Add PUFFS_USE_NO_OP_PERFORMANCE_HACKS.

name                             old speed      new speed      delta

puffs_adler32_10k/gcc            3.26GB/s ± 0%  3.26GB/s ± 0%     ~     (p=0.222 n=5+5)
puffs_adler32_100k/gcc           3.27GB/s ± 0%  3.27GB/s ± 0%   -0.12%  (p=0.032 n=5+5)
puffs_flate_decode_1k/gcc         108MB/s ± 0%   108MB/s ± 0%     ~     (p=0.841 n=5+5)
puffs_flate_decode_10k/gcc        128MB/s ± 0%   128MB/s ± 0%     ~     (p=0.057 n=4+4)
puffs_flate_decode_100k/gcc       136MB/s ± 0%   136MB/s ± 0%     ~     (p=0.222 n=5+5)
puffs_zlib_decode_10k/gcc         123MB/s ± 0%   123MB/s ± 0%     ~     (p=0.730 n=5+4)
puffs_zlib_decode_100k/gcc        130MB/s ± 0%   131MB/s ± 0%     ~     (p=1.000 n=5+5)

puffs_gif_lzw_decode_20k/gcc      137MB/s ± 4%   140MB/s ± 3%     ~     (p=0.095 n=5+5)
puffs_gif_lzw_decode_100k/gcc     177MB/s ± 3%   187MB/s ± 2%   +5.75%  (p=0.008 n=5+5)
puffs_gif_decode_1k/gcc           348MB/s ± 0%   401MB/s ± 1%  +15.34%  (p=0.016 n=4+5)
puffs_gif_decode_10k/gcc          121MB/s ± 5%   154MB/s ± 1%  +27.90%  (p=0.008 n=5+5)
puffs_gif_decode_100k/gcc         106MB/s ± 4%   131MB/s ± 2%  +24.35%  (p=0.008 n=5+5)
puffs_gif_decode_1000k/gcc        109MB/s ± 2%   135MB/s ± 1%  +24.03%  (p=0.008 n=5+5)
This commit is contained in:
Nigel Tao committed 2017-11-02 14:51:29 +11:00
1 parent 4b572702f9
commit ed6ce6fd54
9 files changed
+232 -9

No files matched your search

+24
View File
@@ -48,6 +48,17 @@
// TODO: don't hard code this in base-header.h.
#define PUFFS_VERSION (0x00001)
// PUFFS_USE_NO_OP_PERFORMANCE_HACKS enables code paths that look like
// redundant no-ops, but for reasons to be investigated, can have dramatic
// performance effects with gcc 4.8.4 (e.g. 1.2x on some benchmarks).
//
// TODO: investigate; delete these hacks (without regressing performance).
// The order matters here. Clang also defines "__GNUC__".
#if defined(__clang__)
#elif defined(__GNUC__)
#define PUFFS_USE_NO_OP_PERFORMANCE_HACKS 1
#endif
// puffs_base__slice_u8 is a 1-dimensional buffer (a pointer and length).
//
// A value with all fields NULL or zero is a valid, empty slice.
@@ -68,6 +79,13 @@ typedef struct {
bool closed; // No further writes are expected.
} puffs_base__buf1;
#ifdef PUFFS_USE_NO_OP_PERFORMANCE_HACKS
typedef struct {
void* always_null0;
void* always_null1;
} puffs_base__paired_nulls;
#endif
typedef struct {
// TODO: move buf into private_impl? As it is, it looks like users can modify
// the buf field to point to a different buffer, which can turn the limit and
@@ -78,6 +96,12 @@ typedef struct {
struct {
uint8_t* limit;
uint8_t* mark;
#ifdef PUFFS_USE_NO_OP_PERFORMANCE_HACKS
struct {
puffs_base__paired_nulls* noph0;
uint32_t noph1;
} * no_op_performance_hacks;
#endif
} private_impl;
} puffs_base__reader1;
+42
View File
@@ -48,6 +48,17 @@
// TODO: don't hard code this in base-header.h.
#define PUFFS_VERSION (0x00001)
// PUFFS_USE_NO_OP_PERFORMANCE_HACKS enables code paths that look like
// redundant no-ops, but for reasons to be investigated, can have dramatic
// performance effects with gcc 4.8.4 (e.g. 1.2x on some benchmarks).
//
// TODO: investigate; delete these hacks (without regressing performance).
// The order matters here. Clang also defines "__GNUC__".
#if defined(__clang__)
#elif defined(__GNUC__)
#define PUFFS_USE_NO_OP_PERFORMANCE_HACKS 1
#endif
// puffs_base__slice_u8 is a 1-dimensional buffer (a pointer and length).
//
// A value with all fields NULL or zero is a valid, empty slice.
@@ -68,6 +79,13 @@ typedef struct {
bool closed; // No further writes are expected.
} puffs_base__buf1;
#ifdef PUFFS_USE_NO_OP_PERFORMANCE_HACKS
typedef struct {
void* always_null0;
void* always_null1;
} puffs_base__paired_nulls;
#endif
typedef struct {
// TODO: move buf into private_impl? As it is, it looks like users can modify
// the buf field to point to a different buffer, which can turn the limit and
@@ -78,6 +96,12 @@ typedef struct {
struct {
uint8_t* limit;
uint8_t* mark;
#ifdef PUFFS_USE_NO_OP_PERFORMANCE_HACKS
struct {
puffs_base__paired_nulls* noph0;
uint32_t noph1;
} * no_op_performance_hacks;
#endif
} private_impl;
} puffs_base__reader1;
@@ -1221,6 +1245,16 @@ puffs_gif__status puffs_gif__decoder__decode_id(puffs_gif__decoder* self,
if (a_src.buf) {
size_t n = b_rptr_src - (a_src.buf->ptr + a_src.buf->ri);
a_src.buf->ri += n;
#ifdef PUFFS_USE_NO_OP_PERFORMANCE_HACKS
if (a_src.private_impl.no_op_performance_hacks) {
puffs_base__paired_nulls* pn =
a_src.private_impl.no_op_performance_hacks->noph0;
while (pn && pn->always_null0) {
*((int*)(pn->always_null0)) = 0;
pn = (puffs_base__paired_nulls*)(pn->always_null1);
}
}
#endif
}
puffs_gif__status t_3 = puffs_gif__lzw_decoder__decode(
&self->private_impl.f_lzw, a_dst, v_r);
@@ -1364,6 +1398,14 @@ puffs_gif__status puffs_gif__lzw_decoder__decode(puffs_gif__lzw_decoder* self,
if (a_src.buf) {
b_rptr_src = a_src.buf->ptr + a_src.buf->ri;
uint64_t len = a_src.buf->wi - a_src.buf->ri;
#ifdef PUFFS_USE_NO_OP_PERFORMANCE_HACKS
if (a_src.private_impl.no_op_performance_hacks) {
if (a_src.private_impl.no_op_performance_hacks->noph1 &&
(len > a_src.private_impl.no_op_performance_hacks->noph1)) {
len = a_src.private_impl.no_op_performance_hacks->noph1;
}
}
#endif
b_rend_src = b_rptr_src + len;
if (a_src.private_impl.limit && (b_rend_src > a_src.private_impl.limit)) {
b_rend_src = a_src.private_impl.limit;
+24
View File
@@ -48,6 +48,17 @@
// TODO: don't hard code this in base-header.h.
#define PUFFS_VERSION (0x00001)
// PUFFS_USE_NO_OP_PERFORMANCE_HACKS enables code paths that look like
// redundant no-ops, but for reasons to be investigated, can have dramatic
// performance effects with gcc 4.8.4 (e.g. 1.2x on some benchmarks).
//
// TODO: investigate; delete these hacks (without regressing performance).
// The order matters here. Clang also defines "__GNUC__".
#if defined(__clang__)
#elif defined(__GNUC__)
#define PUFFS_USE_NO_OP_PERFORMANCE_HACKS 1
#endif
// puffs_base__slice_u8 is a 1-dimensional buffer (a pointer and length).
//
// A value with all fields NULL or zero is a valid, empty slice.
@@ -68,6 +79,13 @@ typedef struct {
bool closed; // No further writes are expected.
} puffs_base__buf1;
#ifdef PUFFS_USE_NO_OP_PERFORMANCE_HACKS
typedef struct {
void* always_null0;
void* always_null1;
} puffs_base__paired_nulls;
#endif
typedef struct {
// TODO: move buf into private_impl? As it is, it looks like users can modify
// the buf field to point to a different buffer, which can turn the limit and
@@ -78,6 +96,12 @@ typedef struct {
struct {
uint8_t* limit;
uint8_t* mark;
#ifdef PUFFS_USE_NO_OP_PERFORMANCE_HACKS
struct {
puffs_base__paired_nulls* noph0;
uint32_t noph1;
} * no_op_performance_hacks;
#endif
} private_impl;
} puffs_base__reader1;
+24
View File
@@ -48,6 +48,17 @@
// TODO: don't hard code this in base-header.h.
#define PUFFS_VERSION (0x00001)
// PUFFS_USE_NO_OP_PERFORMANCE_HACKS enables code paths that look like
// redundant no-ops, but for reasons to be investigated, can have dramatic
// performance effects with gcc 4.8.4 (e.g. 1.2x on some benchmarks).
//
// TODO: investigate; delete these hacks (without regressing performance).
// The order matters here. Clang also defines "__GNUC__".
#if defined(__clang__)
#elif defined(__GNUC__)
#define PUFFS_USE_NO_OP_PERFORMANCE_HACKS 1
#endif
// puffs_base__slice_u8 is a 1-dimensional buffer (a pointer and length).
//
// A value with all fields NULL or zero is a valid, empty slice.
@@ -68,6 +79,13 @@ typedef struct {
bool closed; // No further writes are expected.
} puffs_base__buf1;
#ifdef PUFFS_USE_NO_OP_PERFORMANCE_HACKS
typedef struct {
void* always_null0;
void* always_null1;
} puffs_base__paired_nulls;
#endif
typedef struct {
// TODO: move buf into private_impl? As it is, it looks like users can modify
// the buf field to point to a different buffer, which can turn the limit and
@@ -78,6 +96,12 @@ typedef struct {
struct {
uint8_t* limit;
uint8_t* mark;
#ifdef PUFFS_USE_NO_OP_PERFORMANCE_HACKS
struct {
puffs_base__paired_nulls* noph0;
uint32_t noph1;
} * no_op_performance_hacks;
#endif
} private_impl;
} puffs_base__reader1;