mirrors/kitty
1
0
Fork 0
mirror of https://github.com/kovidgoyal/kitty synced 2026-07-25 01:38:02 +02:00
Code Issues Packages Projects Releases Wiki Activity

Use unaligned stores

Browse Source 
Makes no measurable difference in the benchmark. And will eventually
allow us to process larger chunks of data without need to reset a bunch
of vector registers to constant values each time.
This commit is contained in:
Kovid Goyal
2024-02-01 11:29:59 +05:30
parent 77cfd44f24
commit d987ffe49a
1 changed files with 11 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
kitty/simd-string-impl.h
View File

@@ -52,7 +52,7 @@ END_IGNORE_DIAGNOSTIC
#define add_epi8 simde_mm_add_epi8 #define add_epi8 simde_mm_add_epi8
#define load_unaligned simde_mm_loadu_si128 #define load_unaligned simde_mm_loadu_si128
#define load_aligned simde_mm_load_si128 #define load_aligned simde_mm_load_si128
#define store_aligned simde_mm_store_si128 #define store_unaligned simde_mm_storeu_si128
#define cmpeq_epi8 simde_mm_cmpeq_epi8 #define cmpeq_epi8 simde_mm_cmpeq_epi8
#define cmplt_epi8 simde_mm_cmplt_epi8 #define cmplt_epi8 simde_mm_cmplt_epi8
#define cmpgt_epi8 simde_mm_cmpgt_epi8 #define cmpgt_epi8 simde_mm_cmpgt_epi8
@@ -93,7 +93,7 @@ FUNC(is_zero)(const integer_t a) { return simde_mm_testz_si128(a, a); }
#define add_epi8 simde_mm256_add_epi8 #define add_epi8 simde_mm256_add_epi8
#define load_unaligned simde_mm256_loadu_si256 #define load_unaligned simde_mm256_loadu_si256
#define load_aligned simde_mm256_load_si256 #define load_aligned simde_mm256_load_si256
#define store_aligned simde_mm256_store_si256 #define store_unaligned simde_mm256_storeu_si256
#define cmpeq_epi8 simde_mm256_cmpeq_epi8 #define cmpeq_epi8 simde_mm256_cmpeq_epi8
#define cmpgt_epi8 simde_mm256_cmpgt_epi8 #define cmpgt_epi8 simde_mm256_cmpgt_epi8
#define cmplt_epi8(a, b) cmpgt_epi8(b, a) #define cmplt_epi8(a, b) cmpgt_epi8(b, a)
@@ -190,7 +190,7 @@ static inline integer_t shuffle_impl256(const integer_t value, const integer_t s
#define print_register_as_bytes(r) { \ #define print_register_as_bytes(r) { \
printf("%s:\n", #r); \ printf("%s:\n", #r); \
alignas(64) uint8_t data[sizeof(r)]; \ alignas(64) uint8_t data[sizeof(r)]; \
store_aligned((integer_t*)data, r); \ store_unaligned((integer_t*)data, r); \
for (unsigned i = 0; i < sizeof(integer_t); i++) { \ for (unsigned i = 0; i < sizeof(integer_t); i++) { \
uint8_t ch = data[i]; \ uint8_t ch = data[i]; \
if (' ' <= ch && ch < 0x7f) printf("_%c ", ch); else printf("%.2x ", ch); \ if (' ' <= ch && ch < 0x7f) printf("_%c ", ch); else printf("%.2x ", ch); \
@@ -289,26 +289,26 @@ FUNC(output_plain_ascii)(UTF8Decoder *d, integer_t vec, size_t src_sz) {
#if KITTY_SIMD_LEVEL == 128 #if KITTY_SIMD_LEVEL == 128
for (const uint32_t *limit = d->output + src_sz, *p = d->output; p < limit; p += output_increment) { for (const uint32_t *limit = d->output + src_sz, *p = d->output; p < limit; p += output_increment) {
const integer_t unpacked = extract_lower_quarter_as_chars(vec); const integer_t unpacked = extract_lower_quarter_as_chars(vec);
store_aligned((integer_t*)p, unpacked); store_unaligned((integer_t*)p, unpacked);
vec = shift_right_by_bytes128(vec, output_increment); vec = shift_right_by_bytes128(vec, output_increment);
} }
#else #else
const uint32_t *limit = d->output + src_sz, *p = d->output; const uint32_t *limit = d->output + src_sz, *p = d->output;
simde__m128i x = simde_mm256_extracti128_si256(vec, 0); simde__m128i x = simde_mm256_extracti128_si256(vec, 0);
integer_t unpacked = extract_lower_half_as_chars(x); integer_t unpacked = extract_lower_half_as_chars(x);
store_aligned((integer_t*)p, unpacked); p += output_increment; store_unaligned((integer_t*)p, unpacked); p += output_increment;
if (p < limit) { if (p < limit) {
x = shift_right_by_bytes128(x, output_increment); x = shift_right_by_bytes128(x, output_increment);
unpacked = extract_lower_half_as_chars(x); unpacked = extract_lower_half_as_chars(x);
store_aligned((integer_t*)p, unpacked); p += output_increment; store_unaligned((integer_t*)p, unpacked); p += output_increment;
if (p < limit) { if (p < limit) {
x = simde_mm256_extracti128_si256(vec, 1); x = simde_mm256_extracti128_si256(vec, 1);
unpacked = extract_lower_half_as_chars(x); unpacked = extract_lower_half_as_chars(x);
store_aligned((integer_t*)p, unpacked); p += output_increment; store_unaligned((integer_t*)p, unpacked); p += output_increment;
if (p < limit) { if (p < limit) {
x = shift_right_by_bytes128(x, output_increment); x = shift_right_by_bytes128(x, output_increment);
unpacked = extract_lower_half_as_chars(x); unpacked = extract_lower_half_as_chars(x);
store_aligned((integer_t*)p, unpacked); p += output_increment; store_unaligned((integer_t*)p, unpacked); p += output_increment;
} }
} }
} }
@@ -324,7 +324,7 @@ FUNC(output_unicode)(UTF8Decoder *d, integer_t output1, integer_t output2, integ
const integer_t unpacked2 = shift_right_by_one_byte(extract_lower_quarter_as_chars(output2)); const integer_t unpacked2 = shift_right_by_one_byte(extract_lower_quarter_as_chars(output2));
const integer_t unpacked3 = shift_right_by_two_bytes(extract_lower_quarter_as_chars(output3)); const integer_t unpacked3 = shift_right_by_two_bytes(extract_lower_quarter_as_chars(output3));
const integer_t unpacked = or_si(or_si(unpacked1, unpacked2), unpacked3); const integer_t unpacked = or_si(or_si(unpacked1, unpacked2), unpacked3);
store_aligned((integer_t*)p, unpacked); store_unaligned((integer_t*)p, unpacked);
output1 = shift_right_by_bytes128(output1, output_increment); output1 = shift_right_by_bytes128(output1, output_increment);
output2 = shift_right_by_bytes128(output2, output_increment); output2 = shift_right_by_bytes128(output2, output_increment);
output3 = shift_right_by_bytes128(output3, output_increment); output3 = shift_right_by_bytes128(output3, output_increment);
@@ -337,7 +337,7 @@ FUNC(output_unicode)(UTF8Decoder *d, integer_t output1, integer_t output2, integ
const integer_t unpacked1 = extract_lower_half_as_chars(x1); \ const integer_t unpacked1 = extract_lower_half_as_chars(x1); \
const integer_t unpacked2 = shift_right_by_one_byte(extract_lower_half_as_chars(x2)); \ const integer_t unpacked2 = shift_right_by_one_byte(extract_lower_half_as_chars(x2)); \
const integer_t unpacked3 = shift_right_by_two_bytes(extract_lower_half_as_chars(x3)); \ const integer_t unpacked3 = shift_right_by_two_bytes(extract_lower_half_as_chars(x3)); \
store_aligned((integer_t*)p, or_si(or_si(unpacked1, unpacked2), unpacked3)); \ store_unaligned((integer_t*)p, or_si(or_si(unpacked1, unpacked2), unpacked3)); \
p += output_increment; \ p += output_increment; \
} }
#define extract(which) x1 = simde_mm256_extracti128_si256(output1, which); x2 = simde_mm256_extracti128_si256(output2, which); x3 = simde_mm256_extracti128_si256(output3, which); #define extract(which) x1 = simde_mm256_extracti128_si256(output1, which); x2 = simde_mm256_extracti128_si256(output2, which); x3 = simde_mm256_extracti128_si256(output3, which);
@@ -601,7 +601,7 @@ invalid_utf8:
#undef set_epi8 #undef set_epi8
#undef load_unaligned #undef load_unaligned
#undef load_aligned #undef load_aligned
#undef store_aligned #undef store_unaligned
#undef cmpeq_epi8 #undef cmpeq_epi8
#undef cmplt_epi8 #undef cmplt_epi8
#undef cmpgt_epi8 #undef cmpgt_epi8