verilator/include/verilated_vcd_c.cpp

773 lines
27 KiB
C++
Raw Permalink Normal View History

// -*- mode: C++; c-file-style: "cc-mode" -*-
//=============================================================================
//
// Code available from: https://verilator.org
//
2023-01-01 15:18:39 +00:00
// Copyright 2001-2023 by Wilson Snyder. This program is free software; you
// can redistribute it and/or modify it under the terms of either the GNU
// Lesser General Public License Version 3 or the Perl Artistic License
// Version 2.0.
// SPDX-License-Identifier: LGPL-3.0-only OR Artistic-2.0
//
//=============================================================================
///
/// \file
/// \brief Verilated C++ tracing in VCD format implementation code
///
/// This file must be compiled and linked against all Verilated objects
/// that use --trace.
///
/// Use "verilator --trace" to add this to the Makefile for the linker.
///
//=============================================================================
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
// clang-format off
#include "verilatedos.h"
#include "verilated.h"
#include "verilated_vcd_c.h"
#include <algorithm>
#include <cerrno>
#include <fcntl.h>
#if defined(_WIN32) && !defined(__MINGW32__) && !defined(__CYGWIN__)
# include <io.h>
#else
# include <unistd.h>
#endif
2022-06-21 23:28:23 +00:00
#ifndef O_LARGEFILE // WIN32 headers omit this
# define O_LARGEFILE 0
#endif
2022-06-21 23:28:23 +00:00
#ifndef O_NONBLOCK // WIN32 headers omit this
# define O_NONBLOCK 0
#endif
2022-06-21 23:28:23 +00:00
#ifndef O_CLOEXEC // WIN32 headers omit this
# define O_CLOEXEC 0
#endif
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
// clang-format on
// This size comes form VCD allowing use of printable ASCII characters between
// '!' and '~' inclusive, which are a total of 94 different values. Encoding a
// 32 bit code hence needs a maximum of std::ceil(log94(2**32-1)) == 5 bytes.
constexpr unsigned VL_TRACE_MAX_VCD_CODE_SIZE = 5; // Maximum length of a VCD string code
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
// We use 8 bytes per code in a suffix buffer array.
// 1 byte optional separator + VL_TRACE_MAX_VCD_CODE_SIZE bytes for code
// + 1 byte '\n' + 1 byte suffix size. This luckily comes out to a power of 2,
// meaning the array can be aligned such that entries never straddle multiple
// cache-lines.
constexpr unsigned VL_TRACE_SUFFIX_ENTRY_SIZE = 8; // Size of a suffix entry
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
//=============================================================================
// Utility functions: TODO: put these in a common place and share them.
template <size_t N>
static size_t roundUpToMultipleOf(size_t value) {
static_assert((N & (N - 1)) == 0, "'N' must be a power of 2");
size_t mask = N - 1;
return (value + mask) & ~mask;
}
//=============================================================================
// Specialization of the generics for this trace format
#define VL_SUB_T VerilatedVcd
#define VL_BUF_T VerilatedVcdBuffer
#include "verilated_trace_imp.h"
#undef VL_SUB_T
#undef VL_BUF_T
//=============================================================================
//=============================================================================
//=============================================================================
// VerilatedVcdFile
bool VerilatedVcdFile::open(const std::string& name) VL_MT_UNSAFE {
m_fd = ::open(name.c_str(),
O_CREAT | O_WRONLY | O_TRUNC | O_LARGEFILE | O_NONBLOCK | O_CLOEXEC, 0666);
2020-03-02 02:39:23 +00:00
return m_fd >= 0;
}
void VerilatedVcdFile::close() VL_MT_UNSAFE { ::close(m_fd); }
ssize_t VerilatedVcdFile::write(const char* bufp, ssize_t len) VL_MT_UNSAFE {
return ::write(m_fd, bufp, len);
}
//=============================================================================
//=============================================================================
//=============================================================================
// Opening/Closing
VerilatedVcd::VerilatedVcd(VerilatedVcdFile* filep) {
// Not in header to avoid link issue if header is included without this .cpp file
m_fileNewed = (filep == nullptr);
m_filep = m_fileNewed ? new VerilatedVcdFile : filep;
m_wrChunkSize = 8 * 1024;
m_wrBufp = new char[m_wrChunkSize * 8];
m_wrFlushp = m_wrBufp + m_wrChunkSize * 6;
m_writep = m_wrBufp;
}
void VerilatedVcd::open(const char* filename) VL_MT_SAFE_EXCLUDES(m_mutex) {
const VerilatedLockGuard lock{m_mutex};
if (isOpen()) return;
// Set member variables
m_filename = filename; // "" is ok, as someone may overload open
2010-03-13 01:00:08 +00:00
openNextImp(m_rolloverSize != 0);
if (!isOpen()) return;
dumpHeader();
// When using rollover, the first chunk contains the header only.
if (m_rolloverSize) openNextImp(true);
}
void VerilatedVcd::openNext(bool incFilename) VL_MT_SAFE_EXCLUDES(m_mutex) {
// Open next filename in concat sequence, mangle filename if
// incFilename is true.
const VerilatedLockGuard lock{m_mutex};
openNextImp(incFilename);
}
void VerilatedVcd::openNextImp(bool incFilename) {
closePrev(); // Close existing
if (incFilename) {
// Find _0000.{ext} in filename
std::string name = m_filename;
const size_t pos = name.rfind('.');
if (pos > 8 && 0 == std::strncmp("_cat", name.c_str() + pos - 8, 4)
&& std::isdigit(name.c_str()[pos - 4]) && std::isdigit(name.c_str()[pos - 3])
&& std::isdigit(name.c_str()[pos - 2]) && std::isdigit(name.c_str()[pos - 1])) {
// Increment code.
if ((++(name[pos - 1])) > '9') {
name[pos - 1] = '0';
if ((++(name[pos - 2])) > '9') {
name[pos - 2] = '0';
if ((++(name[pos - 3])) > '9') {
name[pos - 3] = '0';
if ((++(name[pos - 4])) > '9') { //
name[pos - 4] = '0';
}
}
}
}
} else {
// Append _cat0000
name.insert(pos, "_cat0000");
}
m_filename = name;
}
2020-05-16 11:43:22 +00:00
if (VL_UNCOVERABLE(m_filename[0] == '|')) {
assert(0); // LCOV_EXCL_LINE // Not supported yet.
} else {
// cppcheck-suppress duplicateExpression
if (!m_filep->open(m_filename)) {
// User code can check isOpen()
m_isOpen = false;
return;
}
}
m_isOpen = true;
fullDump(true); // First dump must be full
m_wroteBytes = 0;
}
bool VerilatedVcd::preChangeDump() {
if (VL_UNLIKELY(m_rolloverSize && m_wroteBytes > m_rolloverSize)) openNextImp(true);
return isOpen();
}
void VerilatedVcd::emitTimeChange(uint64_t timeui) {
printStr("#");
printQuad(timeui);
printStr("\n");
}
void VerilatedVcd::makeNameMap() {
// Take signal information from each module and build m_namemapp
deleteNameMap();
m_namemapp = new NameMap;
Super::traceInit();
// Though not speced, it's illegal to generate a vcd with signals
// not under any module - it crashes at least two viewers.
// If no scope was specified, prefix everything with a "top"
// This comes from user instantiations with no name - IE Vtop("").
bool nullScope = false;
for (const auto& i : *m_namemapp) {
const std::string& hiername = i.first;
if (!hiername.empty() && hiername[0] == '\t') nullScope = true;
}
if (nullScope) {
NameMap* const newmapp = new NameMap;
for (const auto& i : *m_namemapp) {
const std::string& hiername = i.first;
const std::string& decl = i.second;
std::string newname{"top"};
if (hiername[0] != '\t') newname += ' ';
newname += hiername;
newmapp->emplace(newname, decl);
}
deleteNameMap();
m_namemapp = newmapp;
}
}
void VerilatedVcd::deleteNameMap() {
if (m_namemapp) VL_DO_CLEAR(delete m_namemapp, m_namemapp = nullptr);
}
VerilatedVcd::~VerilatedVcd() {
close();
if (m_wrBufp) VL_DO_CLEAR(delete[] m_wrBufp, m_wrBufp = nullptr);
deleteNameMap();
if (m_filep && m_fileNewed) VL_DO_CLEAR(delete m_filep, m_filep = nullptr);
if (parallel()) {
assert(m_numBuffers == m_freeBuffers.size());
for (auto& pair : m_freeBuffers) VL_DO_CLEAR(delete[] pair.first, pair.first = nullptr);
}
}
void VerilatedVcd::closePrev() {
// This function is on the flush() call path
if (!isOpen()) return;
Super::flushBase();
bufferFlush();
m_isOpen = false;
m_filep->close();
}
void VerilatedVcd::closeErr() {
// This function is on the flush() call path
// Close due to an error. We might abort before even getting here,
// depending on the definition of vl_fatal.
if (!isOpen()) return;
// No buffer flush, just fclose
m_isOpen = false;
m_filep->close(); // May get error, just ignore it
}
void VerilatedVcd::close() VL_MT_SAFE_EXCLUDES(m_mutex) {
// This function is on the flush() call path
const VerilatedLockGuard lock{m_mutex};
if (!isOpen()) return;
closePrev();
// closePrev() called Super::flush(), so we just
// need to shut down the tracing thread here.
Super::closeBase();
}
void VerilatedVcd::flush() VL_MT_SAFE_EXCLUDES(m_mutex) {
const VerilatedLockGuard lock{m_mutex};
Super::flushBase();
bufferFlush();
}
void VerilatedVcd::printStr(const char* str) {
// Not fast...
while (*str) {
*m_writep++ = *str++;
bufferCheck();
}
}
void VerilatedVcd::printQuad(uint64_t n) {
constexpr size_t LEN_STR_QUAD = 40;
char buf[LEN_STR_QUAD];
VL_SNPRINTF(buf, LEN_STR_QUAD, "%" PRIu64, n);
printStr(buf);
}
void VerilatedVcd::bufferResize(size_t minsize) {
// minsize is size of largest write. We buffer at least 8 times as much data,
// writing when we are 3/4 full (with thus 2*minsize remaining free)
if (VL_UNLIKELY(minsize > m_wrChunkSize)) {
const char* oldbufp = m_wrBufp;
m_wrChunkSize = roundUpToMultipleOf<1024>(minsize * 2);
m_wrBufp = new char[m_wrChunkSize * 8];
std::memcpy(m_wrBufp, oldbufp, m_writep - oldbufp);
m_writep = m_wrBufp + (m_writep - oldbufp);
m_wrFlushp = m_wrBufp + m_wrChunkSize * 6;
VL_DO_CLEAR(delete[] oldbufp, oldbufp = nullptr);
}
}
void VerilatedVcd::bufferFlush() VL_MT_UNSAFE_ONE {
// This function can be called from the trace offload thread
// This function is on the flush() call path
// We add output data to m_writep.
// When it gets nearly full we dump it using this routine which calls write()
// This is much faster than using buffered I/O
2010-02-03 11:52:02 +00:00
if (VL_UNLIKELY(!isOpen())) return;
const char* wp = m_wrBufp;
while (true) {
const ssize_t remaining = (m_writep - wp);
if (remaining == 0) break;
errno = 0;
const ssize_t got = m_filep->write(wp, remaining);
if (got > 0) {
wp += got;
m_wroteBytes += got;
2020-05-16 00:38:08 +00:00
} else if (VL_UNCOVERABLE(got < 0)) {
if (VL_UNCOVERABLE(errno != EAGAIN && errno != EINTR)) {
// LCOV_EXCL_START
// write failed, presume error (perhaps out of disk space)
const std::string msg
= std::string{"VerilatedVcd::bufferFlush: "} + std::strerror(errno);
VL_FATAL_MT("", 0, "", msg.c_str());
closeErr();
break;
// LCOV_EXCL_STOP
}
}
}
// Reset buffer
m_writep = m_wrBufp;
}
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
//=============================================================================
// VCD string code
char* VerilatedVcd::writeCode(char* writep, uint32_t code) {
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
*writep++ = static_cast<char>('!' + code % 94);
code /= 94;
while (code) {
--code;
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
*writep++ = static_cast<char>('!' + code % 94);
code /= 94;
}
return writep;
}
//=============================================================================
// Definitions
void VerilatedVcd::printIndent(int level_change) {
if (level_change < 0) m_modDepth += level_change;
assert(m_modDepth >= 0);
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
for (int i = 0; i < m_modDepth; i++) printStr(" ");
if (level_change > 0) m_modDepth += level_change;
}
void VerilatedVcd::dumpHeader() {
printStr("$version Generated by VerilatedVcd $end\n");
2022-11-28 01:24:22 +00:00
// Verilator used to put in a $date here. Although $date is shown in
// IEEE examples, and it is common in VCD writers, VCD readers don't
// seem to care about it. Thus, we omit the $date so artifacts are
// more likely to be reproducible. If use cases show up that require
// the $date command to be present, it could be re-added with support
// for the SOURCE_DATE_EPOCH hook.
printStr("$timescale ");
printStr(timeResStr().c_str()); // lintok-begin-on-ref
printStr(" $end\n");
makeNameMap();
// Signal header
assert(m_modDepth == 0);
printIndent(1);
printStr("\n");
// We detect the spaces in module names to determine hierarchy. This
// allows signals to be declared without fixed ordering, which is
// required as Verilog signals might be separately declared from
// SC module signals.
// Print the signal names
const char* lastName = "";
for (const auto& i : *m_namemapp) {
const std::string& hiernamestr = i.first;
const std::string& decl = i.second;
// Determine difference between the old and new names
const char* const hiername = hiernamestr.c_str();
const char* lp = lastName;
const char* np = hiername;
lastName = hiername;
// Skip common prefix, it must break at a space or tab
for (; *np && (*np == *lp); np++, lp++) {}
while (np != hiername && *np && *np != ' ' && *np != '\t') {
--np;
--lp;
}
// printf("hier %s\n lp=%s\n np=%s\n",hiername,lp,np);
// Any extra spaces in last name are scope ups we need to do
bool first = true;
for (; *lp; lp++) {
if (*lp == ' ' || (first && *lp != '\t')) {
printIndent(-1);
printStr("$upscope $end\n");
}
first = false;
}
// Any new spaces are scope downs we need to do
while (*np) {
if (*np == ' ') np++;
if (*np == '\t') break; // tab means signal name starts
printIndent(1);
// Find character after name end
const char* sp = np;
while (*sp && *sp != ' ' && *sp != '\t' && !(*sp & '\x80')) sp++;
printStr("$scope ");
if (*sp & '\x80') {
switch (*sp & 0x7f) {
case VLT_TRACE_SCOPE_STRUCT: printStr("struct "); break;
case VLT_TRACE_SCOPE_INTERFACE: printStr("interface "); break;
case VLT_TRACE_SCOPE_UNION: printStr("union "); break;
default: printStr("module ");
}
} else {
printStr("module ");
}
for (; *np && *np != ' ' && *np != '\t'; np++) {
if (*np == '[') {
printStr("[");
} else if (*np == ']') {
printStr("]");
} else if (!(*np & '\x80')) {
*m_writep++ = *np;
}
}
printStr(" $end\n");
}
printIndent(0);
printStr(decl.c_str());
}
while (m_modDepth > 1) {
printIndent(-1);
printStr("$upscope $end\n");
}
printIndent(-1);
printStr("$enddefinitions $end\n\n\n");
assert(m_modDepth == 0);
// Reclaim storage
deleteNameMap();
}
void VerilatedVcd::declare(uint32_t code, const char* name, const char* wirep, bool array,
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
int arraynum, bool tri, bool bussed, int msb, int lsb) {
const int bits = ((msb > lsb) ? (msb - lsb) : (lsb - msb)) + 1;
const bool enabled = Super::declCode(code, name, bits, tri);
if (m_suffixes.size() <= nextCode() * VL_TRACE_SUFFIX_ENTRY_SIZE) {
m_suffixes.resize(nextCode() * VL_TRACE_SUFFIX_ENTRY_SIZE * 2, 0);
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
}
// Keep upper bound on bytes a single signal can emit into the buffer
m_maxSignalBytes = std::max<size_t>(m_maxSignalBytes, bits + 32);
// Make sure write buffer is large enough, plus header
bufferResize(m_maxSignalBytes + 1024);
if (!enabled) return;
// Split name into basename
// Spaces and tabs aren't legal in VCD signal names, so:
// Space separates each level of scope
// Tab separates final scope from signal name
// Tab sorts before spaces, so signals nicely will print before scopes
// Note the hiername may be nothing, if so we'll add "\t{name}"
std::string nameasstr = namePrefix() + name;
std::string hiername;
std::string basename;
for (const char* cp = nameasstr.c_str(); *cp; cp++) {
if (isScopeEscape(*cp)) {
// Ahh, we've just read a scope, not a basename
if (!hiername.empty()) hiername += " ";
hiername += basename;
basename = "";
} else {
basename += *cp;
}
}
hiername += "\t" + basename;
// Print reference
std::string decl = "$var ";
decl += wirep; // usually "wire"
constexpr size_t bufsize = 1000;
char buf[bufsize];
VL_SNPRINTF(buf, bufsize, " %2d ", bits);
decl += buf;
// Add string code to decl
char* const endp = writeCode(buf, code);
*endp = '\0';
decl += buf;
// Build suffix array entry
char* const entryp = &m_suffixes[code * VL_TRACE_SUFFIX_ENTRY_SIZE];
const size_t length = endp - buf;
assert(length <= VL_TRACE_MAX_VCD_CODE_SIZE);
// 1 bit values don't have a ' ' separator between value and string code
const bool isBit = bits == 1;
entryp[0] = ' '; // Separator
// Use memcpy as we checked size above, and strcpy is flagged unsafe
std::memcpy(entryp + !isBit, buf,
std::strlen(buf)); // Code (overwrite separator if isBit)
entryp[length + !isBit] = '\n'; // Replace '\0' with line termination '\n'
// Set length of suffix (used to increment write pointer)
2022-12-22 17:19:09 +00:00
entryp[VL_TRACE_SUFFIX_ENTRY_SIZE - 1] = static_cast<char>(length + !isBit + 1);
decl += " ";
decl += basename;
if (array) {
VL_SNPRINTF(buf, bufsize, "[%d]", arraynum);
decl += buf;
hiername += buf;
}
if (bussed) {
VL_SNPRINTF(buf, bufsize, " [%d:%d]", msb, lsb);
decl += buf;
}
decl += " $end\n";
m_namemapp->emplace(hiername, decl);
}
void VerilatedVcd::declEvent(uint32_t code, const char* name, bool array, int arraynum) {
declare(code, name, "event", array, arraynum, false, false, 0, 0);
}
void VerilatedVcd::declBit(uint32_t code, const char* name, bool array, int arraynum) {
declare(code, name, "wire", array, arraynum, false, false, 0, 0);
}
void VerilatedVcd::declBus(uint32_t code, const char* name, bool array, int arraynum, int msb,
int lsb) {
declare(code, name, "wire", array, arraynum, false, true, msb, lsb);
}
void VerilatedVcd::declQuad(uint32_t code, const char* name, bool array, int arraynum, int msb,
int lsb) {
declare(code, name, "wire", array, arraynum, false, true, msb, lsb);
}
void VerilatedVcd::declArray(uint32_t code, const char* name, bool array, int arraynum, int msb,
int lsb) {
declare(code, name, "wire", array, arraynum, false, true, msb, lsb);
}
void VerilatedVcd::declDouble(uint32_t code, const char* name, bool array, int arraynum) {
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
declare(code, name, "real", array, arraynum, false, false, 63, 0);
}
//=============================================================================
// Get/commit trace buffer
VerilatedVcd::Buffer* VerilatedVcd::getTraceBuffer() {
VerilatedVcd::Buffer* const bufp = new Buffer{*this};
if (parallel()) {
// Note: This is called from VerilatedVcd::dump, which already holds the lock
// If no buffer available, allocate a new one
if (m_freeBuffers.empty()) {
constexpr size_t pageSize = 4096;
// 4 * m_maxSignalBytes, so we can reserve 2 * m_maxSignalBytes at the end for safety
size_t startingSize = roundUpToMultipleOf<pageSize>(4 * m_maxSignalBytes);
m_freeBuffers.emplace_back(new char[startingSize], startingSize);
++m_numBuffers;
}
// Grab a buffer
const auto pair = m_freeBuffers.back();
m_freeBuffers.pop_back();
// Initialize
bufp->m_writep = bufp->m_bufp = pair.first;
bufp->m_size = pair.second;
bufp->adjustGrowp();
}
// Return the buffer
return bufp;
}
void VerilatedVcd::commitTraceBuffer(VerilatedVcd::Buffer* bufp) {
if (parallel()) {
// Note: This is called from VerilatedVcd::dump, which already holds the lock
// Resize output buffer. Note, we use the full size of the trace buffer, as
// this is a lot more stable than the actual occupancy of the trace buffer.
// This helps us to avoid re-allocations due to small size changes.
bufferResize(bufp->m_size);
// Compute occupancy of buffer
const size_t usedSize = bufp->m_writep - bufp->m_bufp;
// Copy to output buffer
std::memcpy(m_writep, bufp->m_bufp, usedSize);
// Adjust write pointer
m_writep += usedSize;
// Flush if necessary
bufferCheck();
// Put buffer back on free list
m_freeBuffers.emplace_back(bufp->m_bufp, bufp->m_size);
} else {
// Needs adjusting for emitTimeChange
m_writep = bufp->m_writep;
}
delete bufp;
}
//=============================================================================
// VerilatedVcdBuffer implementation
//=============================================================================
// Trace rendering primitives
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
static void VerilatedVcdCCopyAndAppendNewLine(char* writep,
const char* suffixp) VL_ATTR_NO_SANITIZE_ALIGN;
static void VerilatedVcdCCopyAndAppendNewLine(char* writep, const char* suffixp) {
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
// Copy the whole suffix (this avoid having hard to predict branches which
// helps a lot). Note: The maximum length of the suffix is
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
// VL_TRACE_MAX_VCD_CODE_SIZE + 2 == 7, but we unroll this here for speed.
#ifdef VL_X86_64
// Copy the whole 8 bytes in one go, this works on little-endian machines
// supporting unaligned stores.
*reinterpret_cast<uint64_t*>(writep) = *reinterpret_cast<const uint64_t*>(suffixp);
#else
// Portable variant
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
writep[0] = suffixp[0];
writep[1] = suffixp[1];
writep[2] = suffixp[2];
writep[3] = suffixp[3];
writep[4] = suffixp[4];
writep[5] = suffixp[5];
writep[6] = '\n'; // The 6th index is always '\n' if it's relevant, no need to fetch it.
#endif
}
void VerilatedVcdBuffer::finishLine(uint32_t code, char* writep) {
const char* const suffixp = m_suffixes + code * VL_TRACE_SUFFIX_ENTRY_SIZE;
VL_DEBUG_IFDEF(assert(suffixp[0]););
VerilatedVcdCCopyAndAppendNewLine(writep, suffixp);
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
// Now write back the write pointer incremented by the actual size of the
// suffix, which was stored in the last byte of the suffix buffer entry.
m_writep = writep + suffixp[VL_TRACE_SUFFIX_ENTRY_SIZE - 1];
if (m_owner.parallel()) {
// Double the size of the buffer if necessary
if (VL_UNLIKELY(m_writep >= m_growp)) {
// Compute occupied size of current buffer
const size_t usedSize = m_writep - m_bufp;
// We are always doubling the size
m_size *= 2;
// Allocate the new buffer
char* const newBufp = new char[m_size];
// Copy from current buffer to new buffer
std::memcpy(newBufp, m_bufp, usedSize);
// Delete current buffer
delete[] m_bufp;
// Make new buffer the current buffer
m_bufp = newBufp;
// Adjust write pointer
m_writep = m_bufp + usedSize;
// Adjust resize limit
adjustGrowp();
}
} else {
// Flush the write buffer if there's not enough space left for new information
// We only call this once per vector, so we need enough slop for a very wide "b###" line
if (VL_UNLIKELY(m_writep > m_wrFlushp)) {
m_owner.m_writep = m_writep;
m_owner.bufferFlush();
m_writep = m_owner.m_writep;
}
}
}
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
//=============================================================================
// emit* trace routines
// Note: emit* are only ever called from one place (full* in
// verilated_trace_imp.h, which is included in this file at the top),
// so always inline them.
VL_ATTR_ALWINLINE
void VerilatedVcdBuffer::emitEvent(uint32_t code, VlEvent newval) {
const bool triggered = newval.isTriggered();
// TODO : It seems that untriggered events are not filtered
// should be tested before this last step
2022-11-23 09:08:02 +00:00
if (triggered) {
// Don't prefetch suffix as it's a bit too late;
char* wp = m_writep;
*wp++ = '1';
finishLine(code, wp);
}
}
VL_ATTR_ALWINLINE
void VerilatedVcdBuffer::emitBit(uint32_t code, CData newval) {
// Don't prefetch suffix as it's a bit too late;
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
char* wp = m_writep;
*wp++ = '0' | static_cast<char>(newval);
finishLine(code, wp);
}
VL_ATTR_ALWINLINE
void VerilatedVcdBuffer::emitCData(uint32_t code, CData newval, int bits) {
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
char* wp = m_writep;
*wp++ = 'b';
cvtCDataToStr(wp, newval << (VL_BYTESIZE - bits));
finishLine(code, wp + bits);
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
}
VL_ATTR_ALWINLINE
void VerilatedVcdBuffer::emitSData(uint32_t code, SData newval, int bits) {
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
char* wp = m_writep;
*wp++ = 'b';
cvtSDataToStr(wp, newval << (VL_SHORTSIZE - bits));
finishLine(code, wp + bits);
}
VL_ATTR_ALWINLINE
void VerilatedVcdBuffer::emitIData(uint32_t code, IData newval, int bits) {
char* wp = m_writep;
*wp++ = 'b';
cvtIDataToStr(wp, newval << (VL_IDATASIZE - bits));
finishLine(code, wp + bits);
}
VL_ATTR_ALWINLINE
void VerilatedVcdBuffer::emitQData(uint32_t code, QData newval, int bits) {
char* wp = m_writep;
*wp++ = 'b';
cvtQDataToStr(wp, newval << (VL_QUADSIZE - bits));
finishLine(code, wp + bits);
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
}
VL_ATTR_ALWINLINE
void VerilatedVcdBuffer::emitWData(uint32_t code, const WData* newvalp, int bits) {
int words = VL_WORDS_I(bits);
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
char* wp = m_writep;
*wp++ = 'b';
// Handle the most significant word
const int bitsInMSW = VL_BITBIT_E(bits) ? VL_BITBIT_E(bits) : VL_EDATASIZE;
cvtEDataToStr(wp, newvalp[--words] << (VL_EDATASIZE - bitsInMSW));
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
wp += bitsInMSW;
// Handle the remaining words
while (words > 0) {
cvtEDataToStr(wp, newvalp[--words]);
wp += VL_EDATASIZE;
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
}
finishLine(code, wp);
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
}
VL_ATTR_ALWINLINE
void VerilatedVcdBuffer::emitDouble(uint32_t code, double newval) {
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
char* wp = m_writep;
// Buffer can't overflow before VL_SNPRINTF; we sized during declaration
VL_SNPRINTF(wp, m_maxSignalBytes, "r%.16g", newval);
wp += std::strlen(wp);
finishLine(code, wp);
Improve tracing performance. (#2257) * Improve tracing performance. Various tactics used to improve performance of both VCD and FST tracing: - Both: Change tracing functions to templates to take variable widths as template parameters. For VCD, subsequently specialize these to the values used by Verilator. This avoids redundant instructions and hard to predict branches. - Both: Check for value changes via direct pointer access into the previous signal value buffer. This eliminates a lot of simple pointer arithmetic instructions form the tracing code. - Both: Verilator provides clean input, no need to mask out used bits. - VCD: pre-compute identifier codes and use memory copy instead of re-computing them every time a code is emitted. This saves a lot of instructions and hard to predict branches. The added D-cache misses are cheaper than the removed branches/instructions. - VCD: re-write the routines emitting the changes to be more efficient. - FST: Use previous signal value buffer the same way as the VCD tracing code, and only call the FST API when a change is detected. Performance as measured on SweRV EH1, with the pre-canned CoreMark benchmark running from DCCM/ICCM, clang 6.0.0, Intel i7-3770 @ 3.40GHz, and IO to ramdisk: +--------------+---------------+----------------------+ | VCD | FST | FST separate thread | | (--trace) | (--trace-fst) | (--trace-fst-thread) | ------------+-----------------------------------------------------+ Before | 30.2 s | 121.1 s | 69.8 s | ============+==============+===============+======================+ After | 24.7 s | 45.7 s | 32.4 s | ------------+--------------+---------------+----------------------+ Speedup | 22 % | 256 % | 215 % | ------------+--------------+---------------+----------------------+ Rel. to VCD | 1 x | 1.85 x | 1.31 x | ------------+--------------+---------------+----------------------+ In addition, FST trace size for the above reduced by 48%.
2020-04-13 23:13:10 +00:00
}