3/22/18 timer queue

	HANDLE	m_hTimerQueue;
	HANDLE	m_hTimer;

	static	VOID CALLBACK TimerCallback(PVOID lpParameter, BOOLEAN bTimerFired);

	m_hTimerQueue = CreateTimerQueue();
	CreateTimerQueueTimer(&m_hTimer, m_hTimerQueue, TimerCallback, this, 0, 50, 0);

VOID CALLBACK CPolymeterDlgTestDlg::TimerCallback(PVOID lpParameter, BOOLEAN bTimerFired)
{
	CWnd	*pWnd = reinterpret_cast<CWnd*>(lpParameter);
	pWnd->PostMessage(WM_TIMER, 100);
}


Seems to interfere with MIDI stream thread, causing unacceptable delays queueing output.

4/5/18 short duration related glitching and dropouts

"duration 1 bug organ offset -1 is worse"

Seems to be OK on SD-20 with latency of 20 or 16 but not 10ms; what are the corresponding callback lengths?
Looks the issue is that 10ms at 24 PPQ gets you a callback length of 1 tick, which mmsystem doesn't like
latency >= 16ms gets you at least 2 ticks
with duration = -9 and PPQ = 24, dropouts and glitches are still occurring even at latency = 35 ms (3 ticks), on both Win GS and SD-20
sequencer dump doesn't seem to show anything suspicious
need a hypothesis! sure acts like some type of fencepost error, or issue with event sorting

4/9/18 attempted piano masking for testdoc2b edirol; still too busy?

In more recent versions, you'll need this: int nEvtStart = iPairQuant;

if (iTrack == 12 || iTrack == 14 || iTrack == 17) {
	int	a = (nEvtStart * 2) % (4 * 12);
	if (iTrack == 12 && a >= (4 * 8 - 1) && a < (4 * 12 - 2))
		goto yack;
	if (iTrack == 14 && a >= (4 * 9 - 1) && a < (4 * 11))
		goto yack;
	if (iTrack == 17 && a >= (4 * 9 - 1) && a < (4 * 11))
		goto yack;
}
...
						m_arrNoteOff.InsertSorted(evt);
					}
yack:;//@@@ NOTE position!
				}
				iEvt++;


the mask on the C# disappoints... even resuming at 4 * 10 feels weak in places... perhaps it's too sparse to mask effectively!

4/10 WAY BETTER mask the middle F# instead (and maybe octave up the C# and the high F# every now and then)

if (iTrack == 12 || iTrack == 13 || iTrack == 14) {
	int	a = (nEvtStart * 2) % (4 * 12);
	if (iTrack == 12 && a >= (4 * 8 - 1) && a < (4 * 12 - 2))
		goto yack;
	if (iTrack == 13 && a >= (4 * 0) && a < (4 * 4 - 2))
		goto yack;
	if (iTrack == 14 && a >= (4 * 9 - 1) && a < (4 * 11))
		goto yack;
}

4/11 switch low piano note between F#2 and D2 is nice; also octave up G#

int n = trk.m_nNote;
if (iTrack == 12) {
	int	a = (nEvtStart * 2) % (4 * 12);
	if (a >= (4 * 8 - 1) && a < (4 * 12 - 2))
		goto yack;
	if (a >= (4 * 6) && a < (4 * 8 - 1))
		n -= 4;
} else if (iTrack == 13) {
	int	a = (nEvtStart * 2) % (4 * 12);
	if (a >= (4 * 0) && a < (4 * 4 - 2))
		goto yack;
} else if (iTrack == 14) {
	int	a = (nEvtStart * 2) % (4 * 12);
	if (a >= (4 * 9 - 1) && a < (4 * 11))
		goto yack;
} else if (iTrack == 16) {
	int	a = (nEvtStart * 2) % (4 * 12);
	if (a >= (4 * 10 - 2) && a < (4 * 10 - 1))
		n += 12;
}

***NOTE*** The shortest piano loop's duration (30) causes overlapped notes which glitch on the GS synth but not the Edirol.
This problem should go away when you implement legato tracks; the piano durations won't be needed anymore and will be zero.

Works OK with tied notes! Transpose down a step, "testdoc2b test ties F- fixed agogo.plm"
Also add a mask to change Ab to Bb when the low note drops from F to Db.

int n = trk.m_nNote;
if (iTrack == 12) {
	int	a = (nEvtStart * 2) % (4 * 12);
	if (a >= (4 * 8 - 1) && a < (4 * 12 - 2))
		goto yack;
	if (a >= (4 * 6) && a < (4 * 8 - 1))
		n -= 4;
} else if (iTrack == 13) {
	int	a = (nEvtStart * 2) % (4 * 12);
	if (a >= (4 * 0) && a < (4 * 4 - 2))
		goto yack;
} else if (iTrack == 14) {
	int	a = (nEvtStart * 2) % (4 * 12);
	if (a >= (4 * 9 - 1) && a < (4 * 11))
		goto yack;
} else if (iTrack == 15) {
	int	a = (nEvtStart * 2) % (4 * 12);
	if (a >= (4 * 6) && a < (4 * 10 - 1))
		n += 2;
} else if (iTrack == 16) {
	int	a = (nEvtStart * 2) % (4 * 12);
	if (a >= (4 * 9) && a < (4 * 10 - 1))
		n += 12;
}

4/9/18 super nasty undo CString bug

_CrtSetDbgFlag(_CRTDBG_ALLOC_MEM_DF | _CRTDBG_CHECK_ALWAYS_DF | _CRTDBG_CHECK_CRT_DF);//@@@ DEBUG 

32-bit version doesn't seem to have this bug; alignment related?

Good guess! The CRefObj struct's 32-bit reference count was causing confusion over the 
string's address. CRefObj claimed to have a size of 20 bytes, with m_Str at offset 12,
but the CStringT constructor wrote the 8-byte string pointer at offset 16, munging memory
and causing m_Str (at offset 12) to contain a bogus address consisting of four bytes of 
trash in the low DWORD and the low half of the string address in the high DWORD.

Why this is only being discovered now is unclear. Do similar reference counted objects
in other x64 apps also suffer from this bug, but we're somehow getting away with it?
Better check PotterDraw at least. The bug was not always reliably repeatable.

class CRefString : public CRefObj {
public:
	CString m_str; 
};

previous x64 layout:
offset	size	var
0	8 	vtable pointer
8	4 	ref count
12	8 	CString pointer
total size: 20

new x64 layout:
offset	size	var
0	8 	vtable pointer
8	8 	ref count
16	8 	CString pointer
total size: 24

x32 layout:
offset	size	var
0	4 	vtable pointer
4	4 	ref count
8	4 	CString pointer
total size: 12

HAH. It's the #pragma pack(push, 1) at the top of UndoState.h. This explains
why there's no issue in PotterDraw. The failure only occurs when the string 
undo class is defined within CUndoState. The pragma is a legacy from when the
undo code and ID were 16-bit. Removing it fixes the bug. Moral: avoid using
non-default packing.

4/11/18 thread-safe trigger class; don't need it but it could come in handy

class CTrigger {
public:
	CTrigger();
	void	Reset();
	void	Set();
	bool	Get();

protected:
	int		m_nWrites;		// number of writes
	int		m_nReads;		// number of reads
};

inline CTrigger::CTrigger()
{
	Reset();
}

inline void CTrigger::Reset()
{
	m_nWrites = 0;
	m_nReads = 0;
}

inline void CTrigger::Set()
{
	m_nWrites++;
}

inline bool CTrigger::Get()
{
	if (m_nWrites == m_nReads)
		return false;
	m_nReads = m_nWrites;
	return true;
}


4/11/18 test sequencer position conversion methods

for (int i = 0; i < 10000000; i++) {
	LONGLONG	nPos = rand() - RAND_MAX / 2;
	LONGLONG	nTick, nBeat, nPos2;
	m_Seq.PositionToBeat(nPos, nBeat, nTick);
	m_Seq.BeatToPosition(nBeat, nTick, nPos2);
	if (nPos2 != nPos) {
		printf("%lld != %lld \n", nPos2, nPos);
		printf("%lld:%03lld\n", nTick, nBeat);
	}
}

4/12/18 improved sequencer track event loop also fixes positive offset & swing bug

The old event loop handled positive offset with positive swing incorrectly, swinging even beats instead of odd ones.
The bug went undected for a while because positive offset is rare, but it's obvious when sequencer dumps are compared.
It helps to use a big latency (e.g. one second) because this increases the number of events per callback.
The new method fixes this bug, and also handles negative song position. This was a total bear!

NOTE that prior to version 33c, the sequencer dump had a SEVERE bug which caused it to output the entire MIDI
output buffer every time, regardless of how many events the buffer actually contained, resulting in enormous 
sequencer dump files consisting of mostly garbage. Bug in the debug code, boo!

4/14/18 naming confusion

Stop confusing beats with steps! Beats are QUARTER NOTES. It's a STEP SEQUENCER, not a beat sequencer.

4/15/18 can't open a MIDI output device for live playback while it's open for streaming, duh!

4/17/18	keeping output device open all the time

The main justification is that MIDI thru currently only works during playback, which is unexpected and annoying.
A secondary justification is it might significantly reduce the lag between pressing play and playback starting.

The output callback would move from the sequencer to the app. Instead of passing the sequencer's
"this" to midiStreamOpen, the callback would get the sequencer pointer some other way, presumably
a sequencer pointer in the app. The app would also need a dummy CSequencer containing no tracks,
which the sequencer pointer would point as a last resort if no documents exist. Everything would
stay the same in CSequencer except the MIDI device handle would move to the app, and the sequencer
would no longer open or close the device, though it would still start and stop it.

benchmarks (Release version, running standalone, not from VS2012 IDE)
all time in milliseconds

device		device open	other open	device close	other close
MS GS WT Synth	130 to 150	0.15 to 0.2 	20 to 50	1.5 to 2
Edirol SD-20	0.2 to 3	6		22		22

Drastic differences! For the Edirol, stopping takes 10x longer than starting, the opposite of the
MS Synth. This means there would be no gain in terms of playback startup lag. Also, stopping the 
device takes just as long as closing it. Bizarre. This is why we benchmark things.

This result considerably weakens the secondary justification. Now the argument comes down to how
inconvenient it actually is that MIDI thru doesn't work all the time. But keep in mind that we
would still need to restart the device on play/stop, which means there would still be a glitch
in the MIDI thru coverage. However device state (such as patches, volumes, etc.) probably won't
be lost, though this theory should definitely be tested before proceeding.

4/20/18 popup edit controls and editing command accelerators (Ctrl+X etc.)

Currently, using an editing key (such as Ctrl+X) in a popup edit control (e.g. in the channels
bar) unexpectedly operates on the track array. This is a bug but it must be tolerated for now.
The view handlers (and update UI handlers) for the editing commands (cut, copy, paste, etc.) 
need to call the corresponding CFocusEdit members, but that's currently not possible because
CTrackDlg displays edit controls all the time. If CFocusEdit handling were added now, the track
editing accelerators wouldn't work when the focus was on a track edit control (nearly always).
However when the CTrackDlg row dialog scheme gets replaced with a grid control, it will become
possible to call CFocusEdit and the problem will go away. This accelerator routing confusion is
another reason (besides slow performance) why row dialogs suck and grid controls are always
preferable to them.

4/21/18 installer GUIDs

ProductName: Polymeter (x86)
ProductCode: {0C9FC2D8-0E07-4DEA-A384-A9BF1E267258}
UpgradeCode: {5B628FFB-0D67-4F8C-AC44-626C4AE61209}

ProductName: Polymeter
ProductCode: {B18FD079-013F-41D1-95D8-FEE34A4BEC8A}
UpgradeCode: {4A526D06-C4C2-4208-B716-C687C92A7A71}

4/24/18 test for time to repeat method

Note that this is impractically slow for large sets of big numbers, even in Release

void TestTimeToRepeat()
{
	srand(GetTickCount());	// seed random
	int	nPasses = 1000;
	int	nTimeDiv = 120;
	int	nSels = 3;
	for (int iPass = 0; iPass < nPasses; iPass++) {
		CArrayEx<ULONGLONG, ULONGLONG>	arrDuration;
		arrDuration.Add(nTimeDiv);	// avoids fractions when LCM result is converted from ticks back to beats
		for (int iSel = 0; iSel < nSels; iSel++) {
			ULONGLONG	nDuration = (rand() & 0x1ff) + 1;
			if (arrDuration.BinarySearch(nDuration) < 0)	// eliminate duplicates to avoid needless LCM
				arrDuration.InsertSorted(nDuration);	// ascending sort may improve LCM performance
		}
		int	nDurs = arrDuration.GetSize();
		ULONGLONG	nLCM = arrDuration[0];
		printf("%d: %llu", iPass, arrDuration[0]);
		for (int iDur = 1; iDur < nDurs; iDur++) {	// for each track duration
			printf(", %llu", arrDuration[iDur]);
			nLCM = CNumberTheory::LeastCommonMultiple(nLCM, arrDuration[iDur]);
		}
		printf(" = %llu\n", nLCM);
		ULONGLONG	nCnt = 1;
		while (1) {
			int	iDur;
			for (iDur = 0; iDur < nDurs; iDur++) {	// for each track duration
				if (nCnt % arrDuration[iDur])	// if count isn't evenly divisible by duration
					break;	// eliminate this count
			}
			if (iDur >= nDurs)	// if count is evenly divisible by all durations
				break;	// we're done
			nCnt++;
		}
		if (nCnt != nLCM) {
			printf("FAIL! %llu\n", nCnt);
			break;
		}
	}
}

4/25/18 MIDI note

in CTrackView.cpp:

#include "Note.h"

		case COL_Note:
			{
				CNote	note(pDoc->m_Seq.GetNote(iTrack));
				_tcscpy_s(item.pszText, item.cchTextMax, note.MidiName());
			}
			break;

		#define TRACKDEF_EXCLUDE_NOTE	// exclude note
		#include "TrackDef.h"	// generate column definitions

		case COL_Note:
			pEdit->SetRange(0, 127);
			pEdit->SetNoteEntry(true);
			break;

4/27/18 reordering grid control columns

Setting LVS_EX_HEADERDRAGDROP works, but if column order is changed from default, tabbing is out of order.

4/30/18	MFC CArray array resizing algorithm

Instead of wading through the code every time, here it is:

if (nGrowBy == 0)
{
	// heuristically determine growth when nGrowBy == 0
	//  (this avoids heap fragmentation in many situations)
	nGrowBy = m_nSize / 8;
	nGrowBy = (nGrowBy < 4) ? 4 : ((nGrowBy > 1024) ? 1024 : nGrowBy);
}

In other words, grow by an eighth of current size (12.5%), but never less than four or more than 1K elements.

5/1/18 convert binary notes to default velocities

in CPolymeterDoc::ReadProperties:
		CPersist::GetBinary(sTrkID, RK_TRACK_EVENT, trk.m_arrEvent.GetData(), &nUsed);

		if (m_nFileVersion == 0) {//@@@
			int	nEvents = trk.m_arrEvent.GetSize();
			for (int iEvt = 0; iEvt < nEvents; iEvt++) {
				if (trk.m_arrEvent[iEvt] == 1)
					trk.m_arrEvent[iEvt] = DEFAULT_VELOCITY;
			}
		}

5/1/18 gradient fill looks awesome but it's around 20x to 50x slower than FillSolidRect (no surprise)

	struct TRICOLOR {//@@@
		COLOR16	r;
		COLOR16	g;
		COLOR16	b;
	};
	TRICOLOR arrTriColor[_countof(clrStep)];
	for (int i = 0; i < _countof(clrStep); i++) {
		arrTriColor[i].r = COLOR16(round(GetRValue(clrStep[i]) / 255.0 * 0xffff));
		arrTriColor[i].g = COLOR16(round(GetGValue(clrStep[i]) / 255.0 * 0xffff));
		arrTriColor[i].b = COLOR16(round(GetBValue(clrStep[i]) / 255.0 * 0xffff));
	}
	static GRADIENT_RECT	rGrad = {0, 1};//@@@


CBenchmark b;//@@@
#if 0
						const TRICOLOR& clr1 = arrTriColor[iStepColor];
						const TRICOLOR& clr2 = arrTriColor[iCurPosColor];
						TRIVERTEX	tv[2] = {
							{rStep.left, rStep.top, clr1.r, clr1.g, clr1.b}, 
							{rStep.right, rStep.bottom, clr2.r, clr2.g, clr2.b},
						};
						pDC->GradientFill(tv, 2, &rGrad, 1, GRADIENT_FILL_RECT_H);
#else
						rStep.right = x1 + nStepWidth / 2;
						pDC->FillSolidRect(rStep, clrStep[iStepColor]);
						rStep.left = rStep.right;
						rStep.right = x2;
						pDC->FillSolidRect(rStep, clrStep[iCurPosColor]);
#endif
printf("%f\n", b.Elapsed());

Tried doing gradient fill to off-screen bitmap and then blitting it into place but that's actually SLOWER!
Wonder if triangle fill would be faster? NOPE, pair of triangles is SLOWER esp. for large steps.
HOWEVER benchmarking shows that using gradient fill only slows OnDraw by around 3x. Why? Why isn't it 20x or 50x?
Hypothesis: something else in OnDraw is taking more time than you'd expect.
file: heavy graphics.plm
The FIRST step takes 2x to 4x longer to draw than the others. WTF?
That's because of delayed painting and accumulation of the invalid rectangle.
If you move the current position from the first step to the last step, ALL the steps in between are repainted.
Only two steps actually needed to be painted. There's no easy solution to this.

USHORT MakeColor16(BYTE nColor)//@@@
{
	return static_cast<USHORT>(round(static_cast<double>(nColor) / BYTE_MAX * SHRT_MAX));
}

		static GRADIENT_RECT	rGrad = {0, 1};//@@@
		COLORREF	clr1 = m_arrStepColor[iStepColor];
		COLORREF	clr2 = m_arrStepColor[iCurPosColor];
		TRIVERTEX	tv[2] = {
			{rStep.left, rStep.top, MakeColor16(GetRValue(clr1)), MakeColor16(GetGValue(clr1)), MakeColor16(GetBValue(clr1))}, 
			{rStep.right, rStep.bottom, MakeColor16(GetRValue(clr2)), MakeColor16(GetGValue(clr2)), MakeColor16(GetBValue(clr2))},
		};
		pDC->GradientFill(tv, 2, &rGrad, 1, GRADIENT_FILL_RECT_H);

5/2/18 tied swing collision

simple demo track: tempo = 60, length = 3, quant = 60, offset = 0, swing = 10, all 3 steps set and 1st step tied
set m_nCBLen = 4000 in UpdateCallbackLength to make problem easier to see in sequencer dump file

sacred ground:
if (nDurSteps & 1) {	// if odd duration
	if (bIsOdd)	// if odd step
		nDuration -= nSwing;	// subtract swing from duration
	else	// even step
		nDuration += nSwing;	// add swing to duration
}


5/14/18 tooltips suck

in PreTranslateMessage
	m_ToolTip.RelayEvent(pMsg);//@@@

in OnSize
	m_ToolTip.DelTool(this, 1);//@@@
	m_ToolTip.AddTool(this, LPSTR_TEXTCALLBACK, CRect(0, 0, cx, cy), 1);

but there are too many crap cases

void CVelocityView::UpdateToolTip(CPoint point)
{
	CRect	rc;
	GetClientRect(rc);
	int	nVal = round((1 - static_cast<double>(point.y) / rc.Height()) * MIDI_NOTE_MAX);
	if (nVal != m_nToolTipVal) {
		CString	s;
		s.Format(_T("%d"), nVal);
		m_ToolTip.UpdateTipText(s, this, 1);
//		m_ToolTip.Update();
		m_nToolTipVal = nVal;
	}
}

5/15/18 next/prev pane mystery

Default handling works but only when track view has focus; why doesn't it work for step parent view and its children?
The commands are definitely being dispatched, per the following test in CChildFrm::OnCmdMsg:

	case ID_NEXT_PANE:
		if (nCode == CN_COMMAND)
			printf("OnCmdMsg next pane\n");//@@@
		break;
	case ID_PREV_PANE:
		if (nCode == CN_COMMAND)
			printf("OnCmdMsg prev pane\n");//@@@
		break;

5/17/18 serious bug in CMidiOut::GetDeviceNames

MMRESULT mr = midiOutGetDevCaps(iDev, &caps, sizeof(MIDIINCAPS)); <-- WRONG!

MMRESULT mr = midiOutGetDevCaps(iDev, &caps, sizeof(MIDIOUTCAPS)); <-- CORRECT!

sizeof(MIDIINCAPS) = 76
sizeof(MIDIOUTCAPS) = 84

So you were passing a struct eight bytes too small. This might explain Ben's issue, VirtualMIDISynth showing blank device names.
Amazing it didn't get discovered sooner.

5/17/18 sine wave velocity function

double	x = round((iStep + 0.5) * fStepWidth) - x1;	// compute x at center of bar

double	fPhase = x / nDeltaX;
y = y1 + (sin(fPhase * M_PI * 2) + 1) * nDeltaY / 2;

Input gesture determines amplitude, frequency, and direction (phase is either 0 or 0.5).
Easily adapted for other waveforms.

5/19/18 how to record Windows GS Synth

In Control Panel / Sound:
In Playback tab, select "CABLE Input" and Set Default.
In Recording tab, select "CABLE Output" and Set Default;
then click Properties, and in Listen tab, check "Listen to this device",
and if you're using the USB dongle, set "Playback through this device" to "USB audio codec".
In WaveShop, in Options / Audio, record from "Cable Output".

5/22/18 track unique identifier (UID)

Seems to work and passes undo test.

dump:
	int nTracks = GetTrackCount();
	for (int iTrack = 0; iTrack < nTracks; iTrack++)
		printf("%d:%d ", iTrack, m_Seq.GetID(iTrack));
	printf("\n");

5/23/18 dub test

#if 1 //@@@
		CTrack&	trk = GetAt(0);
		trk.m_arrDub.SetSize(4);
		CDub	dub;
		dub.m_nTime = 0;
		dub.m_bMute = 0;
		trk.m_arrDub[0] = dub;
		dub.m_nTime = 240;
		dub.m_bMute = 1;
		trk.m_arrDub[1] = dub;
		dub.m_nTime = 480;
		dub.m_bMute = 0;
		trk.m_arrDub[2] = dub;
		dub.m_nTime = 960;
		dub.m_bMute = 1;
		trk.m_arrDub[3] = dub;
		trk.m_iDub = 0;//@@@ depends on start pos!
		m_bIsSongPlayback = true;//@@@ test ONLY
#endif//@@@

5/25/18 song position hint; mind keeps changing on this

	class CSongPosHint : public CObject {
	public:
		CSongPosHint(LONGLONG nPos) : m_nPos(nPos) {}
		LONGLONG	m_nPos;		// song position in absolute ticks
	};

			CSongPosHint	hint(nPos);
			UpdateAllViews(NULL, HINT_SONG_POS, &hint);

			const CPolymeterDoc::CSongPosHint *pPosHint = static_cast<CPolymeterDoc::CSongPosHint *>(pHint);

5/31/18 critical path OnTimer benchmarks

sequencer's GetPosition only: 4us
UpdateAllViews, with recipients disabled: 27us, with spikes to 120us or so
add status bar updates: 600us to 1ms
only one status bar update: 580us
step view song pos update enabled, but no status bar update: 150 to 300us (full screen, Ala Aye)
so status bar takes more time than step view!
song view song pos update enabled, but no status bar update: 650us with a spike to 2ms when view scrolls

Conclusions:

Status bar overhead is excessive, and it's probably due to strdup and free in CMFCStatusBar::SetPaneText.
You could override it, and just copy directly to _GetPanePtr(nIndex)->lpszText is the buffer is big enough.
This only helps if most of the overhead is the malloc/free and not the window update. It also only helps
if the position and time strings typically stay the same size in characters during playback (and they do).

Song view is MUCH slower than step view, probably because moving the song position cursor is so inefficient
(it erases the old cursor by repainting all the cells underneath it).

6/1/18 CTrack::FindDub unit test

	CTrack trk;
	CTrack::CDub	a(0, 1);
	trk.m_arrDub.Add(a);
	CTrack::CDub	b(5, 0);
	trk.m_arrDub.Add(b);
	CTrack::CDub	c(10, 1);
	trk.m_arrDub.Add(c);
	for (int i = 0; i < 10; i++) {
		printf("%d %d\n", i, trk.FindDub(i));
	}

0	0
1	0
2	0
3	0
4	0
5	1
6	1
7	1
8	1
9	1
10	2


6/4/18 GM patch names

already added to MidiCtrlDef.h

const LPCTSTR CChannelsBar::m_arrGMPatchName[MIDI_NOTES + 1] = {	// one extra for none
	_T("<None>"),
	#define MIDI_PATCH_DEF(name) _T(name),
	#include "MidiCtrlrDef.h"
};

		case COL_GMPatch:
			{
				int	nPatch = pChan->GetProperty(CChannel::PROP_Patch) + 1;
				ASSERT(nPatch >= 0 && nPatch <= MIDI_NOTES);
				_tcscpy_s(item.pszText, item.cchTextMax, m_arrGMPatchName[nPatch]);
			}
			break;

Watch out for update trouble! Editing patch number must update drop list and vice versa. PITA!

6/10/18	23,000 lines of code and counting! (Cloc1.76)

6/15/18 pack and unpack bool array to bit array: unit test

	CByteArrayEx	src, dst, src2;
	for (int i = 0; i < 100000; i++) {
		int	nBits = (rand() & 0xfff) + 1;
		src.SetSize(nBits);
		for (int j = 0; j < nBits; j++) {
			src[j] = rand() & 1;
		}
		CPreset::PackBools(src, dst);
		CPreset::UnpackBools(dst, nBits, src2);
		if (src != src2)
			printf("bad!\n");
	}

6/19/18 full screen

	EnableMDITabs(false); // this has to be undone on exit
	EnableFullScreenMode(0);
	EnableFullScreenMainMenu(FALSE);
	ShowFullScreen();
	
also must hide popup exit dialog

6/21/18	keep view type consistent across all documents? not sure

	if (0) {
		// propagate view type to other documents
		CAllDocIter	iter;	// iterate all documents
		CPolymeterDoc	*pDoc;
		while ((pDoc = STATIC_DOWNCAST(CPolymeterDoc, iter.GetNextDoc())) != NULL) {
			POSITION	pos = pDoc->GetFirstViewPosition();
			if (pos != NULL) {
				CView	*pView = pDoc->GetNextView(pos);
				CChildFrame	*pFrame = DYNAMIC_DOWNCAST(CChildFrame, pView->GetParentFrame());
				if (pFrame != NULL && pFrame != this) {	// if child frame and not us
					pFrame->SetViewType(nViewType);	// propagate view type
				}
			}
		}
	}

6/22/18 bar graph showing position of each "part" in live view

It's fun to watch but more distracting than helpful. Note that grouped parts aren't 
handled properly; should be computing the time to repeat for all the part's members
(computationally expensive, would have to be done beforehand, not per song position).

		case CPolymeterDoc::HINT_SONG_POS:
			{
				CPolymeterDoc::CSongPosHint	*pSongPosHint = static_cast<CPolymeterDoc::CSongPosHint*>(pHint);
				CClientDC	dc(this);
				int	nItems = m_arrPart.GetSize();
				int	cy = 32;
				for (int iItem = 0; iItem < nItems; iItem++) {
					int	iPart = m_arrPart[iItem];
					int	nLength;
					if (iPart & PART_GROUP_MASK) {
						const CTrackGroup&	part = pDoc->m_arrPart[iPart & ~PART_GROUP_MASK];
						int	iTrack = part.m_arrTrackIdx[0];
						nLength = pDoc->m_Seq.GetLength(iTrack) * pDoc->m_Seq.GetQuant(iTrack);
					} else {
						nLength = pDoc->m_Seq.GetLength(iPart) * pDoc->m_Seq.GetQuant(iPart);
					}
					int	nPos = static_cast<int>(pSongPosHint->m_nSongPos);
					nPos = nPos % nLength;
					nPos = round(double(nPos) / nLength * LIST_WIDTH);
					int	x1 = (LIST_WIDTH + LIST_GUTTER) * LISTS;
					int	y1 = iItem * cy + 35;
					dc.FillSolidRect(x1, y1, nPos, cy, RGB(255, 0, 255)); 
					dc.FillSolidRect(x1 + nPos, y1, LIST_WIDTH - nPos, cy, RGB(0, 255, 255)); 
				}
			}
			break;

It's just too much dynamic and mostly irrelevant information. Probably a simple beat
counter with adjustable meter will give much better results. [Not true! Fix groups and see if it helps?].

7/2/18 modulation cut/paste bug

Flunks undo test & it's easy to see why. To replicate: 
1) open modulation move test doc (two tracks, 2nd modulated by 1st)
2) select all
3) copy
4) paste
5) undo
6) redo
and sure enough the redone paste loses its modulations.
Hopefully solution is: clipboard stores modulations as UIDs rather than track indices.

7/23/18 CDiatonic::ParseMidiNoteName doesn't handle Cb correctly in key of Gb

	for (int iKey = 0; iKey < 12; iKey++) {
		for (int iNote = 0; iNote < 128; iNote++) {
			CString	s(CNote(iNote).MidiName(iKey));
			CNote	n(0);
			n.ParseMidi(s);
			if (n != iNote)	// if round trip error
				_tprintf(_T("%d %d %s\n"), iKey, iNote, s);
		}
	}

above test fails for Cb in key of Gb (parsed Cb is an octave too low); all other keys/notes are fine

9/5/18 Windows GS Synth's channel volume behavior

For a single sustained organ note (Drawbar Organ, E4, velocity 64), doubling its volume in the channel bar 
increases the output by 12 dB, as measured using WaveShop's meter. This holds over a wide range of volumes.
For example 50 -> 100, 30 -> 60, 25 -> 50, 64 -> 127, and so on.
This means that the volume control is linear, not log (log velocity changes give linear change in dB).
Hence adding a constant to all the 
channel volumes is INCORRECT unless they all have the same initial values.
Instead you should SCALE the channel volumes by a PERCENTAGE.
Tests also show that the same principle applies to VELOCITY!

dB	scale
12	2	(2**1)
9	1.6818	(2**0.75)
6	1.4142 	(2**0.5)
3	1.1892	(2**0.25)
1.5	1.0905	(2**0.125)
2	1.1225	(2**0.1666)

So for example if a track has a velocity of 64 and you want it 3dB quieter, 64 / 1.1892 = 54.
Also note panning uses 3dB compensation, i.e. if centered signal is -12dB on both channels, hard left will be -9dB.
This means for hard left or right you may need to add an extra 3dB to get same perceived loudness (see 6dB pan law).

actual relationship is: y = 20 * LOG10(x / 127)
where x is the MIDI velocity and y is the voltage amplitude in decibels. 
Note this is voltage not power (20 not 10) hence each doubling of x adds 6 to y.

9/18/18 song MIDI export is broken

Part of the problem is that dubs must be reset prior to processing each track.
Adding the following to ExportImpl helps, but doesn't solve entire problem:

		int	nPadTime = 0;	// after this line
		if (m_bIsSongMode)
			ChaseDubs(m_nStartPos, true);

Fixed by processing tracks within chunk loop and using same chunk size as when running live, in order to exactly match live output.

10/17/18 For velocity offset to give smooth change in dB, must use exponential waveform

For example if velocity should quadruple from 16 to 64, use ramp with Curviness of 4.

11/9/18 do any existing documents have modulated modulations?

According to the test code below, no, they don't, provided recursive modulation is limited only
to source tracks having a track type of Modulator. This is likely a reasonable limitation anyway.

	CWaitCursor	wc;	// can take a while
	CFileFind	ff;
//	LPCTSTR	szPath = _T("C:\\Chris\\MyProjects\\Polymeter\\docs");
	LPCTSTR	szPath = _T("C:\\Chris\\MyProjects\\Polymeter\\docs\\test");
	BOOL bWorking = ff.FindFile(CString(szPath) + _T("\\*.*"));
	int	nDocs = 0;
	int	nHits = 0;
	while (bWorking) {
		bWorking = ff.FindNextFile();
		CString	sExt(PathFindExtension(ff.GetFileName()));
		if (!sExt.CompareNoCase(_T(".plm"))) {
			CPolymeterDoc	doc;
			doc.ReadProperties(ff.GetFilePath());
			nDocs++;
			int	nModMods = 0;
			int	nTracks = doc.GetTrackCount();
			for (int iTrack = 0; iTrack < nTracks; iTrack++) {
				const CTrack&	trk = doc.m_Seq.GetTrack(iTrack);
				int	nMods = trk.m_arrModulator.GetSize();
				for (int iMod = 0; iMod < nMods; iMod++) {
					const CModulation&	mod = trk.m_arrModulator[iMod];
					if (mod.m_iSource >= 0) {	// if valid modulation source
						const CTrack&	trkMod = doc.m_Seq.GetTrack(mod.m_iSource);
						// if modulation source has track type of modulator and is also modulated
						if (trkMod.m_iType == TT_MODULATOR && trkMod.m_arrModulator.GetSize())
							nModMods++;
					}
				}
			}
			if (nModMods) {
				_tprintf(_T("%d, %s\n"), nModMods, ff.GetFileName());
				nHits++;
			}
		}
	}
	printf("%d docs, %d hits\n", nDocs, nHits);

11/10/18 goto next/previous dub

FindDub could be probably use a binary search (though it's unclear whether duplicate times 
would have be accounted for) but it's doubtful any significant performance gain would result.
ChaseDubs averages about 5 microseconds with a typical song, so not much room for improvement,
considering that it calls FindDub for every track in the song.

test code for CArrayEx_BinarySearchAbove:

	int	data[] = {0, 10, 20, 30, 40};
	CIntArrayEx	arr;
	arr.SetSize(_countof(data));
	for (int i = 0; i < _countof(data); i++)
		arr[i] = data[i];
	while (1) {
		int	nIn;
		scanf_s("%d", &nIn);
		W64INT	iIns = arr.BinarySearchAbove(nIn);
		printf("pos=%d\n", iIns);
		if (iIns < 0)
			iIns = arr.GetSize();
		arr.InsertAt(iIns, nIn);
		for (int i = 0; i < arr.GetSize(); i++)
			printf("%d ", arr[i]);
		printf("\n");
	}

11/14/18 recursion overflow

Infinite recusion overflows after roughly 9,180 calls (in Debug mode).

11/22/18 autoscroll in velocity view; too confusing

but in case you change your mind

in ctor:
	m_nScrollDelta = 0;
	m_nScrollTimer = 0;

in EndDrag:
		KillTimer(m_nScrollTimer);
		m_nScrollTimer = 0;

in OnMouseMove:
		m_ptPrev = point;	// update cached cursor position
		CRect	rClient;
		GetClientRect(rClient);
		CSize	szClient(rClient.Size());
		if (point.x < 0)
			m_nScrollDelta = point.x;
		else if (point.x > szClient.cx)
			m_nScrollDelta = point.x - szClient.cx;
		else
			m_nScrollDelta = 0;
		// if auto-scroll needed and scroll timer not set
		if (m_nScrollDelta && !m_nScrollTimer)
			m_nScrollTimer = SetTimer(SCROLL_TIMER_ID, SCROLL_DELAY, NULL);

void CVelocityView::OnTimer(W64UINT nIDEvent)
{
	if (nIDEvent == SCROLL_TIMER_ID) {	// if scroll timer
		if (m_nScrollDelta) {	// if auto-scrolling
			CPoint	ptScroll(m_pStepView->GetScrollPosition());
			int	nPrevScrollX = ptScroll.x; 
			ptScroll.x += m_nScrollDelta;
			CPoint	ptMaxScroll(m_pStepView->GetMaxScrollPos());
			ptScroll.x = CLAMP(ptScroll.x, 0, ptMaxScroll.x);
			int	nScrollDelta = ptScroll.x - nPrevScrollX; 
			m_pStepView->ScrollToPosition(ptScroll);
			m_pStepView->m_pParent->OnStepScroll(CPoint(m_nScrollDelta, 0));
			CPoint	ptCursor;
			if (GetCursorPos(&ptCursor)) {
				UINT	nFlags = 0;
				if (GetKeyState(VK_CONTROL) & GKS_DOWN)	// if control key down
					nFlags |= MK_CONTROL;
				if (GetKeyState(VK_SHIFT) & GKS_DOWN)	// if shift key down
					nFlags |= MK_SHIFT;
				m_ptPrev.x -= nScrollDelta;
				m_ptAnchor.x -= nScrollDelta;
				ScreenToClient(&ptCursor);
				OnMouseMove(nFlags, ptCursor);
			}
		}
	} else
		CView::OnTimer(nIDEvent);
}

11/27/18 modulation graph via GraphViz; neato looks the best but complex patches are still a rat's nest

	CStdioFile	fout(_T("C:\\temp\\PolymeterModGraph.dot"), CFile::modeCreate | CFile::modeWrite);
	fout.WriteString(_T("digraph G {\n"));
	CSize	szLogPix(72, 72);
	double	fDPI = max(szLogPix.cx, szLogPix.cy);
	CSize	sz(1920, 1080);
	double	fWidth = sz.cx / fDPI;
	double	fHeight = sz.cy / fDPI;
	CString	sGraph;
//	sGraph.Format(_T("graph[rankdir=TB,size=\"%f,%f\",ratio=fill,dpi=%f];\n"), fWidth, fHeight, fDPI), 
//	sGraph.Format(_T("graph[ratio=fill,size=\"17.77,10\",dpi=100];\n")); 
	sGraph.Format(_T("graph[ratio=1.77];\n")); 
	sGraph = _T("overlap=false;\nsplines=true;\n");
//	sGraph = _T("overlap=false;\n");
	fout.WriteString(sGraph);
	int	nTracks = GetTrackCount();
	CMap<CString, LPCTSTR, int, int>	map;
	CStringArrayEx	arrName;
	arrName.SetSize(nTracks);
	for (int iTrack = 0; iTrack < nTracks; iTrack++) {	// for each track
		const CTrack& trk = m_Seq.GetTrack(iTrack);
		CString	sName(trk.m_sName);
		if (sName.IsEmpty())
			sName.Format(_T("%d"), iTrack + 1);
		int	nRefs = 1;
		if (map.Lookup(sName, nRefs)) {
			nRefs++;
			CString	sNum;
			sNum.Format(_T("%d"), nRefs);
			arrName[iTrack] = sName + sNum;
		} else {
			arrName[iTrack] = sName;
		}
		map.SetAt(sName, nRefs);
	}
	LPCTSTR	sModTypeColor[] = {
		_T("black"),
		_T("red"),
		_T("green"),
		_T("blue"),
	};
	for (int iTrack = 0; iTrack < nTracks; iTrack++) {	// for each track
		const CTrack& trk = m_Seq.GetTrack(iTrack);
		int	nMods = trk.m_arrModulator.GetSize();
		for (int iMod = 0; iMod < nMods; iMod++) {
			const CModulation&	mod = trk.m_arrModulator[iMod];
			if (mod.m_iSource >= 0) {
				CString	sLine;
				sLine.Format(_T("\"%s\"->\"%s\"[color=%s]\n"), arrName[mod.m_iSource], arrName[iTrack], sModTypeColor[mod.m_iType]);
				fout.WriteString(sLine); 
			}
		}
	}
	fout.WriteString(_T("}\n"));

12/2/18 test code for recording MIDI input that found round error bugs in ImportMidiFile

#if 0
		static const MIDI_EVENT testevt[8] = {	// test data
			{0,		0x404090},
			{200,	0x4090},
			{240,	0x404190}, // this is generating a WRONG RESULT
			{500,	0x4190},
			{500,	0x404290},
			{750,	0x4290},
			{750,	0x404390},
			{1000,	0x4390},
		};
		theApp.m_arrMidiInEvent.SetSize(_countof(testevt));
		for (int i = 0; i < _countof(testevt); i++)
			theApp.m_arrMidiInEvent[i] = testevt[i];
#endif

in Track.cpp, the bug was integer division where floating point division was intended;
fixed by casting int to double in these three lines:

int	nStart = round(double(info.nStartTime) / nInQuant);
int	nEnd = round(double(nTime) / nInQuant);
int	nMaxSteps = round(double(nMaxTime) / nInQuant);	// compute maximum track length


12/4/18 Algo Rag

12/5/18 Interago
2nd A is at 290:000, ends at 579

12/7/18 showing negative time in song view

		int	nOffset = trk.m_nOffset + m_nTimeShift;

		int	nCellTime = round(iFirstCell * fTicksPerCell) + m_nTimeShift;

			nCellTime = round(iCell * fTicksPerCell) + m_nTimeShift;

	return round(x * GetTicksPerCellImpl() / m_nCellWidth) + m_nTimeShift;

	return round((nSongPos - m_nTimeShift) / GetTicksPerCellImpl() * m_nCellWidth);

But this only fixes the display! For editing to work, the time calculations in the document's 
dub functions have to change too. Generally they're in this form:

	int	nStartTime = round(rSelection.left * fTicksPerCell) + m_nTimeShift;

But the problem is, m_nTimeShift isn't a member of the doc and probably shouldn't be because
multiple song views could (in theory) have different shifts. It would have to be passed in,
as ticks per cell already is. What a gnarly mess!

12/7/18 subtle difference between song MIDI export between Release and Debug

Ala Aye rec A B A B fine tune spread.plm

difference appears at 0x2C26 right after this string: Organ B Hi
Debug has nine bytes, Release has only five

12/10/18 note windowing: a more efficient implementation (than CNote's)
	
nNote = ((nNote - nWindowBase) % OCTAVE) + nWindowBase;
if (nNote < 0)
	nNote += OCTAVE;

argument for windowing: window base is 52
varying it separately for each voice will be awesome

OOPS above implementation is WRONG, but generates awesome results for guitar stuff.
correct (but relatively boring) implementation adds base AFTER correcting for wraparound:
(guitar stuff sounds good if base varies from E3 to G3 to A3 to B3 and back or similar; C4 is good too)

nNote = ((nNote - nWindowBase) % OCTAVE);
if (nNote < 0)
	nNote += OCTAVE;
nNote += nWindowBase;

alternate DropBox, new WAV files:

Algo Rag
https://www.dropbox.com/s/kdehpmolbv2oei0/Algo%20Rag%20panned%20TRIM%20NORM.wav?dl=0

Interago
https://www.dropbox.com/s/m18s3jt7zs4dc0p/interago2%20TRIM%20NORM.wav?dl=0

Ilopo Ferese
https://www.dropbox.com/s/ihlkzam68iiofl3/Ilopo%20Ferese%20NORM.wav?dl=0

piano range
https://www.dropbox.com/s/v2pjowkmbpzbbcd/piano%20range.wav?dl=0

12/12/18 position modulation and change ringing

For the change 2,4,1,3 (AKA 1,3,0,2 in zero origin) the correct position offsets are 2,0,3,1 (in high to low order). Why?
The format typically used to describe change ringing tells you, for each position in the sequence, which bell should ring.
But to do position modulation, we need to know the opposite: for each bell, which position it should have in the sequence!
One can be translated to the other via an array lookup:

for (iPos = 0; iPos < Bells; iPos++) {	// for each bell
	int iBell = arrBell[iPos];	// arrBell contains a bell index for each position
	arrPos[iBell] = iPos;	// arrPos will contain a position index for each bell
}

a[1] = 0
a[3] = 1
a[0] = 2
a[2] = 3

a = 2,0,3,1

12/21/18 combinations of additive modulations

Assume each modulator contains two or more unique values, one of which is zero, the others ranging from [1..6].

(0 a) (0 b)
4 outcomes (2**2): 0, a, b, a + b

(0 a) (0 b) (0 c)
8 outcomes (2**3): 0, a, b, c, a + b, a + c, b + c, a + b + c 

(0 a) (0 b c)
6 outcomes: 0, a, b, c, a + b, a + c

(0 a b) (0 c d)
9 outcomes (3**2): 0, a, b, c, d, a + c, b + c, a + d, b + d

(0 a) (0 b c) (0 d e)
18 outcomes:
0, a, b, c, d, e, b + d, b + e, c + d, c + e, 
a + b, a + c, a + d, a + e,
a + b + d, a + b + e, a + c + d, a + c + e

(0 a b) (0 c d) (0 e f)
27 outcomes (3**3):
0, a, b, c, d, e, f
a + c, a + d, b + c, b + d,
a + e, a + f, b + e, b + f
c + e, c + f, d + e, d + f,
a + c + e, a + c + f, a + d + e, a + d + f,
b + c + e, b + c + f, b + d + e, b + d + f,

The first four cases above should be sufficient for generating scales.

1/2/19 restarting application

bool CMyApp::RestartApp()
{
	if (SaveAllModified())	// save any unsaved documents
		return false;	// save failed or user canceled
	CloseAllDocuments(TRUE);	// end session
	TCHAR	szAppPath[MAX_PATH] = {0};
	::GetModuleFileName(NULL, szAppPath, MAX_PATH);
	STARTUPINFO	si = {0};
	PROCESS_INFORMATION	pi;
	GetStartupInfo(&si);
	if (!CreateProcess(NULL, szAppPath, NULL, NULL, FALSE, 0, NULL, NULL, &si, &pi)) {
		AfxMessageBox(_T("Can't create process.")); // report the error
		return false;
	}
	m_pMainWnd->PostMessage(WM_CLOSE);	// close the app
	return true;
}

1/6/19 output validation for changes since mastering

Compared sequencer dumps between version 0.0.21.000 and version 0.0.16.005: no differences.
Test document was "Manmade\Ala Aye rec A B A B fine tune spread.plm"

1/12/19 export test

void ExportTest()
{
	static const LPCTSTR arrFilename[] = {
		_T("Ala Aye rec A B A B fine tune spread"),
		_T("Asiri rec1e lead B dub spread lead string"),
		_T("Bones rec fine tune cvt for PM 0.0.15 Mike"),
		_T("Buy Me ending Mike double finger"),
		_T("Buy Now double bass finger Mike"),
		_T("Dek Sep Bluso jam breaks Mike closed and pedal hat +10"),
		_T("Fazo Kanto B mix kick +15 Mike 3dB less ride"),
		_T("Iyika rec A B take1 scale up volumes spread"),
		_T("Stencil lower bass spread strings mike"),
		_T("Vizyon rec02 edit etc scale up volumes Mike"),
	};
	LPCTSTR	pszDocFolder = _T("C:\\Chris\\MyProjects\\Polymeter\\docs\\Manmade\\");
	LPCTSTR	pszOutFolder = _T("C:\\Chris\\MyProjects\\Polymeter\\experimental\\export\\");
	for (int iFile = 0; iFile < _countof(arrFilename); iFile++) {
		CString	sFilename(arrFilename[iFile]);
		CString	sDocPath(pszDocFolder + sFilename + _T(".plm"));
		CString	sOutPath(pszOutFolder + sFilename + _T(".mid"));
		CPolymeterDoc *pDoc = DYNAMIC_DOWNCAST(CPolymeterDoc, theApp.OpenDocumentFile(sDocPath, FALSE));
		if (pDoc == NULL) {
			AfxMessageBox(_T("Can't open file."));
			return;
		}
		int	nDur = pDoc->m_Seq.GetSongDurationSeconds();
		pDoc->UpdateChannelEvents();	// queue channel events to be output at start of playback
		if (!pDoc->m_Seq.Export(sOutPath, nDur, true, pDoc->m_nStartPos)) {
			AfxMessageBox(_T("Export failed."));
			return;
		}
		pDoc->OnCloseDocument();
	}
}

at end of InitInstance:
ExportTest();

The version used at ManMade Mastering was 0.0.16.005.
Note that versions before 0.0.19.001 didn't support negative dub times, and this breaks 
export comparison for documents that have a negative start time, including Asiri and Ikiya.
The export not resetting cached parameters bug (fixed in 0.0.18.001) could also cause compare failure, though it doesn't seem to matter.

1/13/19 showing SVG graph via HTML view

Fully functional CHtmlView that displays an SVG file adds 21KB to the app. (5113 vs 5092)
Note that wheel zoom is broken for SVG only: both wheel up and wheel down zoom in, so there's no way to zoom out.

1/19/19 Unicode not working

Adding this to CIniFile::Open fixes the track names in the document:

	if (OpenFlags & CFile::modeWrite) {
		BYTE	bom[2] = {0xFF, 0xFE};
		fp.Write(bom, 2);
	}

BUT this doesn't address track names in exported MIDI files (not doable as MIDI doesn't support Unicode)
or in exported CSV files, or in graphs if that gets implemented. It also doubles the size of the document,
which is dumb considering that only track names, preset names, and part names actually need to be Unicode.
One possible solution would be to store these names in an encoded ASCII format such as Base64.

1/25/19 IWebBrowser2 wheel zoom troubles

Ctrl+wheel only zooms in, never out on Dell M4800 Win7 IE 10
on Aspire E15, Win8 IE 11, in/out is normal but zoom is painfully slow, steps are too small
seems to be no setting for zoom step size in IE
also IE 10 gets confused after zoom is set via ExecWB and makes page too small by 0.2
possible related to ZoomFactor registry key HKCU/Software/Microsoft/Internet Explorer/Zoom/ZoomFactor
it's 100000 on M4800 but 157000 on Aspire E15

1/30/19 Graphviz and IE: graph too small when system text size above 100%

Graphviz expects you to tell it the size of the graph in inches. Now of course the window size is in pixels. 
So I was dividing the window size by the system DPI (in floating point) to get the graph size. This works ONLY
if the system text size is 100%. Why? Because Graphviz has its own internal DPI and it defaults to 96. 96 just
happens to be the DPI you get on a Windows machine when text scaling is 100% (for historical reasons). So, you
might think you could just pass the system DPI to Graphviz, and indeed Graphviz has a DPI parameter. BUT it's 
apparently broken for SVG (and has been for years). What seems to work is using 96 when computing the graph 
size, regardless of the system size. This works because IE is scaling up the entire graph anyway to compensate
for the system text size. The only down side I can see to this is that the font size is sometimes bigger or 
smaller than you might expect. But I think this is minor compared to having the entire graph be the wrong size.

2/6/19 Oily Rock notes

Mostly got done during an all-nighter 2/4 but also all day 2/5
Could use just a bit more excitement in the high end and percussion in places
Don't overdo it though, the minimalism is part of what makes it so spooky
Low end is super fat, but possibly a bit floppy
Stereolab style vocal is the bomb
Vocals are nearly perfect but maybe try adjusting offset of "They'll spit on the ground"
wrong note in female vocal at start of 2nd chorus; D over A? why isn't note mod working? wrong offset? fixed

much adjusting of vox offsets later, they sound funky and sometimes fit the bass line
add more drums: rimshot, tom, cymbal (and don't forget to pan them)
maybe some reverb on the cymbal too, or even on the snare / rimshot, try it

nightmarish hassle to open MIDI file in Reason, so stupid! No easy way to map tracks to existing devices
only way seems to be to copy the MIDI data from one track to another by hand, very error-prone, it sucks

2/6/19 overlapping notes

Overlapping notes means two or more overlapping instances of the same MIDI note on the same MIDI channel. 
This situation often causes problems with synths. The typical behavior is, when one the overlapped notes ends, 
it cuts off the other instances, even though they still have more duration to go. Some synths (e.g.Yamaha) do 
the right thing, i.e. they keep a reference count for each note on each channel, and only cut off the note when 
the count reaches zero. But many synths (including Reason) don't do this so it's not possible to count on it.
But I have this case a LOT. Many techniques I use including position modulation create this case constantly. 
So I propose to handle it in Polymeter. There are two approaches and I might support them both. One is, when a 
note ends and would otherwise cut off another instance of the note, don't send Note Off, instead defer it until 
the other instance ends, and when it does, send BOTH Note Offs at the same time. Obviously this requires reference 
counting and is very likely what my Yamaha does internally.
The second solution is, when a note is about to start and would overlap another instance of the same note, send 
a Note Off for the first instance before sending the Note On for the new one. This effectively prevents overlap 
before it happens. This method very likely won't sound the same as the first one, because the Note Off will 
trigger the release stage of the sound's envelope. But in some cases it may be preferable.

2/6/19 Importing MIDI file into Reason SNAFU

To create a new empty automation lane for a track, select the track and then in the sequencer pane, near the
upper left corner, look on for a little drop list with an icon that resembles a polyline, with markers on the 
vertices, just right of a little white plus sign in a gray circle; the tooltip says Track Parameter Automation.
Select the desired parameter and an automation lane is created. You need to do this before trying to drag 
automation data from imported tracks to device tracks, otherwise you get the dreaded alien automation warning.

2/8/19 more on overlapping notes

The immediate problem is overlapping notes, because the unwanted rests make position modulation much less useful.
I think the proposal is clear: two modes, "poly" and "mono". Poly defers note offs for notes that overlap until 
the reference count (for the given note and channel) transitions to zero, at which point a SINGLE note off is sent.
Mono prevents overlap from occurring, by sending a note off whenever the reference would otherwise transition from
one to two. Using "poly"' with a synth that already does reference counting will likely cause notes to be "stuck" 
on and increasing loss of polyphony, but the solution is, don't use this feature with such synths.

To implement the first method requires a significant change. For reference counts to be meaningful, note offs must
be processed BEFORE note ons, which is the opposite of how it currently works. The only down side of this is that
the note off processing must be moved inside the critical section, becayse note on and note off processing must use
the same current time. As a result the callback holds the critical section longer, but only by a few microseconds.
This change was tested on the Manmade files via ExportTest and generated identical files to 0.0.19.

2/9/19 Oily Rock arrangement notes

Starting point is "oily rock more drums" which is pretty good
Shortened choruses to 2/3 of their previous length (64 measures instead of 96)

Tried adjusting some vocal timing in the 1st verse:
Galaxies spin: -110 to -140
Ignoring our prayers: -105 to -155
Waves and particles: -50 to -60
But decided against it for now, too rushed

changed kick turn length from 5 to 3, makes simpler and more traditional variations
shorted chorus is definitely more snappy
tried putting room reverb on drums: sounds great on snare, maybe on hats too

2/14/19 overlapping notes: this is probably pretty expensive but it works

	if (nEvents) {	// if any events to output
		if (m_bPreventNoteOverlap) {
			FixNoteOverlaps();
			nEvents = m_arrEvent.GetSize();
		}
		CEvent	evt(m_nCBLen, MEVT_NOP << 24);

void CSequencer::FixNoteOverlaps()
{
	int	nEvents = m_arrEvent.GetSize();
	int	iEvent = 0;
	while (iEvent < nEvents) {	// for each event
		DWORD	dwEvent = m_arrEvent[iEvent].m_dwEvent;
		if (MIDI_STAT(dwEvent) == NOTE_ON) {
			int	iChan = MIDI_CHAN(dwEvent);
			int	iNote = MIDI_P1(dwEvent);
			if (MIDI_P2(dwEvent)) {
				if (m_arrNoteRef[iChan][iNote]) {
					if (iEvent && m_arrEvent[iEvent - 1] == m_arrEvent[iEvent]) {
						m_arrEvent.RemoveAt(iEvent);
						nEvents--;
					} else {
						CEvent	evt(m_arrEvent[iEvent]);
						evt.m_dwEvent &= ~0xff0000;
						m_arrEvent.InsertAt(iEvent, evt);
						iEvent++;
						nEvents++;
					}
				}
				m_arrNoteRef[iChan][iNote]++;
			} else {
				if (m_arrNoteRef[iChan][iNote] > 0) {
					m_arrNoteRef[iChan][iNote]--;
					if (m_arrNoteRef[iChan][iNote]) {
						m_arrEvent.RemoveAt(iEvent);
						nEvents--;
					}
				}
			}
		}
		iEvent++;
	}
}

2/20/19 note overlap flunking test app

guitar overlap.plm
first problem is note 66 turned on at 575 and again at 1020 without being shut off

2/25/19 note overlap less but still flunking

range test Reason
first problem is zero-duration note, note 65 velocity 88 and then 0 during same callback time 15603

n=6
0 3a90
0 3f90
0 3f90
0 584190 <-- note 65 vel 88
0 623890
0 624190 <-- note 65 vel 98

problem is overoptimization: must iterate all events with same time
can't stop on finding an event with zero velocity, because we're potentially 
inserting such events as we iterate; now passes all tests fine for up to one hour of events

Enabling note overlap prevention doesn't seem to significantly increase callback times (for Ala Aye)
(maximum time can vary considerably, typical range for Ala Aye is 9.5% to 13%)
0.0.23.5: 12.6%, 0.3%
0.0.24.0: 12.0%, 0.3%

2/25/19 fast move for CArrayEx Fast* methods (FastInsertAt, FastRemoveAt)

probably a bad idea because memmove is so heavily optimized with SSE and other assembler tricks

__forceinline static void FastMove(CArrayEx_TYPE *pDst, const CArrayEx_TYPE *pSrc, size_t nCount)
{
	// if destination below source, or above but non-overlapping
    if (pDst <= pSrc || pDst >= pSrc + nCount) {
		// copy from lower addresses to higher addresses
		while (nCount--) {
			*pDst++ = *pSrc++;
		}
	} else {	// destination above source and overlapping
		// copy from higher addresses to lower addresses
		pDst = pDst + nCount - 1;
        pSrc = pSrc + nCount - 1;
        while (nCount--) {
            *pDst-- = *pSrc--;
        }
	}
}

2/28/19 Reason MIDI import doesn't change tempo

This will be a major potential dumb error in the future. This explains why Oily Rock seemed slow,
and also why the imported version's vocal start modulation didn't seem to match the live version.
Changing from 126 to 120 BPM will do that. WATCH OUT FOR THIS!!!


3/1/19 song playback doesn't always match live output when modulator tracks have big quants

"song big quant bug.plm" demos the issue. Synopsis: During song playback, a track's mute state can
only change on a step boundary, because the sequencer only evaluates track states on step boundaries.
If a track has a large quant, the step boundaries are relatively few and far between. So during song
playback, when a track's mute state can change due to prerecorded dubbing is limited. In other words,
the bigger a track's quant, the less dub timing granularity that track has. During live playback, 
granularity isn't an issue because the user directly updates the track mute flags and hence track
mute state changes are completely asynchronous with respect to the sequencer callback thread.

Dub timing granularity affects all track types, not only modulators, and all modulation types.
Large track offsets make the problem more obvious but the root cause is coarse quants.
The easiest solution is to avoid using big quants, especially for modulator tracks, and especially
with large offsets. Such tracks can be converted to sixteenth notes easily and this will make the
problem go away by increasing the dub timing granularity.

The deeper issue is fundamental to the sequencer design. It might be possible to solve the problem
by changing how dubs are recorded. Arguably when a user unmutes a modulator track with a large quant,
they intend for that track to have immediate effect, regardless of whether they're near the track's
step boundary. So instead of recording the current time, the program could record the time of the 
track's nearest step boundary at or before the current time. But this could easily have unintended
side effects. It also wouldn't make sense for non-modulator tracks. Basically there just isn't any
satisfactory solution, because song view playback and live view use the track mutes differently.

3/28/19 Kahelo

A section string hit is too long; try trimming 4 beats from its start (down to 11 steps of duration); more space [done]
in B section, make string hit one step longer? [done]
erroneous unmute of low guitar part at 22:090 but does NOT affect exported file (tested!)
guitar / piano mix seems dubious
high hat seems excessively biting
possible issues with clipping
organ key click may be cluttering mix

Norah 3/29 on C section: bring shaker in halfway through first pass of C section
also C shaker part is lame, use original shaker part instead (look at 5 hat too) [done]

4/20/19 tristate ("Ona Lile")

Attacks are a bit too sharp sometimes. Lowering velocity loop to 5, -5, -15 seems to help. Lower fader 3 dB to avoid clipping.
Biased about 3 dB to right, not sure why or what to do about it

5/9/19 velocity pane origin discrepency

64 is NOT the exact middle of the pane, because the upper rail is 127.

	int	oy = szClient.cy - round(double(MIDI_NOTES / 2) / MIDI_NOTE_MAX * szClient.cy);
	pDC->FillSolidRect(rClip.left, oy, szClient.cx, 1, clrBeatLine);


5/15/19 Manmade Mastering with Mike

interago2.wav (12/8/18)
Interago rec2.plm

kasita mondo rec take2.wav (1/2/19)
kasita mondo rec take2.plm

5/25/19 Ilopo Farese glitch

Experimentation shows it's definitely caused by insufficient polyphony. The NN-XT sampler's default 
polyphony is only 16 voices. That's cutting it fine considering all the legato overlapping notes. 
Increasing it to 30 makes the problem go away. Made it 64 out of an abundance of caution.

6/7/19 Reason velocity woes

Reason doesn't seem to have any velocity curves that approximate the behavior of Windows GS synth.
GS synth generally is linear in decibels, for example with C4 and the piano sound:
100	-18.5
50	-30.5
25	-42.5
i.e. each doubling of velocity adds 12 dB of amplitude.
Atunwi uses a limited set of velocities: 60, 70, 80, 90, 100. Approximate dB values for these:
vel	abs	delta	rel
100	-18.5	0	0
90	-20.2	-1.7	-1.7
80	-22.3	-2.1	-3.8
70	-24.7	-2.4	-6.2
60	-27.3	-2.6	-8.8

based on the observed values for the NN-19 sampler using a test sine wave 
(velocity sensitivity = 64, output volume = 127):
rel	abs	vel	v.offs
0	-8.9	100	0
-1.7	-10.6	94	-6
-3.8	-12.7	86	-12
-6.2	-15.1	78	-22
-8.8	-17.7	70	-30
So subsituting the above velocities should help match the GS synth behavior.

actual using NN-19, C4
abs	velo	offs
-17.7	100	0
-19.4	92	-8
-21.5	84	-16
-23.9	75	-25
-26.5	65	-35

actual using NN-19, C5
abs	velo	offs
-20.3	100	0
-22.0	93	-7
-24.1	85	-15
-26.5	79	-21
-29.1	70	-30

Still sounds too dynamic with velocity amp = +64, +55 is a bit better but still not right.
Try picking a piano sound and then hand tuning the five velocities until it sounds reasonable.
Get the lowest one right and roughly evenly distribute the others.

NOTE: at end of B section, cut bass one note early so it ends on Db. (rec1b2)

6/9/19 Apologize
aah overlaps start of verse a bit: if necessary, try cutting sampler's amp release from 60 to 55

Singularity
WaveNet-C with speed = 1.2 seems to work pretty well for hardest lines; others need to be slower
wobbilator bass
rubstep
quaver
anodyne
bubstep up
talky

Changing Climate
original untrimmed sample has a 25.04ms delay at the start
25 ms is exactly 7 ticks at 140 BMP (PPQ 120)
25 / (60 / 140 / 120) = 7
coincidence? you be the judge

6/25/19 new hip hop track (Exit Game?) low Kong level

looping Hhp 70 Soufland at velocity 100
Kong output level:
100: -14.77
127:  -8.54

result is a boost of 6.23 dB, still low but it helps

6/30/19 Changing Climate samples
Kevin thinks there are too many at the start. He's probably right, 17 is one too many and that first one is spurious.

7/4/19 break hip hop 2 (Lodidi)

| D-9	F-maj7 	| Bb-9 	Db-maj7 | F#-9 	A-maj7 	| F-9 	Ab-maj7 | 
| Bb-9	Db-maj7 | F#-9 	A-maj7 	| D-9 	F-maj7 	| Db-9 	E-maj7 	| 
| F#-9	A-maj7 	| D-9 	F-maj7 	| Bb-9 	Db-maj7 | A-9 	C-maj7 	|

The B section snare seems a hair too loud, and sounds slightly phased.

8/25/19 Kahelo add open hat length 74 turnaround? Kong Oh_LongHatBSQ volume 95 or so
move extra hat to ch 4, load Open Hat in ch 9, and enable exclusive 8/9 for gating

9/24/19 more hip hop combos
31, 33, 22

10/20/19 interesting SetPerm params:
3 3 3 3, delta 2 (Atunwi)
4 4 4, delta 3 (Atunwi break)
3 4 3 2, delta 2 (Ero Ayo)
5 4 3, delta 2 (Ero Ayo B)
3 3 3, delta 2 (Planet Broke)

10/25/19 Irubo Radical Piano velocity vs Total RMS
vel	L	R
100	-15.30	-13.61
99	-15.51	-13.82
98	-15.89	-14.22
97	-16.15	-14.46
96	-16.51	-14.77
95	-16.75	-15.10
94	-17.12	-15.39
93	-17.11	-15.40
92	-17.41	-15.72
91	-18.12	-16.37
90	-18.30	-16.63

10/30/19 waltz

return to A at around 410, why?

Ero Ayo	5:32
Mo Bulu	4:51
Irubo	4:15
Waltz	4:21

11/20/19 switching from track view to song view sometimes leaves one cell upainted

Using "two triads ... rec1" doc as a test, the unpainted cell's location varies,
but it's usually cell [30,20]. It's intermittent, and apparently not related to 
drawing the song position marker. The window style doesn't seem to affect it either.
It occurs in both Debug and Release. Adding even very sparse printf lines to OnDraw
makes it go away. Adding a Sleep(1) after the FillSolidRect also makes it go away.

This bug is very hard to explain and seems possibly timing-dependent. The oddest 
thing is that the bug never affects more than one cell, and it's usually the same
cell, though not always.

Drawing all cells as solid also hides the bug. Perhaps we're overwhelming the GDI, 
or whatever component of Windows is emulating GDI? Maybe the drawing of occupied
cells (which currently takes five calls to FillSolidRect) is just too inefficient 
and breaks some assumption in GDI.

Occupied cells can alternatively be drawn by just overwriting the cell background. 
This appears to solve the sporadic unpainted cells, possibly because it gets the 
GDI cost of an occupied cell down to two FillSolidRect calls, which is certainly
better than five. It risks flicker, but I don't see any. Maybe the flicker is too
fast to see, or maybe the two fills get coalesced somehow. Let's try it.

11/28/19 I'm unable to reproduce the one cell unpainted bug described above.

11/28/19 Watergate chords

G-7	Bb7	G7	Fo7
Eo7	Bb-7b5	C#-7b5	C#-7
E-7b5	E+7	E-7	E7
G-7b5	Bb-7	G+7	C#7

Eb	Bbsus7	G+7	F-7b5
E-7b5	E+7	E-7	E7
EoM7	Bb-7b5	E-maj7	Emaj7
G-7	Bb7	G7	Fo7

12/5/19 PhaseBar: Direct2D vs GDI

Direct2D is consistently 10x faster, antialiasing and all, surprise surprise (not really).
Less than 100 us on average, vs. more than a millisecond for GDI.

To draw vertical line:
	D2D1_POINT_2F	ptZeroDeg = {ptOrigin.x, float(ptOrigin.y - fOrbitWidth * nOrbits)};
	pRenderTarget->DrawLine(ptOrigin, ptZeroDeg, &brOrbit);
But sadly it doesn't improve perception of convergence all that much.
Changing planet's color when it's near zero degrees just made it more confusing.


for a slightly cleaner appearance:
	float	fZeroY1 = float(ptOrigin.y - fOrbitWidth * 0.5);
	float	fZeroY2 = float(ptOrigin.y - fOrbitWidth * (nOrbits - 0.5));
	pRenderTarget->DrawLine(CD2DPointF(ptOrigin.x, fZeroY1), 
		CD2DPointF(ptOrigin.x, fZeroY2), &brOrbitNormal);

1/2/20 track view row colors

	ON_NOTIFY(NM_CUSTOMDRAW, IDC_TRACK_GRID, OnCustomDraw)

void CTrackView::OnCustomDraw(NMHDR* pNMHDR, LRESULT* pResult)
{
	NMLVCUSTOMDRAW* pLVCD = reinterpret_cast<NMLVCUSTOMDRAW*>(pNMHDR);
	switch(pLVCD->nmcd.dwDrawStage) {
	case CDDS_PREPAINT:
		// could be conditional (only if any tracks need custom color)
		*pResult = CDRF_NOTIFYITEMDRAW;
		break;
	case CDDS_ITEMPREPAINT:
		{
			int	iItem = static_cast<int>(pLVCD->nmcd.dwItemSpec);
			CPolymeterDoc	*pDoc = GetDocument();
			ASSERT(pDoc != NULL);
			int	clrCustom = pDoc->m_Seq.GetColor(iItem);
			if (clrCustom >= 0)
				pLVCD->clrTextBk = clrCustom;
		}
		*pResult = CDRF_DODEFAULT;
		break;
	default:
		*pResult = CDRF_DODEFAULT;
	}
}

Performance impact seems negligible but it flickers a bit, presumably due to the
background being erased in the default color. There's probably no way to solve
that, other than double-buffering the grid control, which creates other issues.

The item notifications are a significant amount of message traffic, and it would
preferable to avoid this traffic unless it's really necessary (the user shouldn't
have to pay for unused features). The initial custom draw notification is always
sent, but the item notifications could be requested only if any of the items to 
be painted have custom colors. Iterating all the tracks would work but it seems 
excessive. In order to know what range of tracks should be checked for custom 
coloring, the following lines are needed:

int	iPageTop = m_grid.GetTopIndex();
int	nPageItems = m_grid.GetCountPerPage();

These lines take about 20us, except once in a long while they take between
500us and 1ms for no obvious reason.

But probably the biggest objection to the coloring idea is that it makes the
UI look inconsistent. Why aren't the steps colored too? Maybe coloring just
the non-empty steps is better?

1/27/20 convergences

input: [2 3 5 7 11]
convergences: 26
6       [2 3]
10      [2 5]
14      [2 7]
15      [3 5]
21      [3 7]
22      [2 11]
30      [2 3 5]
33      [3 11]
35      [5 7]
42      [2 3 7]
55      [5 11]
66      [2 3 11]
70      [2 5 7]
77      [7 11]
105     [3 5 7]
110     [2 5 11]
154     [2 7 11]
165     [3 5 11]
210     [2 3 5 7]
231     [3 7 11]
330     [2 3 5 11]
385     [5 7 11]
462     [2 3 7 11]
770     [2 5 7 11]
1155    [3 5 7 11]
2310    [2 3 5 7 11]

2/18/20 song CSV export omits tempo map

In "two triads combine 2222 rec2b" (AKA "Ibaramu") the reason there's 
an spurious empty track at the end of the export is because the song's
tempo track uses channel 10, which is otherwise only used by modulators.
GetChannelUsage excludes modulator tracks, but not tempo tracks. That's
a minor bug, but the more serious bug is that the export omits the tempo
map altogether. The export uses CMidiFile::ReadTracks which currently
doesn't support reading the tempo map, though it should.

2/19/20 binary objects limited to 2K bytes

GetBinary uses CWinApp::GetProfileString, which calls GetPrivateProfileString
when writing to an INI file, with a buffer size of 4K characters. Each byte of
a binary object occupies two characters in the corresponding INI file string,
hence we're limited to 2K bytes, minus one. This limits tracks to 2047 steps 
and 255 dubs. Disaster strikes!

Overloading CWinApp::GetProfileString allows us to increase the buffer size,
but only up to 64K, due to limitations within GetPrivateProfileString, which
we don't have source code for. Hence binary objects are at best limited to 32K
bytes.

WriteBinary seems to be unlimited, because WritePrivateProfileString can write
strings longer than 64K characters to a profile, though it's slow. It's only 
*reading* the strings that's the problem.

With a 64K buffer, tracks are limited to 32767 steps (verified) and 4095 dubs.
This is a big improvement over 2047 steps and 255 dubs, but still unacceptable,
especially since the limit isn't enforced anywhere and exceeding it results in
silent truncations. But at least it should be issued as a patch.

A survey of alternatives to GetPrivateProfileString suggests that the only 
workable approach in terms of performance is to rewrite it from scratch, using
a caching strategy to avoid reading the file over and over. This is apparently
what Wine did, but their code is way too messy to use directly.

Probably the thing to do is override GetProfileInt and GetProfileBinary also,
and handle the entire mess with a CIniFileReader class. It might make sense to
handle writing too, since much of the code would be shared and doing so might
improve performance for writing very long strings.

2/24/20 new CIniFile that handles I/O

Works fine, no length limitations, and huge performance improvement:
reading is 20x faster, writing is 34x faster

Slight performance gain (< 1ms) from pre-allocating 100 sections.

The performance sucks for large strings! But it's fixable by overloading
CStdioFile::ReadString to increase the buffer size exponentially.

// the following line works because CString::GetBuffer increases the
// buffer size exponentially; see PrepareWrite2 in atlsimpstr.h
int	nBufAvail = rString.GetAllocLength() - nLen;	// performance gain
lpszResult = _fgetts(lpsz, nBufAvail + 1, m_pStream);

2/25/20 short out to stream handle

Amazingly, despite the need for a cast, it works. This changes a lot!

"This parameter can also be the handle of a MIDI stream cast to HMIDIOUT."
"This function might not return until the message has been sent to the output
device. You can send short messages while streams are playing on the same 
device (although you cannot use a running status in this case)."

in CSequencer::OutputLiveEvent(DWORD dwEvent)

return MIDI_SUCCEEDED(midiOutShortMsg(HMIDIOUT(m_hStrm), dwEvent));

The entire MIDI thru and recording implementation needs to be rethought.
The message seems to be output immediately and is completely unaffected
by the sequencer callback latency. But it's not so obvious how to play 
back the recording. The times are relative to when the input device was
opened, which isn't synchronized to the stream callback. Also the input
times must be converted to ticks, but how would that work if the tempo 
changes during recording, changing the duration of a tick? Maybe don't 
support that case?

2/26/20 recording MIDI input: timestamps aren't useful

There doesn't seem to be any reliable way to synchronize MIDI input
timestamps with the sequencer timing. If so, this means the strategy
of recording MIDI event timestamps (in milliseconds since the input
device was opened) and converting them to ticks later is unworkable.
But that strategy would be required if we output live input events
directly to the stream handle instead of queing them to the callback.

Instead, the input recording strategy would have to be built around
the current method of outputting the live input events: queuing them
to the sequencer's callback thread and inserting them into the output
stream. The disadvantage of this method is that the live events are
delayed by a variable amount up to the callback latency (jitter), and
even if they weren't, timing accuracy is relatively poor at PPQ 120.
This is because at tempo 120 e.g. it's 4.16 ms per tick, which means
converting milliseconds to ticks wastes a lot of precision.

The weirdest thing is that if I record the input, while tapping along
to a beat at tempo 120, the input events are all separated by around
487 ms +/- 1. They should be separated by 500 ms! This could happen
if the sequencer isn't actually producing 120 BPM, but I checked its
output in WaveShop and it's precise to within 0.1 ms. So something is
really wrong here. Oddly, the error 500 - 487 = 13 ms corresponds to
about three ticks. No idea why!

2/27/20 Reason tests show time stamps are actually fine

Key is to use Reason for all such tests, because it has only 1ms of
latency. Using GS Synth is dumb because it has tons of latency which
corrupts the tests. Test setup requires two loopMIDI ports and two
instances of Polymeter.

Polymeter_2 -> loopMIDI_2 -> Polymeter_1 -> loopMIDI_1 -> Reason

Both instances of Polymeter have only a simple 1 beat loop.
You start instance 2 playing, and then try to start recording
in instance 1 such that the two loops synchronize perfectly.
When they do synchronize, slight phasing is audible.
Then you stop recording and check the timestamps, and they
should be very close to 3,30,123,153 etc. And they are!
(offset of 3 is because of sequencer's initial buffer)
So it looks like the new scheme will actually work.

2/27/20 closing more safely

These two changes will help prevent crashes regardless of which MIDI
input strategy prevails!

in sequencer:
	HMIDISTRM	hStrm = m_hStrm;
	m_hStrm = NULL;	// so IsOpen fails in callback
	CHECK(midiStreamClose(hStrm));	// close stream

in mainframe:
void CMainFrame::OnDestroy()
{
	theApp.OpenMidiInputDevice(false);
	CMDIFrameWndEx::OnDestroy();
}

3/27/20 mapping property edit undo title

If you specify which property was edited, be sure to delete these two resource strings:
IDS_EDIT_MAPPING_PROPERTY
IDS_EDIT_MULTI_MAPPING_PROPERTY
and replace them with defines to zero, because they will be replaced by cases
in CPolymeterDoc::GetUndoTitle.

Also learning multiple mappings at once requires a second undo code:
UCODE_LEARN_MULTI_MAPPING
with different handling, presumably using CUndoMultiIntegerProp.
Both of the learn edit methods should move into the document.
void MappingLearn(int iMapping, DWORD nInMidiMsg);
void MappingLearn(const CIntArrayEx& arrSelection, DWORD nInMidiMsg);

4/2/20 new way

Tempo started at 110 but I like 107 better; it's more swinging and 
less overwhelming. Position -5:000 makes a song start, and there's
a possible ending around 4:17 of 5:02. And a reentry point around 3:05.

It's mostly C harmonic minor; the turnaround drops B to Bb, D to Db, 
C to Cb and back to the top, which is key of Eb (D-7b5), key of Ab,
key of Ab melodic minor (G altered or Db7#4).

A break could be interesting, with different changes. Try going
to the IV (F harmonic minor) but not every time or for same length.

Maybe first jump to a break around 1:12?

4/3/20 Unreleased tracks

Solo piano, done:
Ero Ayo 	120	5:31 	Happy Thoughts, Joy of Thinking, Device of Joy
Mo Bulu 	125	4:51 	I'm Blue
Irubo 		123	4:15 	Ritual
Ninu Meta 	90	4:21 	In Three
Ibaramu 	140	5:29 	Harmonious
Yipo 		107	4:44 	Rolls, Twist
Fun Ife		100	5:01 	For Love
Roro		120	5:00	Gently
(around 39:10)

filenames (double-check this! [checked])
Ero ABC2	10/20/19 4:13 pm
Mo Bulu		10/21/19 7:24 pm
Irubo 3 cresc	10/25/19 5:58 pm
Ninu Meta B2 run	11/7/19 6:35 pm
Ibaramu		11/21/19 4:09 am
Yipo		4/8/20 8:00 pm
Fun Ife2  	4/28/20 2:01 am
Roro		5/17/20 6:13 pm

(DO NOT USE Roro bass boost)
(CORRECTION Roro Reason file is Roro shorter ritards 8/6/20) 
NOTE that the Mo Bulu Reason file is 120 BPM, not 125 BPM. This was unintentional, alas.

Misc:
Stella By Starlight 3:34

Full band, but needs porting to Reason:
Watergate 	2:44
Tritone 	4:43
Alternation 	6:10
(around 25:20)

LCM 3/17/20: live set 004 rec1c merge B3 no intro pad
Boogie Woogie 2/1/20: range modulation merged rec1
Boogie Woogie 2/17/20: range modulation merged rec1 ac bass
Watergate 11/30/19: watergate1 pos 2c rec1

Full band, done:
Kahelo			120	5:59
Lodidi 			130	6:14
Charlie's Big Break 	130	5:10
Interago 		125	5:11 (mastered)
Kasita Mondo 		130	6:00 (mastered)
More Than Four		120	4:46
Shelter In Bass		120	4:23
Moonchego		123	8:20
Pleasant Mistake	125	5:35
LCM (drums only)	123	5:53	
Heard a Moon 		120	6:08 (was Boogie Woogie)

LCM			123	5:34	

files:
Kahelo fast intro mix	9/14/19 1:48 pm
Lodidi 2b mix		7/9/19 6:56 am
Charlie's Big Break mix	6/30/19 11:39 pm
more than four mix	4/11/20 2:31 am
Shelter in Bass mix	4/16/20 6:59 pm
Moonchego 04 mix	5/7/20 9:44 pm
Pleasant Mistake vox less chorus	5/29/20 3:24 am
LCM Kika		5/23/20 11:59 am
Heard a Moon		5/29/20 9:32 pm

4/4/20 OSX

For Time To Repeat command, GetNumberFormatEx is the problem, see test app.

4/4/20 Scale and chord modulation

Chord modulation works fine, and permits Bergonzi's "numbers" system like
ChordEase. But there's a subtle issue with indexing. Consider just scale
and index. If the index is outside the scale, it currently wraps around.
But this limits the scale to one octave! Arguably it would be more useful
if indexing out of range octaved the scale note up or down as needed.

Why should scales be limited to an octave, unless you explicitly request
that behavior by setting Range Type to octave? You can work around it by 
adding more scale modulators so that the scale covers multiple octaves, 
but that might not be intuitive. It also uses a lot of tracks.

The same issue applies to the Chord indexing; why should chords be limited
to an octave?

Changing this is complicated and could affect preexisting songs, so it's
probably better to just use multi-octave scales and chords in cases where
the difference matters. See the file "chord modulation test.plm" for a
demonstration of multi-octave scales and chords and why they're needed.

Most of time I'm using Range modulation in which case it's a non-issue.

4/8/20 docking bars to tabbed panes: visible but inactive

Re the issue with not being able to dock Piano and MIDI bars to a tabbed pane: 
I looked into that some more. It's a non-trivial issue with pros and cons that 
need consideration. But the short form is this: Currently, keeping many of the 
bars in tabbed panes may make the UI sluggish. This is especially true of the 
Step Values bar!

When a bar is docked to a tabbed pane, Windows considers it "visible" even if 
it's not the active tab, in which case it's not actually visible to the user. 
In my current implementation, such "visible but inactive" bars receive document 
updates (affecting performance), though they don't get "painted".

It seems reasonable that hidden bars shouldn't consume any CPU cycles. When a 
bar gets "shown" my code assumes it's "dirty" (meaning it wasn't being updated 
while it was "hidden"), and updates it. But there's a catch to that work-saving 
strategy: bars that show real-time events (e.g. the Piano and MIDI event bars) 
will miss events that occur while they're hidden. This is why when you show 
the Piano bar, it doesn't display notes that were already in progress by the 
time you showed the bar; it missed those events.

It's also why I had to disable tabbed docking for Piano and MIDI event bars. 
Those bars have big impact, because they oblige the sequencer to send a 
"monitoring" copy of the entire output MIDI sequence to the main thread. If 
those bars were allowed to tab-dock (or auto-hide), they would drag 
performance even though they might not be visible to the user. This seemed 
wrong to me, so I disallowed it.

It's a messy situation with competing design goals. On the one hand, it's 
nice if bars don't consume resources while they're "hidden". On the other 
hand, it's annoying that when you show the Piano or MIDI event bars, events 
that occurred while they were hidden are missing.

4/8/20 Yipo

It's in 33. The feel is 17 + 16, typically 5 + 4 * 7.

4/9/20 Sound design: what Reason drum sounds did I like back in the day?

When It Rains:
Bd_Xtc1 (pitch 27, tone 39)
Clp_Xtc6 (pitch -10, tone 23)
Sd3_Kru (pitch 26)
Tom2_Dubfire (pitch 59)
Hh3_Vintage (pitch 12)
Ride2_Xtc6 (pitch 26 or -17)
Sh_Xfile1 (pitch 22)
Oh_FatBoy (pitch 2)
Tmb_Chemical (pitch 4)

This Is Cheese:
Bd_Xtc1 (pitch 27, tone 11)
Clp_Abstract (pitch -10, tone 23)
Sd2_HardKnox
Sd2_Sexy (pitch 18)
Tom_Xtc1  (pitch 49)
CYM3 1JAZZ (pitch -22)
Sh_Xfile1 (pitch -16)
Hh_Xfile2
Oh2_Xfile2
Clp_Xtc7 (pitch 6)
Sh_Jeepkeys (pitch 4, tone 22)
Bongo_Hardknox (pitch -20 or -36)
Bd2_One2 (pitch -7 or -30)

Save the Planet:
Bd_Xtc1 (pitch 27, tone 39)
Clp_RawDirt (pitch 14, tone 23)
Sd_Numbers (pitch 20)
Sd2_Sexy (pitch 8)
Hh3_Vintage (pitch 12)
Crash_Xtc
Cb_Brassic
Sh_XFile1 (pitch 16)
Oh_FatBoy (pitch 10)
Hh2_Sexy

4/10/20 More Than Four 

Europa master volume is 12.1 dB

Try replacing "Played with eleven" with "Combined with eleven".
It would make the meaning clearer and might sound better too.

The vocal has a lot of big transients. Adding 2.5x compression
helped a lot, but it probably needs de-essing too.

The kick is solid but not sure if there's enough bass. Maybe 
try boosting the lows on the synth line.

4/11/20 Sequencer Callback taking too long

It happend quite a few times while working on "More Than Four."
Callback times as long as nearly 20ms were observed. Today I'm
not seeing it. I looped the track's peak section, and after ten
I see callback times of 31% max and 0.1% average. It's unclear 
what if anything the trouble was. Could the laptop have been 
thermally throttling the CPU? The fan was on pretty high. I 
did have Firefox running in the background without realizing 
it, and that's a big CPU-burner, so maybe that's all it was.

Just to be sure, I benchmarked the scale re-sort after voices
are dropped, to make sure the sort isn't taking too long. Over
500 samples it used 1 microsecond maximum. Not much hope of 
optimizing that!

I did find a few spots where the "fast" versions of CArrayEx
methods could be subsituted. Here's the benchmark code. It 
really ought to be easier to do this.

static STATS stat = {0, 0, DBL_MAX};
CBenchmark b2;///@@@
{
	if (bDropped)	// if any voices were dropped
		m_arrScale.Sort();	// re-sort scale tones
}
double t = b2.Elapsed();///@@@
stat.nCallbacks++;
stat.fCBTimeSum += t;
if (t < stat.fCBTimeMin) stat.fCBTimeMin = t;
if (t > stat.fCBTimeMax) stat.fCBTimeMax = t;
if (stat.nCallbacks == 500) printf("min=%f max=%f avg=%f\n", stat.fCBTimeMin, stat.fCBTimeMax, stat.fCBTimeSum / stat.nCallbacks);

4/16/20 process chord before transposing and applying range

The previous code did the transpose and range window loop
BEFORE the chord processing, but this is dumb. Why waste
time transposing notes that won't be used? Only notes that
get picked by the chord should be processed further. This
this potentially saves quite a bit of work, e.g. with a 
hepatonic scale and tetrachords (the usual case) nearly 
half the scale tones were transposed and range windowed 
needlessly. It's an easy fix, and testing shows no impact
on the output data.

4/16/20 Sequencer Callback taking too long: profile results

AddTrackEvents is consuming 99.9% of the CPU time, no surprise
there. But what IS a surprise is that the new scale / chord 
code doesn't seem to be the problem; it's consuming less than 
1% of the time. So what's taking so long in AddTrackEvents? 

Using the main section of "Shelter in Bass" as a test file.

Let's try SumModulations next. Sure looks like SumModulations
related to the problem.

RecurseModulations is also seems to be part of the problem.
Interesting! In a typical run, the callback max time was 16.5ms,
and the RecurseModulations max time was 6.7ms, or about 40% 
of the total max time.

Next let's try profiling around AddTrackEvents's calls to 
RecurseModulations and SumModulations.

Weird! In this case callback max time is 6.9ms but the 
profiled calls are only 75 microseconds. Something isn't 
adding up here. Could it be a thread issue? Could the 
callback be getting preempted? Better check the callback
thread's priority. Nope, the callback thread priority is 
running at priority 2 (Highest), as expected.

Aha! Examining the threads window reveals that the callback
thread has competition, from an unidentified worker thread in 
winmm.dll running at Time Critical priority. So obviously if
that thread blocks for any significant length of time, the
sequencer callback will take too long. This suggests a test.
Let's try disabling the output, i.e. do all the usual work
but send an empty buffer to the MIDI stream handle. If the
Time Critical thread is consuming CPU time in proportion to
the density of the output MIDI stream, this test should 
eliminate that thread as a source of preemption or blocking.

We still need to output the NOP to get callbacks. Just to be
sure the optimizer doesn't outwit the test somehow, we'll 
output the real buffer to the app's queue for the Piano bar, 
then set the buffer to contain only NOP and output it to 
the MIDI stream.

OK. Sure looks like the Time Critical thread is the culprit!
After five minutes, callback max time is 2.6ms, and profile
around AddTrackEvents (but inside critical section) is 2.1ms.

Let's wait and see if the callback ever takes too long. So,
after 15 minutes, the numbers are 3.9ms (callback) and 3.3ms
(AddTrackEvents), and the "callback took too long" error
hasn't occurred. If these numbers are to be believed, the 
only solution is to increase the latency. For very dense 
sequences, the mysterious Time Critical winmm thread is 
taking long enough to disrupt callback scheduling.

If the profile were to include entering the critical section, 
would the callback and AddTrackEvents benchmarks match even 
more closely? Answer: Yes, now they're consistently within 
a few microseconds of each other. So AddTrackEvents is most
of the app's work, but we already knew that.

Contention for the track critical section can't be the issue,
because the only other contenders are the editing functions 
in CSeqTrackArray, and no editing occurs during this test.

Could it have something to do with the output capture and its 
associated double buffering strategy? We could try hiding the
Piano bar. Doubtful but it's an easy test. Answer: No.

Could it be related to loopMIDI or Reason? Not that it would 
matter since they're both indispensible.

It's probably not preemption or blocking, because we're running
on a quad-core CPU and the multimedia thread doesn't share any
critical sections with our callback. It's more likely that the
start of our callback is somehow delayed, perhaps because the 
multimedia driver is overworked by excessive sequence density.

4/16/20 RecurseModulations optimization

Since we've established it does a lot of the work, further
optimizing this method might be helpful. calling GetStepIndex
before the recursion is wasteful because the recursion
result might do a continue, in which the iModStep returned
by GetStepIndex isn't used.

Seems like we could just move nPosMod2 outside the recursion
scope, init it zero, and then the GetStepIndex can move after
the recursion. Same idea in SumModulations.

Better check the output of this change carefully! [seems ok]

4/17/20 mix notes

Shelter In Bass: string part is a bit too loud according to 
Kim and I tend to agree, maybe -2 dB? Or EQ it differently?
This is a question for Mike. But it's masking the piano 
and percussion too much.

4/18/20 more on callback taking too long: memory usage?

Investigate whether the callback thread is allocating memory. 
It shouldn't be. Once all the various dynamic arrays are big 
enough they should stay put provided the sequence density
doesn't increase.

int MyAllocHook(int nAllocType, void *pvData, size_t nSize, int nBlockUse, 
				long lRequest, const unsigned char * szFileName, int nLine)
{
	// need this to call printf else infinite loop since printf allocates
	if ( nBlockUse == _CRT_BLOCK )
		return( TRUE );
	if (GetThreadPriority(GetCurrentThread()) == THREAD_PRIORITY_HIGHEST) { // only sequencer callback thread
		printf("%lld %s %d\n", nSize, szFileName, nLine);
	}
	return true;
}

// in InitInstance
_CrtSetAllocHook(MyAllocHook);

Result of this test is negative: the sequencer callback thread 
doesn't do any unexpected memory allocations.

To write to console without using printf, can do some variant of this:

		static COORD coord;
		DWORD	out;
		TCHAR	buf[16];
		_i64tot_s(nSize, buf, _countof(buf), 10); // for 64-bit vals
  		WriteConsoleOutputCharacter(GetStdHandle(STD_OUTPUT_HANDLE), buf, _tcslen(buf), coord, &out);
		coord.Y++; // need to handle scrolling or at least wrap around!


4/18/20 epic time wasting in AddTrackEvents

1. Muted tracks process all of their modulators, pointlessly.
2. Modulator tracks process all their modulators, pointlessly.
3. When a mute modulator returns true, processing continues, pointlessly.

No wonder the sequencer callback is taking so long!
All of these are easily fixable with a slight refactoring.

One unexpected result of the enhancements: if a track has many 
modulators, and some of them are mute modulators, performance will 
likely be improved if the mute modulators come first in the modulator
list. This is because when a mute modulator mutes the track, the 
"goto next_step" aborts any further processing for the current step, 
including the (possibly recursive) evaluation of other modulators. 
That's potentially a significant savings. It could be a reason to 
sort modulators by modulation type instead of by source track.

4/18/20 high-frequency memory reallocation

OnUpdateEditUndo and OnUpdateEditRedo reallocate memory whenever
the mouse moves, due to loading the undo or redo prefix from a string 
resource and then appending the edit-specific title to it. This work
can be avoided by overloading CUndoManager::OnUpdateTitles and doing
the prefixing there instead. But that's not all: CToolCmdUI::SetText
also creates a CString, in order to remove the trailing label, and
CToolCmdUI is a private class buried in afxtoolbar.cpp. And it's
renamed to CMFCToolBarCmdUI via #define, nice.

Then there's all the stuff that gets updated during play:

CClientDC in CStepView::UpdateSongPosition: how to avoid this?
Can be avoided by using GetDC and ReleaseDC instead of CClientDC,
but it makes a mess of the drawing code; HDC must be passed around.
CClientDC appears to allocate a large buffer, on the order of 14K.

CClientDC in in CSongView::UpdateSongPos: how to avoid this?
GetDC/ReleaseDC isn't enough because UpdateWindow is called and that
sends a paint message to CView::OnPaint which instantiates a CPaintDC.
Still it might save time to construct only one DC wrapper instead of two.
Also CClientDC is constructed long before it's used, not that it matters.

Updating the song position and time also reallocates memory, due to 
CString. The Format statement alone is between 5 and 10 microseconds.
The sequencer methods could write to fixed-size buffers instead, but
this makes a mess of various things and will attract bugs.

And then there's the mysterious message loop that occurs when the
mouse hovers over a list control, or over a property sheet item that
isn't abbreviated. It allocates at high frequency, no doubt due to
calling FromHandle. It's somehow related to tool tips. There's also
some connection to a mysterious timer message with event ID 3. That's
not an official AFX timer ID, and according to the message hooks it
appears to be posted rather than sent, making the source impossible to
determine. But it's probably sent by a tool tip control, possibly one
associated with the list control. There are dozens of unidentified tool
tip controls associated with the app, but having no parent window.

This timer ID 3 issue or whatever it is also occurs in a stock 2012
app if you hover the mouse over one of the tree controls. The tree
controls also have item tooltips. Another sign it's tooltip related.

4/21/20 Offset modulation

MODTYPEDEF(Offset)

case MT_Offset:
	arrMod[iModType] += nStepVal;	// offset is unsigned
	break;

// declare evt much later! REMOVE these two lines:

if (nDurSteps) {	// if at start of note
	CMidiEvent	evt;
	evt.m_nTime = nEvtTime;

// need this instead:

	nNote = CLAMP(nNote, 0, MIDI_NOTE_MAX);
	CMidiEvent	evt;
	evt.m_nTime = nEvtTime + arrMod[MT_Offset];
	evt.m_dwEvent = MakeMidiMsg(NOTE_ON, trk.m_nChannel, nNote, nVel);
	if (evt.m_nTime < m_nCBLen)	// if note starts during this callback
		m_arrEvent.FastInsertSorted(evt);	// add note to sorted array for output
	else {	// note starts during a subsequent callback
		evt.m_nTime += nCBStart;	// convert to absolute time
		m_arrNoteOff.FastInsertSorted(evt);	// schedule delayed note
	}
	int	nDuration = nDurSteps * nQuant + trk.m_nDuration + arrMod[MT_Duration];	// add duration offset
	if (nDurSteps & 1) {	// if odd duration
		if (bIsOdd)	// if odd step
			nDuration -= nSwing;	// subtract swing from duration
		else	// even step
			nDuration += nSwing;	// add swing to duration
	}
	nDuration = max(nDuration, 1);	// keep duration above zero
	evt.m_nTime = nAbsEvtTime + arrMod[MT_Offset] + nDuration;	// absolute time
	evt.m_dwEvent &= ~0xff0000;	// zero note's velocity
	m_arrNoteOff.FastInsertSorted(evt);	// add pending note off to array
}

5/6/20 Moonchego

Watch out for parameters getting reset by mapping, especially in Reason.
The inital synth levels should be:
drums: 100
pad: 0.0 dB
lead: 93
bass: 100
vox: 100

5/17/20 converting tetrachords from a flat list to a four-row table via Excel

assuming the list of note values is in column A, set cell B1 to this:
=INDEX($A:$A,(COLUMN(B1) - 2) * 4 + MOD(ROW(B1) - 1, 4) + 1,1)
and then copy/paste as needed
To convert cells from note numbers to MIDI note names, use this:
=CONCATENATE(CHOOSE(MOD(A1, 12)+1,"C","Db","D","Eb","E","F","Gb","G","Ab","A","Bb","B"),QUOTIENT(A1,12)-1)

5/21/20 Roro vs. Roro bass boost

In Roro bass boost, the bottom note has the same velocity as the upper three notes, instead of -10.
Kim complained about wanting the lower notes to be a bit louder, but I don't hear it that way.
I like the softness of the original. The bass boost version feels too heavy-handed on the low notes.
The title means "gently" so it should be gentle. But perhaps this choice should be revisited during mastering.

5/21/20 LCM mix

Claps weren't correctly assigned to individual mixer channel in the exported wav file; could affect pan / level.
Pad seems overly loud and attack seems too slow?

6/9/20 override MIDI input channel

Option indices are zero origin, so must reserve zero for none.

if (theApp.m_Options.m_Midi_nInputChannel > 0)
	dwEvent = (dwEvent & ~0xf) | (theApp.m_Options.m_Midi_nInputChannel - 1)

Overriding dwEvent post-mapping seems easy at first.
But it introduces inconsistencies, because dwParam1 is also used:
1. The MIDI input bar will show the original channel. Maybe OK?
2. Recorded MIDI input will have the original channel. Confusing.
The latter is hard to fix because changing the channel for an input message
that was mapped will break the mapping when the recording is played back.
Overriding the channel pre-mapping always breaks the mappings, not only 
during playback of a recording, and breaks learn mode too, so it's a mess either way.

6/11/20 go to convergence

This method has questionable utility and is no substitute for a phase map.
It can lock up the program if nConvSize is big and the song uses a lot of 
different prime lengths. This could be demonstrated with "Ala Aye". If the 
phase bar is hidden, the orbit array must be explicitly updated first.

void CPhaseBar::GotoNextConvergence(int nConvSize, bool bReverse) const
{
	CPolymeterDoc	*pDoc = theApp.GetMainFrame()->GetActiveMDIDoc();
	if (pDoc != NULL) {
		int	nSelected = 0;
		int	nOrbits = m_arrOrbit.GetSize();
		for (int iOrbit = 0; iOrbit < nOrbits; iOrbit++) {
			if (m_arrOrbit[iOrbit].m_bSelected)
				nSelected++;
		}
		int	nPos = static_cast<int>(m_nSongPos);
		int	nDelta = bReverse ? -1 : 1;
		if (nSelected)
			nConvSize = min(nConvSize, nSelected);
		else
			nConvSize = min(nConvSize, nOrbits);
		while (1) {
			nPos += nDelta;
			int	nHits = 0;
			for (int iOrbit = 0; iOrbit < nOrbits; iOrbit++) {
				const COrbit& orbit = m_arrOrbit[iOrbit];
				if (!nSelected || orbit.m_bSelected) {
					if (!(nPos % orbit.m_nPeriod)) {
						nHits++;
						if (nHits >= nConvSize) {
							pDoc->SetPosition(nPos);
							pDoc->UpdateAllViews(NULL, CPolymeterDoc::HINT_CENTER_SONG_POS);
							return;
						}
					}
				}
			}
		}
	}
}


6/13/20 negative input to modulo function, handled correctly

int	m = 5;
for (int i = 10; i >= -10; i--) {
	int	x;
	if (i >= 0)
		x = i - i % m;
	else
		x = i - (i + 1) % m - m + 1;
	printf("%d\t%d\n", i, x);
}

10      10
9       5
8       5
7       5
6       5
5       5
4       0
3       0
2       0
1       0
0       0
-1      -5
-2      -5
-3      -5
-4      -5
-5      -5
-6      -10
-7      -10
-8      -10
-9      -10
-10     -10


6/16/20 Passion for Numbers ( <34#s )

Front cover: 
		double	fMidX = szCanvas.cx / 2 - 30;

		for (int iCol = 0; iCol <= nCols + 1; iCol++) {

	CSize	szBitmap(3780, 3685);

Back cover:
		double	fWidth = arrLen[iRow] * (fXScale * 0.969);


	CSize	szBitmap(3839, 2693);

Text:

CHRIS KORDA

PASSION FOR NUMBERS

A1. Ero Ayo   A2. Mo Bulu   A3. Irubo   A4. Ninu Meta
B1. Ibaramu   B2. Yipo   B3. Fun Ife   B4. Roro

Big text is 48 pt Helvetica Black

Label:

1087 x 1087 (safe area)
then expand to 1229 to reach edge of bleed


6/20/20 Modulation bar selection issues

Delete / Cut: removing selection is no problem, just add Deselect call:

pDoc->UpdateAllViews(NULL, CPolymeterDoc::HINT_MODULATION);
m_grid.Deselect();

Paste / Insert: easy if Show Differences is OFF, but harder if it's on

// this ONLY works if Show Differences is OFF
int	nPrevMods = m_arrModulator.GetSize();
pDoc->UpdateAllViews(NULL, CPolymeterDoc::HINT_MODULATION);
if (iSelItem < 0)	// if no selection
	iSelItem = nPrevMods;	// appended
if (iSelItem >= 0)
	m_grid.SelectRange(iSelItem, nMods);
else
	m_grid.Deselect();

What to do if Show Differences is on? Search m_arrModulator for the first pasted element?

Restoring selection after undo is not currently possible because the track selection 
might be different when the undo occurs. This is yet another reason why track 
selection should be an insignificant edit that's saved systematically instead 
of in an adhoc way. Also the Show Differences flag would need to be saved and 
restored because like track selection it also affects which modulators are shown.


6/23/20 MFC Customize subsystem is incompatible with renumbering resource IDs

The easy experiment: 
1. Add some commands to the standard toolbar. Create some new keyboard shortcuts too.
Now exit the app and edit the project as follows.
2. Create a new menu command called ID_APP_ALPHA (guaranteed to be the first command) 
and renumber the resource IDs. This shifts every app-defined command ID up by one.
3. Rerun the app. The custom toolbar commands still have the same names, but they no
longer trigger the intended commands. They may also move around on the toolbar. The 
keyboard shortcuts are also shifted.
All of this is predictable because the Customize system stores command IDs. So that's
an epic fail. When the app is next upgraded, if the upgrade adds new commands (or 
removes or renames existing commands), any customizations the user made will likely 
be partially or completely scrambled. That sounds like a good reason not to allow it, 
but maybe it could be OK to allow it and put a warning in the manual?

The only other solution I can think of is to strictly adhere to this procedure:
1. Avoid renaming or deleting existing commands, or if it's unavoidable, retain the 
command's original ID by preserving the corresponding hint resource string (though
the string can be empty or just contain a single space, or "X" or whatever).
2. New commands must have a special prefix that provides versioning and forces newly
added commands to the end of the renumbered command ID range. For example, the
prefix ID_a??_ could work. The first upgrade would use this prefix:
ID_a00_A_NEW_COMMAND
ID_a00_NEW_COMMAND_IN_THE_SAME_RELEASE
The next upgrade would use an incremented prefix:
ID_a01_A_NEW_COMMAND_IN_A_SUBSEQUENT_RELEASE
This works so long as command names don't start with a lowercase letter. Since the
version number can use not only 0-9 but also A-Z and a-z, the prefix allows around
100,000 upgrades (26 * 62 * 62) which seems adequate for a production app. If you 
exceed 3,844 upgrades (after ID_azz_) you roll over to ID_b00_ and so on. This
scheme doesn't require any changes to the venerable resource renumbering app.

Another issue: menus and especially context menus often have groups of options (with 
checkboxes or radio buttons) that depend on having a contiguous range of command IDs.
Adding a command to such a group also potentially shifts the app's command IDs. To 
maintain the above procedure in such cases will be significant work: the entire group
has to be copied to a new range of command IDs, leaving the old range reserved. This
will result in a lot of "dead wood" i.e. legacy command hint resource strings. In
practice there seem to be relatively few such cases (search for "ON_COMMAND_RANGE").

One way to check for compatibility is to file compare the old and new Resource.h.
If you see any changes other than appends, it's incompatible. Clearly this approach
would be unworkable during development, but once the app is in production it might 
be reasonable if Customize is important enough to warrant it.

7/5/20 Multiple windows of a document

There are two related issues:
if document has multiple windows, changing view type affects first window instead of active one
if document has multiple windows with different view types, properties may be inconsistent between windows

The first issue might not be solvable with the current architecture. It's unclear.

The second issue isn't too bad with song view and track view. Here the only obvious
symptom is that the Track/Length command doesn't work (it's disabled).
But with live view and track view, there are more problens.
Muting/unmuting tracks in live view doesn't update track view (but the reverse works).
Same for selecting a preset or soloing in live view.

7/10/20 Menu key stops working after switching documents via Ctrl+Tab

To replicate it:
1. run app
2. create a 2nd document
3. select the 1st document via left-clicking its tab
4. select the 2nd document via Ctrl+Tab
5. Press Alt+F4 to exit the app; nothing happens

7/20/20 Undo test with changing view type

	if (!Random(10)) {
		POSITION pos = m_pDoc->GetFirstViewPosition();
		CView	*pView = m_pDoc->GetNextView(pos);
		CChildFrame	*pFrame = DYNAMIC_DOWNCAST(CChildFrame, pView->GetParentFrame());
		pFrame->SetViewType(Random(2)); // track or song view only
	}

Note that including Live view will cause checksum errors, because Live view 
potentially changes track mutes without creating an undoable event, in order 
to keep parts and presets consistent.

7/22/20 Showing dockable bars menu

The work is done by this apparently public method:
CDockingManager::BuildPanesMenu(CMenu& menu, BOOL bToolbarsOnly)

8/1/20 mysterious MIDI error 65

#define MIDIERR_STILLPLAYING  (MIDIERR_BASE + 1)   /* still something playing */

It occurs when stopping the sequencer. It almost never happens under Windows,
but it appears to happen every time on a Mac with Wine Bottler. The error's
definition is "Cannot perform this operation while media data is still playing. 
Reset the device, or wait until the data is finished playing." That should tell
us something. The problem is we don't know which MIDI streams function is
causing it. We stop the stream, unprepare the buffers, and finally close the
device. Hypothesis: we're attempting to unprepare a buffer before the driver
has finished with it.

Does midiStreamStop return before the buffers are returned, or wait? The Windows
documentation is a bit vague on this point. "When you call this function, any 
pending system-exclusive or stream output buffers are returned to the callback
mechanism and the MHDR_DONE bit is set in the dwFlags member of the MIDIHDR 
structure." Notice it doesn't say exactly WHEN the output buffers are returned.
It that's the issue, you'd think making the callback time much longer would 
provoke the error, but I tried that and it didn't make any difference.

If this really is the issue, it's an easy fix. In the MIDI callback's MOM_DONE
case, if the stopping flag is set, increment a count of done buffers. In the Play
function, after stopping the stream, wait for this count to reach the number of
buffers, while sleeping periodically. Or more properly, in the callback, when the
done buffer count reaches the desired number, set an event, and in the Play
function, wait on the event with a timeout. [Note: the number of buffers in
use at any given time is too uncertain for this plan to be safe.]

If the unprepare is doing it, another possible solution would be to loop on
unprepare as long as it returns MIDIERR_STILLPLAYING, with a Sleep to avoid
burning CPU. This is more or less shooting the messenger.

MMRESULT CSequencer::UnprepareHeader(int iBuffer)
{
	int	nTries = 100;	// number of tries
	int	nRetryPause = 10;	// retry pause in milliseconds
	MMRESULT	nResult = MMSYSERR_NOERROR;
	for (int iTry = 0; iTry < nTries; iTry++) {	// for each try
		nResult = midiOutUnprepareHeader(
			reinterpret_cast<HMIDIOUT>(m_hStrm), &m_arrMsgHdr[iBuffer], sizeof(MIDIHDR));
		if (MIDI_SUCCEEDED(nResult) || nResult != MIDIERR_STILLPLAYING)	// if success or fatal error
			break;	// exit loop
		Sleep(nRetryPause);	// buffer still playing; pause and try again
	}
	return nResult;
}

Without a Mac to test against, there's no way to reproduce the issue, so no way
to prove the above solution would help. I would tend to think it's a Mac / Wine
MIDI implementation issue except I have very rarely seen it under Windows. It's
possible it's driver-dependent.

The easiest test is to just put a Sleep(1000) after midiStreamStop and see if
that fixes the problem in MacOS. I give it even odds.

Another hypothesis is that calling midiStreamOut from within the callback is 
actually a terrible idea, even though it seems to work. MSDN discourages any
calls to MIDI functions, but examples in the wild vary. Some examples only set
an event in the callback, others call midiStreamOut as we do. The event is a
royal pain in the ass because it means creating a whole other worker thread
and incurring all the overhead and complexity associated with that.

8/7/20 show step selection in velocity view

	int	iStepSelRight;
	int	iStepSelLeft = 0;	// avoids uninitialized variable warning
	for (int iSel = 0; iSel < nSels; iSel++) {	// for each selected track
...
		if (iTrack < rStepSel.bottom && iTrack >= rStepSel.top) {	// if track within step selection
			iStepSelLeft = rStepSel.left;
			iStepSelRight = rStepSel.right;
		} else
			iStepSelRight = 0;	// left doesn't matter if right is zero
		for (int iStep = iFirstStep; iStep <= iLastStep; iStep++) {	// for each step
...
				COLORREF clrBar;
				if (iStep < iStepSelRight && iStep >= iStepSelLeft)	// if step within step selection
					clrBar = clrBarSelected;
				else
					clrBar = clrBarNormal;
				pDC->FillSolidRect(rBar, clrBar);

It works, but it makes the step view responsible for invalidating the velocity view. 
Depending on how that's done it adds significant overhead especially when dragging
the step selection. It's easy to just invalidate the entire velocity view in all
cases, but it's inefficient.

8/12/20 subtitles

Easiest for YouTube is .sbv format:
hh:mm:ss.000,hh:mm:ss.000
blah blah

hh:mm:ss.000,hh:mm:ss.000
more blah blah

[and so on]

Export the song from Polymeter, using Tools/Export/Song
Isolate the vocals note on events and load them into Excel
Excel funtion to convert MIDI MBT to time in days:
=(a1/120*(60/T)+0.1)/86400
where T is tempo (assuming 120 PPQ)
and 0.1 is time offset in seconds
for Overshoot subtitles were too early, +1/10 of a second was about right
86400 is seconds in a day

Format the times with this Custom format:
[h]:mm:ss.000

Now the hard part: getting the on/off times in pairs, and matching them 
with the lyrics. Couldn't figure out how to do it with Excel, by hand is a PITA.


8/12/20 Apologize chorus lyrics order confusion:

The audio goes like this:

Apologize
For the dying seas
Apologize
For the clear-cut trees
Apologize
For needless birth
Apologize
To what’s left of earth
Apologize
For overpopulation
Apologize
For mass migration
Apologize
To the United Nations
Apologize
To future generations

9/5/20 Note Number input mapping type

It might be helpful to have a note mapping that gets the data byte from the
note number rather than the velocity. To be useful it should also ignore note
off messages. This would permit a keyboard to be used for Preset changes, for 
example. However by default it would utilize the entire range of MIDI notes 
on the given channel, which would be very wasteful in most cases. This could
be addressed by re-purposing the Range Start / End parameters to be the range
of notes included in the mapping, but that's a confusing kludge.

This still doesn't solve the problem of how to use a keyboard for toggling
binary switches, for example Mute states. Here the problem is who owns the
toggle state. For Mute states specifically the Mute state itself could be
be its own toggle state, and this is advantageous because it will remain
consistent with the UI. But for a general solution that won't work. The
same applies to Part mute states.

The toggle problem exists because many common controllers don't support 
reprogramming what their keys or buttons send, or make it very difficult.

Another approach to these problems is instead of trying to make a general 
solution, create a custom implementation of Part and Preset mappings for 
ranges of MIDI notes. This would be more work to develop, because it would
need custom UI, file storage, etc. but would allow the most specificity.

Still another approach would be to add a new Switching attribute to the 
mapping parameters, consisting of Normal or Toggle. In Toggle mode, the
target parameter would toggle between 0 and 127. This would solve control
of Parts via MIDI keyboard, and would also be useful in other cases for
example toggling the steps in a track.

9/9/20 Gray Code

Try using gray code in SetPerm for generating permutational harmony.
It might generate smoother transitions.

uint BinaryToGray(uint num)
{	
	return num ^ (num >> 1);
}

9/23/20 ordering of tracks within a preset

A preset's track indices aren't necessarily in ascending order. This is 
easily demonstrated by making a preset and then reordering its tracks.
In order to interdisperse presets with tracks in track order, it's
essential to know the minimum track index for each preset, and this
will need to be determined by iterating. The only other way around it
would be to keep presets always sorted but this would be much more
costly and there's no need.

Arguably the array class should include methods to find the minimum
or maximum element.

Also, CPart::GetTrackRefs should return the number of tracks that
belong to parts. This would make it possible to preallocate the
part array CLiveView::Update(), instead of using Add.

10/18/20 GraphViz 2.44.1 not working

Tried with both 32-bit and 64-bit installers. In both cases dot.exe fails with this error:
There is no layout engine support for "dot"
Perhaps "dot -c" needs to be run (with installer's privileges) to register the plugins?

Their installer should take care of this but it doesn't.

Also 2.44 often bungles the graph size, causing the graph to be too small or clipped.

Version 2.38 doesn't have either of these problems.

10/25/20 converting BalaGray output to note track

In Excel, use this formula: =IF(A1=B1,164,36)

11/13/20 started new piano piece 

The A section uses using new balanced gray code change ringing for arpeggiation. 
The changes are 4-bell pattern 26, which is a rotation of [6,12,6]A.

The B section (started 11/15) similarly uses 4-bell gray [9,6,9] for arpeggiation.
It also uses the balanced gray code sequence [2222] for alterating the tetrachord,
though the sequence is rotated five steps right, starting with [0010] instead of [0000].
The scale is an octatonic (0235678a) which is potentially quite dissonant, due to
having three semitones in a row; its prime form is 8-22 (0123568a), IV 465562.
The unaltered tetrachord is [0246] which yields [0,3,6,8] which is a diminished
triad plus a b6. The result is surprisingly consonant despite the three semitones,
because the alteration scheme avoids sounding more than two semitones at once 
and also keeps them from abutting each other.

(0235678a) [0246] C

0010	[0256]	(0378)	Abmaj7
0011	[0257]	(037a)	C-7
0111	[0357]	(057a)	C7sus4
0101	[0347]	(056a)	Gbmaj7#4
0001	[0247]	(036a)	C-7b5
0000	[0246]	(0368)	Ab7
1000	[1246]	(2368)	Ab7#4
1100	[1346]	(2568)	Ab13#4
0100	[0346]	(0568)	Ab13
0110	[0356]	(0578)	F-9
1110	[1356]	(2578)	F-6
1111	[1357]	(257a)	G-7
1101	[1347]	(256a)	Gbmaj7#5
1001	[1247]	(246a)	Eb-maj7
1011	[1257]	(247a)	Ebmaj7
1010	[1256]	(2478)	Abmaj7#4

This was a lucky accident. The intention was to use the symmetric diminished scale 
(0235689b), but the last three notes were mistakenly entered sharp by a semitone. 
The diminished scale would have been much more dissonant. Making and recognizing 
"pleasant mistakes" is a crucial aspect of the artistic process. This example also
suggests that even scary-looking pitch class sets with lots of semitones can
produce relatively consonant tonalities if they're handled carefully.

11/17/20 consistency test for CStepValuesBar ScanStep and FormatStep methods

	static const int arrFormat[] = {	// array of formats to test
		STF_SIGNED,
		STF_NOTES | STF_OCTAVES,
		STF_TIES,
		STF_HEX,
	};
	int	nPrevStepFormat = m_nStepFormat;	// save format member
	int	nPerms = 1 << _countof(arrFormat);	// all permutations of test formats
	for (int iPerm = 0; iPerm < nPerms; iPerm++) {	// for each permutation
		m_nStepFormat = 0;
		for (int iFormat = 0; iFormat < _countof(arrFormat); iFormat++) {	// for each format
			if (iPerm & (1 << iFormat))	// if current permutation includes this format
				m_nStepFormat |= arrFormat[iFormat];	// add format to mask
		}
		int	nVals;
		if (m_nStepFormat & STF_TIES)
			nVals = 256;
		else
			nVals = 128;
		int	iVal;
		for (iVal = 0; iVal < nVals; iVal++) {
			TCHAR	sz[10];
			FormatStep(sz, _countof(sz), iVal);
			int	nStep;
			ScanStep(sz, nStep);
			if (nStep != iVal) {
				printf("ERROR: %d != %d\n", iVal, nStep);
				break;
			}
		}
		printf("0x%02x: %s\n", m_nStepFormat, iVal >= nVals ? "pass" : "FAIL");
	}
	m_nStepFormat = nPrevStepFormat;	// restore format member


11/17/20 nasty focus issue in CGridControl

The symptom that started the inquiry: Pressing and releasing the Alt key and then 
left-clicking a grid control item crashes or throws an Improper Argument exception.
It was fixed in CGridControl::OnLButtonDown by calling SetFocus before calling
EditSubitem. But there's more. It turns out SetFocus can fail, for example if 
the window that's losing focus resets focus. One scenario: You're editing tempo
in the master properties, and you enter an invalid tempo, and then click a grid
control item. The MFC property grid's kill focus handler is called and tries to
validate the property using DDV. The property fails validation in this case, so
_AfxFailMinMaxReal displays a modal error dialog and calls CDataExchange::Fail()
which sets the focus to the offending control.

Thus when SetFocus returns, the grid control doesn't actually have focus. Most
likely the modal dialog has focus. If the grid control proceeds to create a popup
edit control, that control will immediately receive a kill focus message after it
tries and fails to wrest focus from the modal dialog. The popup edit's kill focus 
handler will delete the edit control right out from under the rest of its creation
code and we're primed for a crash.

This can happen because modal dialogs create a message subloop, allowing messages
to be processed, but also prevent other windows from receiving focus.

The solution is to verify that SetFocus actually worked, and if didn't, don't
create the popup edit control. The easiest way to do that is by checking that
the result of ::GetDlgCtrlID() equals m_hWnd.

11/18/20 more work on auto-hide and attach

For the history, see the entry from 4/8/20. It's time for a drastically new approach 
to visibility in CMyDockablePane. Instead of determining visibility via the base class
Visible() method, we'll instead use the WS_VISIBLE style, i.e. the unadorned window
visibility. This means CMyDockablePane::OnShowChanged will now be called:

1. When a tabbed pane becomes active or inactive.
2. When an auto-hide pane becomes active or inactive.

Hence we can finally enable tabbing and auto-hide for the Piano and MIDI Input/Output
panes. Note however that these panes will NOT receive MIDI data while inactive. In
terms of output capture, making a tabbed or auto-hide pane active is the same as 
showing the pane from the main menu. Any previous history is lost.

In general performance should improve, because the app will no longer wastefully
send document updates to invisible panes. This is particularly beneficial for the
Step Values pane, which can use considerable CPU time.

On the other hand, switching between tabbed panes may take longer than before
in some cases, because pane's update is now deferred until the pane is shown.
Again this is good overall, especially with tabbed panes, because only the 
visible pane that needs updating is actually updated, as opposed to before,
when panes were frequently updated even though they were invisible.

This change also fixes the Phase pane not animating when it's in auto-hide mode.

11/22/20 looping playback

in OutputMidiBuffer, inside the critical section but before the tempo event processing:

		if (m_bIsLooping && m_nCBTime + m_nCBLen >= m_nLoopTo) {
			SetPosition(m_nLoopFrom);
		}

It's that easy. The hard part is the UI. The loop enable can be a Transport
command and a toolbar button. Ctrl+Alt+L is available. The loop range could
be members of the Properties bar. It would be nice if there were a visual
way of setting the range. You could get the range from the step selection
in Track view, or the cell selection in Song view. This has the advantage
of being inherently quantized. For example if looping is off and the loop
command is given when a rectangular selection exists, in that case you
could set the loop range from the width of the rectangular selection.

The only problem with this idea is that in Track view, a multi-track
rectangular selection doesn't necessarily have a uniform width in terms of
time, for example if the tracks have different Quants. It's a degenerate
case. You could exclude it, or take the range from the first track.

It would also be nice if the ruler indicated the loop range while looping,
for example using a colored rectangle. Extra credit if you can modify the 
range by dragging either end. It's important to indicate the range somehow
because the rectangular selection is emphemeral.

The loop enable and range should be saved in the document.

Note that the caller must ensure that m_nLoopFrom < m_nLoopTo, otherwise
playback will get stuck due to setting m_nLoopFrom on every callback.

11/29/20 resource ID usage in MFC UI customization methods

All of these keys are contained in "Workspace".

"Keyboard-0/Accelerators" contains a binary table of accelerator structs.
The struct has a similar layout to the ACCEL struct except the command ID
is a WORD instead of a DWORD, making the size six bytes instead of eight.

struct REG_ACCEL {
    WORD	flags;
    WORD	key;
    WORD	cmd;
};

"MFCToolBar-59392/Buttons" contains the current toolbar buttons.

"MFCToolBar-593980/Buttons" contains the default menu items.

"MFCToolBar-5939871/Buttons" contains the document menu items.

In each case the "OriginalItems" subkey contains a backup.

These button lists are binary CArchives from serialization.

11/30/20 impact of command ID renumbering on UI customization

It's pretty clear that if existing command IDs change or shift, any UI
customizations will be badly scrambled. There doesn't seem to be any
practical way to correct this, so we're left with only two options.
1) reset customizations every time app starts (the current solution),
or 2) name new commands in such a way that renumbering the command IDs
never causes existing IDs to change or shift.

Command ID ranges are especially problematic because they require a
contiguous range of values, hence should a command need to be added
to the range, the range must expand, unavoidably shifting up any
command IDs above the range. The question is, how many such cases 
currently exist? This is answered by searching the source files for
"D_RANGE(" and the answer is that most of the cases use reserved ID
ranges based on ID_APP_DYNAMIC_SUBMENU_BASE and aren't impacted by
Resource.h. Graph Scope, Graph Layout, Filter Channel, Filter Message,
Output Channel, Piano Size, and Convergence Size are in that category.
That leaves two cases, both in the Step Values bar: Layout Columns 
(Steps vs. Tracks), and Delimiter (Comma vs. Tab). These two cases
could fairly easily be modified to use ID_APP_DYNAMIC_SUBMENU_BASE,
after which all the remaining command IDs in Resource.h would be 
independent of range. Thus there would be no remaining obstacle to
adding new command IDs with a prefix such as "ID_z00" in order to 
ensure that they sort to the end of the command ID list; see 6/23/20.

These two cases are effectively booleans masquerading as multiple
choice, so they could also effectively be ignored.

Enabling persistence of customization is simple, just remove
(or make conditional) these statements in MainFrm.cpp:

	// ck: disable UI customization for now, it's too confusing during development
#if _MFC_VER < 0xb00
	m_wndMenuBar.RestoreOriginalstate();
	m_wndToolBar.RestoreOriginalstate();
#else	// MS fixed typo
	m_wndMenuBar.RestoreOriginalState();
	m_wndToolBar.RestoreOriginalState();
#endif
	theApp.GetKeyboardManager()->ResetAll();
	theApp.GetContextMenuManager()->ResetState();

This was tested, including the addition of several new commands,
with their names prefixed by ID_z00_ prefix, and behavior seems as
expected except that adding new commands to the main menu appears
to remove any main menu customizations.

12/1/20 spurious dubs

The problem is caused by CTrackBase::CDubArray::RemoveDuplicates()
skipping the last dub in the dub array, per this comment: 
"exclude last dub from pruning because it indicates song length."
This behavior can cause a blizzard of confusing spurious dubs,
due to each track potentially having a different last dub time.
One solution is to handle the last dub as a special case, and 
if it's a duplicate, delete the earlier dub (its predecessor)
instead of the later one (the last one in this case). This will
remove the duplicate while preserving the song length.

// after pruning loop!
	int	iLastDub = GetSize() - 1;	// reload array size as it may have changed above
	if (iLastDub > 0 && GetAt(iLastDub - 1).m_bMute == GetAt(iLastDub).m_bMute)	// if last dub is a duplicate
		RemoveAt(iLastDub - 1);	// remove earlier duplicate, preserving song length

Nope, this is a bad idea because it causes erroneous dub lengths.
So only remaining solution is to fix the code that searches for
the next/previous dub so that it ignores these degenerate cases.

1/20/21 device change notification should retry missing MIDI device and close message box if successful

This cheap trick seems to work, but it feels risky for such a minor case.

} else {	// message box already displayed
	// assume missing devices message box has sufficiently unique caption
	HWND	hWnd = FindWindow(NULL, CMidiDevices::GetErrorDlgCaption());
	if (hWnd) {	// if missing devices message box found
		INPUT	inp = {0};	// init struct to zero
		inp.type = INPUT_KEYBOARD;	// send return key to message box
		inp.ki.wVk = VK_RETURN;	// assume default button is Retry
		SendInput(1, &inp, sizeof(inp));	// retry missing device
	}
}


6/7/21 installing Visual Studio 2019 on Windows 7 SP 1 (without automatic updates)

vc_community.exe tries to install the same package five times and then fails with
"Unable to download installation files. Check your internet connection and try again."
The issue is stale certificates. Two need to be obtained from the Microsoft PKI 
(Public Key Infrastructure) page, and properly installed in trusted root.

The URL is: https://www.microsoft.com/pkiops/docs/repository.htm
And the two certificates needed are:
Microsoft Root Certificate Authority 2010
Microsoft Root Certificate Authority 2011

Download the .crt files, and for each, right click and install using Wizard,
but do NOT take the default, select "Place all certificates in the following store"
and then press the Browse button and select: Trusted Root Certification Authorities.
If you install them anywhere else it won't work.

When you install VS2019, you need to include the Windows 10 SDK and MFC.

6/7/21 compilation errors and warnings in Visual Studio 2019

optional: set warning level to 3, or suppress these warnings:
4838;4840;4456;4457;4477;4302;4311;4312

These are new warnings:
4838	conversion from 'type_1' to 'type_2' requires a narrowing conversion
4840	non-portable use of class 'type' as an argument to a variadic function
4456	declaration of 'identifier' hides previous local declaration
4457	declaration of 'identifier' hides function parameter
4477	function' : format string 'string' requires an argument of type 'type', but variadic argument number has type 'type'

These warnings were off prior to VS2015:
4302	'conversion' : truncation from 'type 1' to 'type 2'
4311	'variable' : pointer truncation from 'type' to 'type'
4312	'operation' : conversion from 'type1' to 'type2' of greater size


whitespace between concatenated string literals is now mandatory!
requires three fixes in Options.cpp:
		{_T(#group)_T("_")_T(#name), IDS_OPT_NAME_##group##_##name, IDS_OPT_DESC_##group##_##name, \

S.B.
		{_T(#group) _T("_") _T(#name), IDS_OPT_NAME_##group##_##name, IDS_OPT_DESC_##group##_##name, \
and similar for these:
		RdReg(_T("Options\\") _T(#group), _T(#name), m_##group##_##name);

		WrReg(_T("Options\\") _T(#group), _T(#name), m_##group##_##name);

7/21/21 Allow conversion of a dockable pane to an MDI child

in CMainFrame ctor:
m_bCanConvertControlBarToMDIChild = true;
It's a bad idea because the bar isn't restored properly next time you run the app.
It can lead to the app reporting the bar as visible when it's not in fact visible.

7/21/21 Customizing tabbed pane's context menu

A group of tabbed panes doesn't call CMyDockablePane::OnBeforeShowPaneMenu,
as an individual pane does. That's because the group is a CTabbedPane that's
created dynamically by CreateTabbedPane. To fix this, derive a class from 
CTabbedPane and ensure that every dockable pane calls SetTabbedPaneRTC 
with the derived class. It's not all that useful, but here's how do it.

class CMyTabbedPane : public CTabbedPane
{
	DECLARE_SERIAL(CMyTabbedPane)
protected:
	enum {
		ID_TOGGLE_MAXIMIZE = -1110	// don't conflict with IDs in CPane::OnShowControlBarMenu
	};
	virtual BOOL OnBeforeShowPaneMenu(CMenu& menu);
	virtual BOOL OnAfterShowPaneMenu(int nMenuResult);
};

IMPLEMENT_SERIAL(CMyTabbedPane, CTabbedPane, 1)

BOOL CMyTabbedPane::OnBeforeShowPaneMenu(CMenu& menu)
{
	CPaneFrameWnd*	pParent = GetParentMiniFrame();
	if (pParent != NULL) {	// if parent mini-frame is valid (implies floating)
		int	nItemStrId;
		if (pParent->IsZoomed())	// if parent window is maximized
			nItemStrId = IDS_SC_RESTORE;
		else	// parent window isn't maximized
			nItemStrId = IDS_SC_MAXIMIZE;
		menu.AppendMenu(MF_STRING, static_cast<UINT>(ID_TOGGLE_MAXIMIZE), LDS(nItemStrId));
	}
	return CTabbedPane::OnBeforeShowPaneMenu(menu);
}

BOOL CMyTabbedPane::OnAfterShowPaneMenu(int nMenuResult)
{
	if (nMenuResult == ID_TOGGLE_MAXIMIZE) {	// if toggling maximize
		CPaneFrameWnd*	pParent = GetParentMiniFrame();
		if (pParent != NULL) {	// if parent mini-frame is valid (implies floating)
			int	nShowCmd;
			if (pParent->IsZoomed())	// if parent window is maximized
				nShowCmd = SW_RESTORE;
			else	// parent window isn't maximized
				nShowCmd = SW_MAXIMIZE;
			pParent->ShowWindow(nShowCmd);	// maximize or restore parent window
		}
	}
	return CTabbedPane::OnAfterShowPaneMenu(nMenuResult);
}

int CMyDockablePane::OnCreate(LPCREATESTRUCT lpCreateStruct)
{
	if (CDockablePane::OnCreate(lpCreateStruct) == -1)
		return -1;
	SetTabbedPaneRTC(RUNTIME_CLASS(CMyTabbedPane));	// so CreateTabbedPane uses derived class
	return 0;
}

11/3/21 delaying UI to improve sync with audio

Tested using the Windows GS Synth, at 120 BPM and 120 TPQN, updating view at
50 Hz: nPos - 50 ticks seems to work the best. That's +208ms. It's difficult
to evaluate the sync in track view, but slightly easier in song view. There's
a lot of jitter, possibly as much as +/- 10 ticks. Using Reason, the sync 
difference is imperceptible.

11/5/21 duplicate view painting on app startup

The stock InitInstance does ShowWindow first and then updateWindow, and this
unsurprisingly causes a second paint message. But is it really necessary?
Reversing the order seems to work fine, and definitely gets rid of the 
duplicate paint. There could be a reason why it's necessary to do 
ShowWindow first, but if so it's hard to detect.

	pMainFrame->ShowWindow(m_nCmdShow);
	pMainFrame->UpdateWindow();

This fix does NOT solve the issue of duplicate view painting after a new or 
open document command. For opening an existing document, the duplicate paint
only occurs for Track View. This is revealing. There's something special 
about Track View. But what? Could it be that these are different symptoms
of the same problem? Or is there just additional handling for the primary
view? Whatever it is, it's aggravated when the main window isn't maximized.

Delaying the showing of CTrackView's grid control also helps a lot (in
OnCreate, create the grid control without WS_VISIBLE and post yourself
a message, and then the handler does ShowWindow for the grid control).
But it doesn't change the fact that both new and open document are
somehow causing CTrackView::OnDraw to be called twice.

11/8/21 more on startup flicker

The duplicate view update during startup is solved by moving the mainframe
show/update code (see above) to CMainFrame::OnDelayedCreate. WM_DELAYED_CREATE 
is posted rather than sent (by CMainFrame::OnCreate) and therefore isn't
handled until after the message queue settles down. The ultimate cause is 
probably late resizing of the view due to RecalcLayout or similar methods.
Using PostMessage to defer a message until after the window size stabilizes 
is a common fix for such situations.

11/8/21 deleting main frame in InitInstance is an error

The following cleanup code in the stock InitInstance is erroneous and 
causes the app to crash if LoadFrame fails, due to deleting a pointer
that was already deleted in CFrameWnd::PostNcDestroy.

if (!pMainFrame || !pMainFrame->LoadFrame(IDR_MAINFRAME))
{
//	delete pMainFrame;//@@@ ERROR! frame self-destructs in CFrameWnd::PostNcDestroy()
	return FALSE;
}

11/14/21 Phase bar's image sequence export doesn't support tempo change

This is a complex issue, because the planet angles can no longer be 
simply computed from the frame time; instead we must compensate the
angles for a series of small stepwise tempo changes. But even if we
solved this, it's not certain that our solution would match Reason's
methodology precisely enough to avoid audio/video sync problems.

11/15/21 phase video sync

Ero Ayo (for example):
The first note is at -72:60, which occurs at -0:00:36.250 or 750ms before
position 1:00. The original export starts at exactly 750ms, whereas the 
mastered version starts at 131ms, for a difference of 619ms. At 60 FPS, 
each frame is 16.666ms, and 619 / 16.666 = 37.14, so for correct sync we
should skip the first 37 frames of the image sequence. The start position 
is -73:00, and 1:000 occurs at exactly 0:37:00 (time). In the finished 
video, full convergence occurs at 37 seconds. Success!

Fun Ife: master version starts at 106.209ms
video must start at -6 frames (start pos = -1:000, skip 66 frames)

Ibaramu: master version starts at 116.265ms
Reason version starts at 428.397ms; 312ms offset, video must skip 19 frames

Atunwi: master starts at 202.665ms; initial full convergence at frame 2430
or 40.5 seconds

Irubo: master starts at 132.672ms; original Reason file starts at 0.
Change start pos to 0:000 (gaining .666 seconds or 40 frames at 60 FPS), 
export, and skip the first 32 frames (133 / 16.6 = 8, 40 - 8 = 32)

Ninu Meta: master starts at 098.752ms; video starts at 1:000, start
at 0:000 instead (-666.666ms); skip 567.914ms or 34 frames

Roro: master starts at 109.346ms; video starts at 1:000, start
at 0:000 instead (-500ms); skip 390ms or 23 frames

Yipo: master starts at 120.921ms, Reason starts at 168.704ms;
difference of 47.783ms, skip 3 frames

Mo Bulu: Note that Polymeter file is 125 BPM but master is 120 BPM!
master starts at 96.521ms, video starts at 1:000, start
at 0:000 instead (-500ms); skip 403.479ms or 24 frames


NOTE: phase videos are made with Convergence Size set to 4.

ffmpeg -framerate 60 -i [PATH]\img%%05d.png -i [PATH].wav -c:v libx265 -movflags 
+faststart -vf format=yuv420p -preset veryslow -crf 23 -acodec alac -shortest [PATH].mp4
(for greater compatibility replace -acodec alac with -b:a 320k)

11/16/21 elliptical orbits

Benchmarked pretty carefully and it's definitely not slower, despite
additional calculations for aspect ratio. If anything it's about 6 micros
faster. That's due to not recreating the solid brushes for every draw.

Ero Ayo ABC2: 16 us, 35 us, 26 us

#define BENCH_PHASE_BAR_DRAW 1
#if BENCH_PHASE_BAR_DRAW
static int m_nSamps;//@@@
static double m_fSampMin = 1e10, m_fSampMax, m_fSampSum;
#endif

LRESULT CPhaseBar::OnDrawD2D(WPARAM wParam, LPARAM lParam)
{
#if BENCH_PHASE_BAR_DRAW
	CBenchmark b;//@@@
#endif
	...
#if BENCH_PHASE_BAR_DRAW
	double t = b.Elapsed();//@@@
	m_nSamps++;
	if (t < m_fSampMin)
		m_fSampMin = t;
	if (t > m_fSampMax)
		m_fSampMax = t;
	m_fSampSum += t;
	if (m_nSamps >= 100) {
		printf("%f %f %f\n", m_fSampMin, m_fSampMax, m_fSampSum / m_nSamps);
		m_nSamps = 0;
		m_fSampMin = 1e10;
		m_fSampMax = 0;
		m_fSampSum = 0;
	}
#endif
	return 0;
}

11/16/21 mnemonics in docking windows submenu

#include "MainFrm.h"//@@@
void CMyDockablePane::GetPaneName(CString& strName) const
{
	int	iBar = m_nID - ID_APP_DOCKING_BAR_START - 1;
	static const int arr[] = {
		1, 0, 2, 1, 0, 5, 0, 0, 5, 1, 0, 5, // the horror!
	};
	CDockablePane::GetPaneName(strName);
_tprintf(_T("%s %d %c\n"), strName, iBar, strName[arr[iBar]]);
	strName.Insert(arr[iBar], '&');
}

This is a brutal solution. The hard-coded mnemonics indices are ugly.
It also entagles CMyDockablePane with CMainFrame, possibly solvable
by sending messages to main frame to retrieve the mnemonic indices.
But the more serious issue is that according to MS, GetPaneName is
used by floating bars; the MFC sources indicates otherwise (only
usage is in BuildPaneMenu) but do we want to risk it? Is there
a reliable way of detecting that we're in BuildPaneMenu?

11/19/21 phase bar's failure to support tempo change

First step is determining whether the song has tempo changes.
bool bHasTempoChange = pDoc->m_Seq.GetTracks().FindType(CTrack::TT_TEMPO)) >= 0;
This only considers the presence of tempo tracks, regardless of whether they're muted.

If tempo changes exist, the next step is retrieving the tempo map, via pDoc->ExportMidi.

11/23/21 phase bar tempo change support works

PLM file: phase diagram tempo map 3
Export MIDI file, length 1:40
Import into Reason, export audio
Export phase video, frame rate 30
Create MP4 as follows:
c:\apps\ffmpeg\bin\ffmpeg -framerate 30 -i "C:\Temp\Phase graph\data\img%%05d.png" 
-i "C:\Chris\MyProjects\Polymeter\docs\debug\phase diagram tempo map 3.wav" 
-c:v libx265 -movflags +faststart -vf format=yuv420p -preset veryslow -crf 23 
-b:a 320k -shortest "phase diagram tempo map 3.mp4"

The resulting video is correct. The final crash cymbal occurs at 1:32:834 (92.834 s)
according to Reason, and the phase diagram is fully aligned in frame 2785.
92.834 / 30 = 2785.02

11/24/21 startup benchmarks

Using Release version, after Reset Layout. Times are in milliseconds.
Caching reduces the times by about 15 ms; to avoid the caching variance,
rebuild the app before benchmarking. Total time (nearly half a second)
is measured by adding a CBenchmark member to CPolymeterApp class, and 
then printing the elapsed time at the end of CMainFrame::OnDelayedCreate.

Total: 450
CMainFrame::OnCreate: 220
CPolymeterApp::InitInstance: 280

removing EnableD2DSupport gets total down to 390; moving EnableD2DSupport
to OnShowChanged will save 60, or 13%.

InitInstance up to creating doc template: 3
creating main frame: 270
rest of InitInstance: 20

390 - 270 = 120 unaccounted for; most likely this is from painting windows

in CMainFrame::OnDelayedCreate:
ShowWindow and UpdateWindow: 50
MidiInit: 15
CheckForUpdates: 0
CMDIFrameWndEx::LoadFrame: 195

Most of the easily observable overhead is LoadFrame; not much to do about that.

11/27/21 updated benchmarks

Startup code prior to InitInstance: 0 (40 us)

Startup code was measured by copying AfxWinMain into the project,
and adding as its first line:
QueryPerformanceCounter((LARGE_INTEGER *)&tWinMain);
and then as first line of InitInstance, adding:
QueryPerformanceCounter((LARGE_INTEGER *)&t);
and then subtracting and dividing by performance counter frequency:
__int64	fPCFreq;
QueryPerformanceFrequency((LARGE_INTEGER *)&fPCFreq);
printf("%f\n", double(t - tWinMain) / fPCFreq);

From start of InitInstance to start of OnDelayedCreate: 300 .. 330
From start of InitInstance to end of InitInstance: 300 .. 330
Win32Console::Create: 90 <- this can cause confusion!
CWinAppEx::InitInstance: 0
InitInstance from after Win32Console::Create to end: 230
Creating main frame: 190
CMDIFrameWndEx::LoadFrame: 190

Looks like startup overhead is roughly proportional to the number of docking panes,
so consider carefully before adding more docking panes; they're expensive.

12/8/21 test next/prev convergence finding (pass!)

#include "RandList.h"
void CPhaseBar::TestConvergenceFinding()
{
	int	nPasses = 10000;
	srand(GetTickCount());	// random seed
	for (int iPass = 0; iPass < nPasses; iPass++) {
		int	nMaxMods = 40;
		int	nMods = CRandList::Rand(nMaxMods) + 1;
		CRandList	list(nMods);
		CLongLongArray	arrMod;
		arrMod.SetSize(nMods);
		for (int iMod = 0; iMod < nMods; iMod++) {
			arrMod[iMod] = list.GetNext() + 2;
		}
		LONGLONG	nStartPos = rand();
		if (rand() & 1)
			nStartPos = -nStartPos;
		LONGLONG	a, b;
		int	nConvSize = nMods / 2 + 1;
		a = FindNextConvergence(arrMod, nStartPos, nConvSize);
		b = FindNextConvergenceSlow(arrMod, nStartPos, nConvSize, false);
		if (a != b) {
			printf("FindNextConvergence FAIL\n");
			ASSERT(0);
		}
		a = FindPrevConvergence(arrMod, nStartPos, nConvSize);
		b = FindNextConvergenceSlow(arrMod, nStartPos, nConvSize, true);
		if (a != b) {
			printf("FindPrevConvergence FAIL\n");
			ASSERT(0);
		}
	}
}

12/30/21 archive project: lists of polymeter files used on albums

Akoko Ajeji 
C:\Chris\MyProjects\Polymeter\docs\Manmade
Ala Aye rec A B A B fine tune spread.plm
Asiri rec1e lead B dub spread lead string.plm
Bones rec fine tune cvt for PM 0.0.15 Mike.plm
Buy Me ending Mike double finger.plm
Buy Now double bass finger Mike.plm
Dek Sep Bluso jam breaks Mike closed and pedal hat +10.plm
Fazo Kanto B mix kick +15 Mike 3dB less ride.plm
Iyika rec A B take1 scale up volumes spread.plm
Stencil lower bass spread strings mike.plm
Vizyon rec02 edit etc scale up volumes Mike.plm

Polymeter
C:\Chris\MyProjects\Polymeter\docs\test
Algo Rag Reason.plm
Atunwi rec1b ajust velo for Reason Rad Piano.plm
dark dim (Oju Inu) B rest.plm
Ilopo Ferese rec1 edit2.plm
Itumo shorter.plm
ominira head.plm
ominira rec1.plm
Omioto.plm
Sipeli rec1.plm
tristate (Ona Lile) rec1 edit.plm

Apologize to the Future:
Apologize full verse2 cycle skip static.plm
changing climate rec4c no open hat.plm
exit game panning.plm
oily rock shorter chorus.plm
Overshoot 1e mix.plm
singularity 3 AAB.plm

Passion For Numbers:
C:\Chris\MyProjects\Polymeter\docs\test
Ero Ayo ABC2.plm
fun ife 03 still more.plm
Irubo cresc.plm
Mo Bulu 2 rec1.plm
Ninu Meta (334 343 433) rec1e B2 run.plm
Roro shorter ritard.plm
two triads combine 2222 rec2b.plm
Yipo 04 merge B.plm

More Than Four:
Broke strings.plm
Charlie's Big Break mix2c.plm
Heard a Moon more space.plm
Interago pos mod.plm
Kahelo fast intro.plm
kasita mondo rec take2.plm
LCM drums.plm
Lodidi 2b mix.plm
Moonchego 06 area 51.plm
more than four 02b rec 1b.plm
Pleasant Mistake rec5 chorus less.plm
Shelter in Bass 005d vamp rec1 edit.plm
Ticking 2021b rec1b2c.plm
virtue signal rec C break edit.plm

1/1/22 Proposal to suppress second note off for zero-velocity notes

For note tracks, if the velocity is zero, skip queuing a note off. The reasoning:
a note off was already queued at the start of the note, so queuing a second note 
off at the end of the note is superfluous and unexpected, and also potentially 
interferes with using note off tracks to implement exclusion groups (choking).

All album tracks were checked for impact, using an automated script that exported
MIDI files with and without the fix and compared them. A few files differed:
Asiri
Stencil
Overshoot

The above cases were investigated. In every case, the MIDI file with the fix 
differs from the original in the technical sense of flunking a binary compare,
but no audible change occurs, because for all zero-velocity notes, the note 
being released a second time is already off. To further verify this, both the
fixed and original MIDI files were imported into Reason and then re-exported 
as MIDI, and the Reason-exported MIDI files were found to be identical.

1/21/22 tempo modulation makes UI slip out of sync with playback

This bug has been around for months, and it's been blamed on the Windows
MIDI API, but the fact that manually changing the tempo, even continuously,
doesn't also cause the bug is extremely suspicious. How do the two cases
differ? It suggests a logic error.

DWORD	dwTempo;
MMTIME	time;
if (pPlayingDoc->m_Seq.GetPosition(time, TIME_MS) 
&& pPlayingDoc->m_Seq.GetCurrentTempo(dwTempo))
	printf("%d %d %d %g\n", nPos, time.u.ms, 
	pPlayingDoc->m_Seq.ConvertMillisecondsToPosition(time.u.ms), 
	CMidiFile::MICROS_PER_MINUTE / double(dwTempo));

Hypothesis: the MIDI API doesn't compensate the position for tempo changes. 
Querying the position in milliseconds and then converting it to ticks causes 
even worse splippage, but that's because you can't convert milliseconds to 
ticks without compensating for tempo changes. FALSE!

Experiment: Try building a tempo map, and then use it to correct the tick 
position returned by the MIDI API. If that stays in sync, we can conclude 
that the MIDI API doesn't compensate position for tempo changes. If 
that doesn't stay in sync, then the question remains open. MOOT!

But even if compensating does fix the sync, a practical solution would
take more work. The sequencer would have to model the actual tempo, and
keep the model up to date. This is probably possible, but challenging.

One big problem with the above hypothesis is it doesn't explain why the
UI doesn't appear to lose sync when tempo is changed manually during 
playback, via the tempo edit box up/down buttons, or even rapidly and
continuously via tempo MIDI mapping. This is really puzzling.

What are the differences between the two code paths? SetTempo adds only
a single tempo change message at the start of the output event buffer.
But emulating this behavior in the tempo modulation case doesn't help.

SetTempo also changes the callback length, and this is suspicious, because
callback length is used in GetPosition. But removing callback length from
the GetPosition calculation also doesn't help. And, if the bug was caused
by callback length, you'd think it would drift out of sync as tempo
speeded up, but then catch up again as it returned to baseline.

It looks like some kind of order dependency. Something is different
about adding a tempo event at the start of OutputMidiBuffer, and
adding one after AddEvents, and all that other stuff. A lot happens
between those two points, and any of it could be causing the bug.

Experiment: Try not setting callback length in SetTempo, and see
if the bug appears for mapped tempo change. NO!

Experiment: Try moving the signaled tempo change code to the end of
OutputMidiBuffer, and see if the bug appears. YES!

Moving the set tempo check to the same place where we output events
to the MIDI buffer causes the bug to appear. Continuously moving the
tempo knob makes the UI slip, and the splippage worsens over time.

Is it the track critical section? No, entering the critical section
before checking for set tempo doesn't help.

Found it! Tempo events weren't being added to the event count, which
was being maintained separately from the event array size, in nEvents.
Adding tempo events caused that same number of events to be skipped
at the end of the event buffer, including the final NOP event that 
pads the MIDI buffer out to the callback length. The absence of the
padding NOP caused the position slippage. It looked like the UI was
slipping behind, but it was actually the output sequence that was 
slipping ahead, due to missing chunks of time. Moral of the story:
don't do double bookkeeping. FIXED!

In short, tempo modulation was causing incorrect playback. 
The export was not affected.

1/22/22 showing current tempo in status bar

// in ctor
m_dwCachedTempo = 0;

// in OnTempo
DWORD	dwTempo;
if (pPlayingDoc->m_Seq.GetCurrentTempo(dwTempo)) {	// if valid current tempo
	if (dwTempo != m_dwCachedTempo) {	// if tempo actually changed
		m_dwCachedTempo = dwTempo;
		m_sTempo.Format(_T("%.3f"), CMidiFile::MICROS_PER_MINUTE / double(dwTempo));
		m_wndStatusBar.SetPaneText(SBP_TEMPO, m_sTempo);
	}
}

This code adds at least 250 microseconds in the case where the status
pane actually changes. Because OnTimer is in the critical path, 250 
microseconds is long enough to be scary. Assuming the default refresh
rate of 20 Hz, if the tempo is changing on every OnTimer call, it's 
potentially adding 5ms per second, or half a percent. 

Almost all of that 250 microseconds comes from SetPaneText. SetPaneText 
frees and reallocates its lpszText buffer every time the text changes, 
regardless of whether the existing buffer is big enough for the new 
string. That's dumb, but fixing it doesn't help much, because nearly
all the cost of SetPaneText is InvalidatePaneContent, and the cost
of that method is all InvalidateRect and UpdateWindow.

UpdateWindow is by far the biggest offender, hence performance can be 
significantly improved by calling UpdateWindow only once for all three
dynamic status panes, instead of calling it three times, from within
SetPaneText. This requires a customized implementation of SetPaneText.

By doing the UpdateWindow only once per OnTimer call instead of three 
times, we can save enough time to offset the increased overhead from
showing current tempo. Even better, that overhead is only incurred 
when the tempo actually changes, otherwise there's no cost.

2/16/22 modulation errors

The following snippet checks for errors and outputs them to the console.

CTrack::CModulationErrorArray	arrError;
if (!doc.m_Seq.GetTracks().ValidateModulations(arrError)) {
	for (int iError = 0; iError < arrError.GetSize(); iError++) {
		const CTrack::CModulationError& err = arrError[iError];
		_tprintf(_T("%d\t%d\t%s\t%d\n"), err.m_iTarget + 1, err.m_iSource + 1, 
			CTrack::GetModulationTypeInternalName(err.m_iType), err.m_nError);
	}
}

2/19/22 Unicode in track names

After much research and experimentation, the conclusion: Unicode in track 
names is a bad idea, primarily because the MIDI file format only supports 
8-bit characters in track names, and 7-bit characters are preferable for 
maximum compatibility.

"[Text in MIDI files should be] printable ASCII characters for maximum 
interchange. However, other character codes using the high-order bit may 
be used for interchange of files between different programs on the same 
computer which supports an extended character set."
(Schulich School of Music, McGill University, StandardMIDIfileformat.html)

Other factors include limitations of the INI and CSV file formats and the
APIs employed to read and them.

It's possible to open a text file in Unicode and then attach it to the
CStdioFile instance using this code:

BOOL CIniFile::CFastStdioFile::Open(LPCTSTR pszFilePath, UINT nOpenFlags, CFileException *pException)
{
	static const LPCTSTR szMode[] = {
		_T("rt,ccs=UNICODE"),	// can be UTF-8 instead
		_T("wt,ccs=UNICODE"),
	};
	bool	bIsWrite = (nOpenFlags & modeWrite) != 0;
	FILE	*fStream;
	errno_t	nErr =_tfopen_s(&fStream, pszFilePath, szMode[bIsWrite]);
	if (nErr != 0)
		return false;
	CommonBaseInit(fStream, NULL);
	m_bCloseOnDelete = true;
	return true;
}

This seems to work for reading and writing, but it uses UTF-16 LE, which
makes the files twice as large. It does seem to support the entire Unicode
range, judging from tests with Chinese lorem ipsum characters.

It appears that the file size issue could be solved by using ccs=UTF-8 
instead, but it would still break backwards compatibility, because older
versions of the application would read the UTF-8 BOM as part of the first
string, causing the file format check to fail. One workaround would be to
only create UTF-8 if it's necessary, which means Write would have to check
the track names (and song info!) for non-ASCII characters, and if none are
found, use ANSI instead of UTF-8.

Adding the following test code to OnNewDocument results in a PLM file with 
its first 14 track names set to the entire 8-bit character range:

	for (int i = 0; i < 14; i++) {
		TCHAR	szName[17];
		for (int j = 0; j < 16; j++) {
			szName[j] = static_cast<TCHAR>((i + 2) * 16 + j);
		}
		szName[16] = 0;
		m_Seq.SetName(i, szName);
	}


2/21/22 CLOC (Count Lines of Code)

-----------------------------------------------------------------------------------
Language                         files          blank        comment           code
-----------------------------------------------------------------------------------
C++                                 97           3273           2727          32418
C/C++ Header                       141           2494           3279          12929
Windows Resource File                1            135             61           1753
-----------------------------------------------------------------------------------
SUM:                               239           5902           6067          47100
-----------------------------------------------------------------------------------

PolymeterCL is only 9447 LOC (5367 code + 4080 header)

2/27/22

D-7 D-6 E-7b5 A7b9
D-7 D-6 F#-7b5 B7b9
E-7 E-6 A-7b5 D7b9
D-7b5 G7b9 C-7 D7b9

2. D-7b5 G7b9 G-7b C7b9

F-7 F-7 C-7b5 F7b9
Bb-7 Bb-6 F-7b5 Bb7b9
Eb-7 Eb-6 C#-7b5 F#7b9
F#-7b5 B#7b9 B-7b5 E7b9

4/11/22 MIDI Visualizer

Ibaramu:
relative to original file, snippet starts at: 0:01:27.545684
-35:00 is the MIDI offset; -240:0 is start of file; difference of 205:00 beats
87.85714285714286 seconds at 140 BPM; start at 5271 frames 
first three notes better be: F C Bb
add 60 frames (1 second) of pre-roll (probably a setting in MIDI visualizer)
5271 + 60 + 1 = 5332 frames is exactly right

6/15/22 Ring in the Odd shorter version

current duration is 12 cycles
cycle length is 61:30 (245 1/16th note steps)
total length is 2940 steps (735 beats)
starts slowing after 10 cycles at 613:60

cut duration in half to 6 cycles:
remove from 246:0 to 613:60
minor bob and range mod need adjusting
offset both -44100 ticks (61.25 * 6 * 120)
total length is 1470 steps (367.5 beats)

Seems too short now!

cut duration by a third to 8 cycles:
remove from 368:60 to 613:60
minor bob and range mod need adjusting
offset both -29400 ticks (61.25 * 4 * 120)
total length is 1960 steps (490 beats)

differences between intro and main section:
	intro	main
Fader	0	+3.01
VelRH	33%	0%
VelRL	33%	100%
Ambie	34%	25%
SResL	+0.5	0

6/18/22 show selection in status bar; works fine but questionably useful

	m_szCachedSelection = CSize(0, 0);	// in ctor

	m_wndStatusBar.SetPaneText(SBP_SELECTION, _T(""), FALSE);	// don't update; avoids flicker

	ON_UPDATE_COMMAND_UI(ID_INDICATOR_SELECTION, OnUpdateIndicatorSelection)

void CMainFrame::OnUpdateIndicatorSelection(CCmdUI* pCmdUI)
{
	CSize	szSel;
	if (m_pActiveChildFrame != NULL) {	// if active child frame exists
		// if focus window is child of song parent
		if (m_pActiveChildFrame->IsSongView()) {	// if showing song view
			CRect	rSel;
			m_pActiveChildFrame->m_pSongParent->m_pSongView->GetSelection(rSel); // should return const reference instead!
			szSel = rSel.Size();
		} else {	// not showing song view
			szSel = m_pActiveChildFrame->m_pStepParent->m_pStepView->GetStepSelection().Size();
		}
	} else {	// no active document
		szSel = CSize(0, 0);
	}
	if (szSel != m_szCachedSelection) {	// if selection size changed
		m_szCachedSelection = szSel;
		if (szSel.cx) {	// if selection isn't empty
			CString	s;
			if (szSel.cx == INT_MAX)	// song view reserves this value for select all
				s = _T("*");
			else
				s.Format(_T("%d x %d"), szSel.cx, szSel.cy);
			pCmdUI->SetText(s);
		} else {
			pCmdUI->SetText(_T(""));
		}
	}
	pCmdUI->Enable(true);
}

7/17/22 Center song position

CPolymeterDocument::SetPosition currently calls UpdateAllViews a second time to implement
centering, which is inefficient. Centering is a parameter, and parameters should normally
be implemented via the pHint argument. If the bCenterSongPos argument is true, pass this 
to UpdateAllViews, otherwise pass NULL. In CSongView::OnUpdate, after updating the song 
position, if pHint is non-NULL, call CenterCurrentPosition.

HINT_SONG_POS // song position change; pHint is non-NULL if centering song position

7/17/22 Exclude muted tracks from graph

Filtering is the easy part:
// if excluding muted tracks and modulation's source or target is muted
if (!m_bShowMuted && (pDoc->m_Seq.GetMute(mod.m_iSource) || pDoc->m_Seq.GetMute(mod.m_iTarget)))
	continue;	// skip this modulation

Cases must be added to OnUpdate. Solo, Track Prop / Mute, and Multi Track Prop / Mute
all must start a deferred update if m_bShowMuted is false. Other cases? Ideally these
cases should also rebuild the mute cache to avoid needless graph updates (see below).
Updating the graph automatically during song playback requires adding a song position
case to OnUpdate, as follows. This code takes about 1 microsecond in a typical case.

case CPolymeterDoc::HINT_SONG_POS:
	if (!m_bShowMuted) {	// if excluding muted tracks
		CPolymeterDoc	*pDoc = theApp.GetMainFrame()->GetActiveMDIDoc();
		if (pDoc != NULL) {
			int	nTracks = pDoc->GetTrackCount();
			if (nTracks != m_arrMute.GetSize()) {	// if track count differs from cache size
				RebuildMuteCache();	// cache is stale, so rebuild it
			} else {	// track count matches; assume cache is valid
				bool	bMuteChanged = false;
				for (int iTrack = 0; iTrack < nTracks; iTrack++) {	// for each track
					bool	bIsMuted = pDoc->m_Seq.GetMute(iTrack);
					if (bIsMuted != m_arrMute[iTrack]) {	// if track mute differs from cache
						m_arrMute[iTrack] = bIsMuted;	// update cache
						bMuteChanged = true;	// set change flag
					}
				}
				if (bMuteChanged)	// if one or more track mutes changed state
					StartDeferredUpdate();
			}
		}
	}
	break;
}

void CGraphBar::RebuildMuteCache()
{
	CPolymeterDoc	*pDoc = theApp.GetMainFrame()->GetActiveMDIDoc();
	if (pDoc != NULL) {
		int	nTracks = pDoc->GetTrackCount();
		m_arrMute.FastSetSize(nTracks);
		for (int iTrack = 0; iTrack < nTracks; iTrack++) {	// for each track
			m_arrMute[iTrack] = pDoc->m_Seq.GetMute(iTrack);	// cache mute state 
		}
	} else {
		m_arrMute.RemoveAll();
	}
}

7/19/22 Adding docking bar context menu command IDs to Customize dialog

	struct DOCK_BAR_MENU {
		int		nMenuID; 	// context menu's ID
		LPCTSTR	pszMenuName;		// context menu's name
		CMyDockablePane	CMainFrame::*pPane;	// pointer to dockable pane member
	};

const CMainFrame::DOCK_BAR_MENU CMainFrame::m_arrDockBarMenu[] = {
	{IDR_GRAPH_CTX,			_T("Graph"),		reinterpret_cast<CMyDockablePane CMainFrame::*>(&CMainFrame::m_wndGraphBar)},
	{IDR_MODULATION_CTX,	_T("Modulation"),	reinterpret_cast<CMyDockablePane CMainFrame::*>(&CMainFrame::m_wndModulationsBar)},
	{IDR_MAPPING_CTX,		_T("Mapping"),		reinterpret_cast<CMyDockablePane CMainFrame::*>(&CMainFrame::m_wndMappingBar)},
	{IDR_PIANO_CTX,			_T("Piano"),		reinterpret_cast<CMyDockablePane CMainFrame::*>(&CMainFrame::m_wndPianoBar)},
	{IDR_PARTS_CTX,			_T("Parts"),		reinterpret_cast<CMyDockablePane CMainFrame::*>(&CMainFrame::m_wndPartsBar)},
	{IDR_PRESETS_CTX,		_T("Presets"),		reinterpret_cast<CMyDockablePane CMainFrame::*>(&CMainFrame::m_wndPresetsBar)},
	{IDR_PHASE_CTX,			_T("Phase"),		reinterpret_cast<CMyDockablePane CMainFrame::*>(&CMainFrame::m_wndPhaseBar)},
	{IDR_STEP_VALUES_CTX,	_T("StepValues"),	reinterpret_cast<CMyDockablePane CMainFrame::*>(&CMainFrame::m_wndStepValuesBar)},
};

void CMainFrame::AddDockBarMenuCommands(CMFCToolBarsCustomizeDialog *pDlgCust, int nMenuID, LPCTSTR pszTitle)
{
	CMenu menu;
	menu.LoadMenu(nMenuID);
	CMenu	*pSubMenu = menu.GetSubMenu(0);
	int	nItems = pSubMenu->GetMenuItemCount();
	for (int iItem = nItems - 1; iItem >= 0; iItem--) {
		int	nID = pSubMenu->GetMenuItemID(iItem);
		if (nID >= ID_FILE_NEW || nID == ID_EDIT_INSERT || nID == ID_EDIT_DELETE || nID == ID_EDIT_RENAME) {
			pSubMenu->DeleteMenu(iItem, MF_BYPOSITION);
			nItems--;
		}
	}
	pDlgCust->AddMenuCommands(&menu, FALSE, pszTitle);
}

in CMainFrame::OnViewCustomize:

	for (int iBarMenu = 0; iBarMenu < _countof(m_arrDockBarMenu); iBarMenu++) {
		AddDockBarMenuCommands(pDlgCust, m_arrDockBarMenu[iBarMenu].nMenuID, m_arrDockBarMenu[iBarMenu].pszMenuName);
	}

This could also be done in CPolymeterApp::PreLoadState. The advantage is that the
context menus would show up in the Context page of the Customize dialog as expected,
but the disadvantage is more work at startup time instead of doing it on Customize.
Also menu items added via GetContextMenuManager()->AddMenu can't be assigned keyboard 
accelerators, which kind of defeats the purpose of this whole exercise.

The above code works fine. The remaining problem is adding the docking windows to 
the command routing. That presumably must be done by overriding CMainFrame::OnCmdMsg.
Iterating the frame's child windows wouldn't work reliably due to docking complexities.
Iterating the docking manager's pane list would work, but it's inefficient. Note that
CN_UPDATE_COMMAND_UI must also be handled so that commands are checked or disabled.

BOOL CMainFrame::OnCmdMsg(UINT nID, int nCode, void* pExtra, AFX_CMDHANDLERINFO* pHandlerInfo)
{
	if (CMDIFrameWndEx::OnCmdMsg(nID, nCode, pExtra, pHandlerInfo))
		return true;
	if ((nCode == CN_COMMAND || nCode == CN_UPDATE_COMMAND_UI) && nID < ID_FILE_NEW) {	// if custom command
		for (int iPane = 0; iPane < _countof(m_arrDockBarMenu); iPane++) {	// for each pane
			CMyDockablePane	*pPane = &(this->*m_arrDockBarMenu[iPane].pPane);	// pointer to member syntax
			// if pane is visible, give it a try
			if (pPane->FastIsVisible() && pPane->OnCmdMsg(nID, nCode, pExtra, pHandlerInfo))
				return true;
		}
	}
	return false;
}

7/20/22 OnCmdMsg benchmarks

Average pass uses one or two microseconds. Occasionally spikes as high as 200 micros.
Usage as percentage of available CPU time is normally a tiny fraction of a percent,
and only spikes during continuous mouse movement, in which case it reaches 5% - 10%.

7/26/22 IWebBrowser2 event sink

It's easier than it looks. There's no need to implement DWebBrowserEvents2. But it doesn't 
solve the problem, because receiving NavigateComplete2 apparently doesn't guarantee that 
the browser's window has been painted. This implementation is copied from viewhtml.cpp.

In .h:
	DECLARE_EVENTSINK_MAP()
	virtual void NavigateComplete2(LPDISPATCH /* pDisp */, VARIANT* URL);

In .cpp:
#include <exdispid.h> // needed for browser event IDs

BEGIN_EVENTSINK_MAP(CGraphBar, CMyDockablePane)
	ON_EVENT(CGraphBar, AFX_IDW_PANE_FIRST, DISPID_NAVIGATECOMPLETE2, NavigateComplete2, VTS_DISPATCH VTS_PVARIANT)
END_EVENTSINK_MAP()

void CGraphBar::NavigateComplete2(LPDISPATCH pDisp, VARIANT* URL)
{
	UNREFERENCED_PARAMETER(pDisp);
	ASSERT(V_VT(URL) == VT_BSTR);
	CString str(V_BSTR(URL));
	_tprintf(_T("%s"), str); // print URL to console
}

9/5/22 modulation bar's label tooltip shows stale data if track selection changes

The problem is general. When updating a virtual list control, calling SetItemCountEx
isn't enough; you must also call DeleteAllItems first, otherwise the label tooltip is 
potentially stale. DeleteAllItems adds a few extra milliseconds, needlessly except
for the tooltip issue. This is apparently a known bug, for example see:
http://computer-programming-forum.com/82-mfc/7e5e14a6aaf75079.htm

"My virtual list control, with style LVS_EX_INFOTIP in large icon mode has a
problem. It seems that the control caches the tooltip text for an item and
doesn't requery that text until the cursor floats over a new item.
That is, if the cursor hovers over item 0 and the control gets the tooltip
text from me, then I delete the item from the list and add several new items
and hover over item 0 again, I get the old tooltip text. If I then hover
over item 1 and back to 0 I get the correct text because the control resends
the LVN_GETINFOTIP for item 0."

12/5/22 video that shows only phase diagram and piano

This is a customization of the "Polymeter vid cap" sub-project. In CMainFrame::ExportVideo:

Force exporting MIDI flag true, because we will show piano bar:
	bool	bIsExportingMidi = true;//@@@m_wndPianoBar.IsVisible() || m_wndMidiOutputBar.IsVisible();

And after full screen initiated and screen recorder created:
	//@@@
	CSize	szScr(1920, 1080);
	int	nPianoHeight = szScr.cy / 10;
	CWnd	*pPianoParent = m_wndPianoBar.GetParent();
	CWnd	*pPhaseParent = m_wndPhaseBar.GetParent();
	m_wndPianoBar.ShowPane(true, 0, 0);
	m_wndPhaseBar.ShowPane(true, 0, 0);
	m_wndPhaseBar.FloatPane(CRect(0, 0, 100, 200));
	m_wndPianoBar.FloatPane(CRect(0, 0, 100, 100));
	m_wndPhaseBar.SetParent(GetDesktopWindow());
	m_wndPhaseBar.ModifyStyle(WS_CHILD, WS_POPUP);
	m_wndPhaseBar.SetWindowPos(&CWnd::wndTopMost, 0, 0, szScr.cx, szScr.cy - nPianoHeight, SWP_FRAMECHANGED | SWP_NOCOPYBITS | SWP_SHOWWINDOW);
	m_wndPianoBar.SetParent(GetDesktopWindow());
	m_wndPianoBar.ModifyStyle(WS_CHILD, WS_POPUP);
	m_wndPianoBar.SetWindowPos(&CWnd::wndTopMost, 0, szScr.cy - nPianoHeight, szScr.cx, nPianoHeight, SWP_FRAMECHANGED | SWP_NOCOPYBITS | SWP_SHOWWINDOW);

And at end, try to clean up:
	//@@@
	m_wndPianoBar.ModifyStyle(WS_POPUP, WS_CHILD);
	m_wndPianoBar.SetParent(pPianoParent);
	m_wndPhaseBar.ModifyStyle(WS_POPUP, WS_CHILD);
	m_wndPhaseBar.SetParent(pPhaseParent);

Also in PhaseBar.cpp, disable upper limit on planet size:
 	m_nMaxPlanetWidth = INT_MAX;//@@@MAX_PLANET_WIDTH;

None of this is even close to foolproof but it gets the job done.

12/13/22 quant as fraction

Something like this partially works:

NumEdit.h:
		DF_FRACTION	= 0x04,

CNumEdit::StrToVal:
	int	nNumerator, nDenominator;
	if ((m_nFormat & DF_FRACTION)	// if fractions are enabled
	&& _stscanf_s(Str, _T("%d/%d"), &nNumerator, &nDenominator) == 2	// and both values scanned
	&& nDenominator != 0)	// and denominator is non-zero
		r = nNumerator / double(nDenominator) * m_fScale;
	else
		r = _tstof(Str) * m_fScale;

CTrackView::CTrackGridCtrl::CreateEditCtrl:
		case COL_Quant:
			nDataFormat |= CNumEdit::DF_FRACTION;	// let user enter a fraction
			pEdit->SetScale(pDoc->GetTimeDivisionTicks() * 4);
			break;

But the SetScale breaks the spin control delta. This could be fixed by adding a 
special integer multiplier to CNumEdit that's only used in the fraction case.

1/24/23 Showing pane menu in Live view

void CMainFrame::ShowPanesMenu(CPoint point)
{
	CMenu	menu;
	VERIFY(menu.CreatePopupMenu());
	m_dockManager.BuildPanesMenu(menu, false);
	m_dockManager.OnPaneContextMenu(point);
}

And in CLiveView::OnContextMenu:

	theApp.GetMainFrame()->ShowPanesMenu(point);

2/18/23 live screen recording

The issue with the existing screen recorder code is that the BitBlt takes on the order of 25ms.
That's too slow to keep up with a 30 FPS frame rate, because it doesn't leave enough time for
other work (such as updating the UI, and saving the frame). The reason BitBlt is so slow is 
that we're using the desktop window's device context, which means the OS has to do compositing 
to account for windows other than the application's. If we use the main frame's device context
instead, BitBlt only takes about 4ms, roughly an order of magnitude faster. Hopefully that's
fast enough to keep up. Writing PNG files is much too slow, so it will be necessary to compare
successive frames and only store the differences via run length encoding (presumably in a
worker thread) and then reconstruct the frames and save them as PNGs after recording stops.

Here are the necessary changes:
1. In CScreenRecorder, use the main window instead of the desktop window
//	m_hWnd = GetDesktopWindow();
	m_hWnd = AfxGetMainWnd()->m_hWnd;

//	m_hdcWnd = GetWindowDC(NULL);  no good now that we're using main frame instead!
	m_hdcWnd = GetWindowDC(m_hWnd);

2. In CMainFrame::MakeFullScreen, comment out the InflateRect
//	rScreen.InflateRect(szAdjust.cx, 1, szAdjust.cx, szAdjust.cy);

3. In CMainFrame::ExportVideo, before we call MakeFullScreen, remove the caption bar and frame
	ModifyStyle(WS_CAPTION | WS_THICKFRAME, 0);

The caption bar will be omitted, but that's the price you pay for using the main frame's DC.
On the good side, the main frame will be slightly taller so you'll have more space for UI.
Don't forget to restore the caption bar and frame after recording stops!

3/10/23 Graph text is too small when system text size is greater than 100% (96 dpi)

This has nothing to do with Graphviz. It's caused by the web browser control (IWebBrowser)
which displays the SVG file that Graphviz outputs. The core issue is that Internet Explorer
defaults to being non-DPI aware. This problem and various solutions are discussed here: 
https://stackoverflow.com/questions/38754354/wpf-web-browser-control-and-dpi-scaling
The easiest solution is to create this registry key:
HKEY_CURRENT_USER\SOFTWARE\Microsoft\Internet Explorer\Main\FeatureControl\FEATURE_96DPI_PIXEL
and in that key, create a REG_DWORD value named Polymeter.exe with a value of 1.
If done in HKEY_CURRENT_USER, this fix works for both 64-bit and 32-bit flavors of the app.
To affect all users, the fix must be done in HKEY_LOCAL_MACHINE, but you need admin to do it,
AND you need to do it in two separate places in order to affect both 64-bit and 32-bit apps:
HKEY_LOCAL_MACHINE\SOFTWARE\Microsoft\Internet Explorer\Main\FeatureControl\FEATURE_96DPI_PIXEL
and
HKEY_LOCAL_MACHINE\SOFTWARE\Wow6432Node\Microsoft\Internet Explorer\Main\FeatureControl\FEATURE_96DPI_PIXEL
Note that FEATURE_96DPI_PIXEL is obsolete according to MSDN, so YMMV.
Also note that all of the above assumes you're using 64-bit Windows.

3/10/23 more phase videos for Yoyaku (60 FPS)

Have a Good One
audio starts at 0:00:00.612675
start sequencer at -1:080
convergence size 5

Awesome on Mars
0:00:00.232523
convergence size 7

Not My Problem, I'll Be Dead
0:00:00.204892, PLM starts -0:00:00.462, start at frame 16
convergence size 6

Baby Batter Bingo
0:00:00.215512, start at frame 15 
convergence size 7

NOTE: these three made with UNMASTERED AUDIO so redo them later; start times will change

K35
audio starts at 0:00:00.214669 seconds
first note at -28:60, delay is 0:0.231 seconds, start at frame 1
convergence size 8

Saz
0:00:00.135479, change PLM start to 0:000, start at frame 20
convergence size 8

Primitive Man
0:00:00.222454, start at frame 0
convergence size 5

LCM
0:00:00.107405, start at frame 23
convergence size 8

More Than Four
0:00:00.09801, about five frames; just delay video
convergence size 6

Virtue Signal
0:00:00.097684, about five frames; just delay video
convergence size 6

8/18/23 tetrachord inversions via voicing modulation

close
[1, 0, 0]
 0, 0, 0
-4, 0, 0
-3,-4, 0
-2,-3,-4

drop 2
[3, 1, 0]
 2, 0, 0
 1,-4, 0
 0,-3, 0
-4,-2, 0

drop 3
[4, 1, 0]
 3, 0, 0
 2,-4, 0
 1,-3,-4
 0,-2,-3

9/19/23 optimizing CTrackArray::GetModulationTargets

Preallocating modulation arrays speeds it up, but not as much as hoped.
For documents with many tracks, iterating all of the modulations twice may
cost more than repeatedly reallocating and copying the target array data.
The typical speedup is from times two to times five. Even for complicated 
documents (e.g. Moonchego) the unoptimized version typically takes around
20 microseconds, so optimization is unwarranted.

9/25/23 compiler warnings in Visual Studio 2019

PhaseBar.cpp(904,23): warning C4456: declaration of 'fOrbitWidth' hides previous local declaration
PianoBar.cpp(414,4): warning C4457: declaration of 'iChannel' hides function parameter
PhaseBar.cpp(909): warning C4701: potentially uninitialized local variable 'nSelectedOrbits' used
PhaseBar.cpp(913): warning C4701: potentially uninitialized local variable 'nConvergenceSize' used
Track.cpp(1300,61): warning C4456: declaration of 'mod' hides previous local declaration

12/24/23 binary search that returns the exactly matching item or the nearest item above the target value

W64INT BinarySearchAboveEqual(const CArrayEx_TYPE& val) const 
{
	W64INT iStart = 0;
	W64INT iEnd = m_nSize - 1;
	W64INT iResult = -1;
	while (iStart <= iEnd) {
		W64INT iMid = (iStart + iEnd) / 2;
		if (val == GetAt(iMid))
			return iMid;
		else if (GetAt(iMid) < val)
			iStart = iMid + 1;
		else {
			iResult = iMid;
			iEnd = iMid - 1;
		}
	}
	return iResult;
}

1/9/24 time signature with variable denominator (i.e. other than four)

We assume values are restricted to small positive integer powers of two. 
The goal is to allow time signatures such as 3/8 and 5/16.

The most important changes are in CSequencer::ConvertPositionToBeat and 
CSequencer::ConvertBeatToPosition. In both methods, replace all instances
of m_nTimeDiv with m_nBeatDiv, which is m_nTimeDiv scaled as necessary.

The meter denominator can be stored as a binary exponent of two, as in:
0 == whole, 1 == half, 2 == quarter, 3 == eighth, 4 == sixteenth, etc.
This inherently restricts the denominator to powers of two. Note that
since the denominator shift won't exceed 5 it can be stored in a BYTE.
Given the above scheme, computing beat divisions is straightforward:
m_nBeatDiv = m_nTimeDiv * 4 >> m_nMeterDenominator;

The preceding handles all tick / MBT conversions, excepting the step and
song grid rulers. The rulers are a harder problem. The following edits
are required for both step and song views. Instead of this:

double	fBeatWidth = m_pS???View->GetBeatWidth() * m_pS???View->GetZoom();

substitute this (??? indicates specialization for Step or Song):

int	nBeatWidth = (m_pS???View->GetBeatWidth() * 4) >> GetDocument()->GetMeterDenominator();
double	fBeatWidth = nBeatWidth * m_pS???View->GetZoom();

and in UpdateRulerNumbers, instead of:

m_wndRuler.SetMidiParams(pDoc->GetTimeDivisionTicks(), pDoc->m_nMeter);

substitute this:

m_wndRuler.SetMidiParams(pDoc->m_Seq.GetBeatDivision(), pDoc->m_nMeter);

This is the minimum for correct behavior. The remaining issue is that
the ruler doesn't necessarily align major ticks with the denominator. 
So for example in 3/8 time, the major ticks will be:

1:1  1:3  2:2  3:1  3:3  4:2  5:1  etc.

This is not wrong, but it's certainly not optimal. OTH the rulers already
display unexpected major tick alignment even without variable denominator.
Variable denominator just makes the preexisting issue more obvious.

Also, the MIDI export must use the meter denominator in its time signature.
Replace this:

CMidiFile::TIME_SIGNATURE	sigTime = {static_cast<BYTE>(m_nMeter), 2, 0, 8};	// denominator is quarter note

with this:

CMidiFile::TIME_SIGNATURE	sigTime = {static_cast<BYTE>(m_nMeter), static_cast<BYTE>(m_nMeterDenominator), 0, 8};

1/19/24 on launch, docked panes that split a side have incorrect sizes if app was previously maximized

This bug can be replicated in a stock VS 2012 or VS 2019 MFC app, taking the defaults for all template settings.
1. Run the app.
2. Maximize the app (if it isn't already so).
3. Dock the classes pane below the files pane (if it isn't already so).
4. Change the split point between the files and class panes, via left-click and drag, so that the last item 
shown in the files pane is FakeApp.rc.
5. Exit and re-run the app. Notice that the split point between the files and classes panes has changed.

If you repeat the above test but without maximizing the app, the split point between the panes is correctly restored.

Another way to replicate the error:
1. Run the app.
2. Maximize the app (if it isn't already so).
3. Dock the classes pane below the files pane (if it isn't already so).
4. Change the split point between the files and classes panes, via left-click and drag, so that the last item 
shown in the files pane is FakeApp.rc.
5. Un-maximize the app (restore the app to its default size).
6. Re-maximize the app. Notice again that the split point between the files and classes panes has changed,
and by the same amount as in the first example.

1/23/24 test new more efficient CModulationTargets against original GetModulationTargets

	CModulationArrayArray	arrTarget;
	pDoc->m_Seq.GetTracks().GetModulationTargets(arrTarget);
	for (int iTarg = 0; iTarg < arrTarget.GetSize(); iTarg++) {
		const CModulationArray&	arrMod = arrTarget[iTarg];
		int	nMods = arrMod.GetSize();
		ASSERT(nMods == parrTarget->GetCount(iTarg));
		const CModulationArray	*parrTrackTarget = parrTarget->GetTargets(iTarg);
		for (int iMod = 0; iMod < nMods; iMod++) {
			ASSERT(arrMod[iMod] == parrTarget->GetAt(iTarg, iMod));
			ASSERT(arrMod[iMod] == (*parrTrackTarget)[iMod]);
		}
	}

2/15/24 investigation into possibly excessive update command UI activity

OnIdle calls occur in pairs, as determined by overriding CWinApp::OnIdle.
Only the first call, with the lCount argument == 0, does meaningful work.
According to MSDN and code review, this call "updates command user-interface 
objects such as menu items and toolbar buttons, and it performs internal 
data structure cleanup" (which means releasing temporary objects). This 
first call takes approximately a quarter of a millisecond in Release, 
so there is not much margin for optimization.

The next question is why the call occurs periodically during playback, with
a period of one second. This period doesn't correspond to any application
timer. The only active application timer is the view timer, which is much 
faster. The mysterious OnIdle call is traceable to updating the position and
time in the status bar, via the HINT_SONG_POS case in CMainFrame::OnTimer.
Commenting out UpdateSongPositionDisplay() eliminates the mysterious OnIdle
call. UpdateSongPositionDisplay only calls FastSetPaneText and much time was
apparently spent optimizing that method. It boils down to a conditional buffer
reallocation, a string copy, some rectangle manipulation, and InvalidateRect.

Unsurprisingly, it's the InvalidateRect, but why exactly? Strangely, doing 
m_wndStatusBar.UpdateWindow afterwards doesn't help, nor does RedrawWindow.

2/18/24 compacting CTrack

In the future it may become necessary to add members to CTrack. To avoid
increasing CTrack's memory footprint, it could be advantageous to compact 
CTrack by making its member variables smaller where possible. The following
CTrack member variables are currently int (four bytes) and could be char:
Type, Channel, Note, Velocity, RangeType, RangeStart.

Maximal savings would achieved by making all char members adjacent, but
this would affect their order in the user interface, which is undesirable. 
As Type, Channel and Note are adjacent, as are RangeType, RangeStart.
Making those members char instead of int saves 16 bytes, and leaves 2 bytes
available for new char members. Note that converting the Range members to
char saves more space than expected because they're followed by Mute which
is bool; RangeType, RangeStart and Mute would occupy 4 bytes instead of 12.
3 * sizeof(int) + 3 * sizeof(int) -> sizeof(int) + sizeof(int)
24 -> 8
This gets the size of CTrack from 192 bytes down to 176 bytes. Making
Velocity char wouldn't reduce CTrack's size, but would make 3 more
bytes available for reuse.

For the least impact on existing code, make the CSeqTrackArray attribute
wrappers (e.g. SetType and GetType) use char instead of int. It's also 
necessary to overload CPolymeterDoc::ReadEnum for char as follows:

inline void CPolymeterDoc::ReadEnum(CIniFile& fIni, LPCTSTR pszSection, LPCTSTR pszKey, char& Value, const CProperties::OPTION_INFO *pOption, int nOptions)
{
	int	nIntValue;
	ReadEnum(fIni, pszSection, pszKey, nIntValue, pOption, nOptions);
	Value = static_cast<char>(nIntValue);
}

inline void CPolymeterDoc::WriteEnum(CIniFile& fIni, LPCTSTR pszSection, LPCTSTR pszKey, const char& Value, const CProperties::OPTION_INFO *pOption, int nOptions) const
{
	int	nIntValue = Value;
	WriteEnum(fIni, pszSection, pszKey, nIntValue, pOption, nOptions);
}

and also overload CParseCSV::StrToVal for char:

bool CParseCSV::StrToVal(const CString& str, char& nVal)
{
	int	nIntVal;
	if (_stscanf_s(str, _T("%d"), &nIntVal) != 1)
		return false;
	nVal = static_cast<char>(nIntVal);
	return true;
}

The rest of the changes are minor edits to avoid warnings. Note that this
revision could potentially introduce bugs. It could also potentially impact 
performance, due to the need for extra machine language instructions to 
extend bytes to words, though this impact might be reduced or eliminated in
the 64-bit executable as 32-bit words are probably being extended to 64-bit
words anyway. Also the memory savings from reducing the size of CTrack may
improve cache performance and thereby offset any increased code complexity.

2/24/24 docking bar context menu commands should be available in Customize dialog

(issue 529, 7/22/22)

It would nice if commands from some of the dockable windows could be assigned
keyboard shortcuts (accelerators) or added to the main menu. The Modulation, 
Phase, and Graph bars are the most likely to be useful. It seemed at first 
that it was necessary to override CMainFrame::OnCmdMsg to also search the
message maps of dockable bars, and while this does work, it's an expensive
solution in terms of performance. OnCmdMsg is called at high frequency, and
can potentially bog down the user interface. It's also a wasteful solution,
as most of the commands in the dockable window context menus won't be used
in customization, yet their entire message maps have to be searched every
time OnCmdMsg runs.

It's simpler and much more performant to just add command range handlers 
to CMainFrame's message map for the additional command ranges that should
be available in customization. This will get both accelerators and main
menu additions working. However for menu additions, if you want checked
items to appear checked, it's also necessary to add a command update UI
range handler.

ON_COMMAND_RANGE(ID_MODULATION_INSERT_GROUP, ID_MODULATION_SORT_BY_TYPE, OnModulationCmd)
ON_UPDATE_COMMAND_UI_RANGE(ID_MODULATION_INSERT_GROUP, ID_MODULATION_SORT_BY_TYPE, OnUpdateModulationCmd)

void CMainFrame::OnModulationCmd(UINT nID)
{
	m_wndModulationsBar.SendMessage(WM_COMMAND, nID);
}

void CMainFrame::OnUpdateModulationCmd(CCmdUI *pCmdUI)
{
	pCmdUI->DoUpdate(&m_wndModulationsBar, false);
}

This construct must be repeated for each of the desired context menus.
It's possible to generate the code automatically via the preprocessor.
Other than that, it's only necessary to add the desired context menus
to pDlgCust in OnViewCustomize. It might be useful to filter out edit
commands, or any other commands that lie outside the ranges that were 
added to the message map. This too could be done via the preprocessor.

Delete these items from a context menu before passing it to Customize:
menu.DeleteMenu(nEditCmdID, MF_BYCOMMAND);
ID_EDIT_CUT,
ID_EDIT_COPY,
ID_EDIT_PASTE,
ID_EDIT_INSERT,
ID_EDIT_DELETE,
ID_EDIT_SELECT_ALL,
ID_EDIT_RENAME,

2/25/24 More on the update command UI maze

By overriding CMainFrame::OnCmdMsg it's possible to see which commands
are being left to the application due to the absence of handlers in the 
document, child frame or main frame. Every such message is expensive in
terms of performance because CWinAppEx::OnCmdMsg will search all the
message maps again.

BOOL CMainFrame::OnCmdMsg(UINT nID, int nCode, void* pExtra, AFX_CMDHANDLERINFO* pHandlerInfo)
{
	BOOL bRet = CMDIFrameWndEx::OnCmdMsg(nID, nCode, pExtra, pHandlerInfo);
	if (!bRet) printf("%d %d (%x)\n", nCode, nID, nID);
	return bRet;
}

The results are as follows:
0 57607 (e107) // ID_FILE_PRINT
-1 57664 (e140) // ID_APP_ABOUT
-1 101 (65)
-1 101 (65)
-1 57600 (e100) // ID_FILE_NEW
-1 57601 (e101) // ID_FILE_OPEN
-1 57603 (e103) // ID_FILE_SAVE
-1 57607 (e107) // ID_FILE_PRINT

The file commands are due to their presence on the toolbar, proved
by removing them from the toolbar using View/Customize. Based on
this evidence, it would make sense to remove the non-functional
Print icon from the toolbar, as it's needlessly adding overhead.

Overriding the CPolymeterApp::OnCmdMsg reveals that the three status
bar indicators are being routed there.

-1 32789 (8015)	// ID_INDICATOR_SONG_POS
-1 32790 (8016) // ID_INDICATOR_SONG_TIME
-1 32791 (8017) // ID_INDICATOR_TEMPO
-1 57664 (e140) // ID_APP_ABOUT
-1 101 (65)
-1 101 (65)

This occurs because CMDIFrameWnd::OnCmdMsg tries both the child and 
main frames (in that order) and in both cases the work is done by the
bass class CFrameWnd::OnCmdMsg, which first tries the frame's active 
view if any, then the frame itself, and finally the app. In the first 
CFrameWnd::OnCmdMsg (for the child frame) the indicator messages are 
not handled by the child frame, so the app is then tried, uselessly. 

Standard Command Route:
MDI frame window (CMDIFrameWnd) 
1. Active CMDIChildWnd <-- see CMDIChildWnd routing below
2. This frame window
3. Application (CWinApp object)

Document frame window (CFrameWnd, CMDIChildWnd) 	
1. Active view
2. This frame window
3. Application (CWinApp object) <-- this is how the indicators are getting passed to CPolymeterApp::OnCmdMsg before CMainFrame::OnCmdMsg

The easiest solution is to handle the indicators in the child frame.
Because the main frame explicitly updates the song position, time, and
tempo, they don't need update command UI support for their text. They do
need to be enabled while a document exists, but that can be accomplished
by having CChildFrame::OnCmdMsg simply return TRUE for the indicator IDs.
The main frame doesn't even need to have message map entries for the
indicators, provided OnUpdate blanks the tempo pane when playback stops.

	case CPolymeterDoc::HINT_PLAY:
		m_dwCachedTempo = 0;	// force status bar tempo pane to update when playback starts
		if (theApp.m_pPlayingDoc == NULL)
			m_wndStatusBar.SetPaneText(SBP_TEMPO, _T(""));
		break;

The messages with ID 101 come from the tab controls that the framework
creates when multiple dockable panes are docked together in the same 
mini-frame. There's no obvious solution to that.

The keyboard indicators are presumably handled by the app. Eliminating 
them would boost performance, but they're arguably useful.

59137 (E701) // ID_INDICATOR_CAPS
59138 (E702) // ID_INDICATOR_NUM
59139 (E703) // ID_INDICATOR_SCRL

Summary of proposal:
1. Remove the non-functional Print icon from both toolbars.
2. Return TRUE for the indicators in CChildFrame::OnCmdMsg,
and remove the indicator handlers in CMainFrame, along with
the tempo status pane fix shown above.

2/25/24 Phase bar selected tracks only mode

if selected only mode:
in Update, only add selected tracks, and set m_bSelected false
disable OnLButtonDown (exit from top)
disable OnTrackSelectionChange (call Update and exit from top)
in FindNextConvergence, copy all orbit periods to arrMod

On the good side, this mode gives the user a way of directly
specifying which periods are included in the phase diagram.
This is particularly advantageous for documents that contain 
many different periods which would otherwise cause the diagram
to be excessively complex and hard to read.

There are several disadvantages however:

1. The phase window is empty unless tracks are selected,
which is potentially confusing and looks like a bug.

2. The mode causes the Transport Next / Previous Convergence
commands to behave differently (only the selected tracks are
included in the convergence calculation), which similarly 
could be confusing and look like a bug.

3. Because the track selection is ephemeral and often
changes, for example due to editing or merely switching
between documents, the user might have to re-specify 
the periods frequently, which could get aggravating.

4. The mode doesn't work in Live view, because entering 
Live view clears the track selection