github-mirrors/vortex
1
0
Fork 0
mirror of https://github.com/vortexgpgpu/vortex.git synced 2025-04-23 21:39:10 -04:00
Code Issues Projects Releases Packages Wiki Activity Actions

extending tracing feature for advanced debugging

Browse source
This commit is contained in:
Blaise Tine 2021-08-15 05:10:46 -07:00
parent 640c98a4e8
commit a60bfc5e01
7 changed files with 100 additions and 36 deletions
Download patch file Download diff file Expand all files Collapse all files

32
driver/opae/vlsim/opae_sim.cpp
View file

@ -10,8 +10,12 @@
#define ENABLE_MEM_STALLS
#ifndef TRACE_DELAY
#define TRACE_DELAY 0
#ifndef TRACE_START_TIME
#define TRACE_START_TIME 0ull
#endif
#ifndef TRACE_STOP_TIME
#define TRACE_STOP_TIME -1ull
#endif
#ifndef MEM_LATENCY
@ -30,8 +34,6 @@
#define VERILATOR_RESET_VALUE 2
#endif
uint64_t sim_trace_delay = TRACE_DELAY;
static uint64_t timestamp = 0;
double sc_time_stamp() {
@ -55,6 +57,23 @@ static void __aligned_free(void *ptr) {
///////////////////////////////////////////////////////////////////////////////
static bool trace_enabled = false;
static uint64_t trace_start_time = TRACE_START_TIME;
static uint64_t trace_stop_time = TRACE_STOP_TIME;
bool sim_trace_enabled() {
if (timestamp >= trace_start_time
&& timestamp < trace_stop_time)
return true;
return trace_enabled;
}
void sim_trace_enable(bool enable) {
trace_enabled = enable;
}
///////////////////////////////////////////////////////////////////////////////
opae_sim::opae_sim()
: stop_(false)
, host_buffer_ids_(0)
@ -205,7 +224,7 @@ void opae_sim::step() {
void opae_sim::eval() {
vortex_afu_->eval();
#ifdef VCD_OUTPUT
if (timestamp >= sim_trace_delay) {
if (sim_trace_enabled()) {
trace_->dump(timestamp);
}
#endif
@ -349,7 +368,7 @@ void opae_sim::avs_bus() {
}
/*printf("%0ld: [sim] MEM Wr Req: bank=%d, addr=%x, data=", timestamp, b, base_addr);
for (int i = 0; i < MEM_BLOCK_SIZE; i++) {
printf("%0x", data[(MEM_BLOCK_SIZE-1)-i]);
printf("%02x", data[(MEM_BLOCK_SIZE-1)-i]);
}
printf("\n");*/
}
@ -360,6 +379,7 @@ void opae_sim::avs_bus() {
mem_req.cycles_left = MEM_LATENCY;
for (auto& rsp : mem_reads_[b]) {
if (mem_req.addr == rsp.addr) {
// duplicate requests receive the same cycle delay
mem_req.cycles_left = rsp.cycles_left;
break;
}

3
driver/opae/vlsim/opae_sim.h
View file

@ -1,7 +1,6 @@
#pragma once
#include "verilated.h"
//#include "verilated_stub.h"
#include <verilated.h>
#include "Vvortex_afu_shim.h"
#include "Vvortex_afu_shim__Syms.h"

21
driver/rtlsim/vortex.cpp
View file

@ -21,7 +21,6 @@ inline size_t align_size(size_t size, size_t alignment) {
///////////////////////////////////////////////////////////////////////////////
class vx_device;
class vx_buffer {
public:
vx_buffer(size_t size, vx_device* device)
@ -84,11 +83,11 @@ public:
if (dest_addr + asize > ram_.size())
return -1;
/*printf("VXDRV: upload %d bytes from 0x%lx to 0x%lx", size, (uint8_t*)src + src_offset, dest_addr);
if (size <= 1024) {
printf(": ");
for (int i = asize-1; i >= 0; --i) {
printf("%x", *((uint8_t*)src + src_offset + i));
/*printf("VXDRV: upload %ld bytes from 0x%lx:", size, uintptr_t((uint8_t*)src + src_offset));
for (int i = 0; i < (asize / CACHE_BLOCK_SIZE); ++i) {
printf("\n0x%08lx=", dest_addr + i * CACHE_BLOCK_SIZE);
for (int j = 0; j < CACHE_BLOCK_SIZE; ++j) {
printf("%02x", *((uint8_t*)src + src_offset + i * CACHE_BLOCK_SIZE + CACHE_BLOCK_SIZE - 1 - j));
}
}
printf("\n");*/
@ -104,11 +103,11 @@ public:
ram_.read(src_addr, asize, (uint8_t*)dest + dest_offset);
/*printf("VXDRV: download %d bytes from 0x%lx to 0x%lx", size, src_addr, (uint8_t*)dest + dest_offset);
if (size <= 1024) {
printf(": ");
for (int i = asize-1; i >= 0; --i) {
printf("%x", *((uint8_t*)dest + dest_offset + i));
/*printf("VXDRV: download %ld bytes to 0x%lx:", size, uintptr_t((uint8_t*)dest + dest_offset));
for (int i = 0; i < (asize / CACHE_BLOCK_SIZE); ++i) {
printf("\n0x%08lx=", src_addr + i * CACHE_BLOCK_SIZE);
for (int j = 0; j < CACHE_BLOCK_SIZE; ++j) {
printf("%02x", *((uint8_t*)dest + dest_offset + i * CACHE_BLOCK_SIZE + CACHE_BLOCK_SIZE - 1 - j));
}
}
printf("\n");*/

22
hw/dpi/util_dpi.cpp
View file

@ -15,12 +15,13 @@ extern "C" {
int dpi_register();
void dpi_assert(int inst, bool cond, int delay);
void dpi_trace(const char* format, ...);
void dpi_trace(const char* format, ...);
void dpi_trace_start();
void dpi_trace_stop();
}
double sc_time_stamp();
extern uint64_t sim_trace_delay;
bool sim_trace_enabled();
void sim_trace_enable(bool enable);
class ShiftRegister {
public:
@ -141,12 +142,19 @@ void dpi_idiv(int a, int b, bool is_signed, int* quotient, int* remainder) {
}
}
void dpi_trace(const char* format, ...) {
uint64_t timestamp = (uint64_t)sc_time_stamp();
if (timestamp < sim_trace_delay)
void dpi_trace(const char* format, ...) {
if (!sim_trace_enabled())
return;
va_list va;
va_start(va, format);
vprintf(format, va);
va_end(va);
}
void dpi_trace_start() {
sim_trace_enable(true);
}
void dpi_trace_stop() {
sim_trace_enable(false);
}

2
hw/dpi/util_dpi.vh
View file

@ -8,5 +8,7 @@ import "DPI-C" function int dpi_register();
import "DPI-C" function void dpi_assert(int inst, input logic cond, input int delay);
import "DPI-C" function void dpi_trace(input string format /*verilator sformat*/);
import "DPI-C" function void dpi_trace_start();
import "DPI-C" function void dpi_trace_stop();
`endif

52
hw/simulate/simulator.cpp
View file

@ -5,8 +5,12 @@
#define ENABLE_MEM_STALLS
#ifndef TRACE_DELAY
#define TRACE_DELAY 0
#ifndef TRACE_START_TIME
#define TRACE_START_TIME 0ull
#endif
#ifndef TRACE_STOP_TIME
#define TRACE_STOP_TIME -1ull
#endif
#ifndef MEM_LATENCY
@ -28,14 +32,31 @@
#define VL_WDATA_GETW(lwp, i, n, w) \
VL_SEL_IWII(0, n * w, 0, 0, lwp, i * w, w)
uint64_t sim_trace_delay = TRACE_DELAY;
static uint64_t timestamp = 0;
double sc_time_stamp() {
return timestamp;
}
///////////////////////////////////////////////////////////////////////////////
static bool trace_enabled = false;
static uint64_t trace_start_time = TRACE_START_TIME;
static uint64_t trace_stop_time = TRACE_STOP_TIME;
bool sim_trace_enabled() {
if (timestamp >= trace_start_time
&& timestamp < trace_stop_time)
return true;
return trace_enabled;
}
void sim_trace_enable(bool enable) {
trace_enabled = enable;
}
///////////////////////////////////////////////////////////////////////////////
Simulator::Simulator() {
// force random values for unitialized signals
Verilated::randReset(VERILATOR_RESET_VALUE);
@ -127,7 +148,7 @@ void Simulator::step() {
void Simulator::eval() {
vortex_->eval();
#ifdef VCD_OUTPUT
if (timestamp >= sim_trace_delay) {
if (sim_trace_enabled()) {
trace_->dump(timestamp);
}
#endif
@ -169,7 +190,14 @@ void Simulator::eval_mem_bus() {
if (!mem_rsp_active_) {
if (has_response) {
vortex_->mem_rsp_valid = 1;
std::list<mem_req_t>::iterator mem_rsp_it = mem_rsp_vec_[last_mem_rsp_bank_].begin();
std::list<mem_req_t>::iterator mem_rsp_it = mem_rsp_vec_[last_mem_rsp_bank_].begin();
/*
printf("%0ld: [sim] MEM Rd: bank=%d, addr=%0lx, data=", timestamp, last_mem_rsp_bank_, mem_rsp_it->addr);
for (int i = 0; i < MEM_BLOCK_SIZE; i++) {
printf("%02x", mem_rsp_it->block[(MEM_BLOCK_SIZE-1)-i]);
}
printf("\n");
*/
memcpy((uint8_t*)vortex_->mem_rsp_data, mem_rsp_it->block.data(), MEM_BLOCK_SIZE);
vortex_->mem_rsp_tag = mem_rsp_it->tag;
mem_rsp_vec_[last_mem_rsp_bank_].erase(mem_rsp_it);
@ -214,20 +242,28 @@ void Simulator::eval_mem_bus() {
}
}
} else {
/*
printf("%0ld: [sim] MEM Wr: addr=%0x, byteen=%0lx, data=", timestamp, base_addr, byteen);
for (int i = 0; i < MEM_BLOCK_SIZE; i++) {
printf("%02x", data[(MEM_BLOCK_SIZE-1)-i]);
}
printf("\n");
*/
for (int i = 0; i < MEM_BLOCK_SIZE; i++) {
if ((byteen >> i) & 0x1) {
(*ram_)[base_addr + i] = data[i];
}
}
}
} else {
} else {
mem_req_t mem_req;
mem_req.tag = vortex_->mem_req_tag;
mem_req.addr = vortex_->mem_req_addr;
mem_req.addr = (vortex_->mem_req_addr * MEM_BLOCK_SIZE);
ram_->read(vortex_->mem_req_addr * MEM_BLOCK_SIZE, MEM_BLOCK_SIZE, mem_req.block.data());
mem_req.cycles_left = MEM_LATENCY;
for (auto& rsp : mem_rsp_vec_[req_bank]) {
if (mem_req.addr == rsp.addr) {
// duplicate requests receive the same cycle delay
mem_req.cycles_left = rsp.cycles_left;
break;
}

4
hw/simulate/simulator.h
View file

@ -1,8 +1,8 @@
#pragma once
#include <verilated.h>
#include "VVortex.h"
#include "VVortex__Syms.h"
#include "verilated.h"
#ifdef VCD_OUTPUT
#include <verilated_vcd_c.h>
@ -51,7 +51,7 @@ private:
typedef struct {
int cycles_left;
std::array<uint8_t, MEM_BLOCK_SIZE> block;
uint32_t addr;
uint64_t addr;
uint64_t tag;
} mem_req_t;