector(63 downto 0);
     type dma_data_array is array(natural range <>) of dma_data_type;
 
+    -- 14-bit ADC data.
+    constant adc_data_bits: integer := 14;
+    subtype adc_data_type is std_logic_vector(adc_data_bits - 1 downto 0);
+    type adc_data_array is array(natural range <>) of adc_data_type;
+
+    -- 24-bit averaged data.
+    constant sample_data_bits: integer := 24;
+    subtype sample_data_type is std_logic_vector(sample_data_bits - 1 downto 0);
+    type sample_data_array is array(natural range <>) of sample_data_type;
+
+    -- 48-bit timestamp.
+    constant timestamp_bits: integer := 48;
+
     -- Register addresses.
     constant reg_addr_mask:         std_logic_vector(31 downto 0) := x"0010fffc";
     constant reg_info:              natural := 16#000000#;
@@ -35,6 +52,7 @@ package puzzlefw_pkg is
     constant reg_acq_channel_ctrl:  natural := 16#000214#;
     constant reg_acq_intr_ctrl:     natural := 16#000218#;
     constant reg_test_led:          natural := 16#000404#;
+    constant reg_test_divider:      natural := 16#000408#;
     constant reg_dma_buf_addr:      natural := 16#100000#;
     constant reg_dma_buf_size:      natural := 16#100004#;
 
@@ -48,6 +66,11 @@ package puzzlefw_pkg is
         & std_logic_vector(to_unsigned(fw_version_major, 8))
         & std_logic_vector(to_unsigned(fw_version_minor, 8));
 
+    -- Data stream.
+    constant msg_adc_data:          std_logic_vector(7 downto 0) := x"01";
+    constant msg_trigger:           std_logic_vector(7 downto 0) := x"02";
+    constant msg_overflow:          std_logic_vector(7 downto 0) := x"10";
+
     -- ADC input port type.
     type adc_data_input_type is array(0 to 1) of std_logic_vector(15 downto 0);
 
@@ -55,11 +78,12 @@ package puzzlefw_pkg is
     type registers_control is record
         irq_enable:         std_logic;
         test_led:           std_logic_vector(7 downto 0);
+        test_divider:       std_logic_vector(15 downto 0);
         dma_en:             std_logic;
         acq_addr_start:     std_logic_vector(31 downto 7);
         acq_addr_end:       std_logic_vector(31 downto 7);
         acq_addr_limit:     std_logic_vector(31 downto 7);
-        acq_addr_intr:      std_logic_vector(31 downto 7);
+        acq_addr_intr:      std_logic_vector(31 downto 3);
         acq_channel_en:     std_logic;
         acq_intr_en:        std_logic;
         dma_buf_addr:       std_logic_vector(31 downto 12);
@@ -74,7 +98,8 @@ package puzzlefw_pkg is
         dma_err_write:      std_logic;
         dma_err_address:    std_logic;
         dma_err_any:        std_logic;
-        acq_addr_ptr:       std_logic_vector(31 downto 7);
+        acq_addr_ptr:       std_logic_vector(31 downto 3);
+        acq_channel_busy:   std_logic;
     end record;
 
     -- Trigger registers: write-only access by processor, single-cycle pulse signals to FPGA.
@@ -87,6 +112,7 @@ package puzzlefw_pkg is
     constant registers_control_init: registers_control := (
         irq_enable      => '0',
         test_led        => (others => '0'),
+        test_divider    => (others => '0'),
         dma_en          => '0',
         acq_addr_start  => (others => '0'),
         acq_addr_end    => (others => '0'),

fpga/rtl/puzzlefw_top.vhd

@@ -137,6 +137,7 @@ architecture arch of puzzlefw_top is
     signal s_reg_trigger:   registers_trigger;
 
     signal s_dma_write_cmd_addr:  dma_address_array(0 downto 0);
+    signal s_dma_write_cmd_length: dma_burst_length_array(0 to 0);
     signal s_dma_write_cmd_valid: std_logic_vector(0 downto 0);
     signal s_dma_write_cmd_ready: std_logic_vector(0 downto 0);
     signal s_dma_write_data:      dma_data_array(0 downto 0);
@@ -150,6 +151,7 @@ architecture arch of puzzlefw_top is
     signal r_test_prefetch: std_logic;
     signal r_test_raddr:    std_logic_vector(9 downto 0);
     signal r_test_waddr:    std_logic_vector(9 downto 0);
+    signal r_test_cnt:      unsigned(15 downto 0);
     signal s_test_dout:     std_logic_vector(63 downto 0);
     signal s_test_ren:      std_logic;
 
@@ -281,7 +283,6 @@ begin
     -- AXI master.
     inst_axi_master: entity work.dma_axi_master
         generic map (
-            transfer_size       => 16,
             num_read_channels   => 0,
             num_write_channels  => 1 )
         port map (
@@ -297,11 +298,13 @@ begin
             err_any             => s_reg_status.dma_err_any,
             clear_errors        => s_reg_trigger.dma_clear,
             read_cmd_addr       => (others => (others => '0')),
+            read_cmd_length     => (others => (others => '0')),
             read_cmd_valid      => (others => '0'),
             read_cmd_ready      => open,
             read_data           => open,
             read_data_valid     => open,
             write_cmd_addr      => s_dma_write_cmd_addr,
+            write_cmd_length    => s_dma_write_cmd_length,
             write_cmd_valid     => s_dma_write_cmd_valid,
             write_cmd_ready     => s_dma_write_cmd_ready,
             write_data          => s_dma_write_data,
@@ -350,11 +353,14 @@ begin
     -- DMA Write Channel
     inst_write_channel: entity work.dma_write_channel
         generic map (
-            queue_size_bits     => 10 )
+            transfer_size_bits  => 4,
+            queue_size_bits     => 10,
+            idle_timeout        => 256 )
         port map (
             clk                 => clk_adc,
             reset               => s_reset,
             channel_en          => s_reg_control.acq_channel_en,
+            channel_busy        => s_reg_status.acq_channel_busy,
             channel_init        => s_reg_trigger.acq_channel_init,
             addr_start          => s_reg_control.acq_addr_start,
             addr_end            => s_reg_control.acq_addr_end,
@@ -368,6 +374,7 @@ begin
             in_ready            => s_dma_in_ready,
             in_data             => s_dma_in_data,
             write_cmd_addr      => s_dma_write_cmd_addr(0),
+            write_cmd_length    => s_dma_write_cmd_length(0),
             write_cmd_valid     => s_dma_write_cmd_valid(0),
             write_cmd_ready     => s_dma_write_cmd_ready(0),
             write_data          => s_dma_write_data(0),
@@ -384,10 +391,24 @@ begin
     process (clk_adc) is
     begin
         if rising_edge(clk_adc) then
+            if s_reg_control.test_divider = x"ffff" then
                 s_dma_in_valid <= '1';
                 if s_dma_in_ready = '1' then
-                s_dma_in_data(15 downto 0) <= std_logic_vector(unsigned(s_dma_in_data(15 downto 0)) + 1);
+                    s_dma_in_data(47 downto 0) <= std_logic_vector(unsigned(s_dma_in_data(47 downto 0)) + 1);
-                s_dma_in_data(63 downto 16) <= std_logic_vector(unsigned(s_dma_in_data(63 downto 16)) - 1);
+                    s_dma_in_data(63 downto 48) <= (others => '0');
+                end if;
+            else
+                if s_dma_in_ready = '1' then
+                    s_dma_in_valid <= '0';
+                end if;
+                if r_test_cnt >= unsigned(s_reg_control.test_divider) then
+                    r_test_cnt <= (others => '0');
+                    s_dma_in_valid <= '1';
+                    s_dma_in_data(47 downto 0) <= std_logic_vector(unsigned(s_dma_in_data(47 downto 0)) + 1);
+                    s_dma_in_data(63 downto 48) <= (others => '0');
+                else
+                    r_test_cnt <= r_test_cnt + 1;
+                end if;
             end if;
         end if;
     end process;

fpga/rtl/registers.vhd

@@ -101,11 +101,14 @@ begin
                 when reg_acq_addr_start =>  v.prdata(31 downto 7) := r.reg_control.acq_addr_start;
                 when reg_acq_addr_end =>    v.prdata(31 downto 7) := r.reg_control.acq_addr_end;
                 when reg_acq_addr_limit =>  v.prdata(31 downto 7) := r.reg_control.acq_addr_limit;
-                when reg_acq_addr_intr =>   v.prdata(31 downto 7) := r.reg_control.acq_addr_intr;
+                when reg_acq_addr_intr =>   v.prdata(31 downto 3) := r.reg_control.acq_addr_intr;
-                when reg_acq_addr_ptr =>    v.prdata(31 downto 7) := reg_status.acq_addr_ptr;
+                when reg_acq_addr_ptr =>    v.prdata(31 downto 3) := reg_status.acq_addr_ptr;
-                when reg_acq_channel_ctrl => v.prdata(0) := r.reg_control.acq_channel_en;
+                when reg_acq_channel_ctrl =>
+                    v.prdata(0) := r.reg_control.acq_channel_en;
+                    v.prdata(8) := reg_status.acq_channel_busy;
                 when reg_acq_intr_ctrl =>   v.prdata(0) := r.reg_control.acq_intr_en;
                 when reg_test_led =>        v.prdata(7 downto 0) := r.reg_control.test_led;
+                when reg_test_divider =>    v.prdata(15 downto 0) := r.reg_control.test_divider;
                 when reg_dma_buf_addr =>    v.prdata(31 downto 12) := r.reg_control.dma_buf_addr;
                 when reg_dma_buf_size =>    v.prdata(31 downto 12) := r.reg_control.dma_buf_size;
                 when others =>              null;
@@ -122,7 +125,7 @@ begin
                 when reg_acq_addr_start =>  v.reg_control.acq_addr_start := apb_pwdata(31 downto 7);
                 when reg_acq_addr_end =>    v.reg_control.acq_addr_end := apb_pwdata(31 downto 7);
                 when reg_acq_addr_limit =>  v.reg_control.acq_addr_limit := apb_pwdata(31 downto 7);
-                when reg_acq_addr_intr =>   v.reg_control.acq_addr_intr := apb_pwdata(31 downto 7);
+                when reg_acq_addr_intr =>   v.reg_control.acq_addr_intr := apb_pwdata(31 downto 3);
                 when reg_acq_channel_ctrl =>
                     v.reg_control.acq_channel_en := apb_pwdata(0);
                     v.reg_trigger.acq_channel_init := apb_pwdata(1);
@@ -130,6 +133,7 @@ begin
                     v.reg_control.acq_intr_en := apb_pwdata(0);
                     v.reg_trigger.acq_intr_clear := apb_pwdata(1);
                 when reg_test_led =>        v.reg_control.test_led := apb_pwdata(7 downto 0);
+                when reg_test_divider =>    v.reg_control.test_divider := apb_pwdata(15 downto 0);
                 when reg_dma_buf_addr =>    v.reg_control.dma_buf_addr := apb_pwdata(31 downto 12);
                 when reg_dma_buf_size =>    v.reg_control.dma_buf_size := apb_pwdata(31 downto 12);
                 when others =>              null;