fallensigh/ciciec2026
Archived
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Compute matmul CRC before result store

Browse Source
This commit is contained in:
FallenSigh
2026-07-04 19:45:22 +08:00
parent bf6051c8c6
commit c37f493bcb
5 changed files with 64 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files

67
rtl/ip/matmul/matmul_axi_slave.v
View File

@@ -183,22 +183,21 @@ module matmul_axi_slave (
reg [31:0] crc32_table [0:255];
reg [31:0] crc_reg;
wire [31:0] crc_stream_data = (ep_rd_valid && ep_rd_ready) ? ep_rd_data : s_axi_rdata;
wire crc_axi_valid = (r_state == R_DATA) && s_axi_rvalid && s_axi_rready &&
((r_offset >= ADDR_B0_C && r_offset < (ADDR_B0_C + 12'h0C0)) ||
(r_offset >= ADDR_B1_C && r_offset < (ADDR_B1_C + 12'h0C0)));
wire crc_ep_valid = ep_rd_valid && ep_rd_ready && (ep_rd_word_addr < 8'd48);
wire crc_stream_valid = crc_axi_valid || crc_ep_valid;
wire [31:0] crc_t0, crc_c0, crc_t1, crc_c1, crc_t2, crc_c2, crc_t3, crc_next;
assign crc_t0 = crc32_table[(crc_reg[7:0] ^ crc_stream_data[7:0])];
assign crc_c0 = crc_t0 ^ {8'h00, crc_reg[31:8]};
assign crc_t1 = crc32_table[(crc_c0[7:0] ^ crc_stream_data[15:8])];
assign crc_c1 = crc_t1 ^ {8'h00, crc_c0[31:8]};
assign crc_t2 = crc32_table[(crc_c1[7:0] ^ crc_stream_data[23:16])];
assign crc_c2 = crc_t2 ^ {8'h00, crc_c1[31:8]};
assign crc_t3 = crc32_table[(crc_c2[7:0] ^ crc_stream_data[31:24])];
assign crc_next = crc_t3 ^ {8'h00, crc_c2[31:8]};
function [31:0] crc32_word_next;
input [31:0] crc_in;
input [31:0] data_in;
reg [31:0] t0, c0, t1, c1, t2, c2, t3;
begin
t0 = crc32_table[(crc_in[7:0] ^ data_in[7:0])];
c0 = t0 ^ {8'h00, crc_in[31:8]};
t1 = crc32_table[(c0[7:0] ^ data_in[15:8])];
c1 = t1 ^ {8'h00, c0[31:8]};
t2 = crc32_table[(c1[7:0] ^ data_in[23:16])];
c2 = t2 ^ {8'h00, c1[31:8]};
t3 = crc32_table[(c2[7:0] ^ data_in[31:24])];
crc32_word_next = t3 ^ {8'h00, c2[31:8]};
end
endfunction
initial begin
crc32_table[0 ] = 32'h00000000; crc32_table[1 ] = 32'h77073096;
@@ -510,7 +509,7 @@ module matmul_axi_slave (
end
R_DATA: begin
if (s_axi_rvalid && s_axi_rready) begin
// 【修改点 3】删掉直接赋值 crc_stream_valid 的逻辑，移至专用 Pipeline 块
// CRC is updated from completed compute results, not from C reads.
if (r_cnt == r_len) begin
s_axi_rvalid <= 1'b0;
s_axi_rlast <= 1'b0;
@@ -560,17 +559,6 @@ module matmul_axi_slave (
end
end
// CRC register: update on each C data read beat, clear on write to 0x010
always @(posedge clk) begin
if (!resetn) begin
crc_reg <= 32'hFFFFFFFF;
end else if (crc_clear_pulse) begin
crc_reg <= 32'hFFFFFFFF;
end else if (crc_stream_valid) begin
crc_reg <= crc_next;
end
end
// ==========================================
// 计算引擎 (Compute FSM)
// ==========================================
@@ -595,6 +583,23 @@ module matmul_axi_slave (
reg [3:0] acc_idx_d1, acc_idx_d2, acc_idx_d3, acc_idx_d4;
reg acc_bank_d1, acc_bank_d2, acc_bank_d3, acc_bank_d4;
reg [4:0] pipe_cnt;
wire [65:0] calc_c_result = ({2'b0, mult_p[0]} + {2'b0, mult_p[1]}) +
({2'b0, mult_p[2]} + {2'b0, mult_p[3]});
wire calc_crc_valid = (state == C_RUN) && (pipe_cnt >= 5'd4) && (pipe_cnt < 5'd20);
wire [31:0] crc_calc_0 = crc32_word_next(crc_reg, calc_c_result[31:0]);
wire [31:0] crc_calc_1 = crc32_word_next(crc_calc_0, calc_c_result[63:32]);
wire [31:0] crc_calc_2 = crc32_word_next(crc_calc_1, {30'b0, calc_c_result[65:64]});
// CRC tracks completed C elements directly, independent of result writeback.
always @(posedge clk) begin
if (!resetn) begin
crc_reg <= 32'hFFFFFFFF;
end else if (crc_clear_pulse) begin
crc_reg <= 32'hFFFFFFFF;
end else if (calc_crc_valid) begin
crc_reg <= crc_calc_2;
end
end
always @(posedge clk) begin
if (!resetn) begin
@@ -659,11 +664,9 @@ module matmul_axi_slave (
// Accumulate results from 4 cycles ago (pipe_cnt 4-19)
if (pipe_cnt >= 5'd4 && pipe_cnt < 5'd20) begin
if (acc_bank_d4 == 1'b0)
b0_c[acc_idx_d4] <= ({2'b0, mult_p[0]} + {2'b0, mult_p[1]}) +
({2'b0, mult_p[2]} + {2'b0, mult_p[3]});
b0_c[acc_idx_d4] <= calc_c_result;
else
b1_c[acc_idx_d4] <= ({2'b0, mult_p[0]} + {2'b0, mult_p[1]}) +
({2'b0, mult_p[2]} + {2'b0, mult_p[3]});
b1_c[acc_idx_d4] <= calc_c_result;
end
if (pipe_cnt == 5'd20) begin