fabi/Snappy
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge branch 'google:main' into add_rvv_support

Browse Source
This commit is contained in:
anthony-zy
2025-09-05 16:02:38 +08:00
parent 27d8915d34 6f99459b5b
commit c860cabbc5
4 changed files with 79 additions and 103 deletions
Download Patch File Download Diff File Expand all files Collapse all files

83
snappy.cc
View File

@@ -282,18 +282,6 @@ inline char* IncrementalCopySlow(const char* src, char* op,
// calling MakePatternMaskBytes(0, 6, index_sequence<16>()) and
// MakePatternMaskBytes(16, 6, index_sequence<16>()) respectively.
// Selects the appropriate vector size based on the current architecture
// vuint8m1_t, RISC-V vector type with fixed 128-bit size
// (sizeof not used due to variable-length vector register in RVV)
#if defined(__SSE2__) || defined(SNAPPY_HAVE_SSSE3)
constexpr size_t kVectorSize = sizeof(V128); // __m128i
#elif defined(__ARM_NEON) || defined(SNAPPY_HAVE_NEON)
constexpr size_t kVectorSize = sizeof(uint8x16_t); // uint8x16_t
#elif defined(SNAPPY_HAVE_RVV) || defined(__riscv_vector)
constexpr size_t kVectorSize = 16; // vuint8m1_t
#else
#error "Unsupported architecture. Please define __SSE2__, __ARM_NEON, or SNAPPY_HAVE_RVV/__riscv_vector."
#endif
template <size_t... indexes>
inline constexpr std::array<char, sizeof...(indexes)> MakePatternMaskBytes(
@@ -342,12 +330,6 @@ static inline V128 LoadPattern(const char* src, const size_t pattern_size) {
generation_mask);
}
// vuint8m1_t cannot be used as an element of std::pair
#if SNAPPY_HAVE_RVV
#define LoadPatternAndReshuffleMask(src, pattern_size) \
V128 pattern = LoadPattern(src, pattern_size);\
V128 reshuffle_mask = V128_Load(reinterpret_cast<const V128*>(\
pattern_reshuffle_masks[pattern_size - 1].data()));
#else
// Suppress -Wignored-attributes warning for __m128i in x86 SSE2 environment
// warning: ignoring attributes on template argument 'snappy::internal::V128' {aka '__vector(2) long long int'} [-Wignored-attributes]
@@ -355,7 +337,7 @@ static inline V128 LoadPattern(const char* src, const size_t pattern_size) {
#ifdef __SSE2__
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wignored-attributes"
#endif
SNAPPY_ATTRIBUTE_ALWAYS_INLINE
static inline std::pair<V128 /* pattern */, V128 /* reshuffle_mask */>
@@ -411,14 +393,10 @@ static inline bool Copy64BytesWithPatternExtension(char* dst, size_t offset) {
return true;
}
default: {
#if SNAPPY_HAVE_RVV
LoadPatternAndReshuffleMask(dst - offset, offset)
#else
auto pattern_and_reshuffle_mask =
LoadPatternAndReshuffleMask(dst - offset, offset);
V128 pattern = pattern_and_reshuffle_mask.first;
V128 reshuffle_mask = pattern_and_reshuffle_mask.second;
#endif
for (int i = 0; i < 4; i++) {
V128_StoreU(reinterpret_cast<V128*>(dst + 16 * i), pattern);
pattern = V128_Shuffle(pattern, reshuffle_mask);
@@ -526,14 +504,10 @@ inline char* IncrementalCopy(const char* src, char* op, char* const op_limit,
// Typically, the op_limit is the gating factor so try to simplify the loop
// based on that.
if (SNAPPY_PREDICT_TRUE(op_limit <= buf_limit - 15)) {
#if SNAPPY_HAVE_RVV
LoadPatternAndReshuffleMask(src, pattern_size);
#else
auto pattern_and_reshuffle_mask =
LoadPatternAndReshuffleMask(src, pattern_size);
V128 pattern = pattern_and_reshuffle_mask.first;
V128 reshuffle_mask = pattern_and_reshuffle_mask.second;
#endif
// There is at least one, and at most four 16-byte blocks. Writing four
// conditionals instead of a loop allows FDO to layout the code with
// respect to the actual probabilities of each length.
@@ -556,14 +530,10 @@ inline char* IncrementalCopy(const char* src, char* op, char* const op_limit,
}
char* const op_end = buf_limit - 15;
if (SNAPPY_PREDICT_TRUE(op < op_end)) {
#if SNAPPY_HAVE_RVV
LoadPatternAndReshuffleMask(src, pattern_size);
#else
auto pattern_and_reshuffle_mask =
LoadPatternAndReshuffleMask(src, pattern_size);
V128 pattern = pattern_and_reshuffle_mask.first;
V128 reshuffle_mask = pattern_and_reshuffle_mask.second;
#endif
// This code path is relatively cold however so we save code size
// by avoiding unrolling and vectorizing.
//
@@ -1288,36 +1258,27 @@ void MemCopy64(char* dst, const void* src, size_t size) {
data = _mm256_lddqu_si256(static_cast<const __m256i *>(src) + 1);
_mm256_storeu_si256(reinterpret_cast<__m256i *>(dst) + 1, data);
}
// RVV acceleration available on RISC-V when compiled with -march=rv64gcv
#elif defined(__riscv) & SNAPPY_HAVE_RVV
uint8_t* dst_u8 = (uint8_t*)dst;
const uint8_t* src_u8 = (const uint8_t*)src;
//overlap bwd copy
if (src_u8 < dst_u8 && dst_u8 < src_u8 + size) {
size_t offset = size;
while (offset > 0) {
size_t vl = VSETVL_E8M1(offset);
offset -= vl;
vuint8m1_t vec = VLE8_V_U8M1(src_u8 + offset, vl);
VSE8_V_U8M1(dst_u8 + offset, vec, vl);
}
} else {
size_t vl = VSETVL_E8M1(size);
// if size >vl,use the max_vlen copy
if (vl < size) {
size_t offset = 0;
while (offset < size) {
vl = VSETVL_E8M1(size - offset);
vuint8m1_t vec = VLE8_V_U8M1(src_u8 + offset, vl);
VSE8_V_U8M1(dst_u8 + offset, vec, vl);
offset += vl;
}
} else {
// Copy the rest
vuint8m1_t vec = VLE8_V_U8M1(src_u8, vl);
VSE8_V_U8M1(dst_u8, vec, vl);
}
}
// RVV acceleration available on RISC-V when compiled with -march=rv64gcv
#elif defined(__riscv) && SNAPPY_HAVE_RVV
// Cast pointers to the type we will operate on.
unsigned char* dst_ptr = (unsigned char*)dst;
const unsigned char* src_ptr = (const unsigned char*)src;
size_t remaining_bytes = size;
//Loop as long as there are bytes remaining to be copied.
while (remaining_bytes > 0) {
//Set vector configuration: e8 (8-bit elements), m2 (LMUL=2).
//Use e8m2 configuration to maximize throughput.
size_t vl = VSETVL_E8M2(remaining_bytes);
//Load data from the current source pointer.
vuint8m2_t vec = VLE8_V_U8M2(src_ptr, vl);
//Store data to the current destination pointer.
VSE8_V_U8M2(dst_ptr, vec, vl);
//Update pointers and the remaining count.
src_ptr += vl;
dst_ptr += vl;
remaining_bytes -= vl;
}
#else
std::memmove(dst, src, kShortMemCopy);
//Profiling shows that nearly all copies are short.