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

Avoid DSP in matmul master path

Browse Source
This commit is contained in:
FallenSigh
2026-07-01 23:30:11 +08:00
parent fd03aaaaa1
commit baf3f6f696
3 changed files with 76 additions and 287 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
rtl/ip/matmul/matmul_axi_slave.v
View File

@@ -2,7 +2,7 @@
// 3-stage pipelined multiplier (LUT-based, no DSP)
// latency=3, throughput=1 result/cycle
// ═══════════════════════════════════════════════════════
(* use_dsp = "no" *)
(* use_dsp = "no", use_dsp48 = "no", syn_multstyle = "logic" *)
module pipelined_mult_32x32 (
input clk,
input rstn,
@@ -11,7 +11,23 @@ module pipelined_mult_32x32 (
output reg [63:0] p
);
reg [31:0] a_r, b_r;
(* use_dsp = "no", use_dsp48 = "no", syn_multstyle = "logic" *)
reg [63:0] prod;
function automatic [63:0] lut_mul32;
input [31:0] lhs;
input [31:0] rhs;
integer bit_idx;
reg [63:0] acc;
begin
acc = 64'b0;
for (bit_idx = 0; bit_idx < 32; bit_idx = bit_idx + 1) begin
if (rhs[bit_idx])
acc = acc + ({32'b0, lhs} << bit_idx);
end
lut_mul32 = acc;
end
endfunction
always @(posedge clk) begin
if (!rstn) begin
@@ -20,7 +36,7 @@ module pipelined_mult_32x32 (
end else begin
a_r <= a; // Stage 1: register inputs
b_r <= b;
prod <= a_r * b_r; // Stage 2: LUT-based multiply
prod <= lut_mul32(a_r, b_r); // Stage 2: LUT-based multiply
p <= prod; // Stage 3: register output
end
end
@@ -186,6 +202,8 @@ module matmul_axi_slave (
reg dma_done;
reg [31:0] load_group;
reg [31:0] store_group;
reg [31:0] dma_next_rd_addr;
reg [31:0] dma_next_wr_addr;
reg [31:0] completed_groups;
reg [1:0] bank_loaded;
reg [1:0] bank_store_pending;
@@ -880,6 +898,8 @@ module matmul_axi_slave (
dma_done <= 1'b0;
load_group <= 32'd0;
store_group <= 32'd0;
dma_next_rd_addr <= 32'd0;
dma_next_wr_addr <= 32'd0;
completed_groups <= 32'd0;
bank_loaded <= 2'b00;
bank_store_pending <= 2'b00;
@@ -933,6 +953,8 @@ module matmul_axi_slave (
dma_done <= 1'b0;
load_group <= 32'd0;
store_group <= 32'd0;
dma_next_rd_addr <= dma_src_base;
dma_next_wr_addr <= dma_dst_base;
completed_groups <= 32'd0;
bank_loaded <= 2'b00;
bank_store_pending <= 2'b00;
@@ -978,7 +1000,7 @@ module matmul_axi_slave (
!(state == C_RUN && cur_bank == dma_load_bank)) begin
dma_rd_bank <= dma_load_bank;
dma_rd_count <= 5'd0;
m_axi_araddr <= dma_src_base + (load_group * dma_src_stride);
m_axi_araddr <= dma_next_rd_addr;
m_axi_arlen <= 8'd31;
m_axi_arsize <= 3'd2;
m_axi_arburst <= 2'b01;
@@ -999,6 +1021,7 @@ module matmul_axi_slave (
m_axi_rready <= 1'b0;
bank_loaded[dma_rd_bank] <= 1'b1;
load_group <= load_group + 32'd1;
dma_next_rd_addr <= dma_next_rd_addr + dma_src_stride;
dma_load_bank <= ~dma_load_bank;
dma_rd_state <= DMA_RD_IDLE;
end else begin
@@ -1018,7 +1041,7 @@ module matmul_axi_slave (
dma_wr_count <= 6'd0;
bank_store_pending[dma_store_bank] <= 1'b0;
bank_storing[dma_store_bank] <= 1'b1;
m_axi_awaddr <= dma_dst_base + (store_group * dma_dst_stride);
m_axi_awaddr <= dma_next_wr_addr;
m_axi_awlen <= 8'd47;
m_axi_awsize <= 3'd2;
m_axi_awburst <= 2'b01;
@@ -1055,6 +1078,7 @@ module matmul_axi_slave (
bank_storing[dma_store_bank] <= 1'b0;
dma_store_bank <= ~dma_store_bank;
store_group <= store_group + 32'd1;
dma_next_wr_addr <= dma_next_wr_addr + dma_dst_stride;
completed_groups <= completed_groups + 32'd1;
if ((completed_groups + 32'd1) >= dma_groups) begin
dma_busy <= 1'b0;

300
sdk/axi_ram.mif
View File

@@ -52,11 +52,11 @@
00000011100000000110000000001101
00000100000000000000000110101100
00011100000000000000000000101100
00000010101101001111000110001100
00000010101010001011000110001100
00011100000000000001000000001101
00000010101111001001000110101101
00011100000000000001000000001110
00000010101111011111000111001110
00000010101111010111000111001110
01101100000000000001100110101110
00101000100000000000000110001111
00101001100000000000000110101111
@@ -64,9 +64,9 @@
00000010100000000001000110101101
01101011111111111111000110101110
00011100000000000001000000001100
00000010101111010111000110001100
00000010101111001111000110001100
00011100000000000001000000001101
00000010101111010101000110101101
00000010101111001101000110101101
01101100000000000001000110001101
00101001100000000000000110000000
00000010100000000001000110001100
@@ -74,7 +74,7 @@
00000011100000000110010000001100
00000100000001100000000000101100
00011100000000000001000000001100
00000010101111001100000110001100
00000010101111000001000110001100
00101000100000000000000110001101
00000011100000000001110000001110
00101001000000000000100110101110
@@ -94,7 +94,7 @@
00000000000101010000000000001110
00101001000000000001000110101110
00011100000000000000000000001100
00000010100011010001000110001100
00000010100010100001000110001100
00000100000000000011000000101100
00000011100000000000010000001101
00000100000000010001000000101101
@@ -142,14 +142,11 @@
00000000000000000000000000000000
00000000000000000000000000000000
00000000000000000000000000000000
00000010101111111000000001100011
00010101011111100000000000010000
00101001100000000111000001100001
00101001100000000110000001110111
00101001100000000101000001111000
00000010101111111100000001100011
00000010100000010011010000001100
00011100000000000001000000001111
00000010101101101010000111101111
00000010101101101101000111101111
00000011100000000001011000001101
00101001100000000011000001101100
00101010000000000011000001101110
@@ -160,56 +157,43 @@
00101001000000000000001000001110
00101000000000000000000111101100
01011111111111111110000110000000
01010100000001100111100000000000
00010101011111101010000000001100
00000011100000000100000110010001
00101001100000000000001000100000
00000010100000000101100000010000
00000011100000000100000110001111
00101001100000000000000111100000
00000010100000000101100000010010
00000011100000000001000110001101
00101001100000000000000110110000
00010100001110001000000000001110
00000011100000001000000110001111
00101001100000000000000111101110
00000011100101000000000111000101
00000000000101010000000010011000
00000011100000001001000110000100
00101001100000000000000010000101
00000011100000001010000110010011
00000010100000000010100000010100
00101001100000000000000110110010
00010100001110001000000000010001
00000011100000001000000110010000
00010100001110001001001110001110
00101001100000000000001000010001
00000011100100000000000111001110
00000011100000001001000110000101
00010100000000000000000000101111
00101001100000000000000010101110
00000011100000001010000110000100
00000011100011100010000111101111
00101001100000000000000010001111
00000011100000001011000110010011
00000010100000100000000000010100
00101001100000000000001001110100
00000011100000001011000110010001
00000010100000100000000000010010
00000011100000001100000110010001
00000010100000110000000000010010
00000010100001000000000000010000
00101001100000000000001000110010
00000011100000001100000110001111
00000010100000110000000000010000
00000010100001000000000000001110
00101001100000000000000111110000
00010101011111101010000000010111
00101001100000000000000110001110
00010101011111101010000000001110
00101001100000000000000110010000
00101000100000000000000110101100
00000011010000000100000110001100
01011011111111111111100110000000
00000010100000000100000000001100
00101001100000000000000110101100
01010100000001011111110000000000
00011100000000000001000000001101
00000010101101010100000110101101
00000011100000000100001011110111
00101000100000000000000110110010
00101000100000000000001011110001
00010100000000000001111010001100
00000010100000000100000000001111
00101001100000000000000110101111
00000011100000000100000111001110
00101000100000000000000111010001
00010101011111100000000000010000
00000011100010010000000110001110
00000000000100010110000010000100
00000000001000010011101001010011
01011100000000000000100111000000
00000000001010100000000000000111
00000010100000010011010000001100
00000000001000010100110010010010
01011100000000000000101001100000
00000000001010100000000000000111
00011100000000000001000000001111
00000010101100110010000111101111
00000010101101000010000111101111
00000011100000000001011000001101
00101001100000000011000001101100
00101010000000000011000001101110
@@ -222,25 +206,25 @@
01011111111111111110000110000000
00010101011111100000000000010000
00000010100000000111000000001110
00000010100000000010010000010100
00000010101111111111000000010011
00000010100000000010010000010011
00000010101111111111000000010010
00000011100000000001011000001101
00000011010000000000000000000000
00000000000101111011101000101100
00000011010000000011110110001100
00000010100000010101110110001111
01101000000000000000101010001100
01101000000000000000101001101100
00000010100000001100000110001111
00101010000000000000000110101100
00000011010000001000000110001100
01011011111111111111100110000000
00101001000000000000001000001111
00000010101111111111000111001110
01011111111111111101100111010011
01011111111111111101100111010010
00010101011111100000000000010000
00000010100000000011010000001100
00011100000000000001000000001111
00000010101100010111000111101111
00000010101100100111000111101111
00000011100000000001011000001101
00101001100000000011000001101100
00101010000000000011000001101110
@@ -252,44 +236,12 @@
00101001000000000000001000001110
00101000000000000000000111101100
01011111111111111101110110000000
01010100000000101100100000000000
00010101011111100000000000010000
00000010100000000011010000001100
00011100000000000001000000001111
00000010101100001100000111101111
00000011100000000001011000001101
00101001100000000011000001101100
00101010000000000011000001101110
00000010100000000000010111101111
00000011010000000000000000000000
00101010000000000000000110101100
00000011010000001000000110001100
01011011111111111111100110000000
00101001000000000000001000001110
00101000000000000000000111101100
01011111111111111101110110000000
00000000000101010000001001000100
01010100000000101000010000000000
00010101011111100000000000010000
00000010100000000011010000001100
00011100000000000001000000001111
00000010101100000000000111101111
00000011100000000001011000001101
00101001100000000011000001101100
00101010000000000011000001101110
00000010100000000000010111101111
00101010000000000000000110101100
00000011010000001000000110001100
01011011111111111111100110000000
00101001000000000000001000001110
00101000000000000000000111101100
01011111111111111110000110000000
00101000100000000111000001100001
00101000100000000110000001110111
00101000100000000101000001111000
00000000000101010000000000000100
00000010100000001000000001100011
00000010100000000100000001100011
01001100000000000000000000100000
00000011010000000000000000000000
00000011010000000000000000000000
00000011010000000000000000000000
00000000000000000000000000000000
00000000000000000000000000000000
00000000000000000000000000000000
@@ -430,154 +382,6 @@
00000011010000000000000000000000
00000011010000000000000000000000
00000011010000000000000000000000
01011000000000100001010010000000
00000010100000000010100000001100
00000010101111111100000001100011
00000000001000011011000010001101
01011100000000000000100110000000
00000000001010100000000000000111
00000000001000010011000010010000
01011100000000000000100110000000
00000000001010100000000000000111
00000010100000001100000110101101
00000010100000000010010000001111
00101001000000000001000001101101
00000010100000000000010000001110
01101100000000011010100111100100
00000000001000011011001000001101
01011100000000000000100110000000
00000000001010100000000000000111
00000010100000001100000110101101
00000010100000011001000000001110
00000010100000011000110000001111
00000000001000010011100010010000
01011100000000000000100111000000
00000000001010100000000000000111
00101001000000000001010001101101
00000010100000000000100000001110
01101100000000010111100111100100
00000000001000011011001000001101
01011100000000000000100110000000
00000000001010100000000000000111
00000010100000001100000110101101
00000010100011111010000000001110
00000010100011111001110000001111
00000000001000010011100010010000
01011100000000000000100111000000
00000000001010100000000000000111
00101001000000000001100001101101
00000010100000000000110000001110
01101100000000010100100111100100
00010100000000000000000001001101
00000000001000011011001000001111
01011100000000000000100110000000
00000000001010100000000000000111
00000010100000001100000111101111
00000011100111000100000110110000
00101001000000000001110001101111
00000000001000010100000010010001
01011100000000000000101000000000
00000000001010100000000000000111
00000011100111000011110110101101
00000010100000000001000000001110
01101100000000010001010110100100
00010100000000000000001100001101
00000000001000011011001000101111
01011100000000000000100110000000
00000000001010100000000000000111
00000010100000001100000111101111
00000011100110101000000110110000
00101001000000000010000001101111
00000000001000010100000010010001
01011100000000000000101000000000
00000000001010100000000000000111
00000011100110100111110110101101
00000010100000000001010000001110
01101100000000001110000110100100
00010100000000000001111010001101
00000000001000011011001000101111
01011100000000000000100110000000
00000000001010100000000000000111
00000010100000001100000111101111
00000011100010010000000110110000
00101001000000000010010001101111
00000000001000010100000010010001
01011100000000000000101000000000
00000000001010100000000000000111
00000011100010001111110110101101
00000010100000000001100000001110
01101100000000001010110110100100
00010100000000010011000100101101
00000000001000011011001000101111
01011100000000000000100110000000
00000000001010100000000000000111
00000010100000001100000111101111
00000011100110100000000110110000
00101001000000000010100001101111
00000000001000010100000010010001
01011100000000000000101000000000
00000000001010100000000000000111
00000011100110011111110110101101
00000010100000000001110000001110
01101100000000000111100110100100
00010100000010111110101111001101
00000000001000011011001000101111
01011100000000000000100110000000
00000000001010100000000000000111
00000010100000001100000111101111
00000011100001000000000110110000
00101001000000000010110001101111
00000000001000010100000010001111
01011100000000000000101000000000
00000000001010100000000000000111
00000011100000111111110110101101
00000010100000000010000000001110
01101100000000000100010110100100
00000000001000011011000111101110
01011100000000000000100110000000
00000000001010100000000000000111
00010100011101110011010110001101
00000010100000001100000111001100
00000011101010000000000110101111
00101001000000000011000001101100
00000000001000010011110010001100
01011100000000000000100111100000
00000000001010100000000000000111
00000011101001111111110110101101
00000010100000000010010000001110
01101100000000000001000110100100
00000010100000001100000110001100
00000010100000000010100000001110
00101001000000000011010001101100
00000010100000000001000001110001
00010101011111100000000000010000
00000000000100000011101000101110
00000011100000000001011000001101
00000011010000000000000000000000
00101010001111111111110111001111
00101010000000000000000110101100
00000011010000001000000110001100
01011011111111111111100110000000
00101001000000000000001000001111
00000010101111111111110111001110
01011111111111111110101000101110
00000010100000000100000001100011
01001100000000000000000000100000
00010101011111100000000000001110
00000011100000000001010111001101
00000011010000000000000000000000
00101010000000000000000110101100
00000011010000001000000110001100
01011011111111111111100110000000
00000010100000001100000000001100
00101001000000000000000111001100
01001100000000000000000000100000
00000011010000000000000000000000
00000011010000000000000000000000
00000000000000000110000000000100
01001100000000000000000000100000
00000011010000000000000000000000
00000011010000000000000000000000
00000000000000000000000000000000
00000000000000000000000000000000
00000000000000000000000000000000
@@ -908,18 +712,10 @@
00000000000000000000000000111101
01000001010011010000101000001101
01001100010101010100110101010100
01000011010110010100001101011111
00111101010100110100010101001100
00000000000000000000000000000000
01000001010011010000101000001101
01001100010101010100110101010100
01001101010010010101010001011111
01010011010101010101111101000101
00000000000000000000000000111101
01000001010011010000101000001101
01001100010101010100110101010100
01001110010011110100010001011111
00000000000010100000110101000101
00000001111101111000101001000000
10111111000000000000000000000000
10111111000000000000000000000000
00000000000000000000000000000000
00000000000000000000000000000000
00000000000000000000000000000000
00000000000000000000000000000000

31
sdk/software/examples/asm/user-sample.c
View File

@@ -5,7 +5,6 @@
*/
#include <dma.h>
#include <confreg_time.h>
/* ─── UART 16550 direct access (bypasses slow printf/picolibc) ──────── */
#define UART_THR (*(volatile unsigned char *)0xbf000000)
@@ -29,24 +28,6 @@ static void uart_puthex(unsigned int val) {
}
}
static void uart_putdec(unsigned int val) {
char buf[10];
int i = 0;
if (val == 0) {
uart_tx('0');
return;
}
while (val != 0) {
buf[i++] = (char)('0' + (val % 10));
val /= 10;
}
while (i > 0) {
uart_tx((unsigned char)buf[--i]);
}
}
/* ─── Configurable group count (default 5000) ──────────────────────── */
#ifndef MATMUL_GROUP_NUM
#define MATMUL_GROUP_NUM 5000
@@ -54,9 +35,6 @@ static void uart_putdec(unsigned int val) {
/* ─── BSP globals ─────────────────────────────────────────────────── */
unsigned long UART_BASE = 0xbf000000;
unsigned long CONFREG_TIMER_BASE = 0xbf20f100;
unsigned long CONFREG_CLOCKS_PER_SEC = 50000000L;
unsigned long CORE_CLOCKS_PER_SEC = 33000000L;
/* ─── Memory layout ────────────────────────────────────────────────── */
#define EXTRAM_BASE_PHYS 0x1c400000
@@ -107,11 +85,9 @@ int main(void)
volatile unsigned int *mat_crc = (volatile unsigned int *)MATMUL_CRC;
unsigned int crc;
unsigned int cycle_start, cycle_end, cycles, time_us;
/* ── Start marker (must be first output) ──────────────────── */
uart_puts("MATMUL_START\r\n");
cycle_start = get_clock_count();
*(volatile unsigned int *)MATMUL_CRC = 0;
*mat_status = STATUS_DMA_DONE | STATUS_DONE | STATUS_ERROR;
@@ -123,19 +99,12 @@ int main(void)
*mat_ctrl = CTRL_DMA_START;
while (!(*mat_status & STATUS_DMA_DONE)) {}
*mat_status = STATUS_DMA_DONE;
cycle_end = get_clock_count();
cycles = cycle_end - cycle_start;
time_us = cycles / (CORE_CLOCKS_PER_SEC / 1000000UL);
/* ── Read hardware CRC32 ──────────────────────────────────── */
crc = *mat_crc;
uart_puts("MATMUL_CRC32=");
uart_puthex(crc);
uart_puts("\r\nMATMUL_CYCLES=");
uart_putdec(cycles);
uart_puts("\r\nMATMUL_TIME_US=");
uart_putdec(time_us);
uart_puts("\r\nMATMUL_DONE\r\n");
return 0;