verilator/include/verilated_trace_imp.cpp
Geza Lore aa9cde22c8
Use SIMD intrinsics to render VCD traces (#2289)
Use SIMD intrinsics to render VCD traces.

I have measured 10-40% single threaded performance increase with VCD
tracing on SweRV EH1 and lowRISC Ibex using SSE2 intrinsics to render
the trace. Also helps a tiny bit with FST, but now almost all of the FST
overhead is in the FST library.

I have reworked the tracing routines to use more precisely sized
arguments. The nice thing about this is that the performance without the
intrinsics is pretty much the same as it was before, as we do at most 2x
as much work as necessary, but in exchange there are no data dependent
branches at all.
2020-04-30 00:09:09 +01:00

641 lines
24 KiB
C++

// -*- mode: C++; c-file-style: "cc-mode" -*-
//=============================================================================
//
// THIS MODULE IS PUBLICLY LICENSED
//
// Copyright 2001-2020 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 Implementation of tracing functionality common to all trace formats
///
//=============================================================================
// SPDIFF_OFF
// clang-format off
#ifndef VL_DERIVED_T
# error "This file should be included in trace format implementations"
#endif
#include "verilated_intrinsics.h"
#include "verilated_trace.h"
#if 0
# include <iostream>
# define VL_TRACE_THREAD_DEBUG(msg) std::cout << "TRACE THREAD: " << msg << std::endl
#else
# define VL_TRACE_THREAD_DEBUG(msg)
#endif
// clang-format on
//=============================================================================
// Static utility functions
static double timescaleToDouble(const char* unitp) {
char* endp;
double value = strtod(unitp, &endp);
// On error so we allow just "ns" to return 1e-9.
if (value == 0.0 && endp == unitp) value = 1;
unitp = endp;
for (; *unitp && isspace(*unitp); unitp++) {}
switch (*unitp) {
case 's': value *= 1e1; break;
case 'm': value *= 1e-3; break;
case 'u': value *= 1e-6; break;
case 'n': value *= 1e-9; break;
case 'p': value *= 1e-12; break;
case 'f': value *= 1e-15; break;
case 'a': value *= 1e-18; break;
}
return value;
}
static std::string doubleToTimescale(double value) {
const char* suffixp = "s";
// clang-format off
if (value >= 1e0) { suffixp = "s"; value *= 1e0; }
else if (value >= 1e-3 ) { suffixp = "ms"; value *= 1e3; }
else if (value >= 1e-6 ) { suffixp = "us"; value *= 1e6; }
else if (value >= 1e-9 ) { suffixp = "ns"; value *= 1e9; }
else if (value >= 1e-12) { suffixp = "ps"; value *= 1e12; }
else if (value >= 1e-15) { suffixp = "fs"; value *= 1e15; }
else if (value >= 1e-18) { suffixp = "as"; value *= 1e18; }
// clang-format on
char valuestr[100];
sprintf(valuestr, "%3.0f%s", value, suffixp);
return valuestr; // Gets converted to string, so no ref to stack
}
//=============================================================================
// Internal callback routines for each module being traced.
// Each module that wishes to be traced registers a set of callbacks stored in
// this class. When the trace file is being constructed, this class provides
// the callback routines to be executed.
class VerilatedTraceCallInfo {
public: // This is in .cpp file so is not widely visible
typedef VerilatedTrace<VL_DERIVED_T>::callback_t callback_t;
callback_t m_initcb; ///< Initialization Callback function
callback_t m_fullcb; ///< Full Dumping Callback function
callback_t m_changecb; ///< Incremental Dumping Callback function
void* m_userthis; ///< User data pointer for callback
vluint32_t m_code; ///< Starting code number (set later by traceInit)
// CONSTRUCTORS
VerilatedTraceCallInfo(callback_t icb, callback_t fcb, callback_t changecb, void* ut)
: m_initcb(icb)
, m_fullcb(fcb)
, m_changecb(changecb)
, m_userthis(ut)
, m_code(1) {}
~VerilatedTraceCallInfo() {}
};
#ifdef VL_TRACE_THREADED
//=========================================================================
// Buffer management
template <> vluint32_t* VerilatedTrace<VL_DERIVED_T>::getTraceBuffer() {
vluint32_t* bufferp;
// Some jitter is expected, so some number of alternative trace buffers are
// required, but don't allocate more than 8 buffers.
if (m_numTraceBuffers < 8) {
// Allocate a new buffer if none is available
if (!m_buffersFromWorker.tryGet(bufferp)) {
++m_numTraceBuffers;
// Note: over allocate a bit so pointer comparison is well defined
// if we overflow only by a small amount
bufferp = new vluint32_t[m_traceBufferSize + 16];
}
} else {
// Block until a buffer becomes available
bufferp = m_buffersFromWorker.get();
}
return bufferp;
}
template <> void VerilatedTrace<VL_DERIVED_T>::waitForBuffer(const vluint32_t* buffp) {
// Slow path code only called on flush/shutdown, so use a simple algorithm.
// Collect buffers from worker and stash them until we get the one we want.
std::deque<vluint32_t*> stash;
do { stash.push_back(m_buffersFromWorker.get()); } while (stash.back() != buffp);
// Now put them back in the queue, in the original order.
while (!stash.empty()) {
m_buffersFromWorker.put_front(stash.back());
stash.pop_back();
}
}
//=========================================================================
// Worker thread
template <> void VerilatedTrace<VL_DERIVED_T>::workerThreadMain() {
bool shutdown = false;
do {
vluint32_t* const bufferp = m_buffersToWorker.get();
VL_TRACE_THREAD_DEBUG("");
VL_TRACE_THREAD_DEBUG("Got buffer: " << bufferp);
const vluint32_t* readp = bufferp;
while (true) {
const vluint32_t cmd = readp[0];
const vluint32_t top = cmd >> 4;
// Always set this up, as it is almost always needed
vluint32_t* const oldp = m_sigs_oldvalp + readp[1];
// Note this increment needs to be undone on commands which do not
// actually contain a code, but those are the rare cases.
readp += 2;
switch (cmd & 0xF) {
//===
// CHG_* commands
case VerilatedTraceCommand::CHG_BIT_0:
VL_TRACE_THREAD_DEBUG("Command CHG_BIT_0 " << top);
chgBitImpl(oldp, 0);
continue;
case VerilatedTraceCommand::CHG_BIT_1:
VL_TRACE_THREAD_DEBUG("Command CHG_BIT_1 " << top);
chgBitImpl(oldp, 1);
continue;
case VerilatedTraceCommand::CHG_CDATA:
VL_TRACE_THREAD_DEBUG("Command CHG_CDATA " << top);
// Bits stored in bottom byte of command
chgCDataImpl(oldp, *readp, top);
readp += 1;
continue;
case VerilatedTraceCommand::CHG_SDATA:
VL_TRACE_THREAD_DEBUG("Command CHG_SDATA " << top);
// Bits stored in bottom byte of command
chgSDataImpl(oldp, *readp, top);
readp += 1;
continue;
case VerilatedTraceCommand::CHG_IDATA:
VL_TRACE_THREAD_DEBUG("Command CHG_IDATA " << top);
// Bits stored in bottom byte of command
chgIDataImpl(oldp, *readp, top);
readp += 1;
continue;
case VerilatedTraceCommand::CHG_QDATA:
VL_TRACE_THREAD_DEBUG("Command CHG_QDATA " << top);
// Bits stored in bottom byte of command
chgQDataImpl(oldp, *reinterpret_cast<const QData*>(readp), top);
readp += 2;
continue;
case VerilatedTraceCommand::CHG_WDATA:
VL_TRACE_THREAD_DEBUG("Command CHG_WDATA " << top);
chgWDataImpl(oldp, readp, top);
readp += VL_WORDS_I(top);
continue;
case VerilatedTraceCommand::CHG_FLOAT:
VL_TRACE_THREAD_DEBUG("Command CHG_FLOAT " << top);
chgFloatImpl(oldp, *reinterpret_cast<const float*>(readp));
readp += 1;
continue;
case VerilatedTraceCommand::CHG_DOUBLE:
VL_TRACE_THREAD_DEBUG("Command CHG_DOUBLE " << top);
chgDoubleImpl(oldp, *reinterpret_cast<const double*>(readp));
readp += 2;
continue;
//===
// Rare commands
case VerilatedTraceCommand::TIME_CHANGE:
VL_TRACE_THREAD_DEBUG("Command TIME_CHANGE " << top);
readp -= 1; // No code in this command, undo increment
emitTimeChange(*reinterpret_cast<const vluint64_t*>(readp));
readp += 2;
continue;
//===
// Commands ending this buffer
case VerilatedTraceCommand::END: VL_TRACE_THREAD_DEBUG("Command END"); break;
case VerilatedTraceCommand::SHUTDOWN:
VL_TRACE_THREAD_DEBUG("Command SHUTDOWN");
shutdown = true;
break;
//===
// Unknown command
default:
VL_TRACE_THREAD_DEBUG("Command UNKNOWN");
VL_PRINTF_MT("Trace command: 0x%08x\n", cmd);
VL_FATAL_MT(__FILE__, __LINE__, "", "Unknown trace command");
break;
}
// The above switch will execute 'continue' when necessary,
// so if we ever reach here, we are done with the buffer.
break;
}
VL_TRACE_THREAD_DEBUG("Returning buffer");
// Return buffer
m_buffersFromWorker.put(bufferp);
} while (VL_LIKELY(!shutdown));
}
template <> void VerilatedTrace<VL_DERIVED_T>::shutdownWorker() {
// If the worker thread is not running, done..
if (!m_workerThread) return;
// Hand an buffer with a shutdown command to the worker thread
vluint32_t* const bufferp = getTraceBuffer();
bufferp[0] = VerilatedTraceCommand::SHUTDOWN;
m_buffersToWorker.put(bufferp);
// Wait for it to return
waitForBuffer(bufferp);
// Join the thread and delete it
m_workerThread->join();
m_workerThread.reset(nullptr);
}
#endif
//=============================================================================
// Life cycle
template <> void VerilatedTrace<VL_DERIVED_T>::close() {
#ifdef VL_TRACE_THREADED
shutdownWorker();
while (m_numTraceBuffers) {
delete[] m_buffersFromWorker.get();
m_numTraceBuffers--;
}
#endif
}
template <> void VerilatedTrace<VL_DERIVED_T>::flush() {
#ifdef VL_TRACE_THREADED
// Hand an empty buffer to the worker thread
vluint32_t* const bufferp = getTraceBuffer();
*bufferp = VerilatedTraceCommand::END;
m_buffersToWorker.put(bufferp);
// Wait for it to be returned. As the processing is in-order,
// this ensures all previous buffers have been processed.
waitForBuffer(bufferp);
#endif
}
//=============================================================================
// VerilatedTrace
template <>
VerilatedTrace<VL_DERIVED_T>::VerilatedTrace()
: m_sigs_oldvalp(NULL)
, m_timeLastDump(0)
, m_fullDump(true)
, m_nextCode(0)
, m_numSignals(0)
, m_maxBits(0)
, m_scopeEscape('.')
, m_timeRes(1e-9)
, m_timeUnit(1e-9)
#ifdef VL_TRACE_THREADED
, m_numTraceBuffers(0)
#endif
{
set_time_unit(Verilated::timeunitString());
set_time_resolution(Verilated::timeprecisionString());
}
template <> VerilatedTrace<VL_DERIVED_T>::~VerilatedTrace() {
if (m_sigs_oldvalp) VL_DO_CLEAR(delete[] m_sigs_oldvalp, m_sigs_oldvalp = NULL);
while (!m_callbacks.empty()) {
delete m_callbacks.back();
m_callbacks.pop_back();
}
#ifdef VL_TRACE_THREADED
close();
#endif
}
//=========================================================================
// Internals available to format specific implementations
template <> void VerilatedTrace<VL_DERIVED_T>::traceInit() VL_MT_UNSAFE {
m_assertOne.check();
// Note: It is possible to re-open a trace file (VCD in particular),
// so we must reset the next code here, but it must have the same number
// of codes on re-open
const vluint32_t expectedCodes = nextCode();
m_nextCode = 1;
m_numSignals = 0;
m_maxBits = 0;
// Call all initialize callbacks, which will call decl* for each signal.
for (vluint32_t ent = 0; ent < m_callbacks.size(); ++ent) {
VerilatedTraceCallInfo* cip = m_callbacks[ent];
cip->m_code = nextCode();
(cip->m_initcb)(self(), cip->m_userthis, cip->m_code);
}
if (expectedCodes && nextCode() != expectedCodes) {
VL_FATAL_MT(__FILE__, __LINE__, "",
"Reopening trace file with different number of signals");
}
// Now that we know the number of codes, allocate space for the buffer
// holding previous signal values.
if (!m_sigs_oldvalp) m_sigs_oldvalp = new vluint32_t[nextCode()];
#ifdef VL_TRACE_THREADED
// Compute trace buffer size. we need to be able to store a new value for
// each signal, which is 'nextCode()' entries after the init callbacks
// above have been run, plus up to 2 more words of metadata per signal,
// plus fixed overhead of 1 for a termination flag and 3 for a time stamp
// update.
m_traceBufferSize = nextCode() + numSignals() * 2 + 4;
// Start the worker thread
m_workerThread.reset(new std::thread(&VerilatedTrace<VL_DERIVED_T>::workerThreadMain, this));
#endif
}
template <>
void VerilatedTrace<VL_DERIVED_T>::declCode(vluint32_t code, vluint32_t bits, bool tri) {
if (!code) {
VL_FATAL_MT(__FILE__, __LINE__, "", "Internal: internal trace problem, code 0 is illegal");
}
// Note: The tri-state flag is not used by Verilator, but is here for
// compatibility with some foreign code.
int codesNeeded = VL_WORDS_I(bits);
if (tri) codesNeeded *= 2;
m_nextCode = std::max(m_nextCode, code + codesNeeded);
++m_numSignals;
m_maxBits = std::max(m_maxBits, bits);
}
//=========================================================================
// Internals available to format specific implementations
template <> std::string VerilatedTrace<VL_DERIVED_T>::timeResStr() const {
return doubleToTimescale(m_timeRes);
}
template <> std::string VerilatedTrace<VL_DERIVED_T>::timeUnitStr() const {
return doubleToTimescale(m_timeUnit);
}
//=========================================================================
// External interface to client code
template <> void VerilatedTrace<VL_DERIVED_T>::set_time_unit(const char* unitp) {
m_timeUnit = timescaleToDouble(unitp);
}
template <> void VerilatedTrace<VL_DERIVED_T>::set_time_unit(const std::string& unit) {
set_time_unit(unit.c_str());
}
template <> void VerilatedTrace<VL_DERIVED_T>::set_time_resolution(const char* unitp) {
m_timeRes = timescaleToDouble(unitp);
}
template <> void VerilatedTrace<VL_DERIVED_T>::set_time_resolution(const std::string& unit) {
set_time_resolution(unit.c_str());
}
template <> void VerilatedTrace<VL_DERIVED_T>::dump(vluint64_t timeui) {
m_assertOne.check();
if (VL_UNLIKELY(m_timeLastDump && timeui <= m_timeLastDump)) {
VL_PRINTF_MT("%%Warning: previous dump at t=%" VL_PRI64 "u, requesting t=%" VL_PRI64
"u, dump call ignored\n",
m_timeLastDump, timeui);
return;
}
m_timeLastDump = timeui;
Verilated::quiesce();
// Call hook for format specific behaviour
if (VL_UNLIKELY(m_fullDump)) {
if (!preFullDump()) return;
} else {
if (!preChangeDump()) return;
}
#ifdef VL_TRACE_THREADED
// Currently only incremental dumps run on the worker thread
vluint32_t* bufferp = nullptr;
if (VL_LIKELY(!m_fullDump)) {
// Get the trace buffer we are about to fill
bufferp = getTraceBuffer();
m_traceBufferWritep = bufferp;
m_traceBufferEndp = bufferp + m_traceBufferSize;
// Tell worker to update time point
m_traceBufferWritep[0] = VerilatedTraceCommand::TIME_CHANGE;
*reinterpret_cast<vluint64_t*>(m_traceBufferWritep + 1) = timeui;
m_traceBufferWritep += 3;
} else {
// Update time point
flush();
emitTimeChange(timeui);
}
#else
// Update time point
emitTimeChange(timeui);
#endif
// Run the callbacks
if (VL_UNLIKELY(m_fullDump)) {
m_fullDump = false; // No more need for next dump to be full
for (vluint32_t ent = 0; ent < m_callbacks.size(); ++ent) {
VerilatedTraceCallInfo* cip = m_callbacks[ent];
(cip->m_fullcb)(self(), cip->m_userthis, cip->m_code);
}
} else {
for (vluint32_t ent = 0; ent < m_callbacks.size(); ++ent) {
VerilatedTraceCallInfo* cip = m_callbacks[ent];
(cip->m_changecb)(self(), cip->m_userthis, cip->m_code);
}
}
#ifdef VL_TRACE_THREADED
if (VL_LIKELY(bufferp)) {
// Mark end of the trace buffer we just filled
*m_traceBufferWritep++ = VerilatedTraceCommand::END;
// Assert no buffer overflow
assert(m_traceBufferWritep - bufferp <= m_traceBufferSize);
// Pass it to the worker thread
m_buffersToWorker.put(bufferp);
}
#endif
}
//=============================================================================
// Non-hot path internal interface to Verilator generated code
template <>
void VerilatedTrace<VL_DERIVED_T>::addCallback(callback_t initcb, callback_t fullcb,
callback_t changecb,
void* userthis) VL_MT_UNSAFE_ONE {
m_assertOne.check();
if (VL_UNLIKELY(timeLastDump() != 0)) {
std::string msg = (std::string("Internal: ") + __FILE__ + "::" + __FUNCTION__
+ " called with already open file");
VL_FATAL_MT(__FILE__, __LINE__, "", msg.c_str());
}
VerilatedTraceCallInfo* cip = new VerilatedTraceCallInfo(initcb, fullcb, changecb, userthis);
m_callbacks.push_back(cip);
}
//=========================================================================
// Hot path internal interface to Verilator generated code
// These functions must write the new value back into the old value store,
// and subsequently call the format specific emit* implementations. Note
// that this file must be included in the format specific implementation, so
// the emit* functions can be inlined for performance.
template <> void VerilatedTrace<VL_DERIVED_T>::fullBit(vluint32_t* oldp, CData newval) {
*oldp = newval;
self()->emitBit(oldp - m_sigs_oldvalp, newval);
}
template <>
void VerilatedTrace<VL_DERIVED_T>::fullCData(vluint32_t* oldp, CData newval, int bits) {
*oldp = newval;
self()->emitCData(oldp - m_sigs_oldvalp, newval, bits);
}
template <>
void VerilatedTrace<VL_DERIVED_T>::fullSData(vluint32_t* oldp, SData newval, int bits) {
*oldp = newval;
self()->emitSData(oldp - m_sigs_oldvalp, newval, bits);
}
template <>
void VerilatedTrace<VL_DERIVED_T>::fullIData(vluint32_t* oldp, IData newval, int bits) {
*oldp = newval;
self()->emitIData(oldp - m_sigs_oldvalp, newval, bits);
}
template <>
void VerilatedTrace<VL_DERIVED_T>::fullQData(vluint32_t* oldp, QData newval, int bits) {
*reinterpret_cast<QData*>(oldp) = newval;
self()->emitQData(oldp - m_sigs_oldvalp, newval, bits);
}
template <>
void VerilatedTrace<VL_DERIVED_T>::fullWData(vluint32_t* oldp, const WData* newvalp, int bits) {
for (int i = 0; i < VL_WORDS_I(bits); ++i) oldp[i] = newvalp[i];
self()->emitWData(oldp - m_sigs_oldvalp, newvalp, bits);
}
template <> void VerilatedTrace<VL_DERIVED_T>::fullFloat(vluint32_t* oldp, float newval) {
// cppcheck-suppress invalidPointerCast
*reinterpret_cast<float*>(oldp) = newval;
self()->emitFloat(oldp - m_sigs_oldvalp, newval);
}
template <> void VerilatedTrace<VL_DERIVED_T>::fullDouble(vluint32_t* oldp, double newval) {
// cppcheck-suppress invalidPointerCast
*reinterpret_cast<double*>(oldp) = newval;
self()->emitDouble(oldp - m_sigs_oldvalp, newval);
}
//=========================================================================
// Primitives converting binary values to strings...
// All of these take a destination pointer where the string will be emitted,
// and a value to convert. There are a couple of variants for efficiency.
inline static void cvtCDataToStr(char* dstp, CData value) {
#ifdef VL_HAVE_SSE2
// Similar to cvtSDataToStr but only the bottom 8 byte lanes are used
const __m128i a = _mm_cvtsi32_si128(value);
const __m128i b = _mm_unpacklo_epi8(a, a);
const __m128i c = _mm_shufflelo_epi16(b, 0);
const __m128i m = _mm_set1_epi64x(0x0102040810204080);
const __m128i d = _mm_cmpeq_epi8(_mm_and_si128(c, m), m);
const __m128i result = _mm_sub_epi8(_mm_set1_epi8('0'), d);
_mm_storel_epi64(reinterpret_cast<__m128i*>(dstp), result);
#else
dstp[0] = '0' | static_cast<char>((value >> 7) & 1);
dstp[1] = '0' | static_cast<char>((value >> 6) & 1);
dstp[2] = '0' | static_cast<char>((value >> 5) & 1);
dstp[3] = '0' | static_cast<char>((value >> 4) & 1);
dstp[4] = '0' | static_cast<char>((value >> 3) & 1);
dstp[5] = '0' | static_cast<char>((value >> 2) & 1);
dstp[6] = '0' | static_cast<char>((value >> 1) & 1);
dstp[7] = '0' | static_cast<char>(value & 1);
#endif
}
inline static void cvtSDataToStr(char* dstp, SData value) {
#ifdef VL_HAVE_SSE2
// We want each bit in the 16-bit input value to end up in a byte lane
// within the 128-bit XMM register. Note that x86 is little-endian and we
// want the MSB of the input at the low address, so we will bit-reverse
// at the same time.
// Put value in bottom of 128-bit register a[15:0] = value
const __m128i a = _mm_cvtsi32_si128(value);
// Interleave bytes with themselves
// b[15: 0] = {2{a[ 7:0]}} == {2{value[ 7:0]}}
// b[31:16] = {2{a[15:8]}} == {2{value[15:8]}}
const __m128i b = _mm_unpacklo_epi8(a, a);
// Shuffle bottom 64 bits, note swapping high bytes with low bytes
// c[31: 0] = {2{b[31:16]}} == {4{value[15:8}}
// c[63:32] = {2{b[15: 0]}} == {4{value[ 7:0}}
const __m128i c = _mm_shufflelo_epi16(b, 0x05);
// Shuffle whole register
// d[ 63: 0] = {2{c[31: 0]}} == {8{value[15:8}}
// d[126:54] = {2{c[63:32]}} == {8{value[ 7:0}}
const __m128i d = _mm_shuffle_epi32(c, 0x50);
// Test each bit within the bytes, this sets each byte lane to 0
// if the bit for that lane is 0 and to 0xff if the bit is 1.
const __m128i m = _mm_set1_epi64x(0x0102040810204080);
const __m128i e = _mm_cmpeq_epi8(_mm_and_si128(d, m), m);
// Convert to ASCII by subtracting the masks from ASCII '0':
// '0' - 0 is '0', '0' - -1 is '1'
const __m128i result = _mm_sub_epi8(_mm_set1_epi8('0'), e);
// Store the 16 characters to the un-aligned buffer
_mm_storeu_si128(reinterpret_cast<__m128i*>(dstp), result);
#else
cvtCDataToStr(dstp, value >> 8);
cvtCDataToStr(dstp + 8, value);
#endif
}
inline static void cvtIDataToStr(char* dstp, IData value) {
#ifdef VL_HAVE_AVX2
// Similar to cvtSDataToStr but the bottom 16-bits are processed in the
// top half of the YMM registerss
const __m256i a = _mm256_insert_epi32(_mm256_undefined_si256(), value, 0);
const __m256i b = _mm256_permute4x64_epi64(a, 0);
const __m256i s = _mm256_set_epi8(0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2,
2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3);
const __m256i c = _mm256_shuffle_epi8(b, s);
const __m256i m = _mm256_set1_epi64x(0x0102040810204080);
const __m256i d = _mm256_cmpeq_epi8(_mm256_and_si256(c, m), m);
const __m256i result = _mm256_sub_epi8(_mm256_set1_epi8('0'), d);
_mm256_storeu_si256(reinterpret_cast<__m256i*>(dstp), result);
#else
cvtSDataToStr(dstp, value >> 16);
cvtSDataToStr(dstp + 16, value);
#endif
}
inline static void cvtQDataToStr(char* dstp, QData value) {
cvtIDataToStr(dstp, value >> 32);
cvtIDataToStr(dstp + 32, value);
}
#define cvtEDataToStr cvtIDataToStr