Skip to content

Commit

Permalink
[huf] Improve fast huffman decoding speed in linux kernel
Browse files Browse the repository at this point in the history
gcc in the linux kernel was not unrolling the inner loops of the Huffman
decoder, which was destroying decoding performance. The compiler was
generating crazy code with all sorts of branches. I suspect because of
Spectre mitigations, but I'm not certain. Once the loops were manually
unrolled, performance was restored.

Additionally, when gcc couldn't prove that the variable left shift in
the 4X2 decode loop wasn't greater than 63, it inserted checks to verify
it. To fix this, mask `entry.nbBits & 0x3F`, which allows gcc to eliete
this check. This is a no op, because `entry.nbBits` is guaranteed to be
less than 64.

Lastly, introduce the `HUF_DISABLE_FAST_DECODE` macro to disable the
fast C loops for Issue facebook#3762. So if even after this change, there is a
performance regression, users can opt-out at compile time.
  • Loading branch information
Nick Terrell authored and terrelln committed Nov 20, 2023
1 parent e122fcb commit c7269ad
Show file tree
Hide file tree
Showing 2 changed files with 109 additions and 66 deletions.
4 changes: 4 additions & 0 deletions lib/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -178,6 +178,10 @@ The file structure is designed to make this selection manually achievable for an
`ZSTDERRORLIB_VSIBILITY`, and `ZDICTLIB_VISIBILITY` if unset, for backwards compatibility
with the old macro names.

- The C compiler macro `HUF_DISABLE_FAST_DECODE` disables the newer Huffman fast C
and assembly decoding loops. You may want to use this macro if these loops are
slower on your platform.

#### Windows : using MinGW+MSYS to create DLL

DLL can be created using MinGW+MSYS with the `make libzstd` command.
Expand Down
171 changes: 105 additions & 66 deletions lib/decompress/huf_decompress.c
Original file line number Diff line number Diff line change
Expand Up @@ -34,6 +34,12 @@
* Macros
****************************************************************/

#ifdef HUF_DISABLE_FAST_DECODE
# define HUF_ENABLE_FAST_DECODE 0
#else
# define HUF_ENABLE_FAST_DECODE 1
#endif

/* These two optional macros force the use one way or another of the two
* Huffman decompression implementations. You can't force in both directions
* at the same time.
Expand Down Expand Up @@ -292,6 +298,24 @@ static size_t HUF_initRemainingDStream(BIT_DStream_t* bit, HUF_DecompressFastArg
return 0;
}

/* Calls X(N) for each stream 0, 1, 2, 3. */
#define HUF_4X_FOR_EACH_STREAM(X) \
{ \
X(0) \
X(1) \
X(2) \
X(3) \
}

/* Calls X(N, var) for each stream 0, 1, 2, 3. */
#define HUF_4X_FOR_EACH_STREAM_WITH_VAR(X, var) \
{ \
X(0, (var)) \
X(1, (var)) \
X(2, (var)) \
X(3, (var)) \
}


#ifndef HUF_FORCE_DECOMPRESS_X2

Expand Down Expand Up @@ -706,7 +730,6 @@ void HUF_decompress4X1_usingDTable_internal_fast_c_loop(HUF_DecompressFastArgs*
for (;;) {
BYTE* olimit;
int stream;
int symbol;

/* Assert loop preconditions */
#ifndef NDEBUG
Expand Down Expand Up @@ -753,27 +776,42 @@ void HUF_decompress4X1_usingDTable_internal_fast_c_loop(HUF_DecompressFastArgs*
}
#endif

#define HUF_4X1_DECODE_SYMBOL(_stream, _symbol) \
{ \
int const index = (int)(bits[(_stream)] >> 53); \
int const entry = (int)dtable[index]; \
bits[(_stream)] <<= (entry & 0x3F); \
op[(_stream)][(_symbol)] = (BYTE)((entry >> 8) & 0xFF); \
}

#define HUF_4X1_RELOAD_STREAM(_stream) \
{ \
int const ctz = ZSTD_countTrailingZeros64(bits[(_stream)]); \
int const nbBits = ctz & 7; \
int const nbBytes = ctz >> 3; \
op[(_stream)] += 5; \
ip[(_stream)] -= nbBytes; \
bits[(_stream)] = MEM_read64(ip[(_stream)]) | 1; \
bits[(_stream)] <<= nbBits; \
}

/* Manually unroll the loop because compilers don't consistently
* unroll the inner loops, which destroys performance.
*/
do {
/* Decode 5 symbols in each of the 4 streams */
for (symbol = 0; symbol < 5; ++symbol) {
for (stream = 0; stream < 4; ++stream) {
int const index = (int)(bits[stream] >> 53);
int const entry = (int)dtable[index];
bits[stream] <<= (entry & 63);
op[stream][symbol] = (BYTE)((entry >> 8) & 0xFF);
}
}
/* Reload the bitstreams */
for (stream = 0; stream < 4; ++stream) {
int const ctz = ZSTD_countTrailingZeros64(bits[stream]);
int const nbBits = ctz & 7;
int const nbBytes = ctz >> 3;
op[stream] += 5;
ip[stream] -= nbBytes;
bits[stream] = MEM_read64(ip[stream]) | 1;
bits[stream] <<= nbBits;
}
HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X1_DECODE_SYMBOL, 0)
HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X1_DECODE_SYMBOL, 1)
HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X1_DECODE_SYMBOL, 2)
HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X1_DECODE_SYMBOL, 3)
HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X1_DECODE_SYMBOL, 4)

/* Reload each of the 4 the bitstreams */
HUF_4X_FOR_EACH_STREAM(HUF_4X1_RELOAD_STREAM)
} while (op[3] < olimit);

#undef HUF_4X1_DECODE_SYMBOL
#undef HUF_4X1_RELOAD_STREAM
}

_out:
Expand Down Expand Up @@ -869,7 +907,7 @@ static size_t HUF_decompress4X1_usingDTable_internal(void* dst, size_t dstSize,
}
#endif

if (!(flags & HUF_flags_disableFast)) {
if (HUF_ENABLE_FAST_DECODE && !(flags & HUF_flags_disableFast)) {
size_t const ret = HUF_decompress4X1_usingDTable_internal_fast(dst, dstSize, cSrc, cSrcSize, DTable, loopFn);
if (ret != 0)
return ret;
Expand Down Expand Up @@ -1492,7 +1530,6 @@ void HUF_decompress4X2_usingDTable_internal_fast_c_loop(HUF_DecompressFastArgs*
for (;;) {
BYTE* olimit;
int stream;
int symbol;

/* Assert loop preconditions */
#ifndef NDEBUG
Expand Down Expand Up @@ -1549,54 +1586,56 @@ void HUF_decompress4X2_usingDTable_internal_fast_c_loop(HUF_DecompressFastArgs*
}
#endif

#define HUF_4X2_DECODE_SYMBOL(_stream, _decode3) \
if ((_decode3) || (_stream) != 3) { \
int const index = (int)(bits[(_stream)] >> 53); \
HUF_DEltX2 const entry = dtable[index]; \
MEM_write16(op[(_stream)], entry.sequence); \
bits[(_stream)] <<= (entry.nbBits) & 0x3F; \
op[(_stream)] += (entry.length); \
}

#define HUF_4X2_RELOAD_STREAM(_stream) \
{ \
HUF_4X2_DECODE_SYMBOL(3, 1) \
{ \
int const ctz = ZSTD_countTrailingZeros64(bits[(_stream)]); \
int const nbBits = ctz & 7; \
int const nbBytes = ctz >> 3; \
ip[(_stream)] -= nbBytes; \
bits[(_stream)] = MEM_read64(ip[(_stream)]) | 1; \
bits[(_stream)] <<= nbBits; \
} \
}

/* Manually unroll the loop because compilers don't consistently
* unroll the inner loops, which destroys performance.
*/
do {
/* Do 5 table lookups for each of the first 3 streams */
for (symbol = 0; symbol < 5; ++symbol) {
for (stream = 0; stream < 3; ++stream) {
int const index = (int)(bits[stream] >> 53);
HUF_DEltX2 const entry = dtable[index];
MEM_write16(op[stream], entry.sequence);
bits[stream] <<= (entry.nbBits);
op[stream] += (entry.length);
}
}
/* Do 1 table lookup from the final stream */
{
int const index = (int)(bits[3] >> 53);
HUF_DEltX2 const entry = dtable[index];
MEM_write16(op[3], entry.sequence);
bits[3] <<= (entry.nbBits);
op[3] += (entry.length);
}
/* Do 4 table lookups from the final stream & reload bitstreams */
for (stream = 0; stream < 4; ++stream) {
/* Do a table lookup from the final stream.
* This is interleaved with the reloading to reduce register
* pressure. This shouldn't be necessary, but compilers can
* struggle with codegen with high register pressure.
*/
{
int const index = (int)(bits[3] >> 53);
HUF_DEltX2 const entry = dtable[index];
MEM_write16(op[3], entry.sequence);
bits[3] <<= (entry.nbBits);
op[3] += (entry.length);
}
/* Reload the bistreams. The final bitstream must be reloaded
* after the 5th symbol was decoded.
*/
{
int const ctz = ZSTD_countTrailingZeros64(bits[stream]);
int const nbBits = ctz & 7;
int const nbBytes = ctz >> 3;
ip[stream] -= nbBytes;
bits[stream] = MEM_read64(ip[stream]) | 1;
bits[stream] <<= nbBits;
}
}
/* Decode 5 symbols from each of the first 3 streams.
* The final stream will be decoded during the reload phase
* to reduce register pressure.
*/
HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X2_DECODE_SYMBOL, 0)
HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X2_DECODE_SYMBOL, 0)
HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X2_DECODE_SYMBOL, 0)
HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X2_DECODE_SYMBOL, 0)
HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X2_DECODE_SYMBOL, 0)

/* Decode one symbol from the final stream */
HUF_4X2_DECODE_SYMBOL(3, 1)

/* Decode 4 symbols from the final stream & reload bitstreams.
* The final stream is reloaded last, meaning that all 5 symbols
* are decoded from the final stream before it is reloaded.
*/
HUF_4X_FOR_EACH_STREAM(HUF_4X2_RELOAD_STREAM)
} while (op[3] < olimit);
}

#undef HUF_4X2_DECODE_SYMBOL
#undef HUF_4X2_RELOAD_STREAM

_out:

/* Save the final values of each of the state variables back to args. */
Expand Down Expand Up @@ -1681,7 +1720,7 @@ static size_t HUF_decompress4X2_usingDTable_internal(void* dst, size_t dstSize,
}
#endif

if (!(flags & HUF_flags_disableFast)) {
if (HUF_ENABLE_FAST_DECODE && !(flags & HUF_flags_disableFast)) {
size_t const ret = HUF_decompress4X2_usingDTable_internal_fast(dst, dstSize, cSrc, cSrcSize, DTable, loopFn);
if (ret != 0)
return ret;
Expand Down

0 comments on commit c7269ad

Please sign in to comment.