Merge pull request 'fix: axi dma wrapper' (#2) from dev into master

Reviewed-on: #2

axi/axi_reg/axi4l_reg_map.sv

@@ -12,7 +12,8 @@ module axi4l_reg_map #(
   axi4l_if.slave s_axil,
 
   input  logic [N_REGS-1:0][31:0] reg_i,
-  output logic [N_REGS-1:0][31:0] reg_o
+  output logic [N_REGS-1:0][31:0] reg_o,
+  output logic [N_REGS-1:0][31:0] reg_pulse
 );
   import axi_pkg::*;
 
@@ -105,6 +106,7 @@ module axi4l_reg_map #(
       rdata_q    <= '0;
       reg_o      <= REG_RST;
     end else begin
+      reg_pulse <= '0;
       for (int r = 0; r < N_REGS; r++) begin
         for (int bit_idx = 0; bit_idx < 32; bit_idx++) begin
           if (reg_bit_mode_t'(REG_MODE[r][bit_idx]) == REG_BIT_W1S)
@@ -136,12 +138,33 @@ module axi4l_reg_map #(
           for (b = 0; b < 32; b = b + 1) begin
             if (wr_mask[b]) begin
               unique case (reg_bit_mode_t'(REG_MODE[wr_idx][b]))
-                REG_BIT_RSVD: begin end
-                REG_BIT_RO  : begin bresp_q <= 2'b10; end
-                REG_BIT_RW  : rw_new[b] = wr_data32[b];
-                REG_BIT_W1S : if (wr_data32[b]) rw_new[b] = 1'b1;
-                REG_BIT_W1C : if (wr_data32[b]) rw_new[b] = 1'b0;
-                default     : begin end
+                REG_BIT_RSVD: begin
+                end
+
+                REG_BIT_RO: begin
+                  bresp_q <= 2'b10;
+                end
+
+                REG_BIT_RW: begin
+                  rw_new[b] = wr_data32[b];
+                end
+
+                REG_BIT_W1S: begin
+                  if (wr_data32[b]) begin
+                    rw_new[b]             = 1'b1;
+                    reg_pulse[wr_idx][b] <= 1'b1;
+                  end
+                end
+
+                REG_BIT_W1C: begin
+                  if (wr_data32[b]) begin
+                    rw_new[b]             = 1'b0;
+                    reg_pulse[wr_idx][b] <= 1'b1;
+                  end
+                end
+
+                default: begin
+                end
               endcase
             end
           end
@@ -174,6 +197,7 @@ module axi4l_reg_map #(
               REG_BIT_W1C : rd_word[b] = reg_o[rd_idx][b];
               default     : rd_word[b] = 1'b0;
             endcase
+
           end
         end
 

axi/axi_reg/dma_axil_reg_map_pkg.sv

@@ -1,37 +1,64 @@
 package dma_axil_reg_map_pkg;
 
-    localparam int unsigned DMA_AXIL_REG_MAP_N_REGS = 4;
+    localparam int unsigned DMA_AXIL_REG_MAP_N_REGS = 6;
+
+    localparam int unsigned DMA_WRITE_DESC_CONTROL_REG = 0;
+    localparam int unsigned DMA_WRITE_DESC_ADDR_REG    = 1;
+    localparam int unsigned DMA_WRITE_DESC_LEN_REG     = 2;
+    localparam int unsigned DMA_READ_DESC_CONTROL_REG  = 3;
+    localparam int unsigned DMA_READ_DESC_ADDR_REG     = 4;
+    localparam int unsigned DMA_READ_DESC_LEN_REG      = 5;
 
     localparam logic [2:0] REG_BIT_RSVD = 3'd0;
     localparam logic [2:0] REG_BIT_RO   = 3'd1;
     localparam logic [2:0] REG_BIT_RW   = 3'd2;
     localparam logic [2:0] REG_BIT_W1S  = 3'd3;
-    localparam logic [2:0] REG_BIT_W1C  = 3'd4;
 
-    localparam DMA_WRITE_DESC_CONTROL_REG = 0;
-    localparam DMA_WRITE_DESC_ADDR_REG = 1;
-    localparam DMA_WRITE_DESC_LEN_REG = 2;
-    localparam DMA_READ_DESC_CONTROL_REG = 3;
-    localparam DMA_READ_DESC_ADDR_REG = 4;
-    localparam DMA_READ_DESC_LEN_REG = 5;
+    typedef logic [DMA_AXIL_REG_MAP_N_REGS-1:0][31:0][2:0] reg_mode_map_t;
 
-    localparam logic [DMA_AXIL_REG_MAP_N_REGS-1:0][31:0][2:0] DMA_AXIL_REG_MAP_REG_MODE = '{
+    function automatic reg_mode_map_t make_dma_reg_mode();
+        reg_mode_map_t mode;
 
-        '{REG_BIT_RO, REG_BIT_W1S, default: REG_BIT_RSVD},
-        '{32{REG_BIT_RW}, default: REG_BIT_RSVD},
-        '{32{REG_BIT_RW}, default: REG_BIT_RSVD},
-        '{REG_BIT_RO, REG_BIT_W1S, default: REG_BIT_RSVD},
-        '{32{REG_BIT_RW}, default: REG_BIT_RSVD},
-        '{32{REG_BIT_RW}, default: REG_BIT_RSVD}
-    };
+        mode = '0;
 
-    localparam logic [DMA_AXIL_REG_MAP_N_REGS-1:0][31:0] DMA_AXIL_REG_MAP_REG_RST = '{
-        32'h0000_0000,
-        32'h0000_0000,
-        32'h0000_0000,
-        32'h0000_0000,
-        32'h0000_0000,
-        32'h0000_0000
-    };
+        // По умолчанию всё reserved
+        for (int reg_idx = 0; reg_idx < DMA_AXIL_REG_MAP_N_REGS; reg_idx++) begin
+            for (int bit_idx = 0; bit_idx < 32; bit_idx++) begin
+                mode[reg_idx][bit_idx] = REG_BIT_RSVD;
+            end
+        end
+
+        // WRITE CONTROL
+        mode[DMA_WRITE_DESC_CONTROL_REG][0] = REG_BIT_RO;
+        mode[DMA_WRITE_DESC_CONTROL_REG][1] = REG_BIT_W1S;
+
+        // WRITE ADDR
+        for (int bit_idx = 0; bit_idx < 32; bit_idx++) begin
+            mode[DMA_WRITE_DESC_ADDR_REG][bit_idx] = REG_BIT_RW;
+        end
+
+        // WRITE LEN
+        for (int bit_idx = 0; bit_idx < 32; bit_idx++) begin
+            mode[DMA_WRITE_DESC_LEN_REG][bit_idx] = REG_BIT_RW;
+        end
+
+        // READ CONTROL
+        mode[DMA_READ_DESC_CONTROL_REG][0] = REG_BIT_RO;
+        mode[DMA_READ_DESC_CONTROL_REG][1] = REG_BIT_W1S;
+
+        // READ ADDR
+        for (int bit_idx = 0; bit_idx < 32; bit_idx++) begin
+            mode[DMA_READ_DESC_ADDR_REG][bit_idx] = REG_BIT_RW;
+        end
+
+        // READ LEN
+        for (int bit_idx = 0; bit_idx < 32; bit_idx++) begin
+            mode[DMA_READ_DESC_LEN_REG][bit_idx] = REG_BIT_RW;
+        end
+
+        return mode;
+    endfunction
+
+    localparam reg_mode_map_t DMA_AXIL_REG_MAP_REG_MODE = make_dma_reg_mode();
 
 endpackage

axi/rtl/axi_crossbar_wrapper.sv

@@ -2,7 +2,7 @@ module axi_crossbar_wrapper #(
     parameter int SLAVE_QTY = 3,
     parameter int MASTER_QTY = 3,
     parameter int ADDR_WIDTH = 32,
-    parameter int DATA_WIDTH = 32
+    parameter int DATA_WIDTH = 32,
     parameter int STRB_WIDTH = (DATA_WIDTH/8)
 )(
     input  wire clk,

axi/rtl/axi_dma_wrapper.sv

@@ -271,7 +271,7 @@ module axi_dma_wrapper #(
     // local read descriptor interface -> DMA flat input
     axis_if_to_flat #(
         .DATA_W (AXI_ADDR_WIDTH + LEN_WIDTH + TAG_WIDTH),
-        .KEEP_W ('0),
+        .KEEP_W (AXIS_KEEP_WIDTH),
         .ID_W   (AXIS_ID_WIDTH),
         .DEST_W (AXIS_DEST_WIDTH),
         .USER_W (AXIS_USER_WIDTH)
@@ -282,7 +282,7 @@ module axi_dma_wrapper #(
         .m_axis_tstrb (),
         .m_axis_tlast (),
         .m_axis_tid   (dma_s_axis_read_desc_id),
-        .m_axis_tdest (dma_s_axis_read_desc_tdest),
+        .m_axis_tdest (dma_s_axis_read_desc_dest),
         .m_axis_tuser (dma_s_axis_read_desc_user),
         .m_axis_tvalid (dma_s_axis_read_desc_valid),
         .m_axis_tready (dma_s_axis_read_desc_ready)
@@ -291,10 +291,10 @@ module axi_dma_wrapper #(
     // DMA read descriptor status flat output -> local status interface
     axis_flat_to_if #(
         .DATA_W (TAG_WIDTH + 4),
-        .KEEP_W ('0),
-        .ID_W ('0),
-        .DEST_W ('0),
-        .USER_W ('0)
+        .KEEP_W (AXIS_KEEP_WIDTH),
+        .ID_W   (AXIS_ID_WIDTH),
+        .DEST_W (AXIS_DEST_WIDTH),
+        .USER_W (AXIS_USER_WIDTH)
     ) i_read_desc_status_f2i(
         .s_axis_tdata ({dma_m_axis_read_desc_status_tag, dma_m_axis_read_desc_status_error}),
         .s_axis_tkeep (),
@@ -311,7 +311,7 @@ module axi_dma_wrapper #(
     // local write descriptor interface -> DMA flat input
     axis_if_to_flat #(
         .DATA_W (AXI_ADDR_WIDTH + LEN_WIDTH + TAG_WIDTH),
-        .KEEP_W ('0),
+        .KEEP_W (AXIS_KEEP_WIDTH),
         .ID_W   (AXIS_ID_WIDTH),
         .DEST_W (AXIS_DEST_WIDTH),
         .USER_W (AXIS_USER_WIDTH)
@@ -321,9 +321,9 @@ module axi_dma_wrapper #(
         .m_axis_tkeep (),
         .m_axis_tstrb (),
         .m_axis_tlast (),
-        .m_axis_tid   (dma_s_axis_write_desc_id),
-        .m_axis_tdest (dma_s_axis_write_desc_tdest),
-        .m_axis_tuser (dma_s_axis_write_desc_user),
+        .m_axis_tid   (),
+        .m_axis_tdest (),
+        .m_axis_tuser (),
         .m_axis_tvalid (dma_s_axis_write_desc_valid),
         .m_axis_tready (dma_s_axis_write_desc_ready)
     );
@@ -332,7 +332,7 @@ module axi_dma_wrapper #(
     // DMA write descriptor status flat output -> local status interface
     axis_flat_to_if #(
         .DATA_W (TAG_WIDTH + LEN_WIDTH + 4),
-        .KEEP_W ('0),
+        .KEEP_W (AXIS_KEEP_WIDTH),
         .ID_W (AXIS_ID_WIDTH),
         .DEST_W (AXIS_DEST_WIDTH),
         .USER_W (AXIS_USER_WIDTH)
@@ -448,6 +448,6 @@ module axi_dma_wrapper #(
     .m_axi          (m_axi)
   );
 
-endmodule : axi_dma_if_wrapper
+endmodule : axi_dma_wrapper
 
 `default_nettype wire

axi/rtl/axi_if.sv

@@ -13,7 +13,7 @@ interface axi4_if #(
   typedef logic [DATA_W/8-1:0] strb_t;
   typedef logic [ID_W-1:0]     id_t;
   typedef logic [USER_W-1:0]   user_t;
-  `AXI4_TYPEDEF_ALL(axi, addr_t, data_t, strb_t, id_t, user_t)
+  `AXI4_TYPEDEF_ALL(axi, addr_t, data_t, strb_t, id_t, user_t);
   axi_req_t  req;
   axi_resp_t resp;
   modport master (input aclk, aresetn, output req, input resp);
@@ -34,7 +34,7 @@ interface axi4l_if #(
   typedef logic [DATA_W-1:0]   data_t;
   typedef logic [DATA_W/8-1:0] strb_t;
   typedef logic [USER_W-1:0]   user_t;
-  `AXI4L_TYPEDEF_ALL(axil, addr_t, data_t, strb_t, user_t)
+  `AXI4L_TYPEDEF_ALL(axil, addr_t, data_t, strb_t, user_t);
   axil_req_t  req;
   axil_resp_t resp;
   modport master (input aclk, aresetn, output req, input resp);

axi/rtl/axil_cdc_wrapper.sv

@@ -1,6 +1,6 @@
 module axil_cdc_wrapper #(
-    parameter int ADDR_WIDTH = 32
-    parameter int DATA_WIDTH = 32,
+    parameter int ADDR_WIDTH = 32,
+    parameter int DATA_WIDTH = 32
 )(
     input wire s_clk,
     input wire s_rst,
@@ -83,8 +83,8 @@ module axil_cdc_wrapper #(
     );
 
     axil_cdc #(
-        .ADDR_WIDTH (ADDR_WIDTH)
-        .DATA_WIDTH (DATA_WIDTH),
+        .ADDR_WIDTH (ADDR_WIDTH),
+        .DATA_WIDTH (DATA_WIDTH)
     ) i_axil_cdc (
         .s_clk          (s_clk),
         .s_rst          (s_rst),

axi/tb/axi_dma_axis_compat/Makefile

@@ -44,8 +44,8 @@ VERILOG_SOURCES += $(FORENCICH_AXI_RTL_DIR)/axi_dma_rd.v
 VERILOG_SOURCES += $(FORENCICH_AXI_RTL_DIR)/axi_dma_wr.v
 
 
-VERILOG_SOURCES += $(AXI_IF_RTL_DIR)/axi_dma_if_wrapper.sv
-VERILOG_SOURCES += $(AXI_IF_RTL_DIR)/axi_dma_compat.sv
+VERILOG_SOURCES += $(AXI_IF_RTL_DIR)/axi_dma_wrapper.sv
+
 VERILOG_SOURCES += $(TB_DIR)/tb_axi_dma_axis_compat.sv
 
 COMPILE_ARGS += -I$(AXI_IF_RTL_DIR)

axi/tb/axi_dma_axis_compat/tb_axi_dma_axis_compat.sv

@@ -26,18 +26,14 @@ module tb_axi_dma_axis_compat #(
   parameter int unsigned ENABLE_SG          = 0,
   parameter int unsigned ENABLE_UNALIGNED   = 0,
 
-  parameter int unsigned READ_DESC_WIDTH    = AXI_ADDR_WIDTH + LEN_WIDTH + TAG_WIDTH + AXIS_ID_WIDTH + AXIS_DEST_WIDTH + AXIS_USER_WIDTH,
-  parameter int unsigned WRITE_DESC_WIDTH   = AXI_ADDR_WIDTH + LEN_WIDTH + TAG_WIDTH,
-  parameter int unsigned DESC_CMD_RAW_WIDTH = READ_DESC_WIDTH + WRITE_DESC_WIDTH,
-  parameter int unsigned DESC_CMD_DATA_WIDTH = ((DESC_CMD_RAW_WIDTH + 7) / 8) * 8,
-  parameter int unsigned DESC_CMD_KEEP_WIDTH = DESC_CMD_DATA_WIDTH / 8,
-
-  parameter int unsigned READ_STATUS_WIDTH  = TAG_WIDTH + 4,
-  parameter int unsigned WRITE_STATUS_WIDTH = LEN_WIDTH + TAG_WIDTH + AXIS_ID_WIDTH + AXIS_DEST_WIDTH + AXIS_USER_WIDTH + 4,
-  parameter int unsigned DESC_STATUS_RAW_WIDTH = READ_STATUS_WIDTH + WRITE_STATUS_WIDTH,
-  parameter int unsigned DESC_STATUS_DATA_WIDTH = ((DESC_STATUS_RAW_WIDTH + 7) / 8) * 8,
-  parameter int unsigned DESC_STATUS_KEEP_WIDTH = DESC_STATUS_DATA_WIDTH / 8,
-  parameter int unsigned DESC_STATUS_USER_WIDTH = 2
+  parameter int unsigned READ_DESC_RAW_WIDTH     = AXI_ADDR_WIDTH + LEN_WIDTH + TAG_WIDTH,
+  parameter int unsigned READ_DESC_DATA_WIDTH    = ((READ_DESC_RAW_WIDTH + 7) / 8) * 8,
+  parameter int unsigned WRITE_DESC_RAW_WIDTH    = AXI_ADDR_WIDTH + LEN_WIDTH + TAG_WIDTH,
+  parameter int unsigned WRITE_DESC_DATA_WIDTH   = ((WRITE_DESC_RAW_WIDTH + 7) / 8) * 8,
+  parameter int unsigned READ_STATUS_RAW_WIDTH   = TAG_WIDTH + 4,
+  parameter int unsigned READ_STATUS_DATA_WIDTH  = ((READ_STATUS_RAW_WIDTH + 7) / 8) * 8,
+  parameter int unsigned WRITE_STATUS_RAW_WIDTH  = LEN_WIDTH + TAG_WIDTH + 4,
+  parameter int unsigned WRITE_STATUS_DATA_WIDTH = ((WRITE_STATUS_RAW_WIDTH + 7) / 8) * 8
 )(
   input  logic clk,
   input  logic rst,
@@ -46,27 +42,35 @@ module tb_axi_dma_axis_compat #(
   input  logic write_enable,
   input  logic write_abort,
 
-  // Unified descriptor command AXIS, flat for cocotb
-  input  logic [DESC_CMD_DATA_WIDTH-1:0] s_axis_desc_cmd_tdata,
-  input  logic [DESC_CMD_KEEP_WIDTH-1:0] s_axis_desc_cmd_tkeep,
-  input  logic [DESC_CMD_KEEP_WIDTH-1:0] s_axis_desc_cmd_tstrb,
-  input  logic                           s_axis_desc_cmd_tlast,
-  input  logic                           s_axis_desc_cmd_tid,
-  input  logic                           s_axis_desc_cmd_tdest,
-  input  logic                           s_axis_desc_cmd_tuser,
-  input  logic                           s_axis_desc_cmd_tvalid,
-  output logic                           s_axis_desc_cmd_tready,
+  // Read descriptor AXIS input, flat for cocotb.
+  // tdata layout: {addr, len, tag}; id/dest/user use AXIS sidebands.
+  input  logic [READ_DESC_DATA_WIDTH-1:0] s_axis_read_desc_tdata,
+  input  logic [AXIS_ID_WIDTH-1:0]        s_axis_read_desc_tid,
+  input  logic [AXIS_DEST_WIDTH-1:0]      s_axis_read_desc_tdest,
+  input  logic [AXIS_USER_WIDTH-1:0]      s_axis_read_desc_tuser,
+  input  logic                            s_axis_read_desc_tvalid,
+  output logic                            s_axis_read_desc_tready,
 
-  // Unified descriptor status AXIS, flat for cocotb
-  output logic [DESC_STATUS_DATA_WIDTH-1:0] m_axis_desc_status_tdata,
-  output logic [DESC_STATUS_KEEP_WIDTH-1:0] m_axis_desc_status_tkeep,
-  output logic [DESC_STATUS_KEEP_WIDTH-1:0] m_axis_desc_status_tstrb,
-  output logic                              m_axis_desc_status_tlast,
-  output logic                              m_axis_desc_status_tid,
-  output logic                              m_axis_desc_status_tdest,
-  output logic [DESC_STATUS_USER_WIDTH-1:0] m_axis_desc_status_tuser,
-  output logic                              m_axis_desc_status_tvalid,
-  input  logic                              m_axis_desc_status_tready,
+  // Read descriptor status AXIS output.
+  // tdata layout: {tag, error}.
+  output logic [READ_STATUS_DATA_WIDTH-1:0] m_axis_read_desc_status_tdata,
+  output logic                              m_axis_read_desc_status_tvalid,
+  input  logic                              m_axis_read_desc_status_tready,
+
+  // Write descriptor AXIS input.
+  // tdata layout: {addr, len, tag}.
+  input  logic [WRITE_DESC_DATA_WIDTH-1:0] s_axis_write_desc_tdata,
+  input  logic                             s_axis_write_desc_tvalid,
+  output logic                             s_axis_write_desc_tready,
+
+  // Write descriptor status AXIS output.
+  // tdata layout: {len, tag, error}; id/dest/user use AXIS sidebands.
+  output logic [WRITE_STATUS_DATA_WIDTH-1:0] m_axis_write_desc_status_tdata,
+  output logic [AXIS_ID_WIDTH-1:0]           m_axis_write_desc_status_tid,
+  output logic [AXIS_DEST_WIDTH-1:0]         m_axis_write_desc_status_tdest,
+  output logic [AXIS_USER_WIDTH-1:0]         m_axis_write_desc_status_tuser,
+  output logic                               m_axis_write_desc_status_tvalid,
+  input  logic                               m_axis_write_desc_status_tready,
 
   // Read data AXIS output
   output logic [AXIS_DATA_WIDTH-1:0]    m_axis_read_data_tdata,
@@ -144,23 +148,45 @@ module tb_axi_dma_axis_compat #(
   wire rstn = ~rst;
 
   axis_if #(
-    .DATA_W (DESC_CMD_DATA_WIDTH),
-    .KEEP_W (DESC_CMD_KEEP_WIDTH),
-    .ID_W   (1),
-    .DEST_W (1),
-    .USER_W (1)
-  ) desc_cmd_if (
+    .DATA_W (READ_DESC_DATA_WIDTH),
+    .KEEP_W (READ_DESC_DATA_WIDTH / 8),
+    .ID_W   (AXIS_ID_WIDTH),
+    .DEST_W (AXIS_DEST_WIDTH),
+    .USER_W (AXIS_USER_WIDTH)
+  ) read_desc_if (
     .aclk    (clk),
     .aresetn (rstn)
   );
 
   axis_if #(
-    .DATA_W (DESC_STATUS_DATA_WIDTH),
-    .KEEP_W (DESC_STATUS_KEEP_WIDTH),
+    .DATA_W (READ_STATUS_DATA_WIDTH),
+    .KEEP_W (READ_STATUS_DATA_WIDTH / 8),
     .ID_W   (1),
     .DEST_W (1),
-    .USER_W (DESC_STATUS_USER_WIDTH)
-  ) desc_status_if (
+    .USER_W (1)
+  ) read_desc_status_if (
+    .aclk    (clk),
+    .aresetn (rstn)
+  );
+
+  axis_if #(
+    .DATA_W (WRITE_DESC_DATA_WIDTH),
+    .KEEP_W (WRITE_DESC_DATA_WIDTH / 8),
+    .ID_W   (1),
+    .DEST_W (1),
+    .USER_W (1)
+  ) write_desc_if (
+    .aclk    (clk),
+    .aresetn (rstn)
+  );
+
+  axis_if #(
+    .DATA_W (WRITE_STATUS_DATA_WIDTH),
+    .KEEP_W (WRITE_STATUS_DATA_WIDTH / 8),
+    .ID_W   (AXIS_ID_WIDTH),
+    .DEST_W (AXIS_DEST_WIDTH),
+    .USER_W (AXIS_USER_WIDTH)
+  ) write_desc_status_if (
     .aclk    (clk),
     .aresetn (rstn)
   );
@@ -198,41 +224,79 @@ module tb_axi_dma_axis_compat #(
   );
 
   axis_flat_to_if #(
-    .DATA_W (DESC_CMD_DATA_WIDTH),
-    .KEEP_W (DESC_CMD_KEEP_WIDTH),
-    .ID_W   (1),
-    .DEST_W (1),
-    .USER_W (1)
-  ) u_desc_cmd_flat_to_if (
-    .s_axis_tdata  (s_axis_desc_cmd_tdata),
-    .s_axis_tkeep  (s_axis_desc_cmd_tkeep),
-    .s_axis_tstrb  (s_axis_desc_cmd_tstrb),
-    .s_axis_tlast  (s_axis_desc_cmd_tlast),
-    .s_axis_tid    (s_axis_desc_cmd_tid),
-    .s_axis_tdest  (s_axis_desc_cmd_tdest),
-    .s_axis_tuser  (s_axis_desc_cmd_tuser),
-    .s_axis_tvalid (s_axis_desc_cmd_tvalid),
-    .s_axis_tready (s_axis_desc_cmd_tready),
-    .m_axis        (desc_cmd_if)
+    .DATA_W (READ_DESC_DATA_WIDTH),
+    .KEEP_W (READ_DESC_DATA_WIDTH / 8),
+    .ID_W   (AXIS_ID_WIDTH),
+    .DEST_W (AXIS_DEST_WIDTH),
+    .USER_W (AXIS_USER_WIDTH)
+  ) u_read_desc_flat_to_if (
+    .s_axis_tdata  (s_axis_read_desc_tdata),
+    .s_axis_tkeep  ({(READ_DESC_DATA_WIDTH/8){1'b1}}),
+    .s_axis_tstrb  ({(READ_DESC_DATA_WIDTH/8){1'b1}}),
+    .s_axis_tlast  (1'b1),
+    .s_axis_tid    (s_axis_read_desc_tid),
+    .s_axis_tdest  (s_axis_read_desc_tdest),
+    .s_axis_tuser  (s_axis_read_desc_tuser),
+    .s_axis_tvalid (s_axis_read_desc_tvalid),
+    .s_axis_tready (s_axis_read_desc_tready),
+    .m_axis        (read_desc_if)
   );
 
   axis_if_to_flat #(
-    .DATA_W (DESC_STATUS_DATA_WIDTH),
-    .KEEP_W (DESC_STATUS_KEEP_WIDTH),
+    .DATA_W (READ_STATUS_DATA_WIDTH),
+    .KEEP_W (READ_STATUS_DATA_WIDTH / 8),
     .ID_W   (1),
     .DEST_W (1),
-    .USER_W (DESC_STATUS_USER_WIDTH)
-  ) u_desc_status_if_to_flat (
-    .s_axis        (desc_status_if),
-    .m_axis_tdata  (m_axis_desc_status_tdata),
-    .m_axis_tkeep  (m_axis_desc_status_tkeep),
-    .m_axis_tstrb  (m_axis_desc_status_tstrb),
-    .m_axis_tlast  (m_axis_desc_status_tlast),
-    .m_axis_tid    (m_axis_desc_status_tid),
-    .m_axis_tdest  (m_axis_desc_status_tdest),
-    .m_axis_tuser  (m_axis_desc_status_tuser),
-    .m_axis_tvalid (m_axis_desc_status_tvalid),
-    .m_axis_tready (m_axis_desc_status_tready)
+    .USER_W (1)
+  ) u_read_desc_status_if_to_flat (
+    .s_axis        (read_desc_status_if),
+    .m_axis_tdata  (m_axis_read_desc_status_tdata),
+    .m_axis_tkeep  (),
+    .m_axis_tstrb  (),
+    .m_axis_tlast  (),
+    .m_axis_tid    (),
+    .m_axis_tdest  (),
+    .m_axis_tuser  (),
+    .m_axis_tvalid (m_axis_read_desc_status_tvalid),
+    .m_axis_tready (m_axis_read_desc_status_tready)
+  );
+
+  axis_flat_to_if #(
+    .DATA_W (WRITE_DESC_DATA_WIDTH),
+    .KEEP_W (WRITE_DESC_DATA_WIDTH / 8),
+    .ID_W   (1),
+    .DEST_W (1),
+    .USER_W (1)
+  ) u_write_desc_flat_to_if (
+    .s_axis_tdata  (s_axis_write_desc_tdata),
+    .s_axis_tkeep  ({(WRITE_DESC_DATA_WIDTH/8){1'b1}}),
+    .s_axis_tstrb  ({(WRITE_DESC_DATA_WIDTH/8){1'b1}}),
+    .s_axis_tlast  (1'b1),
+    .s_axis_tid    (1'b0),
+    .s_axis_tdest  (1'b0),
+    .s_axis_tuser  (1'b0),
+    .s_axis_tvalid (s_axis_write_desc_tvalid),
+    .s_axis_tready (s_axis_write_desc_tready),
+    .m_axis        (write_desc_if)
+  );
+
+  axis_if_to_flat #(
+    .DATA_W (WRITE_STATUS_DATA_WIDTH),
+    .KEEP_W (WRITE_STATUS_DATA_WIDTH / 8),
+    .ID_W   (AXIS_ID_WIDTH),
+    .DEST_W (AXIS_DEST_WIDTH),
+    .USER_W (AXIS_USER_WIDTH)
+  ) u_write_desc_status_if_to_flat (
+    .s_axis        (write_desc_status_if),
+    .m_axis_tdata  (m_axis_write_desc_status_tdata),
+    .m_axis_tkeep  (),
+    .m_axis_tstrb  (),
+    .m_axis_tlast  (),
+    .m_axis_tid    (m_axis_write_desc_status_tid),
+    .m_axis_tdest  (m_axis_write_desc_status_tdest),
+    .m_axis_tuser  (m_axis_write_desc_status_tuser),
+    .m_axis_tvalid (m_axis_write_desc_status_tvalid),
+    .m_axis_tready (m_axis_write_desc_status_tready)
   );
 
   axis_if_to_flat #(
@@ -273,7 +337,7 @@ module tb_axi_dma_axis_compat #(
     .m_axis        (write_data_if)
   );
 
-  axi_dma_axis_compat #(
+  axi_dma_wrapper #(
     .AXI_DATA_WIDTH    (AXI_DATA_WIDTH),
     .AXI_ADDR_WIDTH    (AXI_ADDR_WIDTH),
     .AXI_STRB_WIDTH    (AXI_STRB_WIDTH),
@@ -293,16 +357,15 @@ module tb_axi_dma_axis_compat #(
     .LEN_WIDTH         (LEN_WIDTH),
     .TAG_WIDTH         (TAG_WIDTH),
     .ENABLE_SG         (ENABLE_SG),
-    .ENABLE_UNALIGNED  (ENABLE_UNALIGNED),
-    .DESC_CMD_DATA_WIDTH    (DESC_CMD_DATA_WIDTH),
-    .DESC_STATUS_DATA_WIDTH (DESC_STATUS_DATA_WIDTH),
-    .DESC_STATUS_USER_WIDTH (DESC_STATUS_USER_WIDTH)
+    .ENABLE_UNALIGNED  (ENABLE_UNALIGNED)
   ) u_dut (
     .clk                      (clk),
     .rst                      (rst),
-    .s_axis_desc_cmd    (desc_cmd_if),
-    .m_axis_desc_status (desc_status_if),
+    .s_axis_read_desc         (read_desc_if),
+    .m_axis_read_desc_status  (read_desc_status_if),
     .m_axis_read_data         (read_data_if),
+    .s_axis_write_desc        (write_desc_if),
+    .m_axis_write_desc_status (write_desc_status_if),
     .s_axis_write_data        (write_data_if),
     .m_axi                    (m_axi_if),
     .read_enable              (read_enable),

axi/tb/axi_dma_axis_compat/test_axi_dma_axis_compat.py

@@ -3,7 +3,7 @@ import os
 
 import cocotb
 from cocotb.clock import Clock
-from cocotb.triggers import RisingEdge, Combine
+from cocotb.triggers import RisingEdge
 
 from cocotbext.axi import AxiBus, AxiRam
 from cocotbext.axi import AxiStreamBus, AxiStreamFrame, AxiStreamSource, AxiStreamSink
@@ -16,42 +16,33 @@ AXIS_ID_WIDTH = int(os.getenv("PARAM_AXIS_ID_WIDTH", "8"))
 AXIS_DEST_WIDTH = int(os.getenv("PARAM_AXIS_DEST_WIDTH", "8"))
 AXIS_USER_WIDTH = int(os.getenv("PARAM_AXIS_USER_WIDTH", "1"))
 
-READ_DESC_WIDTH = AXI_ADDR_WIDTH + LEN_WIDTH + TAG_WIDTH + \
-    AXIS_ID_WIDTH + AXIS_DEST_WIDTH + AXIS_USER_WIDTH
-WRITE_DESC_WIDTH = AXI_ADDR_WIDTH + LEN_WIDTH + TAG_WIDTH
-DESC_CMD_RAW_WIDTH = READ_DESC_WIDTH + WRITE_DESC_WIDTH
-DESC_CMD_DATA_WIDTH = ((DESC_CMD_RAW_WIDTH + 7) // 8) * 8
-DESC_CMD_BYTES = DESC_CMD_DATA_WIDTH // 8
+READ_DESC_RAW_WIDTH = AXI_ADDR_WIDTH + LEN_WIDTH + TAG_WIDTH
+READ_DESC_DATA_WIDTH = ((READ_DESC_RAW_WIDTH + 7) // 8) * 8
+READ_DESC_BYTES = READ_DESC_DATA_WIDTH // 8
 
-READ_STATUS_WIDTH = TAG_WIDTH + 4
-WRITE_STATUS_WIDTH = LEN_WIDTH + TAG_WIDTH + \
-    AXIS_ID_WIDTH + AXIS_DEST_WIDTH + AXIS_USER_WIDTH + 4
-DESC_STATUS_RAW_WIDTH = READ_STATUS_WIDTH + WRITE_STATUS_WIDTH
-DESC_STATUS_DATA_WIDTH = ((DESC_STATUS_RAW_WIDTH + 7) // 8) * 8
-DESC_STATUS_BYTES = DESC_STATUS_DATA_WIDTH // 8
+WRITE_DESC_RAW_WIDTH = AXI_ADDR_WIDTH + LEN_WIDTH + TAG_WIDTH
+WRITE_DESC_DATA_WIDTH = ((WRITE_DESC_RAW_WIDTH + 7) // 8) * 8
+WRITE_DESC_BYTES = WRITE_DESC_DATA_WIDTH // 8
 
-RD_ADDR_LSB = 0
-RD_LEN_LSB = RD_ADDR_LSB + AXI_ADDR_WIDTH
-RD_TAG_LSB = RD_LEN_LSB + LEN_WIDTH
-RD_ID_LSB = RD_TAG_LSB + TAG_WIDTH
-RD_DEST_LSB = RD_ID_LSB + AXIS_ID_WIDTH
-RD_USER_LSB = RD_DEST_LSB + AXIS_DEST_WIDTH
+READ_STATUS_RAW_WIDTH = TAG_WIDTH + 4
+READ_STATUS_DATA_WIDTH = ((READ_STATUS_RAW_WIDTH + 7) // 8) * 8
 
-WR_BASE_LSB = READ_DESC_WIDTH
-WR_ADDR_LSB = WR_BASE_LSB
-WR_LEN_LSB = WR_ADDR_LSB + AXI_ADDR_WIDTH
-WR_TAG_LSB = WR_LEN_LSB + LEN_WIDTH
+WRITE_STATUS_RAW_WIDTH = LEN_WIDTH + TAG_WIDTH + 4
+WRITE_STATUS_DATA_WIDTH = ((WRITE_STATUS_RAW_WIDTH + 7) // 8) * 8
 
-RD_STS_TAG_LSB = 0
-RD_STS_ERROR_LSB = RD_STS_TAG_LSB + TAG_WIDTH
+# axi_dma_wrapper maps descriptor tdata as {addr, len, tag}.
+DESC_TAG_LSB = 0
+DESC_LEN_LSB = DESC_TAG_LSB + TAG_WIDTH
+DESC_ADDR_LSB = DESC_LEN_LSB + LEN_WIDTH
 
-WR_STS_BASE_LSB = READ_STATUS_WIDTH
-WR_STS_LEN_LSB = WR_STS_BASE_LSB
-WR_STS_TAG_LSB = WR_STS_LEN_LSB + LEN_WIDTH
-WR_STS_ID_LSB = WR_STS_TAG_LSB + TAG_WIDTH
-WR_STS_DEST_LSB = WR_STS_ID_LSB + AXIS_ID_WIDTH
-WR_STS_USER_LSB = WR_STS_DEST_LSB + AXIS_DEST_WIDTH
-WR_STS_ERROR_LSB = WR_STS_USER_LSB + AXIS_USER_WIDTH
+# Read status tdata is {tag, error}.
+READ_STATUS_ERROR_LSB = 0
+READ_STATUS_TAG_LSB = READ_STATUS_ERROR_LSB + 4
+
+# Write status tdata is {len, tag, error}.
+WRITE_STATUS_ERROR_LSB = 0
+WRITE_STATUS_TAG_LSB = WRITE_STATUS_ERROR_LSB + 4
+WRITE_STATUS_LEN_LSB = WRITE_STATUS_TAG_LSB + TAG_WIDTH
 
 
 def mask(width):
@@ -68,59 +59,60 @@ def get(value, lsb, width):
     return (int(value) >> lsb) & mask(width)
 
 
-def pack_desc_cmd(
-    *,
-    read_addr=0,
-    read_len=0,
-    read_tag=0,
-    read_id=0,
-    read_dest=0,
-    read_user=0,
-    write_addr=0,
-    write_len=0,
-    write_tag=0,
-):
+def pack_read_desc(*, addr, length, tag, axis_id=0, dest=0, user=0):
     value = 0
-    value = put(value, RD_ADDR_LSB, AXI_ADDR_WIDTH, read_addr)
-    value = put(value, RD_LEN_LSB, LEN_WIDTH, read_len)
-    value = put(value, RD_TAG_LSB, TAG_WIDTH, read_tag)
-    value = put(value, RD_ID_LSB, AXIS_ID_WIDTH, read_id)
-    value = put(value, RD_DEST_LSB, AXIS_DEST_WIDTH, read_dest)
-    value = put(value, RD_USER_LSB, AXIS_USER_WIDTH, read_user)
+    value = put(value, DESC_TAG_LSB, TAG_WIDTH, tag)
+    value = put(value, DESC_LEN_LSB, LEN_WIDTH, length)
+    value = put(value, DESC_ADDR_LSB, AXI_ADDR_WIDTH, addr)
 
-    value = put(value, WR_ADDR_LSB, AXI_ADDR_WIDTH, write_addr)
-    value = put(value, WR_LEN_LSB, LEN_WIDTH, write_len)
-    value = put(value, WR_TAG_LSB, TAG_WIDTH, write_tag)
+    return AxiStreamFrame(
+        value.to_bytes(READ_DESC_BYTES, byteorder="little"),
+        tid=axis_id,
+        tdest=dest,
+        tuser=user,
+    )
 
-    return value.to_bytes(DESC_CMD_BYTES, byteorder="little")
+
+def pack_write_desc(*, addr, length, tag):
+    value = 0
+    value = put(value, DESC_TAG_LSB, TAG_WIDTH, tag)
+    value = put(value, DESC_LEN_LSB, LEN_WIDTH, length)
+    value = put(value, DESC_ADDR_LSB, AXI_ADDR_WIDTH, addr)
+
+    return AxiStreamFrame(
+        value.to_bytes(WRITE_DESC_BYTES, byteorder="little")
+    )
 
 
 def frame_to_int(frame):
     return int.from_bytes(bytes(frame.tdata), byteorder="little")
 
 
-def frame_tuser(frame):
+def frame_sideband(frame, name):
+    value = getattr(frame, name)
     try:
-        return int(frame.tuser)
+        return int(value)
     except TypeError:
-        return int(frame.tuser[0]) if frame.tuser else 0
+        return int(value[0]) if value else 0
 
 
-def unpack_status(frame):
+def unpack_read_status(frame):
     value = frame_to_int(frame)
-    tuser = frame_tuser(frame)
-
     return {
-        "read_valid": bool(tuser & 0x1),
-        "write_valid": bool(tuser & 0x2),
-        "read_tag": get(value, RD_STS_TAG_LSB, TAG_WIDTH),
-        "read_error": get(value, RD_STS_ERROR_LSB, 4),
-        "write_len": get(value, WR_STS_LEN_LSB, LEN_WIDTH),
-        "write_tag": get(value, WR_STS_TAG_LSB, TAG_WIDTH),
-        "write_id": get(value, WR_STS_ID_LSB, AXIS_ID_WIDTH),
-        "write_dest": get(value, WR_STS_DEST_LSB, AXIS_DEST_WIDTH),
-        "write_user": get(value, WR_STS_USER_LSB, AXIS_USER_WIDTH),
-        "write_error": get(value, WR_STS_ERROR_LSB, 4),
+        "tag": get(value, READ_STATUS_TAG_LSB, TAG_WIDTH),
+        "error": get(value, READ_STATUS_ERROR_LSB, 4),
+    }
+
+
+def unpack_write_status(frame):
+    value = frame_to_int(frame)
+    return {
+        "len": get(value, WRITE_STATUS_LEN_LSB, LEN_WIDTH),
+        "tag": get(value, WRITE_STATUS_TAG_LSB, TAG_WIDTH),
+        "error": get(value, WRITE_STATUS_ERROR_LSB, 4),
+        "id": frame_sideband(frame, "tid"),
+        "dest": frame_sideband(frame, "tdest"),
+        "user": frame_sideband(frame, "tuser"),
     }
 
 
@@ -132,24 +124,45 @@ class TB:
 
         cocotb.start_soon(Clock(dut.clk, 10, units="ns").start())
 
-        self.desc_cmd_source = AxiStreamSource(
-            AxiStreamBus.from_prefix(dut, "s_axis_desc_cmd"), dut.clk, dut.rst
+        self.read_desc_source = AxiStreamSource(
+            AxiStreamBus.from_prefix(dut, "s_axis_read_desc"),
+            dut.clk,
+            dut.rst,
         )
-        self.desc_status_sink = AxiStreamSink(
-            AxiStreamBus.from_prefix(
-                dut, "m_axis_desc_status"), dut.clk, dut.rst
+        self.read_desc_status_sink = AxiStreamSink(
+            AxiStreamBus.from_prefix(dut, "m_axis_read_desc_status"),
+            dut.clk,
+            dut.rst,
+        )
+
+        self.write_desc_source = AxiStreamSource(
+            AxiStreamBus.from_prefix(dut, "s_axis_write_desc"),
+            dut.clk,
+            dut.rst,
+        )
+        self.write_desc_status_sink = AxiStreamSink(
+            AxiStreamBus.from_prefix(dut, "m_axis_write_desc_status"),
+            dut.clk,
+            dut.rst,
         )
 
         self.read_data_sink = AxiStreamSink(
-            AxiStreamBus.from_prefix(dut, "m_axis_read_data"), dut.clk, dut.rst
+            AxiStreamBus.from_prefix(dut, "m_axis_read_data"),
+            dut.clk,
+            dut.rst,
         )
         self.write_data_source = AxiStreamSource(
-            AxiStreamBus.from_prefix(
-                dut, "s_axis_write_data"), dut.clk, dut.rst
+            AxiStreamBus.from_prefix(dut, "s_axis_write_data"),
+            dut.clk,
+            dut.rst,
         )
 
-        self.axi_ram = AxiRam(AxiBus.from_prefix(
-            dut, "m_axi"), dut.clk, dut.rst, size=2**16)
+        self.axi_ram = AxiRam(
+            AxiBus.from_prefix(dut, "m_axi"),
+            dut.clk,
+            dut.rst,
+            size=2**AXI_ADDR_WIDTH,
+        )
 
         dut.read_enable.setimmediatevalue(0)
         dut.write_enable.setimmediatevalue(0)
@@ -166,8 +179,11 @@ class TB:
         await RisingEdge(self.dut.clk)
         await RisingEdge(self.dut.clk)
 
-    async def send_desc_cmd(self, payload):
-        await self.desc_cmd_source.send(AxiStreamFrame(payload))
+    async def send_read_desc(self, **kwargs):
+        await self.read_desc_source.send(pack_read_desc(**kwargs))
+
+    async def send_write_desc(self, **kwargs):
+        await self.write_desc_source.send(pack_write_desc(**kwargs))
 
 
 @cocotb.test()
@@ -184,28 +200,26 @@ async def test_axis_compat_write_path(dut):
 
     tb.axi_ram.write(addr - 16, b"\xaa" * (len(data) + 32))
 
-    cmd = pack_desc_cmd(write_addr=addr, write_len=len(data), write_tag=tag)
-
-    cmd_task = cocotb.start_soon(tb.send_desc_cmd(cmd))
+    desc_task = cocotb.start_soon(
+        tb.send_write_desc(addr=addr, length=len(data), tag=tag)
+    )
     data_task = cocotb.start_soon(
         tb.write_data_source.send(AxiStreamFrame(data, tid=tag))
     )
 
-    await cmd_task
+    await desc_task
     await data_task
 
-    status_frame = await tb.desc_status_sink.recv()
-    status = unpack_status(status_frame)
+    status = unpack_write_status(await tb.write_desc_status_sink.recv())
 
-    assert not status["read_valid"]
-    assert status["write_valid"]
-    assert status["write_tag"] == tag
-    assert status["write_len"] == len(data)
-    assert status["write_id"] == tag
-    assert status["write_error"] == 0
+    assert status["tag"] == tag
+    assert status["len"] == len(data)
+    assert status["id"] == tag
+    assert status["error"] == 0
 
-    assert tb.axi_ram.read(addr - 8, len(data) +
-                           16) == b"\xaa" * 8 + data + b"\xaa" * 8
+    assert tb.axi_ram.read(addr - 8, len(data) + 16) == (
+        b"\xaa" * 8 + data + b"\xaa" * 8
+    )
 
 
 @cocotb.test()
@@ -222,32 +236,26 @@ async def test_axis_compat_read_path(dut):
 
     tb.axi_ram.write(addr, data)
 
-    cmd = pack_desc_cmd(
-        read_addr=addr,
-        read_len=len(data),
-        read_tag=tag,
-        read_id=tag,
-        read_dest=0,
-        read_user=0,
+    await tb.send_read_desc(
+        addr=addr,
+        length=len(data),
+        tag=tag,
+        axis_id=tag,
+        dest=0,
+        user=0,
     )
 
-    await tb.send_desc_cmd(cmd)
-
     data_frame = await tb.read_data_sink.recv()
     assert bytes(data_frame.tdata) == data
-    assert int(data_frame.tid) == tag
+    assert frame_sideband(data_frame, "tid") == tag
 
-    status_frame = await tb.desc_status_sink.recv()
-    status = unpack_status(status_frame)
-
-    assert status["read_valid"]
-    assert not status["write_valid"]
-    assert status["read_tag"] == tag
-    assert status["read_error"] == 0
+    status = unpack_read_status(await tb.read_desc_status_sink.recv())
+    assert status["tag"] == tag
+    assert status["error"] == 0
 
 
 @cocotb.test()
-async def test_axis_compat_read_and_write_same_command(dut):
+async def test_axis_compat_read_and_write_independent_desc_ports(dut):
     tb = TB(dut)
     await tb.reset()
 
@@ -264,51 +272,46 @@ async def test_axis_compat_read_and_write_same_command(dut):
     tb.axi_ram.write(read_addr, read_data)
     tb.axi_ram.write(write_addr - 8, b"\xaa" * (len(write_data) + 16))
 
-    cmd = pack_desc_cmd(
-        read_addr=read_addr,
-        read_len=len(read_data),
-        read_tag=read_tag,
-        read_id=read_tag,
-        write_addr=write_addr,
-        write_len=len(write_data),
-        write_tag=write_tag,
+    read_desc_task = cocotb.start_soon(
+        tb.send_read_desc(
+            addr=read_addr,
+            length=len(read_data),
+            tag=read_tag,
+            axis_id=read_tag,
+            dest=0,
+            user=0,
         )
-
-    cmd_task = cocotb.start_soon(tb.send_desc_cmd(cmd))
-    data_task = cocotb.start_soon(
+    )
+    write_desc_task = cocotb.start_soon(
+        tb.send_write_desc(
+            addr=write_addr,
+            length=len(write_data),
+            tag=write_tag,
+        )
+    )
+    write_data_task = cocotb.start_soon(
         tb.write_data_source.send(AxiStreamFrame(write_data, tid=write_tag))
     )
 
-    await cmd_task
-    await data_task
+    await read_desc_task
+    await write_desc_task
+    await write_data_task
 
     data_frame = await tb.read_data_sink.recv()
     assert bytes(data_frame.tdata) == read_data
-    assert int(data_frame.tid) == read_tag
+    assert frame_sideband(data_frame, "tid") == read_tag
 
-    # status might be in 2 beats, depending on timings
-    seen_read = False
-    seen_write = False
+    read_status = unpack_read_status(await tb.read_desc_status_sink.recv())
+    write_status = unpack_write_status(await tb.write_desc_status_sink.recv())
 
-    for _ in range(2):
-        status_frame = await tb.desc_status_sink.recv()
-        status = unpack_status(status_frame)
+    assert read_status["tag"] == read_tag
+    assert read_status["error"] == 0
 
-        if status["read_valid"]:
-            assert status["read_tag"] == read_tag
-            assert status["read_error"] == 0
-            seen_read = True
+    assert write_status["tag"] == write_tag
+    assert write_status["len"] == len(write_data)
+    assert write_status["id"] == write_tag
+    assert write_status["error"] == 0
 
-        if status["write_valid"]:
-            assert status["write_tag"] == write_tag
-            assert status["write_len"] == len(write_data)
-            assert status["write_error"] == 0
-            seen_write = True
-
-        if seen_read and seen_write:
-            break
-
-    assert seen_read
-    assert seen_write
-    assert tb.axi_ram.read(write_addr - 8, len(write_data) +
-                           16) == b"\xaa" * 8 + write_data + b"\xaa" * 8
+    assert tb.axi_ram.read(write_addr - 8, len(write_data) + 16) == (
+        b"\xaa" * 8 + write_data + b"\xaa" * 8
+    )

axi/tb/axil_reg_cocotb/tb_axi4l_reg_map.sv

@@ -92,6 +92,7 @@ module tb_axi4l_reg_map;
 
   logic [N_REGS-1:0][31:0] reg_i;
   logic [N_REGS-1:0][31:0] reg_o;
+  logic [N_REGS-1:0][31:0] reg_pulse;
 
   // debug probes visible from cocotb.  they are useful for checking W1S pulse-like behavior
   logic [31:0] reg0_o;
@@ -157,7 +158,8 @@ module tb_axi4l_reg_map;
     .rst_n  (rst_n),
     .s_axil (s_axil_if.slave),
     .reg_i  (reg_i),
-    .reg_o  (reg_o)
+    .reg_o  (reg_o),
+    .reg_pulse(reg_pulse)
   );
 
 endmodule

axi/tb/axil_reg_cocotb/test_axi4l_reg_map.py

@@ -1,8 +1,6 @@
-# SPDX-License-Identifier: MIT
-
 import cocotb
 from cocotb.clock import Clock
-from cocotb.triggers import RisingEdge
+from cocotb.triggers import ReadOnly, RisingEdge
 from cocotbext.axi import AxiLiteBus, AxiLiteMaster
 
 OKAY = 0
@@ -35,7 +33,8 @@ def _read_data(resp):
 class TB:
     def __init__(self, dut):
         self.dut = dut
-        self.axil = AxiLiteMaster(AxiLiteBus.from_prefix(dut, "s_axil"), dut.clk, dut.rst)
+        self.axil = AxiLiteMaster(AxiLiteBus.from_prefix(
+            dut, "s_axil"), dut.clk, dut.rst)
 
     async def reset(self):
         self.dut.rst.value = 1
@@ -114,28 +113,175 @@ async def test_rw_and_w1c_bits(dut):
     assert await tb.read32(REG2_CONFIG) == 0xDEADBEEF
 
 
+def _reg0_pulse_value(dut):
+    """Return REG0 pulse bits regardless of how the test wrapper exposes them."""
+    if hasattr(dut, "reg0_pulse"):
+        return int(dut.reg0_pulse.value)
+
+    if hasattr(dut, "reg_pulse"):
+        # Packed SystemVerilog array [N_REGS-1:0][31:0]: REG0 occupies bits 31:0.
+        return int(dut.reg_pulse.value) & 0xFFFF_FFFF
+
+    raise AssertionError(
+        "DUT must expose either reg0_pulse[31:0] or packed reg_pulse"
+    )
+
+
+async def _sample_reg0_pulse_cycles(dut, cycles):
+    """Sample REG0 pulse output after each active clock edge."""
+    samples = []
+
+    for _ in range(cycles):
+        await RisingEdge(dut.clk)
+        await ReadOnly()
+        samples.append(_reg0_pulse_value(dut))
+
+    return samples
+
+
+def _assert_single_cycle_bit_pulse(samples, bit, name):
+    """Check that one selected bit was high for exactly one sampled cycle."""
+    high_cycles = [
+        index for index, value in enumerate(samples)
+        if value & (1 << bit)
+    ]
+    assert len(high_cycles) == 1, (
+        f"{name}: expected one high cycle, got {len(high_cycles)}; "
+        f"samples={[f'0x{value:08x}' for value in samples]}"
+    )
+    return high_cycles[0]
+
+
 @cocotb.test()
-async def test_w1s_seen_on_reg_o_probe(dut):
-    """Check that a W1S write creates a short-lived internal reg_o bit."""
+async def test_reg_pulse_is_zero_after_reset(dut):
+    """No W1 event may be reported after reset without a write."""
     cocotb.start_soon(Clock(dut.clk, 10, units="ns").start())
     tb = TB(dut)
     await tb.reset()
 
-    seen = False
+    samples = await _sample_reg0_pulse_cycles(dut, 4)
+    assert not any(samples), (
+        "reg_pulse was asserted without a W1 write: "
+        f"{[f'0x{value:08x}' for value in samples]}"
+    )
 
-    async def monitor_w1s_bit():
-        nonlocal seen
-        for _ in range(20):
-            await RisingEdge(dut.clk)
-            if int(dut.reg0_o.value) & 0x1:
-                seen = True
 
-    mon = cocotb.start_soon(monitor_w1s_bit())
+@cocotb.test()
+async def test_w1s_generates_single_cycle_reg_pulse(dut):
+    """Writing 1 to REG0[0] W1S must pulse REG0 pulse bit 0 once."""
+    cocotb.start_soon(Clock(dut.clk, 10, units="ns").start())
+    tb = TB(dut)
+    await tb.reset()
+
+    monitor = cocotb.start_soon(_sample_reg0_pulse_cycles(dut, 20))
     await tb.write32(REG0_CTRL, 0x00000001)
-    await mon
+    samples = await monitor
 
-    assert seen, "W1S bit was never observed on reg0_o[0]"
-    assert await tb.read32(REG0_CTRL) == 0x00000004  # W1S reads as 0, W1C reset bit remains set
+    _assert_single_cycle_bit_pulse(samples, bit=0, name="reg_pulse[0][0]")
+
+    # No unrelated W1 bit may pulse.
+    assert not any(value & (1 << 2) for value in samples)
+
+    # W1S is a command bit and is not returned by reads.
+    assert await tb.read32(REG0_CTRL) == 0x00000004
+
+
+@cocotb.test()
+async def test_w1c_generates_single_cycle_reg_pulse(dut):
+    """Writing 1 to REG0[2] W1C must pulse bit 2 once and clear state."""
+    cocotb.start_soon(Clock(dut.clk, 10, units="ns").start())
+    tb = TB(dut)
+    await tb.reset()
+
+    assert await tb.read32(REG0_CTRL) == 0x00000004
+
+    monitor = cocotb.start_soon(_sample_reg0_pulse_cycles(dut, 20))
+    await tb.write32(REG0_CTRL, 0x00000004)
+    samples = await monitor
+
+    _assert_single_cycle_bit_pulse(samples, bit=2, name="reg_pulse[0][2]")
+
+    # No unrelated W1 bit may pulse.
+    assert not any(value & (1 << 0) for value in samples)
+
+    assert await tb.read32(REG0_CTRL) == 0x00000000
+
+
+@cocotb.test()
+async def test_zero_to_w1_and_rw_write_do_not_generate_reg_pulse(dut):
+    """Only written ones in W1 fields may create reg_pulse."""
+    cocotb.start_soon(Clock(dut.clk, 10, units="ns").start())
+    tb = TB(dut)
+    await tb.reset()
+
+    monitor = cocotb.start_soon(_sample_reg0_pulse_cycles(dut, 20))
+
+    # bit1 is RW; both W1 fields receive zero.
+    await tb.write32(REG0_CTRL, 0x00000002)
+    samples = await monitor
+
+    assert not any(samples), (
+        "RW write or zero written to W1 fields generated reg_pulse: "
+        f"{[f'0x{value:08x}' for value in samples]}"
+    )
+    assert await tb.read32(REG0_CTRL) == 0x00000006
+
+
+@cocotb.test()
+async def test_w1s_and_w1c_pulse_together(dut):
+    """W1S and W1C ones in one AXI write must pulse on the same cycle."""
+    cocotb.start_soon(Clock(dut.clk, 10, units="ns").start())
+    tb = TB(dut)
+    await tb.reset()
+
+    monitor = cocotb.start_soon(_sample_reg0_pulse_cycles(dut, 20))
+    await tb.write32(REG0_CTRL, 0x00000005)
+    samples = await monitor
+
+    w1s_cycle = _assert_single_cycle_bit_pulse(
+        samples, bit=0, name="reg_pulse[0][0]"
+    )
+    w1c_cycle = _assert_single_cycle_bit_pulse(
+        samples, bit=2, name="reg_pulse[0][2]"
+    )
+
+    assert w1s_cycle == w1c_cycle, (
+        "W1 bits written by one transaction pulsed on different cycles: "
+        f"W1S={w1s_cycle}, W1C={w1c_cycle}"
+    )
+
+    # Only bits 0 and 2 are allowed to pulse.
+    assert samples[w1s_cycle] == 0x00000005
+    assert await tb.read32(REG0_CTRL) == 0x00000000
+
+
+@cocotb.test()
+async def test_each_w1_write_creates_a_new_pulse(dut):
+    """Two separate W1 writes must create two separate one-cycle pulses."""
+    cocotb.start_soon(Clock(dut.clk, 10, units="ns").start())
+    tb = TB(dut)
+    await tb.reset()
+
+    monitor = cocotb.start_soon(_sample_reg0_pulse_cycles(dut, 40))
+
+    await tb.write32(REG0_CTRL, 0x00000001)
+    for _ in range(3):
+        await RisingEdge(dut.clk)
+    await tb.write32(REG0_CTRL, 0x00000001)
+
+    samples = await monitor
+    high_cycles = [
+        index for index, value in enumerate(samples)
+        if value & 0x1
+    ]
+
+    assert len(high_cycles) == 2, (
+        f"expected two W1S pulses, got cycles {high_cycles}; "
+        f"samples={[f'0x{value:08x}' for value in samples]}"
+    )
+    assert high_cycles[1] > high_cycles[0] + 1, (
+        f"separate writes did not produce separate pulses: {high_cycles}"
+    )
 
 
 @cocotb.test()