65
edits
Changes
Jump to navigation
Jump to search
m
Part I: Structures(Wiki-fied by Halkun)1.FF7FrameHeader2.FF7FrameMiniHeader3.FF7ShortVec4.FF7FrameBuffer(Small additions by Borde)
Part II: Functions and Format1.GetBitsFixed()A. Format2.GetDynamicFrameOffsetBits()3.GetEncryptedRotationBits() Part III: Putting it All Together1.LoadFrames()2.A Sample Loop Part IV: Qhimm’s Input === Part I: Structures===
Now the code to skip to any animation, by index, where “iTarget†is the index. This code assumes you have already opened the animation file (hFile) and you have skipped pasted the first 4 bytes.
// chunk, skip it (it // is counted as part // of the total in the // file). if ( SetFilePointer( hFile, 8, NULL, FILE_CURRENT ) == INVALID_SET_FILE_POINTER ) {
*Read X bits and store as a signed SHORT. *Convert the SHORT to the INT field, adding 0x1000 if negative. *Convert to FLOAT using (INT / 4096 * 360). Apply this FLOAT to your model.
*Read X bits, and add them to the SHORT value from last frame.*Convert the SHORT to the INT field, adding 0x1000 if negative. *Convert to FLOAT using (INT / 4096 * 360). Apply this FLOAT to your model.
// This can never actually happen.
2.==== A Sample Loop ====
magnitude of -1).
signed integer overflow). Small values will be stored using fewer bits, while larger values use more bits. The two smallest values, 0 and -1, are not encodeable but are instead handled using the previously mentioned scheme.
14 revisions imported
== Battle Animation File Format==
File format discovered/decoded by me, L. Spiro.
This file was written by me, L. Spiro, as can be seen by the proper grammar and perfect spelling.
There are 4 basic structures we will use in decoding the file format.
The header of animation data has been considered to be composed of 3 DWORD’s, 3 WORD’s, and one BYTE, however this is not how the header is really intended to be, despite being aligned correctly.
Battle animation files start with a DWORD which tells us how many animations are in the file. This number includes the special animations which are not actually animations at all. Although I have not yet decoded them, I suspect these are sets of keys for actual animation sets; keys that call scripted actions or tell the engine to print the damage numbers.
==== FF7FrameHeader ====
Cloud’s battle animation file (rtda) has 94 (0x 5E) animations in it.
After this number begins each animation.
Each animation begins with a 12-byte header (3 DWORD’s) we will call “FF7FrameHeaderâ€"FF7FrameHeader".To get from one animation to the next, start at offset 0x04 in the animation file and begin reading these headers. For each header, skip “FF7FrameHeader"FF7FrameHeader.dwChunkSize†dwChunkSize" bytes until you get to the index of the file you want to load. When skipping, remember to skip starting at the end of the “FF7FrameHeader†"FF7FrameHeader" header.
I mentioned a type of special animation data set that is in the header file.
These data sets, when filled with the “FF7FrameHeader†"FF7FrameHeader" header, will have a “dwChunkSize†"dwChunkSize" less than eleven, we skip them by jumping over the next 8 bytes that follow.
1.
<code><pre>
typedef struct FF7FrameHeader {
DWORD dwBones; // Bones in the model + 1(unless we're dealing with a weapon animation, in which case it's value is always 1). 0x00 // This field is rather unreliable so better use the number of bones provided by the skeleton file.
DWORD dwFrames; // Frames in the animation. 0x04
DWORD dwChunkSize; // Size of the animation set. 0x08
} * PFF7FrameHeader; // Size = 12 bytes.
</pre></code>
2.
<code><pre>
typedef struct FF7FrameMiniHeader {
//SHORT sBones; // Bones in the animation. 0x00SHORT sFrames;// Apparently, frames in the animation (but sometimes sFrames > dwFrames)
SHORT sSize; // Size of the animation data. 0x02
BYTE bKey; // A key flag used for decoding. 0x04
} * PFF7FrameMiniHeader; // Size = 5 bytes.
</pre></code>
NOTE: sBones should provably be called sFrames since it seems to hold a secondary frames counter. Thus, it should be equal to dwFrames. Unfortunately, it usually isn't. In fact, It's hard to say which one should actually be trusted. Apparently dwFrames is a more conservative value, meaning there will always be at least that many frames in the animation. But there can be more of them. I can't help but wonder if this means the rest of the frames are dummied out information or they serve some sort of purpose. On the other hand, sFrames is sometimes higher than the actual number of frames on the animation chunck.
Anyway, the actual number of frames can be computed by parsing the whole animation chunck.
It's also worth mentioning that there is at least one animation (15th from RSAA, the playable frog) which physically lacks the sFrames field. Instead, sSize is at 0x00 and bKey at 0x02. This animation is more than likely damaged, because FF7 doesn't seems to be able to handle it.
==== FF7FrameMiniHeader ====Most of these members are straight-forward, however there is a very special and VERY important member in the “FF7FrameMiniHeader†"FF7FrameMiniHeader" structure called “bKeyâ€"bKey".
This is used for every rotation-decoding scheme (but one). It determines, essentially, the precision of the rotations and the deltas that follow in successive frames.
The value of “bKey†can only be 0, 2, or 4; the equation "(12 - bKey) " is used to determine the length of each raw (uncompressed) rotation.
After decompression, every rotation must be 12 bits, giving it a range from 0 to 4095.
But if “bKey†"bKey" is 4, for example, then that means uncompressed rotations are stored as 8 bits, which gives them a range from 0 to 255. How is this fixed? After the 8 bits are read, they are then shifted left (up) by “bKeyâ€"bKey". This will place them at 12 bits, but with decreased accuracy.
This loss in accuracy is acceptable since rotations work as deltas and usually only change by a small amount.
Most large rotation deltas are things that are spinning, such as the blades on Aero Combatant. These cases are always a nice round number that can be handled with lower precision (in the case of Aero Combatant, it is 90 degrees even).
==== FF7ShortVec ====
Now the code to skip to any animation, by index, where "iTarget" is the index. This code assumes you have already opened the animation file (hFile) and you have skipped pasted the first 4 bytes.
<code><pre>
FF7FrameHeader fhHeader;
DWORD dwBytesRead;
for ( int I = 0; I < iTarget; I++ ) {
if ( !ReadFile( hFile, &fhHeader, sizeof( fhHeader ), &dwBytesRead, NULL ) ) {
CloseHandle( hFile ); return false; }
if ( fhHeader.dwChunkSize < 11 ) { // If this is a special
CloseHandle( hFile );
return false;
}
continue; // Go on to the next // animation. } // Skip this animation set, whose size is determined by // fhHeader.dwChunkSize. if ( SetFilePointer( hFile, fhHeader.dwChunkSize, NULL, FILE_CURRENT ) == INVALID_SET_FILE_POINTER ) {
CloseHandle( hFile );
return false;
}
}
// Once we come to this point, we are at the very first byte of the
// animation we want to load. Let’s store it into a BYTE array.
if ( !ReadFile( hFile,
&fhHeader, sizeof( fhHeader ), &dwBytesRead, NULL ) ) {
CloseHandle( hFile ); return false;
}
BYTE * pbBuffer = new BYTE[fhHeader.dwChunkSize];
// Now pbBuffer holds the actual animation data, including the 5-byte
// “FF7FrameMiniHeader†header.
</pre></code>
We now have the animation we want loaded into a BYTE array (remember to delete it later).
==== FF7FrameBuffer ====
Now let’s look at the other structures we will use.
3.
<code><pre>
typedef struct FF7ShortVec {
SHORT sX, sY, sZ; // Signed short versions. 0x00
FLOAT fX, fY, fZ; // Float version after math. 0x12
} * PFF7ShortRot; // Size = 30 bytes.
</pre></code>
Each rotation goes through 3 forms. Firstly, everything is stored as 2-byte SHORT’s. These SHORT’s are stored from 0 to 4096, where 0 = 0 degrees and 4096 = 360 degrees. This is the equation to convert one of these SHORT’s into degrees: (SHORT / 4096 * 360). Each frame is based off the previous frame, using the SHORT value as its basis.
Each SHORT is converted to an INT, which is the exact same as the SHORT version, except always positive.
Finally, the FLOAT gets filled with the final value, using the INT version as its base.
So, the sequence is:
First frame…
Next frame…
Repeat…
To load an entire frame’s work of bones, we need this structure:
4.
<code><pre>
typedef struct FF7FrameBuffer {
DWORD dwBones;
FF7ShortVec svPosOffset;
FF7ShortVec * psvRots;
FF7FrameBuffer() {
}
} * PFF7FrameBuffer;
</pre></code>
This structure will allocate enough memory for one frame of rotations. Simply call “FF7FrameBuffer.SetBones†with the number of bones in your animation.
=== Part II: Functions and Format===
First, we need a way to read bits from the BYTE array we have stored.
This is a basic bit-reading function. It reads “dwTotalBits†from “pbBuffer†starting at the “dwStartBitâ€â€™th bit.
==== GetBitsFixed ====
1.
<code><pre>
INT GetBitsFixed( BYTE * pbBuffer, DWORD &dwStartBit,
DWORD dwTotalBits ) {
return iReturn;
}
</pre></code>
Now that we can read the bits in the buffer we have made, it’s time to know what we’re doing!
===== A.=====
The animation data begins with one full frame that is uncompressed, but stored in one of 3 ways. Every frame after that is compressed, but compressed in one of three ways; one way can be decoded using the same method as on the first frame, which is why sometimes the second, third, and even fourth frames can be decoded using the same method as was used on the first frame.
First Frame:
Remember that we stored our animation buffer with a 5-byte “FF7FrameMiniHeader†at the beginning of it? We need this header now!
<code><pre>
PFF7FrameMiniHeader pfmhMiniHeader = (PFF7FrameMiniHeader)pbBuffer;
</pre></code>
After this cast, “pfmhMiniHeader->bKey†will contain a number, either 0, 2, or 4.
To get these bits, we first need to make a pointer point to the correct location. “pbBuffer†points 5 bytes before this data, so let’s make a pointer that points to this data directly.
<code><pre>
BYTE * pbAnimBuffer = &pbBuffer[5];
</pre></code>
When we use “GetBitsFixed()†to get the bits.
<code><pre>
DWORD dwBitStart = 0; // The bits at which to begin reading in the
// stream.
SHORT sY = GetBitsFixed( pbAnimBuffer, dwBitStart, 16 );
SHORT sZ = GetBitsFixed( pbAnimBuffer, dwBitStart, 16 );
</pre></code>
After doing this, we have each of the three offsets, 0, -466, and 0.
For each bone, there are 3 rotations. So, for each bone, we do this:
<code><pre>
SHORT sRotX = GetBitsFixed( pbAnimBuffer, dwBitStart,
12 - pfmhMiniHeader->bKey );
sRotY <<= pfmhMiniHeader->bKey;
sRotZ <<= pfmhMiniHeader->bKey;
</pre></code>
The first rotation is always 0, 0, 0. This is the root rotation and is not actually counted as part of the bone network of the character.
==== GetDynamicFrameOffsetBits ====
The first frame is easy.
Remember that all frames after are stored as relative offsets from the frame before it.
2.
<code><pre>
SHORT GetDynamicFrameOffsetBits( BYTE * pBuffer, DWORD &dwBitStart ) {
DWORD dwFirstByte, dwConsumedBits, dwBitsRemainingToNextByte, dwTemp;
mov ecx, pBuffer
add ecx, dwFirstByte // Go to the first byte that
// has the bit where we // want to begin.
xor edx, edx
mov dl, byte ptr ds:[ecx]
mov ecx, dwBitStart //
mov edx, [ecx] //
add edx, 8 //
mov eax, dwBitStart //
mov [eax], edx // Increase dwBitStart by 0x8 (8).
mov edx, dwBitStart //
mov [edx], ecx // Increase dwBitStart by 0x11
// (17).
End :
return sReturn;
}
</pre></code>
After the first frame, we know that the positional offsets immediately follow.
So to get the positional deltas for the next frame, we would do this:
<code><pre>
SHORT sDeltaX = GetDynamicFrameOffsetBits( pbAnimBuffer, dwBitStart );
SHORT sDeltaY = GetDynamicFrameOffsetBits( pbAnimBuffer, dwBitStart );
SHORT sDeltaZ = GetDynamicFrameOffsetBits( pbAnimBuffer, dwBitStart );
</pre></code>
Now we have the change from the previous frame. In our “FF7ShortVec†structure, these are the SHORT values. To get the position of this frame, we add these offsets to the last frame’s position.
If “I†is this frame and “I-1†is the last frame, we could do something like this:
<code><pre>
FF7FrameBuffer[I].svPosOffset.sX = FF7FrameBuffer[I-1].svPosOffset.sX + sX;
FF7FrameBuffer[I].svPosOffset.sY = FF7FrameBuffer[I-1].svPosOffset.sY + sY;
FF7FrameBuffer[I].svPosOffset.sZ = FF7FrameBuffer[I-1].svPosOffset.sZ + sZ;
</pre></code>
==== GetEncryptedRotationBits ====
Now all that is left is to decode the rotations.
Rotations change size in multiple ways.
3.
<code><pre>
SHORT GetEncryptedRotationBits( BYTE * pBuffer, DWORD &dwBitStart,
INT iKeyBits ) {
INT iBits = GetBitsFixed( pBuffer, dwBitStart, 1 );
__asm mov eax, iBits // If the first bit is 0, return 0
// and continue. It is not necessary // to mov iBits into EAX, but I do it // anyway.
__asm test eax, eax
__asm jnz SecondTest
__asm mov ecx, dwNumBits
__asm mov dwType, ecx // dwType = ecx = dwNumBits = eax =
// (iBits & 7). // When we get to the case, all of // these values are the same.
__asm cmp dwType, 7
__asm ja ReturnZero // Is dwType above 7? If so, return 0.
// Otherwise, use it in a switch case.
case 0 : {
__asm or eax, 0xFFFFFFFF // After this, EAX will
// always be -1.
__asm mov ecx, iKeyBits
__asm shl eax, cl // Shift left by // precision. // (-1 << iKeyBits)
__asm mov sReturn, ax // Return that number.
__asm jmp End
case 6 : {
// Get a number of bits equal to the case switch (1,
// 2, 3, 4, 5, or 6).
iTemp = GetBitsFixed( pBuffer, dwBitStart,
dwNumBits );
__asm mov eax, iTemp
__asm cmp iTemp, 0
// If greater than or equal to 0…
__asm mov ecx, dwNumBits // dwNumBits = (iBits &
// 7) from before. __asm sub ecx, 1 // dwNumBits - 1.
__asm mov eax, 1
__asm shl eax, cl // (1 << (dwNumBits – // 1)).
__asm mov ecx, iTemp
__asm add ecx, eax // iTemp += (1 <<
// (dwNumBits - 1)).
__asm mov iTemp, ecx
__asm jmp AfterTests
IfLessThanZero :
__asm mov ecx, dwNumBits // dwNumBits = (iBits &
// 7) from before. __asm sub ecx, 1 // Decrease it by 1.
__asm mov edx, 1
__asm shl edx, cl // Shift “1†left by // (dwNumBits - 1).
__asm mov eax, iTemp // iTemp still has the
// bits we read // from before.
__asm sub eax, edx // iTemp - (1 <<
// (dwNumBits - 1))
__asm mov iTemp, eax
// Now, whatever we set on iTemp, we need to shift it
// up by the precision value.
AfterTests :
__asm mov eax, iTemp
// Uncompressed bits. Use standard decoding.
iTemp = GetBitsFixed( pBuffer, dwBitStart,
12 - iKeyBits );
__asm mov ecx, iKeyBits
__asm shl eax, cl // iTemp <<= iKeyBits.
return sReturn;
}
</pre></code>
=== Part III: Putting it All Together===
To make life easy, let’s use one function to load an entire frame at a time.
This function will load an entire frame into a “FF7FrameBuffer†structure.
After the function returns, we must translate the rotational INT values to their FLOAT forms (although the function can be modified to do this part itself).
This function will be called in a loop for every frame in the rotation.
==== LoadFrames ====
1.
<code><pre>
DWORD LoadFrames( PFF7FrameBuffer pfbFrameBuffer,
INT iBones,
PFF7FrameMiniHeader pfmhMiniHeader =
(PFF7FrameMiniHeader)pbAnimBuffer; SHORT sSize = pfmhMiniHeader->sSize; BYTE bKeyBits = pfmhMiniHeader->bKey;
// Skip the first 5 bytes because they are part of the frame header.
if ( iBitStart == 0 ) { // First frame?
// The first frame is uncompressed and each value is the
// actual rotation. pfbFrameBuffer->svPosOffset.sX = GetBitsFixed( pbThisBuffer, dwThisBitStart, 16 ); // Always 16 bits here.
pfbFrameBuffer->svPosOffset.sY = GetBitsFixed( pbThisBuffer, dwThisBitStart, 16 );
pfbFrameBuffer->svPosOffset.sZ = GetBitsFixed( pbThisBuffer, dwThisBitStart, 16 );
// This function will set the FLOAT values for the
// positions.
// Any scaling that needs to be done would be done here.
pfbFrameBuffer->svPosOffset.fX = (FLOAT)pfbFrameBuffer->svPosOffset.sX;
pfbFrameBuffer->svPosOffset.fY = (FLOAT)pfbFrameBuffer->svPosOffset.sY;
// If we did not read as many bits as there are in the frame,
// return the location where the bits should start for the
// next frame.
if ( (SHORT)(dwThisBitStart / 8) < sSize ) {
return dwThisBitStart;
return 0;
}
</pre></code>
2.
This is an example loop that could be used to load a full animation.
<code><pre>
FF7FrameHeader fhHeader;
ReadFile( hFile, &fhHeader, sizeof( fhHeader ), &ulBytesRead,
// will load diretly into it.
// Every frame after that will actually use it with the
// offsets loaded to determine the final result of that // frame. iBits = LoadFrames( &fbFrameBuffer, fhHeader.dwBones,iBits, baData );
// Reverse the Y offset (required).
fbFrameBuffer.svPosOffset.fY = 0.0f –fbFrameBuffer.svPosOffset.fY;
// The first rotation set is skipped. It is not part of
// to dynamically make the model point at its target
// or face different directions during battle.
// UPDATE: Although the value of this field is 0,0,0 for most animations, some actually store a base rotation here
// so it shouldn't be ignored.
for ( DWORD I = 0; I < fhHeader.dwBones - 1; I++ ) {
fbFrameBuffer.psvRots[I+1].fX = (FLOAT)fbFrameBuffer.psvRots[I+1].iX / 4096.0f * 360.0f;
delete [] baData;
</pre></code>
== Part IV: Qhimm’s Input==
Qhimm has taken the time to rewrite two of these functions used in decoding, so it is easier to understand for people who know C++ better than they know assembly (despite my comments being in the assembly code).
He has also written a more in-depth look at the logistics behind the rotation compression format and explains its limitations
“GetValueFromStream†is the C/C++ version of my “GetDynamicFrameOffsetBits†and his “GetCompressedDeltaFromStream†is the C++ version of my “GetEncryptedRotationBitsâ€.
<code><pre>
short GetValueFromStream( BYTE *pStreamBytes,
DWORD *pdwStreamBitOffset )
// Collect one more byte from the stream.
dwNextStreamBytes = dwNextStreamBytes << 8 |
pStreamBytes[dwStreamByteOffset + 2];
// Shift the delta value into place.
sValue = (dwNextStreamBytes << (dwCurrentBitsEaten + 1)) >> 8;
{
unsigned int uType = GetBitsFromStream( pStreamBytes,
pdwStreamBitOffset, 3 ) & 7;
switch (uType)
{
case 0:
// Return the smallest possible decrement delta (at given
// precision).
return (-1 << nLoweredPrecisionBits);
// Read a corresponding number of bits from the stream.
int iTemp = GetBitsFromStream( pStreamBytes,
pdwStreamBitOffset, nBits );
// Transform the value into the full seven-bit value, using the bit
// length
// as part of the encoding scheme (see notes).
if (iTemp < 0) iTemp -= 1 << (nBits - 1);
case 7:
// Read an uncompressed value from the stream (at requested
// precision), and return.
iTemp = GetBitsFromStream( pStreamBytes,
pdwStreamBitOffset, 12 - nLoweredPrecisionBits );
return (iTemp << nLoweredPrecisionBits);
default:
encodings available for small deltas. The smallest 128 [-64,63] sizes of deltas
can be stored in compressed form instead of as raw values. This method is
efficient
since a majority of the rotational deltas involved in skeletal character animation
will be small, and thus doubly effective if used with reduced precision. Precision
First, a single bit tells us if the delta is non-zero. If this bit is zero, there
is no delta value (0) and the decoding is done. Otherwise, a 3 bit integer
follows, detailing how the delta value is encoded. This has 8 different meanings, as follows:
Type Meaning
------ -------------------------------------------------------
0 The delta is the smallest possible decrement (under the current precision)
1-6 The delta is encoded using this many bytes
7 The delta is stored in raw form (in the current precision)
The encoding of small deltas works as follows: The encoded delta can be stored
using 1-6 bits, giving us a total of 2+4+8+16+32+64 = 126 possible different
values, which during this explanation will be explained as simple integers (the lowest bits of the delta, in current precision). The values 0 (no change) and -1 (minimal decrement) are already covered, leaving the other 126 values to neatly fill out the entire 7 bit range. We do this by encoding each value like follows:
- The magnitude of the delta is defined as the value of its most significant
have a magnitude of 32. For simplicity, we also define the 'signed magnitude' as
the magnitude multiplied by the sign of the value (so '-2' has a signed
- When encoding a value, we subtract its signed magnitude; essentially pushing
everything down one notch towards zero, setting the most significant value bit
- The transformed value is then stored starting from its magnitude bit (normally,
you would have to start one bit higher to include the sign bit and prevent
When decoding, you only need to know the number of bits of the encoded value, use
*/
</pre></code>
