scope fixes
This commit is contained in:
@@ -17,7 +17,7 @@ CXXFLAGS +=-fstack-protector
|
||||
# Position independent code
|
||||
CXXFLAGS += -fPIC
|
||||
|
||||
CXXFLAGS += -DGLOBAL_BLOCK_SIZE_BYTES=64
|
||||
CXXFLAGS += -DSCOPE
|
||||
|
||||
LDFLAGS += -luuid
|
||||
|
||||
|
||||
@@ -4,6 +4,12 @@
|
||||
#include <unistd.h>
|
||||
#include <assert.h>
|
||||
#include <cmath>
|
||||
#include <thread>
|
||||
#include <future>
|
||||
#include <chrono>
|
||||
#include <fstream>
|
||||
#include <iostream>
|
||||
#include <vector>
|
||||
#include <uuid/uuid.h>
|
||||
#include <opae/fpga.h>
|
||||
#include <vortex.h>
|
||||
@@ -21,16 +27,18 @@
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#define CMD_TYPE_READ AFU_IMAGE_CMD_TYPE_READ
|
||||
#define CMD_TYPE_WRITE AFU_IMAGE_CMD_TYPE_WRITE
|
||||
#define CMD_TYPE_RUN AFU_IMAGE_CMD_TYPE_RUN
|
||||
#define CMD_TYPE_CLFLUSH AFU_IMAGE_CMD_TYPE_CLFLUSH
|
||||
#define CMD_TYPE_READ AFU_IMAGE_CMD_TYPE_READ
|
||||
#define CMD_TYPE_WRITE AFU_IMAGE_CMD_TYPE_WRITE
|
||||
#define CMD_TYPE_RUN AFU_IMAGE_CMD_TYPE_RUN
|
||||
#define CMD_TYPE_CLFLUSH AFU_IMAGE_CMD_TYPE_CLFLUSH
|
||||
|
||||
#define MMIO_CSR_CMD (AFU_IMAGE_MMIO_CSR_CMD * 4)
|
||||
#define MMIO_CSR_STATUS (AFU_IMAGE_MMIO_CSR_STATUS * 4)
|
||||
#define MMIO_CSR_IO_ADDR (AFU_IMAGE_MMIO_CSR_IO_ADDR * 4)
|
||||
#define MMIO_CSR_MEM_ADDR (AFU_IMAGE_MMIO_CSR_MEM_ADDR * 4)
|
||||
#define MMIO_CSR_DATA_SIZE (AFU_IMAGE_MMIO_CSR_DATA_SIZE * 4)
|
||||
#define MMIO_CSR_CMD (AFU_IMAGE_MMIO_CSR_CMD * 4)
|
||||
#define MMIO_CSR_IO_ADDR (AFU_IMAGE_MMIO_CSR_IO_ADDR * 4)
|
||||
#define MMIO_CSR_MEM_ADDR (AFU_IMAGE_MMIO_CSR_MEM_ADDR * 4)
|
||||
#define MMIO_CSR_DATA_SIZE (AFU_IMAGE_MMIO_CSR_DATA_SIZE * 4)
|
||||
#define MMIO_CSR_STATUS (AFU_IMAGE_MMIO_CSR_STATUS * 4)
|
||||
#define MMIO_CSR_SCOPE_CMD (AFU_IMAGE_MMIO_CSR_SCOPE_CMD * 4)
|
||||
#define MMIO_CSR_SCOPE_DATA (AFU_IMAGE_MMIO_CSR_SCOPE_DATA * 4)
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
@@ -57,6 +65,10 @@ inline bool is_aligned(size_t addr, size_t alignment) {
|
||||
return 0 == (addr & (alignment - 1));
|
||||
}
|
||||
|
||||
#ifdef SCOPE
|
||||
std::future<void> future_scope;
|
||||
#endif
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
extern int vx_dev_open(vx_device_h* hdevice) {
|
||||
@@ -122,6 +134,10 @@ extern int vx_dev_close(vx_device_h hdevice) {
|
||||
if (nullptr == hdevice)
|
||||
return -1;
|
||||
|
||||
#ifdef SCOPE
|
||||
future_scope.wait();
|
||||
#endif
|
||||
|
||||
vx_device_t *device = ((vx_device_t*)hdevice);
|
||||
|
||||
fpgaClose(device->fpga);
|
||||
@@ -377,6 +393,171 @@ extern int vx_flush_caches(vx_device_h hdevice, size_t dev_maddr, size_t size) {
|
||||
return 0;
|
||||
}
|
||||
|
||||
static int vx_scope_trace(vx_device_h hdevice) {
|
||||
if (nullptr == hdevice)
|
||||
return -1;
|
||||
|
||||
vx_device_t *device = ((vx_device_t*)hdevice);
|
||||
|
||||
std::ofstream ofs("vx_scope.vcd");
|
||||
|
||||
ofs << "$timescale 1 ns $end" << std::endl;
|
||||
|
||||
int fwidth = 0;
|
||||
|
||||
ofs << "$var reg 1 0 clk $end" << std::endl;
|
||||
|
||||
fwidth += 1;
|
||||
|
||||
ofs << "$var reg 1 1 icache_req_valid $end" << std::endl;
|
||||
ofs << "$var reg 2 2 icache_req_tag $end" << std::endl;
|
||||
ofs << "$var reg 1 3 icache_req_ready $end" << std::endl;
|
||||
ofs << "$var reg 1 4 icache_rsp_valid $end" << std::endl;
|
||||
ofs << "$var reg 2 5 icache_rsp_tag $end" << std::endl;
|
||||
ofs << "$var reg 1 6 icache_rsp_ready $end" << std::endl;
|
||||
|
||||
fwidth += 8;
|
||||
|
||||
ofs << "$var reg 4 7 dcache_req_valid $end" << std::endl;
|
||||
ofs << "$var reg 2 8 dcache_req_tag $end" << std::endl;
|
||||
ofs << "$var reg 1 9 dcache_req_ready $end" << std::endl;
|
||||
ofs << "$var reg 4 10 dcache_rsp_valid $end" << std::endl;
|
||||
ofs << "$var reg 2 11 dcache_rsp_tag $end" << std::endl;
|
||||
ofs << "$var reg 1 12 dcache_rsp_ready $end" << std::endl;
|
||||
|
||||
fwidth += 14;
|
||||
|
||||
ofs << "$var reg 1 13 dram_req_valid $end" << std::endl;
|
||||
ofs << "$var reg 29 14 dram_req_tag $end" << std::endl;
|
||||
ofs << "$var reg 1 15 dram_req_ready $end" << std::endl;
|
||||
ofs << "$var reg 1 16 dram_rsp_valid $end" << std::endl;
|
||||
ofs << "$var reg 29 17 dram_rsp_tag $end" << std::endl;
|
||||
ofs << "$var reg 1 18 dram_rsp_ready $end" << std::endl;
|
||||
|
||||
fwidth += 62;
|
||||
|
||||
ofs << "$var reg 1 19 schedule_delay $end" << std::endl;
|
||||
|
||||
fwidth += 1;
|
||||
|
||||
const int num_signals = 20;
|
||||
|
||||
ofs << "enddefinitions $end" << std::endl;
|
||||
|
||||
uint64_t frame_width, max_frames, data_valid;
|
||||
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_SCOPE_CMD, 2));
|
||||
CHECK_RES(fpgaReadMMIO64(device->fpga, 0, MMIO_CSR_SCOPE_DATA, &frame_width));
|
||||
std::cout << "scope::frame_width=" << frame_width << std::endl;
|
||||
|
||||
assert((fwidth-1)== (int)frame_width);
|
||||
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_SCOPE_CMD, 3));
|
||||
CHECK_RES(fpgaReadMMIO64(device->fpga, 0, MMIO_CSR_SCOPE_DATA, &max_frames));
|
||||
std::cout << "scope::max_frames=" << max_frames << std::endl;
|
||||
|
||||
do {
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_SCOPE_CMD, 0));
|
||||
CHECK_RES(fpgaReadMMIO64(device->fpga, 0, MMIO_CSR_SCOPE_DATA, &data_valid));
|
||||
if (data_valid)
|
||||
break;
|
||||
std::this_thread::sleep_for(std::chrono::milliseconds(1));
|
||||
} while (true);
|
||||
|
||||
std::cout << "scope trace dump begin..." << std::endl;
|
||||
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_SCOPE_CMD, 1));
|
||||
|
||||
std::vector<char> signa_data(frame_width+1);
|
||||
|
||||
uint64_t frame_offset = 0, frame_no = 0;
|
||||
|
||||
int signal_id = 0;
|
||||
int signal_offset = 0;
|
||||
|
||||
auto print_signal = [&] (uint64_t word, int signal_width) {
|
||||
|
||||
int word_offset = frame_offset % 64;
|
||||
|
||||
signa_data[signal_width - signal_offset - 1] = ((word >> word_offset) & 0x1) ? '1' : '0';
|
||||
|
||||
++signal_offset;
|
||||
++frame_offset;
|
||||
|
||||
if (signal_offset == signal_width) {
|
||||
signa_data[signal_width] = 0; // string null termination
|
||||
ofs << 'b' << signa_data.data() << ' ' << (num_signals - signal_id) << std::endl;
|
||||
signal_offset = 0;
|
||||
++signal_id;
|
||||
}
|
||||
|
||||
if (frame_offset == frame_width) {
|
||||
assert(0 == signal_offset);
|
||||
signal_id = 0;
|
||||
frame_offset = 0;
|
||||
++frame_no;
|
||||
if (frame_no != max_frames) {
|
||||
ofs << '#' << (frame_no * 2 + 0) << std::endl;
|
||||
ofs << "b0 0" << std::endl;
|
||||
ofs << '#' << (frame_no * 2 + 1) << std::endl;
|
||||
ofs << "b1 0" << std::endl;
|
||||
++signal_id;
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
ofs << '#' << (frame_no * 2 + 0) << std::endl;
|
||||
ofs << "b0 0" << std::endl;
|
||||
ofs << '#' << (frame_no * 2 + 1) << std::endl;
|
||||
ofs << "b1 0" << std::endl;
|
||||
++signal_id;
|
||||
|
||||
do {
|
||||
if (frame_no == max_frames-1) {
|
||||
// verify last frame is valid
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_SCOPE_CMD, 0));
|
||||
CHECK_RES(fpgaReadMMIO64(device->fpga, 0, MMIO_CSR_SCOPE_DATA, &data_valid));
|
||||
assert(data_valid == 1);
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_SCOPE_CMD, 1));
|
||||
}
|
||||
|
||||
uint64_t word;
|
||||
CHECK_RES(fpgaReadMMIO64(device->fpga, 0, MMIO_CSR_SCOPE_DATA, &word));
|
||||
|
||||
do {
|
||||
switch (num_signals - signal_id) {
|
||||
case 2:
|
||||
case 5:
|
||||
case 8:
|
||||
case 11:
|
||||
print_signal(word, 2);
|
||||
break;
|
||||
case 7:
|
||||
case 10:
|
||||
print_signal(word, 4);
|
||||
break;
|
||||
case 14:
|
||||
case 17:
|
||||
print_signal(word, 29);
|
||||
break;
|
||||
default:
|
||||
print_signal(word, 1);
|
||||
break;
|
||||
}
|
||||
} while ((frame_offset % 64) != 0);
|
||||
|
||||
} while (frame_no != max_frames);
|
||||
|
||||
std::cout << "scope trace dump done!" << std::endl;
|
||||
|
||||
// verify data not valid
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_SCOPE_CMD, 0));
|
||||
CHECK_RES(fpgaReadMMIO64(device->fpga, 0, MMIO_CSR_SCOPE_DATA, &data_valid));
|
||||
assert(data_valid == 0);
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
extern int vx_start(vx_device_h hdevice) {
|
||||
if (nullptr == hdevice)
|
||||
return -1;
|
||||
@@ -387,6 +568,12 @@ extern int vx_start(vx_device_h hdevice) {
|
||||
if (vx_ready_wait(hdevice, -1) != 0)
|
||||
return -1;
|
||||
|
||||
#ifdef SCOPE
|
||||
// launch logic scope
|
||||
future_scope = std::async(std::launch::async, [&]{ vx_scope_trace(hdevice); });
|
||||
#endif
|
||||
|
||||
// start execution
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_CMD, CMD_TYPE_RUN));
|
||||
|
||||
return 0;
|
||||
|
||||
@@ -175,11 +175,11 @@ int main(int argc, char *argv[]) {
|
||||
RT_CHECK(vx_alloc_shared_mem(device, 4096, &dbuf));
|
||||
|
||||
// run tests
|
||||
if (0 == test || -1 == test) {
|
||||
/*if (0 == test || -1 == test) {
|
||||
std::cout << "run memcopy test" << std::endl;
|
||||
RT_CHECK(run_memcopy_test(sbuf, dbuf, DEV_MEM_SRC_ADDR, 0x0badf00d00ff00ff, 1));
|
||||
RT_CHECK(run_memcopy_test(sbuf, dbuf, DEV_MEM_SRC_ADDR, 0x0badf00d40ff40ff, 64));
|
||||
}
|
||||
}*/
|
||||
|
||||
if (1 == test || -1 == test) {
|
||||
std::cout << "run kernel test" << std::endl;
|
||||
|
||||
@@ -1,12 +1,16 @@
|
||||
vortex_afu.json
|
||||
|
||||
#+define+DBG_PRINT_CORE_ICACHE
|
||||
#+define+DBG_PRINT_CORE_DCACHE
|
||||
#+define+DBG_PRINT_CACHE_BANK
|
||||
#+define+DBG_PRINT_CACHE_SNP
|
||||
#+define+DBG_PRINT_CACHE_MSRQ
|
||||
#+define+DBG_PRINT_DRAM
|
||||
#+define+DBG_PRINT_OPAE
|
||||
+define+NDEBUG
|
||||
|
||||
+define+DBG_PRINT_CORE_ICACHE
|
||||
+define+DBG_PRINT_CORE_DCACHE
|
||||
+define+DBG_PRINT_CACHE_BANK
|
||||
+define+DBG_PRINT_CACHE_SNP
|
||||
+define+DBG_PRINT_CACHE_MSRQ
|
||||
+define+DBG_PRINT_DRAM
|
||||
+define+DBG_PRINT_OPAE
|
||||
|
||||
+define+SCOPE
|
||||
|
||||
+incdir+.
|
||||
+incdir+../rtl
|
||||
@@ -65,10 +69,12 @@ vortex_afu.json
|
||||
../rtl/libs/VX_generic_queue.v
|
||||
../rtl/libs/VX_indexable_queue.v
|
||||
../rtl/libs/VX_countones.v
|
||||
../rtl/libs/VX_scope.v
|
||||
|
||||
../rtl/Vortex_Socket.v
|
||||
../rtl/Vortex_Cluster.v
|
||||
../rtl/Vortex.v
|
||||
../rtl/VX_pipeline.v
|
||||
../rtl/VX_front_end.v
|
||||
../rtl/VX_back_end.v
|
||||
../rtl/VX_fetch.v
|
||||
|
||||
@@ -10,7 +10,7 @@
|
||||
"mmio-csr-mem-addr": 14,
|
||||
"mmio-csr-data-size": 16,
|
||||
"mmio-csr-status": 18,
|
||||
"mmio-csr-scope-delay": 20,
|
||||
"mmio-csr-scope-cmd": 20,
|
||||
"mmio-csr-scope-data": 22,
|
||||
|
||||
"cmd-type-read": 1,
|
||||
|
||||
@@ -66,8 +66,8 @@ localparam MMIO_CSR_MEM_ADDR = `AFU_IMAGE_MMIO_CSR_MEM_ADDR;
|
||||
localparam MMIO_CSR_DATA_SIZE = `AFU_IMAGE_MMIO_CSR_DATA_SIZE;
|
||||
localparam MMIO_CSR_STATUS = `AFU_IMAGE_MMIO_CSR_STATUS;
|
||||
|
||||
localparam MMIO_CSR_SCOPE_DELAY = `AFU_IMAGE_MMIO_CSR_SCOPE_DELAY;
|
||||
localparam MMIO_CSR_SCOPE_DATA = `AFU_IMAGE_MMIO_CSR_SCOPE_DATA;
|
||||
localparam MMIO_CSR_SCOPE_CMD = `AFU_IMAGE_MMIO_CSR_SCOPE_CMD;
|
||||
localparam MMIO_CSR_SCOPE_DATA= `AFU_IMAGE_MMIO_CSR_SCOPE_DATA;
|
||||
|
||||
logic [127:0] afu_id = `AFU_ACCEL_UUID;
|
||||
|
||||
@@ -138,7 +138,7 @@ t_ccip_clAddr csr_io_addr;
|
||||
logic[DRAM_ADDR_WIDTH-1:0] csr_mem_addr;
|
||||
logic[DRAM_ADDR_WIDTH-1:0] csr_data_size;
|
||||
|
||||
logic [63:0] csr_scope_delay;
|
||||
logic [63:0] csr_scope_cmd;
|
||||
logic [63:0] csr_scope_data;
|
||||
logic csr_scope_read;
|
||||
logic csr_scope_write;
|
||||
@@ -153,8 +153,8 @@ assign mmio_hdr = t_ccip_c0_ReqMmioHdr'(cp2af_sRxPort.c0.hdr);
|
||||
t_if_ccip_c2_Tx mmio_tx;
|
||||
assign af2cp_sTxPort.c2 = mmio_tx;
|
||||
|
||||
assign csr_scope_delay = 64'(cp2af_sRxPort.c0.data);
|
||||
assign csr_scope_write = cp2af_sRxPort.c0.mmioWrValid && (MMIO_CSR_SCOPE_DELAY == mmio_hdr.address);
|
||||
assign csr_scope_cmd = 64'(cp2af_sRxPort.c0.data);
|
||||
assign csr_scope_write = cp2af_sRxPort.c0.mmioWrValid && (MMIO_CSR_SCOPE_CMD == mmio_hdr.address);
|
||||
assign csr_scope_read = cp2af_sRxPort.c0.mmioRdValid && (MMIO_CSR_SCOPE_DATA == mmio_hdr.address);
|
||||
|
||||
always_ff @(posedge clk)
|
||||
@@ -201,6 +201,11 @@ begin
|
||||
$display("%t: CSR_CMD: %0d", $time, $bits(csr_cmd)'(cp2af_sRxPort.c0.data));
|
||||
`endif
|
||||
end
|
||||
MMIO_CSR_SCOPE_CMD: begin
|
||||
`ifdef DBG_PRINT_OPAE
|
||||
$display("%t: CSR_SCOPE_CMD: %0d", $time, 64'(cp2af_sRxPort.c0.data));
|
||||
`endif
|
||||
end
|
||||
default: begin
|
||||
// user-defined CSRs
|
||||
//if (mmio_hdr.addres >= MMIO_CSR_USER) begin
|
||||
@@ -238,9 +243,9 @@ begin
|
||||
mmio_tx.data <= 64'(state);
|
||||
end
|
||||
MMIO_CSR_SCOPE_DATA: begin
|
||||
mmio_tx.data <= csr_scope_data;
|
||||
mmio_tx.data <= csr_scope_data;
|
||||
`ifdef DBG_PRINT_OPAE
|
||||
$display("%t: scope: data=%0d", $time, csr_scope_data);
|
||||
$display("%t: SCOPE: data=%0d", $time, csr_scope_data);
|
||||
`endif
|
||||
end
|
||||
default: mmio_tx.data <= 64'h0;
|
||||
@@ -790,18 +795,20 @@ end
|
||||
|
||||
`ifdef SCOPE
|
||||
|
||||
`SCOPE_SIGNALS_DECL()
|
||||
`SCOPE_SIGNALS_DECL
|
||||
|
||||
`STATIC_ASSERT($bits({`SCOPE_SIGNALS_LIST}) == 85, "oops!")
|
||||
|
||||
VX_scope #(
|
||||
.DATAW ($bits({`SCOPE_SIGNALS_LIST()})),
|
||||
.DATAW ($bits({`SCOPE_SIGNALS_LIST})),
|
||||
.BUSW (64),
|
||||
.SIZE (1024)
|
||||
.SIZE (256)
|
||||
) scope (
|
||||
.clk (clk),
|
||||
.reset (SoftReset),
|
||||
.start (vx_reset),
|
||||
.data_in ({`SCOPE_SIGNALS_LIST()}),
|
||||
.bus_in (csr_scope_delay),
|
||||
.data_in ({`SCOPE_SIGNALS_LIST}),
|
||||
.bus_in (csr_scope_cmd),
|
||||
.bus_out (csr_scope_data),
|
||||
.bus_read (csr_scope_read),
|
||||
.bus_write(csr_scope_write)
|
||||
@@ -814,7 +821,7 @@ VX_scope #(
|
||||
assign cmd_run_done = !vx_busy;
|
||||
|
||||
Vortex_Socket #() vx_socket (
|
||||
`SCOPE_SIGNALS_ATTACH(),
|
||||
`SCOPE_SIGNALS_ATTACH
|
||||
|
||||
.clk (clk),
|
||||
.reset (vx_reset),
|
||||
|
||||
29
hw/opae/vortex_afu.vh
Normal file
29
hw/opae/vortex_afu.vh
Normal file
@@ -0,0 +1,29 @@
|
||||
`ifndef __VORTEX_AFU__
|
||||
`define __VORTEX_AFU__
|
||||
|
||||
`include "ccip_if_pkg.sv"
|
||||
|
||||
`define PLATFORM_PROVIDES_LOCAL_MEMORY
|
||||
`define PLATFORM_PARAM_LOCAL_MEMORY_ADDR_WIDTH 27
|
||||
`define PLATFORM_PARAM_LOCAL_MEMORY_DATA_WIDTH 512
|
||||
`define PLATFORM_PARAM_LOCAL_MEMORY_BURST_CNT_WIDTH 4
|
||||
|
||||
`include "local_mem_cfg_pkg.sv"
|
||||
|
||||
`define AFU_ACCEL_NAME "vortex_afu"
|
||||
`define AFU_ACCEL_UUID 128'h35f9452b_25c2_434c_93d5_6f8c60db361c
|
||||
`define AFU_IMAGE_CMD_TYPE_CLFLUSH 4
|
||||
`define AFU_IMAGE_CMD_TYPE_READ 1
|
||||
`define AFU_IMAGE_CMD_TYPE_RUN 3
|
||||
`define AFU_IMAGE_CMD_TYPE_WRITE 2
|
||||
`define AFU_IMAGE_MMIO_CSR_CMD 10
|
||||
`define AFU_IMAGE_MMIO_CSR_DATA_SIZE 12
|
||||
`define AFU_IMAGE_MMIO_CSR_IO_ADDR 14
|
||||
`define AFU_IMAGE_MMIO_CSR_MEM_ADDR 16
|
||||
`define AFU_IMAGE_MMIO_CSR_STATUS 18
|
||||
`define AFU_IMAGE_MMIO_CSR_SCOPE_CMD 20
|
||||
`define AFU_IMAGE_MMIO_CSR_SCOPE_DATA 22
|
||||
`define AFU_IMAGE_POWER 0
|
||||
`define AFU_TOP_IFC "ccip_std_afu_avalon_mm"
|
||||
|
||||
`endif
|
||||
149
hw/opae/vortex_afu_sim.v
Normal file
149
hw/opae/vortex_afu_sim.v
Normal file
@@ -0,0 +1,149 @@
|
||||
`include "vortex_afu.vh"
|
||||
|
||||
/* verilator lint_off IMPORTSTAR */
|
||||
import ccip_if_pkg::*;
|
||||
import local_mem_cfg_pkg::*;
|
||||
/* verilator lint_on IMPORTSTAR */
|
||||
|
||||
module vortex_afu_sim #(
|
||||
parameter NUM_LOCAL_MEM_BANKS = 2
|
||||
) (
|
||||
// global signals
|
||||
input clk,
|
||||
input reset,
|
||||
|
||||
// IF signals between CCI and AFU
|
||||
input logic vcp2af_sRxPort_c0_TxAlmFull,
|
||||
input logic vcp2af_sRxPort_c1_TxAlmFull,
|
||||
|
||||
input t_ccip_vc vcp2af_sRxPort_c0_hdr_vc_used,
|
||||
input logic vcp2af_sRxPort_c0_hdr_rsvd1,
|
||||
input logic vcp2af_sRxPort_c0_hdr_hit_miss,
|
||||
input logic [1:0] vcp2af_sRxPort_c0_hdr_rsvd0,
|
||||
input t_ccip_clNum vcp2af_sRxPort_c0_hdr_cl_num,
|
||||
input t_ccip_c0_rsp vcp2af_sRxPort_c0_hdr_resp_type,
|
||||
input t_ccip_mdata vcp2af_sRxPort_c0_hdr_mdata,
|
||||
input t_ccip_clData vcp2af_sRxPort_c0_data,
|
||||
input logic vcp2af_sRxPort_c0_rspValid,
|
||||
input logic vcp2af_sRxPort_c0_mmioRdValid,
|
||||
input logic vcp2af_sRxPort_c0_mmioWrValid,
|
||||
|
||||
input t_ccip_vc vcp2af_sRxPort_c1_hdr_vc_used,
|
||||
input logic vcp2af_sRxPort_c1_hdr_rsvd1,
|
||||
input logic vcp2af_sRxPort_c1_hdr_hit_miss,
|
||||
input logic vcp2af_sRxPort_c1_hdr_format,
|
||||
input logic vcp2af_sRxPort_c1_hdr_rsvd0,
|
||||
input t_ccip_clNum vcp2af_sRxPort_c1_hdr_cl_num,
|
||||
input t_ccip_c1_rsp vcp2af_sRxPort_c1_hdr_resp_type,
|
||||
input t_ccip_mdata vcp2af_sRxPort_c1_hdr_mdata,
|
||||
input logic vcp2af_sRxPort_c1_rspValid,
|
||||
|
||||
output t_ccip_vc af2cp_sTxPort_c0_hdr_vc_sel,
|
||||
output logic [1:0] af2cp_sTxPort_c0_hdr_rsvd1,
|
||||
output t_ccip_clLen af2cp_sTxPort_c0_hdr_cl_len,
|
||||
output t_ccip_c0_req af2cp_sTxPort_c0_hdr_req_type,
|
||||
output logic [5:0] af2cp_sTxPort_c0_hdr_rsvd0,
|
||||
output t_ccip_clAddr af2cp_sTxPort_c0_hdr_address,
|
||||
output t_ccip_mdata af2cp_sTxPort_c0_hdr_mdata,
|
||||
output logic af2cp_sTxPort_c0_valid,
|
||||
|
||||
output logic [5:0] af2cp_sTxPort_c1_hdr_rsvd2,
|
||||
output t_ccip_vc af2cp_sTxPort_c1_hdr_vc_sel,
|
||||
output logic af2cp_sTxPort_c1_hdr_sop,
|
||||
output logic af2cp_sTxPort_c1_hdr_rsvd1,
|
||||
output t_ccip_clLen af2cp_sTxPort_c1_hdr_cl_len,
|
||||
output t_ccip_c1_req af2cp_sTxPort_c1_hdr_req_type,
|
||||
output logic [5:0] af2cp_sTxPort_c1_hdr_rsvd0,
|
||||
output t_ccip_clAddr af2cp_sTxPort_c1_hdr_address,
|
||||
output t_ccip_mdata af2cp_sTxPort_c1_hdr_mdata,
|
||||
output t_ccip_clData af2cp_sTxPort_c1_data,
|
||||
output logic af2cp_sTxPort_c1_valid,
|
||||
|
||||
output t_ccip_tid af2cp_sTxPort_c2_hdr_tid,
|
||||
output logic af2cp_sTxPort_c2_mmioRdValid,
|
||||
output t_ccip_mmioData af2cp_sTxPort_c2_data,
|
||||
|
||||
// Avalon signals for local memory access
|
||||
output t_local_mem_data avs_writedata,
|
||||
input t_local_mem_data avs_readdata,
|
||||
output t_local_mem_addr avs_address,
|
||||
input logic avs_waitrequest,
|
||||
output logic avs_write,
|
||||
output logic avs_read,
|
||||
output t_local_mem_byte_mask avs_byteenable,
|
||||
output t_local_mem_burst_cnt avs_burstcount,
|
||||
input avs_readdatavalid,
|
||||
|
||||
output logic [$clog2(NUM_LOCAL_MEM_BANKS)-1:0] mem_bank_select
|
||||
);
|
||||
|
||||
vortex_afu #(
|
||||
.NUM_LOCAL_MEM_BANKS(NUM_LOCAL_MEM_BANKS)
|
||||
) vortex_afu (
|
||||
.clk(clk),
|
||||
.SoftReset(reset),
|
||||
.cp2af_sRxPort({
|
||||
vcp2af_sRxPort_c0_TxAlmFull,
|
||||
vcp2af_sRxPort_c1_TxAlmFull,
|
||||
|
||||
vcp2af_sRxPort_c0_hdr_vc_used,
|
||||
vcp2af_sRxPort_c0_hdr_rsvd1,
|
||||
vcp2af_sRxPort_c0_hdr_hit_miss,
|
||||
vcp2af_sRxPort_c0_hdr_rsvd0,
|
||||
vcp2af_sRxPort_c0_hdr_cl_num,
|
||||
vcp2af_sRxPort_c0_hdr_resp_type,
|
||||
vcp2af_sRxPort_c0_hdr_mdata,
|
||||
vcp2af_sRxPort_c0_data,
|
||||
vcp2af_sRxPort_c0_rspValid,
|
||||
vcp2af_sRxPort_c0_mmioRdValid,
|
||||
vcp2af_sRxPort_c0_mmioWrValid,
|
||||
|
||||
vcp2af_sRxPort_c1_hdr_vc_used,
|
||||
vcp2af_sRxPort_c1_hdr_rsvd1,
|
||||
vcp2af_sRxPort_c1_hdr_hit_miss,
|
||||
vcp2af_sRxPort_c1_hdr_format,
|
||||
vcp2af_sRxPort_c1_hdr_rsvd0,
|
||||
vcp2af_sRxPort_c1_hdr_cl_num,
|
||||
vcp2af_sRxPort_c1_hdr_resp_type,
|
||||
vcp2af_sRxPort_c1_hdr_mdata,
|
||||
vcp2af_sRxPort_c1_rspValid}
|
||||
),
|
||||
.af2cp_sTxPort({
|
||||
af2cp_sTxPort_c0_hdr_vc_sel,
|
||||
af2cp_sTxPort_c0_hdr_rsvd1,
|
||||
af2cp_sTxPort_c0_hdr_cl_len,
|
||||
af2cp_sTxPort_c0_hdr_req_type,
|
||||
af2cp_sTxPort_c0_hdr_rsvd0,
|
||||
af2cp_sTxPort_c0_hdr_address,
|
||||
af2cp_sTxPort_c0_hdr_mdata,
|
||||
af2cp_sTxPort_c0_valid,
|
||||
|
||||
af2cp_sTxPort_c1_hdr_rsvd2,
|
||||
af2cp_sTxPort_c1_hdr_vc_sel,
|
||||
af2cp_sTxPort_c1_hdr_sop,
|
||||
af2cp_sTxPort_c1_hdr_rsvd1,
|
||||
af2cp_sTxPort_c1_hdr_cl_len,
|
||||
af2cp_sTxPort_c1_hdr_req_type,
|
||||
af2cp_sTxPort_c1_hdr_rsvd0,
|
||||
af2cp_sTxPort_c1_hdr_address,
|
||||
af2cp_sTxPort_c1_hdr_mdata,
|
||||
af2cp_sTxPort_c1_data,
|
||||
af2cp_sTxPort_c1_valid,
|
||||
|
||||
af2cp_sTxPort_c2_hdr_tid,
|
||||
af2cp_sTxPort_c2_mmioRdValid,
|
||||
af2cp_sTxPort_c2_data
|
||||
}),
|
||||
.avs_writedata(avs_writedata),
|
||||
.avs_readdata(avs_readdata),
|
||||
.avs_address(avs_address),
|
||||
.avs_waitrequest(avs_waitrequest),
|
||||
.avs_write(avs_write),
|
||||
.avs_read(avs_read),
|
||||
.avs_byteenable(avs_byteenable),
|
||||
.avs_burstcount(avs_burstcount),
|
||||
.avs_readdatavalid(avs_readdatavalid),
|
||||
.mem_bank_select(mem_bank_select)
|
||||
);
|
||||
|
||||
endmodule
|
||||
@@ -283,7 +283,7 @@
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
`ifdef SCOPE
|
||||
`define SCOPE_SIGNALS_LIST() \
|
||||
`define SCOPE_SIGNALS_LIST \
|
||||
scope_icache_req_valid, \
|
||||
scope_icache_req_tag, \
|
||||
scope_icache_req_ready, \
|
||||
@@ -304,7 +304,7 @@
|
||||
scope_dram_rsp_ready, \
|
||||
scope_schedule_delay
|
||||
|
||||
`define SCOPE_SIGNALS_DECL() \
|
||||
`define SCOPE_SIGNALS_DECL \
|
||||
wire scope_icache_req_valid; \
|
||||
wire [`DCORE_TAG_WIDTH-1:0] scope_icache_req_tag; \
|
||||
wire scope_icache_req_ready; \
|
||||
@@ -325,7 +325,7 @@
|
||||
wire scope_dram_rsp_ready; \
|
||||
wire scope_schedule_delay;
|
||||
|
||||
`define SCOPE_SIGNALS_IO() \
|
||||
`define SCOPE_SIGNALS_IO \
|
||||
/* verilator lint_off UNDRIVEN */ \
|
||||
output wire scope_icache_req_valid, \
|
||||
output wire [`DCORE_TAG_WIDTH-1:0] scope_icache_req_tag, \
|
||||
@@ -345,10 +345,10 @@
|
||||
output wire scope_dram_rsp_valid, \
|
||||
output wire [`VX_DRAM_TAG_WIDTH-1:0] scope_dram_rsp_tag, \
|
||||
output wire scope_dram_rsp_ready, \
|
||||
output wire scope_schedule_delay \
|
||||
output wire scope_schedule_delay, \
|
||||
/* verilator lint_on UNDRIVEN */
|
||||
|
||||
`define SCOPE_SIGNALS_ATTACH() \
|
||||
`define SCOPE_SIGNALS_ATTACH \
|
||||
.scope_icache_req_valid (scope_icache_req_valid), \
|
||||
.scope_icache_req_tag (scope_icache_req_tag), \
|
||||
.scope_icache_req_ready (scope_icache_req_ready), \
|
||||
@@ -361,18 +361,18 @@
|
||||
.scope_dcache_rsp_valid (scope_dcache_rsp_valid), \
|
||||
.scope_dcache_rsp_tag (scope_dcache_rsp_tag), \
|
||||
.scope_dcache_rsp_ready (scope_dcache_rsp_ready), \
|
||||
.scope_dram_req_valid (scope_dram_req_valid), \
|
||||
.scope_dram_req_tag (scope_dram_req_tag), \
|
||||
.scope_dram_req_ready (scope_dram_req_ready), \
|
||||
.scope_dram_rsp_valid (scope_dram_rsp_valid), \
|
||||
.scope_dram_rsp_tag (scope_dram_rsp_tag), \
|
||||
.scope_dram_rsp_ready (scope_dram_rsp_ready), \
|
||||
.scope_schedule_delay (scope_schedule_delay)
|
||||
.scope_dram_req_valid (scope_dram_req_valid), \
|
||||
.scope_dram_req_tag (scope_dram_req_tag), \
|
||||
.scope_dram_req_ready (scope_dram_req_ready), \
|
||||
.scope_dram_rsp_valid (scope_dram_rsp_valid), \
|
||||
.scope_dram_rsp_tag (scope_dram_rsp_tag), \
|
||||
.scope_dram_rsp_ready (scope_dram_rsp_ready), \
|
||||
.scope_schedule_delay (scope_schedule_delay),
|
||||
|
||||
`define SCOPE_ASSIGN(d,s) assign d = s
|
||||
`else
|
||||
`define SCOPE_SIGNALS_IO()
|
||||
`define SCOPE_SIGNALS_ATTACH()
|
||||
`define SCOPE_SIGNALS_IO
|
||||
`define SCOPE_SIGNALS_ATTACH
|
||||
`define SCOPE_ASSIGN(d,s)
|
||||
`endif
|
||||
|
||||
|
||||
@@ -3,7 +3,7 @@
|
||||
module VX_pipeline #(
|
||||
parameter CORE_ID = 0
|
||||
) (
|
||||
`SCOPE_SIGNALS_IO(),
|
||||
`SCOPE_SIGNALS_IO
|
||||
|
||||
// Clock
|
||||
input wire clk,
|
||||
|
||||
@@ -3,7 +3,7 @@
|
||||
module Vortex #(
|
||||
parameter CORE_ID = 0
|
||||
) (
|
||||
`SCOPE_SIGNALS_IO(),
|
||||
`SCOPE_SIGNALS_IO
|
||||
|
||||
// Clock
|
||||
input wire clk,
|
||||
@@ -165,7 +165,7 @@ module Vortex #(
|
||||
VX_pipeline #(
|
||||
.CORE_ID(CORE_ID)
|
||||
) pipeline (
|
||||
`SCOPE_SIGNALS_ATTACH(),
|
||||
`SCOPE_SIGNALS_ATTACH
|
||||
|
||||
.clk(clk),
|
||||
.reset(reset),
|
||||
|
||||
@@ -3,7 +3,7 @@
|
||||
module Vortex_Cluster #(
|
||||
parameter CLUSTER_ID = 0
|
||||
) (
|
||||
`SCOPE_SIGNALS_IO(),
|
||||
`SCOPE_SIGNALS_IO
|
||||
|
||||
// Clock
|
||||
input wire clk,
|
||||
@@ -108,7 +108,7 @@ module Vortex_Cluster #(
|
||||
Vortex #(
|
||||
.CORE_ID(i + (CLUSTER_ID * `NUM_CORES))
|
||||
) vortex_core (
|
||||
`SCOPE_SIGNALS_ATTACH(),
|
||||
`SCOPE_SIGNALS_ATTACH
|
||||
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
|
||||
@@ -1,7 +1,7 @@
|
||||
`include "VX_define.vh"
|
||||
|
||||
module Vortex_Socket (
|
||||
`SCOPE_SIGNALS_IO(),
|
||||
`SCOPE_SIGNALS_IO
|
||||
|
||||
// Clock
|
||||
input wire clk,
|
||||
@@ -64,7 +64,7 @@ module Vortex_Socket (
|
||||
Vortex_Cluster #(
|
||||
.CLUSTER_ID(`L3CACHE_ID)
|
||||
) Vortex_Cluster (
|
||||
`SCOPE_SIGNALS_ATTACH(),
|
||||
`SCOPE_SIGNALS_ATTACH
|
||||
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
@@ -151,7 +151,7 @@ module Vortex_Socket (
|
||||
Vortex_Cluster #(
|
||||
.CLUSTER_ID(i)
|
||||
) Vortex_Cluster (
|
||||
`SCOPE_SIGNALS_ATTACH(),
|
||||
`SCOPE_SIGNALS_ATTACH
|
||||
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
|
||||
@@ -1,87 +1,146 @@
|
||||
`include "VX_define.vh"
|
||||
|
||||
module VX_scope #(
|
||||
parameter DATAW = 64,
|
||||
parameter BUSW = 64,
|
||||
parameter SIZE = 1024
|
||||
parameter SIZE = 256
|
||||
) (
|
||||
input wire clk,
|
||||
input wire reset,
|
||||
input wire start,
|
||||
input wire [DATAW-1:0] data_in,
|
||||
input wire [BUSW-1:0] bus_in,
|
||||
output wire [BUSW-1:0] bus_out,
|
||||
input wire [BUSW-1:0] bus_in,
|
||||
output reg [BUSW-1:0] bus_out,
|
||||
input wire bus_write,
|
||||
input wire bus_read
|
||||
);
|
||||
typedef enum logic[2:0] {
|
||||
CMD_GET_VALID,
|
||||
CMD_GET_DATA,
|
||||
CMD_GET_WIDTH,
|
||||
CMD_GET_DEPTH,
|
||||
CMD_SET_DELAY,
|
||||
CMD_SET_DURATION,
|
||||
CMD_SET_RESERVED1,
|
||||
CMD_SET_RESERVED2
|
||||
} cmd_t;
|
||||
|
||||
typedef enum logic[1:0] {
|
||||
GET_VALID,
|
||||
GET_DATA,
|
||||
GET_WIDTH,
|
||||
GET_DEPTH
|
||||
} cmd_get_t;
|
||||
|
||||
reg [DATAW-1:0] mem [SIZE-1:0];
|
||||
//reg [63:0] offsets [SIZE-1:0];
|
||||
|
||||
reg [`CLOG2(SIZE)-1:0] raddr, waddr;
|
||||
|
||||
reg started, running, done;
|
||||
|
||||
reg [BUSW-1:0] delay_cntr;
|
||||
|
||||
reg data_valid, data_end;
|
||||
reg [`CLOG2(SIZE)-1:0] raddr, waddr, waddr_end;
|
||||
|
||||
reg [`LOG2UP(DATAW)-1:0] read_offset;
|
||||
|
||||
wire [BUSW-3:0] data_part;
|
||||
reg start_wait, recording, data_valid;
|
||||
|
||||
reg [BUSW-3:0] delay_val, delay_cntr;
|
||||
|
||||
reg [1:0] out_cmd;
|
||||
|
||||
wire [2:0] cmd_type;
|
||||
wire [BUSW-4:0] cmd_data;
|
||||
assign {cmd_data, cmd_type} = bus_in;
|
||||
|
||||
always @(posedge clk) begin
|
||||
if (reset) begin
|
||||
raddr <= 0;
|
||||
waddr <= 0;
|
||||
started <= 0;
|
||||
running <= 0;
|
||||
done <= 0;
|
||||
start_wait <= 0;
|
||||
recording <= 0;
|
||||
delay_cntr <= 0;
|
||||
read_offset <= 0;
|
||||
data_valid <= 0;
|
||||
out_cmd <= $bits(out_cmd)'(CMD_GET_VALID);
|
||||
delay_val <= 0;
|
||||
waddr_end <= $bits(waddr)'(SIZE-1);
|
||||
end else begin
|
||||
|
||||
if (bus_write) begin
|
||||
delay_cntr <= bus_in;
|
||||
case (cmd_type)
|
||||
CMD_GET_VALID,
|
||||
CMD_GET_DATA,
|
||||
CMD_GET_WIDTH,
|
||||
CMD_GET_DEPTH: out_cmd <= $bits(out_cmd)'(cmd_type);
|
||||
CMD_SET_DELAY: delay_val <= $bits(delay_val)'(cmd_data);
|
||||
CMD_SET_DURATION: waddr_end <= $bits(waddr)'(cmd_data);
|
||||
default:;
|
||||
endcase
|
||||
end
|
||||
|
||||
if (start) begin
|
||||
started <= 1;
|
||||
end
|
||||
|
||||
if (start || started) begin
|
||||
if (0 == delay_cntr) begin
|
||||
running <= 1;
|
||||
if (start) begin
|
||||
waddr <= 0;
|
||||
if (0 == delay_val) begin
|
||||
start_wait <= 0;
|
||||
recording <= 1;
|
||||
delay_cntr <= 0;
|
||||
end else begin
|
||||
delay_cntr <= delay_cntr - 1;
|
||||
start_wait <= 1;
|
||||
recording <= 0;
|
||||
delay_cntr <= delay_val;
|
||||
end
|
||||
end
|
||||
|
||||
if (running && !done) begin
|
||||
if (start_wait) begin
|
||||
delay_cntr <= delay_cntr - 1;
|
||||
if (1 == delay_cntr) begin
|
||||
start_wait <= 0;
|
||||
recording <= 1;
|
||||
end
|
||||
end
|
||||
|
||||
if (recording) begin
|
||||
mem[waddr] <= data_in;
|
||||
waddr <= waddr + 1;
|
||||
if (waddr == $bits(waddr)'(SIZE-1)) begin
|
||||
done <= 1;
|
||||
if (waddr == waddr_end) begin
|
||||
recording <= 0;
|
||||
data_valid <= 1;
|
||||
end
|
||||
end
|
||||
|
||||
if (bus_read) begin
|
||||
if (DATAW > (BUSW-2)) begin
|
||||
if (read_offset < $bits(read_offset)'(DATAW-(BUSW-2))) begin
|
||||
read_offset <= read_offset + $bits(read_offset)'(BUSW-2);
|
||||
if (bus_read
|
||||
&& (out_cmd == GET_DATA)
|
||||
&& data_valid) begin
|
||||
if (DATAW > BUSW) begin
|
||||
if (read_offset < $bits(read_offset)'(DATAW-BUSW)) begin
|
||||
read_offset <= read_offset + $bits(read_offset)'(BUSW);
|
||||
end else begin
|
||||
read_offset <= 0;
|
||||
raddr <= raddr + 1;
|
||||
end
|
||||
read_offset <= 0;
|
||||
if (raddr == waddr_end) begin
|
||||
data_valid <= 0;
|
||||
end
|
||||
end
|
||||
end else begin
|
||||
raddr <= raddr + 1;
|
||||
end
|
||||
raddr <= raddr + 1;
|
||||
if (raddr == waddr_end) begin
|
||||
data_valid <= 0;
|
||||
end
|
||||
end
|
||||
end
|
||||
end
|
||||
end
|
||||
|
||||
assign data_valid = (waddr != 0) && (raddr <= waddr);
|
||||
always @(*) begin
|
||||
case (out_cmd)
|
||||
GET_VALID : bus_out = BUSW'(data_valid);
|
||||
GET_WIDTH : bus_out = BUSW'(DATAW);
|
||||
GET_DEPTH : bus_out = BUSW'(waddr_end) + BUSW'(1);
|
||||
default : bus_out = (BUSW)'(mem[raddr] >> read_offset);
|
||||
endcase
|
||||
end
|
||||
|
||||
assign data_end = (0 == read_offset) || (raddr == waddr);
|
||||
|
||||
assign data_part = (BUSW-2)'(mem[raddr] >> read_offset);
|
||||
|
||||
assign bus_out = {data_valid, data_end, data_part};
|
||||
always_ff @(posedge clk) begin
|
||||
if (bus_read) begin
|
||||
$display("%t: read: cmd=%0d, out=0x%0h, addr=%0d, off=%0d", $time, out_cmd, bus_out, raddr, read_offset);
|
||||
end
|
||||
end
|
||||
|
||||
endmodule
|
||||
Reference in New Issue
Block a user