Kernel/linux-next
1
0
Fork 0
mirror of https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git synced 2024-12-28 16:52:18 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

zstd: Backport Huffman speed improvement from upstream

Browse Source 
Backport upstream commit c7269ad [0] to improve zstd decoding speed.

Updating the kernel to zstd v1.5.5 earlier in this patch series
regressed zstd decoding speed. This turned out to be because gcc was not
unrolling the inner loops of the Huffman decoder which are executed a
constant number of times [1]. This really hurts performance, as we expect
this loop to be completely branch-free. This commit fixes the issue by
unrolling the loop manually [2].

The commit fixes one more minor issue, which is to mask a variable shift
by 0x3F. The shift was guaranteed to be less than 64, but gcc couldn't
prove that, and emitted suboptimal code.

Finally, the upstream commit added a build macro
`HUF_DISABLE_FAST_DECODE` which is not used in the kernel, but is
maintained to keep a clean import from upstream.

This commit was generated from upstream signed tag v1.5.5-kernel [3] by:

  export ZSTD=/path/to/repo/zstd/
  export LINUX=/path/to/repo/linux/
  cd "$ZSTD/contrib/linux-kernel"
  git checkout v1.5.5-kernel
  make import LINUX="$LINUX"

I ran my benchmark & test suite before and after this commit to measure
the overall decompression speed benefit. It benchmarks zstd at several
compression levels. These benchmarks measure the total time it takes to
read data from the compressed filesystem.

Component,	Level,	Read time delta
Btrfs    ,	    1,	-7.0%
Btrfs    ,	    3,	-3.9%
Btrfs    ,	    5,	-4.7%
Btrfs    ,	    7,	-5.5%
Btrfs    ,	    9,	-2.4%
Squashfs ,	    1,	-9.1%

Link: c7269add7e
Link: https://gist.github.com/terrelln/2e14ff1fb197102a08d7823d8044978d
Link: https://gist.github.com/terrelln/a70bde22a2abc800691fb65c21eabc2a
Link: https://github.com/facebook/zstd/tree/v1.5.5-kernel
Signed-off-by: Nick Terrell <terrelln@fb.com>
This commit is contained in:
Nick Terrell 2023-11-20 12:39:49 -08:00
parent 98988fc8e9
commit 40eb0e915d
1 changed files with 105 additions and 66 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

171
lib/zstd/decompress/huf_decompress.c
View File

@ -35,6 +35,12 @@
* 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.
@ -289,6 +295,24 @@ 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
@ -702,7 +726,6 @@ void HUF_decompress4X1_usingDTable_internal_fast_c_loop(HUF_DecompressFastArgs*
for (;;) {
BYTE* olimit;
int stream;
int symbol;
/* Assert loop preconditions */
#ifndef NDEBUG
@ -749,27 +772,42 @@ 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:
@ -865,7 +903,7 @@ 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;
@ -1487,7 +1525,6 @@ void HUF_decompress4X2_usingDTable_internal_fast_c_loop(HUF_DecompressFastArgs*
for (;;) {
BYTE* olimit;
int stream;
int symbol;
/* Assert loop preconditions */
#ifndef NDEBUG
@ -1544,54 +1581,56 @@ 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. */
@ -1676,7 +1715,7 @@ 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;