Skip to content

Optimize oj dump cstr sse4 refactor #3

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
samyron merged 3 commits into from
Jun 26, 2025
Merged

Optimize oj dump cstr sse4 refactor #3

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
97a60b2
Refactor the code to reduce duplication.
samyron Jun 26, 2025
f9c3768
SSE42 bugfixes.
samyron Jun 26, 2025
8f372a2
Formatting.
samyron Jun 26, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
94 changes: 47 additions & 47 deletions ext/oj/dump.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -201,11 +201,10 @@ void initialize_neon(void) {
}
#endif

#ifdef HAVE_SIMD_X86_64
#ifdef OJ_USE_SSE4_2

static size_t (*hibit_friendly_size_simd)(const uint8_t *str, size_t len) = NULL;
static __m128i hibit_friendly_chars_sse42[8];
static SIMD_Implementation simd_impl = SIMD_X86_64_NONE;

#define SIMD_TARGET __attribute__((target("sse4.2,ssse3")))

Expand All @@ -222,28 +221,29 @@ inline static SIMD_TARGET size_t hibit_friendly_size_sse42(const uint8_t *str, s
for (; i + sizeof(__m128i) <= len; i += sizeof(__m128i), str += sizeof(__m128i)) {
size += sizeof(__m128i);

__m128i chunk = _mm_loadu_si128((__m128i*)str);
__m128i tmp = vector_lookup_sse42(chunk, hibit_friendly_chars_sse42, 8);
size += _mm_sum_epu8(tmp);
__m128i chunk = _mm_loadu_si128((__m128i *)str);
__m128i tmp = vector_lookup_sse42(chunk, hibit_friendly_chars_sse42, 8);
size += _mm_sum_epu8(tmp);
}
size_t total = size + calculate_string_size(str, len - i, hibit_friendly_chars);
return total;
}

void SIMD_TARGET initialize_sse42(void) {
hibit_friendly_size_simd = hibit_friendly_size_sse42;
simd_impl = SIMD_X86_64_SSE42;

for (int i = 0; i < 8; i++) {
hibit_friendly_chars_sse42[i] = _mm_sub_epi8(_mm_loadu_si128((__m128i*)(hibit_friendly_chars + i * sizeof(__m128i))), _mm_set1_epi8('1'));
hibit_friendly_chars_sse42[i] = _mm_sub_epi8(
_mm_loadu_si128((__m128i *)(hibit_friendly_chars + i * sizeof(__m128i))),
_mm_set1_epi8('1'));
}
}

#else
#else

#define SIMD_TARGET
#define SIMD_TARGET

#endif /* HAVE_SIMD_X86_64 */
#endif /* OJ_USE_SSE4_2 */

inline static size_t hibit_friendly_size(const uint8_t *str, size_t len) {
#ifdef HAVE_SIMD_NEON
Expand All @@ -264,7 +264,7 @@ inline static size_t hibit_friendly_size(const uint8_t *str, size_t len) {

size_t total = size + calculate_string_size(str, len - i, hibit_friendly_chars);
return total;
#elif defined(HAVE_SIMD_X86_64)
#elif defined(OJ_USE_SSE4_2)
if (len >= sizeof(__m128i)) {
if (hibit_friendly_size_simd != NULL) {
return hibit_friendly_size_simd(str, len);
Expand Down Expand Up @@ -995,13 +995,13 @@ neon_update(const char *str, uint8x16x4_t *cmap_neon, int neon_table_size, bool
return result;
}

#elif defined(HAVE_SIMD_X86_64)
#elif defined(OJ_USE_SSE4_2)
typedef struct _sse42_match_result {
__m128i actions;
bool needs_escape;
int escape_mask;
bool has_some_hibit;
bool do_unicode_validation;
__m128i actions;
bool needs_escape;
int escape_mask;
bool has_some_hibit;
bool do_unicode_validation;
} sse42_match_result;

static inline SIMD_TARGET sse42_match_result
Expand All @@ -1013,8 +1013,8 @@ sse42_update(const char *str, __m128i *cmap_sse42, int sse42_tab_size, bool do_u
__m128i needs_escape = _mm_xor_si128(_mm_cmpeq_epi8(actions, _mm_setzero_si128()), _mm_set1_epi8(0xFF));
result.actions = _mm_add_epi8(actions, _mm_set1_epi8('1'));

result.escape_mask = _mm_movemask_epi8(needs_escape);
result.needs_escape = result.escape_mask != 0;
result.escape_mask = _mm_movemask_epi8(needs_escape);
result.needs_escape = result.escape_mask != 0;
if (has_hi && do_unicode_validation) {
__m128i has_some_hibit = _mm_and_si128(chunk, _mm_set1_epi8(0x80));
result.has_some_hibit = _mm_movemask_epi8(has_some_hibit) != 0;
Expand Down Expand Up @@ -1102,8 +1102,8 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
#ifdef HAVE_SIMD_NEON
uint8x16x4_t *cmap_neon = NULL;
int neon_table_size;
#elif defined(HAVE_SIMD_X86_64)
__m128i *cmap_sse42 = NULL;
#elif defined(OJ_USE_SSE4_2)
__m128i *cmap_sse42 = NULL;
int sse42_tab_size;
#endif /* HAVE_SIMD_NEON */
const char *orig = str;
Expand Down Expand Up @@ -1173,9 +1173,9 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
#ifdef HAVE_SIMD_NEON
cmap_neon = hibit_friendly_chars_neon;
neon_table_size = 2;
#elif defined(HAVE_SIMD_X86_64)
cmap_sse42 = hibit_friendly_chars_sse42;
sse42_tab_size = 8;
#elif defined(OJ_USE_SSE4_2)
cmap_sse42 = hibit_friendly_chars_sse42;
sse42_tab_size = 8;
#endif /* HAVE_NEON_SIMD */
size = hibit_friendly_size((uint8_t *)str, cnt);
}
Expand Down Expand Up @@ -1203,32 +1203,29 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
if (is_sym) {
*out->cur++ = ':';
}
#ifdef HAVE_SIMD_NEON
const char *chunk_start;
const char *chunk_end;
const char *cursor = str;
bool use_neon = (cmap_neon != NULL && cnt >= (sizeof(uint8x16_t))) ? true : false;
char matches[16];

#if defined(HAVE_SIMD_NEON) || defined(OJ_USE_SSE4_2)

#define SEARCH_FLUSH \
if (str > cursor) { \
APPEND_CHARS(out->cur, cursor, str - cursor); \
cursor = str; \
}
#elif defined(HAVE_SIMD_X86_64)

const char *chunk_start;
const char *chunk_end;
const char *cursor = str;
bool use_simd = (simd_impl == SIMD_X86_64_SSE42 && cmap_sse42 != NULL && cnt >= (sizeof(__m128i))) ? true : false;
const char *cursor = str;
char matches[16];
#define SEARCH_FLUSH \
if (str > cursor) { \
APPEND_CHARS(out->cur, cursor, str - cursor); \
cursor = str; \
}
#endif /* HAVE_SIMD_NEON */
#endif /* HAVE_SIMD_NEON || OJ_USE_SSE4_2 */

#if defined(HAVE_SIMD_NEON)
bool use_simd = (cmap_neon != NULL && cnt >= (sizeof(uint8x16_t))) ? true : false;
#elif defined(OJ_USE_SSE4_2)
bool use_simd = (cmap_sse42 != NULL && cnt >= (sizeof(__m128i))) ? true : false;
#endif

#ifdef HAVE_SIMD_NEON
if (use_neon) {
if (use_simd) {
while (str < end) {
const char *chunk_ptr = NULL;
if (str + sizeof(uint8x16_t) <= end) {
Expand Down Expand Up @@ -1291,7 +1288,9 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
}
SEARCH_FLUSH;
}
#elif defined(HAVE_SIMD_X86_64)
#endif

#ifdef OJ_USE_SSE4_2
if (use_simd) {
while (str < end) {
const char *chunk_ptr = NULL;
Expand All @@ -1309,13 +1308,13 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
break;
}
sse42_match_result result = sse42_update(chunk_ptr,
cmap_sse42,
sse42_tab_size,
do_unicode_validation,
has_hi);
cmap_sse42,
sse42_tab_size,
do_unicode_validation,
has_hi);
if ((result.do_unicode_validation) || result.needs_escape) {
SEARCH_FLUSH;
_mm_storeu_si128((__m128i*)matches, result.actions);
_mm_storeu_si128((__m128i *)matches, result.actions);
while (str < chunk_end) {
long match_index = str - chunk_start;
str = process_character(matches[match_index],
Expand All @@ -1334,7 +1333,8 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
}
SEARCH_FLUSH;
}
#endif /* HAVE_SIMD_NEON */
#endif /* OJ_USE_SSE4_2 */

for (; str < end; str++) {
str = process_character(cmap[(uint8_t)*str], str, end, out, orig, do_unicode_validation, &check_start);
}
Expand Down
4 changes: 0 additions & 4 deletions ext/oj/extconf.rb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,10 +33,6 @@
have_func('rb_enc_interned_str')
have_func('rb_ext_ractor_safe', 'ruby.h')

# X86-64 SIMD Support
have_header('x86intrin.h')
have_type('__m128i', 'x86intrin.h')

dflags['OJ_DEBUG'] = true unless ENV['OJ_DEBUG'].nil?

if with_config('--with-sse42')
Expand Down
9 changes: 3 additions & 6 deletions ext/oj/oj.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2087,10 +2087,7 @@ void Init_oj(void) {
initialize_neon();
#endif /* HAVE_SIMD_NEON */

#ifdef HAVE_SIMD_X86_64
SIMD_Implementation simd_impl = find_simd_implemenation();
if (simd_impl == SIMD_X86_64_SSE42) {
initialize_sse42();
}
#endif
#ifdef OJ_USE_SSE4_2
initialize_sse42();
#endif /* OJ_USE_SSE4_2 */
}
42 changes: 12 additions & 30 deletions ext/oj/simd.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,55 +9,37 @@

#if defined(__x86_64__) && (defined(__GNUC__) || defined(__clang__))

#if defined(HAVE_X86INTRIN_H) && defined(HAVE_TYPE___M128I)
#define HAVE_SIMD_X86_64 1
#if defined(OJ_USE_SSE4_2)
#define SIMD_MINIMUM_THRESHOLD 6
#include <x86intrin.h>

typedef enum {
SIMD_X86_64_NONE,
SIMD_X86_64_SSE42
} SIMD_Implementation;
#include <nmmintrin.h>

extern void initialize_sse42(void);

static inline __attribute__((target("sse4.2,ssse3"))) __m128i vector_lookup_sse42(__m128i input, __m128i *lookup_table, int tab_size) {
static inline __attribute__((target("sse4.2,ssse3"))) __m128i vector_lookup_sse42(__m128i input,
__m128i *lookup_table,
int tab_size) {
// Extract high 4 bits to determine which 16-byte chunk (0-15)
__m128i hi_index = _mm_and_si128(_mm_srli_epi32(input, 4), _mm_set1_epi8(0x0F));

// Extract low 4 bits for index within the chunk (0-15)
__m128i low_index = _mm_and_si128(input, _mm_set1_epi8(0x0F));

// Perform lookups in all 16 tables
__m128i results[16];
for (int i = 0; i < tab_size; i++) {
results[i] = _mm_shuffle_epi8(lookup_table[i], low_index);
}

// Create masks for each chunk and blend results
__m128i final_result = _mm_setzero_si128();

for (int i = 0; i < tab_size; i++) {
__m128i mask = _mm_cmpeq_epi8(hi_index, _mm_set1_epi8(i));
__m128i mask = _mm_cmpeq_epi8(hi_index, _mm_set1_epi8(i));
__m128i masked_result = _mm_and_si128(mask, results[i]);
final_result = _mm_or_si128(final_result, masked_result);
final_result = _mm_or_si128(final_result, masked_result);
}

return final_result;
}

inline static SIMD_Implementation find_simd_implemenation(void) {
#ifdef __GNUC__
__builtin_cpu_init();
#endif

#if defined(__GNUC__) || defined(__clang__)
if (__builtin_cpu_supports("sse4.2") && __builtin_cpu_supports("ssse3")) {
return SIMD_X86_64_SSE42;
}
#endif

return SIMD_X86_64_NONE;
return final_result;
}

#endif /* defined(HAVE_X86INTRIN_H) && defined(HAVE_TYPE___M128I) */
Expand Down