cndm: Peel off queue management logic, store queue state in distributed RAM

Signed-off-by: Alex Forencich <alex@alexforencich.com>
This commit is contained in:
Alex Forencich
2026-03-05 18:04:43 -08:00
parent 8f1c082174
commit 7dbe6df56a
7 changed files with 407 additions and 200 deletions

View File

@@ -4,6 +4,7 @@ cndm_micro_dp_mgr.sv
cndm_micro_port.sv cndm_micro_port.sv
cndm_micro_rx.sv cndm_micro_rx.sv
cndm_micro_tx.sv cndm_micro_tx.sv
cndm_micro_queue_state.sv
cndm_micro_desc_rd.sv cndm_micro_desc_rd.sv
cndm_micro_cpl_wr.sv cndm_micro_cpl_wr.sv
../lib/taxi/src/prim/rtl/taxi_ram_2rw_1c.sv ../lib/taxi/src/prim/rtl/taxi_ram_2rw_1c.sv

View File

@@ -36,6 +36,9 @@ module cndm_micro_core #(
// Structural configuration // Structural configuration
parameter PORTS = 2, parameter PORTS = 2,
// Queue configuration
parameter CQN_W = 5,
// PTP configuration // PTP configuration
parameter logic PTP_TS_EN = 1'b1, parameter logic PTP_TS_EN = 1'b1,
parameter logic PTP_TS_FMT_TOD = 1'b0, parameter logic PTP_TS_FMT_TOD = 1'b0,
@@ -308,8 +311,12 @@ apb_dp_ctrl();
cndm_micro_dp_mgr #( cndm_micro_dp_mgr #(
.PORTS(PORTS), .PORTS(PORTS),
.CQN_W(CQN_W),
.PTP_EN(PTP_TS_EN), .PTP_EN(PTP_TS_EN),
.PTP_BASE_ADDR_DP(0), .PTP_BASE_ADDR_DP(0),
.PORT_BASE_ADDR_DP(PORT_BASE_ADDR_DP), .PORT_BASE_ADDR_DP(PORT_BASE_ADDR_DP),
.PORT_BASE_ADDR_HOST(PORT_BASE_ADDR_HOST) .PORT_BASE_ADDR_HOST(PORT_BASE_ADDR_HOST)
) )
@@ -490,6 +497,10 @@ dma_mux_inst (
for (genvar p = 0; p < PORTS; p = p + 1) begin : port for (genvar p = 0; p < PORTS; p = p + 1) begin : port
cndm_micro_port #( cndm_micro_port #(
// Queue configuration
.CQN_W(CQN_W),
// PTP configuration
.PTP_TS_EN(PTP_TS_EN), .PTP_TS_EN(PTP_TS_EN),
.PTP_TS_FMT_TOD(PTP_TS_FMT_TOD) .PTP_TS_FMT_TOD(PTP_TS_FMT_TOD)
) )

View File

@@ -1,7 +1,7 @@
// SPDX-License-Identifier: CERN-OHL-S-2.0 // SPDX-License-Identifier: CERN-OHL-S-2.0
/* /*
Copyright (c) 2025 FPGA Ninja, LLC Copyright (c) 2026 FPGA Ninja, LLC
Authors: Authors:
- Alex Forencich - Alex Forencich
@@ -15,7 +15,9 @@ Authors:
/* /*
* Corundum-micro completion write module * Corundum-micro completion write module
*/ */
module cndm_micro_cpl_wr module cndm_micro_cpl_wr #(
parameter CQN_W = 5
)
( (
input wire logic clk, input wire logic clk,
input wire logic rst, input wire logic rst,
@@ -42,168 +44,56 @@ module cndm_micro_cpl_wr
output wire logic irq output wire logic irq
); );
localparam AXIL_ADDR_W = s_axil_ctrl_wr.ADDR_W; localparam DMA_ADDR_W = dma_wr_desc_req.DST_ADDR_W;
localparam AXIL_DATA_W = s_axil_ctrl_wr.DATA_W;
localparam APB_ADDR_W = s_apb_dp_ctrl.ADDR_W; logic [CQN_W-1:0] cq_req_cqn_reg = '0;
localparam APB_DATA_W = s_apb_dp_ctrl.DATA_W; logic cq_req_valid_reg = 1'b0;
logic cq_req_ready;
logic [DMA_ADDR_W-1:0] cq_rsp_addr;
logic cq_rsp_phase;
logic cq_rsp_error;
logic cq_rsp_valid;
logic cq_rsp_ready_reg = 1'b0;
logic txcq_en_reg = '0; cndm_micro_queue_state #(
logic [3:0] txcq_size_reg = '0; .QN_W(CQN_W),
logic [63:0] txcq_base_addr_reg = '0; .DQN_W(CQN_W), // TODO
logic rxcq_en_reg = '0; .CPL_SIZE(16),
logic [3:0] rxcq_size_reg = '0; .DMA_ADDR_W(DMA_ADDR_W)
logic [63:0] rxcq_base_addr_reg = '0; )
cq_mgr_inst (
.clk(clk),
.rst(rst),
logic [15:0] txcq_prod_ptr_reg = '0; /*
logic [15:0] rxcq_prod_ptr_reg = '0; * Control register interface
*/
.s_axil_ctrl_wr(s_axil_ctrl_wr),
.s_axil_ctrl_rd(s_axil_ctrl_rd),
logic s_axil_ctrl_awready_reg = 1'b0; /*
logic s_axil_ctrl_wready_reg = 1'b0; * Datapath control register interface
logic s_axil_ctrl_bvalid_reg = 1'b0; */
.s_apb_dp_ctrl(s_apb_dp_ctrl),
logic s_axil_ctrl_arready_reg = 1'b0; /*
logic [AXIL_DATA_W-1:0] s_axil_ctrl_rdata_reg = '0; * CQ management interface
logic s_axil_ctrl_rvalid_reg = 1'b0; */
.req_qn(cq_req_cqn_reg),
assign s_axil_ctrl_wr.awready = s_axil_ctrl_awready_reg; .req_valid(cq_req_valid_reg),
assign s_axil_ctrl_wr.wready = s_axil_ctrl_wready_reg; .req_ready(cq_req_ready),
assign s_axil_ctrl_wr.bresp = '0; .rsp_qn(),
assign s_axil_ctrl_wr.buser = '0; .rsp_dqn(),
assign s_axil_ctrl_wr.bvalid = s_axil_ctrl_bvalid_reg; .rsp_addr(cq_rsp_addr),
.rsp_phase(cq_rsp_phase),
assign s_axil_ctrl_rd.arready = s_axil_ctrl_arready_reg; .rsp_error(cq_rsp_error),
assign s_axil_ctrl_rd.rdata = s_axil_ctrl_rdata_reg; .rsp_valid(cq_rsp_valid),
assign s_axil_ctrl_rd.rresp = '0; .rsp_ready(cq_rsp_ready_reg)
assign s_axil_ctrl_rd.ruser = '0; );
assign s_axil_ctrl_rd.rvalid = s_axil_ctrl_rvalid_reg;
logic s_apb_dp_ctrl_pready_reg = 1'b0;
logic [AXIL_DATA_W-1:0] s_apb_dp_ctrl_prdata_reg = '0;
assign s_apb_dp_ctrl.pready = s_apb_dp_ctrl_pready_reg;
assign s_apb_dp_ctrl.prdata = s_apb_dp_ctrl_prdata_reg;
assign s_apb_dp_ctrl.pslverr = 1'b0;
assign s_apb_dp_ctrl.pruser = '0;
assign s_apb_dp_ctrl.pbuser = '0;
always_ff @(posedge clk) begin
s_axil_ctrl_awready_reg <= 1'b0;
s_axil_ctrl_wready_reg <= 1'b0;
s_axil_ctrl_bvalid_reg <= s_axil_ctrl_bvalid_reg && !s_axil_ctrl_wr.bready;
s_axil_ctrl_arready_reg <= 1'b0;
s_axil_ctrl_rvalid_reg <= s_axil_ctrl_rvalid_reg && !s_axil_ctrl_rd.rready;
s_apb_dp_ctrl_pready_reg <= 1'b0;
if (s_axil_ctrl_wr.awvalid && s_axil_ctrl_wr.wvalid && !s_axil_ctrl_bvalid_reg) begin
s_axil_ctrl_awready_reg <= 1'b1;
s_axil_ctrl_wready_reg <= 1'b1;
s_axil_ctrl_bvalid_reg <= 1'b1;
// case ({s_axil_ctrl_wr.awaddr[9:2], 2'b00})
// 10'h000: begin
// txcq_en_reg <= s_axil_ctrl_wr.wdata[0];
// txcq_size_reg <= s_axil_ctrl_wr.wdata[19:16];
// end
// 10'h008: txcq_base_addr_reg[31:0] <= s_axil_ctrl_wr.wdata;
// 10'h00c: txcq_base_addr_reg[63:32] <= s_axil_ctrl_wr.wdata;
// 10'h100: begin
// rxcq_en_reg <= s_axil_ctrl_wr.wdata[0];
// rxcq_size_reg <= s_axil_ctrl_wr.wdata[19:16];
// end
// 10'h108: rxcq_base_addr_reg[31:0] <= s_axil_ctrl_wr.wdata;
// 10'h10c: rxcq_base_addr_reg[63:32] <= s_axil_ctrl_wr.wdata;
// default: begin end
// endcase
end
if (s_axil_ctrl_rd.arvalid && !s_axil_ctrl_rvalid_reg) begin
s_axil_ctrl_rdata_reg <= '0;
s_axil_ctrl_arready_reg <= 1'b1;
s_axil_ctrl_rvalid_reg <= 1'b1;
// case ({s_axil_ctrl_rd.araddr[9:2], 2'b00})
// 10'h000: begin
// s_axil_ctrl_rdata_reg[0] <= txcq_en_reg;
// s_axil_ctrl_rdata_reg[19:16] <= txcq_size_reg;
// end
// 10'h004: s_axil_ctrl_rdata_reg[15:0] <= txcq_prod_ptr_reg;
// 10'h008: s_axil_ctrl_rdata_reg <= txcq_base_addr_reg[31:0];
// 10'h00c: s_axil_ctrl_rdata_reg <= txcq_base_addr_reg[63:32];
// 10'h100: begin
// s_axil_ctrl_rdata_reg[0] <= rxcq_en_reg;
// s_axil_ctrl_rdata_reg[19:16] <= rxcq_size_reg;
// end
// 10'h104: s_axil_ctrl_rdata_reg[15:0] <= rxcq_prod_ptr_reg;
// 10'h108: s_axil_ctrl_rdata_reg <= rxcq_base_addr_reg[31:0];
// 10'h10c: s_axil_ctrl_rdata_reg <= rxcq_base_addr_reg[63:32];
// default: begin end
// endcase
end
if (s_apb_dp_ctrl.penable && s_apb_dp_ctrl.psel && !s_apb_dp_ctrl_pready_reg) begin
s_apb_dp_ctrl_pready_reg <= 1'b1;
s_apb_dp_ctrl_prdata_reg <= '0;
if (s_apb_dp_ctrl.pwrite) begin
case ({s_apb_dp_ctrl.paddr[9:2], 2'b00})
10'h000: begin
txcq_en_reg <= s_apb_dp_ctrl.pwdata[0];
txcq_size_reg <= s_apb_dp_ctrl.pwdata[19:16];
end
10'h008: txcq_base_addr_reg[31:0] <= s_apb_dp_ctrl.pwdata;
10'h00c: txcq_base_addr_reg[63:32] <= s_apb_dp_ctrl.pwdata;
10'h100: begin
rxcq_en_reg <= s_apb_dp_ctrl.pwdata[0];
rxcq_size_reg <= s_apb_dp_ctrl.pwdata[19:16];
end
10'h108: rxcq_base_addr_reg[31:0] <= s_apb_dp_ctrl.pwdata;
10'h10c: rxcq_base_addr_reg[63:32] <= s_apb_dp_ctrl.pwdata;
default: begin end
endcase
end
case ({s_apb_dp_ctrl.paddr[9:2], 2'b00})
10'h000: begin
s_apb_dp_ctrl_prdata_reg[0] <= txcq_en_reg;
s_apb_dp_ctrl_prdata_reg[19:16] <= txcq_size_reg;
end
10'h004: s_apb_dp_ctrl_prdata_reg[15:0] <= txcq_prod_ptr_reg;
10'h008: s_apb_dp_ctrl_prdata_reg <= txcq_base_addr_reg[31:0];
10'h00c: s_apb_dp_ctrl_prdata_reg <= txcq_base_addr_reg[63:32];
10'h100: begin
s_apb_dp_ctrl_prdata_reg[0] <= rxcq_en_reg;
s_apb_dp_ctrl_prdata_reg[19:16] <= rxcq_size_reg;
end
10'h104: s_apb_dp_ctrl_prdata_reg[15:0] <= rxcq_prod_ptr_reg;
10'h108: s_apb_dp_ctrl_prdata_reg <= rxcq_base_addr_reg[31:0];
10'h10c: s_apb_dp_ctrl_prdata_reg <= rxcq_base_addr_reg[63:32];
default: begin end
endcase
end
if (rst) begin
s_axil_ctrl_awready_reg <= 1'b0;
s_axil_ctrl_wready_reg <= 1'b0;
s_axil_ctrl_bvalid_reg <= 1'b0;
s_axil_ctrl_arready_reg <= 1'b0;
s_axil_ctrl_rvalid_reg <= 1'b0;
s_apb_dp_ctrl_pready_reg <= 1'b0;
end
end
typedef enum logic [1:0] { typedef enum logic [1:0] {
STATE_IDLE, STATE_IDLE,
STATE_RX_CPL, STATE_QUERY_CQ,
STATE_WRITE_DATA STATE_WRITE_DATA
} state_t; } state_t;
@@ -231,13 +121,8 @@ always_ff @(posedge clk) begin
dma_wr_desc_req.req_user <= '0; dma_wr_desc_req.req_user <= '0;
dma_wr_desc_req.req_valid <= dma_wr_desc_req.req_valid && !dma_wr_desc_req.req_ready; dma_wr_desc_req.req_valid <= dma_wr_desc_req.req_valid && !dma_wr_desc_req.req_ready;
if (!txcq_en_reg) begin cq_req_valid_reg <= cq_req_valid_reg && !cq_req_ready;
txcq_prod_ptr_reg <= '0; cq_rsp_ready_reg <= 1'b0;
end
if (!rxcq_en_reg) begin
rxcq_prod_ptr_reg <= '0;
end
irq_reg <= 1'b0; irq_reg <= 1'b0;
@@ -245,29 +130,32 @@ always_ff @(posedge clk) begin
STATE_IDLE: begin STATE_IDLE: begin
dma_wr_desc_req.req_src_addr <= '0; dma_wr_desc_req.req_src_addr <= '0;
if (s_axis_cpl.tdest == 0) begin cq_req_cqn_reg <= CQN_W'(s_axis_cpl.tdest);
dma_wr_desc_req.req_dst_addr <= txcq_base_addr_reg + 64'(16'(txcq_prod_ptr_reg & ({16{1'b1}} >> (16 - txcq_size_reg))) * 16);
phase_tag_reg <= !txcq_prod_ptr_reg[txcq_size_reg]; if (s_axis_cpl.tvalid && !s_axis_cpl.tready) begin
if (s_axis_cpl.tvalid && !s_axis_cpl.tready) begin cq_req_valid_reg <= 1'b1;
txcq_prod_ptr_reg <= txcq_prod_ptr_reg + 1; state_reg <= STATE_QUERY_CQ;
if (txcq_en_reg) begin
dma_wr_desc_req.req_valid <= 1'b1;
state_reg <= STATE_WRITE_DATA;
end else begin
state_reg <= STATE_IDLE;
end
end
end else begin end else begin
dma_wr_desc_req.req_dst_addr <= rxcq_base_addr_reg + 64'(16'(rxcq_prod_ptr_reg & ({16{1'b1}} >> (16 - rxcq_size_reg))) * 16); state_reg <= STATE_IDLE;
phase_tag_reg <= !rxcq_prod_ptr_reg[rxcq_size_reg]; end
if (s_axis_cpl.tvalid && !s_axis_cpl.tready) begin end
rxcq_prod_ptr_reg <= rxcq_prod_ptr_reg + 1; STATE_QUERY_CQ: begin
if (rxcq_en_reg) begin dma_wr_desc_req.req_src_addr <= '0;
dma_wr_desc_req.req_valid <= 1'b1; cq_rsp_ready_reg <= 1'b1;
state_reg <= STATE_WRITE_DATA;
end else begin if (cq_rsp_valid) begin
state_reg <= STATE_IDLE; cq_rsp_ready_reg <= 1'b0;
end
dma_wr_desc_req.req_dst_addr <= cq_rsp_addr;
phase_tag_reg <= cq_rsp_phase;
if (cq_rsp_error) begin
// drop completion
s_axis_cpl.tready <= 1'b1;
state_reg <= STATE_IDLE;
end else begin
dma_wr_desc_req.req_valid <= 1'b1;
state_reg <= STATE_WRITE_DATA;
end end
end end
end end
@@ -285,8 +173,8 @@ always_ff @(posedge clk) begin
if (rst) begin if (rst) begin
state_reg <= STATE_IDLE; state_reg <= STATE_IDLE;
txcq_prod_ptr_reg <= '0; cq_req_valid_reg <= 1'b0;
rxcq_prod_ptr_reg <= '0; cq_rsp_ready_reg <= 1'b0;
irq_reg <= 1'b0; irq_reg <= 1'b0;
end end
end end

View File

@@ -18,8 +18,12 @@ Authors:
module cndm_micro_dp_mgr # module cndm_micro_dp_mgr #
( (
parameter PORTS = 2, parameter PORTS = 2,
parameter CQN_W = 5,
parameter logic PTP_EN = 1'b1, parameter logic PTP_EN = 1'b1,
parameter PTP_BASE_ADDR_DP = 0, parameter PTP_BASE_ADDR_DP = 0,
parameter PORT_BASE_ADDR_DP = 0, parameter PORT_BASE_ADDR_DP = 0,
parameter PORT_BASE_ADDR_HOST = 0 parameter PORT_BASE_ADDR_HOST = 0
) )
@@ -247,9 +251,9 @@ always_comb begin
// determine block base address // determine block base address
case (opcode_reg) case (opcode_reg)
// // EQ
// CMD_OP_CREATE_EQ: // CMD_OP_CREATE_EQ:
// begin // begin
// // EQ
// qn_next = 0; // qn_next = 0;
// dp_ptr_next = DP_APB_ADDR_W'({port_reg, 16'd0} | 'h8000) + DP_APB_ADDR_W'(PORT_BASE_ADDR_DP); // dp_ptr_next = DP_APB_ADDR_W'({port_reg, 16'd0} | 'h8000) + DP_APB_ADDR_W'(PORT_BASE_ADDR_DP);
// host_ptr_next = 32'({port_reg, 16'd0} | 'h8000) + PORT_BASE_ADDR_HOST; // host_ptr_next = 32'({port_reg, 16'd0} | 'h8000) + PORT_BASE_ADDR_HOST;
@@ -258,14 +262,13 @@ always_comb begin
// CMD_OP_QUERY_EQ, // CMD_OP_QUERY_EQ,
// CMD_OP_DESTROY_EQ: // CMD_OP_DESTROY_EQ:
// begin // begin
// // EQ
// dp_ptr_next = DP_APB_ADDR_W'({port_reg, 16'd0} | 'h8000) + DP_APB_ADDR_W'(PORT_BASE_ADDR_DP); // dp_ptr_next = DP_APB_ADDR_W'({port_reg, 16'd0} | 'h8000) + DP_APB_ADDR_W'(PORT_BASE_ADDR_DP);
// host_ptr_next = 32'({port_reg, 16'd0} | 'h8000) + PORT_BASE_ADDR_HOST; // host_ptr_next = 32'({port_reg, 16'd0} | 'h8000) + PORT_BASE_ADDR_HOST;
// end // end
// CQ
CMD_OP_CREATE_CQ: CMD_OP_CREATE_CQ:
begin begin
// CQ cnt_next = 2**CQN_W-1;
cnt_next = 1;
dp_ptr_next = DP_APB_ADDR_W'({port_reg, 16'd0} | 'h8000) + DP_APB_ADDR_W'(PORT_BASE_ADDR_DP); dp_ptr_next = DP_APB_ADDR_W'({port_reg, 16'd0} | 'h8000) + DP_APB_ADDR_W'(PORT_BASE_ADDR_DP);
host_ptr_next = 32'({port_reg, 16'd0} | 'h8000) + PORT_BASE_ADDR_HOST; host_ptr_next = 32'({port_reg, 16'd0} | 'h8000) + PORT_BASE_ADDR_HOST;
end end
@@ -273,13 +276,12 @@ always_comb begin
CMD_OP_QUERY_CQ, CMD_OP_QUERY_CQ,
CMD_OP_DESTROY_CQ: CMD_OP_DESTROY_CQ:
begin begin
// CQ dp_ptr_next = DP_APB_ADDR_W'({port_reg, 16'd0} | 'h8000 | {qn_reg, 5'd00}) + DP_APB_ADDR_W'(PORT_BASE_ADDR_DP);
dp_ptr_next = DP_APB_ADDR_W'({port_reg, 16'd0} | 'h8000 | {qn_reg, 8'd00}) + DP_APB_ADDR_W'(PORT_BASE_ADDR_DP); host_ptr_next = 32'({port_reg, 16'd0} | 'h8000 | {qn_reg, 5'd00}) + PORT_BASE_ADDR_HOST;
host_ptr_next = 32'({port_reg, 16'd0} | 'h8000 | {qn_reg, 8'd00}) + PORT_BASE_ADDR_HOST;
end end
// SQ
CMD_OP_CREATE_SQ: CMD_OP_CREATE_SQ:
begin begin
// SQ
cnt_next = 0; cnt_next = 0;
dp_ptr_next = DP_APB_ADDR_W'({port_reg, 16'd0} | 'h0000) + DP_APB_ADDR_W'(PORT_BASE_ADDR_DP); dp_ptr_next = DP_APB_ADDR_W'({port_reg, 16'd0} | 'h0000) + DP_APB_ADDR_W'(PORT_BASE_ADDR_DP);
host_ptr_next = 32'({port_reg, 16'd0} | 'h0000) + PORT_BASE_ADDR_HOST; host_ptr_next = 32'({port_reg, 16'd0} | 'h0000) + PORT_BASE_ADDR_HOST;
@@ -288,13 +290,12 @@ always_comb begin
CMD_OP_QUERY_SQ, CMD_OP_QUERY_SQ,
CMD_OP_DESTROY_SQ: CMD_OP_DESTROY_SQ:
begin begin
// SQ
dp_ptr_next = DP_APB_ADDR_W'({port_reg, 16'd0} | 'h0000) + DP_APB_ADDR_W'(PORT_BASE_ADDR_DP); dp_ptr_next = DP_APB_ADDR_W'({port_reg, 16'd0} | 'h0000) + DP_APB_ADDR_W'(PORT_BASE_ADDR_DP);
host_ptr_next = 32'({port_reg, 16'd0} | 'h0000) + PORT_BASE_ADDR_HOST; host_ptr_next = 32'({port_reg, 16'd0} | 'h0000) + PORT_BASE_ADDR_HOST;
end end
// RQ
CMD_OP_CREATE_RQ: CMD_OP_CREATE_RQ:
begin begin
// RQ
cnt_next = 0; cnt_next = 0;
dp_ptr_next = DP_APB_ADDR_W'({port_reg, 16'd0} | 'h0100) + DP_APB_ADDR_W'(PORT_BASE_ADDR_DP); dp_ptr_next = DP_APB_ADDR_W'({port_reg, 16'd0} | 'h0100) + DP_APB_ADDR_W'(PORT_BASE_ADDR_DP);
host_ptr_next = 32'({port_reg, 16'd0} | 'h0100) + PORT_BASE_ADDR_HOST; host_ptr_next = 32'({port_reg, 16'd0} | 'h0100) + PORT_BASE_ADDR_HOST;
@@ -303,7 +304,6 @@ always_comb begin
CMD_OP_QUERY_RQ, CMD_OP_QUERY_RQ,
CMD_OP_DESTROY_RQ: CMD_OP_DESTROY_RQ:
begin begin
// RQ
dp_ptr_next = DP_APB_ADDR_W'({port_reg, 16'd0} | 'h0100) + DP_APB_ADDR_W'(PORT_BASE_ADDR_DP); dp_ptr_next = DP_APB_ADDR_W'({port_reg, 16'd0} | 'h0100) + DP_APB_ADDR_W'(PORT_BASE_ADDR_DP);
host_ptr_next = 32'({port_reg, 16'd0} | 'h0100) + PORT_BASE_ADDR_HOST; host_ptr_next = 32'({port_reg, 16'd0} | 'h0100) + PORT_BASE_ADDR_HOST;
end end
@@ -470,7 +470,8 @@ always_comb begin
end else begin end else begin
// queue is active // queue is active
qn_next = qn_reg + 1; qn_next = qn_reg + 1;
dp_ptr_next = dp_ptr_reg + 'h100; dp_ptr_next = dp_ptr_reg + 'h20;
host_ptr_next = host_ptr_reg + 'h20;
if (cnt_reg == 0) begin if (cnt_reg == 0) begin
// no more queues // no more queues
m_axis_rsp_tdata_next = '0; // TODO m_axis_rsp_tdata_next = '0; // TODO

View File

@@ -36,6 +36,9 @@ module cndm_micro_pcie_us #(
// Structural configuration // Structural configuration
parameter PORTS = 2, parameter PORTS = 2,
// Queue configuration
parameter CQN_W = 5,
// PTP configuration // PTP configuration
parameter logic PTP_TS_EN = 1'b1, parameter logic PTP_TS_EN = 1'b1,
parameter logic PTP_TS_FMT_TOD = 1'b0, parameter logic PTP_TS_FMT_TOD = 1'b0,
@@ -510,6 +513,9 @@ cndm_micro_core #(
// Structural configuration // Structural configuration
.PORTS(PORTS), .PORTS(PORTS),
// Queue configuration
.CQN_W(CQN_W),
// PTP configuration // PTP configuration
.PTP_TS_EN(PTP_TS_EN), .PTP_TS_EN(PTP_TS_EN),
.PTP_TS_FMT_TOD(PTP_TS_FMT_TOD), .PTP_TS_FMT_TOD(PTP_TS_FMT_TOD),

View File

@@ -16,6 +16,10 @@ Authors:
* Corundum-micro port module * Corundum-micro port module
*/ */
module cndm_micro_port #( module cndm_micro_port #(
// Queue configuration
parameter CQN_W = 5,
// PTP configuration
parameter logic PTP_TS_EN = 1'b1, parameter logic PTP_TS_EN = 1'b1,
parameter logic PTP_TS_FMT_TOD = 1'b0 parameter logic PTP_TS_FMT_TOD = 1'b0
) )
@@ -387,7 +391,9 @@ cpl_mux_inst (
.m_axis(axis_cpl) .m_axis(axis_cpl)
); );
cndm_micro_cpl_wr cndm_micro_cpl_wr #(
.CQN_W(CQN_W)
)
cpl_wr_inst ( cpl_wr_inst (
.clk(clk), .clk(clk),
.rst(rst), .rst(rst),

View File

@@ -0,0 +1,294 @@
// SPDX-License-Identifier: CERN-OHL-S-2.0
/*
Copyright (c) 2026 FPGA Ninja, LLC
Authors:
- Alex Forencich
*/
`resetall
`timescale 1ns / 1ps
`default_nettype none
/*
* Corundum-micro queue state manager module
*/
module cndm_micro_queue_state #(
parameter QN_W = 5,
parameter DQN_W = 5,
parameter CPL_SIZE = 16,
parameter DMA_ADDR_W = 64
)
(
input wire logic clk,
input wire logic rst,
/*
* Control register interface
*/
taxi_axil_if.wr_slv s_axil_ctrl_wr,
taxi_axil_if.rd_slv s_axil_ctrl_rd,
/*
* Datapath control register interface
*/
taxi_apb_if.slv s_apb_dp_ctrl,
/*
* CQ management interface
*/
input wire logic [QN_W-1:0] req_qn,
input wire logic req_valid,
output wire logic req_ready,
output wire logic [QN_W-1:0] rsp_qn,
output wire logic [DQN_W-1:0] rsp_dqn,
output wire logic [DMA_ADDR_W-1:0] rsp_addr,
output wire logic rsp_phase,
output wire logic rsp_error,
output wire logic rsp_valid,
input wire logic rsp_ready
);
localparam PTR_W = 16;
localparam ADDR_W = QN_W+5;
localparam AXIL_ADDR_W = s_axil_ctrl_wr.ADDR_W;
localparam AXIL_DATA_W = s_axil_ctrl_wr.DATA_W;
localparam APB_ADDR_W = s_apb_dp_ctrl.ADDR_W;
localparam APB_DATA_W = s_apb_dp_ctrl.DATA_W;
// check configuration
if (s_axil_ctrl_rd.DATA_W != 32 || s_axil_ctrl_wr.DATA_W != 32)
$fatal(0, "Error: AXI data width must be 32 (instance %m)");
if (s_axil_ctrl_rd.ADDR_W < ADDR_W || s_axil_ctrl_wr.ADDR_W < ADDR_W)
$fatal(0, "Error: AXI address width is insufficient (instance %m)");
if (s_apb_dp_ctrl.DATA_W != 32)
$fatal(0, "Error: APB data width must be 32 (instance %m)");
if (s_apb_dp_ctrl.ADDR_W < ADDR_W)
$fatal(0, "Error: APB address width is insufficient (instance %m)");
logic s_axil_ctrl_awready_reg = 1'b0, s_axil_ctrl_awready_next;
logic s_axil_ctrl_wready_reg = 1'b0, s_axil_ctrl_wready_next;
logic s_axil_ctrl_bvalid_reg = 1'b0, s_axil_ctrl_bvalid_next;
logic s_axil_ctrl_arready_reg = 1'b0, s_axil_ctrl_arready_next;
logic [AXIL_DATA_W-1:0] s_axil_ctrl_rdata_reg = '0, s_axil_ctrl_rdata_next;
logic s_axil_ctrl_rvalid_reg = 1'b0, s_axil_ctrl_rvalid_next;
assign s_axil_ctrl_wr.awready = s_axil_ctrl_awready_reg;
assign s_axil_ctrl_wr.wready = s_axil_ctrl_wready_reg;
assign s_axil_ctrl_wr.bresp = '0;
assign s_axil_ctrl_wr.buser = '0;
assign s_axil_ctrl_wr.bvalid = s_axil_ctrl_bvalid_reg;
assign s_axil_ctrl_rd.arready = s_axil_ctrl_arready_reg;
assign s_axil_ctrl_rd.rdata = s_axil_ctrl_rdata_reg;
assign s_axil_ctrl_rd.rresp = '0;
assign s_axil_ctrl_rd.ruser = '0;
assign s_axil_ctrl_rd.rvalid = s_axil_ctrl_rvalid_reg;
wire [QN_W-1:0] s_axil_ctrl_awaddr_queue_index = s_axil_ctrl_wr.awaddr[5 +: QN_W];
wire [2:0] s_axil_ctrl_awaddr_reg_index = s_axil_ctrl_wr.awaddr[4:2];
wire [QN_W-1:0] s_axil_ctrl_araddr_queue_index = s_axil_ctrl_rd.araddr[5 +: QN_W];
wire [2:0] s_axil_ctrl_araddr_reg_index = s_axil_ctrl_rd.araddr[4:2];
logic s_apb_dp_ctrl_pready_reg = 1'b0, s_apb_dp_ctrl_pready_next;
logic [AXIL_DATA_W-1:0] s_apb_dp_ctrl_prdata_reg = '0, s_apb_dp_ctrl_prdata_next;
assign s_apb_dp_ctrl.pready = s_apb_dp_ctrl_pready_reg;
assign s_apb_dp_ctrl.prdata = s_apb_dp_ctrl_prdata_reg;
assign s_apb_dp_ctrl.pslverr = 1'b0;
assign s_apb_dp_ctrl.pruser = '0;
assign s_apb_dp_ctrl.pbuser = '0;
wire [QN_W-1:0] s_apb_dp_ctrl_paddr_queue_index = s_apb_dp_ctrl.paddr[5 +: QN_W];
wire [2:0] s_apb_dp_ctrl_paddr_reg_index = s_apb_dp_ctrl.paddr[4:2];
logic req_ready_reg = 1'b0, req_ready_next;
logic [QN_W-1:0] rsp_qn_reg = '0, rsp_qn_next;
logic [DQN_W-1:0] rsp_dqn_reg = '0, rsp_dqn_next;
logic [DMA_ADDR_W-1:0] rsp_addr_reg = '0, rsp_addr_next;
logic rsp_phase_reg = 1'b0, rsp_phase_next;
logic rsp_error_reg = 1'b0, rsp_error_next;
logic rsp_valid_reg = 1'b0, rsp_valid_next;
assign req_ready = req_ready_reg;
assign rsp_qn = rsp_qn_reg;
assign rsp_dqn = rsp_dqn_reg;
assign rsp_addr = rsp_addr_reg;
assign rsp_phase = rsp_phase_reg;
assign rsp_error = rsp_error_reg;
assign rsp_valid = rsp_valid_reg;
logic [2**QN_W-1:0] queue_enable_reg = '0;
(* ram_style = "distributed", ramstyle = "no_rw_check, mlab" *)
logic [3:0] queue_mem_size[2**QN_W] = '{default: '0};
(* ram_style = "distributed", ramstyle = "no_rw_check, mlab" *)
logic [DMA_ADDR_W-1:0] queue_mem_base_addr[2**QN_W] = '{default: '0};
(* ram_style = "distributed", ramstyle = "no_rw_check, mlab" *)
logic [PTR_W-1:0] queue_mem_prod[2**QN_W] = '{default: '0};
logic queue_mem_wr_en;
logic [QN_W-1:0] queue_mem_addr;
wire queue_mem_rd_enable = queue_enable_reg[queue_mem_addr];
wire [3:0] queue_mem_rd_size = queue_mem_size[queue_mem_addr];
wire [DMA_ADDR_W-1:0] queue_mem_rd_base_addr = queue_mem_base_addr[queue_mem_addr];
wire [PTR_W-1:0] queue_mem_rd_prod = queue_mem_prod[queue_mem_addr];
logic queue_mem_wr_enable;
logic [3:0] queue_mem_wr_size;
logic [DMA_ADDR_W-1:0] queue_mem_wr_base_addr;
logic [PTR_W-1:0] queue_mem_wr_prod;
always_comb begin
s_axil_ctrl_awready_next = 1'b0;
s_axil_ctrl_wready_next = 1'b0;
s_axil_ctrl_bvalid_next = 1'b0;
s_axil_ctrl_arready_next = 1'b0;
s_axil_ctrl_rdata_next = s_axil_ctrl_rdata_reg;
s_axil_ctrl_rvalid_next = 1'b0;
s_apb_dp_ctrl_pready_next = 1'b0;
s_apb_dp_ctrl_prdata_next = s_apb_dp_ctrl_prdata_reg;
req_ready_next = 1'b0;
rsp_qn_next = rsp_qn_reg;
rsp_dqn_next = rsp_dqn_reg;
rsp_addr_next = rsp_addr_reg;
rsp_phase_next = rsp_phase_reg;
rsp_error_next = rsp_error_reg;
rsp_valid_next = rsp_valid_reg && !rsp_ready;
queue_mem_wr_en = 1'b0;
queue_mem_addr = '0;
queue_mem_wr_enable = queue_mem_rd_enable;
queue_mem_wr_size = queue_mem_rd_size;
queue_mem_wr_base_addr = queue_mem_rd_base_addr;
queue_mem_wr_prod = queue_mem_rd_prod;
// terminate AXI lite writes
if (s_axil_ctrl_wr.awvalid && s_axil_ctrl_wr.wvalid && !s_axil_ctrl_bvalid_reg) begin
s_axil_ctrl_awready_next = 1'b1;
s_axil_ctrl_wready_next = 1'b1;
s_axil_ctrl_bvalid_next = 1'b1;
end
// terminate AXI lite reads
if (s_axil_ctrl_rd.arvalid && !s_axil_ctrl_rvalid_reg) begin
s_axil_ctrl_rdata_next = '0;
s_axil_ctrl_arready_next = 1'b1;
s_axil_ctrl_rvalid_next = 1'b1;
end
if (s_apb_dp_ctrl.penable && s_apb_dp_ctrl.psel && !s_apb_dp_ctrl_pready_reg) begin
// APB read/write
s_apb_dp_ctrl_pready_next = 1'b1;
s_apb_dp_ctrl_prdata_next = '0;
queue_mem_addr = s_apb_dp_ctrl_paddr_queue_index;
if (s_apb_dp_ctrl.pwrite) begin
queue_mem_wr_en = 1'b1;
case (s_apb_dp_ctrl_paddr_reg_index)
3'd0: begin
queue_mem_wr_enable = s_apb_dp_ctrl.pwdata[0];
queue_mem_wr_size = s_apb_dp_ctrl.pwdata[19:16];
end
3'd1: queue_mem_wr_prod = s_apb_dp_ctrl.pwdata[15:0];
3'd2: queue_mem_wr_base_addr[31:0] = s_apb_dp_ctrl.pwdata;
3'd3: queue_mem_wr_base_addr[63:32] = s_apb_dp_ctrl.pwdata;
default: begin end
endcase
end
case (s_apb_dp_ctrl_paddr_reg_index)
3'd0: begin
s_apb_dp_ctrl_prdata_next[0] = queue_mem_rd_enable;
s_apb_dp_ctrl_prdata_next[19:16] = queue_mem_rd_size;
end
3'd1: s_apb_dp_ctrl_prdata_next[15:0] = queue_mem_rd_prod;
3'd2: s_apb_dp_ctrl_prdata_next = queue_mem_rd_base_addr[31:0];
3'd3: s_apb_dp_ctrl_prdata_next = queue_mem_rd_base_addr[63:32];
default: begin end
endcase
end else if (req_valid && !req_ready && (!rsp_valid || rsp_ready)) begin
// completion enqueue request
req_ready_next = 1'b1;
queue_mem_addr = req_qn;
rsp_qn_next = req_qn;
rsp_dqn_next = '0; // TODO
rsp_addr_next = queue_mem_rd_base_addr + DMA_ADDR_W'(16'(queue_mem_rd_prod & ({16{1'b1}} >> (16 - queue_mem_rd_size))) * CPL_SIZE);
rsp_phase_next = !queue_mem_rd_prod[queue_mem_rd_size];
rsp_error_next = !queue_mem_rd_enable;
rsp_valid_next = 1'b1;
queue_mem_wr_prod = queue_mem_rd_prod + 1;
if (queue_mem_rd_enable) begin
queue_mem_wr_en = 1'b1;
end
end
end
always @(posedge clk) begin
s_axil_ctrl_awready_reg <= s_axil_ctrl_awready_next;
s_axil_ctrl_wready_reg <= s_axil_ctrl_wready_next;
s_axil_ctrl_bvalid_reg <= s_axil_ctrl_bvalid_next;
s_axil_ctrl_arready_reg <= s_axil_ctrl_arready_next;
s_axil_ctrl_rdata_reg <= s_axil_ctrl_rdata_next;
s_axil_ctrl_rvalid_reg <= s_axil_ctrl_rvalid_next;
s_apb_dp_ctrl_pready_reg <= s_apb_dp_ctrl_pready_next;
s_apb_dp_ctrl_prdata_reg <= s_apb_dp_ctrl_prdata_next;
req_ready_reg <= req_ready_next;
rsp_qn_reg <= rsp_qn_next;
rsp_dqn_reg <= rsp_dqn_next;
rsp_addr_reg <= rsp_addr_next;
rsp_phase_reg <= rsp_phase_next;
rsp_error_reg <= rsp_error_next;
rsp_valid_reg <= rsp_valid_next;
if (queue_mem_wr_en) begin
queue_enable_reg[queue_mem_addr] <= queue_mem_wr_enable;
queue_mem_size[queue_mem_addr] <= queue_mem_wr_size;
queue_mem_base_addr[queue_mem_addr] <= queue_mem_wr_base_addr;
queue_mem_prod[queue_mem_addr] <= queue_mem_wr_prod;
end
if (rst) begin
s_axil_ctrl_awready_reg <= 1'b0;
s_axil_ctrl_wready_reg <= 1'b0;
s_axil_ctrl_bvalid_reg <= 1'b0;
s_axil_ctrl_arready_reg <= 1'b0;
s_axil_ctrl_rvalid_reg <= 1'b0;
s_apb_dp_ctrl_pready_reg <= 1'b0;
req_ready_reg <= 1'b0;
rsp_valid_reg <= 1'b0;
queue_enable_reg <= '0;
end
end
endmodule
`resetall