Cool_Tools/vgmstream
1
0
Fork 0
mirror of https://github.com/vgmstream/vgmstream.git synced 2025-01-18 15:54:05 +01:00
Code Issues Packages Projects Releases Wiki Activity

Improve MSADPCM performance (~50%) + doc spec diffs

Browse Source
This commit is contained in:
bnnm 2021-03-27 18:15:43 +01:00
parent 0bb0fc5891
commit faafd5b955
3 changed files with 97 additions and 69 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/coding/coding.h
View File

@ -156,6 +156,8 @@ void free_nwa(nwa_codec_data* data);
STREAMFILE* nwa_get_streamfile(nwa_codec_data* data);
/* msadpcm_decoder */
#define MSADPCM_MAX_BLOCK_SIZE 0x800 /* known max and RIFF spec seems to concur, while MS's encoders may be lower (typical stereo: 0x8c, 0x2C, 0x48, 0x400) */
void decode_msadpcm_stereo(VGMSTREAM* vgmstream, sample_t* outbuf, int32_t first_sample, int32_t samples_to_do);
void decode_msadpcm_mono(VGMSTREAM* vgmstream, sample_t* outbuf, int channelspacing, int32_t first_sample, int32_t samples_to_do, int channel);
void decode_msadpcm_ck(VGMSTREAM* vgmstream, sample_t* outbuf, int channelspacing, int32_t first_sample, int32_t samples_to_do, int channel);

156
src/coding/msadpcm_decoder.c
View File

@ -1,7 +1,9 @@
#include "../util.h"
#include "coding.h"
/* table values are interpreted as fixed point 8.8 signed values */
/* AdaptionTable */
static const int16_t msadpcm_steps[16] = {
230, 230, 230, 230,
307, 409, 512, 614,
@ -9,6 +11,7 @@ static const int16_t msadpcm_steps[16] = {
307, 230, 230, 230
};
/* aCoeff table, normally included with container and (in theory) encoder may add extra coefs but first 7 are preset coefs */
static const int16_t msadpcm_coefs[7][2] = {
{ 256, 0 },
{ 512, -256 },
@ -19,16 +22,23 @@ static const int16_t msadpcm_coefs[7][2] = {
{ 392, -232 }
};
/* Decodes MSADPCM as explained in the spec (RIFFNEW / msadpcm.c).
* Though RIFFNEW spec uses "predictor / 256", msadpcm.c uses "predictor >> 8" = diffs on negs (silly MS).
* SHR is also true in Windows msadp32.acm decoders (up to Win10), that seem to use same code.
* Some non-Windows implementations or engines (like UE4) use DIV though (more accurate).
* On invalid coef index, msadpcm.c returns 0 decoded samples but here we clamp and keep on trucking.
* In theory blocks may be 0-padded and should use samples_per_frame from header, in practice seems to
* decode up to block length or available data. */
void decode_msadpcm_stereo(VGMSTREAM* vgmstream, sample_t* outbuf, int32_t first_sample, int32_t samples_to_do) {
VGMSTREAMCHANNEL *ch1,*ch2;
STREAMFILE *streamfile;
VGMSTREAMCHANNEL *stream1, *stream2;
uint8_t frame[MSADPCM_MAX_BLOCK_SIZE] = {0};
int i, frames_in;
size_t bytes_per_frame, samples_per_frame;
off_t frame_offset;
ch1 = &vgmstream->ch[0];
ch2 = &vgmstream->ch[1];
streamfile = ch1->streamfile;
stream1 = &vgmstream->ch[0];
stream2 = &vgmstream->ch[1];
/* external interleave (variable size), stereo */
bytes_per_frame = vgmstream->frame_size;
@ -36,61 +46,63 @@ void decode_msadpcm_stereo(VGMSTREAM* vgmstream, sample_t* outbuf, int32_t first
frames_in = first_sample / samples_per_frame;
first_sample = first_sample % samples_per_frame;
frame_offset = ch1->offset + frames_in*bytes_per_frame;
frame_offset = stream1->offset + frames_in * bytes_per_frame;
read_streamfile(frame, frame_offset, bytes_per_frame, stream1->streamfile); /* ignore EOF errors */
/* parse frame header */
/* parse frame header (ADPCMBLOCKHEADER) */
if (first_sample == 0) {
ch1->adpcm_coef[0] = msadpcm_coefs[read_8bit(frame_offset+0x00,streamfile) & 0x07][0];
ch1->adpcm_coef[1] = msadpcm_coefs[read_8bit(frame_offset+0x00,streamfile) & 0x07][1];
ch2->adpcm_coef[0] = msadpcm_coefs[read_8bit(frame_offset+0x01,streamfile)][0];
ch2->adpcm_coef[1] = msadpcm_coefs[read_8bit(frame_offset+0x01,streamfile)][1];
ch1->adpcm_scale = read_16bitLE(frame_offset+0x02,streamfile);
ch2->adpcm_scale = read_16bitLE(frame_offset+0x04,streamfile);
ch1->adpcm_history1_16 = read_16bitLE(frame_offset+0x06,streamfile);
ch2->adpcm_history1_16 = read_16bitLE(frame_offset+0x08,streamfile);
ch1->adpcm_history2_16 = read_16bitLE(frame_offset+0x0a,streamfile);
ch2->adpcm_history2_16 = read_16bitLE(frame_offset+0x0c,streamfile);
stream1->adpcm_coef[0] = msadpcm_coefs[get_u8(frame+0x00) & 0x07][0]; /* bPredictor[0] index > iCoef1 */
stream1->adpcm_coef[1] = msadpcm_coefs[get_u8(frame+0x00) & 0x07][1]; /* bPredictor[0] index > iCoef2 */
stream2->adpcm_coef[0] = msadpcm_coefs[get_u8(frame+0x01) & 0x07][0]; /* bPredictor[1] index > iCoef1 */
stream2->adpcm_coef[1] = msadpcm_coefs[get_u8(frame+0x01) & 0x07][1]; /* bPredictor[1] index > iCoef2 */
stream1->adpcm_scale = get_s16le(frame+0x02); /* iDelta[0] */
stream2->adpcm_scale = get_s16le(frame+0x04); /* iDelta[0] */
stream1->adpcm_history1_16 = get_s16le(frame+0x06); /* iSamp1[0] */
stream2->adpcm_history1_16 = get_s16le(frame+0x08); /* iSamp1[0] */
stream1->adpcm_history2_16 = get_s16le(frame+0x0a); /* iSamp2[0] */
stream2->adpcm_history2_16 = get_s16le(frame+0x0c); /* iSamp2[1] */
}
/* write header samples (needed) */
if (first_sample==0) {
outbuf[0] = ch1->adpcm_history2_16;
outbuf[1] = ch2->adpcm_history2_16;
outbuf[0] = stream1->adpcm_history2_16;
outbuf[1] = stream2->adpcm_history2_16;
outbuf += 2;
first_sample++;
samples_to_do--;
}
if (first_sample == 1 && samples_to_do > 0) {
outbuf[0] = ch1->adpcm_history1_16;
outbuf[1] = ch2->adpcm_history1_16;
outbuf[0] = stream1->adpcm_history1_16;
outbuf[1] = stream2->adpcm_history1_16;
outbuf += 2;
first_sample++;
samples_to_do--;
}
/* decode nibbles */
for (i = first_sample; i < first_sample+samples_to_do; i++) {
for (i = first_sample; i < first_sample + samples_to_do; i++) {
int ch;
for (ch = 0; ch < 2; ch++) {
VGMSTREAMCHANNEL *stream = &vgmstream->ch[ch];
int32_t hist1,hist2, predicted;
int sample_nibble = (ch == 0) ? /* L = high nibble first */
get_high_nibble_signed(read_8bit(frame_offset+0x07*2+(i-2),streamfile)) :
get_low_nibble_signed (read_8bit(frame_offset+0x07*2+(i-2),streamfile));
VGMSTREAMCHANNEL* stream = &vgmstream->ch[ch];
int32_t hist1, hist2, predicted;
uint8_t byte = get_u8(frame+0x07*2+(i-2));
int sample_nibble = (ch == 0) ? /* L = high nibble first (iErrorDelta) */
get_high_nibble_signed(byte) :
get_low_nibble_signed (byte);
hist1 = stream->adpcm_history1_16;
hist2 = stream->adpcm_history2_16;
predicted = hist1*stream->adpcm_coef[0] + hist2*stream->adpcm_coef[1];
predicted = predicted / 256;
predicted = predicted + sample_nibble*stream->adpcm_scale;
outbuf[0] = clamp16(predicted);
predicted = hist1 * stream->adpcm_coef[0] + hist2 * stream->adpcm_coef[1];
predicted = predicted / 256; /* 256 = FIXED_POINT_COEF_BASE (though MS code uses SHR) */
predicted = predicted + (sample_nibble * stream->adpcm_scale);
outbuf[0] = clamp16(predicted); /* lNewSample */
stream->adpcm_history2_16 = stream->adpcm_history1_16;
stream->adpcm_history1_16 = outbuf[0];
stream->adpcm_scale = (msadpcm_steps[sample_nibble & 0xf] * stream->adpcm_scale) / 256;
if (stream->adpcm_scale < 0x10)
stream->adpcm_scale = 0x10;
stream->adpcm_scale = (msadpcm_steps[sample_nibble & 0xf] * stream->adpcm_scale) / 256; /* 256 = FIXED_POINT_ADAPTION_BASE */
if (stream->adpcm_scale < 16) /* min delta */
stream->adpcm_scale = 16;
outbuf++;
}
@ -98,7 +110,8 @@ void decode_msadpcm_stereo(VGMSTREAM* vgmstream, sample_t* outbuf, int32_t first
}
void decode_msadpcm_mono(VGMSTREAM* vgmstream, sample_t* outbuf, int channelspacing, int32_t first_sample, int32_t samples_to_do, int channel) {
VGMSTREAMCHANNEL *stream = &vgmstream->ch[channel];
VGMSTREAMCHANNEL* stream = &vgmstream->ch[channel];
uint8_t frame[MSADPCM_MAX_BLOCK_SIZE] = {0};
int i, frames_in;
size_t bytes_per_frame, samples_per_frame;
off_t frame_offset;
@ -109,15 +122,16 @@ void decode_msadpcm_mono(VGMSTREAM* vgmstream, sample_t* outbuf, int channelspac
frames_in = first_sample / samples_per_frame;
first_sample = first_sample % samples_per_frame;
frame_offset = stream->offset + frames_in*bytes_per_frame;
frame_offset = stream->offset + frames_in * bytes_per_frame;
read_streamfile(frame, frame_offset, bytes_per_frame, stream->streamfile); /* ignore EOF errors */
/* parse frame header */
if (first_sample == 0) {
stream->adpcm_coef[0] = msadpcm_coefs[read_8bit(frame_offset+0x00,stream->streamfile) & 0x07][0];
stream->adpcm_coef[1] = msadpcm_coefs[read_8bit(frame_offset+0x00,stream->streamfile) & 0x07][1];
stream->adpcm_scale = read_16bitLE(frame_offset+0x01,stream->streamfile);
stream->adpcm_history1_16 = read_16bitLE(frame_offset+0x03,stream->streamfile);
stream->adpcm_history2_16 = read_16bitLE(frame_offset+0x05,stream->streamfile);
stream->adpcm_coef[0] = msadpcm_coefs[get_u8(frame+0x00) & 0x07][0];
stream->adpcm_coef[1] = msadpcm_coefs[get_u8(frame+0x00) & 0x07][1];
stream->adpcm_scale = get_s16le(frame+0x01);
stream->adpcm_history1_16 = get_s16le(frame+0x03);
stream->adpcm_history2_16 = get_s16le(frame+0x05);
}
/* write header samples (needed) */
@ -135,33 +149,35 @@ void decode_msadpcm_mono(VGMSTREAM* vgmstream, sample_t* outbuf, int channelspac
}
/* decode nibbles */
for (i = first_sample; i < first_sample+samples_to_do; i++) {
int32_t hist1,hist2, predicted;
for (i = first_sample; i < first_sample + samples_to_do; i++) {
int32_t hist1, hist2, predicted;
uint8_t byte = get_u8(frame+0x07+(i-2)/2);
int sample_nibble = (i & 1) ? /* high nibble first */
get_low_nibble_signed (read_8bit(frame_offset+0x07+(i-2)/2,stream->streamfile)) :
get_high_nibble_signed(read_8bit(frame_offset+0x07+(i-2)/2,stream->streamfile));
get_low_nibble_signed (byte) :
get_high_nibble_signed(byte);
hist1 = stream->adpcm_history1_16;
hist2 = stream->adpcm_history2_16;
predicted = hist1*stream->adpcm_coef[0] + hist2*stream->adpcm_coef[1];
predicted = hist1 * stream->adpcm_coef[0] + hist2 * stream->adpcm_coef[1];
predicted = predicted / 256;
predicted = predicted + sample_nibble*stream->adpcm_scale;
predicted = predicted + (sample_nibble * stream->adpcm_scale);
outbuf[0] = clamp16(predicted);
stream->adpcm_history2_16 = stream->adpcm_history1_16;
stream->adpcm_history1_16 = outbuf[0];
stream->adpcm_scale = (msadpcm_steps[sample_nibble & 0xf] * stream->adpcm_scale) / 256;
if (stream->adpcm_scale < 0x10)
stream->adpcm_scale = 0x10;
if (stream->adpcm_scale < 16) /* min delta */
stream->adpcm_scale = 16;
outbuf += channelspacing;
}
}
/* Cricket Audio's MSADPCM, same thing with reversed hist and nibble order
* (their tools may convert to float/others but internally it's all PCM16, from debugging). */
/* Cricket Audio's MSADPCM, same thing with reversed hist and nibble order, reverse engineered from the exe.
* (their tools may convert to float/others but internally it's all PCM16). */
void decode_msadpcm_ck(VGMSTREAM* vgmstream, sample_t* outbuf, int channelspacing, int32_t first_sample, int32_t samples_to_do, int channel) {
VGMSTREAMCHANNEL *stream = &vgmstream->ch[channel];
VGMSTREAMCHANNEL* stream = &vgmstream->ch[channel];
uint8_t frame[MSADPCM_MAX_BLOCK_SIZE] = {0};
int i, frames_in;
size_t bytes_per_frame, samples_per_frame;
off_t frame_offset;
@ -172,15 +188,16 @@ void decode_msadpcm_ck(VGMSTREAM* vgmstream, sample_t* outbuf, int channelspacin
frames_in = first_sample / samples_per_frame;
first_sample = first_sample % samples_per_frame;
frame_offset = stream->offset + frames_in*bytes_per_frame;
frame_offset = stream->offset + frames_in * bytes_per_frame;
read_streamfile(frame, frame_offset, bytes_per_frame, stream->streamfile); /* ignore EOF errors */
/* parse frame header */
if (first_sample == 0) {
stream->adpcm_coef[0] = msadpcm_coefs[read_8bit(frame_offset+0x00,stream->streamfile) & 0x07][0];
stream->adpcm_coef[1] = msadpcm_coefs[read_8bit(frame_offset+0x00,stream->streamfile) & 0x07][1];
stream->adpcm_scale = read_16bitLE(frame_offset+0x01,stream->streamfile);
stream->adpcm_history2_16 = read_16bitLE(frame_offset+0x03,stream->streamfile); /* hist2 first, unlike normal MSADPCM */
stream->adpcm_history1_16 = read_16bitLE(frame_offset+0x05,stream->streamfile);
stream->adpcm_coef[0] = msadpcm_coefs[get_u8(frame+0x00) & 0x07][0];
stream->adpcm_coef[1] = msadpcm_coefs[get_u8(frame+0x00) & 0x07][1];
stream->adpcm_scale = get_s16le(frame+0x01);
stream->adpcm_history2_16 = get_s16le(frame+0x03); /* hist2 first, unlike normal MSADPCM */
stream->adpcm_history1_16 = get_s16le(frame+0x05);
}
/* write header samples (needed) */
@ -200,22 +217,23 @@ void decode_msadpcm_ck(VGMSTREAM* vgmstream, sample_t* outbuf, int channelspacin
/* decode nibbles */
for (i = first_sample; i < first_sample+samples_to_do; i++) {
int32_t hist1,hist2, predicted;
uint8_t byte = get_u8(frame+0x07+(i-2)/2);
int sample_nibble = (i & 1) ? /* low nibble first, unlike normal MSADPCM */
get_high_nibble_signed (read_8bit(frame_offset+0x07+(i-2)/2,stream->streamfile)) :
get_low_nibble_signed(read_8bit(frame_offset+0x07+(i-2)/2,stream->streamfile));
get_high_nibble_signed(byte) :
get_low_nibble_signed (byte);
hist1 = stream->adpcm_history1_16;
hist2 = stream->adpcm_history2_16;
predicted = hist1*stream->adpcm_coef[0] + hist2*stream->adpcm_coef[1];
predicted = predicted >> 8; /* probably no difference vs MSADPCM */
predicted = predicted + sample_nibble*stream->adpcm_scale;
predicted = hist1 * stream->adpcm_coef[0] + hist2 *stream->adpcm_coef[1];
predicted = predicted >> 8; /* not DIV unlike spec */
predicted = predicted + (sample_nibble * stream->adpcm_scale);
outbuf[0] = clamp16(predicted);
stream->adpcm_history2_16 = stream->adpcm_history1_16;
stream->adpcm_history1_16 = outbuf[0];
stream->adpcm_scale = (msadpcm_steps[sample_nibble & 0xf] * stream->adpcm_scale) >> 8;
if (stream->adpcm_scale < 0x10)
stream->adpcm_scale = 0x10;
stream->adpcm_scale = (msadpcm_steps[sample_nibble & 0xf] * stream->adpcm_scale) >> 8; /* not DIV but same here (always >=0) */
if (stream->adpcm_scale < 16)
stream->adpcm_scale = 16;
outbuf += channelspacing;
}
@ -230,7 +248,7 @@ long msadpcm_bytes_to_samples(long bytes, int block_size, int channels) {
/* test if MSADPCM coefs were re-defined (possible in theory but not used in practice) */
int msadpcm_check_coefs(STREAMFILE* sf, off_t offset) {
int i;
int count = read_16bitLE(offset, sf);
int count = read_u16le(offset, sf);
if (count != 7) {
VGM_LOG("MSADPCM: bad count %i at %lx\n", count, offset);
goto fail;
@ -238,8 +256,8 @@ int msadpcm_check_coefs(STREAMFILE* sf, off_t offset) {
offset += 0x02;
for (i = 0; i < 7; i++) {
int16_t coef1 = read_16bitLE(offset + 0x00, sf);
int16_t coef2 = read_16bitLE(offset + 0x02, sf);
int16_t coef1 = read_s16le(offset + 0x00, sf);
int16_t coef2 = read_s16le(offset + 0x02, sf);
if (coef1 != msadpcm_coefs[i][0] || coef2 != msadpcm_coefs[i][1]) {
VGM_LOG("MSADPCM: bad coef %i/%i vs %i/%i\n", coef1, coef2, msadpcm_coefs[i][0], msadpcm_coefs[i][1]);

8
src/vgmstream.c
View File

@ -1530,6 +1530,14 @@ int vgmstream_open_stream_bf(VGMSTREAM* vgmstream, STREAMFILE* sf, off_t start_o
vgmstream->frame_size = vgmstream->interleave_block_size;
}
if ((vgmstream->coding_type == coding_MSADPCM ||
vgmstream->coding_type == coding_MSADPCM_ck ||
vgmstream->coding_type == coding_MSADPCM_int) &&
(vgmstream->frame_size > MSADPCM_MAX_BLOCK_SIZE)) {
VGM_LOG("VGMSTREAM: MSADPCM decoder with wrong frame size %x\n", vgmstream->frame_size);
goto fail;
}
/* big interleaved values for non-interleaved data may result in incorrect behavior,
* quick fix for now since layouts are finicky, with 'interleave' left for meta info
* (certain layouts+codecs combos results in funny output too, should rework the whole thing) */